Introduction to

Machine Learning
with Applications in
Information Security

Introduction to Machine Learning with Applications in Information Security,

Second Edition provides a classroom-tested introduction to a wide variety of
machine learning and deep learning algorithms and techniques, reinforced via
realistic applications. The book is accessible and doesn’t prove theorems, or dwell
on mathematical theory. The goal is to present topics at an intuitive level, with
just enough detail to clarify the underlying concepts.
The book covers core classic machine learning topics in depth, including Hidden
Markov Models (HMM), Support Vector Machines (SVM), and clustering. Additional
machine learning topics include k-Nearest Neighbor (k-NN), boosting, Random
Forests, and Linear Discriminant Analysis (LDA). The fundamental deep learn-
ing topics of backpropagation, Convolutional Neural Networks (CNN), Multilayer
Perceptrons (MLP), and Recurrent Neural Networks (RNN) are covered in depth.
A broad range of advanced deep learning architectures are also presented, includ-
ing Long Short-Term Memory (LSTM), Generative Adversarial Networks (GAN),
Extreme Learning Machines (ELM), Residual Networks (ResNet), Deep Belief
Networks (DBN), Bidirectional Encoder Representations from Transformers (BERT),
and Word2Vec. Finally, several cutting-edge deep learning topics are discussed,
including dropout regularization, attention, explainability, and adversarial attacks.
Most of the examples in the book are drawn from the field of information security,
with many of the machine learning and deep learning applications focused on mal-
ware. The applications presented serve to demystify the topics by illustrating the use
of various learning techniques in straightforward scenarios. Some of the exercises in
this book require programming, and elementary computing concepts are assumed
in a few of the application sections. However, anyone with a modest amount of com-
puting experience should have no trouble with this aspect of the book.
Instructor resources, including PowerPoint slides, lecture videos, and other
relevant material are provided on an accompanying website: http://www.cs.sjsu.
edu/~stamp/ML/.
Chapman & Hall/CRC Machine Learning &
Pattern Recognition

A First Course in Machine Learning

Simon Rogers, Mark Girolami

Statistical Reinforcement Learning: Modern Machine Learning Approaches

Masashi Sugiyama

Sparse Modeling: Theory, Algorithms, and Applications

Irina Rish, Genady Grabarnik

Computational Trust Models and Machine Learning

Xin Liu, Anwitaman Datta, Ee-Peng Lim

Regularization, Optimization, Kernels, and Support Vector Machines

Johan A.K. Suykens, Marco Signoretto, Andreas Argyriou

Machine Learning: An Algorithmic Perspective, Second Edition

Stephen Marsland

Bayesian Programming
Pierre Bessiere, Emmanuel Mazer, Juan Manuel Ahuactzin, Kamel Mekhnacha

Multilinear Subspace Learning: Dimensionality Reduction of

Multidimensional Data
Haiping Lu, Konstantinos N. Plataniotis, Anastasios Venetsanopoulos

Data Science and Machine Learning: Mathematical and Statistical Methods

Dirk P. Kroese, Zdravko Botev, Thomas Taimre, Radislav Vaisman

Deep Learning and Linguistic Representation

Shalom Lappin

Artificial Intelligence and Causal Inference

Momiao Xiong

Introduction to Machine Learning with Applications in Information

Security, Second Edition
Mark Stamp

Entropy Randomization in Machine Learning

Yuri S. Popkov, Alexey Yu. Popkov, Yuri A. Dubno

For more information on this series please visit: https://www.routledge.com/Chapman--

HallCRC-Machine-Learning--Pattern-Recognition/book-series/CRCMACLEAPAT
 Introduction to
Machine Learning
with Applications in
Information Security
Second Edition

Mark Stamp
Second edition published 2023
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742

and by CRC Press

4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

CRC Press is an imprint of Taylor & Francis Group, LLC

© 2023 Mark Stamp

First edition published by CRC Press 2017

Reasonable efforts have been made to publish reliable data and information, but the author and
publisher cannot assume responsibility for the validity of all materials or the consequences of their use.
The authors and publishers have attempted to trace the copyright holders of all material reproduced
in this publication and apologize to copyright holders if permission to publish in this form has not
been obtained. If any copyright material has not been acknowledged please write and let us know so
we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, access www.copyright.com
or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923,
978-750-8400. For works that are not available on CCC please contact mpkbookspermissions@tandf.
co.uk

Trademark notice: Product or corporate names may be trademarks or registered trademarks and are
used only for identification and explanation without intent to infringe.

ISBN: 978-1-032-20492-5 (hbk)

ISBN: 978-1-032-20717-9 (pbk)
ISBN: 978-1-003-26487-3 (ebk)

DOI: 10.1201/9781003264873

Typeset in Latin Modern font

by KnowledgeWorks Global Ltd.

Publisher’s note: This book has been prepared from camera-ready copy provided by the authors.

Access the Support Material: http://www.cs.sjsu.edu/~stamp/ML/

Contents

Preface xi

1 Introduction 1
1.1 Basic sampling concepts . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Population parameters . . . . . . . . . . . . . . . . . . 5
1.1.2 Descriptive statistics vs. inference about a population 6
1.1.3 Random sampling vs. probability sampling . . . . . . 6
1.2 Design-based vs. model-based approach . . . . . . . . . . . . . 7
1.3 Populations used in sampling experiments . . . . . . . . . . 8
1.3.1 Soil organic matter in Voorst, the Netherlands . . . . 10
1.3.2 Poppy fields in Kandahar, Afghanistan . . . . . . . . . . 11
1.3.3 Aboveground biomass in Eastern Amazonia, Brazil . . 12
1.3.4 Annual mean air temperature in Iberia . . . . . . . . . 13

I Probability sampling for estimating population

parameters 17
2 Introduction to probability sampling 19
2.1 Horvitz-Thompson estimator . . . . . . . . . . . . . . . . . . . 21
2.2 Hansen-Hurwitz estimator . . . . . . . . . . . . . . . . . . . 23
2.3 Using models in design-based approach . . . . . . . . . . . . 24

3 Simple random sampling 27

3.1 Estimation of population parameters . . . . . . . . . . . . . 30
3.1.1 Population proportion . . . . . . . . . . . . . . . . . . 34
3.1.2 Cumulative distribution function and quantiles . . . . 34
3.2 Sampling variance of estimator of population parameters . . 36
3.3 Confidence interval estimate . . . . . . . . . . . . . . . . . . 40
3.3.1 Confidence interval for a proportion . . . . . . . . . . 43
3.4 Simple random sampling of circular plots . . . . . . . . . . . 44
3.4.1 Sampling from a finite set of fixed circles . . . . . . . 44
3.4.2 Sampling from an infinite set of floating circles . . . . 45

4 Stratified simple random sampling 49

4.1 Estimation of population parameters . . . . . . . . . . . . . . 51
4.1.1 Population proportion, cumulative distribution
function, and quantiles . . . . . . . . . . . . . . . . . . 54

v
vi Contents

4.1.2 Why should we stratify? . . . . . . . . . . . . . . . . . 56

4.2 Confidence interval estimate . . . . . . . . . . . . . . . . . . 59
4.3 Allocation of sample size to strata . . . . . . . . . . . . . . . 60
4.4 Cum-root-f stratification . . . . . . . . . . . . . . . . . . . . 64
4.5 Stratification with multiple covariates . . . . . . . . . . . . . . 67
4.6 Geographical stratification . . . . . . . . . . . . . . . . . . . 69
4.7 Multiway stratification . . . . . . . . . . . . . . . . . . . . . 73
4.8 Multivariate stratification . . . . . . . . . . . . . . . . . . . . 73

5 Systematic random sampling 79

5.1 Estimation of population parameters . . . . . . . . . . . . . 83
5.2 Approximating the sampling variance of the estimator of the
mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6 Cluster random sampling 93

6.1 Estimation of population parameters . . . . . . . . . . . . . . 97
6.2 Clusters selected with probabilities proportional to size,
without replacement . . . . . . . . . . . . . . . . . . . . . . . . 101
6.3 Simple random sampling of clusters . . . . . . . . . . . . . . 103
6.4 Stratified cluster random sampling . . . . . . . . . . . . . . . 106

7 Two-stage cluster random sampling 109

7.1 Estimation of population parameters . . . . . . . . . . . . . 112
7.2 Primary sampling units selected without replacement . . . . 118
7.3 Simple random sampling of primary sampling units . . . . . 119
7.4 Stratified two-stage cluster random sampling . . . . . . . . . . 121

8 Sampling with probabilities proportional to size 123

8.1 Probability-proportional-to-size sampling with replacement . 125
8.2 Probability-proportional-to-size sampling without replacement 128
8.2.1 Systematic pps sampling without replacement . . . . . 128
8.2.2 The pivotal method . . . . . . . . . . . . . . . . . . . 130

9 Balanced and well-spread sampling 137

9.1 Balanced sampling . . . . . . . . . . . . . . . . . . . . . . . . . 137
9.1.1 Balanced sample vs. balanced sampling design . . . . 138
9.1.2 Unequal inclusion probabilities . . . . . . . . . . . . . 139
9.1.3 Stratified random sampling . . . . . . . . . . . . . . . 145
9.1.4 Multiway stratification . . . . . . . . . . . . . . . . . . 148
9.2 Well-spread sampling . . . . . . . . . . . . . . . . . . . . . . 152
9.2.1 Local pivotal method . . . . . . . . . . . . . . . . . . 152
9.2.2 Generalised random-tessellation stratified sampling . . 154
9.3 Balanced sampling with spreading . . . . . . . . . . . . . . . 158

10 Model-assisted estimation 161

10.1 Generalised regression estimator . . . . . . . . . . . . . . . . 162
Contents vii

10.1.1 Simple and multiple regression estimators . . . . . . . 164

10.1.2 Penalised least squares estimation . . . . . . . . . . . 172
10.1.3 Regression estimator with stratified simple random
sampling . . . . . . . . . . . . . . . . . . . . . . . . . 174
10.2 Ratio estimator . . . . . . . . . . . . . . . . . . . . . . . . . . 181
10.2.1 Ratio estimators with stratified simple random
sampling . . . . . . . . . . . . . . . . . . . . . . . . . 185
10.2.2 Poststratified estimator . . . . . . . . . . . . . . . . . . 187
10.3 Model-assisted estimation using machine learning techniques 190
10.3.1 Predicting with a regression tree . . . . . . . . . . . . 192
10.3.2 Predicting with a random forest . . . . . . . . . . . . 195
10.4 Big data and volunteer data . . . . . . . . . . . . . . . . . . 198

11 Two-phase random sampling 201

11.1 Two-phase random sampling for stratification . . . . . . . . . 203
11.2 Two-phase random sampling for regression . . . . . . . . . . 206

12 Computing the required sample size 211

12.1 Standard error . . . . . . . . . . . . . . . . . . . . . . . . . . 212
12.2 Length of confidence interval . . . . . . . . . . . . . . . . . . 213
12.2.1 Length of confidence interval for a proportion . . . . . 214
12.3 Statistical testing of hypothesis . . . . . . . . . . . . . . . . . 215
12.3.1 Sample size for testing a proportion . . . . . . . . . . 216
12.4 Accounting for design effect . . . . . . . . . . . . . . . . . . . 218
12.5 Bayesian sample size determination . . . . . . . . . . . . . . 219
12.5.1 Bayesian criteria for sample size computation . . . . . 220
12.5.2 Mixed Bayesian-likelihood approach . . . . . . . . . . 223
12.5.3 Estimation of population mean . . . . . . . . . . . . . 223
12.5.4 Estimation of a population proportion . . . . . . . . . 226

13 Model-based optimisation of probability sampling designs 231

13.1 Model-based optimisation of sampling design type and
sample size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
13.1.1 Analytical approach . . . . . . . . . . . . . . . . . . . 233
13.1.2 Geostatistical simulation approach . . . . . . . . . . . 239
13.1.3 Bayesian approach . . . . . . . . . . . . . . . . . . . . 245
13.2 Model-based optimisation of spatial strata . . . . . . . . . . 252

14 Sampling for estimating parameters of domains 263

14.1 Direct estimator for large domains . . . . . . . . . . . . . . . 264
14.2 Model-assisted estimators for small domains . . . . . . . . . . 267
14.2.1 Regression estimator . . . . . . . . . . . . . . . . . . . 268
14.2.2 Synthetic estimator . . . . . . . . . . . . . . . . . . . . 271
14.3 Model-based prediction . . . . . . . . . . . . . . . . . . . . . 273
14.3.1 Random intercept model . . . . . . . . . . . . . . . . . 274
14.3.2 Geostatistical model . . . . . . . . . . . . . . . . . . . 279
viii Contents

14.4 Supplemental probability sampling of small domains . . . . . 283

15 Repeated sample surveys for monitoring population

parameters 289
15.1 Space-time designs . . . . . . . . . . . . . . . . . . . . . . . . 289
15.2 Space-time population parameters . . . . . . . . . . . . . . . . 291
15.3 Design-based generalised least squares estimation of spatial
means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
15.3.1 Current mean . . . . . . . . . . . . . . . . . . . . . . . 296
15.3.2 Change of the spatial mean . . . . . . . . . . . . . . . . 297
15.3.3 Temporal trend of the spatial mean . . . . . . . . . . . 297
15.3.4 Space-time mean . . . . . . . . . . . . . . . . . . . . . 298
15.4 Case study: annual mean daily temperature in Iberia . . . . 299
15.4.1 Static-synchronous design . . . . . . . . . . . . . . . . 300
15.4.2 Independent synchronous design . . . . . . . . . . . . 302
15.4.3 Serially alternating design . . . . . . . . . . . . . . . . 304
15.4.4 Supplemented panel design . . . . . . . . . . . . . . . 305
15.4.5 Rotating panel design . . . . . . . . . . . . . . . . . . 308
15.4.6 Sampling experiment . . . . . . . . . . . . . . . . . . . 310
15.5 Space-time sampling with stratified random sampling in space 312

II Sampling for mapping 315

16 Introduction to sampling for mapping 317
16.1 When is probability sampling not required? . . . . . . . . . . . 317
16.2 Sampling for simultaneously mapping and estimating means 319
16.3 Broad overview of sampling designs for mapping . . . . . . . . 321

17 Regular grid and spatial coverage sampling 325

17.1 Regular grid sampling . . . . . . . . . . . . . . . . . . . . . . 325
17.2 Spatial coverage sampling . . . . . . . . . . . . . . . . . . . . . 327
17.3 Spatial infill sampling . . . . . . . . . . . . . . . . . . . . . . . 331

18 Covariate space coverage sampling 335

18.1 Covariate space infill sampling . . . . . . . . . . . . . . . . . 338
18.2 Performance of covariate space coverage sampling in random
forest prediction . . . . . . . . . . . . . . . . . . . . . . . . . 340

19 Conditioned Latin hypercube sampling 345

19.1 Conditioned Latin hypercube infill sampling . . . . . . . . . . 351
19.2 Performance of conditioned Latin hypercube sampling in
random forest prediction . . . . . . . . . . . . . . . . . . . . . 351

20 Spatial response surface sampling 357

20.1 Increasing the sample size . . . . . . . . . . . . . . . . . . . . 365
20.2 Stratified spatial response surface sampling . . . . . . . . . . . 367
Contents ix

20.3 Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371

21 Introduction to kriging 375

21.1 Ordinary kriging . . . . . . . . . . . . . . . . . . . . . . . . . 375
21.2 Block-kriging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
21.3 Kriging with an external drift . . . . . . . . . . . . . . . . . 383
21.4 Estimating the semivariogram . . . . . . . . . . . . . . . . . . 387
21.4.1 Method-of-moments . . . . . . . . . . . . . . . . . . . 388
21.4.2 Maximum likelihood . . . . . . . . . . . . . . . . . . . 390
21.5 Estimating the residual semivariogram . . . . . . . . . . . . . 391
21.5.1 Iterative method-of-moments . . . . . . . . . . . . . . 392
21.5.2 Restricted maximum likelihood . . . . . . . . . . . . . 394

22 Model-based optimisation of the grid spacing 397

22.1 Optimal grid spacing for ordinary kriging . . . . . . . . . . . . 397
22.2 Controlling the mean or a quantile of the ordinary kriging
variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
22.3 Optimal grid spacing for block-kriging . . . . . . . . . . . . . 403
22.4 Optimal grid spacing for kriging with an external drift . . . 404
22.5 Bayesian approach . . . . . . . . . . . . . . . . . . . . . . . . . 407

23 Model-based optimisation of the sampling pattern 413

23.1 Spatial simulated annealing . . . . . . . . . . . . . . . . . . . 413
23.2 Optimising the sampling pattern for ordinary kriging . . . . 415
23.3 Optimising the sampling pattern for kriging with an external
drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
23.4 Model-based infill sampling for ordinary kriging . . . . . . . 424
23.5 Model-based infill sampling for kriging with an external drift 425

24 Sampling for estimating the semivariogram 429

24.1 Nested sampling . . . . . . . . . . . . . . . . . . . . . . . . . 429
24.2 Independent sampling of pairs of points . . . . . . . . . . . . 436
24.3 Optimisation of sampling pattern for semivariogram estimation 441
24.3.1 Uncertainty about semivariogram parameters . . . . . 442
24.3.2 Uncertainty about the kriging variance . . . . . . . . . 445
24.4 Optimisation of sampling pattern for semivariogram estimation
and mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
24.5 A practical solution . . . . . . . . . . . . . . . . . . . . . . . 456

25 Sampling for validation of maps 459

25.1 Map quality indices . . . . . . . . . . . . . . . . . . . . . . . . 461
25.1.1 Estimation of map quality indices . . . . . . . . . . . 463
25.2 Real-world case study . . . . . . . . . . . . . . . . . . . . . . 464
25.2.1 Estimation of the population mean error and mean
squared error . . . . . . . . . . . . . . . . . . . . . . . 465
x Contents

25.2.2 Estimation of the standard error of the estimator of the

population mean error and mean squared error . . . . . 467
25.2.3 Estimation of model efficiency coefficient . . . . . . . . 469
25.2.4 Statistical testing of hypothesis about population ME
and MSE . . . . . . . . . . . . . . . . . . . . . . . . . 470

26 Design-based, model-based, and model-assisted approach for

sampling and inference 473
26.1 Two sources of randomness . . . . . . . . . . . . . . . . . . . 473
26.2 Identically and independently distributed . . . . . . . . . . . 476
26.3 Bias and variance . . . . . . . . . . . . . . . . . . . . . . . . 480
26.4 Effective sample size . . . . . . . . . . . . . . . . . . . . . . . 482
26.5 Exploiting spatial structure in design-based approach . . . . . 487
26.6 Model-assisted vs. model-dependent . . . . . . . . . . . . . . 488

A Answers to exercises 495

Bibliography 515

Index 529
Preface

Since the start of The R Series of Chapman & Hall/CRC in 2011, numerous
books have been published on the statistical analysis and modelling of data
using R. To date, no book has been published in this series on how these data
can best be collected. From my point of view this was an omission, as scientific
research often starts with data collection. If the data collection is part of the
project, it might be a good idea to start thinking right at project start instead
of after the data have been collected, to make a well-founded decision on how
many data are needed, and on the type of sampling design.
My experience as a statistical consultant is that many researchers pay insuffi-
cient attention to the method for data collection. Too many researchers start
thinking when the data are there. Often, I had to conclude that the way the
data were collected by fellow researchers was suboptimal, or even unsuitable
for their aim. I hope that this new book may help researchers, practitioners,
and students to implement proper sampling designs, tailored to their problems
at hand, so that valuable data are collected that can be used to answer the
research questions.
Over the past decades, numerous wall-to-wall data sets have been collected
by remote sensing devices such as satellites and drones. These remote sensing
images are valuable sources of information on the natural environment and
resources. The question may arise about how useful it still can be in this
big data era to collect data in the field at a restricted number of sampling
locations. Do we really need these data to estimate a population mean or total,
for instance of the aboveground biomass or carbon stocks in the soil, or to map
these study variables? In many cases the answer is that it is indeed still useful
to collect sample data on the study variable, because the remote sensing images
provide only proxies of the study variable. The variables derived from the
remote sensing images can be related to the study variable, but we still need
groundtruth data of the study variable to model this relation. By combining
the wall-to-wall data of covariates and the sample data of the groundtruth, we
can increase the accuracy of the survey result compared to using only one of
these data sources.
The handbook Sampling for Natural Resource Monitoring (SNRM) (de Gruijter
et al., 2006) presents an overview of sampling strategies for the survey of natural
resources at a given point in time, as well as for how these resources can be
monitored through repeated surveys. The book presented here can be seen

xi
xii Preface

as a follow-up on SNRM. In SNRM, spatial sampling designs for survey and

space-time designs for monitoring are described and illustrated with notional
and real-world examples. Estimators for global and local quantities in space
and in space-time, and for the variance of these estimators are presented.
However, neither the computer code for how a sample with a given design can
be selected, nor the code for how the estimates can be computed is presented
in SNRM. The publication at hand fills this gap.
This book describes and illustrates classical, basic sampling designs for a spatial
survey, as well as more recently developed, advanced sampling designs and
estimators. Part I of the book is about random sampling designs for estimating
a mean, total, or proportion of a population or of several subpopulations. Part
II focuses on sampling designs for mapping.
The computer code is written in the popular programming language R (R Core
Team, 2021). There are several good reasons for choosing R as a language.
First, it is open-source, giving users the right to view the source code and
modify it to their needs. Second, as a result of this open-source, numerous
add-on packages have been developed, and this number is still increasing.
Happily enough, also quite a few add-on packages have been published for
sampling design and analysis. All these add-on packages make the writing of
computer code much more simple. Even very advanced statistical methods
for sampling design and statistical analysis are now within the reach of many
scientists: only a few lines of R code are needed to do the work. A risk is
that the appliers of the packages may not fully understand the implemented
statistical method. This understanding is not needed to obtain a result. For
this reason, I decided not to jump to the add-on packages right after the theory,
but to follow a more gradual approach. First, I show in the simplest possible
R code how a sample can be selected with a given sampling design, and how
the population parameters can be estimated. After this, I point out how the
same result can be obtained with an add-on package.
The target group of this book is researchers and practitioners of sample surveys,
as well as students in environmental, ecological, agricultural science or any
other science in which knowledge about a population of interest is collected
through spatial sampling. I have added exercises to most chapters, making
this book suitable as a textbook for students. The answers to the exercises
can be found in the appendix of this book. Large parts of the book are self-
contained, requiring no prior knowledge of statistics. For the chapters in Part I
on more advanced sampling designs, such as balanced sampling, and advanced
estimators of population parameters (model-assisted estimators), knowledge
of matrix algebra and regression analysis is required. For the final chapters
of Part II, basic knowledge of geostatistics is required. This knowledge is also
needed for some chapters in Part I. For this reason, I have added a chapter
introducing the basics of geostatistics (Chapter 21).
Preface xiii

An online version of this book is available at https://dickbrus.github.io/

SpatialSamplingwithR/. You can also find the book updates there. Addition-
ally, the SpatialSamplingwithR1 GitHub repository contains the R scripts of
the exercises. This is also the place where you can report errata and comment
on text and R code.
Information about the version of R and the versions of the R packages used
to compile this book can be found in the file SessionInfo.rds on github. My
intention is to update the book as soon as a new version of an R package is
available on CRAN which requires an update of the R code.

Acknowledgments
In 2006 our handbook Sampling for Natural Resource Monitoring (SNRM) was
published (de Gruijter et al., 2006). Soon after this milestone, Jaap de Gruijter
retired from Wageningen University and Research (WUR). I am now in the
final stage of my career at WUR. For a couple of years I had been thinking of
a revision of our handbook, to repair errors and to include new developments
in sampling design. Then I realised that to increase the impact of SNRM, it
might be a better idea to write a new book, showing by means of computer
code how the sampling designs can be implemented, and how the sample data
can be used in statistical inference.
A nice result of the publication of SNRM was that I was asked to give sampling
courses at many places in the world: China, Ethiopia, Uzbekistan, Australia,
and various countries in the European Union. I have very pleasant memories of
these courses, and they made my life as a scientist very joyful. For these courses
I wrote numerous scripts with computer code, using the popular programming
language R (R Core Team, 2021). My naïve idea was that all I had to do was
bundle these R scripts into an Rmarkdown document (Xie et al., 2020), and
add some text explaining the theory and the R code. As usual, it proved to
be much more work than expected, but I am very happy that I was able to
finish the job just before my retirement.
I could not have written this book without the help of many fellow researchers.
First, I am very grateful for the support I received from the authors of various
packages used in this book: Thomas Lumley for his support with package
survey, Yves Tillé and Alina Gabriela Matei with package sampling, Anton
Grafström with package BalancedSampling, Giulio Barcaroli and Marco
Ballin with package SamplingStrata, Andreas Hill and Alex Massey with
1 https://github.com/DickBrus/SpatialSamplingwithR/tree/master
xiv Preface

package forestinventory, and Martins Liberts with package surveyplan-

ning. I am ultimately responsible for all shortcomings of the R code.
Second, I would like to thank the following researchers for their valuable
comments on parts of the book: Gerard Heuvelink (Wageningen University
and ISRIC World Soil Information, the Netherlands), Yuha Heikkinen (Luke,
Natural Resources Institute, Finland), David Rossiter (Cornell University,
USA, and ISRIC World Soil Information, the Netherlands), Steve Stehman
(SUNY College of Environmental Science and Forestry, USA), Anton Grafström
(Swedish University of Agricultural Sciences), Dennis Walvoort (Wageningen
University and Research, the Netherlands), and Ben Marchant (British Geo-
logical Survey, United Kingdom). Dennis Walvoort also was very supportive
with the building of the R package sswr containing the data sets and some
functions used in this book and improving the R code chunks.
Third, I would like to thank Alexandre Wadoux (University of Sydney) for
preparing the data set of aboveground biomass and numerous environmental
and climatological covariates of Eastern Amazonia, Brazil; Coen Bussink
(United Nations Office on Drugs and Crime) for giving permission to use data
on the occurrence of opium poppy fields in Kandahar, Afghanistan; Akmal
Akramkhanov for providing the data set with measurements of the salinity of
soil at a farm which is part of a regional Cotton Research Station in Khorezm,
Uzbekistan, collected in the ZEF/UNESCO Landscape Restructuring project
in the Khorezm province, with financial support by the German Ministry
for Education and Research (BMBF; project number 0339970A); Lin Yang
(Nanjing University) and A-Xing Zhu (Nanjing Normal University) for giving
permission to use the data on soil organic matter concentration in Xuancheng,
Anhui province, China, collected in a project supported by the National Natural
Science Foundation of China (project numbers 41471178, 41971054, 41431177);
the Ethiopian Agricultural Transformation Agency (my contact was Hailu
Shiferaw) for allowing me to use the soil organic matter data of the western
part of the Amhara region in Ethiopia which I used for training; Siegfried
Hofman (Flemish Institute for Technological Research) for giving permission
to use the nitrate-N data of several agricultural fields in Flanders, Belgium;
and Budiman Minasny (University of Sydney) for giving permission to use the
raster maps with terrain attributes in Hunter Valley, Australia.
Last but not least, I would like to thank my sister-in-law, Marijke Compaijen,
who carefully read the entire manuscript. She identified and corrected all kinds
of errors and made several helpful suggestions. I am very grateful for her
extremely precise work.
1
Introduction

This book is about sampling for spatial surveys. A survey is an inventory of an

object of study about which statistical statements will be made based on data
collected from that object. The object of study is referred to as the population
of interest or target population. Examples are a survey of the organic carbon
stored in the soil of a country, the water quality of a lake, the wood volume
in a forest, the yield of rice in a country, etc. In these examples, soil organic
carbon, water quality, wood volume, and rice yield are the study variables, i.e.,
the variables of which we want to estimate the population mean or some other
parameter, or which we want to map. So, this book is about observational
research, not about experiments. In experiments, observations are done under
controlled circumstances; think of an experiment on crop yields as a function
of application rates of fertiliser. Several levels of fertiliser application rate are
chosen and randomly assigned to experimental plots. In observational research,
factors that influence the study variable are not controlled. This implies that
in observational research no conclusions can be drawn on causal relations.
If the whole population is observed, this is referred to as a census. In general,
we cannot afford such a census. Only some parts of the population are selected
and the study variable is observed (measured) for these selected parts (the
population units in the sample) only. Such a survey is referred to as a sample
survey. The observations are subsequently used to derive characteristics of the
whole population. For instance, to estimate the wood volume in a forest, we
cannot afford to measure the wood volume of every tree in the forest. Instead,
some trees are selected, the wood volume of these trees is measured, and based
on these measurements the total wood volume in the forest is estimated.

1.1 Basic sampling concepts

In this book the populations of interest have a spatial dimension. In selecting
parts of such populations for observation, we may account for the spatial coordi-
nates of the parts, but this is not strictly needed. Examples of spatial sampling
designs are designs selecting sampling units that are spread throughout the

1
2 1 Introduction

study area, often leading to more precise estimates of the population mean or
total as compared to sampling designs resulting in spatial clusters of units.
Two types of populations can be distinguished: discrete and continuous pop-
ulations. Discrete populations consist of discrete natural objects, think of
trees, agricultural fields, lakes, etc. These objects are referred to as popula-
tion units. The total number of population units in a discrete population
is finite. A finite spatial population of discrete units can be denoted by
𝒰 = {𝑢(𝐬1 ), 𝑢(𝐬2 ), … , 𝑢(𝐬u� )}, with 𝑢(𝐬u� ) the unit located at 𝐬u� , where 𝐬
is a vector with spatial coordinates. The population units naturally serve
as the elementary sampling units. In this book the spatial populations are
two-dimensional, so a vector 𝐬 has two coordinates, Easting and Northing.
Other populations may, for the purpose of sampling, be considered as a physical
continuum, e.g., the soil in a region, the water in a lake, the crop on a field.

If interest lies in crop properties per areal unit of the field, the population is
continuous. However, if interest lies in properties per plant, the population
is discrete and finite.

Such continuous spatial populations can be denoted by 𝒰 = {𝑢(𝐬), 𝐬 ∈ 𝒜},

with 𝒜 the study area. Discrete objects that can serve as elementary sampling
units do not exist in continuous populations. Therefore, we must define the
elementary sampling units. The elementary sampling units can be areal units,
e.g., 10 m squares; or circular plots, e.g., with a radius of 5 m; or “points”, i.e.,
units of such a small area, compared to the area of the population, that the
area of the units can be ignored.
In this book a population unit and an elementary sampling unit can be an
individual object of a discrete population as well as an areal sampling unit or
a point of a continuous population.
The size and geometry of the elementary units used in sampling a continuous
population is referred to as the sample support. The total number of elementary
sampling units in a continuous population can be finite, e.g., all 25 m × 25 m
(disjoint) raster cells in an area (raster cells in Figure 1.1), or infinite, e.g., all
points in an area, or all squares or circular plots with a given radius that are
allowed to overlap in an area (circles in Figure 1.1).
Ideally, with areal elementary sampling units, the selected elementary units
are exhaustively observed, so that a measurement of the total or mean of the
study variable within an areal unit is obtained, think for instance of the total
aboveground biomass. In some cases this is not feasible, think for instance
of measuring the mean of some soil property in 25 m squares. In this case, a
sample of points is selected from each selected square, and the measurement is
done at the selected points. These measurements at points are used to estimate
the mean of the squares. Stehman et al. (2018) introduced the concept of a
response design as “the protocol used to determine the reference condition of an
1.1 Basic sampling concepts 3

FIGURE 1.1: Three sample supports: points, squares, and circles. With
disjoint squares, the population is finite. With points, and squares or circles
that are allowed to overlap, the population is infinite.

element of the population”. So, in the case just mentioned the response design
is the sampling design and the estimator for the mean of the soil property of
the 25 m squares.
Ideally, the sample support is constant, but in some situations a varying sample
support cannot be avoided. Think, for instance, of square sampling units in an
irregularly shaped study area. Near the border of the study area, there are
squares that cross the border. The part of a square that falls outside the study
area is not observed. So, the support of the observations of squares crossing
the border is smaller than that of the observations of squares in the interior of
the study area. See also Section 3.4.
To sample a finite spatial population, the population units are listed in a data
frame. This data frame contains the spatial coordinates of the population
units and other information needed for selecting sampling units according to a
specific design. Think, for instance, of the labels of more or less homogeneous
subpopulations (used as strata in stratified random sampling, see Chapter
4) and the labels of clusters of population units, for instance, all units in a
polygon of a map (used in cluster random sampling, see Chapter 6). Besides,
if we have information about covariates possibly related to the study variable,
which we would like to use in selecting the population units, these covariates
are added to the list. The list used for selecting sampling units is referred to
as the sampling frame.
If the elementary sampling units are disjoint square grid cells (sample support
is a square), the population is finite and the grid cells can be selected through
selection of their centres (or any other point that uniquely identifies a grid
cell) listed in the sampling frame.
In this book also continuous populations are sampled using a list as a sampling
frame. The infinite population is discretised by the cells of a fine discretisation
grid. The grid cells are listed in the sampling frame by the spatial coordinates of
the centres of the grid cells. So, the infinite population is represented by a finite
4 1 Introduction

list of points. The advantage of this is that existing R packages for sampling
of finite populations can also be used for sampling infinite populations.

FIGURE 1.2: Sampling of points from discretised infinite population. The

grid cells are randomly selected with replacement. Each time a grid cell is
selected, a point is randomly selected from that grid cell.

If the elementary sampling units are points (sample support is a point), the
population is infinite. In this case, sampling of points can be implemented
by a two-step approach. In the first step, cells of the discretisation grid are
selected with or without replacement, and in the second step one or more
points are selected within the selected grid cells. Figure 1.2 is an illustration of
this two-step approach for simple random sampling of points from a discretised
infinite population. Ten grid cells are selected by simple random sampling
with replacement. Every time a grid cell is selected, one point is randomly
selected from that grid cell. Note that a grid cell can be selected more than
once, so that more than one point will be selected from that grid cell. Note
also that we may select a point that falls outside the boundary of the study
area. This is actually the case with one grid cell in Figure 1.2. The points
1.1 Basic sampling concepts 5

outside the study area are discarded and replaced by a randomly selected new
point inside the study area. Finally, note that near the boundary there are
small areas not covered by a grid cell, so that no points can be selected in
these areas. It is important that the discretisation grid is fine enough to keep
the discretisation error so small that it can be ignored. The alternative is to
extend the discretisation grid beyond the boundaries of the study area so that
the full study area is covered by grid cells.

1.1.1 Population parameters

The sample data are used to estimate characteristics of the whole population,
e.g., the population mean or total; some quantile, e.g., the median or the 90th
percentile; or even the entire cumulative frequency distribution.
A finite population total is defined as

u�
𝑡(𝑧) = ∑ 𝑧u� = ∑ 𝑧u� , (1.1)
u�∈u� u�=1

with 𝑁 the number of population units and 𝑧u� the study variable for population
unit 𝑘. A finite population mean is defined as a finite population total divided
by 𝑁.
An infinite population total is defined as an integral of the study variable over
the study area:

𝑡(𝑧) = ∫ 𝑧(𝐬) d𝐬 . (1.2)

u�∈u�

An infinite population mean is defined as a finite population total divided by

the area, 𝐴, covered by the population.
A finite population proportion is defined as the population mean of an 0/1
indicator 𝑦 with value 1 if the condition is satisfied, and 0 otherwise:

∑u� 𝑦
u�=1 u�
𝑝= . (1.3)
𝑁
A cumulative distribution function (CDF) is defined as

𝐹(𝑧) = ∑ 𝑝(𝑧′ ) , (1.4)

u�′ ≤u�

with 𝑝(𝑧′ ) the proportion of population units whose value for the study variable
equals 𝑧′ .
6 1 Introduction

A population quantile, for instance the population median or the population

90th percentile, is defined as

𝑞u� = 𝐹−1 (𝑝) , (1.5)

where 𝑝 is a number between 0 and 1 (e.g., 0.5 for the median, 0.9 for the 90th
percentile), and 𝐹−1 (𝑝) is the smallest value of the study variable 𝑧 satisfying
𝐹(𝑧) ≥ 𝑝.
In surveys of spatial populations, the aim can also be to make a map of the
population.

The parameters defined in this subsection are parameters of spatial popula-

tions, i.e., populations observed in a relatively short period of time related
to the dynamics of the study variable. We assume that the study variable
does not change during the survey period. In Chapter 15 parameters are
defined for space-time populations.

1.1.2 Descriptive statistics vs. inference about a population

When we observe only a (small) part of the population, we are uncertain about
the population parameter estimates and the map of the population. By using
statistical methods, we can quantify how uncertain we are about these results.
In decision making it can be important to take this uncertainty into account.
An example is a survey of water quality. In Europe the concentration levels of
nutrients are regulated in the European Water Framework Directive. To test
whether the mean concentration of a nutrient complies with its standard, it is
important to account for the uncertainty in the estimated mean. When the
estimated mean is just below the standard, there is still a large probability
that the population mean exceeds the standard. This example shows that it is
important to distinguish computing descriptive statistics from characterising
the population using the sample data. For instance, we can compute the sample
mean (average of the sample data) without error, but if we use this sample
mean as an estimate of the population mean, there is certainly an error in this
estimate.

1.1.3 Random sampling vs. probability sampling

Many sampling methods are available. At the highest level, one may distinguish
random from non-random sampling methods. In random sampling, a subset of
population units is randomly selected from the population, using a (pseudo)
random number generator. In non-random sampling, no such random number
1.2 Design-based vs. model-based approach 7

generator is used. Examples of non-random sampling are (i) convenience

sampling, i.e., sampling at places that are easy to access, e.g., along roads; (ii)
arbitrary sampling, i.e., sampling without a specific purpose in mind; and (iii)
targeted sampling, e.g., at sites suspected of soil pollution.
In the literature, the term random sampling is often used for arbitrary sampling,
i.e., sampling without a specific purpose in mind. To avoid confusion the term
probability sampling is used for random sampling using a (pseudo) random
number generator, so that for any unit in the population the probability
of selecting that unit is known. More precisely, a probability sample is a
sample from a population such that every unit of the population has a positive
probability of being included in the sample. Besides, these inclusion probabilities
must be known, at least for the selected units, as they are needed in estimation.
This is explained in following chapters.

1.2 Design-based vs. model-based approach

The choice between probability or non-probability sampling is closely con-
nected with the choice between the design-based or the model-based approach
for sampling and statistical inference (estimation, hypothesis testing). The
difference between these two approaches is a rather technical subject, so, not
to discourage you already in this very first chapter, I will keep it short. In
Chapter 26 I elaborate on the fundamental difference of these two approaches
and a third approach, the model-assisted approach, which can be seen as a
compromise of the design-based and the model-based approach.
In the design-based approach, units are selected by probability sampling (Table
1.1). Estimates are based on the inclusion probabilities of the sampling units
as determined by the sampling design (design-based inference). No model is
used in estimation. On the contrary, in the model-based approach a statistical
model is used in prediction, i.e., a model with a random error term, for

TABLE 1.1: Statistical approaches for sampling and inference.

Approach Sampling Inference

Design-based Probability sampling Based on sampling distribution
required
(no model used)
Model-based Probability sampling Based on statistical model
not required
8 1 Introduction

instance a regression model. As the model already contains a random error

term, probability sampling is not required in this approach.
Which statistical approach is best largely depends on the aim of the survey, see
Brus and de Gruijter (1997) and de Gruijter et al. (2006). Broadly speaking,
the following aims can be distinguished:

1. estimating parameters for the population;

2. estimating parameters for several subpopulations; and
3. mapping the study variable.

A map of the study variable is obtained by predicting the study variable at

the points of a very fine grid that discretises the study area, or by predicting
the means of the study variable for fine grid cells. Many mapping methods
are available. In this book a statistical model is applied to predict the study
variable, for instance a linear regression model, or a spatial linear mixed
model as used in kriging.

When the aim is to map the study variable, a model-based approach is the
most natural option. This implies that for this aim probability sampling is not
necessarily required. In principle, both approaches are suitable for estimating
(sub)population parameters. The more subpopulations are distinguished, the
more attractive a model-based approach becomes. Model-based estimates of
the subpopulation means or totals are potentially more accurate (depending on
how good the model is) than model-free design-based estimates. On the other
hand, an advantage of design-based estimation is that an objective assessment
of the uncertainty of the estimated mean or total is warranted, and that the
coverage of confidence intervals is (almost) correct.
A probability sample can also be used in model-based inference. This flexibility
can be attractive when we have a dual aim, mapping as well as estimation of
parameters of (sub)populations. When units are not selected by probability
sampling, model-free design-based estimation is impossible, and model-based
prediction is the only option.

1.3 Populations used in sampling experiments

In this book various data sets are used to illustrate the sampling designs. Four
data sets, Voorst, Kandahar, Eastern Amazonia, and the Iberian Peninsula
(Spain and Portugal), are exhaustive, i.e., for all population units, data of
the study variable and ancillary data are available. The first two exhaustive
1.3 Populations used in sampling experiments 9

data sets are obtained through simulation, i.e., by drawing numbers from a
probability distribution. Sample data from these two study areas are used
to calibrate a statistical model. This model is subsequently used to simulate
values of the study variable for all population units. Voorst actually is an
infinite population of points. However, this study area is discretised by the
cells of a fine grid, and the study variable, the soil organic matter (SOM)
concentration, is simulated for all centres of the grid cells. Kandahar is a finite
population consisting of 965 squares of size 5 km × 5 km. The study variable
is the area cultivated with poppy. Eastern Amazonia is a map in raster format,
with a resolution of 1 km × 1 km. The study variable is the aboveground
biomass as derived from remote sensing images. The aboveground biomass
value of a raster cell is treated as the average biomass of that raster cell. The
data set Iberian Peninsula is a time series of four maps in raster format with a
resolution of 30 arc sec. The study variable is the annual mean air temperature
at two metres above the earth surface in ∘ C.
The exhaustive data sets are used in the first part of this book on probability
sampling for estimating population parameters. By taking the population as
the reality, we know the population parameters. Also, for any randomly selected
sample from this population, the study variable values for the selected sampling
units are known, so that we can estimate the population parameters from
this sample. An estimated population parameter can then be compared with
the population parameter. The difference between these two is the sampling
error in the estimated population parameter. This opens up the possibility
of repeating the random selection of samples with a given sampling design a
large number of times, estimating the population parameter for every sample,
so that a frequency distribution of the estimated population parameter is
obtained. Ideally, the mean of this frequency distribution, referred to as the
sampling distribution, is equal to the population parameter (mean sampling
error equals zero), and the variance of the estimated population parameters is
small. Another advantage is that sampling designs can be compared on the
basis of the sampling distribution, for instance the sampling distributions of
the estimator of the population mean with stratified random sampling and
simple random sampling, to evaluate whether the stratification leads to more
accurate estimates of the population mean.
Furthermore, various data sets are used with data for a sample of population
units only. These data sets are described at places where they are first used.
All data sets are available by installing the R package sswr. This package
can be installed from github with function install_github of package remotes
(Csárdi et al., 2021).

library(remotes)
install_github("DickBrus/sswr")
10 1 Introduction

The package can then be loaded. You can see the contents of the package
and of the data files by typing a question mark, followed by the name of the
package or a data file.

library(sswr)
?sswr
?grdVoorst

1.3.1 Soil organic matter in Voorst, the Netherlands

The study area of Voorst is located in the eastern part of the Netherlands. The
size of the study area is 6 km × 1 km. At 132 points, samples of the topsoil
were collected by graduate students of Wageningen University, which were then
analysed for SOM concentrations (in g kg-1 dry soil) in a laboratory. The map
is created by conditional geostatistical simulation of natural logs of the SOM
concentration on a 25 m × 25 m grid, followed by backtransformation, using
a linear mixed model with spatially correlated residuals and combinations of
soil type and land use as a qualitative predictor (factor). Figure 1.3 shows the
simulated map of the SOM concentration.

FIGURE 1.3: Simulated map of the SOM concentration (g kg-1 ) in Voorst.

The frequency distribution of the simulated values at all 7,528 grid cells shows
that the SOM concentration is skewed to the right (Figure 1.4).
Summary statistics are:
Min. 1st Qu. Median Mean 3rd Qu. Max.
10.75 49.45 68.15 81.13 100.63 394.45

The ancillary information consists of a map of soil classes and a land use
map, which are combined to five soil-land use combinations (Figure 1.5). The
first letter in the labels for the combinations stands for the soil type: B for
beekeerdgrond (sandy wetland soil with gleyic properties), E for enkeerdgrond
(sandy soil with thick anthropogenic humic topsoil), P for podzols (sandy soil
with eluviated horizon below the topsoil), R for river clay soil, and X for other
sandy soils. The second letter is for land use: A for agriculture (grassland,
arable land) and F for forest.
1.3 Populations used in sampling experiments 11

FIGURE 1.4: Frequency distribution of the simulated SOM concentration

(g kg-1 ) in Voorst.

FIGURE 1.5: Soil-land use combinations in Voorst.

1.3.2 Poppy fields in Kandahar, Afghanistan

Cultivation of poppy for opium production is a serious problem in Afghanistan.

The United Nations Office on Drugs and Crime (UNODC) monitors the area
cultivated with poppy through detailed analysis of aerial photographs and
satellite images. This is laborious, and therefore the analysis is restricted to
a probability sample of 5 km squares. These sample data are then used to
estimate the total poppy area (Anonymous, 2014).
In 2014 the poppy area within 83 randomly selected squares in the province
of Kandahar was determined, as well as the agricultural area within all 965
squares in this province. These data were used to simulate a map of poppy
area per 5 km square. The map is simulated with an ordinary kriging model
for the logit transform of the proportion of the agricultural area cultivated
with poppy within 5 km squares. For privacy reasons, the field was simulated
unconditionally on these sample data. Figure 1.6 shows the map with the
agricultural area in hectares per 5 km square, and the map with the simulated
12 1 Introduction

poppy area in hectares per square. The frequency distribution of the simulated
poppy area per square shows very strong positive skew (Figure 1.7). For 375
squares, the simulated poppy area was smaller than 1 hectare (ha).

FIGURE 1.6: Agricultural area and simulated poppy area, in ha per 5 km

square, in Kandahar.

1.3.3 Aboveground biomass in Eastern Amazonia, Brazil

This data set consists of data on the live woody aboveground biomass (AGB)
in megatons per ha (Baccini et al., 2012). A rectangular area of 1,642 km × 928
km in Eastern Amazonia, Brazil, was selected from this data set. The data were
aggregated to a map with a resolution of 1 km × 1 km. Besides, a stack of five
ecologically relevant covariates of the same spatial extent was prepared, being
long term mean of MODIS short-wave infrared radiation (SWIR2), primary
production in kg C per m2 (Terra_PP), average precipitation in driest month
1.3 Populations used in sampling experiments 13

FIGURE 1.7: Frequency distribution of simulated poppy area in ha per 5

km square, in Kandahar.

in mm (Prec_dm), elevation in m, and clay content in g kg-1 soil. All covariates

were either resampled by bilinear interpolation or aggregated to conform with
the grid of the aboveground biomass map. Figure 1.8 shows a map of AGB
and SWIR2.
Figure 1.9 shows a matrix of two-dimensional density plots of aboveground
biomass and the five covariates, made with function ggpairs of R package
GGally (Schloerke et al., 2021). The covariate with the strongest correlation
with AGB is SWIR2. The Pearson correlation coefficient with AGB is -0.80.
The relation does not look linear. The correlation of AGB with the covariates
Terra_PP and Prec_dm is weakly positive. All correlations are significant,
but this is not meaningful because of the very large number of data used in
computing the correlation coefficients.

1.3.4 Annual mean air temperature in Iberia

The space-time designs of Chapter 15 are illustrated with the annual mean
air temperature at two metres above the earth surface (TAS) in ∘ C, in Iberia
(Spain and Portugal, islands excluded) for 2004, 2009, 2014, and 2019 (Figure
1.10). These data are part of the data set CHELSA1 (Karger et al., 2017). The
raster files are latitude-longitude grids with a resolution of 30 arc sec. The
data are projected using the Lambert azimuthal equal area (laea) projection.
The resolution of the resulting laea raster file is about 780 m × 780 m.

1 https://chelsa-climate.org/wp-admin/download-page/CHELSA_tech_specification_V2.pdf
14 1 Introduction

FIGURE 1.8: Aboveground biomass (AGB) in 109 kg ha-1 and short-wave

infrared radiation (SWIR2) of Eastern Amazonia.
1.3 Populations used in sampling experiments 15

FIGURE 1.9: Matrix of two-dimensional density plots of AGB and five

covariates of Eastern Amazonia. Terra_PP values are divided by 1000.
16 1 Introduction

FIGURE 1.10: Annual mean air temperature in Iberia for 2004, 2009, 2014,
and 2019.
 Part I

Probability sampling for

estimating population
parameters
2
Introduction to probability sampling

To estimate population parameters like the mean or the total, probability

sampling is most appropriate. Probability sampling is random sampling using
a random number generator such that all population units have a probability
larger than zero of being selected, and these probabilities are known for at
least the selected units.
The probability that a unit is included in the sample, in short the inclusion
probability of that unit, can be calculated as the sum of the selection probabil-
ities over all samples that can be selected with a given sampling design and
that contain this unit. In formula:

𝜋u� = ∑ 𝑝(𝒮) , (2.1)

u�∋u�

where 𝒮 ∋ 𝑘 indicates that the sum is over all samples that contain unit 𝑘, and
𝑝(𝒮) is the selection probability of sample 𝒮. 𝑝(⋅) is called the sampling design.
It is a function that assigns a probability to every possible sample (subset of
population units) that can be selected with a given sample selection scheme
(sampling algorithm). For instance, consider the following sample selection
scheme from a finite population of 𝑁 units:

1. Select with equal probability 1/𝑁 a first unit.

2. Select with equal probability 1/(𝑁 − 1) a second unit from the

remaining 𝑁 − 1 units.

3. Repeat this until an 𝑛th unit is selected with equal probability from
the 𝑁 − (𝑛 − 1) units.

This is a selection scheme for simple random sampling without replacement.

With this scheme the selection probability of any sample of 𝑛 units is 1/(u� u� )
(there are ( u� ) samples of size 𝑛, and each sample has an equal selection
u�

probability), and zero for all other samples. There are (u�−1
u�−1 ) samples of size 𝑛
in which unit 𝑘 is included. The inclusion probability of each unit 𝑘 therefore
is (u�−1
u�−1 )/( u� ) = u� . The sampling design plays a key role in the design-based
u� u�

approach as it determines the sampling distribution of random quantities

19
20 2 Introduction to probability sampling

computed from a sample such as the estimator of the population mean, see
Section 2.1. The number of selected population units is referred to as the
sample size.

In sampling with replacement, each unit can be selected more than once. In
this case, the sample size refers to the number of draws, not to the number
of unique population units in the sample.

A common misunderstanding is that with probability sampling the inclusion

probabilities must be equal. Sampling with unequal inclusion probabilities can
be more efficient than with equal probabilities. Unequal probability sampling
is no problem as long as the inclusion probabilities are known and proper
formulas are used for estimation, see Section 2.1.
There are many schemes for selecting a probability sample. The following
sampling designs are described and illustrated in this book:

1. simple random sampling;

2. stratified random sampling;
3. systematic random sampling;
4. cluster random sampling;
5. two-stage cluster random sampling;
6. sampling with probabilities proportional to size;
7. balanced and well-spread sampling; and
8. two-phase random sampling.

The first five sampling designs are basic sampling designs. Implementation of
these designs is rather straightforward, as well as the associated estimation
of the population mean, total, or proportion, and their sampling variance.
The final three sampling designs are more advanced. Appropriate use of these
designs requires more knowledge of sampling theory and statistics, such as
linear regression.

Arbitrary (haphazard) sampling vs. probability sampling

In publications it is commonly stated that the sampling units were selected
at random (within strata), without further specification of how the sampling
units were precisely selected. In statistical inference, the sampling units are
subsequently treated as if they were selected by (stratified) simple random
sampling. With probability sampling, all units in the population have a positive
probability of being selected, and the inclusion probabilities are known for
2.1 Horvitz-Thompson estimator 21

all units. It is highly questionable whether this also holds for arbitrary and
haphazard sampling. In arbitrary and haphazard sampling, the sampling units
are not selected by a probability mechanism. So, the selection probabilities of
the sampling units and of combinations of sampling units are unknown. Design-
based estimation is therefore impossible, because it requires the inclusion
probabilities of the population units as determined by the sampling design.
The only option for statistical analysis using arbitrarily or haphazardly selected
samples is model-based inference, i.e., a model of the spatial variation must be
assumed.

Exercises

1. Suppose a researcher selects a sample of points from a study area by

throwing darts on a map depicting the study area. Is the resulting
sample a probability sample? If not, why?

2.1 Horvitz-Thompson estimator

For any probability sampling design, the population total can be estimated as
a weighted sum of the observations (measurements) of the study variable on
the selected population units:
𝑡u�̂ (𝑧) = ∑ 𝑤u� 𝑧u� , (2.2)
u�∈u�
22 2 Introduction to probability sampling

with 𝒮 the sample, 𝑧u� the observed study variable for unit 𝑘, and 𝑤u� the
design weight attached to unit 𝑘:

1
𝑤u� = , (2.3)
𝜋u�

with 𝜋u� the inclusion probability of unit 𝑘. The estimator of Equation (2.2)
is referred to as the Horvitz-Thompson estimator or 𝜋 estimator. The 𝑧u� /𝜋u� -
values are referred to as the 𝜋-expanded values. The 𝑧-value of unit 𝑘 in the
sample is multiplied by the reciprocal of the inclusion probability of that unit,
and the sample sum of these 𝜋-expanded values is used as an estimator of the
population total. The inclusion probabilities are determined by the type of
sampling design and the sample size.

An estimator is not the same as an estimate. Whereas an estimate is a

particular value calculated from the sample data, an estimator is a formula
for estimating a parameter. An estimator is a random variable and therefore
has a probability distribution. For this reason, it is not correct, although very
common, to say ‘the variance (standard error) of the estimated population
mean equals …’. It is correct to say ‘the variance (standard error) of the
estimator of the population mean equals …’.

Also for infinite populations, think of points in a continuous population, the

above estimator for the population total can be used, but special attention must
then be paid to the inclusion probabilities. Suppose the infinite population is
discretised by 𝑁 cells of a fine grid, and a simple random sample of 𝑛 cells is
selected. The inclusion probabilities of the grid cells is then 𝑛/𝑁. However,
constraining the sampling points to the centres of the cells of the discretisation
grid is not needed and even undesirable. To account for the infinite number
of points in the population we may adopt a two-step approach, see Figure
1.2. In the first step, 𝑛 cells of the discretisation grid are selected by simple
random sampling with replacement. In the second step, one point is selected
fully randomly from the selected grid cells. If a grid cell is selected more than
once, more points are selected in that grid cell. With this selection procedure
the inclusion probability density is 𝑛/𝐴, with 𝐴 the area of the study area.
This inclusion probability density equals the expected number of sampling
points per unit area, e.g., the expected number of points per ha or per m2 .
The inclusion probability density can be interpreted as the global sampling
intensity. Note that the local sampling intensity may strongly vary; think for
instance of cluster random sampling.
The 𝜋 estimator for the mean of a finite population, 𝑧,̄ is simply the 𝜋 estimator
for the total, divided by the total number of units in the population, 𝑁:
1 1
𝑧u�̄̂ = ∑ 𝑧 . (2.4)
𝑁 u�∈u� 𝜋u� u�
2.2 Hansen-Hurwitz estimator 23

For infinite populations discretised by a finite set of points, the same estimator
can be used.
For infinite populations, the population total can be estimated by multiplying
the estimated population mean by the area of the population 𝐴:

𝑡u�̂ (𝑧) = 𝐴𝑧u�̄̂ . (2.5)

The 𝜋 estimator can be worked out for the different types of sampling design
listed above by inserting the inclusion probabilities as determined by the
sampling design. For simple random sampling, this leads to the unweighted
sample mean (see Chapter 3), and for stratified simple random sampling the 𝜋
estimator is equal to the weighted sum of the sample means per stratum, with
weights equal to the relative size of the strata (see Chapter 4).

2.2 Hansen-Hurwitz estimator

In sampling finite populations, units can be selected with or without replace-
ment. In sampling with replacement after each draw the selected unit is
replaced. As a consequence, a unit can be selected more than once. Sampling
with replacement is less efficient than sampling without replacement. If a
population unit is selected in a given draw, there is no additional information
in this unit if it is selected again. One reason that sampling with replacement
is still used is that it is easier to implement.
The most common estimator used for sampling with replacement is the Hansen-
Hurwitz estimator, referred to as the 𝑝-expanded with replacement (pwr)
estimator by Särndal et al. (1992). With direct unit sampling, i.e., sampling of
individual population units, the pwr estimator is

1 𝑧
̂ (𝑧) =
𝑡pwr ∑ u� , (2.6)
𝑛 u�∈u� 𝑝u�

with 𝑝u� the draw-by-draw selection probability of population unit 𝑘. For instance,
in simple random sampling with replacement the draw-by-draw selection
probability 𝑝 of each unit is 1/𝑁. If we select only one unit 𝑘, the population
total can be estimated by the observation of that unit divided by 𝑝, 𝑡(𝑧) ̂ =
𝑧u� /𝑝u� = 𝑁𝑧u� . If we repeat this 𝑛 times, this results in 𝑛 estimated population
totals. The pwr estimator is the average of these 𝑛 elementary estimates. If
a unit occurs multiple times in the sample 𝒮, this unit provides multiple
elementary estimates of the population total.
24 2 Introduction to probability sampling

A sample obtained by sampling with replacement is referred to as an ordered

sample (Särndal et al., 1992). Selecting the distinct units from this ordered
sample results in the set-sample. Instead of using the ordered sample in the
pwr estimator, we may use the set-sample in the 𝜋 estimator. This requires
computation of the inclusion probabilities for sampling with replacement.
For instance, for simple random sampling with replacement, the inclusion
1 u�
probability of each unit equals 1 − (1 − u� ) , with 𝑛 the number of draws.
This probability is smaller than 𝑛/𝑁, the inclusion probability for simple
random sampling without replacement. There is no general rule on which
estimator is most accurate (Särndal et al., 1992). In this book I only use the
pwr estimator for sampling with replacement.
Sampling with replacement can also be applied at the level of clusters of
population units as in cluster random sampling and two-stage cluster random
sampling. If the clusters are selected with probabilities proportional to their size
and with replacement, estimation of a population parameter is rather simple.
This is a second reason why sampling with replacement can be attractive. With
cluster sampling, the Hansen-Hurwitz estimator is

1 𝑡u� (𝑧)
̂ (𝑧) =
𝑡pwr ∑ , (2.7)
𝑛 u�∈u� 𝑝u�

with 𝑡u� (𝑧) the total of the cluster selected in the 𝑗th draw. If not all population
units of a selected cluster are observed, but only a sample of population units
from a cluster, as in two-stage cluster random sampling, the cluster totals 𝑡u� (𝑧)
are replaced by the estimated cluster totals 𝑡u�̂ (𝑧).

Exercises

2. Consider a population of four units (𝑁 = 4). What is the inclusion

probability of each population unit for simple random sampling
without replacement and simple random sampling with replacement
of two units (𝑛 = 2)?

2.3 Using models in design-based approach

Design-based estimates of population parameters such as the mean, total, or
proportion (areal fraction) are model-free: no use is made of a model for the
spatial variation of the study variable. However, such a model can be used to
optimise the probability sampling design. In Chapter 13 I describe how a model
can be used to compare alternative sampling designs at equal costs or equal
2.3 Using models in design-based approach 25

precision to evaluate which sampling design performs best, to optimise the

sample size given a requirement on the precision of the estimated population
parameter, or to optimise the spatial strata for stratified random sampling.
A model of the spatial variation can also be used at a later stage, after the data
have been collected, in estimating the population parameter of interest. If one
or more ancillary variables that are related to the study variable are available,
these variables can be used in estimation to increase the accuracy. This leads
to alternative estimators, such as the regression estimator, the ratio estimator,
and the poststratified estimator (Chapter 10). These estimators together are
referred to as model-assisted estimators. In model-assisted estimation the
inclusion probabilities, as determined by the random sampling design, play a
key role, but besides, modelling assumptions about how the population might
have been generated are used to work out an efficient estimator. The role of a
model in the model-assisted approach is fundamentally different from its role
in the model-based approach. This is explained in Chapter 26.
For novices in geostatistics, Chapters 10 and 13 can be quite challenging, and
I recommend skipping these chapters and only return to them after having
read the introductory chapter on geostatistics (Chapter 21).
3
Simple random sampling

Simple random sampling is the most basic form of probability sampling. There
are two subtypes:

1. simple random sampling with replacement; and

2. simple random sampling without replacement.

This distinction is irrelevant for infinite populations. In sampling with replace-

ment a population unit may be selected more than once.
In R a simple random sample can be selected with or without replacement by
function sample from the base package. For instance, a simple random sample
without replacement of 10 units from a population of 100 units labelled as
1, 2, … , 100, can be selected by

sample(100, size = 10, replace = FALSE)

[1] 21 7 5 16 58 76 44 100 71 84

The number of units in the sample is referred to as the sample size (𝑛 = 10 in

the code chunk above). Use argument replace = TRUE to select a simple random
sample with replacement.
When the spatial population is continuous and infinite, as in sampling points
from an area, the infinite population is discretised by a very fine grid. Discreti-
sation is not strictly needed (we could also select points directly), but it is used
in this book for reasons explained in Chapter 1. The centres of the grid cells
are then listed in a data frame, which serves as the sampling frame (Chapter
1). In the next code chunk, a simple random sample without replacement of
size 40 is selected from Voorst. The infinite population is represented by the
centres of square grid cells with a side length of 25 m. These centres are listed
in tibble1 grdVoorst.

1 A tibble is a data frame of class tbl_df of package tibble (Müller and Wickham, 2021).

Hereafter, I will use the terms tibble and data frame interchangeably. A traditional data
frame is referred to as a data.frame.

27
28 3 Simple random sampling

n <- 40
N <- nrow(grdVoorst)
set.seed(314)
units <- sample(N, size = n, replace = FALSE)
mysample <- grdVoorst[units, ]
mysample

# A tibble: 40 x 4
s1 s2 z stratum
<dbl> <dbl> <dbl> <chr>
1 206992 464506. 23.5 EA
2 202567 464606. 321. XF
3 205092 464530. 124. XF
4 203367 464556. 53.6 EA
5 205592 465180. 38.4 PA
6 201842 464956. 159. XF
7 201667 464930. 139. XF
8 204317 465306. 59.4 PA
9 203042 464406. 90.5 BA
10 204567 464530. 48.1 PA
# ... with 30 more rows

The result of function sample is a vector with the centres of the selected cells
of the discretisation grid, referred to as discretisation points. The order of the
elements of the vector is the order in which these are selected. Restricting
the sampling points to the discretisation points can be avoided as follows. A
simple random sample of points is selected in two stages. First, n times a grid
cell is selected by simple random sampling with replacement. Second, every
time a grid cell is selected, one point is selected fully randomly from that
grid cell. This selection procedure accounts for the infinite number of points
in the population. In the code chunk below, the second step of this selection
procedure is implemented with function jitter. It adds random noise to the
spatial coordinates of the centres of the selected grid cells, by drawing from a
continuous uniform distribution unif(−𝑐, 𝑐), with 𝑐 half the side length of the
square grid cells. With this selection procedure we respect that the population
actually is infinite.

set.seed(314)
units <- sample(N, size = n, replace = TRUE)
mysample <- grdVoorst[units, ]
cellsize <- 25
mysample$s1 <- jitter(mysample$s1, amount = cellsize / 2)
mysample$s2 <- jitter(mysample$s2, amount = cellsize / 2)
mysample
3.0 Simple random sampling 29

# A tibble: 40 x 4
s1 s2 z stratum
<dbl> <dbl> <dbl> <chr>
1 206986. 464493. 23.5 EA
2 202574. 464609. 321. XF
3 205095. 464527. 124. XF
4 203369. 464556. 53.6 EA
5 205598. 465181. 38.4 PA
6 201836. 464965. 159. XF
7 201665. 464941. 139. XF
8 204319. 465310. 59.4 PA
9 203052. 464402. 90.5 BA
10 204564. 464529. 48.1 PA
# ... with 30 more rows

Variable stratum is not used in this chapter but in the next chapter. The selected
sample is shown in Figure 3.1.

FIGURE 3.1: Simple random sample of size 40 from Voorst.

Dropouts

In practice, it may happen that inspection in the field shows that a selected
sampling unit does not belong to the target population or cannot be observed
for whatever reason (e.g., no permission). For instance, in a soil survey the
sampling unit may happen to fall on a road or in a built-up area. What to do
with these dropouts? Shifting this unit to a nearby unit may lead to a biased
estimator of the population mean, i.e., a systematic error in the estimated
population mean. Besides, knowledge of the inclusion probabilities is lost. This
can be avoided by discarding these units and replacing them by sampling
units from a back-up list, selected in the same way, i.e., by the same type of
sampling design. The order of sampling units in this list must be the order in
which they are selected. In summary, do not replace a deleted sampling unit
by the nearest sampling unit from the back-up list, but by the first unit, not
yet selected, from the back-up list.
30 3 Simple random sampling

3.1 Estimation of population parameters

In simple random sampling without replacement of a finite population, every
possible sample of 𝑛 units has an equal probability of being selected. There
are (u�
u� ) samples of size 𝑛 and ( u�−1 ) samples that contain unit 𝑘. From
u�−1

this it follows that the probability that unit 𝑘 is included in the sample is
u�−1 )/( u� ) = u� (Lohr, 1999). Substituting this in the general 𝜋 estimator
(u�−1 u� u�

for the total (Equation (2.2)) gives for simple random sampling without
replacement (from finite populations)

𝑁
̂ =
𝑡(𝑧) ∑ 𝑧 = 𝑁𝑧u�̄ , (3.1)
𝑛 u�∈u� u�

with 𝑧u�̄ the unweighted sample mean. So, for simple random sampling without
replacement the 𝜋 estimator of the population mean is the unweighted sample
mean:

1
𝑧 ̄̂ = 𝑧u�̄ = ∑𝑧 . (3.2)
𝑛 u�∈u� u�

In simple random sampling with replacement of finite populations, a unit may

occur multiple times in the sample 𝒮. In this case, the population total can be
estimated by the pwr estimator (Särndal et al., 1992)

1 𝑧
̂ =
𝑡(𝑧) ∑ u� , (3.3)
𝑛 u�∈u� 𝑝u�

where 𝑛 is the number of draws (sample size) and 𝑝u� is the draw-by-draw
selection probability of unit 𝑘. With simple random sampling 𝑝u� = 1/𝑁, 𝑘 =
1, … , 𝑁. Inserting this in the pwr estimator yields

𝑁
̂ =
𝑡(𝑧) ∑𝑧 , (3.4)
𝑛 u�∈u� u�

which is equal to the 𝜋 estimator of the population total for simple random
sampling without replacement.
Alternatively, the population total can be estimated by the 𝜋 estimator. With
simple random sampling with replacement the inclusion probability of each
1 u�
unit 𝑘 equals 1 − (1 − u� ) , which is smaller than the inclusion probability
with simple random sampling without replacement of size 𝑛 (Särndal et al.,
1992). Inserting these inclusion probabilities in the general 𝜋 estimator of the
population total (Equation (2.2)), where the sample 𝒮 is reduced to the unique
3.1 Estimation of population parameters 31

units in the sample, yields the 𝜋 estimator of the total for simple random
sampling with replacement.
With simple random sampling of infinite populations, the 𝜋 estimator of the
population mean equals the sample mean. Multiplying this estimator with the
area of the region of interest 𝐴 yields the 𝜋 estimator of the population total:

𝐴
̂ =
𝑡(𝑧) ∑𝑧 . (3.5)
𝑛 u�∈u� u�

As explained above, selected sampling units that do not belong to the target
population must be replaced by a unit from a back-up list if we want to
observe the intended number of units. The question then is how to estimate
the population total and mean. We cannot use the 𝜋 estimator of Equation
(3.1) to estimate the population total, because we do not know the population
size 𝑁. The population size can be estimated by

̂ = 𝑛 − 𝑑 𝑁∗ ,
𝑁 (3.6)
𝑛
with 𝑑 the number of dropouts and 𝑁∗ the supposed population size, i.e., the
number of units in the sampling frame used to select the sample. This yields
the inclusion probability

𝑛 𝑛2
𝜋u� = = . (3.7)
̂ (𝑛 − 𝑑)𝑁∗
𝑁
Inserting this in the 𝜋 estimator of the population total yields

(𝑛 − 𝑑)𝑁∗ (𝑛 − 𝑑)𝑁∗
̂ =
𝑡(𝑧) ∑ 𝑧 u� = ̂𝑧u�̄ .
𝑧u�̄ = 𝑁 (3.8)
𝑛2 u�∈u�
𝑛

A natural estimator of the population mean is

̂
𝑡(𝑧)
𝑧 ̄̂ = = 𝑧u�̄ . (3.9)
𝑁 ̂

This estimator is a so-called ratio estimator: both the numerator and denom-
inator are estimators of totals. See Section 10.2 for more information about
this estimator.
The simple random sample of size 40 selected above is used to estimate the total
mass of soil organic matter (SOM) in the population. First, the population
mean is estimated.
32 3 Simple random sampling

mz <- mean(mysample$z)

The estimated mean SOM concentration is 93.3 g kg-1 . Simply multiplying the
estimated mean by the area 𝐴 to obtain an estimate of the population total is
not very useful, as the dimension of the total then is in g kg-1 m2 . To estimate
the total mass of SOM in the soil layer 0 − 30 cm, first the soil volume in m3
is computed by the total number of grid cells, 𝑁, multiplied by the size of the
grid cells and by the thickness of the soil layer. The total is then estimated by
the product of this volume, the bulk density of soil (1,500 kg m-3 ), and the
estimated population mean (g kg-1 ). This is multiplied by 10-6 to obtain the
total mass of SOM in Mg (1 Mg is 1,000 kg).

vol_soil <- N * 25ˆ2 * 0.3

bd <- 1500
tz <- vol_soil * bd * mz * 10ˆ-6

The estimated total is 197,545 Mg (420 Mg ha-1 ).

Note that a constant bulk density is used. Ideally, this bulk density is
also measured at the sampling points, by collecting soil aliquots of a
constant volume. The measured SOM concentration and bulk density can
then be used to compute the volumetric SOM concentration in kg m-3 at
the sampling points. The estimated population mean of this volumetric
SOM concentration can then be multiplied by the total volume of soil in
the study area, to get an estimate of the total mass of SOM in the study area.

The simulated population is now sampled 10,000 times to see how sampling
affects the estimates. For each sample, the population mean is estimated by
the sample mean. Figure 3.2 shows the approximated sampling distribution
of the 𝜋 estimator of the mean SOM concentration. Note that the sampling
distribution is nearly symmetric, whereas the frequency distribution of the
SOM concentrations in the population is far from symmetric, see Figure 1.4.
The increased symmetry is due to the averaging of 40 numbers.
If we would repeat the sampling an infinite number of times and make the
width of the bins in the histogram infinitely small, then we obtain, after scaling
so that the sum of the area under the curve equals 1, the sampling distribution
of the estimator of the population mean. Important summary statistics of this
sampling distribution are the expectation (mean) and the variance.
When the expectation equals the population mean, there is no systematic error.
The estimator is then said to be design-unbiased. In Chapter 21 another type
of unbiasedness is introduced, model-unbiasedness. The difference between
design-unbiasedness and model-unbiasedness is explained in Chapter 26. In
3.1 Estimation of population parameters 33

FIGURE 3.2: Approximated sampling distribution of the 𝜋 estimator of the

mean SOM concentration (g kg-1 ) in Voorst for simple random sampling of
size 40.

following chapters of Part I unbiased means design-unbiased. Actually, it is

not the estimator which is unbiased, but the combination of a sampling design
and an estimator. For instance, with an equal probability sampling design, the
sample mean is an unbiased estimator of the population mean, whereas it is a
biased estimator in combination with an unequal probability sampling design.
The variance, referred to as the sampling variance, is a measure of the random
error. Ideally, this variance is as small as possible, so that there is a large
probability that for an individual estimate the estimation error is small. The
variance is a measure of the precision of an estimator. An estimator with a
small variance but a strong bias is not a good estimator. To assess the quality
of an estimator, we should look at both. The variance and the bias are often
combined in the mean squared error (MSE), which is the sum of the variance
and the squared bias. An estimator with a small MSE is an accurate estimator.
So, contrary to precision, accuracy also accounts for the bias.
Do not confuse the population variance and the sampling variance. The popu-
lation variance, or spatial variance, is a population characteristic, whereas the
sampling variance is a characteristic of a sampling strategy, i.e., a combina-
tion of a sampling design and an estimator. The sampling variance quantifies
our uncertainty about the population mean. The sampling variance can be
manipulated by changing the sample size 𝑛, the type of sampling design, and
the estimator. This has no effect on the population variance. The average of
34 3 Simple random sampling

the 10,000 estimated population means equals 81.1 g kg-1 , so the difference
with the true population mean equals -0.03 g kg-1 . The variance of the 10,000
estimated population means equals 55.8 (g kg-1 )2 . The square root of this
variance, referred to as the standard error, equals 7.47 g kg-1 . Note that the
standard error has the same units as the study variable, g kg-1 , whereas the
units of the variance are the squared units of the study variable.

3.1.1 Population proportion

In some cases one is interested in the proportion of the population (study area)
satisfying a given condition. Think, for instance, of the proportion of trees in
a forest infected by some disease, the proportion of an area or areal fraction,
in which a soil pollutant exceeds some critical threshold, or the proportion
of an area where habitat conditions are suitable for some endangered species.
Recall that a population proportion is defined as the population mean of
an 0/1 indicator 𝑦 with value 1 if the condition is satisfied, and 0 otherwise
(Subsection 1.1.1). For simple random sampling, this population proportion
can be estimated by the same formula as for the mean (Equation (3.2)):

1
𝑝̂ = ∑𝑦 . (3.10)
𝑛 u�∈u� u�

3.1.2 Cumulative distribution function and quantiles

The population cumulative distribution function (CDF) is defined in Equation

(1.4). A population CDF can be estimated by repeated application of the
indicator technique described in the previous subsection on estimating a
population proportion. A series of threshold values is chosen. Each threshold
results in 𝑛 indicator values having value 1 if the observed study variable 𝑧
of unit 𝑘 is smaller than or equal to the threshold, and 0 otherwise. These
indicator values are then used to estimate the proportion of the population
with a 𝑧-value smaller than or equal to that threshold. For simple random
sampling, these proportions can be estimated with Equation (3.10). Commonly,
the unique 𝑧-values in the sample are used as threshold values, leading to
as many estimated population proportions as there are unique values in the
sample.
Figure 3.3 shows the estimated CDF, estimated from the simple random sample
of 40 units from Voorst. The steps are at the unique values of SOM in the
sample.

ggplot(mysample, mapping = aes(z)) +

stat_ecdf(geom = "step") +
3.1 Estimation of population parameters 35

scale_x_continuous(name = "SOM") +
scale_y_continuous(name = "F")

FIGURE 3.3: Cumulative distribution function of the SOM concentration

(g kg-1 ) in Voorst, estimated from the simple random sample of 40 units.

The estimated population proportions can be used to estimate a population

quantile for any population proportion (cumulative frequency, probability),
for instance the median, first quartile, and third quartile, corresponding to a
population proportion of 0.5, 0.25, and 0.75, respectively. A simple estimator
is the smallest 𝑘th order statistic with an estimated proportion larger than or
equal to the desired cumulative frequency (Hyndman and Fan, 1996).
The estimated CDF shows jumps of size 1/𝑛, so that the estimated population
proportion can be larger than the desired proportion. The estimated population
proportions therefore are often interpolated, for instance linearly. Function
quantile of the stats package can be used to estimate a quantile. With argument
type = 4 linear interpolation is used to estimate the quantiles.

Function quantile actually computes sample quantiles, i.e., it assumes that

the population units are selected with equal inclusion probabilities (as in
simple random sampling), so that the estimators of the population propor-
tions obtained with Equation (3.10) are unbiased. With unequal inclusion
probabilities these probabilities must be accounted for in estimating the
population proportions, see following chapters.
36 3 Simple random sampling

quantile(mysample$z, probs = c(0.25, 0.5, 0.75), type = 4) %>%

round(1)

25% 50% 75%

50.4 65.6 138.7

Note the pipe operator %>% of package magrittr (Bache and Wickham, 2020)
forwarding the result of function quantile to function round.
Package QuantileNPCI (Hutson et al., 2019) can be used to compute a
non-parametric confidence interval estimate of a quantile, using fractional
order statistics (Hutson, 1999). Parameter q specifies the proportion.

library(QuantileNPCI)
res <- quantCI(mysample$z, q = 0.5, alpha = 0.05, method = "exact")

The estimated median equals 66.2 g kg-1 , the lower bound of the 95% confidence
interval equals 54.0 g kg-1 , and the upper bound equals 98.1 g kg-1 .

Exercises

1. Compare the approximated sampling distribution of the 𝜋 estimator

of the mean SOM concentration of Voorst (Figure 3.2) with the
histogram of the 7,528 simulated SOM values (Figure 1.4). Explain
the differences.

2. What happens with the spread in the approximated sampling

distribution (variance of estimated population means) when the
sample size 𝑛 is increased?

3. Suppose we would repeat the sampling 1012 number of times, what

would happen with the difference between the average of the esti-
mated population means and the true population mean?

3.2 Sampling variance of estimator of population

parameters
For simple random sampling of an infinite population and simple random
sampling with replacement of a finite population, the sampling variance of the
3.2 Sampling variance of estimator of population parameters 37

estimator of the population mean equals

𝑆2 (𝑧)
𝑉(𝑧)̄̂ = , (3.11)
𝑛

with 𝑆2 (𝑧) the population variance, also referred to as the spatial variance. For
finite populations, this population variance is defined as (Lohr, 1999)

u�
1
𝑆2 (𝑧) = ∑ (𝑧u� − 𝑧)̄ 2 , (3.12)
𝑁 − 1 u�=1

and for infinite populations as

1
𝑆2 (𝑧) = ∫ (𝑧(𝐬) − 𝑧)̄ 2 d𝐬 , (3.13)
𝐴
u�∈u�

with 𝑧(𝐬) the value of the study variable 𝑧 at a point with two-dimensional
coordinates 𝐬 = (𝑠1 , 𝑠2 ), 𝐴 the area of the study area, and 𝒜 the study area.
In practice, we select only one sample, i.e., we do not repeat the sampling
many times. Still it is possible to estimate the variance of the estimator of
the population mean if we would repeat the sampling. In other words, we can
estimate the sampling variance of the estimator of the population mean from
a single sample. We do so by estimating the population variance from the
sample, and this estimate can then be used to estimate the sampling variance
of the estimator of the population mean. For simple random sampling with
replacement from finite populations, the sampling variance of the estimator of
the population mean can be estimated by

̂
𝑆 2 (𝑧) 1
𝑉(̂ 𝑧)̄̂ = = ∑ (𝑧 − 𝑧u�̄ )2 , (3.14)
𝑛 𝑛 (𝑛 − 1) u�∈u� u�

with 𝑆̂2 (𝑧) the estimated population variance. With simple random sampling,

the sample variance, i.e., the variance of the sample data, is an unbiased
estimator of the population variance. The variance estimator of Equation
(3.14) can also be used for infinite populations. For simple random sampling
without replacement from finite populations, the sampling variance of the
estimator of the population mean can be estimated by

𝑛 𝑆̂2 (𝑧)
𝑉(̂ 𝑧)̄̂ = (1 − ) . (3.15)
𝑁 𝑛
The term 1 − u�
u� is referred to as the finite population correction (fpc).
38 3 Simple random sampling

In the sampling experiment described above, the average of the 10,000 estimated
sampling variances equals 55.7 (g kg-1 )2 . The true sampling variance equals
55.4 (g kg-1 )2 . So, the difference is very small, indicating that the estimator of
the sampling variance, Equation (3.15), is design-unbiased.
The sampling variance of the estimator of the total of a finite population can
be estimated by multiplying the estimated variance of the estimator of the
population mean by 𝑁2 . For simple random sampling without replacement
this estimator thus equals

𝑛 𝑆̂2 (𝑧)
𝑉(̂ 𝑡(𝑧))
̂ = 𝑁2 (1 − ) . (3.16)
𝑁 𝑛
For simple random sampling of infinite populations, the sampling variance of
the estimator of the total can be estimated by

̂
𝑆 2 (𝑧)
𝑉(̂ 𝑡(𝑧))
̂ = 𝐴2 . (3.17)
𝑛
The sampling variance of the estimator of a proportion 𝑝̂ for simple random
sampling without replacement of a finite population can be estimated by

𝑛 𝑝(1
̂ − 𝑝)̂
𝑉(̂ 𝑝)̂ = (1 − ) . (3.18)
𝑁 𝑛−1
The numerator in this estimator is an estimate of the population variance of
the indicator. Note that this estimated population variance is divided by 𝑛 − 1,
and not by 𝑛 as in the estimator of the population mean (Lohr, 1999).
Estimation of the standard error of the estimated population mean in R is
very straightforward. To estimate the standard error of the estimated total in
Mg, the standard error of the estimated population mean must be multiplied
by a constant equal to the product of the soil volume, the bulk density, and
10−6 ; see second code chunk in Section 3.1.

se_mz <- sqrt(var(mysample$z) / n)

se_tz <- se_mz * vol_soil * bd * 10ˆ-6

The estimated standard error of the estimated total equals 20,334 Mg. This
standard error does not account for spatial variation of bulk density.
Although there is no advantage in using package survey (Lumley, 2021) to
compute the 𝜋 estimator and its standard error for this simple sampling
design, I illustrate how this works. For more complex designs and alternative
estimators, estimation of the population mean and its standard error with
functions defined in this package is very convenient, as will be shown in the
following chapters.
3.2 Sampling variance of estimator of population parameters 39

First, the sampling design that is used to select the sampling units is specified
with function svydesign. The first argument specifies the sampling units. In
this case, the centres of the discretisation grid cells are used as sampling
units, which is indicated by the formula id = ~ 1. In Chapter 6 clusters of
population units are used as sampling units, and in Chapter 7 both clusters
and individual units are used as sampling units. Argument probs specifies the
inclusion probabilities of the sampling units. Alternatively, we may specify the
weights with argument weights, which are in this case equal to the inverse of
the inclusion probabilities. Variable pi is a column in tibble mysample, which is
indicated with the tilde in probs = ~ pi.
The population mean is then estimated with function svymean. The first argu-
ment is a formula specifying the study variable. Argument design specifies the
sampling design.

library(survey)
mysample$pi <- n / N
design_si <- svydesign(id = ~ 1, probs = ~ pi, data = mysample)
svymean(~ z, design = design_si)

mean SE
z 93.303 9.6041

For simple random sampling of finite populations without replacement, argu-

ment fpc is used to correct the standard error.

mysample$N <- N
design_si <- svydesign(id = ~ 1, probs = ~ pi, fpc = ~ N, data = mysample)
svymean(~ z, design_si)

mean SE
z 93.303 9.5786

The estimated standard error is smaller now due to the finite population
correction, see Equation (3.15).
Population totals can be estimated with function svytotal, quantiles with
function svyquantile, and ratios of population totals with svyratio, to mention
a few functions that will be used in following chapters.

svyquantile(~ z, design_si, quantile = c(0.5, 0.9))

$z
quantile ci.2.5 ci.97.5 se
0.5 65.56675 53.67764 99.93484 11.43457
0.9 164.36975 153.86258 320.74887 41.25353
40 3 Simple random sampling

attr(,"hasci")
[1] TRUE
attr(,"class")
[1] "newsvyquantile"

Exercises

4. Is the sampling variance for simple random sampling without

replacement larger or smaller than for simple random sampling with
replacement, given the sample size 𝑛? Explain your answer.

5. What is the effect of the population size 𝑁 on this difference?

6. In Section 3.2 the true sampling variance is reported, i.e., the

variance of the estimator of the population mean if we would
repeat the sampling an infinite number of times. How can this true
sampling variance be computed?

7. In reality, we cannot compute the true sampling variance. Why not?

3.3 Confidence interval estimate

A second way of expressing our uncertainty about the estimated total, mean,
or proportion is to present not merely a single number, but an interval. The
wider the interval, the more uncertain we are about the estimate, and vice
versa, the narrower the interval, the more confident we are. To learn how to
compute a confidence interval, I return to the sampling distribution of the
estimator of the mean SOM concentration. Suppose we would like to compute
the bounds of an interval [𝑎, 𝑏] such that 5% of the estimated population means
is smaller than 𝑎, and 5% is larger than 𝑏. To compute the lower bound 𝑎
and the upper bound 𝑏 of this 90% interval, we must specify the distribution
function. When the distribution of the study variable 𝑧 is normal and we know
the variance of 𝑧 in the population, then the sampling distribution of the
estimator of the population mean is also normal, regardless of the sample size.
The larger the sample size, the smaller the effect of the distribution of 𝑧 on the
sampling distribution of the estimator of the population mean. For instance,
even when the distribution of 𝑧 is far from symmetric, then still the sampling
distribution of the estimator of the population mean is approximately normal
if the sample size is large, say 𝑛 > 100. This is the essence of the central limit
3.3 Confidence interval estimate 41

theorem. Above, we already noticed that the sampling distribution is much

less asymmetric than the frequency distribution of the simulated values, and
looks much more like a normal distribution. Assuming a normal distribution,
the bounds of the 90% interval are given by

𝑧 ̄̂ ± 𝑢(0.10/2) ⋅ √𝑉(𝑧)̄̂ , (3.19)

where 𝑢(0.10/2) is the 0.95 quantile of the standard normal distribution, i.e.,
the value of 𝑢 having a tail area of 0.05 to its right. Note that in this equation
the sampling variance of the estimator of the population mean 𝑉(𝑧)̄̂ is used. In
practice, this variance is unknown, because the population variance is unknown,
and must be estimated from the sample (Equations (3.14) and (3.15)). To
account for the unknown sampling variance, the standard normal distribution
is replaced by Student’s t distribution (hereafter shortly referred to as the t
distribution), which has thicker tails than the standard normal distribution.
This leads to the following bounds of the 100(1 − 𝛼)% confidence interval
estimate of the mean:

(3.20)
(u�−1)
𝑧 ̄̂ ± 𝑡u�/2 ⋅ √𝑉(̂ 𝑧)̄̂ ,

where 𝑡u�/2 is the (1 − 𝛼/2) quantile of the t distribution with (𝑛 − 1) degrees

(u�−1)

of freedom. The quantity (1 − 𝛼) is referred to as the confidence level. The

larger the number of degrees of freedom (𝑛 − 1), the closer the t distribution is
to the standard normal distribution. The quantity 𝑡 ⋅ √𝑉(̂ 𝑧)̄̂ is referred
(u�−1)
1−u�/2
to as the margin of error.
Function qt computes a quantile of a t distribution, given the degrees of freedom
and the cumulative probability. The bounds of the confidence interval can then
be computed as follows.

alpha <- 0.05

margin <- qt(1 - alpha / 2, n - 1, lower.tail = TRUE) * se_mz
lower <- mz - margin
upper <- mz + margin

More easily, we can use method confint of package survey to compute the
confidence interval.

confint(svymean(~ z, design_si), df = degf(design_si), level = 0.95)

2.5 % 97.5 %
z 73.92817 112.6771
42 3 Simple random sampling

The interpretation of a confidence interval is not straightforward. A common

misinterpretation is that if the 0.90 confidence interval estimate of the
population mean equals [𝑎, 𝑏], then the probability that the population mean
is in this interval equals 0.90. In classical sampling theory, this cannot be a
correct interpretation, because the population mean is not a random variable,
and consequently the probability that the population mean is in an interval
does not exist. However, the estimated bounds of the confidence interval
are random variables, because the estimated population mean and also the
estimated sampling variance vary among samples drawn with a probability
sampling design. Therefore, it does make sense to attach a probability to
this interval.

Figure 3.4 shows the 90% confidence interval estimates of the mean SOM
concentration for the first 100 simple random samples drawn above. Note that
both the location and the length of the intervals differ between samples. For
each sample, I determined whether this interval covers the population mean.

FIGURE 3.4: Estimated confidence intervals of the mean SOM concentration

(g kg-1 ) in Voorst, estimated from 100 simple random samples of size 40. The
vertical red line is at the true population mean (81.1 g kg-1 ).
3.3 Confidence interval estimate 43

Out of the 10,000 samples, 1,132 samples do not cover the population mean,
i.e., close to the specified 10%. So, a 90% confidence interval is a random
interval that contains in the long run the population mean 90% of the time.

3.3.1 Confidence interval for a proportion

Ideally, a confidence interval for a population proportion is based on the

binomial distribution of the number of sampling units satisfying a condition
(the number of successes). The binomial distribution is a discrete distribution.
There are various methods for computing coverage probabilities of confidence
intervals for a binomial proportion, see Brown et al. (2001) for a discussion.
A common method for computing the confidence interval of a proportion is
the Clopper-Pearson method. Function BinomCI of package DescTools can be
used to compute confidence intervals for proportions (Signorell, 2021).

library(DescTools)
n <- 50
k <- 5
print(p.est <- BinomCI(k, n, conf.level = 0.95, method = "clopper-pearson"))

est lwr.ci upr.ci

[1,] 0.1 0.03327509 0.2181354

The confidence interval is not symmetric around the estimated proportion of

0.1. As can be seen below, the upper bound is the proportion at which the
probability of 5 or fewer successes is 0.025,

pbinom(q = k, size = n, prob = p.est[3])

[1] 0.025

and the lower bound of the confidence interval is the proportion at which the
probability of 5 or more successes is also equal to 0.025. Note that to compute
the upper tail probability, we must assign 𝑘 − 1 = 4 to argument q, because
with argument lower.tail = FALSE function pbinom computes the probability of
𝑋 > 𝑥, not of 𝑋 ≥ 𝑥.

pbinom(q = k - 1, size = n, prob = p.est[2], lower.tail = FALSE)

[1] 0.025

For large sample sizes and for proportions close to 0.5, the confidence interval
can be computed with a normal distribution as an approximation to the
binomial distribution, using Equation (3.18) for the variance estimator of the
44 3 Simple random sampling

estimator of a proportion:

𝑝(1
̂ − 𝑝)̂
𝑝̂ ± 𝑢u�/2 √ . (3.21)
𝑛−1

This interval is referred to as the Wald interval. It is a fact that unless 𝑛 is

very large, the actual coverage probability of the Wald interval is poor for 𝑝
near 0 or 1. A rule of thumb is that the Wald interval should be used only
when 𝑛 ⋅ 𝑚𝑖𝑛{𝑝, (1 − 𝑝)} is at least 5 or 10. For small 𝑛, Brown et al. (2001)
recommend the Wilson interval and for larger 𝑛 the Agresti-Coull interval.
These intervals can be computed with function BinomCI of package DescTools.

3.4 Simple random sampling of circular plots

In forest inventory, vegetation surveys, and agricultural surveys, circular sam-
pling plots are quite common. Using circular plots as sampling units is not
entirely straightforward, because the study area cannot be partitioned into
a finite number of circles that fully cover the study area. The use of circular
plots as sampling units can be implemented in two ways (De Vries, 1986):

1. sampling from a finite set of fixed circles; and

2. sampling from an infinite set of floating circles.

3.4.1 Sampling from a finite set of fixed circles

Sampling from a finite set of fixed circles is simple, but as we will see this
requires an assumption about the distribution of the study variable in the
population. In this implementation, the sampling units consist of a finite set of
slightly overlapping or non-overlapping fixed circular plots (Figure 3.5). The
circles can be constructed as follows. A grid with squares is superimposed
on the study area, so that it fully covers the study area. These squares are
then substituted by circles with an area equal to the area of the squares, or
by non-overlapping tangent circles inscribed in the squares. The radius of the
partly overlapping circles equals √𝑎/𝜋, with√ 𝑎 the area of the squares, the
radius of the non-overlapping circles equals 𝑎/2. In both implementations,
the infinite population is replaced by a finite population of circles that does
not fully tessellate the study area. When using the partly overlapping circles as
sampling units we may avoid overlap by selecting a systematic sample (Chapter
5) of circular plots. The population total can then be estimated by Equation
(3.1), substituting 𝐴/𝑎 for 𝑁, and where 𝑧u� is the total of the 𝑘th circle (sum
of observations of all population units in 𝑘th circle). However, no unbiased
3.4 Simple random sampling of circular plots 45

estimator of the sampling variance of the estimator of the population total or

mean is available for this sampling design, see Chapter 5.

FIGURE 3.5: Simple random sample of ten circular plots from a square
discretised by a finite set of partly overlapping or non-overlapping circular
plots.

With simple random sampling without replacement of non-overlapping circular

plots, the finite population total can be estimated by Equation (3.1) and its
sampling variance by Equation (3.16). However, the circular plots do not cover
the full study area, and as a consequence the total of the infinite population is
underestimated. A corrected estimate can be obtained by estimating the mean
of the finite population and multiplying this estimated population mean by
𝐴/𝑎 (De Vries, 1986):

𝐴 ̂
̂ =
𝑡(𝑧) 𝑧̄, (3.22)
𝑎

with 𝑧 ̄̂ the estimated mean of the finite population. The variance can be
estimated by the variance of the estimator of the mean of the finite population,
multiplied by the square of 𝐴/𝑎. However, we still need to assume that the
mean of the finite population is equal to the mean of the infinite population.
This assumption can be avoided by sampling from an infinite set of floating
circles.

3.4.2 Sampling from an infinite set of floating circles

A simple random sample of floating circular plots can be selected by simple

random sampling of the centres of the plots. The circular plots overlap if two
selected points are separated by a distance smaller than the diameter of the
46 3 Simple random sampling

circular plots. Besides, when a plot is selected near the border of the study area,
a part of the plot is outside the study area. This part is ignored in estimating
the population mean or total. To select the centres, the study area must be
extended by a zone with a width equal to the radius of the circular plots. This
is illustrated in Figure 3.6, showing a square study area of 100 m × 100 m.
To select ten circular plots with a radius of 5 m from this square, ten points
are selected by simple random sampling, using function runif, with -5 as lower
limit and 105 as upper limit of the uniform distribution.

set.seed(129)
s1 <- runif(10, min = -5, max = 105)
s2 <- runif(10, min = -5, max = 105)

Two points are selected outside the study area, in the extended zone. For both
points, a small part of the circular plot is inside the square. To determine the
study variable for these two sampling units, only the part of the plot inside
the square is observed. In other words, these two observations have a smaller
support than the observations of the other eight plots, see Chapter 1.

FIGURE 3.6: Simple random sample of ten floating circular plots from a
square.
3.4 Simple random sampling of circular plots 47

In the upper left corner, two sampling units are selected that largely overlap.
The intersection of the two circular plots is used twice, to determine the study
variable of both sampling units.
Given the observations of the selected circular plots, the population total can
be estimated by (De Vries, 1986)

𝐴1
̂ =
𝑡(𝑧) ∑𝑧 , (3.23)
𝑎 𝑛 u�∈u� u�

with 𝑎 the area of the circle and 𝑧u� the observed total of sampling unit 𝑘
(circle). The same estimate of the total is obtained if we divide the observations
by 𝑎 to obtain a mean per sampling unit:

1 𝑧
̂ =𝐴
𝑡(𝑧) ∑ u� . (3.24)
𝑛 u�∈u� 𝑎

The sampling variance of the estimator of the total can be estimated by

̂
𝐴 2𝑆2 (𝑧)
𝑉(̂ 𝑡(𝑧))
̂ =( ) , (3.25)
𝑎 𝑛

with 𝑆̂ 2 (𝑧) the estimated population variance of the totals per population unit

(circle).

Exercises

8. Write an R script to select a simple random sample of size 40 from

Voorst.
• Use the selected sample to estimate the population mean of
the SOM concentration (𝑧 in the data frame) and its standard
error.
• Compute the lower and the upper bound of the 90% confi-
dence interval using the t distribution, and check whether the
population mean is covered by the interval.
• Compare the length of the 90% confidence interval with the
length of the 95% interval. Explain the difference in width.
• Use the selected sample to estimate the total mass of SOM in
Mg in the topsoil (0 − 30 cm) of Voorst. Use as a bulk density
1,500 kg m-3 . The size of the grid cells is 25 m × 25 m.
• Estimate the standard error of the estimated total.
• Do you think this standard error is a realistic estimate of the
uncertainty about the estimated total?
4
Stratified simple random sampling

In stratified random sampling the population is divided into subpopulations,

for instance, soil mapping units, areas with the same land use or land cover,
administrative units, etc. The subareas are mutually exclusive, i.e., they do not
overlap, and are jointly exhaustive, i.e., their union equals the entire population
(study area). Within each subpopulation, referred to as a stratum, a probability
sample is selected by some sampling design. If these probability samples are
selected by simple random sampling, as described in the previous chapter,
the design is stratified simple random sampling, the topic of this chapter. If
sampling units are selected by cluster random sampling, then the design is
stratified cluster random sampling.
Stratified simple random sampling is illustrated with Voorst (Figure 4.1).
Tibble grdVoorst with simulated data contains variable stratum. The strata are
combinations of soil class and land use, obtained by overlaying a soil map and
a land use map. To select a stratified simple random sample, we set the total
sample size 𝑛. The sampling units must be apportioned to the strata. I chose
to apportion the units proportionally to the size (area, number of grid cells) of
the strata (for details, see Section 4.3). The larger a stratum, the more units
are selected from this stratum. The sizes of the strata, i.e., the total number
of grid cells, are computed with function tapply.

N_h <- tapply(grdVoorst$stratum, INDEX = grdVoorst$stratum, FUN = length)

w_h <- N_h / sum(N_h)
n <- 40
print(n_h <- round(n * w_h))

BA EA PA RA XF
13 8 9 4 7

The sum of the stratum sample sizes is 41; we want 40, so we reduce the largest
stratum sample size by 1.

n_h[1] <- n_h[1] - 1

The stratified simple random sample is selected with function strata of package
sampling (Tillé and Matei, 2021). Argument size specifies the stratum sample
sizes.

49
50 4 Stratified simple random sampling

The stratum sample sizes must be in the order the strata are encountered in
tibble grdVoorst, which is determined first with function unique.

Within the strata, the grid cells are selected by simple random sampling with
replacement (method = "srswr"), so that in principle more than one point can
be selected within a grid cell, see Chapter 3 for a motivation of this. Function
getdata extracts the observations of the selected units from the sampling
frame, as well as the spatial coordinates and the stratum of these units. The
coordinates of the centres of the selected grid cells are jittered by an amount
equal to half the side of the grid cells. In the next code chunk, this is done
with function mutate of package dplyr (Wickham et al., 2021) which is part of
package tidyverse (Wickham et al., 2019). We have seen the pipe operator %>%
of package magrittr (Bache and Wickham, 2020) before in Subsection 3.1.2.
If you are not familiar with tidyverse I recommend reading the excellent book
R for Data Science (Wickham and Grolemund, 2017).

library(sampling)
ord <- unique(grdVoorst$stratum)
set.seed(314)
units <- sampling::strata(
grdVoorst, stratanames = "stratum", size = n_h[ord], method = "srswr")
mysample <- getdata(grdVoorst, units) %>%
mutate(s1 = s1 %>% jitter(amount = 25 / 2),
s2 = s2 %>% jitter(amount = 25 / 2))

The name of the package is added to function strata (sampling::strata), as

strata is also a function of another package. Not adding the name of the
package may result in an error message.

Figure 4.1 shows the selected sample.

FIGURE 4.1: Stratified simple random sample of size 40 from Voorst. Strata
are combinations of soil class and land use.
4.1 Estimation of population parameters 51

4.1 Estimation of population parameters

With simple random sampling within strata, the estimator of the population
mean for simple random sampling (Equation (3.2)) is applied at the level of
the strata. The estimated stratum means are then averaged, using the relative
sizes or areas of the strata as weights:

u�
𝑧 ̄̂ = ∑ 𝑤ℎ 𝑧ℎ̄̂ , (4.1)
ℎ=1

where 𝐻 is the number of strata, 𝑤ℎ is the relative size (area) of stratum ℎ

(stratum weight): 𝑤ℎ = 𝑁ℎ /𝑁, and 𝑧ℎ̄̂ is the estimated mean of stratum ℎ,
estimated by the sample mean for stratum ℎ:

1
𝑧ℎ̄̂ = ∑ 𝑧 , (4.2)
𝑛ℎ u�∈u� u�
ℎ

with 𝒮ℎ the sample selected from stratum ℎ.

The same estimator is found when the 𝜋 estimator is worked out for stratified
simple random sampling. With stratified simple random sampling without
replacement and different sampling fractions for the strata, the inclusion
probabilities differ among the strata and equal 𝜋u� = 𝑛ℎ /𝑁ℎ for all 𝑘 in
stratum ℎ, with 𝑛ℎ the sample size of stratum ℎ and 𝑁ℎ the size of stratum ℎ.
Inserting this in the 𝜋 estimator of the population mean (Equation (2.4)) gives

1 u� 𝑧 1 u� 𝑁ℎ u�
𝑧 ̄̂ = ∑ ∑ u� = ∑ ∑ 𝑧u� = ∑ 𝑤ℎ 𝑧ℎ̄̂ . (4.3)
𝑁 ℎ=1 u�∈u� 𝜋u� 𝑁 ℎ=1 𝑛ℎ u�∈u� ℎ=1
ℎ ℎ

The sampling fractions are usually slightly different, even with proportional
allocation (Section 4.3), because 𝑛ℎ /𝑁ℎ cannot be made exactly equal for
all strata. Sample sizes necessarily are integers, so 𝑛ℎ /𝑁ℎ must be rounded
to integers.

The sampling variance of the estimator of the population mean is estimated by

first estimating the sampling variances of the estimated stratum means, followed
by computing the weighted average of the estimated sampling variances of the
estimated stratum means. Note that we must square the stratum weights:

u�
𝑉(̂ 𝑧)̄̂ = ∑ 𝑤2ℎ 𝑉(̂ 𝑧ℎ̄̂ ) , (4.4)
ℎ=1
52 4 Stratified simple random sampling

TABLE 4.1: Size (Nh), sample size (nh), estimated mean (Mean), estimated
variance (Variance), and estimated standard error of estimator of mean (se) of
the five strata in Voorst.

Stratum Nh nh Mean Variance se

BA 2,371 12 91.1 946.3 8.9
EA 1,442 8 58.3 555.5 8.3
PA 1,710 9 59.4 214.7 4.9
RA 659 4 103.2 2,528.8 25.1
XF 1,346 7 133.9 3,807.3 23.3

where 𝑉(̂ 𝑧ℎ̄̂ ) is the estimated sampling variance of 𝑧ℎ̄̂ :

𝑛 𝑆 ̂2 (𝑧)
𝑉(̂ 𝑧ℎ̄̂ ) = (1 − ℎ ) ℎ , (4.5)
𝑁ℎ 𝑛ℎ

with 𝑆
̂2 (𝑧) the estimated variance of 𝑧 within stratum ℎ:
ℎ

1 2
̂
𝑆2 (𝑧) =
ℎ ∑ (𝑧u� − 𝑧ℎ̄̂ ) . (4.6)
𝑛ℎ − 1 u�∈u�
ℎ

For stratified simple random sampling with replacement of finite populations

and stratified simple random sampling of infinite populations the fpcs 1 −
(𝑛ℎ /𝑁ℎ ) can be dropped.

mz_h <- tapply(mysample$z, INDEX = mysample$stratum, FUN = mean)

mz <- sum(w_h * mz_h)
S2z_h <- tapply(mysample$z, INDEX = mysample$stratum, FUN = var)
v_mz_h <- S2z_h / n_h
se_mz <- sqrt(sum(w_hˆ2 * v_mz_h))

Table 4.1 shows per stratum the estimated mean, variance, and sampling
variance of the estimated mean of the SOM concentration. We can see large
differences in the within-stratum variances. For the stratified sample of Figure
4.1, the estimated population mean equals 86.3 g kg-1 , and the estimated
standard error of this estimator equals 5.8 g kg-1 .
The population mean can also be estimated directly using the basic 𝜋 estimator
(Equation (2.4)). The inclusion probabilities are included in data.frame mysample,
obtained with function getdata (see code chunk above), as variable Prob.

head(mysample)
4.1 Estimation of population parameters 53

s1 s2 z stratum ID_unit Prob Stratum

1135 202554.8 464556.7 186.99296 XF 1135 0.005189017 1
2159 204305.5 464738.9 75.04809 XF 2159 0.005189017 1
4205 203038.3 465057.0 138.14617 XF 4205 0.005189017 1
4503 202381.1 465096.7 69.43803 XF 4503 0.005189017 1
5336 203610.4 465237.8 78.02003 XF 5336 0.005189017 1
5853 205147.1 465315.5 164.55224 XF 5853 0.005189017 1

The population total is estimated first, and by dividing this estimated total by
the total number of population units 𝑁 an estimate of the population mean is
obtained.

tz <- sum(mysample$z / mysample$Prob)

print(mz <- tz / sum(N_h))

[1] 86.53333

The two estimates of the population mean are not exactly equal. This is due to
rounding errors in the inclusion probabilities. This can be shown by computing
the sum of the inclusion probabilities over all population units. This sum
should be equal to the sample size 𝑛 = 40, but as we can see below, this sum
is slightly smaller.

pi_h <- tapply(mysample$Prob, INDEX = mysample$stratum, FUN = unique)

print(sum(pi_h * N_h))

[1] 39.90711

Now suppose we ignore that the sample data come from a stratified sampling
design and we use the (unweighted) sample mean as an estimate of the
population mean.

print(mean(mysample$z))

[1] 86.11247

The sample mean slightly differs from the proper estimate of the population
mean (7.238). The sample mean is a biased estimator, but the bias is small. The
bias is only small because the stratum sample sizes are about proportional to
the sizes of the strata, so that the inclusion probabilities (sampling intensities)
are about equal for all strata: 0.0050494, 0.0055344, 0.0052509, 0.006056,
0.005189. The probabilities are not exactly equal because the stratum sample
sizes are necessarily rounded to integers and because we reduced the largest
sample size by one unit. The bias would have been substantially larger if an
equal number of units would have been selected from each stratum, leading
to much larger differences in the inclusion probabilities among the strata.
Sampling intensity in stratum BA, for instance, then would be much smaller
54 4 Stratified simple random sampling

compared to the other strata, and so would be the inclusion probabilities of the
units in this stratum as compared to the other strata. Stratum BA then would
be underrepresented in the sample. This is not a problem as long as we account
for the difference in inclusion probabilities of the units in the estimation of the
population mean. The estimated mean of stratum BA then gets the largest
weight, equal to the inverse of the inclusion probability. If we do not account
for these differences in inclusion probabilities, the estimator of the mean will
be seriously biased.
The next code chunk shows how the population mean and its standard error
can be estimated with package survey (Lumley, 2021). Note that the stratum
weights 𝑁ℎ /𝑛ℎ must be passed to function svydesign using argument weight.
These are first attached to data.frame mysample by creating a look-up table lut,
which is then merged with function merge to data.frame mysample.

library(survey)
labels <- sort(unique(mysample$stratum))
lut <- data.frame(stratum = labels, weight = N_h / n_h)
mysample <- merge(x = mysample, y = lut)
design_stsi <- svydesign(
id = ~ 1, strata = ~ stratum, weight = ~ weight, data = mysample)
svymean(~ z, design_stsi)

mean SE
z 86.334 5.8167

4.1.1 Population proportion, cumulative distribution

function, and quantiles

The proportion of a population satisfying some condition can be estimated

by Equations (4.1) and (4.2), substituting for the study variable 𝑧u� an 0/1
indicator 𝑦u� with value 1 if for unit 𝑘 the condition is satisfied, and 0 otherwise
(Subsection 3.1.1). In general, with stratified simple random sampling the
inclusion probabilities are not exactly equal, so that the estimated population
proportion is not equal to the sample proportion.
These unequal inclusion probabilities must also be accounted for when esti-
mating the cumulative distribution function (CDF) and quantiles (Subsection
3.1.2), as shown in the next code chunk for the CDF.

thresholds <- sort(unique(mysample$z))

cumfreq <- numeric(length = length(thresholds))
for (i in seq_len(length(thresholds))) {
ind <- mysample$z <= thresholds[i]
4.1 Estimation of population parameters 55

mh_ind <- tapply(ind, INDEX = mysample$stratum, FUN = mean)

cumfreq[i] <- sum(w_h * mh_ind)
}
df <- data.frame(x = thresholds, y = cumfreq)

Figure 4.2 shows the estimated CDF, estimated from the stratified simple
random sample of 40 units from Voorst (Figure 4.1).

FIGURE 4.2: Estimated cumulative distribution function of the SOM con-

centration (g kg-1 ) in Voorst, estimated from the stratified simple random
sample of 40 units.

The estimated proportions, or cumulative frequencies, are used to estimate

a quantile. These estimates are easily obtained with function svyquantile of
package survey.

svyquantile(~ z, design_stsi, quantile = c(0.5, 0.8))

$z
quantile ci.2.5 ci.97.5 se
0.5 69.56081 65.70434 84.03993 4.515916
0.8 117.73877 102.75359 161.88611 14.563887

attr(,"hasci")
[1] TRUE
attr(,"class")
[1] "newsvyquantile"
56 4 Stratified simple random sampling

4.1.2 Why should we stratify?

There can be two reasons for stratifying the population:

1. we are interested in the mean or total per stratum; or

2. we want to increase the precision of the estimated mean or total for
the entire population.

Figure 4.3 shows boxplots of the approximated sampling distributions of the

𝜋 estimator of the mean SOM concentration for stratified simple random
sampling and simple random sampling, both of size 40, obtained by repeating
the random sampling with each design and estimation 10,000 times.

FIGURE 4.3: Approximated sampling distribution of the 𝜋 estimator of

the mean SOM concentration (g kg-1 ) in Voorst for stratified simple random
sampling (STSI) and simple random sampling (SI) of size 40.

The approximated sampling distributions of the estimators of the population

mean with the two designs are not very different. With stratified random sam-
pling, the spread of the estimated means is somewhat smaller. The horizontal
red line is the population mean (81.1 g kg-1 ). The gain in precision due to
the stratification, referred to as the stratification effect, can be quantified by
the ratio of the variance with simple random sampling and the variance with
stratified simple random sampling. So, when this variance ratio is larger than
1, stratified simple random sampling is more precise than simple random sam-
pling. For Voorst the stratification effect with proportional allocation (Section
4.3) equals 1.310. This means that with simple random sampling we need 1.310
4.1 Estimation of population parameters 57

more sampling units than with stratified simple random sampling to obtain
an estimate of the same precision.
The stratification effect can be computed from the population variance 𝑆2 (𝑧)
(Equation (3.12)) and the variances within the strata 𝑆2ℎ (𝑧). In the sampling
experiment, these variances are known without error because we know the
𝑧-values for all units in the population. In practice, we only know the 𝑧-values
for the sampled units. However, a design-unbiased estimator of the population
variance is (de Gruijter et al., 2006)

̂2 2
̂
𝑆 2 (𝑧) = 𝑧 − (𝑧)̄̂ + 𝑉(̂ 𝑧)̄̂ , (4.7)

̂
where 𝑧2 denotes the estimated population mean of the study variable squared
(𝑧 ), obtained in the same way as 𝑧 ̄̂ (Equation (4.1)), but using squared values,
2

and 𝑉(̂ 𝑧)̄̂ denotes the estimated variance of the estimator of the population
mean (Equation (4.4)).
The estimated population variance is then divided by the sum of the stratum
sample sizes to get an estimate of the sampling variance of the estimator of
the mean with simple random sampling of an equal number of units:

̂
𝑆2 (𝑧)
𝑉(̂ 𝑧SI
̄̂ ) = . (4.8)
∑u� 𝑛
ℎ=1 ℎ

The population variance can be estimated with function s2 of package survey-

planning (Breidaks et al., 2020). However, this function is an implementation
of an alternative, consistent estimator of the population variance (Särndal
et al., 1992):

𝑁−1 𝑛 1 (𝑧 − 𝑧u�̄̂ )2
̂
𝑆2 (𝑧) = ∑ u� . (4.9)
𝑁 𝑛 − 1 𝑁 − 1 u�∈u� 𝜋u�

An estimator is consistent if it converges in probability to the true value of

the parameter as the sample size tends to infinity (Särndal et al., 1992).

library(surveyplanning)
S2z <- s2(mysample$z, w = mysample$weight)

The design effect is defined as the variance of an estimator of the population

mean with the sampling design under study divided by the variance of the 𝜋
estimator of the mean with simple random sampling of an equal number of
units (Section 12.4). So, the design effect of stratified random sampling is the
58 4 Stratified simple random sampling

reciprocal of the stratification effect. For the stratified simple random sample
of Figure 4.1, the design effect can then be estimated as follows. Function SE
extracts the estimated standard error of the estimator of the mean from the
output of function svymean. The extracted standard error is then squared to
obtain an estimate of the sampling variance of the estimator of the population
with stratified simple random sampling. Finally, this variance is divided by
the variance with simple random sampling of an equal number of units.

v_mz_SI <- S2z / n

res <- svymean(~ z, design_stsi)
SE(res)ˆ2 / v_mz_SI

z
z 0.6903965

The same value is obtained with argument deff of function svymean.

design_stsi <- svydesign(

id = ~ 1, strata = ~ stratum, weight = ~ weight, data = mysample)
svymean(~ z, design_stsi, deff = "replace")

mean SE DEff
z 86.3340 5.8167 0.6904

So, when using package survey, estimation of the population variance is not
needed to estimate the design effect. I only added this to make clear how
the design effect is computed with functions in package survey. In following
chapters I will skip the estimation of the population variance.
The estimated design effect as estimated from the stratified sample is smaller
than 1, showing that stratified simple random sampling is more efficient than
simple random sampling. The reciprocal of the estimated design effect (1.448)
is somewhat larger than the stratification effect as computed in the sampling
experiment, but this is an estimate of the design effect from one stratified
sample only. The estimated population variance varies among stratified samples,
and so does the estimated design effect.
Stratified simple random sampling with proportional allocation (Section 4.3)
is more precise than simple random sampling when the sum of squares of the
stratum means is larger than the sum of squares within strata (Lohr, 1999):

𝑆𝑆𝐵 > 𝑆𝑆𝑊 , (4.10)

with SSB the weighted sum-of-squares between the stratum means:

u�
𝑆𝑆𝐵 = ∑ 𝑁ℎ (𝑧ℎ̄ − 𝑧)̄ 2 , (4.11)
ℎ=1
4.2 Confidence interval estimate 59

and SSW the sum over the strata of the weighted variances within strata
(weights equal to 1 − 𝑁ℎ /𝑁):

u�
𝑁ℎ 2
𝑆𝑆𝑊 = ∑ (1 − )𝑆 . (4.12)
ℎ=1
𝑁 ℎ

In other words, the smaller the differences in the stratum means and the larger
the variances within the strata, the smaller the stratification effect will be.
Figure 4.4 shows a boxplot of the SOM concentration per stratum (soil-land
use combination). The stratum means are equal to 83.0, 49.0, 68.8, 92.7, 122.3
g kg-1 . The stratum variances are 1799.2, 238.4, 1652.9, 1905.4, 2942.8 (g kg-1 )2 .
The large stratum variances explain the modest gain in precision realised by
stratified simple random sampling compared to simple random sampling in
this case.

FIGURE 4.4: Boxplots of the SOM concentration (g kg-1 ) for the five strata
(soil-land use combinations) in Voorst.

4.2 Confidence interval estimate

The 100(1 − 𝛼)% confidence interval for 𝑧 ̄ is given by

𝑧 ̄̂ ± 𝑡u�/2,u�u� ⋅ √𝑉(̂ 𝑧)̄̂ , (4.13)

60 4 Stratified simple random sampling

where 𝑡u�/2,u�u� is the (1 − 𝛼/2) quantile of a t distribution with 𝑑𝑓 degrees

of freedom. The degrees of freedom 𝑑𝑓 can be approximated by 𝑛 − 𝐻, as
proposed by Lohr (1999). This is the number of the degrees of freedom if the
variances within the strata are equal. With unequal variances within strata, 𝑑𝑓
can be approximated by Sattherwaite’s method (Nanthakumar and Selvavel,
2004):

̂2
2
(∑u� 𝑤2 u� ℎ (u�) )
ℎ=1 ℎ u�
(4.14)
ℎ
𝑑𝑓 ≈ 2
.
u� ̂2
∑ℎ=1 𝑤4ℎ ( u� u�ℎ (u�) ) 1
u�ℎ −1
ℎ

A confidence interval estimate of the population mean can be extracted with

method confint of package survey. It uses 𝑛 − 𝐻 degrees of freedom.

res <- svymean(~ z, design_stsi)

df_stsi <- degf(design_stsi)
confint(res, df = df_stsi, level = 0.95)

2.5 % 97.5 %
z 74.52542 98.14252

4.3 Allocation of sample size to strata

After we have decided on the total sample size 𝑛, we must decide how to
apportion the units to the strata. It is reasonable to allocate more sampling
units to large strata and fewer to small strata. The simplest way to achieve
this is proportional allocation:

𝑁ℎ
𝑛ℎ = 𝑛 ⋅ , (4.15)
∑ 𝑁ℎ

with 𝑁ℎ the total number of population units (size) of stratum ℎ. With infinite
populations 𝑁ℎ is replaced by the area 𝐴ℎ . The sample sizes computed with
this equation are rounded to the nearest integers.
If we have prior information on the variance of the study variable within the
strata, then it makes sense to account for differences in variance. Heterogeneous
strata should receive more sampling units than homogeneous strata, leading
4.3 Allocation of sample size to strata 61

to Neyman allocation:

𝑁ℎ 𝑆ℎ (𝑧)
𝑛ℎ = 𝑛 ⋅ u�
, (4.16)
∑ 𝑁ℎ 𝑆ℎ (𝑧)
ℎ=1

with 𝑆ℎ (𝑧) the standard deviation (square root of variance) of the study
variable 𝑧 in stratum ℎ.
Finally, the costs of sampling may differ among the strata. It can be relatively
expensive to sample nearly inaccessible strata, and we do not want to sample
many units there. This leads to optimal allocation:

u�ℎ u�ℎ (u�)

√
u�ℎ
𝑛ℎ = 𝑛 ⋅ u�
, (4.17)
∑ u�ℎ√u�u�ℎ (u�)
ℎ
ℎ=1

with 𝑐ℎ the costs per sampling unit in stratum ℎ. Optimal means that given the
total costs this allocation type leads to minimum sampling variance, assuming
a linear costs model:

u�
𝐶 = 𝑐 0 + ∑ 𝑛 ℎ 𝑐ℎ , (4.18)
ℎ=1

with 𝑐0 overhead costs. So, the more variable a stratum and the lower the
costs, the more units will be selected from this stratum.

S2z_h <- tapply(X = grdVoorst$z, INDEX = grdVoorst$stratum, FUN = var)

n_h_Neyman <- round(n * N_h * sqrt(S2z_h) / sum(N_h * sqrt(S2z_h)))

These optimal sample sizes can be computed with function optsize of package
surveyplanning.

labels <- sort(unique(mysample$stratum))

res <- optsize(labels, n, N_h, S2z_h)
round(res$nh, 0)

[1] 14 3 9 4 10

Table 4.2 shows the proportional and optimal sample sizes for the five strata of
the study area Voorst, for a total sample size of 40. Stratum XF is the one-but-
smallest stratum and therefore receives only seven sampling units. However,
the standard deviation in this stratum is the largest, and as a consequence
with optimal allocation the sample size in this stratum is increased by three
points, at the cost of stratum EA which is relatively homogeneous.
62 4 Stratified simple random sampling

TABLE 4.2: Proportional and Neyman sample sizes in stratified simple

random sampling of Voorst with a total sample size of 40.

Stratum Nh Sh nhprop nhNeyman

BA 2,371 42.4 12 14
EA 1,442 15.4 8 3
PA 1,710 40.7 9 9
RA 659 43.7 4 4
XF 1,346 54.2 7 10
Nh: stratum size; Sh: stratum standard
deviation.

Figure 4.5 shows the standard error of the 𝜋 estimator of the mean SOM con-
centration as a function of the total sample size, for simple random sampling
and for stratified simple random sampling with proportional and Neyman allo-
cation. A small extra gain in precision can be achieved using Neyman allocation
instead of proportional allocation. However, in practice often Neyman allo-
cation is not achievable, because we do not know the standard deviations of
the study variable within the strata. If a quantitative covariate 𝑥 is used for
stratification (see Sections 4.4 and 13.2), the standard deviations 𝑆ℎ (𝑧) are
approximated by 𝑆ℎ (𝑥), resulting in approximately optimal stratum sample
sizes. The gain in precision compared to proportional allocation is then partly
or entirely lost.
Optimal allocation and Neyman allocation assume univariate stratification, i.e.,
the stratified simple random sample is used to estimate the mean of a single
study variable. If we have multiple study variables, optimal allocation becomes
more complicated. In Bethel allocation, the total sampling costs, assuming a
linear costs model (Equation (4.18)), are minimised given a constraint on the
precision of the estimated mean for each study variable (Bethel, 1989), see
Section 4.8. Bethel allocation can be computed with function bethel of package
SamplingStrata (Barcaroli et al., 2020).

Exercises

1. Use function strata of package sampling to select a stratified simple

random sample with replacement of size 40 from Voorst, using
proportional allocation. Check that the sum of the stratum sample
sizes is 40.
• Estimate the population mean and the standard error of the
estimator.
• Compute the true standard error of the estimator. Hint: com-
pute the population variances of the study variable 𝑧 per stra-
tum, and divide these by the stratum sample sizes.
4.3 Allocation of sample size to strata 63

FIGURE 4.5: Standard error of the 𝜋 estimator of the mean SOM concentra-
tion (g kg-1 ) as a function of the total sample size, for simple random sampling
(SI) and for stratified simple random sampling with proportional (STSI(prop))
and Neyman allocation (STSI(Neyman)) for Voorst.

• Compute a 95% confidence interval estimate of the population

mean, using function confint of package survey.

2. Looking at Figure 4.4, which strata do you expect can be merged

without losing much precision of the estimated population mean?

3. Use function fct_collapse of package forcats (Wickham, 2021) to

merge the strata EA and PA.
• Compute the true sampling variance of the estimator of the
mean for this new stratification, for a total sample size of 40
and proportional allocation.
• Compare this true sampling variance with the true sampling
variance using the original five strata (same sample size,
proportional allocation). What is your conclusion about the
new stratification?

4. Proof that the sum of the inclusion probabilities over all population
units with stratified simple random sampling equals the sample size
𝑛.
64 4 Stratified simple random sampling

4.4 Cum-root-f stratification

When we have a quantitative covariate 𝑥 related to the study variable 𝑧 and 𝑥
is known for all units in the population, strata can be constructed with the
cum-root-f method using this covariate as a stratification variable, see Dalenius
and Hodges (1959) and Cochran (1977). Population units with similar values
for the covariate (stratification variable) are grouped into a stratum. Strata
are computed as follows:

1. Compute a frequency histogram of the stratification variable using

a large number of bins.

2. Compute the square root of the frequencies.

3. Cumulate the square root of the frequencies, i.e., compute √𝑓1 ,

√𝑓1 + √𝑓2 , √𝑓1 + √𝑓2 + √𝑓3 , etc.

4. Divide the cumulative sum of the last bin by the number of strata,
multiply this value by 1, 2, … , 𝐻 − 1, with H the number of strata,
and select the boundaries of the histogram bins closest to these
values.

In cum-root-f stratification, it is assumed that the covariate values are nearly

perfect predictions of the study variable, so that the prediction errors do not
affect the stratification. Under this assumption the stratification is optimal.
Cum-root-f stratification is illustrated with the data of Xuancheng in China.
We wish to estimate the mean organic matter concentration in the topsoil
(SOM, g kg-1 ) of this area. Various covariates are available that are correlated
with SOM, such as elevation, yearly average temperature, slope, and various
other terrain attributes. Elevation, the name of this variable in the tibble is
dem, is used as as a single stratification variable, see Figure 4.6. The correlation
coefficient of SOM and elevation in a sample of 183 observations is 0.59. The
positive correlation can be explained as follows. Temperature is decreasing
with elevation, leading to a smaller decomposition rate of organic matter in
the soil.
The strata can be constructed with the package stratification (Baillargeon
and Rivest, 2011). Care should be taken that the data are sorted in ascending
order by the variable used for stratification, see help of function strata.cumrootf.
Argument n of this function is the total sample size, but this value has no
effect on the stratification. Argument Ls is the number of strata. I arbitrarily
chose to construct five strata. Argument nclass is the number of bins of the
histogram. The output object of function strata.cumrootf is a list containing
4.4 Cum-root-f stratification 65

FIGURE 4.6: Elevation used as a stratification variable in cum-root-f strati-

fication of Xuancheng.

amongst others a numeric vector with the stratum bounds (bh) and a factor
with the stratum levels of the grid cells (stratumID). Finally, note that the
values of the stratification variable must be positive. The minimum elevation
is -5 m, so I added the absolute value of this minimum to elevation.

library(stratification)
grdXuancheng <- grdXuancheng %>%
arrange(dem) %>%
mutate(dem_new = dem + abs(min(dem)))
crfstrata <- strata.cumrootf(
x = grdXuancheng$dem_new, n = 100, Ls = 5, nclass = 500)
bh <- crfstrata$bh
grdXuancheng$crfstrata <- crfstrata$stratumID

Stratum bounds are threshold values of the stratification variable elevation;

these stratum bounds are equal to 46.7, 108.3, 214.5, 384.4. Note that the
number of stratum bounds is one less than the number of strata. The resulting
stratification is shown in Figure 4.7. Note that most strata are not single
polygons, but are made up of many smaller polygons. This may be even more
so if the stratification variable shows a noisy spatial pattern. This is not a
problem at all, as a stratum is just a collection of population units (raster
cells) and need not be spatially contiguous.
66 4 Stratified simple random sampling

FIGURE 4.7: Stratification of Xuancheng obtained with the cum-root-f

method, using elevation as a stratification variable.

Exercises

5. Write an R script to compute five cum-root-f strata for Eastern

Amazonia (grdAmazonia in package sswr) to estimate the population
mean of aboveground biomass (AGB), using log-transformed short-
wave infrared radiation (SWIR2) as a stratification variable.
• Compute ten cum-root-f strata, using function strata of pack-
age sampling. Sort the units first in ascending order on
lnSWIR2. Use the stratum sample sizes as computed by func-
tion strata.cumrootf. What allocation is used for computing the
stratum sample sizes?
• Select a stratified simple random sample of 100 units. First,
compute the stratum sample sizes for proportional allocation.
• Estimate the population mean of AGB and its sampling vari-
ance.
• Compute the true sampling variance of the estimator of the
mean for this sampling design (see Exercise 1 for a hint).
• Compute the stratification effect (gain in precision). Hint: com-
pute the sampling variance for simple random sampling by
computing the population variance of AGB, and divide this by
the total sample size.
4.5 Stratification with multiple covariates 67

4.5 Stratification with multiple covariates

If we have multiple variables that are possibly related to the study variable, we
may want to use them all or a subset of them as stratification variables. Using
the quantitative variables one-by-one in cum-root-f stratification, followed
by overlaying the maps with univariate strata, may lead to numerous cross-
classification strata.
A simple solution is to construct homogeneous groups, referred to as clusters,
of population units (raster cells). The units within a cluster are more similar
to each other than to the units in other clusters. Various clustering techniques
are available. Here, I use hard k-means.
This is illustrated again with the Xuancheng case study. Five quantitative
covariates are used for constructing the strata. Besides elevation, which was used
as a single stratification variable in the previous section, now also temperature,
slope, topographic wetness index (twi), and profile curvature are used to
construct clusters that are used as strata in stratified simple random sampling.
To speed up the computations, a subgrid with a spacing of 0.4 km is selected,
using function spsample of package sp, see Chapter 5 (Bivand et al., 2013).

library(sp)
gridded(grdXuancheng) <- ~ s1 + s2
subgrd <- spsample(
grdXuancheng, type = "regular", cellsize = 400, offset = c(0.5, 0.5))
subgrd <- data.frame(coordinates(subgrd), over(subgrd, grdXuancheng))

Five clusters are computed with k-means using as clustering variables the five
covariates mentioned above. The scale of these covariates is largely different,
and for this reason they must be scaled before being used in clustering. The
k-means algorithm is a deterministic algorithm, i.e., the same initial clustering
will end in the same final, optimised clustering. This final clustering can be
suboptimal, and therefore it is recommended to repeat the clustering as many
times as feasible, with different initial clusterings. Argument nstart is the
number of initial clusterings. The best clustering, i.e., the one with the smallest
within-cluster sum-of-squares, is kept.

x <- c("dem", "temperature", "slope", "profile.curvature", "twi")

set.seed(314)
myClusters <- kmeans(
scale(subgrd[, x]), centers = 5, iter.max = 1000, nstart = 100)
subgrd$cluster <- myClusters$cluster
68 4 Stratified simple random sampling

Figure 4.8 shows the five clusters obtained by k-means clustering of the raster
cells. These clusters can be used as strata in random sampling.

FIGURE 4.8: Five clusters obtained by k-means clustering of the raster cells
of Xuancheng, using five scaled covariates in clustering.

The size of the clusters used as strata is largely different (Table 4.3). This
table also shows means of the unscaled covariates used in clustering.
Categorical variables can be accommodated in clustering using the technique
proposed by Huang (1998), implemented in package clustMixType (Szepan-
nek, 2018).
In the situation that we already have some data of the study variable, an
alternative solution is to calibrate a model for the study variable, for instance

TABLE 4.3: Size (Nh) and means of clustering variables of the five strata of
Xuancheng obtained with k-means clustering of raster cells.

Stratum Nh Elevation Temperature Slope Profilecurv Twi

1 1,581 300 14.45 12.21 -0.00147 6.34
2 15,987 54 15.44 2.11 0.00001 9.24
3 4,255 181 14.64 11.11 0.00068 7.75
4 4,955 20 15.59 0.47 0.00006 17.08
5 1,625 416 13.70 21.21 0.00012 6.45
4.6 Geographical stratification 69

a multiple linear regression model, using the covariates as predictors, and to

use the predictions of the study variable as a single stratification variable in
cum-root-f stratification or in optimal spatial stratification, see Section 13.2.

4.6 Geographical stratification

When no covariate is available, we may still decide to apply a geographical
stratification. For instance, a square study area can be divided into 4 × 4
equal-sized subsquares that are used as strata. When we select one or two
points per subsquare, we avoid strong spatial clustering of the sampling points.
Geographical stratification improves the spatial coverage. When the study
variable is spatially structured, think for instance of a spatial trend, then
geographical stratification will lead to more precisely estimated means (smaller
sampling variances).
A simple method for constructing geographical strata is k-means clustering
(Brus et al., 1999). See Section 17.2 for a simple illustrative example of how
geographical strata are computed with k-means clustering. In this approach,
the study area is discretised by a large number of grid cells. These grid cells are
the objects that are clustered. The clustering variables are simply the spatial
coordinates of the centres of the grid cells. This method leads to compact
geographical strata, briefly referred to as geostrata. Geostrata can be computed
with function kmeans, as shown in Section 4.5. The two clustering variables have
the same scale, so they should not be scaled because this would lead to an
arbitrary distortion of geographical distances. The geostrata generally will not
have the same number of grid cells. Geostrata of equal size can be attractive,
as then the sample becomes selfweighting, i.e., the sample mean is an unbiased
estimator of the population mean.
Geostrata of the same size can be computed with function stratify of the
package spcosa (Walvoort et al. (2020), Walvoort et al. (2010)), with argument
equalArea = TRUE.

If the total number of grid cells divided by the number of strata is an integer,
the stratum sizes are exactly equal; otherwise, the difference is one grid cell.
Walvoort et al. (2010) describe the k-means algorithms implemented in this
package in detail. Argument object of function stratify specifies a spatial ob-
ject of the population units. In the R code below grdVoorst is converted to a
SpatialPixelsDataFrame with function gridded of the package sp. The spatial ob-
ject can also be of class SpatialPolygons. In that case, either argument nGridCells
or argument cellSize must be set, so that the vector map in object can be
discretised by a finite number of grid cells. Argument nTry specifies the number
of initial stratifications in k-means clustering, and therefore is comparable with
70 4 Stratified simple random sampling

argument nstart of function kmeans. For more details on spatial stratification

using k-means clustering, see Section 17.2. The k-means algorithm used with
equalArea = TRUE takes much more computing time than the one used with
equalArea = FALSE.

library(spcosa)
library(sp)
set.seed(314)
gridded(subgrd) <- ~ x1 + x2
mygeostrata <- stratify(
object = subgrd, nStrata = 50, nTry = 1, equalArea = TRUE)

Function spsample of package spcosa is used to select from each geostratum a

simple random sample of two points.

set.seed(314)
mysample <- spcosa::spsample(mygeostrata, n = 2)

Figure 4.9 shows fifty compact geostrata of equal size for Xuancheng with the
selected sampling points. Note that the sampling points are reasonably well
spread throughout the study area1 .
Once the observations are done, the population mean can be estimated with
function estimate. For Xuancheng I simulated data from a normal distribution,
just to illustrate estimation with function estimate. Various statistics can be
estimated, among which the population mean (spatial mean), the standard
error, and the CDF. The CDF is estimated by transforming the data into
indicators (Subsection 3.1.2).

library(spcosa)
mysample <- spcosa::spsample(mygeostrata, n = 2)
mydata <- data.frame(z = rnorm(100, mean = 10, sd = 2))
mean <- estimate("spatial mean", mygeostrata, mysample, data = mydata)
se <- estimate("standard error", mygeostrata, mysample, data = mydata)
cdf <- estimate("scdf", mygeostrata, mysample, data = mydata)

The estimated population mean equals 9.8 with an estimated standard error
of 0.2.
1 The compact geostrata and the sample are plotted with package ggplot2. A simple

alternative is to use method plot of spcosa: plot(mygeostrata, mysample).

4.6 Geographical stratification 71

FIGURE 4.9: Compact geostrata of equal size for Xuancheng and stratified
simple random sample of two points per stratum.

Exercises

6. Why is it attractive to select at least two points per geostratum?

7. The alternative to 50 geostrata and two points per geostratum

is 100 geostrata and one point per geostratum. Which sampling
strategy will be more precise?

8. The geostrata in Figure 4.9 have equal size (area), which can
be enforced by argument equalArea = TRUE. Why are equal sizes
attractive? Work out the estimator of the population mean for
strata of equal size.

9. Write an R script to construct 20 compact geographical strata of

equal size for agricultural field Leest. The geopackage file of this field
can be read with sf function read_sf(system.file("extdata/leest.gpkg",
package = "sswr")). Remove the projection attributes with
st_set_crs(NA_crs_), and convert the simple feature object to
a spatial object with method as_Spatial. Select two points per
72 4 Stratified simple random sampling

geostratum, using function spsample of package spcosa. Repeat this

with 40 strata of equal size, and randomly select one point per
stratum.
• If only one point per stratum is selected, the sampling variance
can be approximated by the collapsed strata estimator. In this
method, pairs of strata are formed, and the two strata of a
pair are joined. In each new stratum we now have two points.
With an odd number of strata there will be one group of three
strata and three points. The sample is then analysed as if it
were a random sample from the new collapsed strata. Suppose
we group the strata on the basis of the measurements of the
study variable. Do you think this is a proper way of grouping?
• In case you think this is not a proper way of grouping the
strata, how would you group the strata?
• Is the sampling variance estimator unbiased? If not, is the
sampling variance overestimated or underestimated?

10. Laboratory costs for measuring the study variable can be saved by
bulking the soil aliquots (composite sampling). There are two options:
bulking all soil aliquots from the same stratum (bulking within strata)
or bulking by selecting one aliquot from each stratum (bulking across
strata). In spcosa bulking across strata is implemented. Write an
R script to construct 20 compact geographical strata for study area
Voorst. Use argument equalArea = TRUE. Select four points per stratum
using argument type = "composite", and convert the resulting object
to SpatialPoints. Extract the 𝑧-values in grdVoorst at the selected
sampling points using function over. Add a variable to the resulting
data frame indicating the composite (points 1 to 4 are from the first
stratum, points 5 to 8 from the second stratum, etc.), and estimate
the means for the four composites using function tapply. Finally,
estimate the population mean and its standard error.
• Can the sampling variance of the estimator of the mean be
estimated for bulking within the strata?
• The alternative to analysing the concentration of four composite
samples obtained by bulking across strata is to analyse all 20
× 4 aliquots separately. The strata have equal size, so the
inclusion probabilities are equal. As a consequence, the sample
mean is an unbiased estimator of the population mean. Is the
precision of this estimated population mean equal to that of
the estimated population mean with composite sampling? If
not, is it smaller or larger, and why?
• If you use argument equalArea = FALSE in combination with ar-
gument type = "composite", you get an error message. Why does
this combination of arguments not work?
4.8 Multiway stratification 73

4.7 Multiway stratification

In Section 4.5 multiple continuous covariates are used to construct clusters
of raster cells using k-means. These clusters are then used as strata. This
section considers the case where we have multiple categorical and/or con-
tinuous variables that we would like to use as stratification variables. The
continuous stratification variables are first used to compute strata based on
that stratification variable, e.g., using the cum-root-f method. What could
be done then is to compute the cross-classification of each unit and use these
cross-classifications as strata in random sampling. However, this may lead to
numerous strata, maybe even more than the intended sample size. To reduce
the total number of strata, we may aggregate cross-classification strata with
similar means of the study variable, based on our prior knowledge.
An alternative to aggregation of cross-classification strata is to use the separate
strata, i.e., the strata based on an individual stratification variable, as marginal
strata in random sampling. How this works is explained in Subsection 9.1.4.

4.8 Multivariate stratification

Another situation is where we have multiple study variables and would like to
optimise the stratification and allocation for estimating the population means
of all study variables. Optimal stratification for multiple study variables is only
relevant if we would like to use different stratification variables for the study
variables. In many cases, we do not have reliable prior information about the
different study variables justifying the use of multiple stratification variables.
We are already happy to have one stratification variable that may serve to
increase the precision of the estimated means of all study variables.
However, in case we do have multiple stratification variables tailored to different
study variables, the objective is to partition the population in strata, so that
for a given allocation, the total sampling costs, assuming a linear costs model
(Equation (4.18)), are minimised given a constraint on the precision of the
estimated mean for each study variable.
Package SamplingStrata (Barcaroli et al., 2020) can be used to optimise
multivariate strata. Barcaroli (2014) gives details about the objective function
and the algorithm used for optimising the strata. Sampling units are allocated
to the strata by Bethel allocation (Bethel, 1989). The required precision is
specified in terms of a coefficient of variation, one per study variable.
74 4 Stratified simple random sampling

Multivariate stratification is illustrated with the Meuse data set of package

gstat (Pebesma, 2004). The prior data of heavy metal concentrations of Cd and
Zn are used in spatial prediction to create maps of these two study variables.
The maps of natural logs of the two metal concentrations are created by kriging
with an external drift, using the square root of the distance to the Meuse river
as a predictor for the mean, see Section 21.3 for how this spatial prediction
method works.
Figure 4.10 shows the map with the predicted log Cd and log Zn concentrations.

FIGURE 4.10: Kriging predictions of natural logs of Cd and Zn concen-

trations in the study area Meuse, used as stratification variables in bivariate
stratification.

The predicted log concentrations of the two heavy metals are used as stratifi-
cation variables in designing a new sample for design-based estimation of the
population means of Cd and Zn. For the log of Cd, there are negative predicted
concentrations (Figure 4.10). This leads to an error when running function
optimStrata. The minimum predicted log Cd concentration is -1.7, so I added
2 to the predictions. A variable indicating the domains of interest is added
to the data frame. The value of this variable is 1 for all grid cells, so that a
sample is designed for estimating the mean of the entire population. As a first
step, function buildFrameDF is used to create a data frame that can be handled
by function optimStrata. Argument X specifies the stratification variables, and
argument Y the study variables. In our case, the stratification variables and
the study variables are the same. This is typical for the situation where the
stratification variables are obtained by mapping the study variables.

library(SamplingStrata)
df <- data.frame(cd = lcd_kriged$var1.pred + 2,
4.8 Multivariate stratification 75

zn = lzn_kriged$var1.pred,
dom = 1,
id = seq_len(nrow(lcd_kriged)))
frame <- buildFrameDF(
df = df, id = "id",
X = c("cd", "zn"), Y = c("cd", "zn"),
domainvalue = "dom")

Next, a data frame with the precision requirements for the estimated means is
created. The precision requirement is given as a coefficient of variation, i.e.,
the standard error of the estimated population mean, divided by the estimated
mean. The study variables as specified in Y are used to compute the estimated
means and the standard errors for a given stratification and allocation.

cv <- as.data.frame(list(DOM = "DOM1", CV1 = 0.02, CV2 = 0.02, domainvalue = 1))

Finally, the multivariate stratification is optimised by searching for the optimal

stratum bounds using a genetic algorithm (Gershenfeld, 1999).

set.seed(314)
res <- optimStrata(
method = "continuous", errors = cv, framesamp = frame, nStrata = 5,
iter = 50, pops = 20, showPlot = FALSE)

A summary of the strata can be obtained with function summaryStrata.

smrstrata <- summaryStrata(res$framenew, res$aggr_strata, progress = FALSE)

Stratum Population Allocation Lower_X1 Upper_X1 Lower_X2 Upper_X2

1 1 717 7 0.266 1.421 4.502 5.576
2 2 694 5 1.421 2.090 4.950 6.010
3 3 597 3 2.091 2.630 5.163 6.358
4 4 704 6 2.630 3.476 5.472 6.802
5 5 391 5 3.480 4.781 6.234 7.527

Column Population contains the sizes of the strata, i.e., the number of grid cells.
The total sample size equals 26. The sample size per stratum is computed with
Bethel allocation, see Section 4.3. The last four columns contain the lower and
upper bounds of the orthogonal intervals.
Figure 4.11 shows a 2D-plot of the bivariate strata. The strata can be plotted
as a series of nested rectangles. All population units in the smallest rectangle
belong to stratum 1; all units in the one-but-smallest rectangle that are not
in the smallest rectangle belong to stratum 2, etc. If we have more than two
76 4 Stratified simple random sampling

stratification variables, the strata form a series of nested hyperrectangles or

boxes. The strata are obtained as the Cartesian product of orthogonal intervals.

plt <- plotStrata2d(res$framenew, res$aggr_strata,

domain = 1, vars = c("X1", "X2"), labels = c("Cd", "Zn"))

FIGURE 4.11: 2D-plot of optimised bivariate strata of the study area Meuse.

It may happen that after the optimisation of the stratum bounds in some
resulting strata, no units are contained. If the stratification with a smaller
number of strata requires fewer sampling units so that the sampling costs
are lower (and still the precision requirement is met), then this is retained
as the optimal stratification.

Figure 4.12 shows a map of the optimised strata.

The expected coefficient of variation can be extracted with function expected_CV.

expected_CV(res$aggr_strata)
4.8 Multivariate stratification 77

FIGURE 4.12: Map of optimised bivariate strata of the study area Meuse.

cv(Y1) cv(Y2)
DOM1 0.02 0.0087401

The coefficient of variation of Cd is indeed equal to the desired level of 0.02,

for Zn it is smaller. So, in this case Cd is the study variable that determines
the total sample size of 26 units.
Note that these coefficients of variation are computed from the stratification
variables, which are predictions of the study variable. Errors in these predictions
are not accounted for. It is well known that kriging is a smoother, so that the
variance of the predicted values within a stratum is smaller than the variance of
the true values. As a consequence, the coefficients of variation of the predictions
underestimate the coefficients of variation of the study variables. See Section
13.2 for how prediction errors and spatial correlation of prediction errors can
be accounted for in optimal stratification. An additional problem is that I
added a value of 2 to the log Cd concentrations. This does not affect the
standard error of the estimated mean, but does affect the estimated mean, so
that also for this reason the coefficient of variation of the study variable Cd is
underestimated.
5
Systematic random sampling

A simple way of drawing probability samples whose units are spread uniformly
over the study area is systematic random sampling (SY), which from a two-
dimensional spatial population entails the selection of a regular grid randomly
placed on the area. A systematic sample can be selected with function spsample
of package sp with argument type = "regular" (Bivand et al., 2013). Argument
offset is not used, so that the grid is randomly placed on the study area.
This is illustrated with Voorst. First, data.frame grdVoorst is converted to
SpatialPixelsDataFrame with function gridded.

library(sp)
gridded(grdVoorst) <- ~ s1 + s2
n <- 40
set.seed(777)
mySYsample <- spsample(x = grdVoorst, n = n, type = "regular") %>%
as("data.frame")

Figure 5.1 shows the randomly selected systematic sample. The shape of the
grid is square, and the orientation is East-West (E-W), North-South (N-S).
There is no strict need for random selection of the orientation of the grid.
Random placement of the grid on the study area suffices for design-based
estimation.

FIGURE 5.1: Systematic random sample (square grid) from Voorst.

Argument n in function spsample is used to set the sample size. Note that this
is the expected sample size, i.e., on average over repeated sampling the sample
size is 40. In Figure 5.1 the number of selected sampling points equals 38. Given
the expected sample size, the spacing of the square grid can be computed with
√𝐴/𝑛, with 𝐴 the area of the study area. This area 𝐴 can be computed by

79
80 5 Systematic random sampling

the total number of cells of the discretisation grid multiplied by the area of a
grid cell. Note that the area of the study area is smaller than the number of
grid cells in the horizontal direction multiplied by the number of grid cells in
the vertical direction multiplied by the grid cell area, as we have non-availables
(built-up areas, roads, etc.).

cell_size <- 25
A <- nrow(grdVoorst) * cell_sizeˆ2
(spacing <- sqrt(A / n))

[1] 342.965

Instead of argument n we may use argument cell_size to select a grid with

a specified spacing. The expected sample size of a square grid can then be
computed with 𝐴/𝑠𝑝𝑎𝑐𝑖𝑛𝑔2 .
The spatial coverage with random grid sampling is better than that with strat-
ified random sampling using compact geographical strata (Section 4.6), even
with one sampling unit per geostratum. Consequently, in general systematic
random sampling results in more precise estimates of the mean or total.
However, there are also two disadvantages of systematic random sampling
compared to geographically stratified random sampling. First, for systematic
random sampling, no design-unbiased estimator of the sampling variance exists.
Second, the number of sampling units with random grid sampling is not fixed,
but varies among randomly drawn samples. We may choose the grid spacing
such that on average the number of sampling units equals the required (allowed)
number of sampling units, but for the actually drawn sample, this number can
be smaller or larger. In Voorst the variation of the sample size is quite large.
The approximated sampling distribution, obtained by repeating the sampling
10,000 times, is bimodal (Figure 5.2). The smaller sample sizes are of square
grids with only two E-W oriented rows of points instead of three rows.
A large variation in sample size over repeated selection with the sampling
design under study is undesirable and should be avoided when possible. In the
case of Voorst, a simple solution is to select a rectangular grid instead of a
square grid, with a spacing in the N-S direction that results in a fixed number
of E-W oriented rows of sampling points over repeated selection of grids. This
is achieved with a N-S spacing equal to the dimension of the study area in N-S
direction divided by an integer. The spacing in E-W direction is then adapted
so that on average a given number of sampling points is selected. As the N-S
dimension of Voorst is 1,000 m, a N-S spacing of 1,000/3 m is chosen, so that
the number of E-W oriented rows of sampling points in the systematic sample
equals three for any randomly selected rectangular grid.
5.0 Systematic random sampling 81

FIGURE 5.2: Approximated sampling distribution of the sample size of

systematic random samples from Voorst. The expected sample size is 40.

dy <- 1000 / 3
dx <- A / (n * dy)
mySYsample_rect <- spsample(
x = grdVoorst, cellsize = c(dx, dy), type = "regular")

The E-W spacing is somewhat larger than the N-S spacing: 352.875 m
vs. 333.333 m. The variation in sample size with the random rectangular
grid is much smaller than that of the square grid. The sample size now ranges
from 33 to 46, whereas with the square grid the range varies from 20 to 48.

summary(sampleSizes)

Min. 1st Qu. Median Mean 3rd Qu. Max.

33.00 38.00 40.00 39.99 42.00 46.00

An alternative shape for the sampling grid is triangular. Triangular grids

can be selected with argument type = "hexagonal". The centres of hexagonal
sampling grid cells form a triangular grid. The triangular grid was shown to
82 5 Systematic random sampling

yield most precise estimates of the population mean given the expected sample
size (Matérn, 1986). Given the spacing of a triangular grid, the expected sample
size can be computed by the area 𝐴 of the study area divided by the area of
hexagonal grid cells
√ with the sampling points at their centres. The area of a
hexagon equals 6 3/4 𝑟2 , with 𝑟 the radius of the circle circumscribing the
hexagon (distance from centre to a corner of the hexagon). So, by choosing a
√
radius of √𝐴/(6 3/4) 𝑛 the expected sample equals 𝑛. The distance between
neighbouring
√ points of the triangular
√ grid in the E-W direction, 𝑑𝑥, then equals
𝑟 3. The N-S distance equals 3/2 𝑑𝑥.

cnst <- 6 * sqrt(3) / 4

r <- sqrt(A / (cnst * n))
dx <- r * sqrt(3)
dy <- sqrt(3) / 2 * dx

Function does not work properly in combination with argument type

spsample
= "hexagonal".
Over repeated sampling, the average sample size is not equal to
the chosen sample size passed to function spsample with argument n. The same
problem remains when using argument cellsize.
Min. 1st Qu. Median Mean 3rd Qu. Max.
18.00 23.00 26.00 28.39 35.00 41.00

The following code can be used for random selection of triangular grids.

SY_triangular <- function(dx, grd) {

dy <- sqrt(3) / 2 * dx
#randomly select offset
offset_x <- runif(1, min = 0, max = dx)
offset_y <- runif(1, min = 0, max = dy)
#compute x-coordinates of 1 row and y-coordinates of 1 column
bbox <- bbox(grd)
nx <- ceiling((bbox[1, 2] - bbox[1, 1]) / dx)
ny <- ceiling((bbox[2, 2] - bbox[2, 1]) / dy)
x <- (-1:nx) * dx + offset_x
y <- (0:ny) * dy + offset_y
#compute coordinates of rectangular grid
xy <- expand.grid(x, y)
names(xy) <- c("x", "y")
#shift points of even rows in horizontal direction
units <- which(xy$y %in% y[seq(from = 2, to = ny, by = 2)])
xy$x[units] <- xy$x[units] + dx / 2
#add coordinates of origin
xy$x <- xy$x + bbox[1, 1]
xy$y <- xy$y + bbox[2, 1]
5.1 Estimation of population parameters 83

#overlay with grid

coordinates(xy) <- ~ x + y
mysample <- data.frame(coordinates(xy), over(xy, grd))
#delete points with NA
mysample <- mysample[!is.na(mysample[, 3]), ]
}
set.seed(314)
mySYsample_tri <- SY_triangular(dx = dx, grd = grdVoorst)

Figure 5.3 shows a triangular grid, selected randomly from Voorst with an
expected sample size of 40. The selected triangular grid has 42 points.

FIGURE 5.3: Systematic random sample (triangular grid) from Voorst.

5.1 Estimation of population parameters

With systematic random sampling, all units have the same inclusion probabil-
ity, equal to 𝐸[𝑛]/𝑁, with 𝐸[𝑛] the expected sample size. Consequently, the
population total can be estimated by

𝑧u� 𝑧
̂ =∑
𝑡(𝑧) = 𝑁 ∑ u� . (5.1)
𝜋
u�∈u� u� u�∈u�
𝐸[𝑛]

The population mean can be estimated by dividing this 𝜋 estimator of the

population total by the population size:

𝑧u�
𝑧 ̄̂ = ∑ . (5.2)
u�∈u�
𝐸[𝑛]

In this 𝜋 estimator of the population mean the sample sum of the observations
is not divided by the number of selected units 𝑛, but by the expected number
of units 𝐸[𝑛].
84 5 Systematic random sampling

An alternative estimator is obtained by dividing the 𝜋 estimator of the popu-

lation total by the 𝜋 estimator of the population size:

1 𝑁
𝑁̂ = ∑ =𝑛 . (5.3)
𝜋
u�∈u� u�
𝐸[𝑛]

This yields the ratio estimator of the population mean:

̂
𝑡(𝑧) 1
̄̂
𝑧ratio = = ∑ 𝑧u� . (5.4)
𝑁̂ 𝑛 u�∈u�

So, the ratio estimator of the population total is equal to the unweighted
sample mean. In general, the variance of this ratio estimator is smaller than
that of the 𝜋 estimator. On the other hand the 𝜋 estimator is design-unbiased,
whereas the ratio estimator is not, although its bias can be negligibly small.
Only in the very special case where the sample size with systematic random
sampling is fixed, the two estimators are equivalent.
Recall that for Voorst we have exhaustive knowledge of the study variable 𝑧:
values of the soil organic matter (SOM) concentration were simulated for all
grid cells. To determine the 𝑧-values at the selected sampling points, an overlay
of the systematic random sample and the SpatialPixelsDataFrame is made, using
function over of package sp.

res <- over(mySYsample, grdVoorst)

mySYsample <- as(mySYsample, "data.frame")
mySYsample$z <- res$z
mz_HT <- sum(mySYsample$z) / n
mz_ratio <- mean(mySYsample$z)

Using the systematic random sample of Figure 5.1, the 𝜋 estimated mean SOM
concentration equals 69.8 g kg-1 , the ratio estimate equals 73.5 g kg-1 . The
ratio estimate is larger than the 𝜋 estimate, because the size of the selected
sample is two units smaller (38) than the expected sample size (40).

5.2 Approximating the sampling variance of the

estimator of the mean
An unbiased estimator of the sampling variance of the estimator of the mean
is not available. A simple, often applied procedure is to calculate the sampling
variance as if the sample were a simple random sample (Equation (3.14) or
5.2 Approximating the sampling variance of the estimator of the mean 85

(3.15)). In general, this procedure overestimates the sampling variance, so that

we are on the safe side.

av_SI_mz <- var(mySYsample$z) / nrow(mySYsample)

The approximated variance equals 50.3 (g kg-1 )2 .

Alternatively, the sampling variance can be estimated by treating the systematic
random sample as if it were a stratified simple random sample (Equation (4.4)).
The sampling units are clustered on the basis of their spatial coordinates into
𝐻 = 𝑛/2 clusters (𝑛 even) or 𝐻 = (𝑛 − 1)/2 clusters (𝑛 odd). In the next code
chunk, a simple k-means function is defined to cluster the sampling units of the
grid into equal-sized clusters. Arguments s1 and s2 are the spatial coordinates
of the sampling units, k is the number of clusters. First, in this function the
ids of equal-sized clusters are randomly assigned to the sampling units on the
nodes of the sampling grid (initial clustering). Next, the centres of the clusters,
i.e., the means of the spatial coordinates of the clusters (initial cluster centres),
are computed. There are two for-loops. In the inner loop it is determined
whether the cluster id of the unit selected in the outer loop should be swopped
with the cluster id of the next unit. If both units have the same cluster id the
next unit is selected, until a unit of a different cluster is found. The cluster ids
of the two units are swopped when the sum of the squared distances of the
two units to their corresponding cluster centres is reduced. When the cluster
ids are swopped, the centres are recomputed. The two loops are repeated until
no swops are made anymore.

.kmeans_equal_size <- function(s1, s2, k) {

n <- length(s1)
cluster_id <- rep(1:k, times = ceiling(n / k))
cluster_id <- cluster_id[1:n]
cluster_id <- cluster_id[sample(n, size = n)]
s1_c <- tapply(s1, INDEX = cluster_id, FUN = mean)
s2_c <- tapply(s2, INDEX = cluster_id, FUN = mean)
repeat {
n_swop <- 0
for (i in 1:(n - 1)) {
ci <- cluster_id[i]
for (j in (i + 1):n) {
cj <- cluster_id[j]
if (ci == cj) {
next
}
d1 <- (s1[i] - s1_c[ci])ˆ2 + (s2[i] - s2_c[ci])ˆ2 +
(s1[j] - s1_c[cj])ˆ2 + (s2[j] - s2_c[cj])ˆ2
d2 <- (s1[i] - s1_c[cj])ˆ2 + (s2[i] - s2_c[cj])ˆ2 +
86 5 Systematic random sampling

(s1[j] - s1_c[ci])ˆ2 + (s2[j] - s2_c[ci])ˆ2

if (d1 > d2) {
#swop cluster ids and recompute cluster centres
cluster_id[i] <- cj; cluster_id[j] <- ci
s1_c <- tapply(s1, cluster_id, mean)
s2_c <- tapply(s2, cluster_id, mean)
n_swop <- n_swop + 1
break
}
}
}
if (n_swop == 0) {
break
}
}
D <- fields::rdist(x1 = cbind(s1_c, s2_c), x2 = cbind(s1, s2))
dmin <- apply(D, MARGIN = 2, FUN = min)
MSSD <- mean(dminˆ2)
list(clusters = cluster_id, MSSD = MSSD)
}

The clustering is repeated 100 times (ntry = 100). The clustering with the
smallest mean of the squared distances of the sampling units to their cluster
centres (mean squared shortest distance, MSSD) is selected.

kmeans_equal_size <- function(s1, s2, k, ntry) {

res_opt <- NULL
MSSD_min <- Inf
for (i in 1:ntry) {
res <- .kmeans_equal_size(s1, s2, k)
if (res$MSSD < MSSD_min) {
MSSD_min <- res$MSSD
res_opt <- res
}
}
res_opt
}
n <- nrow(mySYsample); k <- floor(n / 2)
set.seed(314)
res <- kmeans_equal_size(s1 = mySYsample$x1 / 1000, s2 = mySYsample$x2 / 1000,
k = k, ntry = 100)
mySYsample$cluster <- res$clusters
5.2 Approximating the sampling variance of the estimator of the mean 87

Figure 5.4 shows the clustering of the systematic random sample of Figure 5.1.
The two or three sampling units of a cluster are treated as a simple random
sample from a stratum, and the variance estimator for stratified random
sampling is used. The weights are computed by 𝑤ℎ = 𝑛ℎ /𝑛. With 𝑛 even the
stratum weight is 1/𝐻 for all strata. For more details on variance estimation
with stratified simple random sampling, refer to Section 4.1.

S2z_h <- tapply(mySYsample$z, INDEX = mySYsample$cluster, FUN = var)

nh <- tapply(mySYsample$z, INDEX = mySYsample$cluster, FUN = length)
v_mz_h <- S2z_h / nh
w_h <- nh / sum(nh)
av_STSI_mz <- sum(w_hˆ2 * v_mz_h)

FIGURE 5.4: Clustering of grid points for approximating the variance of

the ratio estimator of the mean SOM concentration in Voorst.

This method yields an approximated variance of 51.2 (g kg-1 )2 , which is for

the selected triangular grid slightly larger than the simple random sample
approximation. Hereafter, we will see that on average the stratified simple ran-
dom sample approximation of the variance is smaller than the simple random
sample approximation. For an individual sample, the reverse can be true.
A similar approach for approximating the variance was proposed by Matérn
(Matérn, 1947) a long time ago. In this approach the variance is approximated
by computing the squared difference of two local means. A local mean is
computed by linear interpolation of the observations at the two nodes on the
diagonal of a square sampling grid cell. The four corners of a sampling grid
cell serve as a group. Every sampling grid node belongs to four groups, and
so the observation at a sampling grid node is used four times in computing a
local mean. Near the edges of the study area, we have incomplete groups: one,
two, or even three observations are missing. To compute a squared difference,
these missing values are replaced by the sample mean. This results in as many
squared differences as we have groups. Note that the number of groups is larger
88 5 Systematic random sampling

than the sample size. The squared differences are computed by

𝑧u�,u� + 𝑧u�+1,u�+1 𝑧u�+1,u� + 𝑧u�,u�+1 2

𝑑2u�,u� = ( − )
2 2
(5.5)
(𝑧u�,u� − 𝑧u�+1,u� − 𝑧u�,u�+1 + 𝑧u�+1,u�+1 )2
= ,
4

with 𝑟 = 0, 1, … , 𝑅 an index for the column number and 𝑠 = 0, 1, … , 𝑆 an

index for the row number of the extended grid. The variance of the estimator
of the mean (sample mean) is then approximated by the sum of the squared
differences divided by the squared sample size:

∑u� 𝑑2
u�=1 u�
𝑉(̂ 𝑧u�̄ ) = , (5.6)
𝑛2

with 𝑑2u� the squared difference of group unit 𝑔, and 𝐺 the total number of
groups.
To approximate the variance with Matérn’s method, a function is defined.

matern <- function(s) {

g_11 <- within(s, {gr <- i; gs <- j; z11 <- z})[, c("gr", "gs", "z11")]
g_12 <- within(s, {gr <- i; gs <- j - 1; z12 <- z})[, c("gr", "gs", "z12")]
g_21 <- within(s, {gr <- i - 1; gs <- j; z21 <- z})[, c("gr", "gs", "z21")]
g_22 <- within(s, {gr <- i - 1; gs <- j - 1; z22 <- z})[, c("gr", "gs", "z22")]
g <- Reduce(function(x, y) merge(x = x, y = y, by = c("gr", "gs"), all = TRUE),
list(g_11, g_12, g_21, g_22))
g[is.na(g)] <- mean(s$z)
g <- within(g, T <- (z11 - z12 - z21 + z22)ˆ2 / 4)
sum(g$T) / ((nrow(s))ˆ2)
}

Before using this function the data frame with the sample data must be
extended with two variables: an index 𝑖 for the column number and an index 𝑗
for the row number of the square grid.

mySYsample <- mySYsample %>%

mutate(
i = round((x1 - min(x1)) / spacing),
j = round((x2 - min(x2)) / spacing))
matern(mySYsample)

[1] 41.63163
5.2 Approximating the sampling variance of the estimator of the mean 89

Figure 5.5 shows the approximated sampling distributions of estimators of the

mean SOM concentration for systematic random sampling, using a randomly
placed square grid with fixed orientation and an expected sample size of 40,
and for simple random sampling, obtained by repeating the random sampling
with each design and estimation 10,000 times. To estimate the population
mean from the systematic random samples, both the 𝜋 estimator and the ratio
estimator are used.

FIGURE 5.5: Approximated sampling distribution of estimators of the mean

SOM concentration (g kg-1 ) in Voorst, for systematic random sampling (square
grid) and simple random sampling and an expected sample size of 40. With
systematic random sampling, both the 𝜋 estimator (SY.HT) and the ratio
estimator (SY.ratio) are used in estimation.

The boxplots of the estimated means indicate that systematic random sam-
pling in combination with the ratio estimator is more precise than simple
random sampling. The variance of the 10,000 ratio estimates equals 49.0 (g
kg-1 )2 , whereas for simple random sampling this variance equals 55.4 (g kg-1 )2 .
Systematic random sampling in combination with the 𝜋 estimator performs
very poorly: the variance equals 142.6 (g kg-1 )2 . This can be explained by the
strong variation in sample size (Figure 5.2), which is not accounted for in the
𝜋 estimator.
The mean of the 10,000 ratio estimates is 81.2 g kg-1 , which is about equal to
the population mean 81.1 g kg-1 , showing that in this case the design-bias of
the ratio estimator is negligibly small indeed.
90 5 Systematic random sampling

The average of the 10,000 approximated variances treating the systematic

sample as a simple random sample equals 56.4 (g kg-1 )2 . This is larger than the
variance of the ratio estimator (49.0 (g kg-1 )2 ). The stratified simple random
sample approximation of the variance somewhat underestimates the variance:
the mean of this variance approximation equals 47.1 (g kg-1 )2 . Also with
Matérn’s method, the variance is underestimated in this case: the mean of
the 10,000 variances equals 45.6 (g kg-1 )2 . Figure 5.6 shows boxplots of the
approximated standard error of the ratio estimator of the population mean.
The horizontal red line is at the standard deviation of the 10,000 ratio estimates
of the population mean. Differences between the three approximation methods
are small in this case.

FIGURE 5.6: Sampling distribution of the approximated standard error of

the ratio estimator of the mean SOM concentration (g kg-1 ) in Voorst, with
systematic random sampling (square grid) and an expected sample size of 40.
Approximations are obtained by treating the systematic sample as a simple
random sample (SI) or a stratified simple random sample (STSI), and with
Matérn’s method (Mat).

The variance of the 10,000 ratio estimates of the population mean with the
triangular grid and an expected sample size of 40 equals 46.9 (g kg-1 )2 . Treating
the triangular grid as a simple random sample strongly overestimates the
variance: the average approximated variance equals 60.1 (g kg-1 )2 . The stratified
simple random sample approximation performs much better in this case: the
average of the 10,000 approximated variances equals 46.8 (g kg-1 )2 . Matérn’s
method cannot be used to approximate the variance with a triangular grid.
5.2 Approximating the sampling variance of the estimator of the mean 91

The approximated variance for this clustering equals round(av_STSI_mz_SYtri,1).

Brus and Saby (2016) compared various variance approximations for systematic
random sampling, among which model-based prediction of the variance, using
a semivariogram that is estimated from the systematic sample, see Chapter 13.

Exercises

1. One solution to the problem of variance estimation with systematic

random sampling is to select multiple systematic random samples
independently from each other. So, for instance, instead of one
systematic random sample with an expected sample size of 40, we
may select two systematic random samples with an expected size of
20.
• Write an R script to select two systematic random samples
(random square grids) both with an expected size of 20 from
Voorst.
• Use each sample to estimate the population mean, so that you
obtain two estimated means. Overlay the points of each sample
with grdVoorst, using function over and extract the 𝑧-values.
• Use the two estimated means to estimate the sampling variance
of the estimator of the mean for systematic random sampling
with an expected sample size of 20.
• Use the two estimated means to compute a single, final estimate
of the population mean, as estimated from two systematic
random samples, each with an expected sample size of 20.
• Estimate the sampling variance of the final estimate of the
population mean.

2. Do you like this solution? What about the variance of the estimator
of the mean, obtained by selecting two systematic random samples
of half the expected size, as compared with the variance of the
estimator of the mean, obtained with a single systematic random
sample? Hint: plot the two random square grids. What do you think
of the spatial coverage of the two samples?
6
Cluster random sampling

With stratified random sampling using geographical strata and systematic

random sampling, the sampling units are well spread throughout the study
area. In general, this leads to an increase of the precision of the estimated
mean (total). This is because many spatial populations show spatial structure,
so that the values of the study variable at two close units are more similar than
those at two distant units. With large study areas the price to be paid for this
is long travel times, so that fewer sampling units can be observed in a given
survey time. In this situation, it can be more efficient to select spatial clusters
of population units. In cluster random sampling, once a cluster is selected, all
units in this cluster are observed. Therefore, this design is also referred to as
single-stage cluster random sampling. The clusters are not subsampled as in
two-stage cluster random sampling (see Chapter 7).
In spatial sampling, a popular cluster shape is a transect. This is because the
individual sampling units of a transect can easily be located in the field, which
was in particular an advantage in the pre-GPS era.
The implementation of cluster random sampling is not straightforward. Fre-
quently this sampling design is improperly implemented. A proper selection
technique is as follows (de Gruijter et al., 2006). In the first step, a starting
unit is selected, for instance by simple random sampling. Then the remaining
units of the cluster to which the starting unit belongs are identified by making
use of the definition of the cluster. For instance, with clusters defined as E-W
oriented transects with a spacing of 100 m between the units of a cluster, all
units E and W of the starting unit at a distance of 100 m, 200 m, etc. that
fall inside the study area are selected. These two steps are repeated until the
required number of clusters, not the number of units, is selected.
A requirement of a valid selection method is that the same cluster is selected,
regardless of which of its units is used as a starting unit. In the example above,
this is the case: regardless of which of the units of the transect is selected first,
the final set of units selected is the same because, as stated above, all units E
and W of the starting unit are selected.

An example of an improper implementation of cluster random sampling is

the following selection procedure. A cluster is defined as an E-W oriented
transect of four units with a mutual spacing of 100 m. A cluster is selected by

93
94 6 Cluster random sampling

randomly selecting a starting unit. The remaining three units of the cluster
are selected E of this starting unit. Units outside the study area are ignored.
With this selection method, the set of selected units is not independent of
the starting unit, and therefore this selection method is invalid.

Note that the size, i.e., the number of units, of a cluster need not be constant.
With the proper selection method described above, the selection probability of
a cluster is proportional to its size. With irregularly shaped study areas, the
size of the clusters can vary strongly. The size of the clusters can be controlled
by subdividing the study area into blocks, for instance, stripes perpendicular
to the direction of the transects, or square blocks in case the clusters are grids.
In this case, the remaining units are identified by extending the transect or
grid to the boundary of the block. With irregularly shaped areas, blocking will
not entirely eliminate the variation in cluster sizes.
Cluster random sampling is illustrated with the selection of E-W oriented
transects in Voorst. In order to delimit the length of the transects, the study
area is split into six 1 km × 1 km zones. In this case, the zones have an equal
size, but this is not needed. Note that these zones do not serve as strata. When
used as strata, from each zone, one or more clusters would be selected, see
Section 6.4.
In the code chunk below, function findInterval of the base package is used to
determine for all discretisation points in which zone they fall.

cell_size <- 25
w <- 1000 #width of zones
grdVoorst <- grdVoorst %>%
mutate(zone = s1 %>% findInterval(min(s1) + 1:5 * w + 0.5 * cell_size))

As a first step in the R code below, variable cluster is added to grdVoorst

indicating to which cluster a unit belongs. Note that each unit belongs exactly
to one cluster. The operator %% computes the modulus of the s1-coordinate and
the spacing of units within a transect (cluster). Function stringr of package
stringr (Wickham, 2019) joins the resulting vector, the vector with the s2-
coordinate, and the vector with the zone into a single character vector. The
sizes of the clusters are computed with function tapply.

spacing <- 100

grdVoorst <- grdVoorst %>%
mutate(
cluster = str_c(
(s1 - min(s1)) %% spacing,
s2 - min(s2),
6.0 Cluster random sampling 95

zone, sep = "_"),

unit = row_number())
M_cl <- tapply(grdVoorst$z, INDEX = grdVoorst$cluster, FUN = length)

FIGURE 6.1: Frequency distribution of the size of clusters in Voorst. Clusters

are E-W oriented transects within zones, with a spacing of 100 m between
units.

In total, there are 960 clusters in the population. Figure 6.1 shows the frequency
distribution of the size of the clusters.
Clusters are selected with probabilities proportional to their size and with
replacement (ppswr). So, the sizes of all clusters must be known, which explains
that all clusters must be enumerated. Selection of clusters by ppswr can be
done by simple random sampling with replacement of elementary units (centres
of grid cells) and identifying the clusters to which these units belong. Finally,
all units of the selected clusters are included in the sample. In the code chunk
below, a function is defined for selecting clusters by ppswr. Note variable cldraw,
that has value 1 for all units selected in the first draw, value 2 for all units
selected in the second draw, etc. This variable is needed in estimating the
population mean, as explained in Section 6.1.
96 6 Cluster random sampling

cl_ppswr <- function(sframe, n) {

units <- sample(nrow(sframe), size = n, replace = TRUE)
units_cl <- sframe$cluster[units]
mysamples <- NULL
for (i in seq_len(length(units_cl))) {
mysample <- sframe[sframe$cluster %in% units_cl[i], ]
mysample$start <- 0
mysample$start[mysample$unit %in% units[i]] <- 1
mysample$cldraw <- rep(i, nrow(mysample))
mysamples <- rbind(mysamples, mysample)
}
mysamples
}

Function cl_ppswr is now used to select six times a cluster by ppswr.

n <- 6
set.seed(314)
mysample <- cl_ppswr(sframe = grdVoorst, n = n)

As our population actually is infinite, the centres of the selected grid cells are
jittered to a random point within the selected grid cells. Note that the same
noise is added to all units of a given cluster.

mysample <- mysample %>%

group_by(cldraw) %>%
mutate(s1 = s1 + runif(1, min = -12.5, max = 12.5),
s2 = s2 + runif(1, min = -12.5, max = 12.5))

Figure 6.2 shows the selected sample. Note that in this case the second west-
most zone has two transects (clusters), whereas one zone has none, showing
that the zones are not used as strata. The total number of selected points equals
50. Similar to systematic random sampling, with cluster random sampling the
total sample size is random, so that we do not have perfect control of the total
sample size. This is because in this case the size, i.e., the number of points, of
the clusters is not constant but varies.
The output data frame of function cl has a variable named start. This is
an indicator with value 1 if this point of the cluster is selected first, and 0
otherwise. When in the field, it appears that the first selected point of a cluster
does not belong to the target population, all other points of that cluster are
also discarded. This is to keep the selection probabilities of the clusters exactly
proportional to their size. Column cldraw is needed in estimation because
clusters are selected with replacement. In case a cluster is selected more than
once, multiple means of that cluster are used in estimation, see next section.
6.1 Estimation of population parameters 97

FIGURE 6.2: Cluster random sample from Voorst selected by ppswr.

6.1 Estimation of population parameters

With ppswr sampling of clusters, the population total can be estimated by the
pwr estimator:

1 𝑡u� (𝑧)
̂ =
𝑡(𝑧) ∑ , (6.1)
𝑛 u�∈u� 𝑝u�

with 𝑛 the number of cluster draws, 𝑝u� the draw-by-draw selection probability
of cluster 𝑗, and 𝑡u� (𝑧) the total of cluster 𝑗:

u�u�

𝑡u� (𝑧) = ∑ 𝑧u�u� , (6.2)

u�=1

with 𝑀u� the size (number of units) of cluster 𝑗 and 𝑧u�u� the study variable
value of unit 𝑘 in cluster 𝑗.
The draw-by-draw selection probability of a cluster equals

𝑀u�
𝑝u� = , (6.3)
𝑀
with 𝑀 the total number of population units (for Voorst 𝑀 equals 7,528).
Inserting this in Equation (6.1) yields

𝑀 𝑡u� (𝑧) 𝑀
̂ =
𝑡(𝑧) ∑ = ∑𝑧̄ , (6.4)
𝑛 u�∈u� 𝑀u� 𝑛 u�∈u� u�

with 𝑧u�̄ the mean of cluster 𝑗. Note that if a cluster is selected more than once,
multiple means of that cluster are used in the estimator.
98 6 Cluster random sampling

Dividing this estimator by the total number of population units, 𝑀, yields the
estimator of the population mean:
1
𝑧 ̄̂ = ∑𝑧̄ . (6.5)
𝑛 u�∈u� u�

Note the two bars in 𝑧,̄̂ indicating that the observations are averaged twice.
For an infinite population of points discretised by the centres of a finite number
of grid cells, 𝑧u�u� in Equation (6.2) is the study variable value at a randomly
selected point within the grid cell multiplied by the area of the grid cell. The
estimated population total thus obtained is equal to the estimated population
mean (Equation (6.5)) multiplied by the area of the study area.
The sampling variance of the estimator of the mean with ppswr sampling of
clusters is equal to (Cochran (1977), equation (9A.6))

1 u� 𝑀u�
𝑉(𝑧)̄̂ = ∑ (𝑧 ̄ − 𝑧)̄ 2 , (6.6)
𝑛 u�=1 𝑀 u�

with 𝑁 the total number of clusters (for Voorst, 𝑁 = 960), 𝑧u�̄ the mean
of cluster 𝑗, and 𝑧 ̄ the population mean. Note that 𝑀u� /𝑀 is the selection
probability of cluster 𝑗.
This sampling variance can be estimated by (Cochran (1977), equation (9A.22))

̂2
̂ 𝑧)̄̂ = 𝑆 (𝑧)̄ ,
𝑉( (6.7)
𝑛

where 𝑆̂2 (𝑧)̄ is the estimated variance of cluster means (the between-cluster

variance):
1
̂
𝑆2 (𝑧)̄ = ∑(𝑧 ̄ − 𝑧)̄̂ 2 . (6.8)
𝑛 − 1 u�∈u� u�

In R the population mean and the sampling variance of the estimator of the
population means can be estimated as follows.

est <- mysample %>%

group_by(cldraw) %>%
summarise(mz_cl = mean(z)) %>%
summarise(mz = mean(mz_cl),
se_mz = sqrt(var(mz_cl) / n()))

The estimated mean equals 87.1 g kg-1 , and the estimated standard error
equals 17.4 g kg-1 . Note that the size of the clusters does not appear in these
formulas. This simplicity is due to the fact that the clusters are selected with
6.1 Estimation of population parameters 99

probabilities proportional to size. The effect of the cluster size on the variance
is implicitly accounted for. To understand this, consider that larger clusters
result in smaller variance among their means.
The same estimates are obtained with functions svydesign and svymean of package
survey (Lumley, 2021). Argument weights specifies the weights of the sampled
clusters equal to 𝑀/(𝑀u� 𝑛) (Equation (6.4)).

library(survey)
M <- nrow(grdVoorst)
mysample$weights <- M / (M_cl[mysample$cluster] * n)
design_cluster <- svydesign(id = ~ cldraw, weights = ~ weights, data = mysample)
svymean(~ z, design_cluster, deff = "replace")

mean SE DEff
z 87.077 17.428 4.0767

The design effect DEff as estimated from the selected cluster sample is consid-
erably larger than 1. About 4 times more sampling points are needed with
cluster random sampling compared to simple random sampling to estimate
the population mean with the same precision.
A confidence interval estimate of the population mean can be computed with
method confint. The number of degrees of freedom equals the number of cluster
draws minus 1.

confint(svymean(~ z, design_cluster, df = degf(design_cluster), level = 0.95))

2.5 % 97.5 %
z 52.91908 121.2347

Figure 6.3 shows the approximated sampling distribution of the pwr estimator
of the mean soil organic matter (SOM) concentration with cluster random
sampling and of the 𝜋 estimator with simple random sampling, obtained by
repeating the random sampling with each design and estimation 10,000 times.
The size of the simple random samples is equal to the expected sample size of
the cluster random sampling design (rounded to nearest integer).
The variance of the 10,000 estimated population means with cluster random
sampling equals 126.2 (g kg-1 )2 . This is considerably larger than with simple
random sampling: 44.8 (g kg-1 )2 . The large variance is caused by the strong
spatial clustering of points. This may save travel time in large study areas,
but in Voorst the saved travel time will be very limited, and therefore cluster
random sampling in Voorst is not a good idea. The average of the estimated
variances with cluster random sampling equals 125.9 (g kg-1 )2 . The difference
with the variance of the 10,000 estimated means is small because the estimator
of the variance, Equation (6.7), is unbiased. Figure 6.4 shows the approximated
100 6 Cluster random sampling

FIGURE 6.3: Approximated sampling distribution of the pwr estimator of

the mean SOM concentration (g kg-1 ) in Voorst with cluster random sampling
(Cl) and of the 𝜋 estimator with simple random sampling (SI). In Cl six clusters
are selected by ppswr. The (expected) sample size is 49 units.

sampling distribution of the sample size. The expected sample size can be
computed as follows:

p <- M_cl / sum(M_cl)

print(m_n <- n * sum(p * M_cl))

[1] 49.16844

So, the unequal draw-by-draw selection probabilities of the clusters are ac-
counted for in computing the expected sample size.

Exercises

1. Write an R script to compute the true sampling variance of the

estimator of the mean SOM concentration in Voorst for cluster ran-
dom sampling and clusters selected with ppswr, 𝑛 = 6, see Equation
(6.6). Compare the sampling variance for cluster random sampling
with the sampling variance for simple random sampling with a sam-
ple size equal to the expected sample size of cluster random sampling.
6.2 Clusters selected with probabilities proportional to size 101

FIGURE 6.4: Approximated sampling distribution of the sample size with

cluster random sampling from Voorst, for six clusters selected by ppswr.

2. As an alternative we may select three times a transect, using three 2

km × 1 km zones obtained by joining two neighbouring 1 km × 1 km
zones of Figure 6.2. Do you expect that the sampling variance of the
estimator of the population mean is equal to, larger or smaller than
that of the sampling design with six transects of half the length?

6.2 Clusters selected with probabilities proportional to

size, without replacement
In the previous section the clusters were selected with replacement (ppswr).
The advantage of sampling with replacement is that this keeps the statistical
inference simple, more specifically the estimation of the standard error of the
estimator of the population mean. However, in sampling from finite populations,
cluster sampling with replacement is less efficient than cluster sampling without
replacement, especially with large sampling fractions of clusters, i.e., if 1 − 𝑛/𝑁
102 6 Cluster random sampling

is small, with 𝑁 being the total number of clusters and 𝑛 the sample size, i.e.,
the number of cluster draws. If a cluster is selected more than once, there is
less information about the population mean in this sample than in a sample
with all clusters different. Selection of clusters with probabilities proportional
to size without replacement (ppswor) is not straightforward.

The problem is the computation of the inclusion probabilities of the clusters.

After we have selected a first cluster, we must adapt the sum of the sizes of
the 𝑁 − 1 remaining clusters and recompute the selection probabilities of
the remaining clusters in the second draw, etc. Section 6.4 of Lohr (1999)
nicely describes how the inclusion probabilities of the 𝑁 clusters in a cluster
random sample of size two, selected by ppswor, can be computed.

Many algorithms have been developed for ppswor sampling, see Tillé (2006) for
an overview, and quite a few of them are implemented in package sampling
(Tillé and Matei, 2021). In the next code chunk, function UPpivotal is used
to select a cluster random sample with ppswor. For an explanation of this
algorithm, see Subsection 8.2.2.

library(sampling)
n <- 6
pi <- n * M_cl / M
set.seed(314)
eps <- 1e-6
sampleind <- UPpivotal(pik = pi, eps = eps)
clusters <- sort(unique(grdVoorst$cluster))
clusters_sampled <- clusters[sampleind == 1]
mysample <- grdVoorst[grdVoorst$cluster %in% clusters_sampled, ]

The population mean can be estimated by the 𝜋 estimator by writing a few

lines of R code yourself or by using function svymean of package survey as
shown hereafter. Estimation of the sampling variance in pps sampling of
clusters without replacement is difficult1 . A simple solution is to treat the
cluster sample as a ppswr sample and to estimate the variance with Equation
(6.7). With small sampling fractions, this variance approximation is fine: the
overestimation of the variance is negligible. For larger sampling fractions,
various alternative variance approximations are developed, see Berger (2004)
for details. One of the methods is Brewer’s method, which is implemented in
function svydesign.

mysample$pi <- n * M_cl[mysample$cluster] / M

design_clppswor <- svydesign(

1 The problem is the computation of the joint inclusion probabilities of pairs of points.
6.3 Simple random sampling of clusters 103

id = ~ cluster, data = mysample, pps = "brewer", fpc = ~ pi)

svymean(~ z, design_clppswor)

mean SE
z 96.83 13.454

Another variance estimator implemented in function svydesign is the Hartley-

Rao estimator. The two estimated standard errors are nearly equal.

p2sum <- sum((n * M_cl[mysample$cluster] / M)ˆ2) / n

design_hr <- svydesign(
id = ~ cluster, data = mysample, pps = HR(p2sum), fpc = ~ pi)
svymean(~ z, design_hr)

mean SE
z 96.83 13.436

6.3 Simple random sampling of clusters

Suppose the clusters have unequal size, but we do not know the size of the
clusters, so that we cannot select the clusters with probabilities proportional to
their size. In this case, we may select the clusters by simple random sampling
without replacement. The inclusion probability of a cluster equals 𝑛/𝑁 with
𝑛 the number of selected clusters and 𝑁 the total number of clusters in the
population. This yields the following 𝜋 estimator of the population total:

𝑁
̂ =
𝑡(𝑧) ∑ 𝑡 (𝑧) . (6.9)
𝑛 u�∈u� u�

The population mean can be estimated by dividing this estimator of the

population total by the total number of units in the population 𝑀:

̂
𝑡(𝑧)
𝑧u�̄̂ (𝑧) = . (6.10)
𝑀
Alternatively, we may estimate the population mean by dividing the estimate
of the population total by the estimated population size:

𝑀u� 𝑁
̂= ∑
𝑀 = ∑ 𝑀u� . (6.11)
u�∈u�
𝜋u� 𝑛 u�∈u�
104 6 Cluster random sampling

This leads to the ratio estimator of the population mean:

̂
𝑡(𝑧)
̄̂
𝑧ratio (𝑧) = . (6.12)
̂
𝑀
The 𝜋 estimator and the ratio estimator are equal when the clusters are selected
with probabilities proportional to size. This is because the estimated population
size is equal to the true population size.

print(M_HT <- sum(1 / mysample$pi))

[1] 7528

However, when clusters of different size are selected with equal probabilities,
the two estimators are different. This is shown below. Six clusters are selected
by simple random sampling without replacement.

set.seed(314)
clusters <- sort(unique(grdVoorst$cluster))
units_cl <- sample(length(clusters), size = n, replace = FALSE)
clusters_sampled <- clusters[units_cl]
mysample <- grdVoorst[grdVoorst$cluster %in% clusters_sampled, ]

The 𝜋 estimate and the ratio estimate of the population mean are computed
for the selected sample.

N <- length(clusters)
mysample$pi <- n / N
tz_HT <- sum(mysample$z / mysample$pi)
mz_HT <- tz_HT / M
M_HT <- sum(1 / mysample$pi)
mz_ratio <- tz_HT / M_HT

The 𝜋 estimate equals 68.750 g kg-1 , and the ratio estimate equals 70.319 g
kg-1 . The 𝜋 estimate of the population mean can also be computed by first
computing totals of clusters, see Equations (6.9) and (6.10).

tz_cluster <- tapply(mysample$z, INDEX = mysample$cluster, FUN = sum)

pi_cluster <- n / N
tz_HT <- sum(tz_cluster / pi_cluster)
print(mz_HT <- tz_HT / M)

[1] 68.74994
6.3 Simple random sampling of clusters 105

The variance of the 𝜋 estimator of the population mean can be estimated by

first estimating the variance of the estimator of the total:

𝑛 𝑆̂2 (𝑡(𝑧))
𝑉(̂ 𝑡(𝑧))
̂ = 𝑁2 (1 − ) , (6.13)
𝑁 𝑛
and dividing this variance by the squared number of population units:

1
𝑉(̂ 𝑧)̄̂ = 2 𝑉(̂ 𝑡(𝑧))
̂ . (6.14)
𝑀

fpc <- 1 - n / N
v_tz <- Nˆ2 * fpc * var(tz_cluster) / n
se_mz_HT <- sqrt(v_tz / Mˆ2)

The estimated standard error equals 11.5 g kg-1 .

To compute the variance of the ratio estimator of the population mean, we
first compute residuals of cluster totals:

̂ u� ,
𝑒u� = 𝑡u� (𝑧) − 𝑏𝑀 (6.15)

with 𝑏̂ the ratio of the estimated population mean of the cluster totals to the
estimated population mean of the cluster sizes:

1
u� ∑u�∈u� 𝑡u�
𝑏̂ = 1
. (6.16)
u� ∑u�∈u� 𝑀u�

The variance of the ratio estimator of the population mean can be estimated
by

𝑛 1 ̂
𝑆2
̄̂
𝑉(̂ 𝑧ratio ) = (1 − ) 1 u�
, (6.17)
𝑁 ( u� ∑ 𝑀u� )2 𝑛
u�∈u�

with 𝑆
̂2 the estimated variance of the residuals.
u�

m_M_cl <- mean(M_cl[unique(mysample$cluster)])

b <- mean(tz_cluster) / m_M_cl
e_cl <- tz_cluster - b * M_cl[sort(unique(mysample$cluster))]
S2e <- var(e_cl)
print(se_mz_ratio <- sqrt(fpc * 1 / m_M_clˆ2 * S2e / n))

[1] 12.39371
106 6 Cluster random sampling

The ratio estimate can also be computed with function svymean of package
survey, which also provides an estimate of the standard error of the estimated
mean.

design_SIC <- svydesign(

id = ~ cluster, probs = ~ pi, fpc = ~ pi, data = mysample)
svymean(~ z, design_SIC)

mean SE
z 70.319 12.394

6.4 Stratified cluster random sampling

The basic sampling designs stratified random sampling (Chapter 4) and cluster
random sampling can be combined into stratified cluster random sampling.
So, instead of selecting simple random samples from the strata, within each
stratum clusters are randomly selected. Figure 6.5 shows a stratified cluster
random sample from Voorst. The strata consist of three 2 km × 1 km zones,
obtained by joining two neighbouring 1 km × 1 km zones (Figure 6.2). The
clusters are the same as before, i.e., E-W oriented transects within 1 km ×
1 km zones, with an inter-unit spacing of 100 m. Within each stratum, two
times a cluster is selected by ppswr. The stratification avoids the clustering of
the selected transects in one part of the study area. Compared to (unstratified)
cluster random sampling, the geographical spreading of the clusters is improved,
which may lead to an increase of the precision of the estimated population
mean. In Figure 6.5 and in the most western stratum, the two selected transects
are in the same 1 km × 1 km zone. The alternative would be to use the six
zones as strata, leading to an improved spreading of the clusters, but there
is also a downside with this design, see Exercise 3. Note that the selection
probabilities are now equal to

𝑝u�ℎ = 𝑀u� /𝑀ℎ , (6.18)

with 𝑀ℎ the total number of population units of stratum ℎ.

grdVoorst$zone_stratum <- as.factor(grdVoorst$zone)

levels(grdVoorst$zone_stratum) <- rep(c("a", "b", "c"), each = 2)
n_h <- c(2, 2, 2)
set.seed(324)
stratumlabels <- unique(grdVoorst$zone_stratum)
6.4 Stratified cluster random sampling 107

mysample <- NULL

for (i in 1:3) {
grd_h <- grdVoorst[grdVoorst$zone_stratum == stratumlabels[i], ]
mysample_h <- cl_ppswr(sframe = grd_h, n = n_h[i])
mysample <- rbind(mysample, mysample_h)
}

FIGURE 6.5: Stratified cluster random sample from Voorst, with three
strata. From each stratum two times a cluster is selected by ppswr.

The population mean is estimated by first estimating the stratum means using
Equation (6.5) at the level of the strata, followed by computing the weighted
average of the estimated stratum means using Equation (4.3). The variance of
the estimator of the population mean is estimated in the same way, by first
estimating the variance of the estimator of the stratum means using Equations
(6.7) and (6.8) at the level of the strata, followed by computing the weighted
average of the estimated variances of the estimated stratum means (Equation
(4.4)).

strata_size <- grdVoorst %>%

group_by(zone_stratum) %>%
summarise(M_h = n()) %>%
mutate(w_h = M_h / sum(M_h))
est <- mysample %>%
group_by(zone_stratum, cldraw) %>%
summarise(mz_cl = mean(z), .groups = "drop_last") %>%
summarise(mz_h = mean(mz_cl),
v_mz_h = var(mz_cl) / n()) %>%
left_join(strata_size, by = "zone_stratum") %>%
summarise(mz = sum(w_h * mz_h),
se_mz = sqrt(sum(w_hˆ2 * v_mz_h)))

The estimated mean equals 82.8 g kg-1 , and the estimated standard error equals
4.7 g kg-1 . The same estimates are obtained with function svymean. Weights
for the clusters are computed as before, but now at the level of the strata.
Note argument nest = TRUE, which means that the clusters are nested within
the strata.
108 6 Cluster random sampling

mysample$weights <- strata_size$M_h[mysample$zone_stratum] /

(M_cl[mysample$cluster] * n_h[mysample$zone_stratum])
design_strcluster <- svydesign(id = ~ cldraw, strata = ~ zone_stratum,
weights = ~ weights, data = mysample, nest = TRUE)
svymean(~ z, design_strcluster)

mean SE
z 82.796 4.6737

Exercises

3. Why is it attractive in stratified random cluster sampling to select

at least two clusters per stratum?
7
Two-stage cluster random sampling

As opposed to cluster random sampling in which all population units of a

cluster are observed (Chapter 6), in two-stage cluster random sampling not
all units of the selected clusters are observed, but only some. In two-stage
cluster random sampling the clusters will generally be contiguous groups of
units, for instance all points in a map polygon (the polygons on the map are
the clusters), whereas in single-stage cluster random sampling the clusters
generally are non-contiguous. The units to be observed are selected by random
subsampling of the randomly selected clusters. In two-stage cluster sampling,
the clusters are commonly referred to as primary sampling units (PSUs) and
the units selected in the second stage as the secondary sampling units (SSUs).
As with cluster random sampling, two-stage cluster random sampling may lead
to a strong spatial clustering of the selected population units in the study area.
This may save considerable time for fieldwork, and more population units can
be observed for the same budget. However, due to the spatial clustering the
estimates will generally be less precise compared to samples of the same size
selected by a design that leads to a much better spreading of the sampling
units throughout the study area, such as systematic random sampling.
In two-stage cluster random sampling, in principle any type of sampling design
can be used at the two stages, leading to numerous combinations. An example is
(SI,SI), in which both PSUs and SSUs are selected by simple random sampling.
Commonly, the PSUs have unequal size, i.e., the number of SSUs (finite
population) or the area (infinite population) are not equal for all PSUs. Think
for instance of the agricultural fields, forest stands, lakes, river sections, etc.
in an area. If the PSUs are of unequal size, then PSUs can best be selected
with probabilities proportional to their size (pps). Recall that in (one-stage)
cluster random sampling, I also recommended to select the clusters with
probabilities proportional to their size, see Chapter 6. If the total of the
study variable of a PSU is proportional to its size, then pps sampling leads
to more precise estimates compared to simple random sampling of PSUs.
Also, with pps sampling of PSUs, the estimation of means or totals and of
their sampling variances is much simpler compared to selection with equal
probabilities. Implementation of selection with probabilities proportional to
size is easiest when units are replaced (pps with replacement, ppswr). This
implies that a PSU might be selected more than once, especially if the total

109
110 7 Two-stage cluster random sampling

number of PSUs in the population is small compared to the number of PSU

draws (large sampling fraction in first stage).
Using a list as a sampling frame, the following algorithm can be used to select
𝑛 times a PSU by ppswr from a total of 𝑁 PSUs in the population:

1. Select randomly one SSU from the list with 𝑀 = ∑u�u�=1

𝑀u� SSUs
(𝑀u� is the number of SSUs of PSU 𝑗), and determine the PSU of
the selected SSU.

2. Repeat step 1 until 𝑛 selections have been made.

In the first stage, an SSU is selected in order to select a PSU. This may
seem unnecessarily complicated. The reason for this is that this procedure
automatically adjusts for the size of the PSUs (number of SSUs within a PSU),
i.e., a PSU is selected with probability proportional to its size. In the second
stage, a pre-determined number of SSUs, 𝑚u� , is selected every time PSU 𝑗 is
selected.
Note that the SSU selected in the first step of the two algorithms primarily
serves to identify the PSU, but these SSUs can also be used as selected SSUs.
The selection of a two-stage cluster random sample is illustrated again with
Voorst. Twenty-four 0.5 km squares are constructed that serve as PSUs.

Due to built-up areas, roads, etc., the PSUs in Voorst have unequal size, i.e.,
the number of SSUs (points, in our case) within the PSUs varies among the
PSUs.

cell_size <- 25
w <- 500 #width of zones
grdVoorst <- grdVoorst %>%
mutate(zone_s1 = s1 %>% findInterval(min(s1) + 1:11 * w + 0.5 * cell_size),
zone_s2 = s2 %>% findInterval(min(s2) + w + 0.5 * cell_size),
psu = str_c(zone_s1, zone_s2, sep = "_"))

In the next code chunk, a function is defined to select a two-stage cluster

random sample from an infinite population, discretised by a finite number of
points, being the centres of grid cells.

twostage <- function(sframe, psu, n, m) {

units <- sample(nrow(sframe), size = n, replace = TRUE)
mypsusample <- sframe[units, psu]
7.0 Two-stage cluster random sampling 111

ssunits <- NULL

for (psunit in mypsusample) {
ssunit <- sample(
x = which(sframe[, psu] == psunit), size = m, replace = TRUE)
ssunits <- c(ssunits, ssunit)
}
psudraw <- rep(c(1:n), each = m)
mysample <- data.frame(ssunits, sframe[ssunits, ], psudraw)
mysample
}

Note that both the PSUs and the SSUs are selected with replacement. If a
grid cell centre is selected, one point is selected fully randomly from that grid
cell. This is done by shifting the centre of the grid cell to a random point
within the selected grid cell with function jitter, see code chunk hereafter. In
every grid cell, there is an infinite number of points, so we must select the
grid cell centres with replacement. If a grid cell is selected more than once,
more than one point is selected from the associated grid cell. Column psudraw
in the output data frame of function twostage is needed in estimation because
PSUs are selected with replacement. In case a PSU is selected more than once,
multiple estimates of the mean of that PSU are used in estimation, see next
section.
In the next code chunk, function twostage is used to select four times a PSU
(𝑛 = 4), with probabilities proportional to size and with replacement (ppswr).
The second stage sample size equals 10 for all PSUs (𝑚u� = 10, 𝑗 = 1, … , 𝑁).
These SSUs are selected by simple random sampling.

n <- 4
m <- 10
cell_size <- 25
set.seed(314)
mysample <- grdVoorst %>%
twostage(psu = "psu", n = n, m = m) %>%
mutate(s1 = s1 %>% jitter(amount = cell_size / 2),
s2 = s2 %>% jitter(amount = cell_size / 2))

Figure 7.1 shows the selected sample.

112 7 Two-stage cluster random sampling

FIGURE 7.1: Two-stage cluster random sample from Voorst. PSUs are 0.5
km squares, built-up areas, roads, etc. excluded. Four times a PSU is selected
by ppswr. Each time a PSU is selected, ten SSUs (points) are selected from
that PSU by simple random sampling.

7.1 Estimation of population parameters

The population total can be estimated by substituting the estimated cluster
(PSU) totals in Equation (6.4). This yields the following estimator for the
population total:

𝑀 𝑡u�̂ (𝑧) 𝑀
̂ =
𝑡(𝑧) ∑ = ∑ 𝑧 ̄̂ , (7.1)
𝑛 u�∈u� 𝑀u� 𝑛 u�∈u� u�

where 𝑛 is the number of PSU selections and 𝑀u� is the total number of SSUs in
PSU 𝑗. This shows that the mean of cluster 𝑗, 𝑧u�̄ , is replaced by the estimated
mean of PSU 𝑗, 𝑧u�̄̂ . Dividing this estimator by the total number of population
units 𝑀 gives the pwr estimator of the population mean:

1
𝑧 ̄̂ = ∑ 𝑧 ̄̂ , (7.2)
𝑛 u�∈u� u�

with 𝑧u�̄̂ the estimated mean of the PSU 𝑗. With simple random sampling of
SSUs, this mean can be estimated by the sample mean of this PSU. Note the
two bars in 𝑧,̄̂ indicating that the population mean is estimated as the mean
of estimated PSU means. When 𝑚u� is equal for all PSUs, the sampling design
is self-weighting, i.e., the average of 𝑧 over all selected SSUs is an unbiased
estimator of the population mean.
For an infinite population of points, the population total is estimated by
multiplying the estimated population mean (Equation (7.2)) by the area of
the study area.
The sampling variance of the estimator of the mean with two-stage cluster
random sampling, PSUs selected with probabilities proportional to size with
7.1 Estimation of population parameters 113

replacement, SSUs selected by simple random sampling, with replacement in

case of finite populations, and 𝑚u� = 𝑚, 𝑗 = 1, … , 𝑁, is equal to (Cochran
(1977), equation (11.33)1 )

𝑆2 𝑆2
𝑉(𝑧)̄̂ = b + w , (7.3)
𝑛 𝑛𝑚
with

u�
2
𝑆2b = ∑ 𝑝u� (𝑧u�̄ − 𝑧)̄ (7.4)
u�=1

and

u�
𝑆2w = ∑ 𝑝u� 𝑆2u� , (7.5)
u�=1

with 𝑁 the total number of PSUs in the population, 𝑝u� = 𝑀u� /𝑀 the draw-by-
draw selection probability of PSU 𝑗, 𝑧u�̄ the mean of PSU 𝑗, 𝑧 ̄ the population
mean of 𝑧, and 𝑆2u� the variance of 𝑧 within PSU 𝑗:

u�u�
1
𝑆2u� = ∑ (𝑧 − 𝑧u�̄ )2 . (7.6)
𝑀u� u�=1 u�u�

The first term of Equation (7.3) is equal to the variance of Equation (6.6). This
variance component accounts for the variance of the true PSU means within
the population. The second variance component quantifies our additional
uncertainty about the population mean, as we do not observe all SSUs of
the selected PSUs, but only a subset (sample) of these units.

The sampling variance of the estimator of the population mean can simply be
estimated by

̂2 ̂
̂ 𝑧)̄̂ = 𝑆 (𝑧)̄ ,
𝑉( (7.7)
𝑛

with 𝑆
̂2 (𝑧)̄̂ the estimated variance of the estimated PSU means:

1
̂
𝑆2 (𝑧)̄̂ = ∑(𝑧 ̄̂ − 𝑧)̄̂ 2 , (7.8)
𝑛 − 1 u�∈u� u�

1 Equation (11.33) in Cochran (1977) is the variance estimator for the estimator of the
population total. In Exercise 5 you are asked to derive the variance estimator for the estimator
of the population mean from this variance estimator.
114 7 Two-stage cluster random sampling

with 𝑧u�̄̂ the estimated mean of PSU 𝑗 and 𝑧 ̄̂ the estimated population mean
(Equation (7.2)).

Neither the sizes of the PSUs, 𝑀u� , nor the secondary sample sizes 𝑚u� occur
in Equations (7.7) and (7.8). This simplicity is due to the fact that the PSUs
are selected with replacement and with probabilities proportional to their
size. The effect of the secondary sample sizes on the variance is implicitly
accounted for. To understand this, note that the larger 𝑚u� , the less variable
𝑧u�̄̂ , and the smaller its contribution to the variance.

Let us assume a linear model for the total costs: 𝐶 = 𝑐0 + 𝑐1 𝑛 + 𝑐2 𝑛𝑚, with
𝑐0 the fixed costs, 𝑐1 the costs per PSU, and 𝑐2 the costs per SSU. We want
to minimise the total costs, under the constraint that the variance of the
estimator of the population mean may not exceed 𝑉max . The total costs can
then be minimised by selecting (de Gruijter et al., 2006)

1 𝑐2
𝑛= (𝑆w 𝑆b √ + 𝑆2b ) (7.9)
𝑉max 𝑐1

PSUs and

𝑆w 𝑐1
𝑚= √ (7.10)
𝑆b 𝑐2

SSUs per PSU.

Conversely, given a budget 𝐶max , the optimal number of PSU selections can
be computed with (de Gruijter et al., 2006)

𝐶max 𝑆b
𝑛= √ , (7.11)
𝑆w 𝑐1 𝑐2 + 𝑆b 𝑐1

and 𝑚 as above.
In R the population mean and the sampling variance of the estimator of the
mean can be estimated as follows.

est <- mysample %>%

group_by(psudraw) %>%
summarise(mz_psu = mean(z)) %>%
summarise(mz = mean(mz_psu),
se_mz = sqrt(var(mz_psu) / n()))
7.1 Estimation of population parameters 115

The estimated mean equals 48.6 g kg-1 , and the estimated standard error
equals 0.0 g kg-1 . The sampling design is self-weighting, and so the estimated
mean is equal to the sample mean.

print(mean(mysample$z))

[1] 48.55792

The same estimate is obtained with functions svydesign and svymean of package
survey (Lumley, 2021). The estimator of the population total can be written
as a weighted sum of the observations with all weights equal to 𝑀/(𝑛 𝑚).
These weights are passed to function svydesign with argument weight.

library(survey)
M <- nrow(grdVoorst)
mysample$weights <- M / (n * m)
design_2stage <- svydesign(
id = ~ psudraw + ssunits, weight = ~ weights, data = mysample)
svymean(~ z, design_2stage, deff = "replace")

mean SE DEff
z 48.558 0.000 0

Similar to (one-stage) cluster random sampling, the estimated design effect is

much larger than 1.
A confidence interval estimate of the population mean can be computed with
method confint. The number of degrees of freedom equals the number of PSU
draws minus 1.

confint(svymean(~ z, design_2stage, df = degf(design_2stage), level = 0.95))

2.5 % 97.5 %
z 48.55792 48.55792

Figure 7.2 shows the approximated sampling distribution of the pwr estimator
of the mean soil organic matter (SOM) concentration with two-stage cluster
random sampling and of the 𝜋 estimator with simple random sampling from
Voorst, obtained by repeating the random sampling with each design and
estimation 10,000 times. For simple random sampling the sample size is equal
to 𝑛 × 𝑚.
The variance of the 10,000 means with two-stage cluster random sampling
equals 179.6 (g kg-1 )2 . This is considerably larger than with simple random
116 7 Two-stage cluster random sampling

FIGURE 7.2: Approximated sampling distribution of the pwr estimator of

the mean SOM concentration (g kg-1 ) in Voorst with two-stage cluster random
sampling (Twostage) and of the 𝜋 estimator with simple random sampling (SI).
The sample size with both sampling designs is 40. In two-stage sampling, four
times a PSU is selected by ppswr and ten SSUs (points) are selected per PSU
draw by simple random sampling.

sampling: 56.3 (g kg-1 )2 . The average of the estimated variances with two-stage
cluster random sampling equals 182.5 (g kg-1 )2 .
Optimal sample sizes for two-stage cluster random sampling (ppswr in first
stage, simple random sampling without replacement in second stage) can be
computed with function clusOpt2 of R package PracTools (Valliant et al.
(2021), Valliant et al. (2018)). This function requires as input various variance
measures, which can be computed with function BW2stagePPS, in case the study
variable is known for the whole population or estimated from a sample with
function BW2stagePPSe. This is left as an exercise (Exercise 5).

Exercises

1. Write an R script to compute for Voorst the true sampling variance

of the estimator of the mean SOM concentration for two-stage
cluster random sampling, PSUs selected by ppswr, 𝑛 = 4, and
𝑚 = 10, see Equation (7.3).
7.1 Estimation of population parameters 117

2. Do you expect that the standard error of the estimator of the

population mean with ten PSU draws (𝑛 = 10) and four SSUs per
PSU draw (𝑚 = 4) is larger or smaller than with four PSU draws
(𝑛 = 4) and ten SSUs per PSU draw (𝑚 = 10)?

3. Compute the optimal sample sizes 𝑛 and 𝑚 for a maximum variance

of the estimator of the mean SOM concentration of 1, 𝑐1 = 2 mone-
tary units, and 𝑐2 = 1 monetary unit, see Equations (7.9) and (7.10).

4. Compute the optimal sample sizes 𝑛 and 𝑚 for a budget of 100

monetary units, 𝑐1 = 2 monetary units, and 𝑐2 = 1 monetary unit,
see Equations (7.11) and (7.10).

5. Use function clusOpt2 of R package PracTools to compute optimal

sample sizes given the precision requirement for the estimated
population mean of Exercise 3 and given the budget of Exercise 4.
Use function BW2stagePPS to compute the variance measures needed
as input for function clusOpt2. Note that the precision requirement
of function clusOpt2 is the coefficient of variation of the estimated
population total, i.e., the standard deviation of the estimated
population total divided by the population total. Compute this
coefficient of variation from the maximum variance of the estimator
of the population mean used in Exercise 3.

6. The variance of the estimator for the population total is (Cochran,

1977):

2
1 u� 𝑡u� (𝑧) 1 u� 𝑀2u� (1 − 𝑓2u� )𝑆2u�
̂
𝑉(𝑡(𝑧)) = ∑ 𝑝u� ( − 𝑡(𝑧)) + ∑ ,
𝑛 u�=1 𝑝u� 𝑛 u�=1 𝑚u� 𝑝u�

(7.12)

with 𝑡(𝑧)
̂ and 𝑡(𝑧) the estimated and the true population total of
𝑧, respectively, 𝑡u� (𝑧) the total of PSU 𝑗, and 𝑝u� = 𝑀u� /𝑀. Use
𝑚u� = 𝑚, 𝑗 = 1, … , 𝑁, and 𝑓2u� = 0, i.e., sampling from infinite
population, or sampling of SSUs within PSUs by simple random
sampling with replacement from finite population. Derive the vari-
ance of the estimator for the population mean, Equation (7.3), from
Equation (7.12).
118 7 Two-stage cluster random sampling

7.2 Primary sampling units selected without

replacement
Similar to cluster random sampling, we may prefer to select the PSUs without
replacement. This leads to less strong spatial clustering of the sampling points,
especially with large sampling fractions of PSUs. Sampling without replacement
of PSUs can be done with function UPpivotal of package sampling (Tillé and
Matei, 2021), see Subsection 8.2.2. The second stage sample of SSUs is selected
with function strata of the same package, using the PSUs as strata.

library(sampling)
M_psu <- tapply(grdVoorst$z, INDEX = grdVoorst$psu, FUN = length)
n <- 6
pi <- n * M_psu / M
set.seed(314)
sampleind <- UPpivotal(pik = pi, eps = 1e-6)
psus <- sort(unique(grdVoorst$psu))
sampledpsus <- psus[sampleind == 1]
mysample_stage1 <- grdVoorst[grdVoorst$psu %in% sampledpsus, ]
units <- sampling::strata(mysample_stage1, stratanames = "psu",
size = rep(m, n), method = "srswor")
mysample <- getdata(mysample_stage1, units)
mysample$ssunits <- units$ID_unit
mysample$pi <- n * m / M
print(mean_HT <- sum(mysample$z / mysample$pi) / M)

[1] 100.039

The population mean can be estimated with function svymean of package survey.
To estimate the variance, a simple solution is to treat the two-stage cluster
random sample as a pps sample with replacement, so that variance can be
estimated with Equation (7.7). With small sampling fractions of PSUs, the
overestimation of the variance is negligible. With larger sampling fractions,
Brewer’s method is recommended, see Berger (2004) (option 2).

mysample$fpc1 <- n * M_psu[mysample$psu] / M

mysample$fpc2 <- m / M_psu[mysample$psu]
design_2stageppswor <- svydesign(id = ~ psu + ssunits, data = mysample,
pps = "brewer", fpc = ~ fpc1 + fpc2)
svymean(~ z, design_2stageppswor)

mean SE
z 100.04 19.883
7.3 Simple random sampling of primary sampling units 119

7.3 Simple random sampling of primary sampling units

Suppose the PSUs are for some reason not selected with probabilities propor-
tional to their size, but by simple random sampling without replacement. The
inclusion probabilities of the PSUs then equal 𝜋u� = 𝑛/𝑁, 𝑗 = 1, … , 𝑁, and the
population total can be estimated by (compare with Equation (6.9))

u�
𝑡u�̂ (𝑧) 𝑁 u�
̂ =∑
𝑡(𝑧) = ∑ 𝑡u�̂ (𝑧) , (7.13)
u�=1
𝜋u� 𝑛 u�=1

with 𝑡u�̂ (𝑧) an estimator of the total of PSU 𝑗. The population mean can be
estimated by dividing this estimator by the population size 𝑀.
Alternatively, we may estimate the population mean by dividing the estimate
of the population total by the estimated population size. The population size
can be estimated by the 𝜋 estimator, see Equation (6.11). The 𝜋 estimator
and the ratio estimator are equal when the PSUs are selected by ppswr, but
not so when the PSUs of different size are selected with equal probabilities.
This is shown below. First, a sample is selected by selecting both PSUs and
SSUs by simple random sampling without replacement.

library(sampling)
set.seed(314)
psus <- sort(unique(grdVoorst$psu))
ids_psu <- sample(length(psus), size = n, replace = FALSE)
sampledpsus <- psus[ids_psu]
mysample_stage1 <- grdVoorst[grdVoorst$psu %in% sampledpsus, ]
units <- sampling::strata(mysample_stage1, stratanames = "psu",
size = rep(m, n), method = "srswor")
mysample <- getdata(mysample_stage1, units)
mysample$ssunits <- units$ID_unit

The population mean is estimated by the 𝜋 estimator and the ratio estimator.

N <- length(unique(grdVoorst$psu))
M_psu <- tapply(grdVoorst$z, INDEX = grdVoorst$psu, FUN = length)
pi_psu <- n / N
pi_ssu <- m / M_psu[mysample$psu]
est <- mysample %>%
mutate(pi = pi_psu * pi_ssu,
z_piexpanded = z / pi) %>%
summarise(tz_HT = sum(z_piexpanded),
120 7 Two-stage cluster random sampling

mz_HT = tz_HT / M,
M_HT = sum(1 / pi),
mz_ratio = tz_HT / M_HT)

The 𝜋 estimate equals 79.0 g kg-1 , and the ratio estimate equals 79.8 g kg-1 . The
𝜋 estimate of the population mean can also be computed by first estimating
totals of PSUs, see Equation (7.13).

tz_psu <- tapply(mysample$z / pi_ssu, INDEX = mysample$psu, FUN = sum)

tz_HT <- sum(tz_psu / pi_psu)
(mz_HT <- tz_HT / M)

[1] 78.99646

The variance of the 𝜋 estimator of the population mean can be estimated by

first estimating the variance of the estimator of the PSU totals:

𝑛 𝑆̂2 (𝑡 ̂ (𝑧))
𝑉(̂ 𝑡(𝑧))
̂ = 𝑁2 (1 − ) u�
, (7.14)
𝑁 𝑛
and dividing this variance by the squared number of population units:

1
𝑉(̂ 𝑧)̄̂ = 2 𝑉(̂ 𝑡(𝑧))
̂ , (7.15)
𝑀
as shown in the code chunk below (the final line computes the standard error).

fpc <- 1 - n / N
v_tz <- Nˆ2 * fpc * var(tz_psu) / n
(se_mz_HT <- sqrt(v_tz / Mˆ2))

[1] 9.467406

The ratio estimator of the population mean and its standard error can be
computed with function svymean of package survey.

mysample$fpc1 <- N
mysample$fpc2 <- M_psu[mysample$psu]
design_2stage <- svydesign(
id = ~ psu + ssunits, fpc = ~ fpc1 + fpc2, data = mysample)
svymean(~ z, design_2stage)

mean SE
z 79.845 7.7341
7.4 Stratified two-stage cluster random sampling 121

The estimated standard error of the ratio estimator is slightly smaller than
the standard error of the 𝜋 estimator.

7.4 Stratified two-stage cluster random sampling

The basic sampling designs stratified random sampling (Chapter 4) and two-
stage cluster random sampling can be combined into stratified two-stage cluster
random sampling. Figure 7.3 shows a stratified two-stage cluster random sample
from Voorst. The strata are groups of eight PSUs within 2 km × 1 km blocks,
as before in stratified cluster random sampling (Figure 6.2). The PSUs are 0.5
km squares (built-up areas, roads, etc. excluded), as before in (unstratified)
two-stage cluster random sampling (Figure 7.1). Within each stratum two
times a PSU is selected by ppswr, and every time a PSU is selected, six SSUs
(points) are selected by simple random sampling. The stratification avoids the
clustering of the selected PSUs in one part of the study area. Compared to
(unstratified) two-stage cluster random sampling, the geographical spreading
of the PSUs is somewhat improved, which may lead to an increase of the
precision of the estimated population mean.

n_h <- rep(2, 3)

m <- 6
set.seed(314)
stratumlabels <- unique(grdVoorst$zone_stratum)
mysample <- NULL
for (i in 1:3) {
grd_h <- grdVoorst[grdVoorst$zone_stratum == stratumlabels[i], ]
mysample_h <- twostage(sframe = grd_h, psu = "psu", n = n_h[i], m = m)
mysample <- rbind(mysample, mysample_h)

FIGURE 7.3: Stratified two-stage random sample from Voorst. Strata are
groups of eight PSUs (0.5 km squares) within 2 km × 1 km blocks. From
each stratum two times a PSU is selected by ppswr, and six SSUs (points) are
selected per PSU draw by simple random sampling.
122 7 Two-stage cluster random sampling

}
mysample$s1 <- jitter(mysample$s1, amount = cell_size / 2)
mysample$s2 <- jitter(mysample$s2, amount = cell_size / 2)

The population mean can be estimated in much the same way as with stratified
cluster random sampling. With function svymean this is an easy task.

N_h <- tapply(grdVoorst$psu, INDEX = grdVoorst$zone_stratum,

FUN = function(x) {
length(unique(x))
})
M_h <- tapply(grdVoorst$z, INDEX = grdVoorst$zone_stratum, FUN = length)
mysample$w1 <- N_h[mysample$zone_stratum]
mysample$w2 <- M_h[mysample$zone_stratum]
design_str2stage <- svydesign(id = ~ psudraw + ssunits, strata = ~ zone_stratum,
weights = ~ w1 + w2, data = mysample, nest = TRUE)
svymean(~ z, design_str2stage)

mean SE
z 73.654 0
8
Sampling with probabilities proportional to
size

In simple random sampling, the inclusion probabilities are equal for all popu-
lation units. The advantage of this is simple and straightforward statistical
inference. With equal inclusion probabilities the unweighted sample mean is
an unbiased estimator of the spatial mean, i.e., the sampling design is self-
weighting. However, in some situations equal probability sampling is not very
efficient, i.e., given the sample size the precision of the estimated mean or total
will be relatively low. An example is the following. In order to estimate the
total area of a given crop in a country, a raster of square cells of, for instance,
1 km × 1 km is constructed and projected on the country. The square cells
are the population units, and these units serve as the sampling units. Note
that near the country border cells cross the border. Some of them may contain
only a few hectares of the target population, the country under study. We do
not want to select many of these squares with only a few hectares of the study
area, as intuitively it is clear that this will result in a low precision of the
estimated crop area. In such situation it can be more efficient to select units
with probabilities proportional to the area of the target population within
the squares, so that small units near the border have a smaller probability of
being selected than interior units. Actually, the sampling units are not the
square cells, but the pieces of land obtained by overlaying the cells and the GIS
map of the country under study. As a consequence, the sampling units have
unequal size. The sampling units of unequal size are selected by probabilities
proportional to their size (pps).

In Chapters 6 and 7 pps sampling was already used to select clusters (primary
sampling units) of population units. In this chapter the individual population
units (elementary sampling units) are selected with probabilities proportional
to size.

If we have a GIS map of land use categories such as agriculture, built-up areas,
water bodies, forests, etc., we may use this file to further adapt the selection
probabilities. The crop will be grown in agricultural areas only, so we expect
small crop areas in cells largely covered by non-agricultural land. As a size
measure in computing the selection probabilities, we may use the agricultural
area, as represented in the GIS map, in the country under study within the

123
124 8 Sampling with probabilities proportional to size

cells. Note that size now has a different meaning. It does not refer to the area
of the sampling units anymore, but to an ancillary variable that we expect to
be related to the study variable, i.e., the crop area. When the crop area per
cell is proportional to the agricultural area per cell, then the precision of the
estimated total area of the crop can be increased by selecting the cells with
probabilities proportional to the agricultural area.
In this example the sampling units have an area. However, sampling with
probabilities proportional to size is not restricted to areal sampling units, but
can also be used for selecting points. If we have a map of an ancillary variable
that is expected to be positively related to the study variable, this ancillary
variable can be used as a size measure. For instance, in areas where soil organic
matter shows a positive relation with elevation, it can be efficient to select
sampling points with a selection probability proportional to this environmental
variable. The ancillary variable must be strictly positive for all points.
Sampling units can be selected with probabilities proportional to their size
(pps) with or without replacement. This distinction is immaterial for infinite
populations, as in sampling points from an area. pps sampling with replacement
(ppswr) is much easier to implement than pps sampling without replacement
(ppswor). The problem with ppswor is that after each draw the selected unit
is removed from the sampling frame, so that the sum of the size variable
over all remaining units changes and as a result the draw-by-draw selection
probabilities of the units.
pps sampling is illustrated with the simulated map of poppy area per 5 km
square in the province of Kandahar (Figure 1.6). The first six rows of the data
frame are shown below. Variable poppy is the study variable, variable agri is
the agricultural area within the 5 km squares, used as a size variable.

grdKandahar

# A tibble: 965 x 4
s1 s2 poppy agri
<dbl> <dbl> <dbl> <dbl>
1 809232. 3407627. 0.905 65.7
2 814232. 3412627. 0.00453 15.6
3 794232. 3417627. 11.3 17.6
4 809232. 3417627. 0.110 14.0
5 814232. 3417627. 0.0344 22.2
6 819232. 3417627. 0.143 13.3
7 794232. 3422627. 3.66 34.1
8 799232. 3422627. 3.66 6.12
9 809232. 3422627. 0.688 10.6
10 814232. 3422627. 4.79 130.
# ... with 955 more rows
8.1 Probability-proportional-to-size sampling with replacement 125

8.1 Probability-proportional-to-size sampling with

replacement
In the first draw, a sampling unit is selected with probability 𝑝u� = 𝑥u� /𝑡(𝑥),
with 𝑥u� the size variable for unit 𝑘 and 𝑡(𝑥) = ∑u�=1 𝑥u� the population total
u�

of the size variable. The selected unit is then replaced, and these two steps are
repeated 𝑛 times. Note that with this sampling design population units can be
selected more than once, especially with large sampling fractions 𝑛/𝑁.
The population total can be estimated by the pwr estimator:

1 𝑧
̂ =
𝑡(𝑧) ∑ u� , (8.1)
𝑛 u�∈u� 𝑝u�

where 𝑛 is the sample size (number of draws). The population mean can be
estimated by the estimated population total divided by the population size
𝑁. With independent draws, the sampling variance of the estimator of the
population total can be estimated by

2
1 𝑧
𝑉(̂ 𝑡(𝑧))
̂ = ∑ ( u� − 𝑡(𝑧))
̂ . (8.2)
𝑛 (𝑛 − 1) u�∈u� 𝑝u�

The sampling variance of the estimator of the mean can be estimated by the
variance of the estimator of the total divided by 𝑁2 .
As a first step, I check whether the size variable is strictly positive in our
case study of Kandahar. The minimum equals 0.307 m2 , so this is the case. If
there are values equal to or smaller than 0, these values must be replaced by a
small number, so that all units have a positive probability of being selected.
Then the draw-by-draw selection probabilities are computed, and the sample
is selected using function sample.

grdKandahar$p <- grdKandahar$agri / sum(grdKandahar$agri)

N <- nrow(grdKandahar)
n <- 40
set.seed(314)
units <- sample(N, size = n, replace = TRUE, prob = grdKandahar$p)
mysample <- grdKandahar[units, ]

To select the units, computing the selection probabilities is not strictly needed.
Exactly the same units are selected when the agricultural area within the
units (variable agri in the data frame) is used in argument prob of sample.
126 8 Sampling with probabilities proportional to size

Four units are selected twice.

table_frq <- table(units) %>% data.frame()

print(table_frq[table_frq$Freq > 1, ])

units Freq
9 278 2
13 334 2
14 336 2
24 439 2

Figure 8.1 shows the selected sampling units, plotted on a map of the agricul-
tural area within the units which is used as a size variable.

FIGURE 8.1: Sample of size 40 from Kandahar, selected with probabilities

proportional to agricultural area with replacement. Four units are selected
twice, so that the number of distinct units is 36.

The next code chunk shows how the population total of the poppy area can be
estimated, using Equation (8.1), as well as the standard error of the estimator
of the population total (square root of estimator of Equation (8.2)). As a
first step, the observations are inflated, or expanded, through division of the
observations by the selection probabilities of the corresponding units.

z_pexpanded <- mysample$poppy / mysample$p

tz <- mean(z_pexpanded)
se_tz <- sqrt(var(z_pexpanded) / n)
8.1 Probability-proportional-to-size sampling with replacement 127

The estimated total equals 65,735 ha, with a standard error of 12,944 ha. The
same estimates are obtained with package survey (Lumley, 2021).

library(survey)
mysample$weight <- 1 / (mysample$p * n)
design_ppswr <- svydesign(id = ~ 1, data = mysample, weights = ~ weight)
svytotal(~ poppy, design_ppswr)

total SE
poppy 65735 12944

In ppswr sampling, a sampling unit can be selected more than once, especially
with large sampling fractions 𝑛/𝑁. This may decrease the sampling efficiency.
With large sampling fractions, the alternative is pps sampling without replace-
ment (ppswor), see next section.
The estimators of Equations (8.1) and (8.2) can also be used for infinite
populations. For infinite populations, the probability that a unit is selected
more than once is zero.

Exercises

1. Write an R script to select a pps with replacement sample from

Eastern Amazonia (grdAmazonia in package sswr) to estimate the pop-
ulation mean of aboveground biomass (AGB), using log-transformed
short-wave infrared radiation (SWIR2) as a size variable.
• The correlation of AGB and lnSWIR2 is negative. The first step
is to compute an appropriate size variable, so that the larger
the size variable, the larger the selection probability. Multiply
the lnSWIR2 values by -1. Then add a small value, so that the
size variable becomes strictly positive.
• Select in a for-loop 1,000 times a ppswr sample of size 100
(𝑛 = 100), and estimate from each sample the population mean
of AGB with the pwr estimator (Hansen-Hurwitz estimator)
and its sampling variance. Compute the variance of the 1,000 es-
timated population means and the mean of the 1,000 estimated
variances. Make a histogram of the 1,000 estimated means.
• Compute the true sampling variance of the 𝜋 estimator with
simple random sampling with replacement and the same sample
size.
• Compute the gain in precision by the ratio of the variance of
the estimator of the mean with simple random sampling to the
variance with ppswr.
128 8 Sampling with probabilities proportional to size

8.2 Probability-proportional-to-size sampling without

replacement
The alternative to pps sampling with replacement (ppswr) is pps sampling
without replacement (ppswor). In ppswor sampling the inclusion probabilities
are proportional to a size variable, not the draw-by-draw selection probabilities
as in ppswr. For this reason, ppswor sampling is referred to as 𝜋ps sampling by
Särndal et al. (1992). ppswor sampling starts with assigning target inclusion
probabilities to all units in the population. With inclusion probabilities pro-
portional to a size variable 𝑥 the target inclusion probabilities are computed
by 𝜋u� = 𝑛 𝑥u� / ∑u�=1 𝑥u� , 𝑘 = 1, … , 𝑁.
u�

8.2.1 Systematic pps sampling without replacement

Many algorithms are available for ppswor sampling, see Tillé (2006) for an
overview. A simple, straightforward method is systematic ppswor sampling. Two
subtypes can be distinguished, systematic ppswor sampling with fixed frame
order and systematic ppswor sampling with random frame order (Rosén, 1997).
Given some order of the units, the cumulative sum of the inclusion probabilities
is computed. Each population unit is then associated with an interval of
cumulative inclusion probabilities. The larger the inclusion probability of
a unit, the wider the interval. Then a random number from the uniform
distribution is drawn, which serves as the start of a one-dimensional systematic
sample of size 𝑛 with an interval of 1. Finally, the units are determined for
which the systematic random values are in the interval of cumulative inclusion
probabilities, see Figure 8.2 for ten population units and a sample size of four.
The units selected are 2, 5, 7, and 9. Note that the sum of the interval lengths
equals the sample size. Further note that a unit cannot be selected more than
once because the inclusion probabilities are < 1 and the sampling interval
equals 1.

library(sampling)
set.seed(314)
N <- 10
n <- 4
x <- rnorm(N, mean = 20, sd = 5)
pi <- inclusionprobabilities(x, n)
print(data.frame(id = seq_len(N), x, pi))

id x pi
1 1 13.55882 0.3027383
8.2 Probability-proportional-to-size sampling without replacement 129

2 2 23.63731 0.5277684
3 3 15.83538 0.3535687
4 4 16.48162 0.3679978
5 5 20.63624 0.4607613
6 6 18.32529 0.4091630
7 7 16.50655 0.3685545
8 8 20.06336 0.4479702
9 9 22.94495 0.5123095
10 10 11.15957 0.2491684

cumsumpi <- c(0, cumsum(pi))

start <- runif(1, min = 0, max = 1)
sys <- 0:(n - 1) + start
print(units <- findInterval(sys, cumsumpi))

[1] 2 5 7 9

FIGURE 8.2: Systematic random sample along a line with unequal inclusion
probabilities.

Sampling efficiency can be increased by ordering the units by the size variable
(Figure 8.3). With this design, the third, fourth, fifth, and second units in the
original frame are selected, with sizes 15.8, 16.5, 20.6, and 23.6, respectively.
Ordering the units by size leads to a large within-sample and a small between-
sample variance of the size variable 𝑥. If the study variable is proportional to
the size variable, this results in a smaller sampling variance of the estimator of
the mean of the study variable. A drawback of systematic ppswor sampling
with fixed order is that no unbiased estimator of the sampling variance exists.

FIGURE 8.3: Systematic random sample along a line with unequal inclusion
probabilities. Units are ordered by size.

A small simulation study is done next to see how much gain in precision can be
achieved by ordering the units by size. A size variable 𝑥 and a study variable 𝑧
are simulated by drawing 1,000 values from a bivariate normal distribution with
a correlation coefficient of 0.8. Function mvrnorm of package MASS (Venables
and Ripley, 2002) is used for the simulation.
130 8 Sampling with probabilities proportional to size

library(MASS)
rho <- 0.8
mu1 <- 10; sd1 <- 2
mu2 <- 15; sd2 <- 4
mu <- c(mu1, mu2)
sigma <- matrix(
data = c(sd1ˆ2, rep(sd1 * sd2 * rho, 2), sd2ˆ2),
nrow = 2, ncol = 2)
N <- 1000
set.seed(314)
dat <- as.data.frame(mvrnorm(N, mu = mu, Sigma = sigma))
names(dat) <- c("z", "x")
head(dat)

z x
1 9.462930 9.149784
2 12.605847 17.306046
3 7.892686 11.979986
4 7.945021 12.567608
5 11.004325 15.165744
6 10.369943 13.258177

Twenty units are selected by systematic ppswor sampling with random order
and ordered by size. This is repeated 10,000 times.
The standard deviation of the 10,000 estimated means with systematic ppswor
sampling with random order is 0.336, and when ordered by size 0.321. So,
a small gain in precision is achieved through ordering the units by size. For
comparison, I also computed the standard error for simple random sampling
without replacement (SI) of the same size. The standard error with this basic
sampling design is 0.424.

8.2.2 The pivotal method

Another interesting algorithm for ppswor sampling is the pivotal method

(Deville and Tillé, 1998). A nice adaptation of this algorithm, the local piv-
otal method, leading to samples with improved (geographical) spreading, is
described in Section 9.2. In the pivotal method, the 𝑁-vector with inclusion
probabilities is successively updated to a vector with indicators. If the indicator
value for sampling unit 𝑘 becomes 1, then this sampling unit is selected, if it
becomes 0, then it is not selected. The updating algorithm can be described
as follows:
8.2 Probability-proportional-to-size sampling without replacement 131

1. Select randomly two units 𝑘 and 𝑙 with 0 < 𝜋u� < 1 and 0 < 𝜋u� < 1.

2. If 𝜋u� + 𝜋u� < 1, then update the probabilities by

(0, 𝜋u� + 𝜋u� ) with probability u� u�+u�

u�

(𝜋′u� , 𝜋′u� ) = { u� u�
, (8.3)
(𝜋u� + 𝜋u� , 0) with probability u� +u�
u�u�
u� u�

and if 𝜋u� + 𝜋u� ≥ 1, update the probabilities by

(1, 𝜋u� + 𝜋u� − 1) with probability 2−(u�

1−u�u�
(𝜋′u� , 𝜋′u� ) = { u� +u�u� )
. (8.4)
(𝜋u� + 𝜋u� − 1, 1) with probability 2−(u�
1−u�u�
+u� )
u� u�

3. Replace (𝜋u� , 𝜋u� ) by (𝜋′u� , 𝜋′u� ), and repeat the first two steps until
each population unit is either selected (inclusion probability equals
1) or not selected (inclusion probability equals 0).

In words, when the sum of the inclusion probabilities is smaller than 1, the
updated inclusion probability of one of the units will become 0, which means
that this unit will not be sampled. The inclusion probability of the other unit
will become the sum of the two inclusion probabilities, which means that the
probability increases that this unit will be selected in one of the subsequent
iterations. The probability of a unit of being excluded from the sample is
proportional to the inclusion probability of the other unit, so that the larger
the inclusion probability of the other unit, the larger the probability that it
will not be selected.
When the sum of the inclusion probabilities of the two units is larger than or
equal to 1, then one of the units is selected (updated inclusion probability is
one), while the inclusion probability of the other unit is lowered by 1 minus
the inclusion probability of the selected unit. The probability of being selected
is proportional to the complement of the inclusion probability of the other
unit. After the inclusion probability of a unit has been updated to either 0 or
1, this unit cannot be selected anymore in the next iteration.
With this ppswor design, the population total can be estimated by the 𝜋
estimator, Equation (2.2). The 𝜋 estimator of the mean is simply obtained by
dividing the estimator for the total by the population size 𝑁.

The inclusion probabilities 𝜋u� used in the 𝜋 estimator are not the final
probabilities obtained with the local pivotal method, which are either 0 or 1,
but the initial inclusion probabilities.
132 8 Sampling with probabilities proportional to size

An alternative estimator of the population mean is the ratio estimator, also

known as the Hájek estimator:

∑u�∈u� 𝑤u� 𝑧u�

̄̂
𝑧Hajek = , (8.5)
∑u�∈u� 𝑤u�

with 𝑤u� = 1/𝜋u� . The denominator is an estimator of the population size 𝑁.

The Hájek estimator of the population total is obtained by multiplying the
Hájek estimator of the mean with the population size 𝑁. Recall that the ratio
estimator of the population mean was presented before in the chapters on
systematic random sampling (Equation (5.4)), cluster random sampling with
simple random sampling of clusters (Equation (6.12)), and two-stage cluster
random sampling with simple random sampling of PSUs. These sampling
designs have in common that the sample size (for cluster random sampling,
the number of SSUs) is random.
Various functions in package sampling (Tillé and Matei, 2021) can be used to
select a ppswor sample. In the code chunk below, I use function UPrandompivotal.
With this function, the order of the population units is randomised before
function UPpivotal is used. Argument pi is a numeric with the inclusion proba-
bilities. These are computed with function inclusionprobabilities. Recall that
𝜋u� = 𝑛 𝑥u� /𝑡(𝑥). The sum of the inclusion probabilities should be equal to the
sample size 𝑛. Function UPpivotal returns a numeric of length 𝑁 with elements
1 and 0, 1 if the unit is selected, 0 if it is not selected.

library(sampling)
n <- 40
size <- ifelse(grdKandahar$agri < 1E-12, 0.1, grdKandahar$agri)
pi <- inclusionprobabilities(size, n)
set.seed(314)
sampleind <- UPrandompivotal(pik = pi)
mysample <- data.frame(grdKandahar[sampleind == 1, ], pi = pi[sampleind == 1])
nrow(mysample)

[1] 39

As can be seen, not 40 but only 39 units are selected. The reason is that function
UPrandompivotal uses a very small number that can be set with argument eps. If
the updated inclusion probability of a unit is larger than the complement of
this small number eps, the unit is treated as being selected. The default value
of eps is 10−6 . If we replace sampleind == 1 by sampleind > 1 - eps, 40 units are
selected.

eps <- 1e-6

mysample <- data.frame(
8.2 Probability-proportional-to-size sampling without replacement 133

grdKandahar[sampleind > 1 - eps, ], pi = pi[sampleind > 1 - eps])

nrow(mysample)

[1] 40

The total poppy area can be estimated from the ppswor sample by

tz_HT <- sum(mysample$poppy / mysample$pi)

tz_Hajek <- N * sum(mysample$poppy / mysample$pi) / sum(1 / mysample$pi)

The total poppy area as estimated with the 𝜋 estimator equals 88,501 ha. The
Hájek estimator results in a much smaller estimated total: 62,169 ha.
The 𝜋 estimate can also be computed with function svytotal of package survey,
which also provides an approximate estimate of the standard error. Various
methods are implemented in function svydesign for approximating the standard
error. These methods differ in the way the pairwise inclusion probabilities are
approximated from the unitwise inclusion probabilities. These approximated
pairwise inclusion probabilities are then used in the 𝜋 variance estimator or the
Yates-Grundy variance estimator. In the next code chunks, Brewer’s method is
used, see option 2 of Brewer’s method in Berger (2004), as well as Hartley-Rao’s
method for approximating the variance.

library(survey)
design_ppsworbrewer <- svydesign(
id = ~ 1, data = mysample, pps = "brewer", fpc = ~ pi)
svytotal(~ poppy, design_ppsworbrewer)

total SE
poppy 88501 14046

p2sum <- sum(mysample$piˆ2) / n

design_ppsworhr <- svydesign(
id = ~ 1, data = mysample, pps = HR(p2sum), fpc = ~ pi)
svytotal(~ poppy, design_ppsworhr)

total SE
poppy 88501 14900

In package samplingVarEst (Escobar et al., 2019), also various functions are

available for approximating the variance: VE.Hajek.Total.NHT, VE.HT.Total.NHT,
and VE.SYG.Total.NHT. The first variance approximation is the Hájek-Rosén
variance estimator, see equation (4.3) in Rosén (1997). The latter two functions
require the pairwise inclusion probabilities, which can be estimated by function
Pkl.Hajek.s.
134 8 Sampling with probabilities proportional to size

library(samplingVarEst)
se_tz_Hajek <- sqrt(VE.Hajek.Total.NHT(mysample$poppy, mysample$pi))
pikl <- Pkl.Hajek.s(mysample$pi)
se_tz_HT <- sqrt(VE.HT.Total.NHT(mysample$poppy, mysample$pi, pikl))
se_tz_SYG <- sqrt(VE.SYG.Total.NHT(mysample$poppy, mysample$pi, pikl))

The three approximated standard errors are 14,045, 14,068, and 14,017 ha.
The differences are small when related to the estimated total.
Figure 8.4 shows the approximated sampling distribution of estimators of
the total poppy area with ppswor sampling and simple random sampling
without replacement of size 40, obtained by repeating the random sampling
with each design and estimation 10,000 times. With the ppswor samples, the
total poppy area is estimated by the 𝜋 estimator and the Hájek estimator. For
each ppswor sample, the variance of the 𝜋 estimator is approximated by the
Hájek-Rosén variance estimator (using function VE.Hajek.Total.NHT of package
samplingVarEst).

FIGURE 8.4: Approximated sampling distribution of the 𝜋 estimator

(ppswor.HT) and the Hájek estimator (ppswor.Hajek) of the total poppy
area (ha) in Kandahar with ppswor sampling of size 40, and of the 𝜋 estimator
with simple random sampling without replacement (SI) of size 40.

Sampling design ppswor in combination with the 𝜋 estimator is clearly much

more precise than simple random sampling. The standard deviation of the
10,000 𝜋 estimates of the total poppy area with ppswor equals 11,684 ha. The
8.2 Probability-proportional-to-size sampling without replacement 135

average of the square root of the Hájek-Rosén approximated variances equals

12,332 ha.
Interestingly, with ppswor sampling the variance of the 10,000 Hájek estimates
is much larger than that of the 𝜋 estimates. The standard deviation of the
10,000 Hájek estimates with ppswor sampling is about equal to that of the 𝜋
estimates with simple random sampling: 31,234 ha and 33,178 ha, respectively.

Exercises

2. A field with poppy was found outside Kandahar in a selected

sampling unit crossing the boundary. Should this field be included
in the sum of the poppy area of that sampling unit?

3. In another sampling unit, a poppy field was encountered in Kandahar

but in the area represented as non-agricultural in the GIS map.
Should this field be included in the sum of that sampling unit?
9
Balanced and well-spread sampling

In this chapter two related but fundamentally different sampling designs are
described and illustrated. The similarity and difference are shortly outlined
below, but hopefully will become clearer in following sections.
Roughly speaking, for a balanced sample the sample means of covariates are
equal to the population means of these covariates. When the covariates are
linearly related to the study variable, this may yield a more precise estimate
of the population mean or total of the study variable.
A well-spread sample is a sample with a large range of values for the covariates,
from small to large values, but also including intermediate values. In more
technical terms: the sampling units are well-spread along the axes spanned
by the covariates. If the spatial coordinates are used as covariates (spreading
variables), this results in samples that are well-spread in geographical space.
Such samples are commonly referred to as spatially balanced samples, which is
somewhat confusing, as the geographical spreading is not implemented through
balancing on the geographical coordinates. On the other hand, the averages of
the spatial coordinates of a sample well-spread in geographical space will be
close to the population means of the coordinates. Therefore, the sample will
be approximately balanced on the spatial coordinates (Grafström and Schelin,
2014). The reverse is not true: with balanced sampling, the spreading of the
sampling units in the space spanned by the balancing variables can be poor. A
sample with all values of a covariate used in balancing near the population
mean of that variable has a poor spreading along the covariate axis, but can
still be perfectly balanced.

9.1 Balanced sampling

Balanced sampling is a sampling method that exploits one or more quantitative
covariates that are related to the study variable. The idea behind balanced
sampling is that, if we know the mean of the covariates, then the sampling
efficiency can be increased by selecting a sample whose averages of the covariates
must be equal to the population means of the covariates.

137
138 9 Balanced and well-spread sampling

Let me illustrate balanced sampling with a small simulation study. The sim-
ulated population shown in Figure 9.1 shows a linear trend from West to
East and besides a trend from South to North. Due to the West-East trend,
the simulated study variable 𝑧 is correlated with the covariate Easting and,
due to the South-North trend, also with the covariate Northing. To estimate
the population mean of the simulated study variable, intuitively it is attractive
to select a sample with an average of the Easting coordinate that is equal to
the population mean of Easting (which is 10). Figure 9.1 (subfigure on the
left) shows such a sample of size four; we say that the sample is ‘balanced’ on
the covariate Easting. The sample in the subfigure on the right is balanced on
Easting as well as on Northing.

FIGURE 9.1: Sample balanced on Easting (E) and on Easting and Northing
(E and N).

9.1.1 Balanced sample vs. balanced sampling design

We must distinguish a balanced sample from a balanced sampling design. A

sampling design is balanced on a covariate 𝑥 when all possible samples that
can be generated by the design are balanced on 𝑥.

Simple random sampling is not a balanced sampling design, because for many
simple random samples the sample mean of the balancing variable 𝑥 is not
equal to the population mean of 𝑥. Only the expectation of the sample mean
of 𝑥, i.e., the mean of the sample means obtained by selecting an infinite
number of simple random samples, equals the population mean of 𝑥.
9.1 Balanced sampling 139

Figure 9.2 shows for 1,000 simple random samples the squared error of the
estimated population mean of the study variable 𝑧 against the difference
between the sample mean of balancing variable Easting and the population
mean of Easting. Clearly, the larger the absolute value of the difference, the
larger on average the squared error. So, to obtain a precise and accurate
estimate of the population mean of 𝑧, we better select samples with a difference
close to 0.

FIGURE 9.2: Squared error in the estimated population mean of 𝑧 against

the difference between the sample mean and the population mean of Easting,
for 1,000 simple random samples of size four selected from the population
shown in Figure 9.1.

Using only Easting as a balancing variable reduces the sampling variance of the
estimator of the mean substantially. Using Easting and Northing as balancing
variables further reduces the sampling variance. See Table 9.1.

9.1.2 Unequal inclusion probabilities

Until now we have assumed that the inclusion probabilities of the population
units are equal, but this is not a requirement for balanced sampling designs. A
140 9 Balanced and well-spread sampling

TABLE 9.1: Sampling variance of the 𝜋 estimator of the mean for simple
random sampling (SI) and balanced sampling of four units.

Sampling design Balancing variables Sampling variance

SI - 39.70
Balanced Easting 14.40
Balanced Easting and Northing 9.77

more general definition of a balanced sampling design is as follows. A sampling

design is balanced on variable 𝑥 when for all samples generated by the design
the 𝜋 estimate of the population total of 𝑥 equals the population total of 𝑥:

u�
𝑥u�
∑ = ∑ 𝑥u� . (9.1)
𝜋
u�∈u� u� u�=1

Similar to the regression estimator explained in the next chapter (Subsection

10.1.1), balanced sampling exploits the linear relation between the study
variable and one or more covariates. When using the regression estimator, this
is done at the estimation stage. Balanced sampling does so at the sampling
stage. For a single covariate the regression estimator of the population total
equals (see Equation (10.10))

̂
𝑡regr (𝑧) = 𝑡u�̂ (𝑧) + 𝑏̂ (𝑡(𝑥) − 𝑡u�̂ (𝑥)) , (9.2)

with 𝑡u�̂ (𝑧) and 𝑡u�̂ (𝑥) the 𝜋 estimators of the population total of the study
variable 𝑧 and the covariate 𝑥, respectively, 𝑡(𝑥) the population total of the
covariate, and 𝑏̂ the estimated slope parameter (see hereafter). With a perfectly
balanced sample the second term in the regression estimator, which adjusts
the 𝜋 estimator, equals zero.
Balanced samples can be selected with the cube algorithm of Deville and Tillé
(2004). The population total and mean can be estimated by the 𝜋 estimator.
The approximated variance of the 𝜋 estimator of the population mean can be
estimated by (Deville and Tillé (2005), Grafström and Tillé (2013))

2
1 𝑛 𝑒
𝑉(̂ 𝑧)̄̂ = 2 ∑ 𝑐u� ( u� ) , (9.3)
𝑁 𝑛 − 𝑝 u�∈u� 𝜋u�

with 𝑝 the number of balancing variables, 𝑐u� a weight for unit 𝑘 (see hereafter),
and 𝑒u� the residual of unit 𝑘 given by

𝑒u� = 𝑧u� − 𝐱T ̂
u� 𝐛 , (9.4)
9.1 Balanced sampling 141

with 𝐱u� a vector of length 𝑝 with the balancing variables for unit 𝑘, and 𝐛̂ the
estimated population regression coefficients, given by

−1
𝐱 𝐱 T 𝐱u� 𝑧u�
𝐛̂ = (∑ 𝑐u� u� u� ) ∑ 𝑐u� . (9.5)
u�∈u�
𝜋u� 𝜋u� u�∈u�
𝜋u� 𝜋u�

Working this out for balanced sampling without replacement with equal inclu-
sion probabilities, 𝜋u� = 𝑛/𝑁, 𝑘 = 1, … , 𝑁, yields

1
𝑉(̂ 𝑧)̄̂ = ∑ 𝑐 𝑒2 . (9.6)
𝑛(𝑛 − 𝑝) u�∈u� u� u�

Deville and Tillé (2005) give several formulas for computing the weights 𝑐u� ,
one of which is 𝑐u� = (1 − 𝜋u� ).
Balanced sampling is now illustrated with aboveground biomass (AGB) data
of Eastern Amazonia, see Figure 1.8. Log-transformed short-wave infrared
radiation (lnSWIR2) is used as a balancing variable. The samplecube function of
the sampling package (Tillé and Matei, 2021) implements the cube algorithm.
Argument X of this function specifies the matrix of ancillary variables on which
the sample must be balanced. The first column of this matrix is filled with
ones, so that the sample size is fixed. To speed up the computations, a 5 km
× 5 km subgrid of grdAmazonia is used.

Recall that a sample is balanced on a covariate 𝑥 if the 𝜋 estimate of the

population total of 𝑥 is equal to the known true population total of 𝑥
(Equation (9.1)). If we know the total number of units in a population, 𝑁,
we can balance the sample on this known total using a constant with value
1 as a balancing variable. Only for samples of size 𝑛 the 𝜋 estimate of the
total number of population units equals 𝑁: ∑u�∈u� 1/𝜋u� = 𝑁 for |𝒮| = 𝑛.

Equal inclusion probabilities are used, i.e., for all population units the inclusion
probability equals 𝑛/𝑁.

grdAmazonia <- grdAmazonia %>%

mutate(lnSWIR2 = log(SWIR2))
library(sampling)
N <- nrow(grdAmazonia)
n <- 100
X <- cbind(rep(1, times = N), grdAmazonia$lnSWIR2)
pi <- rep(n / N, times = N)
sample_ind <- samplecube(X = X, pik = pi, comment = FALSE, method = 1)
eps <- 1e-6
142 9 Balanced and well-spread sampling

units <- which(sample_ind > (1 - eps))

mysample <- grdAmazonia[units, ]

The population mean can be estimated by the sample mean.

mz <- mean(mysample$AGB)

To estimate the variance, a function is defined for estimating the population

regression coefficients.

estimate_b <- function(z, X, c) {

cXX <- matrix(nrow = ncol(X), ncol = ncol(X), data = 0)
cXz <- matrix(nrow = 1, ncol = ncol(X), data = 0)
for (i in seq_len(length(z))) {
x <- X[i, ]
cXX_i <- c[i] * (x %*% t(x))
cXX <- cXX + cXX_i
cXz_i <- c[i] * t(x) * z[i]
cXz <- cXz + cXz_i
}
b <- solve(cXX, t(cXz))
return(b)
}

The next code chunk shows how the estimated variance of the 𝜋 estimator of
the population mean can be computed.

pi <- rep(n / N, n)
c <- (1 - pi)
b <- estimate_b(z = mysample$AGB / pi, X = X[units, ] / pi, c = c)
zpred <- X %*% b
e <- mysample$AGB - zpred[units]
v_tz <- n / (n - ncol(X)) * sum(c * (e / pi)ˆ2)
v_mz <- v_tz / Nˆ2

Figure 9.3 shows the selected balanced sample. Note the spatial clustering
of some units. The estimated population mean (as estimated by the sample
mean) of AGB equals 224.5 109 kg ha-1 . The population mean of AGB equals
225.3 109 kg ha-1 . The standard error of the estimated mean equals 6.1 109 kg
ha-1 .
Figure 9.4 shows the approximated sampling distribution of the 𝜋 estimator of
the mean AGB with balanced sampling and simple random sampling, obtained
9.1 Balanced sampling 143

FIGURE 9.3: Sample balanced on lnSWIR2 from Eastern Amazonia.

by repeating the random sampling with both designs and estimation 1,000
times.

FIGURE 9.4: Approximated sampling distribution of the 𝜋 estimator of

the mean AGB (109 kg ha-1 ) in Eastern Amazonia, with balanced sampling,
balanced on lnSWIR2 and using equal inclusion probabilities, and simple
random sampling without replacement (SI) of 100 units.
144 9 Balanced and well-spread sampling

The variance of the 1,000 estimates of the population mean of the study
variable AGB equals 28.8 (109 kg ha-1 )2 . The gain in precision compared
to simple random sampling equals 2.984 (design effect is 0.335), so with
simple random sampling about three times more sampling units are needed
to estimate the population mean with the same precision. The mean of the
1,000 estimated variances equals 26.4 (109 kg ha-1 )2 , indicating that the
approximated variance estimator somewhat underestimates the true variance
in this case. The population mean of the balancing variable lnSWIR2 equals
6.414. The sample mean of lnSWIR2 varies a bit among the samples. Figure 9.5
shows the approximated sampling distribution of the sample mean of lnSWIR2.
In other words, many samples are not perfectly balanced on lnSWIR2. This is
not exceptional; in most cases perfect balance is impossible.

FIGURE 9.5: Approximated sampling distribution of the sample mean of

balancing variable lnSWIR2 in Eastern Amazonia with balanced sampling of
size 100 and equal inclusion probabilities.

Exercises

1. Select a sample of size 100 balanced on lnSWIR2 and Terra_PP

from Eastern Amazonia, using equal inclusion probabilities for all
units.
9.1 Balanced sampling 145

• First, select a subgrid of 5 km × 5 km using function spsample,

see Chapter 5.
• Estimate the population mean.
• Estimate the standard error of the 𝜋 estimator. First, estimate
the regression coefficients, using function estimate_b defined
above, then compute the residuals, and finally compute the
variance.

2. Spatial clustering of sampling units is not avoided in balanced

sampling. What effect do you expect of this spatial clustering on
the precision of the estimated mean? Can you think of a situation
where this effect does not occur?

9.1.3 Stratified random sampling

Much in the same way as we controlled in the previous subsection the sample
size 𝑛 by balancing the sample on the known total number of population units 𝑁,
we can balance a sample on the known total number of units in subpopulations.
A sample balanced on the sizes of subpopulations is a stratified random sample.
Figure 9.6 shows four subpopulations or strata. These four strata can be used
in balanced sampling by constructing the following design matrix 𝐗 with as
many columns as there are strata and as many rows as there are population
units:

1 0 0 0
⎡1 0 0 0⎤
⎢ ⎥
⎢1 0 0 0⎥
⎢1 0 0 0⎥
𝐗 = ⎢0 1 0 0⎥ . (9.7)
⎢0 1 0 0⎥
⎢ ⎥
⎢0 0 1 0⎥
⎢⋮ ⋮ ⋮ ⋮⎥
⎣0 0 0 1⎦

The first four rows refer to the four leftmost bottom row population units in
Figure 9.6. These units belong to class A, which explains that the first column
for these units contain ones. The other three columns for these rows contain
all zeroes. The fifth and sixth unit belong to stratum B, so that the second
column for these rows contain ones, and so on. The final row is the upperright
sampling unit in stratum D, so the first three columns contain zeroes, and the
fourth column is filled with a one. The sum of the indicators in the columns is
the total number of population units in the strata.
In the next code chunk, the inclusion probabilities are computed by 𝜋ℎu� =
𝑛ℎ /𝑁ℎ , 𝑘 = 1, … , 𝑁ℎ , with 𝑛ℎ = 5 for all four strata. The stratum sample
146 9 Balanced and well-spread sampling

FIGURE 9.6: Sample balanced on a categorical variable with four classes.

sizes are equal, but the number of population units of a stratum differ among
the strata, so the inclusion probabilities also differ among the strata.

N_h <- tapply(mypop$s1, INDEX = mypop$stratum, FUN = length)

n_h <- rep(5, times = 4)
n <- sum(n_h)
print(pi_h <- n_h / N_h)

A B C D
0.06250000 0.12500000 0.03125000 0.04166667

The inclusion probabilities are added to tibble mypop with the 400 population
units, using a look-up table lut and function left_join. The ten leftmost units
on the bottom row of Figure 9.6 are shown below. Variables s1 and s2 are the
spatial coordinates of the centres of the units.

lut <- data.frame(stratum = sort(unique(mypop$stratum)), pi = n_h / N_h)

mypop <- mypop %>%
as.tibble(mypop) %>%
left_join(lut, by = "stratum") %>%
arrange(s2, s1)
mypop
9.1 Balanced sampling 147

# A tibble: 400 x 4
stratum s1 s2 pi
<fct> <dbl> <dbl> <dbl>
1 A 0.5 0.5 0.0625
2 A 1.5 0.5 0.0625
3 A 2.5 0.5 0.0625
4 A 3.5 0.5 0.0625
5 B 4.5 0.5 0.125
6 B 5.5 0.5 0.125
7 C 6.5 0.5 0.0312
8 C 7.5 0.5 0.0312
9 C 8.5 0.5 0.0312
10 C 9.5 0.5 0.0312
# ... with 390 more rows

Next, the design matrix 𝐗 is computed with function model.matrix, expanding

the factor stratum to a set of dummy variables, see code chunk below. By
adding - 1 to the formula, the usual first column with ones in the design
matrix is omitted. The design matrix has four columns with dummy variables
(indicators), indicating to which stratum a unit belongs.
The columns in the design matrix with dummy variables are multiplied by
the vector with inclusion probabilities, using function sweep, resulting in the
following design matrix:

0.0625 0 0 0
⎡0.0625 0 0 0 ⎤
⎢ ⎥
⎢ 0.0625 0 0 0 ⎥
⎢0.0625 0 0 0 ⎥
𝐗=⎢ 0 0.125 0 0 ⎥ . (9.8)
⎢ 0 0.125 0 0 ⎥
⎢ ⎥
⎢ 0 0 0.03125 0 ⎥
⎢ ⋮ ⋮ ⋮ ⋮ ⎥
⎣ 0 0 0 0.04167⎦

The multiplication by the inclusion probabilities is not strictly needed. Using

the design matrix with dummy variables implies that we balance the sample on
the known total number of population units in the strata, 𝑁ℎ . For samples with
stratum sample sizes equal to 𝑛ℎ , the sample sums of the dummy variables
used in balancing, divided by the inclusion probability, are equal to 𝑁ℎ .
Multiplication of the dummy variables by the vector with inclusion probabilities
implies that we balance the sample on the population totals of the inclusion
probabilities, which are equal to the targeted stratum sample sizes. For samples
with stratum sample sizes 𝑛ℎ equal to these targeted sample sizes, the sample
sums of the balancing variables, divided by the inclusion probability (having
value 𝜋ℎu� /𝜋ℎu� = 1 or 0), are equal to the targeted sample sizes.
148 9 Balanced and well-spread sampling

X <- model.matrix(~ stratum - 1, data = mypop)

X <- sweep(X, MARGIN = 1, mypop$pi, `*`)
set.seed(314)
sample_ind <- samplecube(X = X, pik = mypop$pi, comment = FALSE, method = 1)
mysample <- mypop[sample_ind > (1 - eps), ]

In the above example all units in a stratum have the same inclusion probability,
yielding a stratified simple random sample. We may also use variable inclusion
probabilities, for instance proportional to a size measure of the units, yielding
a stratified ppswor random sample (Section 8.2).
The advantage of selecting a stratified random sample by balancing the sample
on a categorical variable becomes clear in case we have multiple classifications
that we would like to use in stratification, and we cannot afford to use all
cross-classifications as strata. This is the topic of the next subsection.

9.1.4 Multiway stratification

Falorsi and Righi (2008) describe how a multiway stratified sample can be
selected as a balanced sample. Multiway stratification is of interest when one
has multiple stratification variables, each stratification variable leading to
several strata, so that the total number of cross-classification strata becomes
so large that the stratum sample sizes are strongly disproportional to their size
or even exceed the total sample size. For instance, suppose we have three maps
with 4, 3, and 5 map units. Further, suppose that all combinations of map
units are non-empty, so that we have 4 × 3 × 5 = 60 combinations. We may
not like to use all combinations (cross-classifications) as strata. The alternative
is then to use the 4 + 3 + 5 = 12 map units as strata.
The sample sizes of the marginal strata can be controlled using a design matrix
with as many columns as there are strata. The units of an individual map used
for stratification are referred to as marginal strata. Each row 𝑘 = 1, … , 𝑁 in
the design matrix 𝐗 has as many non-zero values as we have maps, in entries
corresponding to the cross-classification map unit of population unit 𝑘, and
zeroes in the remaining entries. The non-zero value is the inclusion probability
of that unit. Each column of the design matrix has non-zero values in entries
corresponding to the population units in that marginal stratum and zeroes in
all other entries.
Two-way stratified random sampling is illustrated with a simulated population
of 400 units (Figure 9.7). Figure 9.8 shows two classifications of the population
units. Classification A consists of four classes (map units), classification B
of three classes. Instead of using 4 × 3 = 12 cross-classifications as strata in
random sampling, only 4 + 3 = 7 marginal strata are used in two-way stratified
random sampling.
9.1 Balanced sampling 149

FIGURE 9.7: Simulated population used for illustration of two-way stratified

random sampling.

As a first step, the inclusion probabilities are added to data.frame mypop with
the spatial coordinates and simulated values. To keep it simple, I computed
inclusion probabilities equal to 2 divided by the number of population units in
a cross-classification stratum. Note that this does not imply that a sample is
selected with two units per cross-classification stratum. As we will see later, it
is possible that in some cross-classification strata no units are selected at all,
while in other cross-classification strata more than two units are selected. In
multiway stratified sampling, the marginal stratum sample sizes are controlled.
The inclusion probabilities should result in six selected units for all four units
of map A and eight selected units for all three units of map B.

mypop <- mypop %>%

group_by(A, B) %>%
summarise(N_h = n(), .groups = "drop") %>%
mutate(pih = rep(2, 12) / N_h) %>%
right_join(mypop, by = c("A", "B"))

The next step is to create the design matrix. Two submatrices are computed, one
per stratification. The two submatrices are joined columnwise, using function
cbind. The columns are multiplied by the vector with inclusion probabilities.
150 9 Balanced and well-spread sampling

XA <- model.matrix(~ A - 1, mypop)

XB <- model.matrix(~ B - 1, mypop)
X <- cbind(XA, XB)
X <- sweep(X, MARGIN = 1, mypop$pih, `*`)

Matrix 𝐗 can be reduced by one column if in the first column the inclusion
probabilities of all population units are inserted. This first column contains
no zeroes. Balancing on this variable implies that the total sample size is
controlled. Now there is no need anymore to control the sample sizes of all
marginal strata. It is sufficient to control the sample sizes of three marginal
strata of map A (A2, A3, and A4) and two marginal strata of map B (B2 and
B3). Given the total sample size, the sample sizes of map units A1 and B1
then cannot be chosen freely anymore.

X <- model.matrix(~ A + B, mypop)

colnames(X)

[1] "(Intercept)" "AA2" "AA3" "AA4" "BB2"

[6] "BB3"

X <- sweep(X, MARGIN = 1, mypop$pih, `*`)

This reduced design matrix is not strictly needed for selecting a multiway
stratified sample, but it must be used in estimation. If in estimation, as many
balancing variables are used as we have marginal strata, the matrix with the
sum of squares of the balancing variables (first sum in Equation (9.5)) cannot
be inverted (the matrix is singular), and as a consequence the population
regression coefficients cannot be estimated.
Finally, the two-way stratified random sample is selected with function
samplecube of package sampling.

sample_ind <- samplecube(X = X, pik = mypop$pih, method = 1, comment = FALSE)

units <- which(sample_ind > (1 - eps))
mysample <- mypop[units, ]

Figure 9.8 shows the selected sample.

All marginal stratum sample sizes of map A are six and all marginal stratum
sample sizes of map B are eight, as expected. The sample sizes of the cross-
classification strata vary from zero to four.

addmargins(table(mysample$A, mysample$B))
9.1 Balanced sampling 151

FIGURE 9.8: Two-way stratified sample.

B1 B2 B3 Sum
A1 2 0 4 6
A2 2 3 1 6
A3 3 1 2 6
A4 1 4 1 6
Sum 8 8 8 24

The population mean can be estimated by the 𝜋 estimator.

N <- nrow(mypop)
print(mean <- sum(mysample$z / mysample$pih) / N)

[1] 8.688435

The variance is estimated as before (Equation (9.3)).

c <- (1 - mysample$pih)
b <- estimate_b(
z = mysample$z / mysample$pih, X = X[units, ] / mysample$pih, c = c)
zpred <- X %*% b
e <- mysample$z - zpred[units]
n <- nrow(mysample)
v_tz <- n / (n - ncol(X)) * sum(c * (e / mysample$pih)ˆ2)
print(v_mz <- v_tz / Nˆ2)

[1] 0.1723688
152 9 Balanced and well-spread sampling

9.2 Well-spread sampling

With balanced sampling, the spreading of the sampling units in the space
spanned by the balancing variables can be poor. For instance, in Figure 9.1 the
Easting coordinates of all units of a sample balanced on Easting can be equal
or close to the population mean of 10. So, in this example, balancing does not
guarantee a good geographical spreading. Stated more generally, a balanced
sample can be selected that shows strong clustering in the space spanned
by the balancing variables. This clustering may inflate the standard error of
the estimated population total and mean. The clustering in geographical or
covariate space can be avoided by the local pivotal method (Grafström et al.,
2012) and the spatially correlated Poisson sampling method (Grafström, 2012).

For spreading in geographical space various other designs are available. A

simple design is stratified random sampling from compact geographical
strata, see Section 4.6. Alternative designs are generalised random-tessellation
stratified sampling (Stevens and Olson, 2004), see Subsection 9.2.2, and
balanced acceptance sampling (Robertson et al., 2013).

9.2.1 Local pivotal method

The local pivotal method (LPM) is a modification of the pivotal method

explained in Subsection 8.2.2. The only difference with the pivotal method
is the selection of the pairs of units. In the pivotal method, at each step two
units are selected, for instance, the first two units in the vector with inclusion
probabilities after randomising the order of the units. In the local pivotal
method, the first unit is selected fully randomly, and the nearest neighbour
of this unit is used as its counterpart. Recall that when one unit of a pair is
included in the sample, the inclusion probability of its counterpart is decreased.
This leads to a better spreading of the sampling units in the space spanned by
the spreading variables.
LPM can be used for arbitrary inclusion probabilities. The inclusion probabili-
ties can be equal, but as with the pivotal method these probabilities may also
differ among the population units.
Selecting samples with LPM can be done with functions lpm, lpm1, or lpm2 of
package BalancedSampling (Grafström and Lisic, 2019). Functions lpm1 and
lpm2 only differ in the selection of neighbours that are allowed to compete, for
details see Grafström et al. (2012). For most populations, the two algorithms
perform similarly. The algorithm implemented in function lpm is only recom-
mended when the population size is too large for lpm1 or lpm2. It only uses a
9.2 Well-spread sampling 153

subset of the population in search for nearest neighbours and is thus not as
good. Another function lpm2_kdtree of package SamplingBigData (Lisic and
Grafström, 2018) is developed for big data sets.
Inclusion probabilities are computed with function inclusionprobabilities of
package sampling. A matrix 𝐗 must be defined with the values of the spread-
ing variables of the population units. Figure 9.9 shows a ppswor sample of 40
units selected from the sampling frame of Kandahar, using the spatial coordi-
nates of the population units as spreading variables. Inclusion probabilities
are proportional to the agricultural area within the population units. The
geographical spreading is improved compared with the sample shown in Figure
8.1.

library(BalancedSampling)
library(sampling)
n <- 40
pi <- inclusionprobabilities(grdKandahar$agri, n)
X <- cbind(grdKandahar$s1, grdKandahar$s2)
set.seed(314)
units <- lpm1(pi, X)
myLPMsample <- grdKandahar[units, ]

FIGURE 9.9: Spatial ppswor sample of size 40 from Kandahar, selected by

the local pivotal method, using agricultural area as a size variable.

The total poppy area can be estimated with the 𝜋 estimator (Equation (2.2)).
154 9 Balanced and well-spread sampling

myLPMsample$pi <- pi[units]

tz_HT <- sum(myLPMsample$poppy / myLPMsample$pi)

The estimated total poppy area equals 62,232 ha. The sampling variance of
the estimator of the population total with the local pivotal method can be
estimated by (Grafström and Schelin, 2014)

2
1 𝑧 𝑧u�
𝑉(̂ 𝑡(𝑧))
̂ = ∑ ( u� − u� ) , (9.9)
2 u�∈u� 𝜋u� 𝜋u�u�

with 𝑘u� the nearest neighbour of unit 𝑘 in the sample. This variance estimator
is for the case where we have only one nearest neighbour.
Function vsb of package BalancedSampling is an implementation of a more
general variance estimator that accounts for more than one nearest neighbour
(equation (6) in Grafström and Schelin (2014)). We expect a somewhat smaller
variance compared to pps sampling, so we may use the variance of the pwr
estimator (Equation (8.2)) as a conservative variance estimator.

Xsample <- X[units, ]

se_tz_HT <- sqrt(vsb(pi[units], myLPMsample$poppy, Xsample))
pk <- myLPMsample$pi / n
se_tz_pwr <- sqrt(var(myLPMsample$poppy / pk) / n)

The standard error obtained with function vsb equals 12,850 ha, the standard
error of the pwr estimator equals 14,094 ha.
As explained above, the LPM design can also be used to select a probability
sample well-spread in the space spanned by one or more quantitative covariates.
Matrix 𝐗 then should contain the values of the scaled (standardised) covariates
instead of the spatial coordinates.

Exercises

3. Geographical spreading of the sampling units can also be achieved

by random sampling from compact geographical strata (geostrata)
(Section 4.6). Can you think of one or more advantages of LPM
sampling over random sampling from geostrata?

9.2.2 Generalised random-tessellation stratified sampling

Generalised random-tessellation stratified (GRTS) sampling is designed for

sampling discrete objects scattered throughout space, think for instance of
9.2 Well-spread sampling 155

the lakes in Finland, segments of hedgerows in England, etc. Each object is

represented by a point in 2D-space. It is a complicated design, and for sampling
points from a continuous universe or raster cells from a finite population, I
recommend more simple designs such as the local pivotal method (Subsection
9.2.1), balanced sampling with geographical spreading (Section 9.3), or sam-
pling from compact geographical strata (Section 4.6). Let me try to explain
the GRTS design with a simple example of a finite population of point objects
in a circular study area (Figure 9.10). For a more detailed description of this
design, refer to Hankin et al. (2019). As a first step, a square bounding box of
the study area is constructed. This bounding box is recursively partitioned into
square grid cells. First, 2 × 2 grid cells are constructed. These grid cells are
numbered in a predefined order. In Figure 9.10 this numbering is from lower
left, lower right, upper left to upper right. Each grid cell is then subdivided
into four subcells; the subcells are numbered using the same order. This is
repeated until at most one population unit occurs in each subcell. For our
population only two iterations were needed, leading to 4 × 4 subcells. Note
that two subcells are empty. Each address of a subcell consists of two digits:
the first digit refers to the grid cell, the second digit to the subcell.

FIGURE 9.10: Numbering of grid cells and subcells for GRTS sampling.

The next step is to place the 16 subcells on a line in a random order. The
randomisation is done hierarchically. First, the four grid cells at the highest
level are randomised. In our example, the randomised order is 1, 2, 3, 0 (Figure
9.11). Next, within each grid cell, the order of the subcells is randomised. This
is done independently for the grid cells. In our example, for grid cell 1 the
randomised order of the subcells is 2, 1, 3, 0 (Figure 9.11). Note that the empty
subcells (0,0) and (3,3) are removed from the line.
156 9 Balanced and well-spread sampling

set.seed(314)
ord <- sample(4, 4)
myfinpop_rand <- NULL
for (i in ord) {
units <- which(myfinpop$partit1 == i)
units_rand <- sample(units, size = length(units))
myfinpop_rand <- rbind(myfinpop_rand, myfinpop[units_rand, ])
}

After the subcells have been placed on a line, a one-dimensional systematic

random sample is selected (Figure 9.11), see also Subsection 8.2.1. This can be
done either with equal or unequal inclusion probabilities. With equal inclusion
probabilities, the length of the intervals representing the population units
is constant. With unequal inclusion probabilities, the interval lengths are
proportional to a size variable. For a sample size of 𝑛, the total line is divided
into 𝑛 segments of equal length. A random point is selected in the first segment,
and the other points of the systematic sample are determined. Finally, the
population units corresponding to the selected systematic sample are identified.
With equal probabilities, the five selected units are the units in subcells 11, 23,
22, 32, and 03 (Figure 9.11).

size <- rep(1, N)

n <- 5
interval <- sum(size) / n
start <- round(runif(1) * interval, 2)
mysys <- c(start, 1:(n - 1) * interval + start)

FIGURE 9.11: Systematic random sample along a line with equal inclusion
probabilities.

Figure 9.12 shows a systematic random sample along a line with unequal
inclusion probabilities. The inclusion probabilities are proportional to a size
variable, with values 1, 2, 3, or 4. The selected population units are the units
in subcells 10, 20, 31, 01, and 02.
GRTS samples can be selected with function grts of package spsurvey
(Dumelle et al., 2021). The next code chunk shows the selection of a GRTS
sample of 40 units from Kandahar. Tibble grdKandahar is first converted to an
sf object with function st_as_sf of package sf (Pebesma, 2018). The data set
is using a UTM projection (zone 41N) with WGS84 datum. This projection is
9.2 Well-spread sampling 157

FIGURE 9.12: Systematic random sample along a line with inclusion proba-
bilities proportional to size.

passed to function st_as_df with argument crs (crs = 32641). Argument sframe
of function grts specifies the sampling frame. The sample size is passed to
function grts with argument n_base. Argument seltype is set to proportional to
select units with probabilities proportional to an ancillary variable which is
passed to function grts with argument aux_var.

library(spsurvey)
library(sf)
sframe_sf <- st_as_sf(grdKandahar, coords = c("s1", "s2"), crs = 32641)
set.seed(314)
res <- grts(
sframe = sframe_sf, n_base = 40, seltype = "proportional", aux_var = "agri")
myGRTSsample <- res$sites_base

Function cont_analysis computes the ratio estimator of the population mean

and its standard error. The estimated mean is multiplied by the total number
of population units to obtain a ratio estimate of the population total.

myGRTSsample$siteID <- paste0("siteID", seq_len(nrow(myGRTSsample)))

res <- cont_analysis(myGRTSsample,
vars = "poppy", siteID = "siteID", weight = "wgt", statistics = "Mean")
tz_ratio <- res$Mean$Estimate * nrow(grdKandahar)
se_tz_ratio <- res$Mean$StdError * nrow(grdKandahar)

The estimated total poppy area is 109,809 ha, and the estimated standard
error is 24,327 ha.
The alternative is to estimate the total poppy area by the 𝜋 estimator. Function
vsb of package BalancedSampling can be used to estimate the standard error
of the 𝜋 estimator.

tz_HT <- sum(myGRTSsample$poppy / myGRTSsample$ip)

Xsample <- st_coordinates(myGRTSsample)
se_tz_HT <- sqrt(vsb(myGRTSsample$ip, myGRTSsample$poppy, Xsample))

The estimated total is 71,634 ha, and the estimated standard error is 12,593
ha.
158 9 Balanced and well-spread sampling

9.3 Balanced sampling with spreading

As mentioned in the introduction to this chapter, a sample balanced on a
covariate still may have a poor spreading along the axis spanned by the covariate.
Grafström and Tillé (2013) presented a method for selecting balanced samples
that are also well-spread in the space spanned by the covariates, which they refer
to as doubly balanced sampling. If we take one or more covariates as balancing
variables and, besides, Easting and Northing as spreading variables, this leads
to balanced samples with good geographical spreading. When the residuals
of the regression model show spatial structure (are spatially correlated), the
estimated population mean of the study variable becomes more precise thanks
to the improved geographical spreading. Balanced samples with spreading
can be selected with function lcube of package BalancedSampling. This is
illustrated with Eastern Amazonia, using as before lnSWIR2 for balancing the
sample.

library(BalancedSampling)
N <- nrow(grdAmazonia)
n <- 100
Xbal <- cbind(rep(1, times = N), grdAmazonia$lnSWIR2)
Xspread <- cbind(grdAmazonia$x1, grdAmazonia$x2)
pi <- rep(n / N, times = N)
set.seed(314)
units <- lcube(Xbal = Xbal, Xspread = Xspread, prob = pi)
mysample <- grdAmazonia[units, ]

The selected sample is shown in Figure 9.13. Comparing this sample with the
balanced sample of Figure 9.3 shows that the geographical spreading of the
sample is improved, although there still are some close points. I used equal
inclusion probabilities, so the 𝜋 estimate of the mean is equal to the sample
mean, which is equal to 225.6 109 kg ha-1 .
The variance of the 𝜋 estimator of the mean can be estimated by (equation (7)
of Grafström and Tillé (2013))
2
𝑛 𝑝 𝑒
𝑉(̂ 𝑧)̄̂ = ∑(1 − 𝜋u� ) ( u� − 𝑒u�̄ ) , (9.10)
𝑛 − 𝑝 𝑝 + 1 u�∈u� 𝜋u�

with 𝑝 the number of balancing variables, 𝑒u� the regression model residual of
unit 𝑘 (Equation (9.4)), and 𝑒u�̄ the local mean of the residuals of this unit,
computed by
u�
∑u�+1 (1 − 𝜋u� ) u�u�
(9.11)
u�=1 u�
𝑒u�̄ = u�+1
.
∑u�=1 (1 − 𝜋u� )
9.3 Balanced sampling with spreading 159

FIGURE 9.13: Sample balanced on lnSWIR2 with geographical spreading

from Eastern Amazonia.

The variance estimator can be computed with functions localmean_weight and

localmean_var of package spsurvey (Dumelle et al., 2021).

library(spsurvey)
pi <- rep(n / N, n)
c <- (1 - pi)
b <- estimate_b(z = mysample$AGB / pi, X = Xbal[units, ] / pi, c = c)
zpred <- Xbal %*% b
e <- mysample$AGB - zpred[units]
weights <- localmean_weight(x = mysample$x1, y = mysample$x2, prb = pi, nbh = 3)
v_mz <- localmean_var(z = e / pi, weight_1st = weights) / Nˆ2

The estimated standard error is 2.8 109 kg ha-1 , which is considerably smaller
than the estimated standard error of the balanced sample without geographical
spreading.
10
Model-assisted estimation

In many cases ancillary information is available that could be useful to increase

the accuracy of the estimated mean or total of the study variable. The ancillary
variable(s) can be qualitative (i.e., classifications) or quantitative. As we have
seen before, both types of ancillary variable can be used at the design stage, i.e.,
in selecting the sampling units, to improve the performance of the sampling
strategy, for instance by stratification (Chapter 4), selecting sampling units
with probabilities proportional to size (Chapter 8), or through balancing and/or
spreading the sample on the covariates (Chapter 9). This chapter explains
how these covariates can be used at the stage of estimation, once the data are
collected.
In the design-based approach for sampling, various estimators are developed
that exploit one or more covariates. These estimators are derived from different
superpopulation models of the study variable. A superpopulation model is a
statistical model that can be used to generate an infinite number of populations,
a superpopulation, through simulation. An example is the simulation of spatial
populations using a geostatistical model, see Chapter 13. A superpopulation
is a construct, it does not exist in reality. We assume that the population of
interest is one of the populations that can be generated with the chosen model.
The combination of probability sampling and estimators that are built on a
superpopulation model is referred to as the model-assisted approach. Also
in the model-based approach a superpopulation model is used; however, its
role is fundamentally different from that in the model-assisted approach, see
Chapter 26. To stress the different use of the superpopulation model in the
model-assisted approach, this model is referred to as the “working model”, i.e.,
the superpopulation model that is used to derive a model-assisted estimator.
Breidt and Opsomer (2017) present an overview of model-assisted estimators
derived from a general working model:

𝑍u� = 𝜇(𝐱u� ) + 𝜖u� , (10.1)

with 𝜇(𝐱u� ) the model-mean for population unit 𝑘 which is a function of the
covariate values of that unit collected in vector 𝐱u� = (1, 𝑥1,u� , … , 𝑥u�,u� )T and
𝜖u� a random variable with zero mean. Note that I use uppercase 𝑍 to distinguish
the random variable 𝑍u� of unit 𝑘 from one realisation of this random variable
for unit 𝑘 in the population of interest, 𝑧u� . The model-mean 𝜇(𝐱u� ) can be

161
162 10 Model-assisted estimation

a linear or a non-linear combination of the covariates. If the study variable

and the covariate values were observed for all population units, all these data
could be used to compute a so-called hypothetical population fit of the model
parameters. These model parameters can then be used to compute estimates
of the model-means 𝜇(𝐱u� ), denoted by 𝑚(𝐱u� ), for all population units. For
instance, with a (multiple) regression model 𝑚(𝐱u� ) = 𝐱T u� 𝐛, with 𝐛 the vector
with regression coefficients estimated from observations of the study variable 𝑧
and the covariates on all population units. In practice, we have a sample only,
which is used to estimate 𝑚(𝐱u� ) by 𝑚(𝐱̂ u� ). For the multiple regression model
̂ u� ) = 𝐱T
𝑚(𝐱 u�
̂ with 𝐛̂ the vector with regression coefficients estimated from
𝐛,
the sample data. This leads to the generalised difference estimator (Wu and
Sitter, 2001):

1 u� 1 𝑧 − 𝑚(𝐱
̂ u� )
̄̂ =
𝑧dif ∑ 𝑚(𝐱̂ u� ) + ∑ u� , (10.2)
𝑁 u�=1 𝑁 u�∈u� 𝜋u�

with 𝜋u� the inclusion probability of unit 𝑘. The first term is the population
mean of model predictions of the study variable, and the second term is the 𝜋
estimator of the population mean of the residuals.
A wide variety of model-assisted estimators have been developed and tested
over the past decades. They differ in the working model used to obtain the
estimates 𝑚(𝐱
̂ u� ) in Equation (10.2). The best known class of model-assisted
estimators is the generalised regression estimator that uses a linear model in
prediction (Särndal et al., 1992). Alternative model-assisted estimators are
the estimators using machine learning techniques for prediction. In the era
of big data with a vastly increasing number of exhaustive data sets and a
rapid development of machine learning techniques, these estimators have great
potentials for spatial sample survey.

10.1 Generalised regression estimator

The working model of the generalised regression estimator is the heteroscedastic
multiple linear regression model:

𝑍u� = 𝐱T
u� 𝛽 + 𝜖u� , (10.3)

with 𝜖u� uncorrelated residuals, with zero mean and variance 𝜎2 (𝜖u� ). Note
that the variance of the residuals 𝜎2 (𝜖u� ) need not be constant but may differ
among the population units. If {𝑧u� , 𝑥1,u� , … , 𝑥u�,u� } were observed for all units
10.1 Generalised regression estimator 163

𝑘 = 1, … , 𝑁 in the population, the regression coefficients 𝛽 would be estimated

by

u� −1 u�
𝐱u� 𝐱T 𝐱u� 𝑧u�
𝐛 = (∑ u�
2 (𝜖 )
) ∑ 2 (𝜖 )
, (10.4)
u�=1
𝜎 u� u�=1
𝜎 u�

with 𝐱u� the vector (1, 𝑥1,u� , … , 𝑥u�,u� )T and 𝜎2 (𝜖u� ) the variance of the residual
of unit 𝑘. Similar to the distinction between model-mean and population mean
(see Chapter 26), here the model regression coefficients 𝛽 are distinguished
from the population regression coefficients 𝐛. The means 𝑚(𝐱u� ) would then
be computed by

𝑚(𝐱u� ) = 𝐱T
u� 𝐛 . (10.5)

If we have a probability sample from the population of interest, 𝐛 is estimated

by replacing the population totals in Equation (10.4) by their 𝜋 estimators:

−1
𝐱u� 𝐱T 𝐱u� 𝑧u�
𝐛̂ = (∑ 2 (𝜖 )𝜋
u�
) ∑ 2
. (10.6)
u�∈u�
𝜎 u� u� u�∈u�
𝜎 (𝜖u� )𝜋u�

With unequal inclusion probabilities, the design-based estimators of the

population regression coefficients differ from the usual ordinary least squares
(OLS) estimators of the regression coefficients defined as model parameters.
The values 𝑏̂u� are estimates of the population parameters 𝑏u� .

The mean values 𝑚(𝐱u� ) are now estimated by

̂ u� ) = 𝐱T
𝑚(𝐱 ̂
u� 𝐛 . (10.7)

Plugging Equation (10.7) into the generalised difference estimator, Equation

(10.2), leads to the generalised regression estimator for the population mean:

1 u� T ̂ 𝑧 − 𝐱T ̂
1 u� 𝐛
̄̂
𝑧regr = ∑ 𝐱u� 𝐛 + ∑ u� . (10.8)
𝑁 u�=1 𝑁 u�∈u� 𝜋u�

This estimator can also be written as

u�
̄̂
𝑧regr = 𝑧u�̄̂ + ∑ 𝑏̂u� (𝑥u�̄ − 𝑥u�,u�
̄̂ ) , (10.9)
u�=1
164 10 Model-assisted estimation

with 𝑧u�̄̂ and 𝑥u�,u�

̄̂ the 𝜋 estimator of the study variable and the 𝑗th covariate,
respectively, 𝑥u�̄ the population mean of the 𝑗th covariate, and 𝑏̂u� the estimated
slope coefficient associated with the 𝑗th covariate. So, the generalised regression
estimate is equal to the 𝜋 estimate when the estimated means of the covariates
are equal to the population means. This is the rationale of balanced sampling
(Chapter 9).
The alternative formulation of the regression estimator (Equation (10.9))
shows that we do not need to know the covariate values for all population
units. Knowledge of the population means of the covariates is sufficient. This
is because a linear relation is assumed between the study variable and the
covariates. On the contrary, for non-linear working models such as a random
forest model, exhaustive knowledge of the covariates is needed so that the
estimated mean 𝑚(𝐱̂ u� ) in Equation (10.2) can be computed for every unit in
the population.
Särndal et al. (1992) worked out the generalised regression estimator for various
superpopulation models, such as the simple and multiple linear regression
model, the ratio model, and the analysis of variance (ANOVA) model.

10.1.1 Simple and multiple regression estimators

The working model of the simple and the multiple regression estimator is
the homoscedastic linear regression model. The only difference with the het-
eroscedastic model is that the variance of the residuals is assumed constant:
𝜎2 (𝜖u� ) = 𝜎2 (𝜖), 𝑘 = 1, … , 𝑁.
In the simple linear regression model, the mean is a linear function of a single
covariate, 𝜇(𝑥u� ) = 𝛼 + 𝛽 𝑥u� . The simple linear regression model leads to the
simple regression estimator. With simple random sampling, this estimator for
the population mean is

̄̂
𝑧regr = 𝑧u�̄ + 𝑏̂ (𝑥̄ − 𝑥u�̄ ) , (10.10)

where 𝑧u�̄ and 𝑥u�̄ are the sample means of the study variable and the covariate,
respectively, 𝑥̄ is the population mean of the covariate, and 𝑏̂ is the estimated
slope coefficient:

∑u�∈u� (𝑥u� − 𝑥u�̄ )(𝑧u� − 𝑧u�̄ )

𝑏̂ = . (10.11)
∑u�∈u� (𝑥u� − 𝑥u�̄ )2

The rationale of the regression estimator is that when the estimated mean of
the covariate is, for instance, smaller than the population mean of the covariate,
then with a positive correlation between study variable and covariate, also
the estimated mean of the study variable is expected to be smaller than the
10.1 Generalised regression estimator 165

population mean of the study variable. The difference between the population
mean and the estimated mean of the covariate can be used to improve the 𝜋
estimate of the mean of 𝑧 (which is for simple random sampling equal to the
sample mean 𝑧u�̄ ), by adding a term proportional to the difference between the
estimated mean and the population mean of the covariate. As a scaling factor,
the estimated slope of the fitted regression line is used.
The sampling variance of this regression estimator can be estimated by com-
puting first the regression residuals 𝑒u� = 𝑧u� − 𝑧u�̂ , 𝑘 = 1, … , 𝑛 at the sampling
units, with 𝑧u�̂ = 𝑎̂ + 𝑏𝑥
̂ u� the predicted value for unit 𝑘. Note that I use symbol
𝜖 (Equation (10.3)) for the residuals from the model with the model regression
coefficients 𝛽 , whereas for the residuals from the model with the estimated
population regression coefficients 𝐛̂ I use symbol 𝑒. To compute the residuals 𝑒
also an estimate of the intercept 𝑎 is needed. With simple random sampling,
this intercept can be estimated by

𝑎̂ = 𝑧u�̄ − 𝑏̂ 𝑥u�̄ . (10.12)

The sampling variance of the regression estimator of the population mean is

approximately equal to the sampling variance of the 𝜋 estimator of the mean
of the model residuals:

𝑛 𝑆̂2 (𝑒)
𝑉(̂ 𝑧regr
̄̂ ) = (1 − ) , (10.13)
𝑁 𝑛

with 𝑆
̂2 (𝑒) the estimated population variance of the regression residuals:

1
̂
𝑆2 (𝑒) = ∑ 𝑒2 . (10.14)
𝑛 − 1 u�∈u� u�

The variance estimator is an approximation because the regression coefficients

are also estimated from the sample, which makes the regression estimator non-
linear. The approximation of the variance is based on a Taylor linearisation
of the regression estimator (Särndal et al. (1992), p. 235).

For simple random sampling with replacement from finite populations and
simple random sampling of infinite populations, the finite population correction
factor 1 − 𝑛/𝑁 must be dropped, see Chapter 3.
In the multiple linear regression model, the mean is a linear function of multiple
covariates. This model leads to the multiple regression estimator. With simple
random sampling, the population regression coefficients of this estimator can
166 10 Model-assisted estimation

be estimated by
−1

𝐛̂ = (∑ 𝐱u� 𝐱T
u� ) ∑ 𝐱u� 𝑧u� . (10.15)
u�∈u� u�∈u�

Comparison with the general estimator of the population regression coefficients

(Equation (10.6)) shows that the variance of the residuals, 𝜎2 (𝜖u� ), is missing as
they are assumed constant. Besides, the inclusion probabilities 𝜋u� are missing
because they are equal for all population units with simple random sampling.
The simple regression estimator is illustrated with Eastern Amazonia. The
population mean of the aboveground biomass (AGB) is estimated by the
simple regression estimator, using natural logs of MODIS short-wave infrared
radiation (SWIR2) as a covariate.
A simple random sample without replacement of 100 units is selected using
slice_sample of package dplyr, and the two population regression coefficients
are estimated with Equation (10.15).

grdAmazonia <- grdAmazonia %>%

mutate(lnSWIR2 = log(SWIR2))
set.seed(321)
n <- 100
mysample <- grdAmazonia %>%
dplyr::select(AGB,lnSWIR2) %>%
slice_sample(n = n)
X <- matrix(nrow = n, ncol = 2, data = 1)
X[, 2] <- mysample$lnSWIR2
XXinv <- solve(t(X) %*% X)
Xz <- t(X) %*% mysample$AGB
print(ab <- t(XXinv %*% Xz))

[,1] [,2]
[1,] 1751.636 -237.1379

The same estimates are obtained by ordinary least squares (OLS) fitting of
the model with function lm.

lm_sample <- lm(AGB ~ lnSWIR2, data = mysample)

print(ab_mb <- coef(lm_sample))

(Intercept) lnSWIR2
1751.6363 -237.1379

As already stressed above, the design-based estimates of the population re-

gression coefficients are only equal to the model-based OLS estimates of the
regression coefficients for equal probability sampling designs. Also be aware
10.1 Generalised regression estimator 167

that the variance of the design-based estimates of the population regression

coefficients is not equal to the model-based variance of the model regression
coefficients. See section (11.2.2.1) in Lohr (1999) for how to estimate the
variance of the design-based estimates of the population regression coefficients.
Figure 10.1 shows the scatter plot for the simple random sample and the fitted
simple linear regression model.

FIGURE 10.1: Scatter plot of AGB (109 kg ha-1 ) against lnSWIR2 of a

simple random sample of size 100 from Eastern Amazonia and the fitted simple
linear regression model for AGB.

The simple random sample is used to estimate the population mean of the
study variable AGB by the simple regression estimator and to approximate the
sampling variance of the regression estimator. The residuals of the fitted model
can be extracted with function residuals because in this case the OLS estimates
of the regression coefficients are equal to the design-based estimates. With
unequal inclusion probabilities, the residuals must be computed by predicting
the study variable for the selected units, using the design-based estimates
of the regression coefficients, and subtracting the observations of the study
variable.

mx_pop <- mean(grdAmazonia$lnSWIR2)

mx_sam <- mean(mysample$lnSWIR2)
mz_sam <- mean(mysample$AGB)
mz_regr <- mz_sam + ab[2] * (mx_pop - mx_sam)
e <- residuals(lm_sample)
S2e <- var(e)
168 10 Model-assisted estimation

N <- nrow(grdAmazonia)
se_mz_regr <- sqrt((1 - n / N) * S2e / n)

The difference 𝛿(𝑥) between the population mean of the covariate lnSWIR2
(6.415) and its estimated mean (6.347) equals 0.068. We may expect the
difference between the unknown population mean of the study variable AGB
and its sample mean (246.510) to be equal to 𝛿(𝑥), multiplied by the estimated
slope of the line, which equals -237.1. The result, -16.1039, is added to the
simple random sample estimate, so that the ultimate regression estimate is
adjusted downward to 230.4 109 kg ha-1 .
The estimated approximate standard error of the regression estimator equals
4.458 109 kg ha-1 . The approximated variance is a simplification of a more
complicated approximation derived from writing the regression estimator of the
population total as a weighted sum of the 𝜋-expanded observations (Särndal
et al. (1992), equation (6.5.9)):

1 𝑧
̄̂
𝑧regr = ∑ 𝑔 u� , (10.16)
𝑁 u�∈u� u� 𝜋u�

with 𝑔u� the weight for unit 𝑘. For simple random sampling, the weights are
equal to (Särndal et al. (1992), equation (6.5.12))

(𝑥̄ − 𝑥u�̄ )(𝑥u� − 𝑥u�̄ )

𝑔u� = 1 + . (10.17)
̂
𝑆 2 (𝑥)

These weights are now computed.

S2x <- sum((mysample$lnSWIR2 - mean(mysample$lnSWIR2))ˆ2) / n

g <- 1 + ((mx_pop - mx_sam) * (mysample$lnSWIR2 - mx_sam)) / S2x

The sample mean of the weights equals 1,

mean(g)

[1] 1

and the sample mean of the product of the weights and the covariate 𝑥 equals
the population mean of the covariate.

all.equal(mean(g * mysample$lnSWIR2), mean(grdAmazonia$lnSWIR2))

[1] TRUE
10.1 Generalised regression estimator 169

In other words, the sample is calibrated on the known population means.

The variance of the regression estimator of the population mean can be
approximated by (Särndal et al. (1992), section 6.6)

𝑛 ∑ 𝑔2u� 𝑒2u�
𝑉(̂ 𝑧regr
̄̂ ) = (1 − ) u�∈u� . (10.18)
𝑁 𝑛(𝑛 − 1)

Comparing this with Equation (10.13) shows that in the first approximation
we assumed that all weights are equal to 1.
The alternative approximate standard error is computed in the next code
chunk.

S2ge <- sum(gˆ2 * eˆ2) / (n - 1)

print(se_mz_regr <- sqrt((1 - n / N) * S2ge / n))

[1] 4.546553

The regression estimator and its standard error can be computed with package
survey (Lumley, 2021). After specifying the sampling design with function
svydesign, function calibrate is used to calibrate the sample on the known
population totals 𝑁 and 𝑡(𝑥) = ∑u�=1 𝑥u� , with 𝑥u� the value of covariate
u�

lnSWIR2 for unit 𝑘.

library(survey)
mysample$fpc <- N
design_si <- svydesign(id = ~ 1, data = mysample, fpc = ~ fpc)
populationtotals <- c(N, sum(grdAmazonia$lnSWIR2))
mysample_cal <- calibrate(design_si, formula = ~ lnSWIR2,
population = populationtotals, calfun = "linear")

The calibrated weights can be extracted with function weights.

The calibrated weights are divided by the inclusion probabilities 𝜋 = 𝑛/𝑁,

so that the sample sum of the weights equals 𝑁 and not the sample size 𝑛
(as in the code chunk above).

g <- weights(mysample_cal)
all.equal(sum(g), N)

[1] TRUE

The sample sum of the product of the weights and the covariate equals the
population total of the covariate.
170 10 Model-assisted estimation

all.equal(sum(g * mysample$lnSWIR2), sum(grdAmazonia$lnSWIR2))

[1] TRUE

Finally, the population mean can be estimated with function svymean. This is
simply the sample sum of the product of the weights and the study variable
AGB, divided by 𝑁.

svymean(~ AGB, mysample_cal)

mean SE
AGB 230.41 4.5466

The standard error is computed with Equation (10.18). Figure 10.2 shows
the sampling distribution of the simple regression estimator along with the
distribution of the 𝜋 estimator, obtained by repeating simple random sampling
of 100 units and estimation 10,000 times.

FIGURE 10.2: Approximated sampling distribution of the simple regression

estimator (Regression) and the 𝜋 estimator (HT) of the mean AGB (109 kg
ha-1 ) in Eastern Amazonia, for simple random sampling without replacement
of size 100.

The average of the 10,000 regression estimates equals 224.9 109 kg ha-1 . The
population mean of the study variable AGB equals 225.0 109 kg ha-1 , so the
estimated bias of the regression estimator equals -0.1 109 kg ha-1 , which is
negligibly small related to the estimated population mean. The variance of
10.1 Generalised regression estimator 171

the 10,000 regression estimates equals 26.70 (109 kg ha-1 )2 , and the average
of the 10,000 estimated approximate variances using Equation (10.18) equals
26.86 (109 kg ha-1 )2 . The gain in precision due to the regression estimator,
quantified by the ratio of the variance of the 𝜋 estimator to the variance of
the regression estimator equals 3.192.
For simple random sampling, the ratio of the variances of the simple regression
estimator and the 𝜋 estimator is independent of the sample size and equals
1 − 𝑟2 , with 𝑟 the correlation coefficient of the study variable and the covariate
(Särndal et al. (1992), p. 274).
Using multiple covariates in the regression estimator is straightforward with
function calibrate of package survey. As a first step, the best model is selected
with function regsubsets of package leaps (Lumley, 2020).

library(leaps)
n <- 100
set.seed(321)
mysample <- grdAmazonia %>%
dplyr::select(AGB, lnSWIR2, Terra_PP, Prec_dm, Elevation, Clay) %>%
slice_sample(n = n)
models <- regsubsets(AGB ~ ., data = mysample, nvmax = 4)
res_sum <- summary(models)
res_sum$outmat

lnSWIR2 Terra_PP Prec_dm Elevation Clay

1 ( 1 ) "*" " " " " " " " "
2 ( 1 ) "*" "*" " " " " " "
3 ( 1 ) "*" "*" " " "*" " "
4 ( 1 ) "*" "*" "*" "*" " "

The best model with one predictor is the model with lnSWIR2, the best
model with two predictors is the one with lnSWIR2 and Terra_PP, etc. Of
these models, the third model, i.e., the model with lnSWIR2, Terra_PP, and
Elevation, is the best when using adjusted 𝑅2 as a selection criterion.

which.max(res_sum$adjr2)

[1] 3

The standard error of the estimated mean AGB is somewhat reduced by adding
the covariates Terra_PP and Elevation to the regression estimator.

mysample$fpc <- nrow(grdAmazonia)

design_si <- svydesign(id = ~ 1, data = mysample, fpc = ~ fpc)
totals <- c(nrow(grdAmazonia), sum(grdAmazonia$lnSWIR2),
172 10 Model-assisted estimation

sum(grdAmazonia$Terra_PP), sum(grdAmazonia$Elevation))
mysample_cal <- calibrate(design_si, formula = ~ lnSWIR2 + Terra_PP + Elevation,
population = totals, calfun = "linear")
svymean(~ AGB, mysample_cal)

mean SE
AGB 230.54 4.2224

Another interesting package for model-assisted estimation is package mase

(McConville et al., 2021). The regression estimate can be computed with
function greg.

library(mase)
covars <- c("lnSWIR2", "Terra_PP", "Elevation")
res <- greg(y = mysample$AGB, xsample = mysample[covars],
xpop = grdAmazonia[covars], pi = rep(n / N, n),
var_est = TRUE, var_method = "LinHTSRS", model = "linear")
res$pop_mean

[1] 230.5407

The multiple regression estimate is equal to the estimate obtained with function
calibrate of package survey. The estimated standard error equals

sqrt(res$pop_mean_var)

[,1]
[1,] 4.207809

which is slightly smaller than the standard error computed with package
survey. The standard error obtained with function greg is computed by
ignoring the g-weights (McConville et al., 2020). In an exercise, the two
approximate standard errors are compared in a sampling experiment.

10.1.2 Penalised least squares estimation

In the previous subsection, I first selected a best subset of covariates before

using these covariates in estimating the population regression coefficients. The
alternative is to skip the selection of the best model and to estimate the
population regression coefficients of all covariates by penalised least squares
(PLS) estimation. In PLS estimation a penalty equal to the sum of the absolute
or squared values of the population regression coefficients is added to the
minimisation criterion, see McConville et al. (2020) for details. PLS estimation
is implemented in function gregElasticNet of package mase.
10.1 Generalised regression estimator 173

covars <- c("lnSWIR2", "Terra_PP", "Prec_dm", "Elevation", "Clay")

res <- gregElasticNet(
y = mysample$AGB, xsample = mysample[covars],
xpop = grdAmazonia[covars], pi = rep(n / N, n),
var_est = TRUE, var_method = "LinHTSRS", model = "linear",
lambda = "lambda.min", cvfolds = 100)
signif(res$coefficients, 4)

(Intercept) lnSWIR2 Terra_PP Prec_dm Elevation Clay

-71.07000 -2.83800 0.02204 0.52710 -0.03507 0.13630

All five covariates are used in prediction, but the coefficients associated with
these predictors are small except for lnSWIR2.
As shown below, the estimated standard error is considerably larger than the
standard error obtained with lnSWIR2, Terra_PP, and Elevation as predictors.
In this case, the elastic net regression estimator does not work as well as the
multiple regression estimator using the best subset of the covariates.

sqrt(res$pop_mean_var)

s1
s1 6.207637

Exercises

1. Write an R script to select a simple random sample without replace-

ment of size 50 from Eastern Amazonia. Compute the regression
estimator of the population mean of AGB and its standard error
“by hand”, i.e., without using package survey, using lnSWIR2 as a
covariate. Use Equation (10.13) to estimate the standard error.
• Use the same sample to compute the regression estimator
with functions calibrate and svymean of package survey. The
regression estimate and its standard error can be extracted
from the output object of svymean with methods coef and SE,
respectively.
• Repeat both estimation procedures 1,000 times (for-loop).
Check that the 2 × 1,000 estimated population means obtained
with both estimators are all equal (use function all.equal), and
compute summary statistics of the two approximated standard
errors. Which approximate standard error estimator has the
largest mean value?
174 10 Model-assisted estimation

2. Write an R script to compute the sampling variance of the simple

regression estimator of the mean AGB, using lnSWIR2 as a covari-
ate, for simple random sampling and sample sizes 10, 25, and 100,
assuming that the population regression coefficients are perfectly
known. Hint: fit a simple linear regression model on all data, and
compute the population variance of the residuals.
• Next, select 10,000 times a simple random sample with replace-
ment of size 10 (for-loop). Use each sample to estimate the
population mean of AGB with the simple regression estimator
(using sample estimates of the population regression coeffi-
cients), and estimate the approximate variance of the estimator
of the mean. Compute the variance of the 10,000 regression
estimates and the average of the 10,000 approximate variance
estimates. Repeat this for sample sizes 25 and 100.
• Compute for each sample size the difference between the exper-
imental variance (variance of the 10,000 regression estimates)
and the variance obtained with the population fit of the re-
gression model as a proportion of the experimental variance.
Explain what you see.
• Compute for each sample size the difference between the average
of the 10,000 approximated variances and the experimental
variance as a proportion of the experimental variance. Explain
what you see.

10.1.3 Regression estimator with stratified simple random

sampling

With stratified simple random sampling there are two regression estimators:
separate and combined. In the first estimator, the regression estimator for
simple random sampling is applied at the level of the strata. This implies that
for each stratum separately a vector with population regression coefficients 𝐛ℎ
is estimated. The regression estimates of the stratum means are then combined
by computing the weighted average, using the relative sizes of the strata as
weights:
u�
̄̂
𝑧sregr ̄̂
= ∑ 𝑤ℎ 𝑧regr,ℎ , (10.19)
ℎ=1

with, for the simple linear estimator

̄̂
𝑧regr,ℎ ̄ + 𝑏̂ℎ (𝑥ℎ
= 𝑧u�ℎ ̄ − 𝑥u�ℎ
̄ ) , (10.20)

with 𝑧u�ℎ
̄ and 𝑥u�ℎ
̄ the stratum sample means of the study variable and the
covariate, respectively, 𝑥ℎ
̄ the mean of the covariate in stratum ℎ, and 𝑏̂ℎ the
estimated slope coefficient for stratum ℎ.
10.1 Generalised regression estimator 175

The variance of this separate regression estimator of the population mean can
be estimated by first estimating the variances of the regression estimators of
the stratum means using Equation (10.13), and then combining these variances
using Equation (4.4).
The separate regression estimator is illustrated with Eastern Amazonia. Biomes
are used as strata. There are four biomes, the levels of which are given short
names using function levels.

grdAmazonia$Biome <- as.factor(grdAmazonia$Biome)

levels(grdAmazonia$Biome)

[1] "Mangroves"
[2] "Tropical & Subtropical Dry Broadleaf Forests"
[3] "Tropical & Subtropical Grasslands, Savannas & Shrublands"
[4] "Tropical & Subtropical Moist Broadleaf Forests"

biomes <- c("Mangrove", "Forest_dry", "Grassland", "Forest_moist")

levels(grdAmazonia$Biome) <- biomes

Moist forest is by far the largest stratum, it covers 92% of the area. Mangrove,
Forest_dry, and Grassland cover 0.4, 2.3, and 5.5% of the area, respectively. A
stratified simple random sample of size 100 is selected using function strata
of package sampling, see Chapter 4. I chose five units as a minimum sample
size. Note that the stratum sample sizes are not proportional to their size.

library(sampling)
N_h <- table(grdAmazonia$Biome)
n_h <- c(5, 5, 5, 85)
set.seed(314)
units <- sampling::strata(grdAmazonia, stratanames = "Biome",
size = n_h[unique(grdAmazonia$Biome)], method = "srswor")
mysample <- getdata(grdAmazonia, units)

As a first step in estimation, for each stratum the mean of the covariate over
all units in a stratum (population mean per stratum) and the sample means
of the study variable and the covariate are computed.

mx_h_pop <- tapply(grdAmazonia$lnSWIR2, INDEX = grdAmazonia$Biome, FUN = mean)

mzh_sam <- tapply(mysample$AGB, INDEX = mysample$Biome, FUN = mean)
mx_h_sam <- tapply(mysample$lnSWIR2, INDEX = mysample$Biome, FUN = mean)

The next step is to estimate the regression coefficients (intercept and slope)
per stratum. This is done in a for-loop. The estimated slope coefficient is used
176 10 Model-assisted estimation

to compute the regression estimator per stratum. The residuals are extracted
to approximate the variance of the regression estimator per stratum.

b_h <- mz_h_regr <- v_mz_h_regr <- numeric(length = 4)

for (i in 1:4) {
subsam <- subset(mysample, mysample$Biome == levels(grdAmazonia$Biome)[i])
lm_sample <- lm(AGB ~ lnSWIR2, data = subsam)
b_h[i] <- coef(lm_sample)[2]
mz_h_regr[i] <- mzh_sam[i] + b_h[i] * (mx_h_pop[i] - mx_h_sam[i])
e <- residuals(lm_sample)
S2e_h <- var(e)
v_mz_h_regr[i] <- (1 - n_h[i] / N_h[i]) * S2e_h / n_h[i]
}

Finally, the separate regression estimate is computed as a weighted average of

the regression estimates per stratum.

w_h <- N_h / sum(N_h)

print(mz_sepreg <- sum(w_h * mz_h_regr))

[1] 223.9426

The standard error of the separate regression estimator is computed by the

square root of the pooled variances of the regression estimator per stratum,
using the squared relative sizes of the strata as weights.

sum(w_hˆ2 * v_mz_h_regr) %>% sqrt(.)

[1] 5.077558

The separate regression estimator can be computed with package survey.

The computation of the population totals merits special attention. For the
simple regression estimator using simple random sampling, these totals are
the total number of population units and the population total of the covariate
lnSWIR2. These are the population totals associated with the columns of the
design matrix that is constructed with function lm to estimate the regression
coefficients. The column with ones results in an estimated intercept, the column
with lnSWIR2 values in an estimated slope.
The model that is fitted now is an analysis of covariance (ANCOVA) model
with factor Biome and covariate lnSWIR2.

ancova <- lm(AGB ~ Biome * lnSWIR2, data = mysample)

R uses the so-called cornerstone representation of the ANCOVA model. The

reference level is stratum Mangrove. The question is what population totals
10.1 Generalised regression estimator 177

must be passed to function calibrate with this ANCOVA model. This can be
determined by printing the design matrix that is used to fit the ANCOVA
model. Only the first two rows are printed.

designmat <- model.matrix(ancova, mysample)

(Intercept) BiomeForest_dry BiomeGrassland BiomeForest_moist lnSWIR2

[1,] 1 0 0 1 6.307024
[2,] 1 0 0 1 6.236278
BiomeForest_dry:lnSWIR2 BiomeGrassland:lnSWIR2 BiomeForest_moist:lnSWIR2
[1,] 0 0 6.307024
[2,] 0 0 6.236278

With this model formulation, the first population total is the total number
of population units. The second, third, and fourth population totals are the
number of population units in stratum levels 2, 3, and 4. The fifth population
total is the population total of covariate lnSWIR2 and the sixth, seventh, and
eighth population totals are the totals of covariate lnSWIR2 in stratum levels
2, 3, and 4.

N_h <- as.numeric(N_h)

lut <- data.frame(Biome = biomes, N_h)
mysample <- merge(x = mysample, y = lut)
design_stsi <- svydesign(
id = ~ 1, strata = ~ factor(Biome), data = mysample, fpc = ~ N_h)
tx_pop <- sum(grdAmazonia$lnSWIR2)
tx_h_pop <- N_h * mx_h_pop
totals <- c(sum(N_h), N_h[c(2, 3, 4)], tx_pop, tx_h_pop[c(2, 3, 4)])
names(totals) <- names(coef(ancova))
mysample_cal <- calibrate(
design_stsi, formula = ~ Biome * lnSWIR2,
population = totals, calfun = "linear")
svymean(~ AGB, mysample_cal)

mean SE
AGB 223.94 5.8686

The line names(totals) <- names(coef(ancova)) is not strictly needed. This is

just to suppress a warning that the names of the numeric with the population
totals does not match the names of the columns of the design matrix. As a
consequence, we do not need to fit the ANCOVA model either.

Alternatively, we may use the following formula in function lm.

178 10 Model-assisted estimation

ancova2 <- lm(AGB ~ 0 + Biome / lnSWIR2, data = mysample)

designmat <- model.matrix(ancova, mysample)

(Intercept) BiomeForest_dry BiomeGrassland BiomeForest_moist lnSWIR2

[1,] 1 0 0 1 6.307024
[2,] 1 0 0 1 6.236278
BiomeForest_dry:lnSWIR2 BiomeGrassland:lnSWIR2 BiomeForest_moist:lnSWIR2
[1,] 0 0 6.307024
[2,] 0 0 6.236278

With this formula, the population totals are the number of population units
in stratum levels 1, 2, 3, and 4, as well as the population totals of covariate
lnSWIR2 of the strata.

totals <- c(N_h, tx_h_pop)

names(totals) <- names(coef(ancova2))
mysample_cal <- calibrate(
design_stsi, formula = ~ 0 + Biome / lnSWIR2, population = totals,
calfun = "linear")
svymean(~ AGB, mysample_cal)

mean SE
AGB 223.94 5.8686

10.1.3.1 Combined regression estimator

The alternative to the separate simple regression estimator is the combined

simple regression estimator:

̄̂
𝑧cregr = 𝑧u�̄̂ + 𝑏̂ (𝑥̄ − 𝑥u�
̄̂ ) , (10.21)

with 𝑏̂ the estimated slope coefficient, estimated by Equation (10.6), discarding

the variance of the residuals 𝜎2 (𝜖u� ) as they are assumed constant, and using
the appropriate inclusion probabilities which differ among the strata, and 𝑧u�̄̂
and 𝑥u�̄̂ the 𝜋 estimators of the population mean of the study variable and
the covariate for stratified simple random sampling, respectively. Working
Equation (10.6) out for stratified simple random sampling yields

̂
𝑤2 𝑆2 (𝑧, 𝑥)
𝑏̂ = ℎ ℎ , (10.22)
2 ̂
𝑤ℎ 𝑆2 (𝑥)
ℎ

with 𝑆̂2 (𝑧, 𝑥) the estimated covariance of the study variable and the covariate
ℎ
in stratum ℎ and 𝑆 ̂ 2 (𝑥) the estimated variance of the covariate.
ℎ
10.1 Generalised regression estimator 179

In the combined simple regression estimator only one regression coefficient 𝑏

is estimated, the slope coefficient for the entire population. This combined
regression estimator is recommended when the stratum sample sizes are so
small, as in our case, that the estimated regression coefficients per stratum,
𝑏̂ℎ , become unreliable.

Estimator (10.22) is for infinite populations and for stratified simple random
sampling with replacement of finite populations. For sampling without re-
placement from finite populations, finite population corrections 1 − 𝑛ℎ /𝑁ℎ
must be added to the numerator and denominator of 𝑏̂ (Cochran (1977),
p. 202).

The approximate variance of the combined regression estimator can be esti-

mated as follows:

1. Compute residuals: 𝑒u� = 𝑧u� − (𝑎̂ + 𝑏𝑥

̂ u� ), with 𝑎̂ and 𝑏̂ the estimated
regression coefficients for the whole population.

2. Estimate for each stratum the variance of the estimator of the mean
of the residuals: 𝑉(̂ 𝑒ℎ̄̂ ) = 𝑆
̂2 (𝑒)/𝑛 , with 𝑆
ℎ ℎ
̂2 (𝑒) the estimated
ℎ
variance of the residuals in stratum ℎ.

3. Combine the estimated variances per stratum: 𝑉(̂ 𝑧cregr

̄̂ ) =
u� 2 ̂ ̂
∑ 𝑤ℎ 𝑉(𝑒ℎ̄ ).
ℎ=1

The next code chunks show the estimation procedure. First, the population
means of the study variable AGB and of the covariate lnSWIR2 are estimated
by the 𝜋 estimator, see Chapter 4.

mz_h_HT <- tapply(mysample$AGB, INDEX = mysample$Biome, FUN = mean)

mx_h_HT <- tapply(mysample$lnSWIR2, INDEX = mysample$Biome, FUN = mean)
mz_HT <- sum(w_h * mz_h_HT)
mx_HT <- sum(w_h * mx_h_HT)

The next step is to estimate the population regression coefficients, using

Equation (10.6) in which the variances 𝜎2 (𝜖u� ) can be dropped, as these are
assumed constant. The inclusion probabilities are in column Prob of mysample.

W <- diag(x = 1 / mysample$Prob, nrow = n, ncol = n)

X <- matrix(nrow = n, ncol = 2, data = 1)
X[, 2] <- mysample$lnSWIR2
XWX <- t(X) %*% W %*% X
180 10 Model-assisted estimation

XWz <- t(X) %*% W %*% mysample$AGB

print(ab <- t(solve(XWX, XWz)))

[,1] [,2]
[1,] 1678.268 -226.6772

Note that the same estimates are obtained by model-based estimation, using
weighted least squares, based on the assumption that the variances 𝜎2 (𝜖u� ) are
proportional to the inclusion probabilities (which is a weird assumption).

lm_wls <- lm(AGB ~ lnSWIR2, weights = 1 / Prob, data = mysample)

coef(lm_wls)

(Intercept) lnSWIR2
1678.2684 -226.6772

In model-based estimation, the weights differ among the units because of

assumed differences in the variance of the residuals, whereas in design-based
estimation we assign different weights to the observations because the units
have different inclusion probabilities (Lohr, 1999).

Finally, the combined regression estimate is computed.

print(mz_combreg <- mz_HT + ab[2] * (mx_pop - mx_HT))

[1] 224.1433

To approximate the variance of the combined regression estimator, first the

residuals are computed. Then these residuals are used to estimate the spatial
variance of the residuals within the strata, 𝑆 ̂2 (𝑒), and the variance of the
ℎ
estimator of the mean of the residuals, 𝑉(𝑒ℎ̄ ). Finally, by taking the square
̂ ̂
root, the estimated standard error is obtained.

mysample$e <- mysample$AGB - (ab[1] + ab[2] * mysample$lnSWIR2)

v_me_h <- numeric(length = 4)
for (i in 1:4) {
subsam <- subset(mysample, mysample$Biome == levels(grdAmazonia$Biome)[i])
S2e_h <- var(subsam$e)
v_me_h[i] <- (1 - n_h[i] / N_h[i]) * S2e_h / n_h[i]
}
print(se_mz_combreg <- sqrt(sum(w_hˆ2 * v_me_h)))

[1] 5.122518
10.2 Ratio estimator 181

Computing the combined regression estimator with package survey proceeds

as follows.

design_stsi <- svydesign(

id = ~ 1, strata = ~ factor(Biome), data = mysample, fpc = ~ N_h)
totals <- c(nrow(grdAmazonia), sum(grdAmazonia$lnSWIR2))
mysample_cal <- calibrate(
design_stsi, formula = ~ lnSWIR2, population = totals,
calfun = "linear")
svymean(~ AGB, mysample_cal)

mean SE
AGB 224.14 5.8707

Function calibrate computes the regression estimate and its standard error
with the calibrated weights 𝑔u� (Särndal et al. (1992), equation (6.5.12)). This
explains the difference between the two standard errors.

10.2 Ratio estimator

In some cases it is reasonable to assume that the fitted line passes through the
origin. An example is the case study on poppy area in Kandahar (Chapter 8).
The covariate is the agricultural area within the 5 km squares that serve as
sampling units. It is reasonable to assume that when the covariate equals zero,
also the poppy area is zero. So, if we have an estimate of the ratio of the total
poppy area in the population to the total agricultural area in the population
and besides know the total agricultural area in the population, the total poppy
area in the population can be estimated by multiplying the estimated ratio
with the known population total agricultural area:

𝑡u�̂ (𝑧)
̂
𝑡ratio (𝑧) = 𝑡(𝑥) = 𝑏̂ 𝑡(𝑥) , (10.23)
𝑡u�̂ (𝑥)

with 𝑡u�̂ (𝑧) and 𝑡u�̂ (𝑥) the 𝜋 estimators of the total of the study variable (poppy
area) and the ancillary variable (agricultural area), respectively, and 𝑡(𝑥) the
total of the ancillary variable, which must be known.
182 10 Model-assisted estimation

The working model of the ratio estimator is a heteroscedastic model, i.e., a

model with non-constant variance, without intercept (see Exercise 3 hereafter):

𝑍(𝑥u� ) = 𝛽 𝑥u� + 𝜖u�

(10.24)
𝜎2 (𝜖u� ) = 𝜎2 𝑥u� ,

with 𝛽 the slope of the line and 𝜎2 a constant (variance of residual for 𝑥u� = 1).
The residual variance is assumed proportional to the covariate 𝑥.

The ratio estimator was applied before to estimate the population mean or
population total from a systematic random sample (Chapter 5), a one-stage
and two-stage cluster random sample (Sections 6.3 and 7.3), and a ppswor
sample (Section 8.2). By taking 𝑥u� = 1, 𝑘 = 1, … , 𝑁, 𝑡u�̂ (𝑥) in Equation
(10.23) is equal to 𝑁,̂ and 𝑡(𝑥) is equal to 𝑁. For (two-stage) cluster random
sampling 𝑀 is used for the total number of population units (𝑁 is the total
number of clusters or primary sampling units in the population) and therefore
𝑡u�̂ (𝑥) = 𝑀̂ and 𝑡(𝑥) = 𝑀. This yields the ratio estimators of the population
total appropriate for these sampling designs.

Equation (10.23) is a general estimator that can be used for any probability
sampling design, not only for simple random sampling. For simple random
sampling, the coefficient 𝑏 is estimated by the ratio of the sample means of 𝑧
and 𝑥.
For simple random sampling, the sampling variance of the ratio estimator of
the population total can be approximated by

̂
𝑆2 (𝑒)
𝑉(̂ 𝑡ratio
̂ (𝑧)) = 𝑁2 , (10.25)
𝑛

with 𝑆
̂2 (𝑒) the estimated variance of the residuals 𝑒 = 𝑧 − 𝑏𝑥
u� u�
̂ u� :

1
̂
𝑆2 (𝑒) = ∑ 𝑒2 . (10.26)
𝑛 − 1 u�∈u� u�

For simple random sampling without replacement from finite populations,

Equation (10.25) must be multiplied by (1 − u�
u�
).
In R the ratio estimator for the total poppy area and the estimator of its
variance for a simple random sample without replacement can be computed as
follows.

n <- 50
N <- nrow(grdKandahar)
units <- sample(N, size = n, replace = FALSE)
10.2 Ratio estimator 183

mysample <- grdKandahar[units, ]

b <- mean(mysample$poppy) / mean(mysample$agri)
tx_pop <- sum(grdKandahar$agri)
print(tz_ratio <- b * tx_pop)

[1] 55009.69

e <- mysample$poppy - b * mysample$agri

print(se_tz_ratio <- sqrt(Nˆ2 * (1 - (n / N)) * var(e) / n))

[1] 18847.31

An improved variance approximation is obtained with Equation (10.18). For

the ratio model and simple random sampling, the calibrated weights are equal
to (Särndal et al. (1992), p. 248)

𝑡(𝑥)
𝑔= , (10.27)
𝑡u�̂ (𝑥)

with 𝑡(𝑥) the population total of the covariate and 𝑡u�̂ (𝑥) the 𝜋 estimate of the
population total of the covariate.

pi <- n / N
tx_HT <- sum(mysample$agri / pi)
g <- tx_pop / tx_HT
S2ge <- sum(gˆ2 * eˆ2) / (n - 1)
print(se_tz_ratio <- sqrt(Nˆ2 * (1 - n / N) * S2ge / n))

[1] 17149.62

The ratio estimate and the estimated standard error of the ratio estimator can
be computed with package survey as follows.

mysample$N <- N
design_si <- svydesign(id = ~ 1, data = mysample, fpc = ~ N)
b <- svyratio(~ poppy, ~ agri, design = design_si)
predict(b, total = tx_pop)

$total
agri
poppy 55009.69

$se
agri
poppy 17149.62
184 10 Model-assisted estimation

FIGURE 10.3: Approximated sampling distribution of the ratio estimator

(Ratio) and the 𝜋 estimator (HT) of the total poppy area (ha) in Kandahar
with simple random sampling without replacement of size 50.

Figure 10.3 shows the sampling distribution of the ratio estimator and the
𝜋 estimator, obtained by repeating simple random sampling of size 50 and
estimation 10,000 times. The average of the 10,000 ratio estimates of the total
poppy area equals 62,512 ha. The population total of poppy equals 63,038
ha, so the estimated bias of the ratio estimator equals -526 ha. The boxplots
in Figure 10.3 show that the ratio estimator has less extreme outliers. The
standard deviation of the 10,000 ratio estimates equals 24,177 ha. The gain in
precision due to the ratio estimator, quantified by the ratio of the variance of
the 𝜋 estimator to the variance of the ratio estimator, equals 1.502.

Exercises

3. Write an R script to compute the ratio of the population total

poppy area to the population total agricultural area (𝑡(𝑧)/𝑡(𝑥)).
Then use all data to fit a linear model without intercept for the
poppy area, using the agricultural area as a covariate, assuming
that the variance of the residuals is proportional to the agricultural
area (heteroscedastic model). Hint: use function lm with argument
formula = poppy ~ agri - 1 and argument weights = 1 / agri. Also fit
a model without intercept, assuming a constant variance of the
residuals (homoscedastic model). Compare the estimated slopes of
10.2 Ratio estimator 185

the two models with the ratio of the total poppy area to the total
agricultural area.

10.2.1 Ratio estimators with stratified simple random

sampling

With stratified simple random sampling, there are, similar to the regression
estimator, two options for estimating a population parameter: either estimate
the ratios separately for the strata or estimate a combined ratio. The separate
ratio estimator of the population total is

u�
̂
𝑡sratio ̂
(𝑧) = ∑ 𝑡ratio,ℎ (𝑧) , (10.28)
ℎ=1

with

̂ (𝑧)
𝑡u�,ℎ
̂
𝑡ratio,ℎ (𝑧) = 𝑡 (𝑥) , (10.29)
̂ (𝑥) ℎ
𝑡u�,ℎ

in which 𝑡u�,ℎ
̂ (𝑧) and 𝑡u�,ℎ
̂ (𝑥) are the 𝜋 estimators of the population total of
the study variable and the covariate for stratum ℎ, respectively.
The combined ratio estimator is

∑u� 𝑡 ̂ (𝑧)
ℎ=1 u�,ℎ
̂
𝑡cratio (𝑧) = 𝑡(𝑥) . (10.30)
∑u� 𝑡 ̂ (𝑥)
ℎ=1 u�,ℎ

The code chunks below show how the combined and separate regression estimate
can be computed with package survey. First, two equal-sized strata are
computed using the median of the covariate agri as a stratum bound. Stratum
sample sizes are computed, and a stratified simple random sample without
replacement is selected.

median_agri <- quantile(grdKandahar$agri, probs = 0.5)

grdKandahar$stratum <- findInterval(grdKandahar$agri, median_agri) + 1
N_h <- table(grdKandahar$stratum)
n_h <- round(n * N_h / sum(N_h))
set.seed(314)
units <- sampling::strata(grdKandahar, stratanames = "stratum",
size = n_h, method = "srswor")
mysample <- getdata(grdKandahar, units)
186 10 Model-assisted estimation

The stratum sizes N_h are added to mysample, function svydesign specifies the
sampling design, function svyratio estimates the population ratio and its
variance, and finally function predict estimates the population total.

lut <- data.frame(stratum = c(1, 2), N_h)

mysample <- merge(x = mysample, y = lut)
design_stsi <- svydesign(
id = ~ 1, strata = ~ stratum, data = mysample, fpc = ~ Freq)
common <- svyratio(~ poppy, ~ agri, design_stsi, separate = FALSE)
predict(common, total = sum(grdKandahar$agri))

$total
agri
poppy 28389.02

$se
agri
poppy 8845.847

The same estimate is obtained with function calibrate.

mysample_cal <- calibrate(

design_stsi, ~ agri - 1, population = tx_pop, variance = 1)
svytotal(~ poppy, mysample_cal)

total SE
poppy 28389 8845.8

Computing the separate ratio estimator goes along the same lines. Function
svyratio with argument separate = TRUE estimates the ratio and its variance for
each stratum separately. To predict the population total, the stratum totals of
the covariate must be passed to function predict using argument total.

separate <- svyratio(~ poppy, ~ agri, design_stsi, separate = TRUE)

tx_h_pop <- tapply(grdKandahar$agri, INDEX = grdKandahar$stratum, FUN = sum)
predict(separate, total = tx_h_pop)

$total
agri
poppy 28331.32

$se
agri
poppy 8882.492
10.2 Ratio estimator 187

10.2.2 Poststratified estimator

In stratified random sampling (Chapter 4), the population is divided into

several disjoint subpopulations and from each subpopulation a probability
sample is selected. The subpopulations then serve as strata. The larger the
difference in the stratum means and the smaller the variance within the strata,
the larger the gain in precision compared to simple random sampling, see
Subsection 4.1.2.
The alternative to using the subpopulations as strata at the stage of sampling
is to use them as poststrata in estimating the population mean. For instance,
if we have selected a simple random sample from a spatial population and we
have a map of subpopulations possibly related to the study variable, then these
subpopulations still can be used in the poststratified estimator. What needs
to be done only is to classify the selected units. Hereafter, the subpopulations
that serve as poststrata are referred to as groups.
For any probability sampling design, the population mean can be estimated by

u�u�
u�
𝑡u�̂ (𝑧) u� ∑u�∈u� u�u�
̄̂
𝑧pos = ∑ 𝑤u� = ∑ 𝑤u� u�
1
, (10.31)
u�=1 𝑁u�̂ u�=1 ∑ u�∈u�u� u�u�

where 𝒮u� is the sample from group 𝑔, 𝑤u� = 𝑁u� /𝑁 is the relative size of group
𝑔, 𝑡u�̂ (𝑧) is the estimated total of the study variable for group 𝑔, 𝑁 ̂u� is the
estimator of the size of group 𝑔, and 𝜋u� is the inclusion probability of unit
𝑘. The estimated group means are weighted by their relative sizes 𝑤u� , which
are assumed to be known. In spite of this, the group means are estimated by
dividing the estimated group totals by their estimated size, 𝑁 ̂u� , because this
ratio estimator is more precise than the 𝜋 estimator of the group mean.
The poststratified estimator is the natural estimator for the one-way ANOVA
model,

𝑍u� = 𝜇u� + 𝜖u�

(10.32)
𝜎2u� = 𝜎2u� ,

with 𝜇u� the mean for group (subpopulation) 𝑔 = 1, … , 𝐺 and 𝜎2u� the variance
of the study variable of group 𝑔.
For simple random sampling, the poststratified estimator reduces to

u�
̄̂ = ∑ 𝑤u� 𝑧u�̄ u� ,
𝑧pos (10.33)
u�=1
188 10 Model-assisted estimation

where 𝑧u�̄ u� is the sample mean of group 𝑔. If for all groups we have at least
two sampling units, 𝑛u� ≥ 2, the variance of this poststratified estimator of the
mean can be estimated by

u� ̂
𝑆2
(10.34)
u�
𝑉(̂ 𝑧pos
̄̂ |𝐧u� ) = ∑ 𝑤2u� ,
u�=1
𝑛u�

where 𝑛u� is the number of sampling units in group 𝑔, and 𝑆

̂2 is the estimated
u�
spatial variance of 𝑧 in group 𝑔, which for simple random sampling can be
estimated by

1
̂
𝑆2 =
u� ∑ (𝑧 − 𝑧u�̄ u� )2 . (10.35)
𝑛u� − 1 u�∈u� u�
u�

This is an estimator of the conditional sampling variance, i.e., the variance of

the poststratified estimator over all simple random samples with group sample
sizes, collected in the vector 𝐧u� , equal to the group sample sizes in the sample
actually selected. The poststratified estimator requires that the sizes (areas)
of the strata are known. See Section 11.1 for a sampling strategy that does
not require known stratum sizes.
The poststratified estimator is illustrated with study area Voorst. We consider
the situation that we do not have the map with the five combinations of soil
type and land use that served as strata in Chapter 4. The soil-land use classes
(groups) used in the poststratified estimator are only observed at the selected
sampling units. Only three poststrata are distinguished: the original strata
BA, EA, and PA are merged into one stratum SA with function fct_collapse
of package forcats (Wickham, 2021). The sizes of these poststrata must be
known.

library(forcats)
grdVoorst$poststratum <- fct_collapse(
grdVoorst$stratum, SA = c("BA", "EA", "PA"))
print(N_g <- tapply(grdVoorst$z, INDEX = grdVoorst$poststratum, FUN = length))

SA RA XF
5523 659 1346

One hundred points are selected by simple random sampling with replacement.
The expected sample sizes per group are proportional to the size of the groups,
𝐸(𝑛u� /𝑛) = 𝑁u� /𝑁, but for a single sample the sample proportions may deviate
considerably from the population proportions.
10.2 Ratio estimator 189

n <- 100
N <- nrow(grdVoorst)
set.seed(314)
units <- sample(N, size = n, replace = TRUE)
mysample <- grdVoorst[units, ]
n_g <- tapply(mysample$z, INDEX = mysample$poststratum, FUN = length)
print(n_g)

SA RA XF
71 6 23

The population mean is estimated by first computing the sample means per
group, followed by computing the weighted average of the sample means, using
the relative sizes of the groups as weights.

mz_g <- tapply(mysample$z, INDEX = mysample$poststratum, FUN = mean)

w_g <- N_g / N
print(mz_pst <- sum(w_g * mz_g))

[1] 85.11039

The variance of the estimator of the mean is estimated by computing the

sample variances per group, dividing these by the sample sizes per group, and
computing the weighted average, using as weights the squared relative group
sizes. This estimated sampling variance is the variance of the estimator of the
mean over all simple random samples with 71 units of group SA, 6 units of
group RA, and 23 units of XF.

S2z_g <- tapply(mysample$z, INDEX = mysample$poststratum, FUN = var)

v_mz_g <- S2z_g / as.numeric(n_g)
print(condse_mz_pst <- sqrt(sum(w_gˆ2 * v_mz_g)))

[1] 4.49969

Note that this variance estimator can only be computed with at least two units
per group. For this reason, I recommend using a limited number of groups,
especially for small sample sizes.
Function postStratify of package survey can be used to compute the post-
stratified estimator and its standard error.

mysample$weights <- N / n
design_si <- svydesign(id = ~ 1, weights = ~ weights, data = mysample)
pop <- data.frame(poststratum = c("SA", "RA", "XF"), Freq = N_g)
mysample_pst <- postStratify(
190 10 Model-assisted estimation

design_si, strata = ~ poststratum, population = pop)

svymean(~ z, mysample_pst)

mean SE
z 85.11 4.3942

Lohr (1999) warns about data snooping. By defining groups after analysing
the data, arbitrarily small sampling variances of the estimated mean can be
obtained.

10.3 Model-assisted estimation using machine learning

techniques
Breidt and Opsomer (2017) review model-assisted estimators based on machine
learning techniques. Of special interest is the general approach proposed by Wu
and Sitter (2001) for incorporating non-linear predictions in the model-assisted
estimator. They show how non-linear predictions of the study variable, for
instance obtained by a regression tree or random forest, can be used in the
model-calibration estimator:

1 1 1 u� 1 𝑚(𝐱
̂ u� )
̄̂
𝑧MC = 𝑧u�̄̂ + 𝑎̂ (1 − ∑ ) + 𝑏̂ ( ∑ 𝑚(𝐱 ̂ u� ) − ∑ ) ,
𝑁 u�∈u� 𝜋u� 𝑁 u�=1 𝑁 u�∈u� 𝜋u�
(10.36)

with 𝑏̂ a slope coefficient estimated by

̂ u� ) − 𝑚̄̂ u� }{𝑧u� − 𝑧u�̄̂ }

∑u�∈u� 1/𝜋u� {𝑚(𝐱
𝑏̂ = , (10.37)
∑ ̂ u� ) − 𝑚̄̂ u� }2
1/𝜋u� {𝑚(𝐱
u�∈u�

with 𝑧u�̄̂ the 𝜋 estimator of the population mean of the study variable, 𝑚̄̂ u� the
𝜋 estimator of the population mean of the predicted values, and 𝑎̂ an intercept
estimated by

1 𝑧
𝑎̂ = (1 − 𝑏)̂ ( ∑ u� ) . (10.38)
𝑁 u�∈u� 𝜋u�
10.3 Model-assisted estimation using machine learning techniques 191

The second term in Equation (10.36) cancels for all sampling designs for which
the sum of the design weights, i.e., the sum of the reciprocal of the inclusion
probabilities, equals the population size: ∑u�∈u� 1/𝜋u� = 𝑁. Only for some
unequal probability sampling designs this may not be the case.
The alternative is to plug the fitted values 𝑚(𝐱
̂ u� ) into the generalised difference
estimator, Equation (10.2). If we drop the second term, the model-calibration
estimator can be rewritten as

1 u� ̂ 1 𝑧 − 𝑏̂ 𝑚(𝐱
̂ u� )
̄̂
𝑧MC = ̂ u� ) + ∑ u�
∑ 𝑏 𝑚(𝐱 . (10.39)
𝑁 u�=1 𝑁 u�∈u� 𝜋u�

Comparison with the generalised difference estimator, Equation (10.2), shows

that these two estimators are equivalent when 𝑏̂ = 1. For non-linear working
models, generally 𝑏̂ ≠ 1, so that these two estimators are not the same. Wu
(2003) shows that the calibration estimator has a general optimality property.
In case you are confused by all these model-assisted estimators, let me clarify.
The most general estimator is the model-calibration estimator. If we take for 𝑏̂
the value 1, this estimator is equivalent to the generalised difference estimator
(Equation (10.2)). The predictions 𝑚(𝐱̂ u� ) in these estimators can be computed
either by a linear model or a non-linear model. If a linear model is used in
the generalised difference estimator, this estimator is equal to the generalised
regression estimator (Equation (10.8)). With linear models, 𝑏̂ in Equation
(10.36) equals 1, so that all three estimators are equal.
For simple random sampling, the inclusion probabilities of the units are the
same for all units: 𝜋u� = 𝑛/𝑁, reducing Equations (10.36) and (10.37) to

1 1 u� 1
̄̂
𝑧MC = ∑ 𝑧u� + 𝑏̂SI ( ∑ 𝑚(𝐱 ̂ u� ) − ∑ 𝑚(𝐱 ̂ u� )) , (10.40)
𝑛 u�∈u� 𝑁 u�=1 𝑛 u�∈u�

with 𝑏̂SI equal to

∑u�∈u� {𝑚(𝐱
̂ u� ) − 𝑚̄ u� }{𝑧u� − 𝑧u�̄ }
𝑏̂SI = , (10.41)
̂ u� ) − 𝑚̄ u� }2
∑u�∈u� {𝑚(𝐱
with 𝑚̄ u� the sample mean of the predicted values.
An estimator of the variance of the model-assisted calibration estimator is

𝑉(̂ 𝑧MC
̄̂ ) = 𝑉(̂ 𝑒u�̄̂ ) , (10.42)

with 𝑒u�̄̂ the 𝜋 estimator of the population mean of the residuals 𝑒. For sampling
designs with fixed sample size, these residuals are equal to 𝑒u� = 𝑧u� − 𝑏̂ 𝑚(𝐱
̂ u� ).
For simple random sampling with replacement from finite populations and
192 10 Model-assisted estimation

simple random sampling from infinite populations, the variance estimator

equals

̂
𝑆2 (𝑒)
𝑉(̂ 𝑧MC
̄̂ ) = , (10.43)
𝑛

with 𝑆
̂2 (𝑒) the estimated population variance of the residuals.

An estimator of the variance of the generalised difference estimator is

̄̂ ) = 𝑉(̂ 𝑑u�̄̂ ) ,
𝑉(̂ 𝑧dif (10.44)

with 𝑑u�̄̂ the 𝜋 estimator of the population mean of the differences 𝑑u� =
𝑧u� − 𝑚(𝐱
̂ u� ).
The data of Eastern Amazonia are used to illustrate model-assisted estimation
of AGB, using five environmental covariates in predicting AGB. First, a
regression tree is used for prediction, after that a random forest is used for
prediction. For an introduction to regression trees and random forest modelling,
see this blog1 . In this blog the study variable is a categorical variable, whereas
in our example the study variable is quantitative and continuous. This is not
essential. The only difference is the measure for quantifying how good a split
is. With a quantitative study variable, this is quantified by the following sum
of squares:

2
𝑆𝑆 = ∑ ∑ (𝑧u�u� − 𝑧u�̄ u� )2 , (10.45)
u�=1 u�∈u�u�

with 𝑧u�̄ u� the sample mean of group 𝑔.

10.3.1 Predicting with a regression tree

A simple random sample without replacement of size 100 is selected.

N <- nrow(grdAmazonia)
n <- 100
set.seed(314)
units <- sample(N, size = n, replace = FALSE)
covs <- c("SWIR2", "Terra_PP", "Prec_dm", "Elevation", "Clay")
mysample <- grdAmazonia[units, c("AGB", covs)]

1 https://victorzhou.com/blog/intro-to-random-forests/
10.3 Model-assisted estimation using machine learning techniques 193

Package rpms (Toth, 2021) is used to build a regression tree for AGB, using
all five covariates as predictors. Note that I now use the original untransformed
SWIR2 as a predictor. Transforming predictors so that the relation with the
study variable becomes linear is not needed when fitting a non-linear model
such as a regression tree. Figure 10.4 shows the fitted tree.

library(rpms)
tree <- rpms(
rp_equ = AGB ~ SWIR2 + Terra_PP + Prec_dm + Elevation + Clay,
data = as.data.frame(mysample), pval = 0.05)

FIGURE 10.4: Regression tree for AGB (109 kg ha-1 ) calibrated on a simple
random sample of size 100 from Eastern Amazonia.

The regression tree is used to predict AGB for all population units.

AGBpred <- predict(tree, newdata = grdAmazonia)

The population mean is then estimated by the generalised difference estimator.

d <- grdAmazonia$AGB[units] - AGBpred[units]

mean(AGBpred) + mean(d)

[1] 226.7433

Its standard error is estimated by the square root of the variance of the
estimator of the mean differences.

S2d <- var(d)

sqrt((1 - n / N) * S2d / n)

[1] 3.212222
194 10 Model-assisted estimation

This estimation procedure is implemented in function gregTree of package mase

(McConville et al., 2021).

library(mase)
pi <- rep(n / N, n)
res <- gregTree(
mysample$AGB, xsample = mysample[, covs],
xpop = grdAmazonia[, covs], pi = pi,
var_est = TRUE, var_method = "LinHTSRS")
res$pop_mean

[1] 226.7433

sqrt(res$pop_mean_var)

[1] 3.212222

The variance of the estimator of the mean can also be estimated by bootstrap-
ping the sample (Lohr (1999), section 9.3.3).

res <- gregTree(

mysample$AGB, xsample = mysample[, covs],
xpop = grdAmazonia[, covs], pi = pi,
var_est = TRUE, var_method = "bootstrapSRS", B = 100)

sqrt(res$pop_mean_var)

[,1]
[1,] 4.076

The standard error obtained by the bootstrap is considerably larger than the
previous standard error based on a Taylor linearisation of the estimator of the
mean. As we will see hereafter, the Taylor linearisation seriously underestimates
the true standard error.
The simple random sampling of 100 units and the model-assisted estimation
are repeated 500 times, using a regression tree for prediction. The variance is
estimated by Taylor linearisation (var_method = LinHTSRS) and by bootstrapping
(var_method = bootstrapSRS) using 100 bootstrap samples.
The variance of the 500 estimated population means of AGB is 20.8 (109
kg ha-1 )2 . Estimation of the variance through Taylor linearisation strongly
underestimates the variance: the average of the 500 estimated variances equals
14.5 (109 kg ha-1 )2 . On the contrary, the bootstrap variance estimator overesti-
mates the variance: the average of the 500 estimated variances equals 23.4 (109
kg ha-1 )2 . I prefer to overestimate my uncertainty about the mean, instead
10.3 Model-assisted estimation using machine learning techniques 195

of being overoptimistic, and so I would recommend to report the bootstrap

variance.

10.3.2 Predicting with a random forest

The package ranger (Wright and Ziegler, 2017) is used to fit a random forest
(RF) model for AGB using the five environmental covariates as predictors
and the simple random sample of size 100 selected in the previous subsection.
Function importance shows how often the covariates are used in a binary splitting.
All five covariates are used, SWIR2 by far most often.

library(ranger)
set.seed(314)
forest.sample <- ranger(
AGB ~ ., data = mysample, num.trees = 1000, importance = "impurity")
importance(forest.sample)

SWIR2 Terra_PP Prec_dm Elevation Clay

466318.39 214716.35 136505.65 62429.66 34612.68

Out-of-bag predictions for the selected units are saved in element predictions
of the output object of function ranger. The fitted model is also used to predict
AGB at all units (raster cells), using function predict.

AGBpred_OOB <- forest.sample$predictions

res <- predict(forest.sample, data = grdAmazonia, type = "response")
AGBpred <- res$predictions

Finally, the model-calibration estimate and the generalised difference estimate

are computed. Both estimators and their variances are computed in two ways.
They differ in how the study variable AGB is predicted for the sampling units:

1. using all trees of the forest (the predictions obtained with function
predict); or
2. using only the trees calibrated on bootstrap samples that do not
include the sampling unit used as a prediction unit. These out-of-bag
predictions are stored in element predictions of the output object of
function ranger.

The next code chunk shows how the model-calibration estimate can be com-
puted with the AGB data of the simple random sample and the RF predictions
of AGB. First, all trees are used.
196 10 Model-assisted estimation

numer <- sum((AGBpred[units] - mean(AGBpred[units])) *

(mysample$AGB - mean(mysample$AGB)))
denom <- sum((AGBpred[units] - mean(AGBpred[units]))ˆ2)
b <- numer / denom
mz_MC <- mean(mysample$AGB) + b * (mean(AGBpred) - mean(AGBpred[units]))
u <- mysample$AGB - AGBpred[units] * b
v_mz_MC <- (1 - n / N) * var(u) / n

Next, the out-of-bag predictions are used.

numer <- sum((AGBpred_OOB - mean(AGBpred_OOB)) *

(mysample$AGB - mean(mysample$AGB)))
denom <- sum((AGBpred_OOB - mean(AGBpred_OOB))ˆ2)
b_OOB <- numer / denom
mz_MC_OOB <- mean(mysample$AGB) + b_OOB * (mean(AGBpred) - mean(AGBpred_OOB))
u_OOB <- mysample$AGB - AGBpred_OOB * b_OOB
v_mz_MC_OOB <- (1 - n / N) * var(u_OOB) / n

The two calibration estimates are about equal: 226.8 109 kg ha-1 using sample
predictions obtained with function predict, and 227.3 109 kg ha-1 with the
out-of-bag sample predictions. However, their estimated variances are largely
different: 2.35 (109 kg ha-1 )2 and 12.46 (109 kg ha-1 )2 , respectively.
In the next code chunk, the generalised difference estimate (Equation (10.2))
is computed. Similar to the model-calibration estimate, this difference estimate
is computed from predictions based on all trees and from the out-of-bag
predictions.

d <- mysample$AGB - AGBpred[units]

mz_MD <- mean(AGBpred) + mean(d)
v_mz_MD <- (1 - n / N) * var(d) / n
#using out-of-bag predictions
d_OOB <- mysample$AGB - AGBpred_OOB
mz_MD_OOB <- mean(AGBpred) + mean(d_OOB)
v_mz_MD_OOB <- (1 - n / N) * var(d_OOB) / n

For the difference estimator the results are very similar. The two difference
estimates are 226.6 109 kg ha-1 and 227.0 109 kg ha-1 , and their estimated
variances are 2.70 (109 kg ha-1 )2 and 12.80 (109 kg ha-1 )2 , respectively. The
model-calibration estimate and the generalised difference estimate are nearly
equal.
The sampling and estimation are repeated: 1,000 simple random samples with-
out replacement of size 100 are selected. Each sample is used to calibrate a RF,
each forest consisting of 1,000 trees. This results in 2 × 1,000 model-calibration
10.3 Model-assisted estimation using machine learning techniques 197

estimates and their estimated variances as well as 2 × 1,000 difference estimates

and their estimated variances. To limit the computing time, a 5 km × 5 km
subgrid of grdAmazonia is used for selecting the simple random samples and for
predicting AGB with the RF.
For each estimator, the 1,000 estimated means are used to compute the relative
bias:

1
∑1000 𝑧u�̄̂ − 𝑧 ̄
(10.46)
1000 u�=1
𝑏𝑖𝑎𝑠 = .
𝑧̄
Besides, for each estimator the variance of the 1,000 estimates is computed,
which can be compared with the mean of the 1,000 estimated variances. The
mean of the estimated variances is used to compute the variance to mean
squared error ratio:

1
∑1000 𝑉(̂ 𝑧u�̄̂ )
(10.47)
1000 u�=1
𝑅= ,
𝑀𝑆𝐸
with

1 1000 ̂
𝑀𝑆𝐸 = ∑ (𝑧 ̄ − 𝑧)̄ 2 . (10.48)
1000 u�=1 u�

Ideally, this ratio equals 1. If it is smaller than 1, the variance estimator

underestimates the mean squared error.
Finally, the relative efficiency is computed as the ratio of the MSE of the
𝜋 estimator and the MSE of a model-assisted estimator. The 𝜋 estimator is
unbiased, so the MSE equals the variance, which can be computed without
error by the population variance divided by the sample size. If the relative
efficiency is larger than 1, the model-assisted estimator is more accurate than
the 𝜋 estimator.
The variance of the 1,000 model-calibration estimates (experimental variance),
using all trees to predict AGB for the sampled units, equals 15.0. This is 5.65
times larger than the average of the 1,000 estimated variances, which equals
2.66. When using the out-of-bag sample predictions, the experimental variance
is about equal to the average of the 1,000 variance estimates: 15.5 vs. 15.0.
The reason of underestimation of the variance when predicting AGB at sample
units with function predict is that all 1,000 trees are used in prediction,
including the trees calibrated on bootstrap samples that contain the unit in the
sample to be predicted. On the contrary, the out-of-bag predicted values are
computed as the average of predictions from the trees calibrated on bootstrap
samples that do not contain the sample unit to be predicted. The default
sample fraction is 0.632, see argument sample.fraction of function ranger, so
198 10 Model-assisted estimation

TABLE 10.1: Summary statistics of 1,000 calibration estimates of the mean

AGB (109 kg ha-1 ) and their estimated variances in Eastern Amazonia, using
out-of-bag sample predictions from a RF model with five covariates as a
working model, for simple random sampling without replacement and sample
sizes 50, 100, and 250.

50 100 250
Relative bias -0.0014 -0.0006 -0.0012
Experimental variance 33.9075 15.5255 5.6111
Mean variance estimates 36.1512 14.9750 5.0467
Variance to MSE ratio 1.0642 0.9642 0.8882
Relative efficiency 5.0066 5.4755 5.9866

with 1,000 trees these predictions are the average of, on average, 368 tree
predictions. This explains why the out-of-bag prediction errors are larger
than the prediction errors obtained with function predict. In other words, the
variance of the out-of-bag differences 𝑑 and of the out-of-bag residuals 𝑢 are
larger than those obtained with predicting using all trees.
Hereafter, I report the results obtained with the out-of-bag samples only.
The relative bias is negligibly small for all sample sizes (Table 10.1). For
𝑛 = 250 and 100 the average of the 1,000 estimated variances is smaller than
the variance of the 1,000 estimated means, whereas for 𝑛 = 50 the average of
the 1,000 variance estimates is larger than the experimental variance of the
estimated means. The variance to MSE ratio is smaller than 1 for 𝑛 = 250 and
100, but larger than 1 for 𝑛 = 50. The model-calibration estimator is much
more accurate than the 𝜋 estimator for all three sample sizes, as shown by the
high relative efficiencies. The relative efficiency increases with the sample size.
The summary statistics for the performance of the generalised difference
estimator are very similar to those of the model-calibration estimator (Table
10.2).

10.4 Big data and volunteer data

Over the past decades numerous large data sets have become available, and
this number will further increase in the future; think of data sets collected
by satellites. These data sets may contain valuable information about study
variables, so that they can be used in model-assisted estimation of global (this
chapter) or local means and totals (Chapter 14) of these study variables.
10.4 Big data and volunteer data 199

TABLE 10.2: Summary statistics of 1,000 generalised difference estimates

of the mean AGB (109 kg ha-1 ) and their estimated variances in Eastern
Amazonia, using out-of-bag sample predictions from a RF model with five
covariates as a working model, for simple random sampling without replacement
and and sample sizes 50, 100, and 250.

50 100 250
Relative bias -0.0008 -0.0006 -0.0013
Experimental variance 34.5188 16.2339 5.6891
Mean variance estimates 38.4629 15.4568 5.1125
Variance to MSE ratio 1.1144 0.9520 0.8867
Relative efficiency 4.9275 5.2379 5.8997

Another interesting source of information is the geographic data collected by

volunteers. Although these data typically are from non-probability samples,
they can nevertheless render valuable information about the global or local
means of study variables.
When the volunteer data are supplemented by a probability sample with
observations of the study variable, the volunteer data can be used at the design
stage and/or at the estimation stage. As an example of the first approach, the
volunteer data are used to predict the study variable at a fine discretisation
grid. These predictions are then used to construct strata for supplemental
sampling, for instance by the cum-root-f method (Section 4.4) or using the
approach described in Section 13.2.
At the estimation stage, the volunteer data are used to predict the study
variable at points of the supplemental probability sample and at the nodes of a
discretisation grid. These predictions are then used in model-assisted estimation,
using the generalised difference or regression estimator, as explained in this
chapter. For more details and a simulation study, see Brus and de Gruijter
(2003).
Stehman et al. (2018) compared the model-assisted estimation approach with
a certainty stratum approach for estimating the area covered by land cover
classes and the accuracy of land cover maps. The volunteer data are treated
as the data of a stratum of which all units are observed. A probability sample
is selected from the remaining units not observed by the volunteers. The total
(area of a land cover class) of the certainty stratum is added to the estimated
total of the subpopulation not observed by the volunteers.
The model-assisted approach requires a supplemental sample with observations
of the study variable 𝑧. Kim and Wang (2019) described an approach in which
the big data sample is combined with a probability sample with observations
of one or more ancillary variables.
200 10 Model-assisted estimation

Kim and Wang (2019) also describe an alternative approach that does not
require a probability sample at all. In this approach, the big data sample is
subsampled to correct the selection bias in the big data sample. The subsample
is selected by inverse sampling, using data on an ancillary variable 𝑥, either from
a census or a probability sample. The subsample is selected with conditional
inclusion probabilities equal to the subsample size multiplied by an importance
weight (Kim and Wang (2019), equation 4).
11
Two-phase random sampling

The regression and ratio estimators of Chapter 10 require that the means
of the ancillary variables are known. If these are unknown, but the ancillary
variables can be measured cheaply, one may decide to estimate the population
means of the ancillary variables from a large sample. The study variable is
measured in a random subsample of this large sample only. This technique
is known in the sampling literature as two-phase random sampling or double
sampling. Another application of two-phase sampling is two-phase sampling
for stratification. Stratified random sampling (Chapter 4) requires a map with
the strata. The poststratified estimator of Subsection 10.2.2 requires that
the sizes of the strata are known. With two-phase sampling for stratification,
neither a map of the strata nor knowledge of the stratum sizes is required.
Note that the term ‘phase’ does not refer to a period of time; all data can be
collected in one sampling campaign. Let me also explain the difference with
two-stage cluster sampling (Chapter 7). In two-stage cluster random sampling,
we have two types of sampling units, clusters of population units and individual
population units. In two-phase sampling, we have one type of sampling unit
only, the objects of a discrete population or the elementary sampling units of
a continuous population (Section 1.1).
In two-phase sampling for regression and two-phase sampling for stratification,
the two phases have the same aim, i.e., to estimate the population mean of
the study variable. The observations of the covariate(s) and/or strata in the
first phase are merely done to increase the precision of the estimated mean of
the study variable. Another application of two-phase sampling is subsampling
an existing probability sample designed for a different aim. So, in this case the
study variable observed in the second-phase sample may not be related to the
variables observed in the first-phase sample.
An example is LUCAS-Topsoil (Ballabio et al., 2019) which is a subsample of
approximately 22,000 units sampled from a much larger sample, the LUCAS
sample, designed for estimating totals of land use and land cover classes across
the European Union. It was not feasible to observe the soil properties at all
sites of the LUCAS sample, and for this reason a subsample was selected.
Regrettably, this subsample is not a probability sample from the LUCAS
sample: the inclusion probabilities are either zero or unknown. Design-based
or model-assisted estimation of means of soil properties for domains of interest
is not feasible. The only option is model-based prediction.

201
202 11 Two-phase random sampling

In case the subsample is a probability subsample from the first-phase sample

and no variable observed in the first-phase sample is of use for estimating the
total or mean of the study variable observed in the subsample, the population
total can be estimated by the 𝜋 estimator:

𝑧u� 𝑧
̂ = ∑
𝑡(𝑧) = ∑ u�∗ , (11.1)
u�∈u�2
𝜋1u� 𝜋u�|u�1 u�∈u�
𝜋 u�
2

with 𝜋1u� the probability that unit 𝑘 is selected in the first phase, and 𝜋u�|u�1
the probability that unit 𝑘 is selected in the second phase, given the first-phase
sample 𝒮1 . This general 𝜋 estimator for two-phase sampling, referred to as
the 𝜋∗ estimator by Särndal et al. (1992), can be used for any combination of
probability sampling designs in the first and second phase.
To derive the variance, it is convenient to write the total estimation error as
the sum of two errors:

𝑧u� 𝑧 𝑧
̂ − 𝑡(𝑧) = ( ∑
𝑡(𝑧) − 𝑡(𝑧)) + ( ∑ u�∗ − ∑ u� )
u�∈u�1
𝜋1u� u�∈u�
𝜋 u�2u�∈u�
𝜋 1u� 1
(11.2)
= 𝑒 1 + 𝑒2 .

The first error 𝑒1 is the error in the estimated population total, as estimated
by the usual 𝜋 estimator using the study variable values for the units in
the first-phase sample. This estimator cannot be computed in practice, as the
study variable values are only known for a subset of the units in the first-phase
sample. The second error 𝑒2 is the difference between the 𝜋∗ estimator using the
study variable values for the units in the subsample only, and the 𝜋 estimator
using the study variable values for all units in the first-phase sample.
The variance of the 𝜋∗ estimator can be decomposed into the variance of these
two errors as follows:

𝑉u�1 ,u�2 (𝑡)̂ = 𝑉u�1 𝐸u�2 (𝑡|𝒮 ̂ 1 ) = 𝑉u� (𝑒1 )+𝐸u� 𝑉u� (𝑒2 |𝒮1 ) , (11.3)
̂ 1 )+𝐸u� 𝑉u� (𝑡|𝒮
1 2 1 1 2

with 𝑉u�1 and 𝐸u�1 the variance and expectation of the estimator for the
population total over repeated sampling with the design of the first phase,
respectively, and 𝑉u�2 and 𝐸u�2 the variance and expectation of the estimator
for the population total over repeated sampling with the design of the second
phase, respectively. The population mean can be estimated by the estimated
total divided by the population size 𝑁.
11.1 Two-phase random sampling for stratification 203

11.1 Two-phase random sampling for stratification

In two-phase sampling for stratification, in the first phase a large sample is
taken and the selected sampling units are all classified. The classes thus formed
are then used as strata in the second sampling phase. A stratified subsample is
selected, and the study variable is observed on the units in the subsample only.

This sampling design is applied, for instance, to monitor land use and land
cover in the European Union by the LUCAS monitoring network mentioned
above. In the first phase, a systematic random sample is selected, consisting
of the nodes of a square sampling grid with a spacing of 2 km. Land use and
land cover (LULC) are then determined at the selected grid nodes, using
orthophotographs, satellite imagery, and fieldwork. The idea is that this
procedure results in a more accurate classification of LULC at the selected
units than by overlaying the grid nodes with an existing LULC map such
as the Corine Land Cover map. The site-specific determinations of LULC
classes are then used to select a stratified random subsample (second-phase
sample). In 2018 the monitoring network was redesigned (Ballin et al., 2018).

Two-phase sampling for stratification is now illustrated with study area Voorst.
A map with five combinations of soil type and land use is available of this
study area. These combinations were used as strata in Chapter 4, and the
stratum sizes were used in the poststratified estimator of Subsection 10.2.2.
Here, we consider the situation that we do not have this map and that we do
not know the sizes of these strata either. In the first phase, a simple random
sample of size 100 is selected. In the field, the soil-land use combination is
determined for the selected points, see Figure 11.1. This time we assume that
the field determinations are equal to the classes as shown on the map.

n1 <- 100
set.seed(123)
N <- nrow(grdVoorst)
units <- sample(N, size = n1, replace = FALSE)
mysample <- grdVoorst[units, ]

The simple random sample is subsampled by stratified simple random sampling,

using the soil-land use classes as strata. The total sample size of the second
phase is set to 40. The number of points in the simple random sample per
stratum is determined. Then the subsample size per stratum is computed for
proportional allocation. Finally, function strata of package sampling (Tillé
and Matei, 2021) is used to select a stratified simple random sample without
204 11 Two-phase random sampling

replacement, see Chapter 4 for details. At the 40 points of the second-phase,

the soil organic matter (SOM) concentration is measured.

library(sampling)
n2 <- 40
n1_h <- tapply(mysample$z, INDEX = mysample$stratum, FUN = length)
n2_h <- round(n1_h / n1 * n2, 0)
units <- sampling::strata(mysample, stratanames = "stratum",
size = n2_h[unique(mysample$stratum)], method = "srswor")
mysubsample <- getdata(mysample, units)
table(mysubsample$stratum)

BA EA PA RA XF
15 8 6 4 7

FIGURE 11.1: Two-phase random sample for stratification from Voorst.

Coloured dots: first-phase sample of 100 points selected by simple random
sampling, with observations of the soil-land use combination. Triangles: second-
phase sample of 40 points selected by stratified simple random subsampling
of the first-phase sample, using the soil-land use combinations as strata, with
measurements of SOM.

With simple random sampling in the first phase and stratified simple random
sampling in the second phase, the population mean can be estimated by

u�u� 1
𝑛
𝑧 ̄̂ = ∑ 1ℎ 𝑧u�̄ 2ℎ , (11.4)
ℎ=1
𝑛1

where 𝐻u�1 is the number of strata used for stratification of the first-phase
sample, 𝑛1ℎ is the number of units in the first-phase sample that form stratum
ℎ in the second phase, 𝑛1 is the total number of units of the first-phase sample,
and 𝑧u�̄ 2ℎ is the mean of the subsample from stratum ℎ.

mz_h_subsam <- tapply(mysubsample$z, INDEX = mysubsample$stratum, FUN = mean)

mz <- sum(n1_h / n1 * mz_h_subsam, na.rm = TRUE)
11.1 Two-phase random sampling for stratification 205

The estimated population mean equals 85.6 g kg-1 . The sampling variance over
repeated sampling with both designs can be approximated1 by (Särndal et al.
(1992), equation at bottom of p. 353)

u�u�1 u�u�
𝑛
2 ̂
𝑆2
1 1
𝑛 2
u�2ℎ
𝑉(̂ 𝑧)̄̂ = ∑ ( 1ℎ ) + ∑ 1ℎ (𝑧u�̄ 2ℎ − 𝑧)̄̂ , (11.5)
ℎ=1
𝑛1 𝑛2ℎ 𝑛1 ℎ=1 𝑛1

with 𝑆
̂2
u�2ℎ the variance of 𝑧 in the subsample from stratum ℎ.

S2z_h_subsam <- tapply(mysubsample$z, INDEX = mysubsample$stratum, FUN = var)

w1_h <- n1_h / n1
v_mz_1 <- sum(w1_hˆ2 * S2z_h_subsam / n2_h)
v_mz_2 <- 1 / n1 * sum(w1_h * (mz_h_subsam - mz)ˆ2)
se_mz <- sqrt(v_mz_1 + v_mz_2)

The estimated standard error equals 7.1 g kg-1 .

The mean and its standard error can be estimated with functions twophase and
svymean of package survey (Lumley, 2021). A data frame with the first-phase
sample is passed to function twophase using argument data. A variable in this
data frame, passed to function twophase with argument subset, is an indicator
with value TRUE if this unit is selected in the second phase, and FALSE otherwise.

library(survey)
lut <- data.frame(stratum = sort(unique(mysample$stratum)), fpc2 = n1_h)
mysample <- mysample %>%
mutate(ind = FALSE,
fpc1 = N) %>%
left_join(lut, by = "stratum")
mysample$ind[units$ID_unit] <- TRUE
design_2phase <- survey::twophase(
id = list(~ 1, ~ 1), strata = list(NULL, ~ stratum),
data = mysample, subset = ~ ind, fpc = list(~ fpc1, ~ fpc2))
svymean(~ z, design_2phase)

mean SE
z 85.606 7.033

As shown in the next code chunk, the standard error is computed with the
original variance estimator, without approximation (Särndal et al. (1992),
equation (9.4.14)).

1 In the approximation, it is assumed that u� is much larger than u� , and (u�

1 1ℎ −1)/(u�1 −1)
is replaced by u�1ℎ /u�1 .
206 11 Two-phase random sampling

v_mz_1 <- 1 / Nˆ2 * N * (N - 1) *

sum((((n1_h - 1) / (n1 - 1)) - ((n2_h - 1) / (N - 1)))
* w1_h * S2z_h_subsam / n2_h)
v_mz_2 <- 1 / Nˆ2 * (N * (N - n1)) / (n1 - 1) *
sum(w1_h * (mz_h_subsam - mz)ˆ2)
sqrt(v_mz_1 + v_mz_2)

[1] 7.032964

11.2 Two-phase random sampling for regression

The simple regression estimator of Equation (10.10) requires that the popula-
tion mean of the ancillary variable 𝑥 is known. This section, however, is about
applying the regression estimator in situations where the mean of 𝑥 is unknown.
A possible application is estimating the soil organic carbon (SOC) stock in an
area. To estimate this carbon stock, soil samples are collected and analysed
in a laboratory. The laboratory measurements can be very accurate, but also
expensive. Proximal sensors can be used to derive soil carbon concentrations
from the spectra. Compared to laboratory measurements of soil the proximal
sensor determinations are much cheaper, but also less accurate. If there is a
relation between the laboratory and the proximal sensing determinations of
SOC, then we expect that the regression estimator of the carbon stock will
be more accurate than the 𝜋 estimator which does not exploit the proximal
sensing measurements. However, the population mean of the proximal sensing
determinations is unknown. What we can do is estimate this mean from a large
sample. Additionally, for a subsample of this large sample, SOC concentration
is also measured in the laboratory. This is another example of two-phase
sampling.
Intuitively, we understand that with two-phase sampling the variance of the
regression estimator of the total carbon stock is larger than when the population
mean of the proximal sensing determinations is known. There is a sampling
error in the estimated population mean of the proximal sensing determinations,
estimated from the large first-phase sample, and this error propagates to the
error in the estimated total carbon stock.
Two-phase sampling for regression is now illustrated with Eastern Amazonia
(Subsection 1.3.3). The study variable is the aboveground biomass (AGB),
and lnSWIR2 is used here as a covariate. We do have a full coverage map of
lnSWIR2, so two-phase sampling with a large first-phase sample to estimate the
population mean of lnSWIR2 is not needed. Nevertheless, hereafter a two-phase
11.2 Two-phase random sampling for regression 207

sample is selected, and the population mean of lnSWIR2 is estimated from the
first-phase sample. In doing so, the effect of ignorance of the population mean
of the covariate on the variance of the regression estimator becomes apparent.
In the next code chunk, a first-phase sample of 250 units, the dots in the plot,
is selected by simple random sampling without replacement. In the second
phase a subsample of 100 units, the triangles in the plot, is selected from the
250 units by simple random sampling without replacement. At all 250 units of
the first-phase sample the covariate lnSWIR2 is measured, whereas AGB is
measured at the 100 subsample units only.

grdAmazonia <- grdAmazonia %>%

mutate(lnSWIR2 = log(SWIR2))
n1 <- 250; n2 <- 100
set.seed(314)
units_1 <- sample(nrow(grdAmazonia), size = n1, replace = FALSE)
mysample <- grdAmazonia[units_1, ]
units_2 <- sample(n1, size = n2, replace = FALSE)
mysubsample <- mysample[units_2, ]

Figure 11.2 shows the selected two-phase sample.

FIGURE 11.2: Two-phase random sample for the regression estimator of

the mean AGB in Eastern Amazonia. Coloured dots: simple random sample
without replacement of 250 units with measurements of covariate lnSWIR2
(first-phase sample). Triangles: simple random subsample without replacement
of 100 units with measurements of AGB (second-phase sample).

Estimation of the population mean or total by the regression estimator from

a two-phase sample is very similar to estimation when the covariate mean is
known, as described in Subsection 10.1.1 (Equation (10.10)). The observations
208 11 Two-phase random sampling

of the subsample can be used to estimate the regression coefficient 𝑏. The true
population mean of the ancillary variable, 𝑥̄ in Equation (10.10), is unknown
now. This true mean is replaced by the mean as estimated from the relatively
large first-phase sample, 𝑥u�̄ 1 . The estimated mean of the covariate, 𝑥u�̄ in
Equation (10.10), is estimated from the subsample, 𝑥u�̄ 2 . This leads to the
following estimator:

𝑧 ̄̂ = 𝑧u�̄ 2 + 𝑏̂ (𝑥u�̄ 1 − 𝑥u�̄ 2 ) , (11.6)

where 𝑧u�̄ 2 is the subsample mean of the study variable, and 𝑥u�̄ 1 and 𝑥u�̄ 2 are
the means of the covariate in the first-phase sample and the subsample (i.e.,
the second-phase sample), respectively.
The sampling variance is larger than that of the regression estimator with
known mean of 𝑥. The variance can be decomposed into two components. The
first component is equal to the sampling variance of the 𝜋 estimator of the
mean of 𝑧 with the sampling design of the first phase (in this case, simple
random sampling without replacement), supposing that the study variable is
observed on all units of the first-phase sample. The second component is equal
to the sampling variance of the regression estimator of the mean of 𝑧 in the
first-phase sample, with the design of the second-phase sample (again simple
random sampling without replacement in this case):

𝑛 𝑆 ̂2 (𝑧) 𝑛 𝑆 ̂2 (𝑒)
𝑉(̂ 𝑧)̄̂ = (1 − 1 ) + (1 − 2 ) , (11.7)
𝑁 𝑛1 𝑛1 𝑛2

with 𝑆
̂ 2 (𝑒) the variance of the regression residuals as estimated from the

subsample:

1
̂
𝑆2 (𝑒) = ∑ 𝑒2 . (11.8)
(𝑛2 − 1) u�∈u� u�
2

The ratios (1 − 𝑛1 /𝑁) and (1 − 𝑛2 /𝑛1 ) in Equation (11.7) are finite popula-
tion corrections (fpcs). These fpcs account for the reduced variance due to
sampling the finite population and subsampling the first-phase sample without
replacement.

lm_subsample <- lm(AGB ~ lnSWIR2, data = mysubsample)

ab <- coef(lm_subsample)
mx_sam <- mean(mysample$lnSWIR2)
mx_subsam <- mean(mysubsample$lnSWIR2)
mz_subsam <- mean(mysubsample$AGB)
mz_reg2ph <- mz_subsam + ab[2] * (mx_sam - mx_subsam)
11.2 Two-phase random sampling for regression 209

The estimated population mean equals 228.1 109 kg ha-1 . The standard error
can be approximated as follows.

e <- residuals(lm_subsample)
S2e <- sum(eˆ2) / (n2 - 1)
S2z <- var(mysubsample$AGB)
N <- nrow(grdAmazonia)
se_mz_reg2ph <- sqrt((1 - n1 / N) * S2z / n1 + (1 - n2 / n1) * S2e / n2)

The estimated standard error equals 6.35 109 kg ha-1 .

The regression estimator for two-phase sampling and its standard error can
also be computed with package survey, as shown below. The standard error
differs from the standard error computed above because it is computed with the
g-weights, see Subsection 10.1.1. Note argument fpc = list(~ N, NULL). There
is no need to add the first-phase sample size as a second element of the list,
because this sample size is simply the number of rows of the data frame. Setting
the second element of the list to NULL does not mean that the standard error is
computed for sampling with replacement in the second phase. Function twophase
assumes that the second-phase units are always selected without replacement.

mysample <- mysample %>%

mutate(id = row_number(),
N = N,
ind = id %in% units_2)
design_2phase <- survey::twophase(
id = list(~ 1, ~ 1), data = mysample, subset = ~ ind, fpc = list(~ N, NULL))
mysample_cal <- calibrate(
design_2phase, formula = ~ lnSWIR2, calfun = "linear", phase = 2)
svymean(~ AGB, mysample_cal)

mean SE
AGB 228.07 7.2638

Exercises

1. Write an R script to select a simple random sample without replace-

ment of 250 units from Eastern Amazonia and a subsample of 100
units by simple random sampling without replacement. Repeat this
1,000 times in a for-loop.
• Use each sample selected in the first phase (sample of 250 units)
to estimate the population mean of AGB by the regression
estimator (Equation (10.10) in Chapter 10). Assume that the
AGB data are known for all units selected in the first phase.
Use lnSWIR2 as a covariate.
210 11 Two-phase random sampling

• Compute the variance of the 10,000 regression estimates of the

population mean of AGB.
• Use each two-phase sample to compute the regression estimator
of AGB for two-phase sampling (Equation (11.6)). Now only
use the AGB data of the subsample. Estimate the population
mean of lnSWIR2 from the first-phase sample of 250 units.
Approximate the variance of the regression estimator for two-
phase sampling (Equation (11.7)).
• Compute the variance of the 10,000 regression estimates of the
population mean of AGB for the two-phase sampling design.
• Compare the two variances and explain the difference.
• Compute the average of the 10,000 approximate variances and
compare the result with the variance of the 10,000 estimated
means, as estimated by the regression estimator for two-phase
sampling.
12
Computing the required sample size

An important decision in designing sampling schemes is the number of units

to select. In other words, what should the sample size be? If a certain budget
is available for sampling, we can determine the affordable sample size from
this budget. A costs model is then needed.
The alternative to deriving the affordable sample size from the budget is to start
from a requirement on the quality of the survey result obtained by statistical
inference. Two types of inference are distinguished, estimation and testing. The
required sample size depends on the type of sampling design. With stratified
random sampling, we expect that we need fewer sampling units compared
to simple random sampling to estimate the population mean with the same
precision, whereas with cluster random sampling and two-stage cluster random
sampling in general we need more sampling units. To compute the sample size
given some quality requirement, we may start with computing the required
sample size for simple random sampling and then correct this sample size to
account for the design effect. Therefore, I start with presenting formulas for
computing the required sample size for simple random sampling. Section 12.4
describes how the required sample sizes for other types of sampling design can
be derived.
Hereafter, formulas for computing the required sample size are presented for
simple random sampling with replacement of finite populations and simple
random sampling of infinite populations. For simple random sampling without
replacement (SI) of finite populations, these sample sizes can be corrected by
(Lohr, 1999)

𝑛SIR
𝑛SI = , (12.1)
1 + u�u�
SIR

with 𝑛SIR the required sample size for simple random sampling with replace-
ment.

211
212 12 Computing the required sample size

12.1 Standard error

A first option is to set a limit on the variance of the 𝜋 estimator of the
population mean, see Equation (3.11), or on the square root of this variance,
the standard error of the estimator. Given a chosen limit for the standard error
𝑠𝑒max , the required sample size for simple random sampling with replacement
can be computed by

2
𝑆∗ (𝑧)
𝑛=( ) , (12.2)
𝑠𝑒max

with 𝑆∗ (𝑧) a prior estimate of the population standard deviation. The required
sample size 𝑛 should be rounded to the nearest integer greater than the
right-hand side of Equation (12.2). This also applies to the following equations.
For the population proportion (areal fraction) as the parameter of interest,
the required sample size can be computed by (see Equation (3.18))

2
√𝑝∗ (1 − 𝑝∗ )
𝑛=( ) +1, (12.3)
𝑠𝑒max

with 𝑝∗ a prior estimate of the population proportion.

To determine the required sample size for estimating the population propor-
tion, we need a prior estimate of the population parameter of interest itself,
whereas for the population mean a prior estimate is needed of the population
standard deviation. The parameter of which a prior estimate is needed for
sample size determination is referred to as the design parameter.

Alternatively, we may require that the relative standard error, i.e., the standard
error of the estimator divided by the population mean, may not exceed a given
limit 𝑟𝑠𝑒max . In this case the required sample size can be computed by

2
𝑐𝑣∗
𝑛=( ) , (12.4)
𝑟𝑠𝑒max

with 𝑐𝑣∗ a prior estimate of the population coefficient of variation 𝑆(𝑧)/𝑧.̄

For a constraint on the relative standard error of the population proportion
estimator, we obtain

2 2
𝑝∗ (1 − 𝑝∗ ) 1 − 𝑝∗
𝑛=( ) + 1 = ( ) +1. (12.5)
𝑟𝑠𝑒max 𝑝∗ 𝑟𝑠𝑒max
12.2 Length of confidence interval 213

12.2 Length of confidence interval

Another option is to require that the length of the confidence interval of the
population mean may not exceed a given limit 𝑙max :

𝑆(𝑧)
2 𝑡u�/2,u�−1 √ ≤ 𝑙max , (12.6)
𝑛

with 𝑡u�/2,u�−1 the (1 − 𝛼/2) quantile of the t distribution with 𝑛 − 1 degrees

of freedom, 𝑆(𝑧) the population standard deviation of the study variable, and
𝑛 the sample size. The problem is that we do not know the degrees of freedom
(we want to determine the sample size 𝑛). Therefore, 𝑡u�/2,u�−1 is replaced by
𝑢u�/2 , the (1 − 𝛼/2) quantile of the standard normal distribution. Rearranging
yields

2
𝑆∗ (𝑧)
𝑛 = (𝑢u�/2 ) . (12.7)
𝑙max /2

The requirement can also be formulated as

𝑃(|𝑧 ̄̂ − 𝑧|̄ ≤ 𝑑max ) ≤ 1 − 𝛼 , (12.8)

with 𝑑max the margin of error: 𝑑max = 𝑙max /2.

An alternative is to require that with a large probability 1 − 𝛼 the absolute
value of the relative error of the estimated mean may not exceed a given limit
𝑟max . In formula:

|𝑧 ̄̂ − 𝑧|̄
𝑃( ≤ 𝑟max ) ≤ 1 − 𝛼 . (12.9)
𝑧̄

Noting that the absolute error equals 𝑟max 𝑧 ̄ and inserting this in Equation
(12.7) gives

2
𝑐𝑣∗
𝑛 = (𝑢u�/2 ) . (12.10)
𝑟max

As an example, the required sample size is computed for estimating the

population mean of the soil organic matter concentration in Voorst. The
requirement is that with a probability of 95% the absolute value of the relative
error does not exceed 10%. A prior estimate of 0.5 for the population coefficient
of variation is used.
214 12 Computing the required sample size

cv <- 0.5
rmax <- 0.1
u <- qnorm(p = 1 - 0.05 / 2, mean = 0, sd = 1)
n <- ceiling((u * cv / rmax)ˆ2)

The required sample size is 97. The same result is obtained with function
nContMoe of package PracTools (Valliant et al. (2021), Valliant et al. (2018)).

library(PracTools)
print(ceiling(nContMoe(moe.sw = 2, e = rmax, alpha = 0.05, CVpop = cv)))

[1] 97

12.2.1 Length of confidence interval for a proportion

Each of the methods for computing a confidence interval of a proportion

described in Subsection 3.3.1 can be used to compute the required sample size
given a limit for the length of the confidence interval of a proportion. The
most simple option is to base the required sample size on the Wald interval
(Equation (3.21)), so that the required sample size can be computed by

2
√𝑝∗ (1 − 𝑝∗ )
𝑛 = (𝑢u�/2 ) +1. (12.11)
𝑙max /2

The Wald interval approximates the discrete binomial distribution by a nor-

mal distribution. See the rule of thumb in Subsection 3.3.1 for when this
approximation is reasonable.
Package binomSamSize (Höhle, 2017) has quite a few functions for computing
the required sample size. Function ciss.wald uses the normal approximation.
In the next code chunk, the required sample sizes are computed for a prior
estimate of the population proportion 𝑝∗ of 0.2.

Argument d in the functions below is half the length of the confidence

interval.

library(binomSamSize)
p_prior <- 0.2
n_prop_wald <- ciss.wald(p0 = p_prior, d = 0.1, alpha = 0.05)
n_prop_agrcll <- ciss.agresticoull(p0 = p_prior, d = 0.1, alpha = 0.05)
n_prop_wilson <- ciss.wilson(p0 = p_prior, d = 0.1, alpha = 0.05)
12.3 Statistical testing of hypothesis 215

The required sample sizes are 62, 58, and 60, for the Wald, Agresti-Coull,
and Wilson approximation of the binomial proportion confidence interval,
respectively. The required sample size with function ciss.wald is one unit
smaller than as computed with Equation (12.11), as shown in the code chunk
below.

ceiling((qnorm(0.975) * sqrt(p_prior * (1 - p_prior)) / 0.1)ˆ2 + 1)

[1] 63

12.3 Statistical testing of hypothesis

The required sample size for testing a population mean with a two-sided
alternative hypothesis can be computed by (Ott and Longnecker, 2015)

𝑆2 (𝑧)
𝑛= (𝑢u�/2 + 𝑢u� )2 , (12.12)
Δ2
with Δ the smallest relevant difference of the population mean from the test
value, 𝛼 the tolerable probability of a type I error, i.e., the probability of
rejecting the null hypothesis when the population mean is equal to the test
value, 𝛽 the tolerable probability of a type II error, i.e., the probability of
not rejecting the null hypothesis when the population mean is not equal to
the test value, 𝑢u�/2 as before, and 𝑢u� the (1 − 𝛽) quantile of the standard
normal distribution. The quantity 1 − 𝛽 is the power of a test: the probability
of correctly rejecting the null hypothesis. For a one-sided test, 𝑢u�/2 must be
replaced by 𝑢u� .
In the next code chunk, the sample size required for a given target power is
computed with the standard normal distribution (Equation (12.12)), as well as
with the t distribution using function pwr.t.test of package pwr (Champely,
2020)1 . This requires some iterative algorithm, as the degrees of freedom of
the t distribution are a function of the sample size. The required sample size
is computed for a one-sample test and a one-sided alternative hypothesis.

library(pwr)
sd <- 4; delta <- 1; alpha <- 0.05; beta <- 0.2
n_norm <- (sd / delta)ˆ2 * (qnorm(1 - alpha) + qnorm(1 - beta))ˆ2
n_t <- pwr.t.test(

1 The same result is obtained with function power.t.test of the stats package.
216 12 Computing the required sample size

d = delta / sd, sig.level = alpha, power = (1 - beta),

type = "one.sample", alternative = "greater")

In this example, the required sample size computed with the t distribution is
two units larger than that obtained with the standard normal distribution:
101 vs. 99. Package pwr has various functions for computing the power of a
test given the sample size, or reversely, the sample size for a given power, such
as for the two independent samples t test, binomial test (for one proportion),
test for two proportions, etc.

12.3.1 Sample size for testing a proportion

For testing a proportion, a graph is computed with the power of a binomial

test against the sample size. This is illustrated with a one-sided alternative
𝐻u� ∶ 𝑝 > 0.20 and a smallest relevant difference of 0.10.

p_test <- 0.20; alpha <- 0.10; delta <- 0.10

n <- 1:150
power <- k_min <- numeric(length = length(n))
for (i in seq_len(length(n))) {
k_min[i] <- qbinom(p = 1 - alpha, size = n[i], prob = p_test)
power[i] <- pbinom(
q = k_min[i], size = n[i], prob = p_test + delta, lower.tail = FALSE)
}

As can be seen in the R code, as a first step for each total sample size the
smallest number of successes 𝑘min is computed at which the null hypothesis
is rejected. Then the binomial probability is computed of 𝑘min + 1 or more
successes for a probability of success equal to 𝑝test + Δ. Note that there is no
need to add 1 to k_min as with argument lower.tail = FALSE the value specified
by argument q is not included.
Figure 12.1 shows that the power does not increase monotonically with the
sample size. The graph shows a saw-toothed behaviour. This is caused by
the stepwise increase of the critical number of successes (k_min) with the total
sample size.
The required sample size can be computed in two ways. The first option is to
compute the smallest sample size for which the power is larger than or equal
to the required power 1 − 𝛽. The alternative is to compute the smallest sample
size for which the power is larger than or equal to 1 − 𝛽 for all sample sizes
larger than this.
12.3 Statistical testing of hypothesis 217

FIGURE 12.1: Power of right-tail binomial test (test proportion: 0.2; signifi-
cance level: 0.10).

n1 <- min(n[which(power > 1 - beta)])

ind <- (power > 1 - beta)
for (i in seq_len(length(n))) {
if (ind[length(n) - i] == FALSE)
break
}
n2 <- n[length(n) - i + 1]

The smallest sample size at which the desired level of 0.8 is reached is 88.
However, as can be seen in Figure 12.1, for sample sizes 89, 90, 93, 94, and 97,
the power drops below the desired level of 0.80. The smallest sample size at
which the power stays above the level of 0.8 is 98.
Alternatively, we may use function pwr.p.test of package pwr. This is an
approximation, using an arcsine transformation of proportions. The first step
is to compute Cohen’s ℎ, which is a measure of the distance between two
√ √
proportions: ℎ = 2 𝑎𝑟𝑐𝑠𝑖𝑛( 𝑝1 ) − 2 𝑎𝑟𝑐𝑠𝑖𝑛( 𝑝2 ). This can be done with
function ES.h. The value of ℎ must be positive, which is achieved when the
proportion specified by argument p1 is larger than the proportion specified by
argument p2.

h <- ES.h(p1 = 0.30, p2 = 0.20)

n_approx <- pwr.p.test(
h, power = (1 - beta), sig.level = alpha, alternative = "greater")
218 12 Computing the required sample size

The approximated sample size equals 84, which is somewhat smaller than the
required sample sizes computed above.

Exercises

1. Write an R script to compute the required sample size, given a

requirement in terms of the half-length of the confidence interval of
a population proportion. Use a normal approximation for computing
the confidence interval. Use a range of values for half the length
of the interval: 𝑑 = (0.01, 0.02, … , 0.49). Use a prior (anticipated)
proportion of 0.1 and a significance level of 0.95. Plot the required
sample size against 𝑑. Explain what you see. Why is it needed to
provide a prior proportion?

2. Do the same for a single value for the half-length of the confidence
interval of 0.2 and a range of values for the prior proportion 𝑝∗ =
(0.01, 0.02, … , 0.49). Explain what you see. Why is it not needed to
compute the required sample size for prior proportions > 0.5?

12.4 Accounting for design effect

The required sample sizes computed in the previous sections are all for simple
random sampling in combination with the 𝜋 estimator of the population mean.
But what is the required sample size for other types of sampling design, such as
stratified (simple) random sampling, systematic random sampling, two-stage
cluster random sampling, and cluster random sampling? Broadly speaking, we
expect that with stratified random sampling and systematic random sampling
the sampling variance of the estimator of the mean will be smaller than with
simple random sampling of the same number of units, whereas with two-
stage cluster random sampling and cluster random sampling we expect larger
sampling variances. Therefore, reversely, for the first two types of sampling
design, we expect that the sample size required to achieve the same level of
accuracy or confidence will be smaller than with simple random sampling,
and for the latter two design types this sample size will be larger. The design
effect is commonly quantified by the ratio of two sampling variances of the
population mean estimator (Lohr, 1999):

𝑉u� (𝑧)̄̂ 𝑉u� (𝑧)̄̂

𝑑𝑒(𝑝, 𝑧)̄̂ = = 2 , (12.13)
𝑉SI (𝑧u�̄̂ ) 𝑆 (𝑧)/𝑛
12.5 Bayesian sample size determination 219

with 𝑉u� (𝑧)̄̂ the sampling variance of an estimator (𝜋 estimator, regression

estimator) of the population mean with sampling design 𝑝 and 𝑉SI (𝑧u�̄̂ ) the
sampling variance of the 𝜋 estimator of the population mean with simple
random sampling. Given an estimate of this design effect, the required sample
size for a more complex sampling strategy (combination of sampling design
and estimator), given a constraint on the standard error or the half-length of
a confidence interval, can be computed by

𝑛(𝑝, 𝑧)̄̂ = √𝑑𝑒(𝑝, 𝑧)̄̂ 𝑛(SI, 𝜋) . (12.14)

The design effect can also be quantified by the ratio of two standard errors.
Then there is no need to take the square root of the design effect, as done in
Equation (12.14), to compute the required sample size for a more complex
design, given a constraint on the standard error or the half-length of a
confidence interval.

12.5 Bayesian sample size determination

A serious drawback of the classical frequentist approach of computing the
required sample size explained in the previous sections is that the required
sample sizes are sensitive to the design parameters 𝑆∗ , 𝑝∗ , and 𝑐𝑣∗ . We are
rather uncertain about these parameters, and therefore it is attractive to
replace a single value for these parameters by a probability distribution. This
leads to a different statistical approach of computing the required sample size,
the Bayesian approach. This Bayesian approach also offers the possibility of
accommodating existing information about the population mean or proportion.
In this section I show how this approach can be used to compute the required
sample size for estimating a population mean or proportion.
But before going into details, let me explain the basics of the Bayesian approach
of statistical inference. In the previous sections, the statistical inference was
from the frequentist perspective. How does the frequency distribution of
the estimator of the population mean look like if we repeat the selection of a
sample with a given sampling design? Is the mean of this frequency distribution,
referred to as the sampling distribution, equal to the population mean, and
what is the variance of this sampling distribution?
The Bayesian approach is fundamentally different. The frequency distribution
of the frequentist approach is replaced by a probability distribution of the
population mean reflecting our belief about the population mean. Note that
220 12 Computing the required sample size

expressing our belief in terms of a probability distribution implies that in the

Bayesian approach, contrary to the frequentist approach, the population mean
is a random variable. Where in the frequentist approach, it is incorrect to say
that the probability that the population mean is inside the 95% confidence
interval equals 95% (see Section 3.3), this is perfectly fine in the Bayesian
approach. The term confidence interval is replaced by the term credible interval
to underline the fundamental different meaning of the interval.
The first step in the Bayesian approach of statistical inference is to postulate
a prior distribution for the population parameter of interest. This prior distri-
bution expresses our belief and uncertainty about the parameter before the
sample data are taken into account.
The next step is to formalise a theory about the data. This boils down to
making an assumption about the type of distribution function of the data. Can
we safely assume that the data follow a normal or a binomial distribution?
Once the type of distribution has been specified, we can write an equation for
the probability of the data as a function of the parameter. This function is
referred to as the likelihood function.
The final step is to revise our prior belief about the population parameter of
interest, using the data and our theory about the data as expressed in the
likelihood function. This results in the posterior distribution of the parameter.
Our revised or updated belief is computed with Bayes’ rule:

𝑓(𝜃)𝑓(𝐳|𝜃)
𝑓(𝜃|𝐳) = , (12.15)
𝑓(𝐳)

with 𝑓(𝜃|𝐳) the posterior distribution, i.e., the probability density2 of the
parameter given the sample data, 𝑓(𝜃) our prior belief in the parameters
specified by a probability distribution (prior distribution), 𝑓(𝐳|𝜃) the likelihood
of the data, and 𝑓(𝐳) the probability distribution of the data.

12.5.1 Bayesian criteria for sample size computation

Equation (12.15) shows that the posterior distribution of the population

parameter of interest depends on the probability distribution of the new data
𝑓(𝐳). The problem is that these new data are not yet known. We are designing
a sample, and the data yet are to be collected, so at first glance this might seem
an unsolvable problem. However, what we could do is to simulate with the
prior probability density function a large number of values of the population
parameter(s), and with each parameter a large number of possible vectors with
𝑛 data. Each data vector is then used to update the prior to the posterior,
using Bayes’ rule (Equation (12.15)). For each posterior either the length of the
2I assume here that the parameter of interest u� is a continuous random variable.
12.5 Bayesian sample size determination 221

highest posterior density (HPD) interval with a coverage probability of 1 − 𝛼 is

computed, or reversely, the coverage probability of the HPD interval of length
𝑙max . Finally, the average of the lengths of the HPD intervals or the average
of the coverage probabilities is computed, and these averages are compared
with our precision requirement. If the average length is larger than 𝑙max , or
the coverage probability of intervals of length 𝑙max is smaller than 1 − 𝛼, then
we must increase 𝑛; if the average length is smaller than 𝑙max , or the coverage
probability of intervals of length 𝑙max is larger than 1 − 𝛼, then we must
decrease 𝑛. This whole procedure is repeated until our precision requirement is
met. Simulation is one option to compute the sample size, (partly) analytical
approaches are also available.
More formally, the procedure is as follows. The prior probability density
function on the population parameter(s) 𝜃 is used to compute for a given
sample size 𝑛 the predictive distribution of the data:

𝑓(𝐳|𝑛) = ∫ 𝑓(𝐳|𝜃, 𝑛)𝑓(𝜃)d𝜃 , (12.16)

Θ

with Θ the parameter space for 𝜃 containing all possible values of 𝜃. This
predictive distribution is also named the preposterior distribution, stressing
that the data are not yet accounted for in the distribution.
Even if 𝜃 would be fixed, we do not have only one vector 𝐳 with 𝑛 data values
but a probability distribution, from which we can simulate possible data vectors,
referred to as the data space 𝒵. In case of a binomial probability and sample
size 𝑛, the data space 𝒵 (in the form of the number of observed successes given
sample size 𝑛) can be written as the set {0, 1, … , 𝑛}, i.e., one vector of length
𝑛 with all “failures”, 𝑛 vectors of length 𝑛 with one success, (u�2 ) vectors with
two successes, etc. Each data vector is associated with a probability density
(for continuous data) or probability mass (for discrete data). As a consequence,
we do not have only one posterior distribution function 𝑓(𝜃|𝐳), but as many
as we have data vectors in the data space. For each posterior distribution
function the coverage of the HPD interval of a given length can be computed,
or reversely, the length of the HPD interval for a given coverage. This leads
to various criteria for computing the required sample size, among which are
the average length criterion (ALC), the average coverage criterion (ACC), and
the worst outcome criterion (WOC) (Joseph et al. (1995), Joseph and Bélisle
(1997)).

12.5.1.1 Average length criterion

For a fixed posterior HPD interval coverage of 100(1−𝛼)%, the smallest sample
size 𝑛 is determined such that

∫ 𝑙(𝐳, 𝑛)𝑓(𝐳|𝑛)d𝐳 ≤ 𝑙max , (12.17)

u�
222 12 Computing the required sample size

where 𝑓(𝐳|𝑛) is the predictive distribution of the data (Equation (12.16)) and
𝑙(𝐳, 𝑛) is the length of the 100(1 − 𝛼)% HPD interval for data 𝐳 and sample
size 𝑛, obtained by solving

u�+u�(u�,u�)
∫ 𝑓(𝜃|𝐳, 𝑛)d𝜃 = 1 − 𝛼 , (12.18)
u�

for 𝑙(𝐳, 𝑛), for each possible data set 𝐳 ∈ 𝒵. 𝑓(𝜃|𝐳, 𝑛) is the posterior density
of the population parameter of interest given the data 𝐳 and sample size 𝑛.
ALC ensures that the average length of 100(1 − 𝛼)% posterior HPD intervals,
weighted by 𝑓(𝐳|𝑛), is at most 𝑙max .

12.5.1.2 Average coverage criterion

For a fixed posterior HPD interval of length 𝑙max , the smallest sample size 𝑛
is determined such that

u�+u�max
∫ {∫ 𝑓(𝜃|𝐳, 𝑛)d𝜃} 𝑓(𝐳|𝑛)d𝐳 ≥ 1 − 𝛼 . (12.19)
u� u�

ACC ensures that the average coverage of HPD intervals of length 𝑙max is
at least 1 − 𝛼. The integral inside the curly brackets is the integral of the
posterior density of the population parameter of interest over the HPD interval
(𝑣, 𝑣 + 𝑙max ), given a data vector 𝐳 of size 𝑛. The mean of this integrated
posterior density of the parameter of interest 𝜃 is obtained by multiplying the
integrated density with the predictive probability of the data and integrating
over all possible data sets in 𝒵.

12.5.1.3 Worst outcome criterion

Neither ALC nor ACC guarantee that for a particular data set 𝐳 the criterion
is met, as both are defined as averages over all possible data sets in 𝒵. A more
conservative sample size can be computed by requiring that for all data sets
𝒵 both criteria are met. Joseph and Bélisle (1997) modified this criterion by
restricting the data sets to a subset 𝒲 of most likely data sets. The criterion
thus obtained is referred to as the modified worst outcome criterion, or for
short, the worst outcome criterion (WOC). So, the criterion is

u�+u�(u�,u�)
infu�∈u� {∫ 𝑓(𝜃|𝐳, 𝑛)d𝜃} ≥ 1 − 𝛼 . (12.20)
u�

The smallest sample size satisfying this condition is used as the sample size.
For instance, if the 95% most likely data sets are chosen as subspace 𝒲, WOC
12.5 Bayesian sample size determination 223

guarantees that there is 95% assurance that the length of the 100(1 − 𝛼)%
posterior HPD intervals will be at most 𝑙max . The fraction of most likely data
sets in subspace 𝒲 is referred to as the worst level.

12.5.2 Mixed Bayesian-likelihood approach

Besides the fully Bayesian approach, Joseph and Bélisle (1997) describe a
mixed Bayesian-likelihood approach for determining the sample size. In the
mixed Bayesian-likelihood approach of sample size determination, the prior
distribution of the parameter or parameters is only used to derive the predictive
distribution of the data (Equation (12.16)), not the posterior distributions of
the parameter of interest for each data vector. For analysis of the posterior
distribution, an uninformative prior is therefore used. This mixed approach
is of interest when, after the data have been collected, we prefer to estimate
the population mean from these data only, using the frequentist approach
described in previous sections.
An example of a situation where the mixed Bayesian-likelihood approach can
be attractive is the following. Suppose some data of the study variable from
the population of interest are already available, but we would like to collect
more data so that we will be more confident about the (current) population
mean once these new data are collected. The legacy data are used to construct
a prior distribution. We have doubts about the quality of the legacy data
because they were collected a long time ago and the study variable might
have changed in the meantime. In that case, the mixed Bayesian-likelihood
approach can be a good option – we are willing to use the legacy data to plan
the sampling, but not to make statements about the current population.
No closed formula for computing the required sample size exists for this ap-
proach because the posterior density function 𝑓(𝜃|𝑧, 𝑛) is not a well-defined
distribution as before. However, the required sample size still can be approxi-
mated by simulation.

12.5.3 Estimation of population mean

The three criteria (ALC, ACC, and WOC) described above are now used to
compute the required sample size for estimating the population mean, assuming
that the data come from a normal distribution. As we are uncertain about
the population standard deviation 𝜎 (𝑆∗ (𝑧) in Equation (12.7) is only a prior
point estimate of 𝜎), a prior distribution is assigned to this parameter. It
is convenient to assign a gamma distribution as a prior distribution to the
reciprocal of the population variance, referred to as the precision parameter
𝜆 = 1/𝜎2 . More precisely, a prior bivariate normal-gamma distribution is
224 12 Computing the required sample size

assigned to the population mean and the precision parameter3 . With this prior
distribution, the posterior distribution of the population mean is fully defined,
i.e., both the type of distribution and its parameters are known. The prior
distribution is so-called conjugate with the normal distribution.
The gamma distribution has two parameters: 𝑎 and 𝑏. Figure 12.2 shows the
gamma distribution for 𝑎 = 5 and 𝑏 = 100.

a <- 5; b <- 100

x <- seq(from = 0, to = 0.2, length = 1000)
dg <- dgamma(x = x, shape = a, scale = 1 / b)
plot(x = x, y = dg, type = "l", ylab = "Density", xlab = "Precision")

FIGURE 12.2: Prior gamma distribution for the precision parameter for a
shape parameter 𝑎 = 5 and a scale parameter 1/𝑏 = 1/100.

The mean of the precision parameter 𝜆 is given by 𝑎/𝑏 and its standard
deviation by √𝑎/𝑏2 .
The normal-gamma prior is used to compute the predictive distribution for the
data. For ACC the required sample size can then be computed with (Adcock,
1988)

4𝑏 2
𝑛= 𝑡 − 𝑛0 , (12.21)
𝑎 𝑙2max 2u�;u�/2

with 𝑡22u�;u�/2 the squared (1 − 𝛼/2) quantile of the (usual, i.e., neither shifted
nor scaled) t distribution with 2𝑎 degrees of freedom and 𝑛0 the number
of prior points. The prior sample size 𝑛0 is only relevant if we have prior
3 This is equal to a normal-inverse gamma distribution to the population mean and

population variance.
12.5 Bayesian sample size determination 225

information about the population mean and an informative prior is used for
this population mean. If we have no information about the population mean,
a non-informative prior is used and 𝑛0 equals 0. Note that as 𝑎/𝑏 is the prior
mean of the inverse of the population variance, Equation (12.21) is similar
to Equation (12.7). The only difference is that a quantile from the standard
normal distribution is replaced by a quantile from a t distribution with 2𝑎
degrees of freedom.
No closed-form formula exists for computing the smallest 𝑛 satisfying ALC,
but the solution can be found by a bisectional search algorithm (Joseph and
Bélisle, 1997).
Package SampleSizeMeans (Joseph and Bélisle, 2012) is used to compute
Bayesian required sample sizes, using both criteria, ACC and ALC, for the fully
Bayesian and the mixed Bayesian-likelihood approach. The gamma distribution
plotted in Figure 12.2 is used as a prior distribution for the precision parameter
𝜆. As a reference, also the frequentist required sample size is computed.

library(SampleSizeMeans)
lmax <- 2
n_freq <- mu.freq(len = lmax, lambda = a / b, level = 0.95)
n_alc <- mu.alc(len = lmax, alpha = a, beta = b, n0 = 0, level = 0.95)
n_alcmbl <- mu.mblalc(len = lmax, alpha = a, beta = b, level = 0.95)
n_acc <- mu.acc(len = lmax, alpha = a, beta = b, n0 = 0, level = 0.95)
n_accmbl <- mu.mblacc(len = lmax, alpha = a, beta = b, level = 0.95)
n_woc <- mu.modwoc(
len = lmax, alpha = a, beta = b, n0 = 0, level = 0.95, worst.level = 0.95)
n_wocmbl <- mu.mblmodwoc(
len = lmax, alpha = a, beta = b, level = 0.95, worst.level = 0.95)

Table 12.1 shows that all six required sample sizes are larger than the frequentist
required sample size. This makes sense, as the frequentist approach does not
account for uncertainty in the population variance parameter. The mixed
approach leads to slightly larger required sample sizes than the fully Bayesian

TABLE 12.1: Required sample sizes for estimating a normal mean, computed
with three criteria for the fully Bayesian and the mixed Bayesian-likelihood
(MBL) approach.

Freq ALC ALC-MBL ACC ACC-MBL WOC WOC-MBL

77 92 93 100 102 194 201
Freq: required sample size computed with the frequentist
approach; ALC: average length criterion; ACC: average coverage
criterion; WOC: worst outcome criterion.
226 12 Computing the required sample size

approach. This is because in the mixed approach the prior distribution of the
precision parameter is not used. Apparently, we do not lose much information
by ignoring this prior. With WOC the required sample sizes are about twice
the sample sizes obtained with the other two criteria, but this depends of
course on the size of the subspace 𝒲. If, for instance the 80% most likely data
sets are chosen as subspace 𝒲, the required sample sizes are much smaller.

n_woc80 <- mu.modwoc(

len = lmax, alpha = a, beta = b, n0 = 0, level = 0.95, worst.level = 0.80)
n_wocmbl80 <- mu.mblmodwoc(
len = lmax, alpha = a, beta = b, level = 0.95, worst.level = 0.80)

The required sample sizes with this criterion are 124 and 128 using the fully
Bayesian and the mixed Bayesian-likelihood approach, respectively.

12.5.4 Estimation of a population proportion

The same criteria can be used to estimate the proportion of a population or,
in case of an infinite population, the areal fraction, satisfying some condition
(Joseph et al., 1995). With simple random sampling, this boils down to esti-
mating the probability-of-success parameter 𝑝 of a binomial distribution. In
this case the space of possible outcomes 𝒵 is the number of successes, which is
discrete: 𝒵 = {0, 1, … , 𝑛} with 𝑛 the sample size.
The conjugate prior for the binomial likelihood is the beta distribution:

1
𝑝∼ 𝜋u�−1 (1 − 𝜋)u�−1 , (12.22)
𝐵(𝑐, 𝑑)

where 𝐵(𝑐, 𝑑) is the beta function. The two parameters 𝑐 and 𝑑 correspond
to the number of ‘successes’ and ‘failures’ in the problem context. The larger
these numbers, the more the prior information, and the more sharply defined
the probability distribution. The plot below shows this distribution for 𝑐 = 0.6
and 𝑑 = 2.4.

c <- 0.6; d <- 2.4

x <- seq(from = 0, to = 1, length = 1000)
dbt <- dbeta(x = x, shape1 = c, shape2 = d)
plot(x = x, y = dbt, type = "l", ylab = "Density", xlab = "Proportion")

The mean of the binomial proportion equals 𝑐/(𝑐+𝑑) and its standard deviation
√𝑐𝑑/{(𝑐 + 𝑑 + 1)(𝑐 + 𝑑)2 }.
12.5 Bayesian sample size determination 227

The preposterior marginal distribution of the data is the beta-binomial distri-

bution:
𝑛 𝐵(𝑧 + 𝑐, 𝑛 − 𝑧 + 𝑑)
𝑓(𝑧|𝑛) = ( ) , (12.23)
𝑧 𝐵(𝑐, 𝑑)

and for a given number of successes 𝑧 out of 𝑛 trials the posterior distribution
of 𝑝 equals

1
𝑓(𝑝|𝑧, 𝑛, 𝑐, 𝑑) = 𝑝u�+u�−1 (1 − 𝑝)u�−u�+u�−1 . (12.24)
𝐵(𝑧 + 𝑐, 𝑛 − 𝑧 + 𝑑)

For the binomial parameter, criterion ALC (Equation (12.17)) can be written
as u�
∑ 𝑙(𝑧, 𝑛)𝑓(𝑧, 𝑛) ≤ 𝑙max . (12.25)
u�=0

To compute the smallest 𝑛 satisfying this condition, for each value of 𝑧 and
each 𝑛, 𝑙(𝑧, 𝑛) must be computed so that

u�+u�(u�,u�)
∫ 𝑓(𝑝|𝑧, 𝑛, 𝑐, 𝑑)d𝑝 = 1 − 𝛼 , (12.26)
u�

with 𝑣 the lower bound of the HPD credible set given the sample size and the
observed number of successes 𝑧.
For the binomial parameter, criterion ACC (Equation (12.19)) can be written
as

u�
∑ Pr{𝑝 ∈ (𝑣, 𝑣 + 𝑙max )}𝑓(𝑧, 𝑛) ≥ 1 − 𝛼 , (12.27)
u�=0

with

u�+u�max
Pr{𝑝 ∈ (𝑣, 𝑣 + 𝑙max )} ∝ ∫ 𝑝u� (1 − 𝑝)u�−u� 𝑓(𝑝)d𝑝 , (12.28)
u�

with 𝑓(𝑝) the prior density of the binomial parameter.

For more details about ACC and ALC, and about how the required sample
size can be computed with WOC in case of the binomial parameter 𝑝, refer to
Joseph et al. (1995).
The required sample sizes for ALC, ACC, and WOC described in the previous
subsection, using the fully Bayesian approach or the mixed Bayesian-likelihood
approach, can be computed with package SampleSizeBinomial4 (Joseph
4 http://www.medicine.mcgill.ca/epidemiology/Joseph/software/Bayesian-Sample-Size.html
228 12 Computing the required sample size

FIGURE 12.3: Prior beta distribution for the binomial proportion for a beta
function 𝐵(0.6, 2.4).

et al., 2018). This package is used to compute the required sample sizes us-
ing the beta distribution shown in Figure 12.3 as a prior for the population
proportion. Note that argument len of the various functions of package Sam-
pleSizeBinomial specifies the total length of the confidence interval, not half
the length as passed to function ciss.wald using argument d.

library(SampleSizeBinomial)
n_alc <- prop.alc(
len = 0.2, alpha = c, beta = d, level = 0.95, exact = TRUE)$n
n_alcmbl <- prop.mblalc(
len = 0.2, alpha = c, beta = d, level = 0.95, exact = TRUE)$n
n_acc <- prop.acc(
len = 0.2, alpha = c, beta = d, level = 0.95, exact = TRUE)$n
n_accmbl <- prop.mblacc(
len = 0.2, alpha = c, beta = d, level = 0.95, exact = TRUE)$n
n_woc <- prop.modwoc(
len = 0.2, alpha = c, beta = d, level = 0.95, exact = TRUE,
worst.level = 0.80)$n
n_wocmbl <- prop.mblmodwoc(
len = 0.2, alpha = c, beta = d, level = 0.95, exact = TRUE,
worst.level = 0.80)$n
library(binomSamSize)
n_freq <- ciss.wald(p0 = c / (c + d), d = 0.1, alpha = 0.05)

The required sample sizes are shown in Table 12.2.

12.5 Bayesian sample size determination 229

TABLE 12.2: Required sample sizes for estimating a binomial proportion,

computed with three criteria for the fully Bayesian and the mixed Bayesian-
likelihood (MBL) approach.

Freq ALC ALC-MBL ACC ACC-MBL WOC WOC-MBL

62 33 38 50 53 80 81
Freq: required sample size computed with the frequentist
approach; ALC: average length criterion; ACC: average coverage
criterion; WOC: worst outcome criterion.
13
Model-based optimisation of probability
sampling designs

Chapter 10 on model-assisted estimation explained how a linear regression

model or a non-linear model obtained with a machine learning algorithm can
be used to increase the precision of design-based estimates of the population
mean or total using the data collected by a given probability sampling design.
This chapter will explain how a model of the study variable can be used at
an earlier stage to optimise probability sampling designs. It will show how
a model can help in choosing between alternative sampling design types, for
instance between systematic random sampling, spreading the sampling units
throughout the study area, and two-stage cluster random sampling, resulting
in spatial clusters of sampling units. Besides, the chapter will explain how to
use a model to optimise the sample size of a given sampling design type, for
instance, the number of primary and secondary sampling units with two-stage
cluster random sampling. The final section of this chapter is about how a model
can be used to optimise spatial strata for stratified simple random sampling.
The models used in this chapter are all geostatistical models of the spatial
variation. Chapter 21 is an introduction to geostatistical modelling. Several
geostatistical concepts explained in that chapter are needed here to predict
the sampling variance.
A general geostatistical model of the spatial variation is

𝑍(𝐬) = 𝜇(𝐬) + 𝜖(𝐬)

𝜖(𝐬) ∼ 𝒩(0, 𝜎2 ) (13.1)
Cov(𝜖(𝐬), 𝜖(𝐬′ )) = 𝐶(𝐡) ,

with 𝑍(𝐬) the study variable at location 𝐬, 𝜇(𝐬) the mean at location 𝐬, 𝜖(𝐬)
the residual at location 𝐬, and 𝐶(𝐡) the covariance of the residuals at two
locations separated by vector 𝐡 = 𝐬 − 𝐬′ . The residuals are assumed to have a
normal distribution with zero mean and a constant variance 𝜎2 (𝒩(0, 𝜎2 )).
The model of the spatial variation has several parameters. In case of a model in
which the mean is a linear combination of covariates, these are the regression
coefficients associated with the covariates and the parameters of a semivari-
ogram describing the spatial dependence of the residuals. A semivariogram is

231
232 13 Model-based optimisation of probability sampling designs

a model for half the expectation of the squared difference of the study variable
or the residuals of a model at two locations, referred to as the semivariance,
as a function of the length (and direction) of the vector separating the two
locations (Chapter 21).
Using the model to predict the sampling variance of a design-based estimator of
a population mean requires prior knowledge of the semivariogram. When data
from the study area of interest are available, these data can be used to choose
a semivariogram model and to estimate the parameters of the model. If no
such data are available, we must make a best guess, based on data collected in
other areas. In all cases I recommend keeping the model as simple as possible.

13.1 Model-based optimisation of sampling design type

and sample size
Chapter 12 presented methods and formulas for computing the required sample
size given various measures to quantify the quality of the survey result. These
required sample sizes are for simple random sampling. For other types of
sampling design, the required sample size can be approximated by multiplying
the required sample size for simple random sampling with the design effect, see
Section 12.4. An alternative is to use a model of the spatial variation to predict
the sampling variance of the estimator of the mean for the type of sampling
design under study and a range of sample sizes, plotting the predicted variance,
or standard error, against the sample size, and using this plot inversely to
derive the required sample size given a constraint on the sampling variance or
standard error.
The computed required sample size applies to several given parameters of the
sampling design. For instance, for stratified random sampling, the sample size
is computed for a given stratification and sample size allocation scheme, for
cluster random sampling for given clusters, and for two-stage cluster random
sampling for given primary sampling units (PSUs) and the number of secondary
sampling units (SSUs) selected per PSU draw. However, the model can also
be used to optimise these sampling design parameters. For stratified random
sampling, the optimal allocation can be computed by predicting the population
standard deviations within strata and using these predicted standard deviations
in Equation (4.17). Even the stratification can be optimised, see Section 13.2.
If we have a costs model for two-stage cluster random sampling, the number
of PSU draws and the number of SSUs per PSU draw can be optimised.
Model-based prediction of the sampling variance can also be useful to compare
alternative types of sampling design at equal total costs or equal variance of
the estimated population mean or total. For instance, to compare systematic
13.1 Model-based optimisation of sampling design type and sample size 233

random sampling, leading to good spatial coverage, and two-stage cluster

random sampling, resulting in spatial clusters of observations.
Three approaches for model-based prediction of the sampling variance of a
design-based estimator of the population mean, or total, are described: the
analytical approach (Subsection 13.1.1), the geostatistical simulation approach
(Subsection 13.1.2), and the Bayesian approach (Subsection 13.1.3). In the
analytical approach we assume that the mean, 𝜇(𝐬) in Equation (13.1), is
everywhere the same. This assumption is relaxed in the geostatistical simulation
approach. This simulation approach can also accommodate models in which
the mean is a linear combination of covariates and/or spatial coordinates to
predict the sampling variance. Furthermore, this simulation approach can
be used to predict the sampling variance of the estimator of the mean of a
trans-Gaussian variable, i.e., a random variable that can be transformed to a
Gaussian variable.
The predicted sampling variances of the estimated population mean obtained
with the analytical and the geostatistical simulation approach are conditional
on the model of the spatial variation. Uncertainty about this model is not
accounted for. On the contrary, in the Bayesian approach, we do account for
our uncertainty about the assumed model, and we analyse how this uncertainty
propagates to the sampling variance of the estimator of the mean.
In some publications, the sampling variance as predicted with a statistical
model is referred to as the anticipated variance (Särndal et al. (1992), p. 450).

13.1.1 Analytical approach

In the analytical approach, the sampling variance of the estimator of the

mean is derived from mean semivariances within the study area and mean
semivariances within the sample. Assuming isotropy, the mean semivariances
are a function of the separation distance between pairs of points.
The sampling variance of the 𝜋 estimator of the population mean can be
predicted by (Domburg et al. (1994), de Gruijter et al. (2006))

̄̂ = 𝛾̄ − 𝐸u� (𝜆
𝐸u� {𝑉u� (𝑧)} 𝜆TΓ u�𝜆 ) , (13.2)

where 𝐸u� (⋅) is the statistical expectation over realisations from the model
𝜉, 𝐸u� (⋅) is the statistical expectation over repeated sampling with sampling
design 𝑝, 𝑉u� (𝑧)̄̂ is the variance of the 𝜋 estimator of the population mean over
repeated sampling with sampling design 𝑝, 𝛾̄ is the mean semivariance of the
random variable at two randomly selected locations in the study area, 𝜆 is the
vector of design-based weights of the units of a sample selected with design 𝑝,
and Γ u� is the matrix of semivariances between the units of a sample 𝒮 selected
with design 𝑝.
234 13 Model-based optimisation of probability sampling designs

The mean semivariance 𝛾̄ is a model-based prediction of the population variance,

or spatial variance, i.e., the model-expectation of the population variance:

𝛾̄ = 𝐸u� {𝜎2 (𝑧)} . (13.3)

The mean semivariance 𝛾̄ can be calculated by discretising the study area by a

fine square grid, computing the matrix with geographical distances between
the discretisation nodes, transforming this into a semivariance matrix, and
computing the average of all elements of the semivariance matrix. The second
term 𝐸u� (𝜆
𝜆TΓ u�𝜆 ) can be evaluated by Monte Carlo simulation, repeatedly
selecting a sample according to design 𝑝, calculating 𝜆 TΓ u�𝜆 , and averaging.

The semivariance at zero distance (same location) is 0, so the diagonal of a

semivariance matrix is filled with zeroes. If a semivariogram with nugget
is assumed, the zeroes on the diagonal must be replaced by the nugget to
compute 𝛾.̄ The same holds for the diagonal zeroes in Γ u� .

This generic procedure is still computationally demanding, but it is the only

option for complex spatial sampling designs. For basic sampling designs, the
general formula can be worked out. For simple random sampling, the sampling
variance can be predicted by

̄̂ = 𝛾/𝑛
𝐸u� {𝑉SI (𝑧)} ̄ , (13.4)

and for stratified simple random sampling by

u�
̄̂ = ∑ 𝑤2ℎ 𝛾ℎ̄ /𝑛ℎ ,
𝐸u� {𝑉STSI (𝑧)} (13.5)
ℎ=1

with 𝐻 the number of strata, 𝑤ℎ the stratum weight (relative size), 𝛾ℎ

̄ the
mean semivariance of stratum ℎ, and 𝑛ℎ the number of sampling units of
stratum ℎ.

The model-based predictions of the variances within the strata, 𝛾ℎ

̄ , can also
be used to compute the sample sizes for Neyman allocation, which are the
optimal sample sizes when the mean costs per unit are equal for the strata.
To compute these sample sizes, the standard deviations 𝑆ℎ (𝑧) in Equation
√
(4.16) are replaced by 𝛾ℎ
̄ .

For systematic random sampling, i.e., sampling on a randomly placed grid, the
variance can be predicted by

̄̂ = 𝛾̄ − 𝐸SY (𝛾SY
𝐸u� {𝑉SY (𝑧)} ̄ ), (13.6)
13.1 Model-based optimisation of sampling design type and sample size 235

with 𝐸SY (𝛾SY

̄ ) the design-expectation, i.e., the expectation over repeated
systematic sampling, of the mean semivariance within the grid. With systematic
random sampling, the number of grid points within the study area can vary
among samples, as well as the spatial pattern of the points (Chapter 5).
Therefore, multiple systematic random samples must be selected, and the
average of the mean semivariance within the systematic sample must be
computed.
The analytical approach is illustrated with the data of agricultural field Leest
(Hofman and Brus, 2021). Nitrate-N (NO3 -N) in kg ha-1 in the layer 0 − 90 cm,
using a standard soil density of 1,500 kg m-3 , is measured at 30 points. In the
next code chunk, these data are used to compute a sample semivariogram using
the method-of-moments with function variogram, see Chapter 21. A spherical
model without nugget is fitted to the sample semivariogram using function
fit.variogram. The numbers in this plot are the numbers of pairs of points used
to compute the semivariances.

library(gstat)
coordinates(sampleLeest) <- ~ s1 + s2
vg <- variogram(N ~ 1, data = sampleLeest)
vgm_MoM <- fit.variogram(
vg, model = vgm(model = "Sph", psill = 2000, range = 20))

The few data lead to a very noisy sample semivariogram. For the moment, I
ignore my uncertainty about the semivariogram parameters; Subsection 13.1.3
will show how we can account for our uncertainty about the semivariogram
parameters in model-based prediction of the sampling variance. A spherical
semivariogram model without nugget is fitted to the sample semivariogram,
i.e., the intercept is 0. The fitted range of the model is 45 m, and the fitted sill
equals 966 (kg ha-1 )2 . The fitted semivariogram is used to predict the sampling
variance for three sampling designs: simple random sampling, stratified simple
random sampling, and systematic random sampling. The costs for these three
design types will be about equal, as the study area is small, so that the
access time of the sampling points selected with the three designs is about
equal. The sample size of the evaluated sampling designs is 25 points. As for
systematic random sampling, the number of points varies among the samples,
this sampling design has an expected sample size of 25 points.

Simple random sampling

For simple random sampling, we must compute the mean semivariance within
the field (Equation (13.4)). As shown in the next code chunk, the mean
semivariance is approximated by discretising the field by a square grid of 2,000
points, computing the 2,000 × 2,000 matrix with distances between all pairs
of discretisation nodes, transforming this distance matrix into a semivariance
236 13 Model-based optimisation of probability sampling designs

matrix using function variogramLine of package gstat (Pebesma, 2004), and

finally averaging the semivariances. Note that in this case we do not need to
replace the zeroes on the diagonal of the semivariance matrix by the nugget,
as a model without nugget is fitted. The geopackage file is read with function
read_sf of package sf (Pebesma, 2018), resulting in a simple feature object. The
projection attributes of this object are removed with function st_set_crs. The
centres of square grid cells are selected with function st_make_grid. The centres
inside the field are selected with mygrid[field], and finally the coordinates are
extracted with function st_coordinates.

field <- read_sf(system.file("extdata/leest.gpkg", package = "sswr")) %>%

st_set_crs(NA_crs_)
mygrid <- st_make_grid(field, cellsize = 2, what = "centers")
mygrid <- mygrid[field] %>%
st_coordinates(mygrid)
H <- as.matrix(dist(mygrid))
G <- variogramLine(vgm_MoM, dist_vector = H)
m_semivar_field <- mean(G)
n <- 25
Exi_V_SI <- m_semivar_field / n

The model-based prediction of the sampling variance of the estimator of the

mean with this design equals 35.0.

Stratified simple random sampling

The strata of the stratified simple random sampling design are compact geo-
graphical strata of equal size (Section 4.6). The number of geostrata is equal
to the sample size, 25 points, so that we have one point per stratum. With
this design, the sampling points are reasonably well spread over the field, but
not as good as with systematic random sampling. To predict the sampling
variance, we must compute the mean semivariances within the geostrata, see
Equation (13.5). Note that the stratum weights are constant as the strata have
equal size, 𝑤ℎ = 1/𝑛, and that 𝑛ℎ = 1. Therefore, Equation (13.5) reduces to

1 u�
̄̂ =
𝐸u� {𝑉STSI (𝑧)} ∑𝛾 ̄ . (13.7)
𝑛2 ℎ=1 ℎ

The next code chunk shows the computation of the mean semivariance per
stratum and the model-based prediction of the sampling variance of the
estimator of the mean. The matrix with the coordinates is first converted to a
tibble with function as_tibble.
13.1 Model-based optimisation of sampling design type and sample size 237

library(spcosa)
mygrid <- mygrid %>%
as_tibble() %>%
setNames(c("x1", "x2"))
gridded(mygrid) <- ~ x1 + x2
mygeostrata <- stratify(mygrid, nStrata = n, equalArea = TRUE, nTry = 10) %>%
as("data.frame")
m_semivar_geostrata <- numeric(length = n)
for (i in 1:n) {
ids <- which(mygeostrata$stratumId == (i - 1))
mysubgrd <- mygeostrata[ids, ]
H_geostratum <- as.matrix(dist(mysubgrd[, c(2, 3)]))
G_geostratum <- variogramLine(vgm_MoM, dist_vector = H_geostratum)
m_semivar_geostrata[i] <- mean(G_geostratum)
}
Exi_V_STSI <- sum(m_semivar_geostrata) / nˆ2

The model-based prediction of the sampling variance with this design equals
13.5 (kg ha-1 )2 , which is much smaller than with simple random sampling.
The large stratification effect can be explained by the assumed strong spatial
structure of NO3 -N in the agricultural field and the improved geographical
spreading of the sampling points, see Figure 13.1.

FIGURE 13.1: Sample semivariogram and fitted spherical model for NO3 -N
in field Leest. The numbers refer to point-pairs used in computing semivari-
ances.
238 13 Model-based optimisation of probability sampling designs

Systematic random sampling

To predict the sampling variance for systematic random sampling with an

expected sample size of 25 points, we must compute the design-expectation
of the mean semivariance within the systematic sample (Equation (13.6)). As
shown in the next code chunk, I approximated this expectation by selecting 100
systematic random samples, computing the mean semivariance for each sample,
and averaging. Finally, the model-based prediction of the sampling variance
is computed by subtracting the average of the mean semivariances within
a systematic sample from the mean semivariance within the field computed
above.

set.seed(314)
m_semivar_SY <- numeric(length = 100)
for (i in 1:100) {
mySYsample <- spsample(x = mygrid, n = n, type = "regular") %>%
as("data.frame")
H_SY <- as.matrix(dist(mySYsample))
G_SY <- variogramLine(vgm_MoM, dist_vector = H_SY)
m_semivar_SY[i] <- mean(G_SY)
}
Exi_V_SY <- m_semivar_field - mean(m_semivar_SY)

The model-based prediction of the sampling variance of the estimator of the

mean with this design equals 8.3 (kg ha-1 )2 , which is smaller than that of
stratified simple random sampling. This can be explained by the improved
geographical spreading of the sampling points with systematic random sampling
as compared to stratified simple random sampling with compact geographical
strata.

13.1.1.1 Bulking soil aliquots into a composite sample

If the soil aliquots collected at the points of the stratified random sample
are bulked into a composite, as is usually done in soil testing of agricultural
fields, the procedure for predicting the variance of the estimator of the mean
is slightly different. Only the composite sample is analysed in a laboratory on
NO3 -N, not the individual soil aliquots. This implies that the contribution of
the measurement error to the total uncertainty about the population mean
is larger. To predict the sampling variance in this situation, we need the
semivariogram of errorless measurements of NO3 -N, i.e., of the true NO3 -N
contents of soil aliquots collected at points. The sill of this semivariogram will
be smaller than the sill of the semivariogram of measured NO3 -N data. A
simple option is to subtract an estimate of the measurement error variance
from the semivariogram of measured NO3 -N data that contain a measurement
error. So, the measurement error variance is subtracted from the nugget. This
13.1 Model-based optimisation of sampling design type and sample size 239

may lead to negative nuggets, which is not allowed (a variance cannot be

negative). The preferable alternative is to add the measurement error variance
to the diagonal of the covariance matrix of the data in fitting the model with
maximum likelihood, see function ll in Subsection 13.1.3.

Exercises

1. Write an R script to predict the sampling variance of the estimator

of the mean of NO3 -N of agricultural field Leest, for simple random
sampling and a sample size of 25 points. Use in prediction a
spherical semivariogram with a nugget of 483, a partial sill of 483,
and a range of 44.6 m. The sum of the nugget and partial sill (966)
is equal to the sill of the semivariogram used above in predicting
sampling variances. Compare the predicted sampling variance with
the predicted sampling variance for the same sampling design,
obtained with the semivariogram without nugget. Explain the
difference.

2. Write an R script to compute the required sample size for simple

random sampling of agricultural field Leest, for a maximum length
of a 95% confidence interval of 20. Use the semivariogram without
nugget in predicting the sampling variance. See Section 12.2
(Equation (12.7)) for how to compute the required sample size given
a prior estimate of the standard deviation of the study variable in
the population.

3. Do the same for systematic random sampling. Note that for this
sampling design, no such formula is available. Predict for a series
of expected sample sizes, 𝑛 = 5, 6, … , 40, the sampling variance of
the estimator of the mean, using Equation (13.6). Approximate
̄ ) from ten repeated selections. Compute the length of
𝐸SY (𝛾SY
the confidence interval from the predicted sampling variances, and
plot the interval length against the sample size. Finally, determine
the required sample size for a maximum length of 20. What is the
design effect for an expected sample size of 34 points (the required
sample size for simple random sampling)? See Equation (12.13). Also
compute the design effect for expected sample sizes of 5, 6, … , 40.
Explain why the design effect is not constant.

13.1.2 Geostatistical simulation approach

The alternative to the analytical approach is to use a geostatistical simulation

approach. It is computationally more demanding, but an advantage of this
approach is its flexibility. It can also be used to predict the sampling variance
240 13 Model-based optimisation of probability sampling designs

of the estimator of the mean using a geostatistical model with a non-constant

mean. And besides, this approach can also handle trans-Gaussian variables,
i.e., variables whose distribution can be transformed into a normal distribution.
In Subsection 13.1.3, geostatistical simulation is used to predict the variance
of the estimator of the mean of a lognormal variable.
The geostatistical simulation approach for predicting the sampling variance of
a design-based estimator of the population mean involves the following steps:

1. Select a large number 𝑆 of random samples with sampling design 𝑝.

2. Use the model to simulate values of the study variable for all
sampling points.

3. Estimate for each sample the population mean, using the design-
based estimator of the population mean for sampling design 𝑝. This
results in 𝑆 estimated population means.

4. Compute the variance of the 𝑆 estimated means.

5. Repeat steps 1 to 4 𝑅 times, and compute the mean of the 𝑅

variances.

This approach is illustrated with the western part of the Amhara region in
Ethiopia (hereafter referred to as West-Amhara) where a large sample is
available with organic matter data in the topsoil (SOM) in decagram per kg
dry soil (dag kg-1 ; 1 decagram = 10 gram). The soil samples are collected
along roads (see Figure 17.5). It is a convenience sample, not a probability
sample, so these sample data cannot be used in design-based or model-assisted
estimation of the mean or total soil carbon stock in the study area. However,
the data can be used to model the spatial variation of the SOM concentration,
and this geostatistical model can then be used to design a probability sample
for design-based estimation of the total mass of SOM. Apart from the point
data of the SOM concentration, maps of covariates are available, such as
a digital elevation model and remote sensing reflectance data. In the next
code chunk, four covariates are selected to model the mean of the SOM
concentration: elevation (dem), average near infrared reflectance (rfl-NIR),
average red reflectance (rfl-red), and average land surface temperature (lst). I
assume a normal distribution for the residuals of the linear model. The model
parameters are estimated by restricted maximum likelihood (REML), using
package geoR (Ribeiro Jr et al., 2020), see Subsection 21.5.2 for details on
REML estimation of a geostatistical model. As a first step, the projected
coordinates of the sampling points are changed from m into km using function
mutate. Using coordinates in m in function likfit could not find an optimal
estimate for the range.
13.1 Model-based optimisation of sampling design type and sample size 241

library(geoR)
sampleAmhara <- sampleAmhara %>%
mutate(s1 = s1 / 1000, s2 = s2 / 1000)
dGeoR <- as.geodata(obj = sampleAmhara, header = TRUE,
coords.col = c("s1", "s2"), data.col = "SOM",
covar.col = c("dem", "rfl_NIR", "rfl_red", "lst"))
vgm_REML <- likfit(geodata = dGeoR,
trend = ~ dem + rfl_NIR + rfl_red + lst,
cov.model = "spherical", ini.cov.pars = c(1, 50), nugget = 0.2,
lik.method = "REML", messages = FALSE)

The estimated parameters of the residual semivariogram of the SOM concen-

tration are shown in Table 22.2. The estimated regression coefficients are 12.9
for the intercept, 0.922 for elevation (dem), 7.41 for NIR reflectance, -10.42 for
red reflectance, and -0.039 for land surface temperature.
Package gstat is used for geostatistical simulation, and therefore first the
REML estimates of the semivariogram parameters are passed to function vgm
using arguments nugget, psill, and range of function vgm of this package.

vgm_REML_gstat <- vgm(model = "Sph", nugget = vgm_REML$tausq,

psill = vgm_REML$sigmasq, range = vgm_REML$phi)

The fitted model of the spatial variation of the SOM concentration is used to
compare systematic random sampling and two-stage cluster random sampling
at equal variances of the estimator of the mean.

Systematic random sampling

One hundred systematic random samples (𝑆 = 100) with an expected sam-

ple size of 50 points (𝐸[𝑛] = 50) are selected. The four covariates at the
selected sampling points are extracted by overlaying the SpatialPointsDataFrame
mySYsamples and the SpatialPixelsDataFrame grdAmhara with function over of pack-
age sp (Pebesma and Bivand, 2005). Values at the sampling points are sim-
ulated by sequential Gaussian simulation (Goovaerts, 1997), using function
krige with argument nsim = 1 of package gstat. Argument dummy is set to TRUE
to enforce unconditional simulation.

The alternative is conditional simulation, using the data of the convenience

sample as conditioning data. Conditional simulation is only recommended
if the quality of these legacy data is sufficient, and we may trust that the
study variable at the legacy points has not changed since these legacy data
have been collected.
242 13 Model-based optimisation of probability sampling designs

Note that by first drawing 100 samples, followed by simulating values of 𝑧 at

the selected sampling points, instead of first simulating values of 𝑧 at the nodes
of a discretisation grid, followed by selecting samples and overlaying with the
simulated field, the simulated values of points in the same discretisation cell
differ, so that we account for the infinite number of points in the population.
With systematic random sampling, the sample mean is an approximately
unbiased estimator of the population mean (Chapter 5). Therefore, of each
sample the mean of the simulated values is computed, using function tapply.
Finally, the variance of the 100 sample means is computed. This is a conditional
variance, conditional on the simulated values. In the code chunk below, the
whole procedure is repeated 100 times (𝑅 = 100), leading to 100 conditional
variances of sample means.

grdAmhara <- grdAmhara %>%

mutate(s1 = s1 / 1000, s2 = s2 / 1000)
gridded(grdAmhara) <- ~ s1 + s2
S <- R <- 100
v_mzsim_SY <- numeric(length = R)
set.seed(314)
for (i in 1:R) {
mySYsamples <- NULL
for (j in 1:S) {
xy <- spsample(x = grdAmhara, n = 50, type = "regular")
mySY <- data.frame(
s1 = xy$x1, s2 = xy$x2, sample = rep(j, length(xy)))
mySYsamples <- rbind(mySYsamples, mySY)
}
coordinates(mySYsamples) <- ~ s1 + s2
res <- over(mySYsamples, grdAmhara)
mySYs <- data.frame(mySYsamples, res[, c("dem", "rfl_NIR", "rfl_red", "lst")])
coordinates(mySYs) <- ~ s1 + s2
zsim <- krige(
dummy ~ dem + rfl_NIR + rfl_red + lst,
locations = mySYs, newdata = mySYs,
model = vgm_REML_gstat, beta = vgm_REML$beta,
nmax = 20, nsim = 1,
dummy = TRUE,
debug.level = 0) %>% as("data.frame")
m_zsim <- tapply(zsim$sim1, INDEX = mySYs$sample, FUN = mean)
v_mzsim_SY[i] <- var(m_zsim)
}
grdAmhara <- as_tibble(grdAmhara)
13.1 Model-based optimisation of sampling design type and sample size 243

Exi_vmz_SY <- mean(v_mzsim_SY)

The mean of the 100 conditional variances equals 0.015 (dag kg-1 )2 . This is
a Monte Carlo approximation of the model-based prediction of the sampling
variance of the ratio estimator of the mean for systematic random sampling
with an expected sample size of 50.

Two-stage cluster random sampling

Due to the geographical spreading of the sampling points with systematic

random sampling, the accuracy of the estimated mean is expected to be high
compared to that of other sampling designs of the same size. However, with
large areas, the time needed for travelling to the sampling points can become
substantial, lowering the sampling efficiency. With large areas, sampling designs
leading to spatial clusters of sampling points can be an attractive alternative.
One option then is two-stage cluster random sampling, see Chapter 7. The
question is whether this alternative design is more efficient than systematic
random sampling.
In the next code chunk, 100 compact geostrata (see Section 4.6) are computed
for West-Amhara. Here, these geostrata are not used as strata in stratified
random sampling, but as PSUs in two-stage cluster random sampling. The
difference is that in stratified random sampling from each geostratum at
least one sampling unit is selected, whereas in two-stage cluster random
sampling only a randomly selected subset of the geostrata is sampled. The
compact geostrata, used as PSUs, are computed with function kmeans, and as a
consequence the PSUs do not have equal size, see output of code chunk below.
This is not needed in two-stage cluster random sampling, see Chapter 7. If
PSUs of equal size are preferred, then these can be computed with function
stratify of package spcosa with argument equalArea = TRUE, see Section 4.6.

set.seed(314)
res <- kmeans(
grdAmhara[, c("s1", "s2")], iter.max = 1000, centers = 100, nstart = 100)
mypsus <- res$cluster
psusize <- as.numeric(table(mypsus))
summary(psusize)

Min. 1st Qu. Median Mean 3rd Qu. Max.

79.0 103.8 109.0 108.4 113.0 131.0

In the next code chunks, I assume that the PSUs are selected with probabilities
proportional to their size and with replacement (ppswr sampling), see Chapter
7. In Section 7.1 formulas are presented for computing the optimal number
of PSU draws and SSU draws per PSU draw. The optimal sample sizes are a
244 13 Model-based optimisation of probability sampling designs

function of the pooled variance of PSU means, 𝑆2b , and the pooled variance
of secondary units (points) within the PSUs, 𝑆2w . In the current subsection,
these variance components are predicted with the geostatistical model.
As a first step, a large number of maps are simulated.

grdAmhara$psu <- mypsus

coordinates(grdAmhara) <- ~ s1 + s2
set.seed(314)
zsim <- krige(
dummy ~ dem + rfl_NIR + rfl_red + lst,
locations = grdAmhara, newdata = grdAmhara,
model = vgm_REML_gstat, beta = vgm_REML$beta,
nmax = 20, nsim = 1000,
dummy = TRUE, debug.level = 0) %>% as("data.frame")
zsim <- zsim[, -c(1, 2)]

For each simulated field, the means of the PSUs and the variances of the
simulated values within the PSUs are computed using function tapply in
function apply.

m_zsim_psu <- apply(zsim, MARGIN = 2, FUN = function(x)

tapply(x, INDEX = grdAmhara$psu, FUN = mean))
v_zsim_psu <- apply(zsim, MARGIN = 2, FUN = function(x)
tapply(x, INDEX = grdAmhara$psu, FUN = var))

Next, for each simulated field, the pooled variance of PSU means and the
pooled variance within PSUs are computed, and finally these pooled variances
are averaged over all simulated fields. The averages are approximations of
the model-expectations of the pooled between unit and within unit variances,
𝐸u� [𝑆2b ] and 𝐸u� [𝑆2w ].

p_psu <- psusize / sum(psusize)

S2b <- apply(m_zsim_psu, MARGIN = 2, FUN = function(x)
sum(p_psu * (x - sum(p_psu * x))ˆ2))
S2w <- apply(v_zsim_psu, MARGIN = 2, FUN = function(x)
sum(p_psu * x))
Exi_S2b <- mean(S2b)
Exi_S2w <- mean(S2w)

The optimal sample sizes are computed for a simple linear costs model: 𝐶 =
𝑐0 + 𝑐1 𝑛 + 𝑐2 𝑛𝑚, with 𝑐0 the fixed costs, 𝑐1 the access costs per PSU, including
the access costs of the SSUs (points) within a given PSU, and 𝑐2 the observation
costs per SSU. In the next code chunk, I use 𝑐1 = 2 and 𝑐2 = 1. For the optimal
sample sizes only the ratio of 𝑐1 and 𝑐2 is important, not their absolute values.
13.1 Model-based optimisation of sampling design type and sample size 245

Given values for 𝑐1 and 𝑐2 , the optimal number of PSU draws 𝑛 and the
optimal number of SSU draws per PSU draw 𝑚 are computed, required for a
sampling variance of the estimator of the mean equal to the sampling variance
with systematic random sampling of 50 points, see Equations (7.9) and (7.10).

c1 <- 2; c2 <- 1
nopt <- 1 / Exi_vmz_SY * (sqrt(Exi_S2w * Exi_S2b) * sqrt(c2 / c1) + Exi_S2b)
mopt <- sqrt(Exi_S2w / Exi_S2b) * sqrt(c1 / c2)

The optimal number of PSU draws is 26, and the optimal number of points
per PSU draw equals 5. The total number of sampling points is 26 × 5 =
130. This is much larger than the sample size of 50 obtained with systematic
random sampling. The total observation costs therefore are substantially larger.
However, the access time can be substantially smaller due to the spatial
clustering of sampling points. To answer the question of whether the costs
saved by this reduced access time outweigh the extra costs of observation, the
model for the access costs and observation costs must be further developed.

13.1.3 Bayesian approach

The model-based prediction of the variance of the design-based estimator

of the population mean for a given sampling design is conditional on the
model. If we change the model type or the model parameters, the predicted
sampling variance also changes. In most situations we are quite uncertain
about the model, even in situations where we have data that can be used to
estimate the model parameters, as in the West-Amhara case study. Instead
of using the best estimated model to predict the sampling variance as done
in the previous sections, we may prefer to account for the uncertainty about
the model parameters. This can be done through a Bayesian approach, in
which the legacy data are used to update a prior distribution of the model
parameters to a posterior distribution. For details about a Bayesian approach
for estimating model parameters, see Section 22.5. A sample from the posterior
distribution of the model parameters is used one-by-one to predict the sampling
variance. This can be done either analytically, as described in Subsection
13.1.1, or through geostatistical simulation, as described in Subsection 13.1.2.
Both approaches result in a distribution of sampling variances, reflecting our
uncertainty about the sampling variance of the estimator of the population
mean due to uncertainty about the model parameters. The mean or median of
the distribution of sampling variances can be used as the predicted sampling
variance.
The Bayesian approach is illustrated with a case study on predicting the
sampling variance of NO3 -N in agricultural field Melle in Belgium (Hofman
and Brus, 2021). As for field Leest used in Subsection 13.1.1, data of NO3 -N
are available at 30 points. The sampling points are approximately on the nodes
246 13 Model-based optimisation of probability sampling designs

of a square grid with a spacing of about 4.5 m. As a first step, I check whether
we can safely assume that the data come from a normal distribution.

ggplot(sampleMelle, aes(sample = N)) +

geom_qq() +
geom_qq_line()

FIGURE 13.2: Q-Q plot of NO3 -N of field Melle.

pvalue <- shapiro.test(sampleMelle$N)$p.value

The Q-Q plot (Figure 13.2) shows that a normal distribution is not very likely:
there are too many large values, i.e., the distribution is skewed to the right.
Also the p-value of the Shapiro-Wilk test shows that we should reject the
null hypothesis of a normal distribution for the data: 𝑝 = 0.0028. I therefore
proceed with the natural log of NO3 -N, in short lnN.

sampleMelle$lnN <- log(sampleMelle$N)

As a first step, the semivariogram of lnN is estimated by maximum likelihood

(Subsection 21.4.2). An exponential semivariogram model is assumed, see
Equation (21.13).

The parameters that are estimated are the reciprocal of the sill 𝜆, the ratio
of spatial dependence 𝜉, defined as the partial sill divided by the sill, and the
distance parameter 𝜙. This parameterisation of the semivariogram is chosen
13.1 Model-based optimisation of sampling design type and sample size 247

because hereafter in the Bayesian approach prior distributions are chosen for
these parameters.

The likelihood function is defined, using a somewhat unusual parameterisation,

tailored to the Markov chain Monte Carlo (MCMC) sampling from the posterior
distribution of the semivariogram parameters. In MCMC a Markov chain of
sampling units (vectors with semivariogram parameters) is generated using the
previous sampling unit to randomly generate the next sampling unit (Gelman
et al. (2013), chapter 11). In MCMC sampling, the probability of accepting a
proposed sampling unit 𝜃 ∗ is a function of the ratio of the posterior density of the
proposed sampling unit and that of the current sampling unit, 𝑓(𝜃𝜃∗ |𝐳)/𝑓(𝜃𝜃u�−1 |𝐳),
so that the normalising constant, the denominator of Equation (12.15), cancels.

library(mvtnorm)
ll <- function(thetas) {
sill <- 1 / thetas[1]
psill <- thetas[2] * sill
nugget <- sill - psill
vgmodel <- vgm(
model = model, psill = psill, range = thetas[3], nugget = nugget)
C <- variogramLine(vgmodel, dist_vector = D, covariance = TRUE)
XCX <- crossprod(X, solve(C, X))
XCz <- crossprod(X, solve(C, z))
betaGLS <- solve(XCX, XCz)
mu <- as.numeric(X %*% betaGLS)
logLik <- dmvnorm(x = z, mean = mu, sigma = C, log = TRUE)
logLik
}

Next, initial estimates of the semivariogram parameters are computed by

maximising the likelihood, using function optim.

lambda.ini <- 1 / var(sampleMelle$lnN)

xi.ini <- 0.5
phi.ini <- 20
pars <- c(lambda.ini, xi.ini, phi.ini)
D <- as.matrix(dist(sampleMelle[, c("s1", "s2")]))
X <- matrix(1, nrow(sampleMelle), 1)
z <- sampleMelle$lnN
model <- "Exp"
vgML <- optim(pars, ll, control = list(fnscale = -1),
lower = c(1e-6, 0, 1e-6), upper = c(1000, 1, 150), method = "L-BFGS-B")
248 13 Model-based optimisation of probability sampling designs

The maximum likelihood (ML) estimates of the semivariogram parameters

are used as initial values in MCMC sampling. A uniform prior is used for the
inverse of the sill parameter, 𝜆 = 1/𝜎2 , with a lower bound of 10−6 and an
upper bound of 1. For the relative nugget, 𝜏2 /𝜎2 , a uniform prior is assumed
with a lower bound of 0 and an upper bound of 1. For the distance parameter
𝜙 of the exponential semivariogram a uniform prior is assumed, with a lower
bound of 10−6 m and an upper bound of 150 m.
These priors can be defined by function createUniformPrior of package
BayesianTools (Hartig et al., 2019). Function createBayesianSetup is then
used to define the setup of the MCMC sampling, specifying the likelihood
function, the prior, and the vector with best prior estimates of the model
parameters, passed to function createBayesianSetup using argument best. Argu-
ment sampler of function runMCMC specifies the type of MCMC sampler. I used
the differential evolution algorithm of ter Braak and Vrugt (2008). Argument
start of function getSample specifies the burn-in period, i.e., the number of first
samples that are discarded to diminish the influence of the initial semivari-
ogram parameter values. Argument numSamples specifies the sample size, i.e.,
the number of saved vectors with semivariogram parameter values, drawn from
the posterior distribution.

library(BayesianTools)
priors <- createUniformPrior(lower = c(1e-6, 0, 1e-6),
upper = c(1000, 1, 150))
bestML <- c(vgML$par[1], vgML$par[2], vgML$par[3])
setup <- createBayesianSetup(likelihood = ll, prior = priors,
best = bestML, names = c("lambda", "xi", "phi"))
set.seed(314)
res <- runMCMC(setup, sampler = "DEzs")
MCMCsample <- getSample(res, start = 1000, numSamples = 1000) %>% data.frame()

Figure 13.3 shows several semivariograms, sampled by MCMC from the poste-
rior distribution of the estimated semivariogram parameters.
The evaluated sampling design is the same as used in Subsection 13.1.1 for
field Leest: stratified simple random sampling, using compact geographical
strata of equal size, a total sample size of 25 points, and one point per stratum.
The next step is to simulate with each of the sampled semivariograms a
large number of maps of lnN. This is done by sequential Gaussian simulation,
conditional on the available data. The simulated values are backtransformed.
Each simulated map is then used to compute the variance of the simulated
values within the geostrata 𝑆2ℎ . These stratum variances are used to compute
the sampling variance of the estimator of the mean. Plugging 𝑤ℎ = 1/𝑛 (all
strata have equal size) into Equation (4.4) and using 𝑛ℎ = 1 in Equation (4.5)
13.1 Model-based optimisation of sampling design type and sample size 249

FIGURE 13.3: Semivariograms of the natural log of NO3 -N for field Melle
obtained by MCMC sampling from posterior distribution of the estimated
semivariogram parameters.

yield (compare with Equation (13.7))

1 u� 2
𝑉(𝑧)̄̂ = ∑𝑆 . (13.8)
𝑛2 ℎ=1 ℎ

In the code chunk below, I use the first 100 sampled semivariograms to simulate
with each semivariogram 100 maps.

V <- matrix(data = NA, nrow = 100, ncol = 100)

coordinates(sampleMelle) <- ~ s1 + s2
set.seed(314)
for (i in 1:100) {
sill <- 1 / MCMCsample$lambda[i]
psill <- MCMCsample$xi[i] * sill
nug <- sill - psill
range <- MCMCsample$phi[i]
vgmdl <- vgm(model = "Exp", nugget = nug, psill = psill, range = range)
ysim <- krige(
lnN ~ 1, locations = sampleMelle, newdata = mygrid,
model = vgmdl,
nmax = 20, nsim = 100,
debug.level = 0) %>% as("data.frame")
zsim <- exp(ysim[, -c(1, 2)])
250 13 Model-based optimisation of probability sampling designs

S2h <- apply(zsim, MARGIN = 2, FUN = function(x)

tapply(x, INDEX = as.factor(mygeostrata$stratumId), FUN = var))
V[i, ] <- 1 / nˆ2 * apply(S2h, MARGIN = 2, FUN = sum)
}

Figure 13.4 shows 16 maps simulated with the first four semivariograms. The
four maps in a row (a to d) are simulated with the same semivariogram. All maps
show that the simulated data have positive skew, which is in agreement with
the prior data. The data obtained by simulating from a lognormal distribution
are always strictly positive. This is not guaranteed when simulating from a
normal distribution.
The sampling variances of the estimated mean of NO3 -N obtained with these
16 maps are shown below.

a b c d
1 1.364 0.831 0.878 0.847
2 1.379 1.151 0.991 1.162
3 0.669 0.594 0.522 0.530
4 0.932 1.949 0.878 0.739

The sampling variance shows quite strong variation among the maps. The
frequency distribution of Figure 13.5 shows our uncertainty about the sampling
variance, due to uncertainty about the semivariogram, and about the spatial
distribution of NO3 -N within the agricultural field given the semivariogram
and the available data from that field.
As a model-based prediction of the sampling variance, we can take the mean or
the median of the sampling variances over all 100 × 100 simulated maps, which
are equal to 0.728 (dag kg-1 ) and 0.666 (dag kg-1 ), respectively. If we want to
be more safe, we can take a high quantile, e.g., the P90 of this distribution as
the predicted sampling variance, which is equal to 1.100 (dag kg-1 ).
I used the 30 available NO3 -N data as conditioning data in geostatistical
simulation. Unconditional simulation is recommended if we cannot rely on the
quality of the legacy data, for instance due to a temporal change in lnN since
the time the legacy data have been observed. For NO3 -N this might well be
the case. I believe that, although the effect of 30 observations on the simulated
fields and on the uncertainty distribution of the sampling variance will be
very small, one still may prefer unconditional simulation. With unconditional
simulation, we must assign the model-mean 𝜇 to argument beta of function
krige. The estimated model-mean can be estimated by generalised least squares,
see function ll above.
13.1 Model-based optimisation of sampling design type and sample size 251

FIGURE 13.4: Maps of NO3 -N of field Melle simulated with four semivari-
ograms (rows). Each semivariogram is used to simulate four maps (columns
a-d).
252 13 Model-based optimisation of probability sampling designs

FIGURE 13.5: Frequency distribution of simulated sampling variances of the

𝜋 estimator of the mean of NO3 -N of field Melle, for stratified simple random
sampling, using 25 compact geostrata of equal size, and one point per stratum.

13.2 Model-based optimisation of spatial strata

In this section a spatial stratification method is described that uses model
predictions of the study variable as a stratification variable. As opposed to
cum-root-f stratification, this spatial stratification method accounts for errors
in the predictions, as well as for spatial correlation of the prediction errors (de
Gruijter et al., 2015).
The Julia package Ospats is an implementation of this stratification method.
In Ospats the stratification is optimised through iterative reallocation of the
raster cells to the strata. Recently, this stratification method was implemented
in the package SamplingStrata (Barcaroli (2014), Barcaroli et al. (2020)).
However, the algorithm used to optimise the strata differs from that in Ospats.
In SamplingStrata the stratification is optimised by optimising the bounds
(splitting points) on the stratification variable with a genetic algorithm. Opti-
misation of the strata through optimisation of the bounds on the stratification
variable necessarily leads to non-overlapping strata, while with iterative reallo-
cation the strata may overlap, i.e., when the strata are sorted on the mean of
the stratification variable, the upper bound of a stratum can be larger than
the lower bound of the next stratum. As argued by de Gruijter et al. (2015)
optimisation of strata through optimisation of the stratum bounds can be
13.2 Model-based optimisation of spatial strata 253

suboptimal. On the other hand, optimisation of the stratum bounds needs

fewer computations and therefore is quicker.
The situation considered in this section is that prior data are available, either
from the study area itself or from another similar area, that can be used to fit
a linear regression model for the study variable, using one or more quantitative
covariates and/or factors as predictors. These predictors must be available
in the study area so that the fitted model can be used to map the study
variable in the study area. We wish to collect more data by stratified simple
random sampling, to be used in design-based estimation of the population
mean or total of the study variable. The central research question then is how
to construct these strata.
Recall the variance estimator of the mean estimator for stratified simple random
sampling (Equations (4.4) and (4.5)):

u�
𝑆2ℎ (𝑧)
𝑉(𝑧)̄̂ = ∑ 𝑤2ℎ . (13.9)
ℎ=1
𝑛ℎ

Plugging the stratum sample sizes under optimal allocation (Equation (4.17))
into Equation (13.9) yields

u� u�
1 √ 𝑤 𝑆 (𝑧)
𝑉(𝑧)̄̂ = ( ∑ 𝑤ℎ 𝑆ℎ (𝑧) 𝑐ℎ ∑ ℎ√ ℎ ) . (13.10)
𝑛 ℎ=1 ℎ=1
𝑐ℎ

So, given the total sample size 𝑛, the variance of the estimator of the mean is
minimal when the criterion

u� u�
√ 𝑤 𝑆 (𝑧)
𝑂 = ∑ 𝑤ℎ 𝑆ℎ (𝑧) 𝑐ℎ ∑ ℎ√ ℎ (13.11)
ℎ=1 ℎ=1
𝑐ℎ

is minimised.
Assuming that the costs are equal for all population units, so that the mean
costs are the same for all strata, the minimisation criterion reduces to

u� 2

𝑂 = ( ∑ 𝑤ℎ 𝑆ℎ (𝑧)) . (13.12)
ℎ=1

In practice, we do not know the values of the study variable 𝑧. de Gruijter et al.
(2015) consider the situation where we have predictions of the study variable
from a linear regression model: 𝑧 ̂ = 𝑧 + 𝜖, with 𝜖 the prediction error. So, this
implies that we do not know the population standard deviations within the
strata, 𝑆ℎ (𝑧) of Equation (13.10). What we do have are the stratum standard
deviations of the predictions of 𝑧: 𝑆ℎ (𝑧).
̂ With many statistical models, such
254 13 Model-based optimisation of probability sampling designs

as regression and kriging models, the standard deviation of the predictions is

smaller than that of the study variable: 𝑆ℎ (𝑧)̂ < 𝑆ℎ (𝑧). This is known as the
smoothing or levelling effect.
The stratum standard deviations in the minimisation criterion are replaced by
model-expectations of the stratum standard deviations, i.e., by model-based
predictions of the stratum standard deviations, 𝐸u� [𝑆ℎ (𝑧)]. This leads to the
following minimisation criterion:

u� 2

𝐸u� [𝑂] = ( ∑ 𝑤ℎ 𝐸u� [𝑆ℎ (𝑧)]) . (13.13)

ℎ=1

The stratum variances are predicted by

u� −1
ℎ u�ℎ
1
𝐸u� [𝑆2ℎ (𝑧)] = ∑ ∑ 𝐸u� [𝑑2u�u� ] , (13.14)
𝑁2ℎ u�=1 u�=u�+1

with 𝑑2u�u� = (𝑧u� − 𝑧u� )2 the squared difference of the study variable values at two
nodes of a discretisation grid. The model-expectation of the squared differences
are equal to

𝐸u� [𝑑2u�u� ] = (𝑧u�̂ − 𝑧u�̂ )2 + 𝑆2 (𝜖u� ) + 𝑆2 (𝜖u� ) − 2𝑆2 (𝜖u� , 𝜖u� ) , (13.15)

with 𝑆2 (𝜖u� ) the variance of the prediction error at node 𝑖 and 𝑆2 (𝜖u� , 𝜖u� ) the
covariance of the prediction errors at nodes 𝑖 and 𝑗. The authors then argue
that for smoothers, such as kriging and regression, the first term must be
divided by the squared correlation coefficient 𝑅2 :

(𝑧u�̂ − 𝑧u�̂ )2
𝐸u� [𝑑2u�u� ] = + 𝑆2 (𝜖u� ) + 𝑆2 (𝜖u� ) − 2𝑆2 (𝜖u� , 𝜖u� ) . (13.16)
𝑅2
The predicted stratum standard deviations are approximated by the square root
of Equation (13.16). Plugging these model-based predictions of the stratum
standard deviations into the minimisation criterion, Equation (13.12), yields

u� −1 u�ℎ 1/2
1 u� ℎ
(𝑧u�̂ − 𝑧u�̂ )2
𝐸u� [𝑂] = ∑( ∑ ∑ + 𝑆2 (𝜖u� ) + 𝑆2 (𝜖u� ) − 2𝑆2 (𝜖u� , 𝜖u� )) .
𝑁 ℎ=1 u�=1 u�=u�+1
𝑅2

(13.17)

Optimal spatial stratification with package SamplingStrata is illustrated

with a survey of the SOM concentration (g kg-1 ) in the topsoil (A horizon)
of Xuancheng (China). Three samples are available. These three samples are
merged. The total number of sampling points is 183. This sample is used to fit
13.2 Model-based optimisation of spatial strata 255

a simple linear regression model for the SOM concentration, using the elevation
of the surface (dem) as a predictor. Function lm of the stats package is used
to fit the simple linear regression model.

lm_SOM <- lm(SOM_A_hori ~ dem, data = sampleXuancheng)

In fitting a linear regression model, we assume that the relation is linear, the
residual variance is constant (independent of the fitted value), and the residuals
have a normal distribution. These assumptions are checked with a scatter plot
of the residuals against the fitted value and a Q-Q plot, respectively (Figure
13.6).

FIGURE 13.6: Scatter plot of residuals against fitted value and Q-Q plot of
residuals, for a simple linear regression model of the SOM concentration in
Xuancheng, using elevation as a predictor.

The scatter plot shows that the first assumption is realistic. No pattern can
be seen: at all fitted values, the residuals are scattered around the horizontal
line. However, the second and third assumptions are questionable: the residual
variance clearly increases with the fitted value, and the distribution of the
residuals has positive skew, i.e., it has a long upper tail. There clearly is some
evidence that these two assumptions are violated. Possibly these problems can
be solved by fitting a model for the natural log of the SOM concentration.
The variance of the residuals is more constant (Figure 13.7), and the Q-Q plot
is improved, although we now have too many strong negative residuals for
a normal distribution. I proceed with the model for natural-log transformed
SOM (lnSOM). The fitted linear regression model is used to predict lnSOM at
the nodes of a 200 m × 200 m discretisation grid.
256 13 Model-based optimisation of probability sampling designs

FIGURE 13.7: Scatter plot of residuals against fitted value and Q-Q plot of
residuals, for a simple linear regression model of the natural log of the SOM
concentration in Xuancheng, using elevation as a predictor.

res <- predict(lm_lnSOM, newdata = as(grdXuancheng, "data.frame"), se.fit = TRUE)

grdXuancheng <- within(grdXuancheng, {
lnSOMpred <- res$fit; varpred <- res$se.fitˆ2})

The predictions and their standard errors are shown in Figure 13.8.
Let us check now whether the spatial structure of the study variable lnSOM is
fully captured by the mean, modelled as a linear function of elevation. This
can be checked by estimating the semivariogram of the model residuals. If
the semivariogram of the residuals is pure nugget (the semivariance does not
increase with distance), then we can assume that the prediction errors are
independent. In that case, we do not need to account for a covariance of
the prediction errors in optimisation of the spatial strata. However, if the
semivariogram does show spatial structure, we must account for a covariance
of the prediction errors. Figure 13.9 shows the sample semivariogram of the
residuals computed with function variogram of package gstat.

library(gstat)
sampleXuancheng <- sampleXuancheng %>%
mutate(s1 = s1 / 1000, s2 = s2 / 1000)
coordinates(sampleXuancheng) <- ~ s1 + s2
vg <- variogram(lnSOM ~ dem, data = sampleXuancheng)

The sample semivariogram does not show much spatial structure, but the first
two points in the semivariogram have somewhat smaller values. This indicates
13.2 Model-based optimisation of spatial strata 257

FIGURE 13.8: Predicted natural log of the SOM concentration (g kg-1 ) in

the topsoil of Xuancheng and the standard error (se) of the predictor, obtained
with a linear regression model with elevation as a predictor.
258 13 Model-based optimisation of probability sampling designs

FIGURE 13.9: Sample semivariogram of the residuals of a simple linear

regression model for the natural log of the SOM concentration in Xuancheng.
Numbers refer to point-pairs used in computing semivariances.

TABLE 13.1: Estimated regression coefficients (intercept and slope for dem)
and parameters of an exponential semivariogram for the natural log of the
SOM concentration (g kg-1 ) in Xuancheng.

Int dem Nugget Partial sill Distance parameter (km)

2.771 0.00222 0.085 0.061 2.588

that the residuals at two close points, say, < ±5 km, are not independent,
whereas if the distance between the two points > ±5 km, they are independent.
This spatial dependency of the residuals can be modelled, e.g., by an exponential
function. The exponential semivariogram has three parameters, the nugget
variance 𝑐0 , the partial sill 𝑐1 , and the distance parameter 𝜙. The total number
of model parameters now is five: two regression coefficients (intercept and slope
for elevation) and three semivariogram parameters. All five parameters can best
be estimated by restricted maximum likelihood, see Subsection 21.5.2. Table
13.1 shows the estimated regression coefficients and semivariogram parameters.
Up to a distance of about three times the estimated distance parameter 𝜙,
which is about 8 km, the residuals are spatially correlated; beyond that distance,
they are hardly correlated anymore.
We conclude that the errors in the regression model predictions are not inde-
pendent, although the correlation will be weak in this case, and that we must
account for this correlation in optimising the spatial strata.
13.2 Model-based optimisation of spatial strata 259

The discretisation grid with predicted lnSOM consists of 115,526 nodes. These
are too many for function optimStrata. The grid is therefore thinned to a grid
with a spacing of 800 m × 800 m, resulting in 7,257 nodes.
The first step in optimisation of spatial strata with package SamplingStrata
is to build the sampling frame with function buildFrameSpatial. Argument X
specifies the stratification variables, and argument Y specifies the study variables.
In our case, we have only one stratification variable and one study variable,
and these are the same variable. Argument variance specifies the variance of
the prediction error of the study variable. Variable dom is an identifier of the
domain of interest of which we want to estimate the mean or total. I assign
the value 1 to all population units, see code chunk below, which implies that
the stratification is optimised for the entire population. If we have multiple
domains of interest, the stratification is optimised for each domain separately.
Finally, as a preparatory step we must specify how precise the estimated
mean should be. This precision must be specified in terms of the coefficient of
variation (cv), i.e., the standard error of the estimated mean divided by the
mean. I use a cv of 0.005. In case of multiple domains of interest and multiple
study variables, a cv must be specified per domain and per study variable.
This precision requirement is used to compute the sample size for Neyman
allocation (Equation (4.16))1 . The optimal stratification is independent of the
precision requirement.

library(SamplingStrata)
subgrd$id <- seq_len(nrow(subgrd))
subgrd$dom <- rep(1, nrow(subgrd))
frame <- buildFrameSpatial(df = subgrd, id = "id", X = c("lnSOMpred"),
Y = c("lnSOMpred"), variance = c("varpred"), lon = "x1", lat = "x2",
domainvalue = "dom")
cv <- as.data.frame(list(DOM = "DOM1", CV1 = 0.005, domainvalue = 1))

The optimal spatial stratification can be computed with function optimStrata,

with argument method = "spatial". The 𝑅2 value of the linear regression model,
used in the minimisation criterion (Equation (13.17)), can be specified with
argument fitting.

I am not sure that the correction factor 𝑅2 in Equation (13.16) is really

needed. I believe that the smoothing effect is already accounted for by the
variances and covariances of the prediction errors. I used an 𝑅2 value of 1.

Arguments range and kappa are parameters of an exponential semivariogram,

needed for computing the covariance of the prediction errors. Function
1 For multivariate stratification, i.e., stratification with multiple study variables, Bethel

allocation is used to compute the required sample size.

260 13 Model-based optimisation of probability sampling designs

optimStrata uses an extra parameter in the exponential semivariogram: 𝑐0 +

𝑐1 exp(−𝜅ℎ/𝜙). So, for the usual exponential semivariogram (Equation (21.13))
kappa equals 1.

res <- optimStrata(

framesamp = frame, method = "spatial", errors = cv, nStrata = 5,
fitting = 1, range = c(vgm_REML$phi), kappa = 1, showPlot = FALSE)

A summary of the optimised strata can be obtained with function summaryStrata.

print(smr_strata <- summaryStrata(

res$framenew, res$aggr_strata, progress = FALSE))

Domain Stratum Population Allocation SamplingRate Lower_X1 Upper_X1

1 1 1 3090 12 0.004035 2.767950 2.862539
2 1 2 1931 10 0.005087 2.864846 2.991733
3 1 3 1133 10 0.008555 2.994040 3.215515
4 1 4 769 11 0.014398 3.217822 3.600789
5 1 5 334 11 0.033443 3.605403 5.098052

In the next code chunk, it is checked whether the coefficient of variation is

indeed equal to the desired value of 0.005.

strata <- res$aggr_strata

framenew <- res$framenew
N_h <- strata$N
w_h <- N_h / sum(N_h)
se <- sqrt(sum(w_hˆ2 * strata$S1ˆ2 / strata$SOLUZ))
se / mean(framenew$Y1)

[1] 0.005033349

The coefficient of variation can also be computed with function expected_CV.

expected_CV(strata)

cv(Y1)
DOM1 0.005

Figure 13.10 shows the optimised strata. I used the stratum bounds in data.frame
smr_strata,to compute the stratum for all raster cells of the original 200 m ×
200 m grid.
13.2 Model-based optimisation of spatial strata 261

FIGURE 13.10: Model-based optimal strata for estimating the mean of the
natural log of the SOM concentration in Xuancheng.
14
Sampling for estimating parameters of
domains

This chapter is about probability sampling and estimation of means or totals

of subpopulations (subareas, subregions). In the sampling literature, these
subpopulations are referred to as domains of interest, or shortly domains.
Ideally, at the stage of designing a sample, these domains are known, and of
every population unit we know to which domain it belongs. In that situation
it is most convenient to use these domains as strata in random sampling, so
that we can control the sample size in each domain (Chapter 4).
If we have multiple maps with domains, think for instance of a soil class map,
a map with land cover classes, and a map with countries, we can make an
overlay of these maps to construct the cross-tabulation strata. However, this
may result in numerous strata, in some cases even more than the sample size.
In this situation, an attractive solution is multiway stratification (Subsection
9.1.4). With this design, the domains of interest are used as strata, not their
cross-classification, and the sample sizes of these marginal strata are controlled.
Even with a multiway stratified sample, resulting in controlled sample sizes
for each domain, the sample size of a domain can be too small for a reliable
estimate of the mean or total when using just the data of this domain. In
this case we may use model-assisted estimators (Chapter 10) to increase the
precision. Not only the data collected from a given domain are used to estimate
the mean or total, but also data outside the domain (Section 14.2) (Chaudhuri
(1994), Rao (2003), Falorsi and Righi (2008)).
We may also wish to estimate the mean or total of domains that are not used
as strata or marginal strata. The sample size in these domains is then not
controlled and varies among samples selected with the sampling design. As
before with multiway stratified sampling, the mean can either be estimated
with the direct estimator, using the data from the domain only (Section 14.1),
or a model-assisted estimator, also using data from outside the domain (Section
14.2).

263
264 14 Sampling for estimating parameters of domains

14.1 Direct estimator for large domains

If the sample size of a domain 𝑑 is considered large enough to obtain a reliable
estimate of the mean and, besides, the size of the domain is known, the mean
of that domain can be estimated by the direct estimator:
1 𝑧
𝑧u�̄̂ = ∑ u�u� , (14.1)
𝑁u� u�∈u� 𝜋u�u�
u�

where 𝑁u� is the size of the domain, 𝑧u�u� is the value for unit 𝑘 of domain 𝑑,
and 𝜋u�u� is the inclusion probability of this point.
When the domain is not used as a (marginal) stratum, so that the sample size
of the domain is random, the mean of the domain can be estimated best by
the ratio estimator:
u�u�u�
𝑡u�̂ (𝑧) ∑u�∈u� u�u�u�
̄̂
𝑧ratio,u� = = u�
1
. (14.2)
𝑁u�̂ ∑ u�u�u�
u�∈u� u�

The ratio estimator can also be used when the size of the domain is unknown.
An example of this is estimating the mean of soil classes as observed in the
field, not as depicted on a soil map. A soil map is impure, i.e., the map units
contain patches with soil classes that differ from the soil class as indicated
on the map. The area of a given true soil class is not known.

For simple random sampling without replacement, 𝜋u�u� = 𝑛/𝑁. Inserting this
in Equation (14.2) gives
1
̄̂
𝑧ratio,u� = ∑𝑧 . (14.3)
𝑛u� u�∈u� u�u�
u�

The mean of the domain is simply estimated by the mean of the 𝑧-values
observed in the domain, i.e., the sample mean in domain 𝑑. The variance of
this estimator can be estimated by
1 1
𝑉(̂ 𝑧ratio,u�
̄̂ )= 2 ⋅ ∑ (𝑧 − 𝑧u�̄ u� )2 , (14.4)
𝑎u�̂ 𝑛 (𝑛 − 1) u�∈u� u�u�
u�

where 𝑧u�̄ u� is the sample mean in domain 𝑑, and 𝑎u�̂ is the estimated relative
size of domain 𝑑:
𝑛
𝑎u�̂ = u� . (14.5)
𝑛
14.1 Direct estimator for large domains 265

Refer to section (8.2.2) in de Gruijter et al. (2006) for the ratio estimator and
its standard error with stratified simple random sampling, in case the domains
cut across the strata, and other sampling designs.
The ratio estimator and its standard error can be computed with function svyby
of package survey (Lumley, 2021). This is illustrated with Eastern Amazonia.
We wish to estimate the mean aboveground biomass (AGB) of the 16 ecoregions
from a simple random sample of 200 units.

library(survey)
set.seed(314)
n <- 200
mysample <- grdAmazonia %>%
mutate(N = n()) %>%
slice_sample(n = n)
design_si <- svydesign(id = ~ 1, data = mysample, fpc = ~ N)
res <- svyby(~ AGB, by = ~ Ecoregion, design = design_si, FUN = svymean)

The ratio estimates of the mean AGB are shown in Table 14.1. Two ecoregions
are missing in the table: no units are selected from these ecoregions so that a
direct estimate is not available. There are three ecoregions with an estimated
standard error of 0.0. These ecoregions have less than two sampling units only,
so that the standard error cannot be estimated.

TABLE 14.1: Ratio estimates and estimated standard errors of the ratio
estimator of the mean AGB (109 kg ha-1 ) of ecoregions in Eastern Amazonia,
with simple random sampling without replacement of size 200.

Ecoregion AGB se
Cerrado 99.7 15.6
Guianan highland moist forests 296.0 0.0
Guianan lowland moist forests 263.0 0.0
Gurupa varzea 80.0 0.0
Madeira-Tapajos moist forests 286.0 14.0
Marajo varzea 116.6 27.2
Maranhao Babassu forests 90.0 9.6
Mato Grosso tropical dry forests 177.0 90.7
Monte Alegre varzea 189.5 89.0
Purus-Madeira moist forests 145.5 47.1
Tapajos-Xingu moist forests 288.6 8.9
Tocantins/Pindare moist forests 176.3 17.9
Uatuma-Trombetas moist forests 274.7 6.7
Xingu-Tocantins-Araguaia moist forests 223.1 16.6
The estimated standard errors of 0.0 are non-availables.
266 14 Sampling for estimating parameters of domains

TABLE 14.2: Standard deviations of 1,000 𝜋 estimates (HT) and 1,000 ratio
estimates (Ratio) of the mean AGB (109 kg ha-1 ) of ecoregions in Eastern
Amazonia, with simple random sampling without replacement of size 200.

Ecoregion HT Ratio n
Amazon-Orinoco-Southern Caribbean mangroves 136.9 36.9 0.87
Cerrado 38.6 18.2 8.16
Guianan highland moist forests 271.4 29.5 0.71
Guianan lowland moist forests 160.0 17.9 2.51
Guianan savanna 101.5 67.1 2.86
Gurupa varzea 103.7 57.8 0.99
Madeira-Tapajos moist forests 55.9 14.2 22.96
Marajo varzea 48.6 24.7 10.80
Maranhao Babassu forests 59.3 25.7 4.61
Mato Grosso tropical dry forests 72.7 41.1 2.98
Monte Alegre varzea 120.2 63.3 2.27
Purus-Madeira moist forests 288.0 55.5 0.45
Tapajos-Xingu moist forests 45.5 12.0 31.51
Tocantins/Pindare moist forests 31.5 15.3 28.39
Uatuma-Trombetas moist forests 33.4 8.1 52.37
Xingu-Tocantins-Araguaia moist forests 42.4 17.2 27.56
n: expected sample size.

The simple random sampling is repeated 1,000 times, and every sample is used
to estimate the mean AGB of the ecoregions both with the 𝜋 estimator and
the ratio estimator. As can be seen in Table 14.2 the standard deviation of the
ratio estimates is much smaller than that of the 𝜋 estimates. The reason is
that the number of sampling units in an ecoregion varies among samples, i.e.,
the sample size of an ecoregion is random. When many units are selected from
an ecoregion, the estimated total of that ecoregion is large. The estimated
mean as obtained with the 𝜋 estimator then is large too, because the estimated
total is divided by the fixed size (total number of population units, 𝑁u� ) of
the ecoregion. However, in the ratio estimator the size of an ecoregion is
estimated from the same sample, although we know its size, see Equation
(14.2). With many units selected from an ecoregion, the estimated size of that
ecoregion, 𝑁̂u� , is also large. By dividing the large estimated total by the large
estimated size, a more stable estimate of the mean of the domain is obtained.
For quite a few ecoregions the standard deviations are very large, especially
of the 𝜋 estimator. These are the ecoregions with very small average sample
sizes. With simple random sampling, the expected sample size can simply be
computed by 𝐸[𝑛] = 𝑛 𝑁u� /𝑁. In the following section, alternative estimators
are described for these ecoregions with small expected sample sizes. To speed
14.2 Model-assisted estimators for small domains 267

up the computations, I used a 5 km × 5 km subgrid of grdAmazonia in this

sampling experiment.
No covariates are used in the ratio estimator. If we wish to exploit covariates,
the mean of a domain can be estimated best by the ratio of the regression
estimate of the domain total and the estimated size of the domain:

̂
𝑡regr,u� (𝑧)
̄̂
𝑧ratio,u� = . (14.6)
̂u�
𝑁

For a large domain with a reasonable sample size, the regression estimate can be
computed from the data of that domain (Chapter 10). For small domains, also
the data from outside these domains can be used to estimate the population
regression coefficients. This is explained in Subsection 14.2.1.

14.2 Model-assisted estimators for small domains

When the domains are not well represented in the sample, the direct estimators
from the previous section lead to large standard errors. In this situation, we
may try to increase the precision by also using observations from outside the
domain. If we have covariates related to the study variable, we may exploit
this ancillary information by fitting a regression model relating the study
variable to the covariates and using the fitted model to predict the study
variable for all population units (nodes of discretisation grid), see Chapter
10. However, for a small domain, we may have too few sampled units in that
domain to fit a separate regression model. The alternative then is to use the
entire sample to estimate the regression coefficients, and to use this global
regression model to estimate the means of the domains. This introduces a
systematic error, a design-bias, in the estimator. However, this extra error is
potentially outweighed by the reduction of the random error due to the use of
the globally estimated regression coefficients. If one or more units are selected
from a domain, the observations of the study variable of these units can be
used to correct for the bias. This leads to the regression estimator for small
domains. In the absence of such data, the mean of the domain can still be
estimated by the so-called synthetic estimator.
There are quite a few packages for model-assisted estimation of means of
small areas, the maSAE package (Cullman, 2021), the JoSAE package (Brei-
denbach, 2018), the rsae package (Schoch, 2014), and the forestinventory
package (Hill et al., 2021). I use package forestinventory for model-assisted
estimation (Subsections 14.2.1 and 14.2.2) and package JoSAE for model-
based prediction of the means of small areas (Section 14.3).
268 14 Sampling for estimating parameters of domains

14.2.1 Regression estimator

In the regression estimator, the potential bias due to the globally estimated
regression coefficients can be eliminated by adding the 𝜋 estimator of the
mean of the regression residuals to the mean of the predictions in the domain
(compare with Equation (10.8)) (Mandallaz (2007), Mandallaz et al. (2013)):

u� u�
1 1 𝑒 1 𝑒
̄̂
𝑧regr,u� = ∑ 𝐱T 𝐛̂ + ∑ u�u� = 𝐱̄ T ̂
u� 𝐛 + ∑ u�u� , (14.7)
𝑁u� u�=1 u�u� 𝑁u� u�∈u� 𝜋u�u� 𝑁u� u�∈u� 𝜋u�u�
u� u�

with 𝐱u�u� the vector with covariate values for unit 𝑘 in domain 𝑑, 𝐛̂ the vector
with globally estimated regression coefficients, 𝑒u�u� the residual for unit 𝑘 in
domain 𝑑, 𝜋u�u� the inclusion probability of that unit, and 𝐱̄ u� the mean of the
covariates in domain 𝑑. Alternatively, the mean of the residuals in a domain is
estimated by the ratio estimator:

1 𝑒
̄̂
𝑧regr,u� = 𝐱̄ T ̂
u� 𝐛 + ∑ u�u� , (14.8)
̂ 𝜋
𝑁u� u�∈u�u� u�u�

with 𝑁̂u� the estimated size of domain 𝑑, see Equation (14.2). The regression
coefficients can be estimated by Equation (10.15). With simple random sam-
pling, the second term in Equation (14.8) is equal to the sample mean of the
residuals, so that the estimator reduces to

̄̂
𝑧regr,u� = 𝐱̄ T ̂
u� 𝐛 + 𝑒u�
̄ u� , (14.9)

with 𝑒u�̄ u� the sample mean of the residuals in domain 𝑑.

A regression estimate can only be computed if we have at least one observation
of the study variable in the domain 𝑑. The variance of the regression estimator
of the mean for a small domain can be estimated by (Hill et al., 2021)

𝑉(̂ 𝑧regr,u�
̄̂ ) = 𝐱̄ T ̂ ̂ ̂ ̄̂ ) ,
u� 𝐂(𝐛)𝐱̄ u� + 𝑉(𝑒u� (14.10)

with 𝐂(
̂ 𝐛)̂ the matrix with estimated sampling variances and covariances of
the regression coefficients. The first variance component is the contribution
due to uncertainty about the regression coefficients, the second component
accounts for the uncertainty about the mean of the residuals in the domain. For
simple random sampling, the sampling variance of the 𝜋 estimator of the mean
of the residuals in a domain can be estimated by the sample variance of the
residuals in that domain divided by the sample size 𝑛u� . This variance estimator
is presented in Hill et al. (2021). If the domain is not used as a stratum and the
14.2 Model-assisted estimators for small domains 269

domain mean of the residuals is estimated by the ratio estimator, the second
variance component can be estimated by

2
𝑛 1
𝑉(̂ 𝑒ratio,u�
̄̂ )=( ) ⋅ ∑ (𝑒 − 𝑒u�̄ u� )2 . (14.11)
𝑛u� 𝑛 (𝑛 − 1) u�∈u� u�u�
u�

With simple random sampling, the sampling variances and covariances of the
estimated regression coefficients can be estimated by (equation 2 in Hill et al.
(2021))

−1 u� −1
̂ = 1 (∑ 𝐱u� 𝐱T )
̂ 𝐛)
𝐂(
1
( 2 ∑ 𝑒2u� 𝐱u� 𝐱T
1
(∑ 𝐱u� 𝐱T . (14.12)
u� u� ) u� )
𝑛 u�∈u� 𝑛 u�∈u� 𝑛 u�∈u�

The sampling variances and covariances of the estimators of the population

regression coefficients are not equal to the model-variances and covariances
as obtained with multiple linear regression, using functions lm and vcov, see
Section 10.1 and Chapter 26.
Function twophase of package forestinventory (Hill et al., 2021) can be used
to compute the regression estimate for small domains and its standard error.

The function name ‘twophase’ is somewhat confusing. It suggests that we

have a large sample which is subsampled in a second phase, as described in
Chapter 11. This is not the case here. However, upon considering infinite
populations, Hill et al. (2021) treat the grid that discretises the infinite
population as the first-phase sample. The sampling error introduced by this
discretisation grid can then be accounted for. I ignore this sampling error,
which will be very small anyway, because the number of grid cells is very
large.

By assigning the domain means of the covariates to argument exhaustive of

function twophase the sampling error of the first phase is ignored. Function
twophase assumes simple random sampling (unless optional argument cluster is
used). Note that for the unobserved population units (not selected units) the
AGB values are changed into non-availables. In package survey also a function
twophase is defined, for this reason the name of the package is made explicit by
forestinventory::twophase. With argument psmall = TRUE and element unbiased =
TRUE in the list small_area the regression estimate is computed. Log-transformed
SWIR2 is used as a covariate in the regression estimator of the mean AGB of
the ecoregions.
270 14 Sampling for estimating parameters of domains

library(forestinventory)
n <- 200
set.seed(314)
units <- sample(nrow(grdAmazonia), size = n, replace = FALSE)
grdAmazonia <- grdAmazonia %>%
mutate(lnSWIR2 = log(SWIR2),
id = row_number(),
ind = as.integer(id %in% units))
grdAmazonia$AGB[grdAmazonia$ind == 0L] <- NA
mx_eco_pop <- tapply(
grdAmazonia$lnSWIR2, INDEX = grdAmazonia$Ecoregion, FUN = mean)
mX_eco_pop <- data.frame(
Intercept = rep(1, length(mx_eco_pop)), lnSWIR2 = mx_eco_pop)
ecos_in_sam <- unique(mysample$Ecoregion)
res <- forestinventory::twophase(AGB ~ lnSWIR2,
data = as.data.frame(grdAmazonia),
phase_id = list(phase.col = "ind", terrgrid.id = 1),
small_area = list(sa.col = "Ecoregion",
areas = sort(ecos_in_sam),
unbiased = TRUE),
psmall = TRUE, exhaustive = mX_eco_pop)
regr <- res$estimation

The alternative is to save the selected units (sample) in a data frame, passed
to function twophase with argument data. The results are identical because the
true means of the covariate 𝑥 specified with argument exhaustive contains all
required information at the population level.
For two ecoregions, no regression estimate of the mean AGB is obtained (Table
14.3). No units are selected from these domains. The estimated variance of
the estimated domain mean is in the column g_var. In the estimated variance
ext_var the first variance component of Equation (14.10) is ignored. Note that
for the ecoregions with a sample size of one unit (the sample size per domain
is in column n2G), no estimate of the variance is available, because the variance
of the estimated mean of the residuals cannot be estimated from one unit.
Figure 14.1 shows the regression estimates plotted against the ratio estimates.
The intercept of the line, fitted with ordinary least squares (OLS), is larger
than 0, and the slope is smaller than 1. Using the regression model predictions
in the estimation of the means leads to some smoothing.
I quantified the gain in precision of the estimated mean AGB due to the use
of the regression model by the variance of the ratio estimator divided by the
variance of the regression estimator (Table 14.4). For ratios larger than 1,
there is a gain in precision. Both variances are estimated from 1,000 repeated
ratio and regression estimates obtained with simple random sampling without
14.2 Model-assisted estimators for small domains 271

TABLE 14.3: Regression estimates of the mean AGB (109 kg ha-1 ) of ecore-
gions in Eastern Amazonia, for simple random sample without replacement of
size 200, using lnSWIR2 as a predictor.

Ecoregion AGB ext_var g_var n2G

Cerrado 105.5 51.6 82.6 12
Guianan highland moist forests 281.8 NA NaN 1
Guianan lowland moist forests 280.3 NA NaN 1
Gurupa varzea 57.8 NA NaN 1
Madeira-Tapajos moist forests 296.1 91.2 107.3 19
Marajo varzea 163.1 535.2 546.0 10
Maranhao Babassu forests 114.2 157.1 178.1 7
Mato Grosso tropical dry forests 144.6 959.6 981.5 2
Monte Alegre varzea 209.1 4,105.7 4,117.3 2
Purus-Madeira moist forests 225.7 2,428.9 2,441.1 2
Tapajos-Xingu moist forests 277.8 23.4 36.9 38
Tocantins/Pindare moist forests 157.8 75.8 87.7 30
Uatuma-Trombetas moist forests 270.6 33.3 47.5 46
Xingu-Tocantins-Araguaia moist forests 223.6 51.1 61.6 29
For explanation of variances of regression estimator, see text. n2G:
sample size of ecoregion; NA: not available; NaN: not a number.

replacement of size 200. For all but two small ecoregions, there is a gain. For
quite a few ecoregions, the gain is quite large. These are the ecoregions where
the globally fitted regression model explains a large part of the spatial variation
of AGB.

14.2.2 Synthetic estimator

For small domains from which no units are selected, the mean can still be
estimated by the synthetic estimator, also referred to as the synthetic regression
estimator, by dropping the second term in Equation (14.7):

̄̂
𝑧syn,u� = 𝐱̄ T ̂
u� 𝐛 . (14.13)

The variance can be estimated by

𝑉(̂ 𝑧syn,u�
̄̂ ) = 𝐱̄ T ̂ ̂
u� 𝐂(𝐛)𝐱̄ u� . (14.14)

This is equal to the first variance component of Equation (14.10). The synthetic
estimate can be computed with function twophase, with argument psmall = FALSE
and element unbiased = FALSE in the list small_area.
272 14 Sampling for estimating parameters of domains

FIGURE 14.1: Scatter plot of the ratio and the regression estimates of the
mean AGB (109 kg ha-1 ) of ecoregions in Eastern Amazonia for simple random
sample without replacement of size 200. In the regression estimate, lnSWIR2
is used as a predictor. The line is fitted by ordinary least squares.

res <- forestinventory::twophase(AGB ~ lnSWIR2,

data = as.data.frame(grdAmazonia),
phase_id = list(phase.col = "ind", terrgrid.id = 1),
small_area = list(sa.col = "Ecoregion", areas = ecoregions, unbiased = FALSE),
psmall = FALSE, exhaustive = mX_eco_pop)
synt <- res$estimation

For all ecoregions, also the unsampled ones, a synthetic estimate of the mean
AGB is obtained (Table 14.5). For the sampled ecoregions, the synthetic
estimate differs from the regression estimate. This difference can be quite large
for ecoregions with a small sample size. Averaged over all sampled ecoregions,
the difference, computed as synthetic estimate minus regression estimate,
equals 14.9 109 kg ha-1 . The variance of the regression estimator is always
much larger than the variance of the synthetic estimator. The difference is
the variance of the estimator of the domain mean of the residuals. However,
recall that the regression estimator is design-unbiased, whereas the synthetic
14.3 Model-based prediction 273

TABLE 14.4: Estimated gain in precision of the estimated mean AGB of

ecoregions in Eastern Amazonia, as quantified by the ratio of the estimated
variance of the ratio estimator (no covariate used) to the estimated variance
of the regression estimator (using lnSWIR2 as a predictor), for simple random
sampling without replacement of size 200.

Ecoregion Gain
Amazon-Orinoco-Southern Caribbean mangroves 0.57
Cerrado 1.63
Guianan highland moist forests 1.01
Guianan lowland moist forests 1.26
Guianan savanna 6.70
Gurupa varzea 0.65
Madeira-Tapajos moist forests 1.44
Marajo varzea 1.87
Maranhao Babassu forests 1.70
Mato Grosso tropical dry forests 2.81
Monte Alegre varzea 1.21
Purus-Madeira moist forests 1.74
Tapajos-Xingu moist forests 2.73
Tocantins/Pindare moist forests 2.56
Uatuma-Trombetas moist forests 2.10
Xingu-Tocantins-Araguaia moist forests 4.06

estimator is not. A more fair comparison is on the basis of the root mean
squared error (RMSE) (Table 14.6). For the regression estimator, the RMSE
is equal to its standard error and therefore not shown in the table.
In the synthetic estimator and the regression estimator, both quantitative
covariates and categorical variables can be used. If one or more categorical
variables are included in the estimator, the variable names in the data frame
with the true means of the ancillary variables per domain, specified with
argument exhaustive, must correspond to the column names of the design
matrix that is generated with function lm, see Subsection 10.1.3.

14.3 Model-based prediction

The alternative for design-based and model-assisted estimation of the means or
totals of small domains is model-based prediction. The fundamental difference
between model-assisted estimation and model-based prediction is explained
274 14 Sampling for estimating parameters of domains

TABLE 14.5: Synthetic estimates of the mean AGB (109 kg ha-1 ) of ecore-
gions in Eastern Amazonia, for simple random sample without replacement of
size 200, using lnSWIR2 as a predictor.

Ecoregion AGB g_var n2G

Amazon-Orinoco-Southern Caribbean mangroves 225.3 10.6 0
Cerrado 81.6 31.0 12
Guianan highland moist forests 278.6 16.0 1
Guianan lowland moist forests 277.3 15.8 1
Guianan savanna 171.8 12.2 0
Gurupa varzea 218.2 10.4 1
Madeira-Tapajos moist forests 279.2 16.1 19
Marajo varzea 230.6 10.8 10
Maranhao Babassu forests 118.3 20.9 7
Mato Grosso tropical dry forests 114.0 21.9 2
Monte Alegre varzea 242.6 11.6 2
Purus-Madeira moist forests 250.4 12.3 2
Tapajos-Xingu moist forests 261.1 13.4 38
Tocantins/Pindare moist forests 175.7 11.9 30
Uatuma-Trombetas moist forests 267.2 14.2 46
Xingu-Tocantins-Araguaia moist forests 221.9 10.5 29
n2G: sample size of ecoregion.

in Chapter 26. The models used in this section are linear mixed models. In
a linear mixed model, the mean of the study variable is modelled as a linear
combination of covariates, similar to a linear regression model. The difference
with a linear regression model is that the residuals of the mean are not assumed
independent. The dependency of the residuals is also modelled. Two types of
linear mixed model are described: a random intercept model and a geostatistical
model.

14.3.1 Random intercept model

A basic linear mixed model that can be used for model-based prediction of
means of small domains is the random intercept model:
𝑍u�u� = 𝐱T
u�u�𝛽 + 𝑣u� + 𝜖u�u�
𝑣u� ∼ 𝒩(0, 𝜎2u� ) (14.15)
𝜖u�u� ∼ 𝒩(0, 𝜎2u� ) .

Two random variables are now involved, both with a normal distribution with
mean zero: 𝑣u� , a random intercept at the domain level with variance 𝜎2u� , and
the residuals 𝜖u�u� at the unit level with variance 𝜎2u� . The variance 𝜎2u� can
14.3 Model-based prediction 275

TABLE 14.6: Estimated standard error (se), bias, and root mean squared
error (RMSE) of the regression estimator (reg) and the synthetic estimator
(syn) of the mean AGB of ecoregions in Eastern Amazonia. The regression
estimator is design-unbiased, so the RMSE of the regression estimator is equal
to its standard error.

Ecoregion RMSE reg se syn bias syn RMSE syn

Amazon-Orinoco Carib. mangroves 48.8 3.7 96.2 96.3
Cerrado 14.2 7.1 -17.3 18.7
Guianan highland moist forests 29.4 4.9 -1.2 5.0
Guianan lowland moist forests 15.9 4.8 -7.2 8.7
Guianan savanna 25.9 4.1 12.3 13.0
Gurupa varzea 71.8 3.7 119.4 119.5
Madeira-Tapajos moist forests 11.8 4.9 -2.2 5.4
Marajo varzea 18.1 3.8 74.4 74.5
Maranhao Babassu forests 19.7 5.7 3.9 6.9
Mato Grosso tropical dry forests 24.5 5.8 3.4 6.7
Monte Alegre varzea 57.5 3.9 50.6 50.8
Purus-Madeira moist forests 42.1 4.2 32.9 33.1
Tapajos-Xingu moist forests 7.3 4.4 -18.5 19.0
Tocantins/Pindare moist forests 9.5 4.0 14.7 15.3
Uatuma-Trombetas moist forests 5.6 4.5 -14.2 14.9
Xingu-Toc.-Arag. moist forests 8.5 3.7 -2.5 4.4

be interpreted as a measure of the heterogeneity among the domains after

accounting for the fixed effect (Breidenbach and Astrup, 2012). With this
model, the mean of a domain can be predicted by

̄̂
𝑧mb,u� = 𝐱̄ T ̂
u� 𝛽 + 𝑣u�
̂ , (14.16)

with 𝛽̂ the best linear unbiased estimates (BLUE) of the regression coefficients
and 𝑣u�̂ the best linear unbiased prediction (BLUP) of the intercept for domain
𝑑, 𝑣u� . The model-based predictor can also be written as

1
̄̂
𝑧mb,u� = 𝐱̄ T ̂
u� 𝛽 + 𝜆u� ( ∑𝜖 ) , (14.17)
𝑛u� u�∈u� u�u�
u�

with 𝜆u� a weight for the second term that corrects for the bias of the synthetic
estimator. This weight is computed by

𝜎̂2u�
𝜆u� = . (14.18)
𝜎̂2u� + 𝜎̂2u� /𝑛u�
276 14 Sampling for estimating parameters of domains

This equation shows that the larger the estimated residual variance 𝜎̂2u� , the
smaller the weight for the bias correction factor, and the larger the sample size
𝑛u� , the larger the weight. Comparing Equations (14.16) and (14.17) shows
that the random intercept of a domain is predicted by the sample mean of the
residuals of that domain, multiplied by a weight factor computed by Equation
(14.18).
The means of the small domains can be computed with function eblup.mse.f.wrap
of package JoSAE (Breidenbach, 2018). It requires as input a linear mixed
model generated with function lme of package nlme (Pinheiro et al., 2021). The
simple random sample of size 200 selected before is used to fit the linear mixed
model, with lnSWIR2 as a fixed effect, i.e., the effect of lnSWIR2 on the mean
AGB. The random effect is added by assigning another formula to argument
random. The formula ~ 1 | Ecoregion means that the intercept is treated as a
random variable and that it varies among the ecoregions. This linear mixed
model is referred to as a random intercept model: the intercepts are allowed to
differ among the small domains, whereas the effects of the covariates, lnSWIR2
in our case, is equal for all domains.
Tibble grdAmazonia is converted to a data.frame to avoid problems with function
eblup.mse.f.wrap hereafter. A simple random sample with replacement of size
200 is selected.

grdAmazonia <- grdAmazonia %>%

as.data.frame() %>%
mutate(x1 = x1 /1000,
x2 = x2 / 1000)
set.seed(314)
mysample <- grdAmazonia %>%
dplyr::select(x1, x2, AGB, lnSWIR2, Ecoregion) %>%
slice_sample(n = n, replace = TRUE)

library(nlme)
library(JoSAE)
lmm_AGB <- lme(fixed = AGB ~ lnSWIR2, data = mysample, random = ~ 1 | Ecoregion)

The fixed effects of the linear mixed model can be extracted with function
fixed.effects.

fixed_lmm <- fixed.effects(lmm_AGB)

The fixed effects of the linear mixed model differ somewhat from the fixed
effects in the simple linear regression model (fixed_lm):
14.3 Model-based prediction 277

fixed_lm fixed_lmm
(Intercept) 1778.1959 1667.9759
lnSWIR2 -241.1567 -225.6561

The random effect can be extracted with function random.effects.

random.effects(lmm_AGB)

(Intercept)
Cerrado 21.439891
Guianan highland moist forests 6.397816
Guianan lowland moist forests 5.547995
Gurupa varzea -52.985839
Madeira-Tapajos moist forests 27.479921
Marajo varzea -50.587786
Maranhao Babassu forests -1.702322
Mato Grosso tropical dry forests 18.812207
Monte Alegre varzea -12.201148
Purus-Madeira moist forests -9.320683
Tapajos-Xingu moist forests 28.760508
Tocantins/Pindare moist forests -8.940962
Uatuma-Trombetas moist forests 16.165017
Xingu-Tocantins-Araguaia moist forests 11.135385

The random intercepts are added to the fixed intercept; the coefficient of
lnSWIR2 is the same for all ecoregions:

coef(lmm_AGB)

(Intercept) lnSWIR2
Cerrado 1689.416 -225.6561
Guianan highland moist forests 1674.374 -225.6561
Guianan lowland moist forests 1673.524 -225.6561
Gurupa varzea 1614.990 -225.6561
Madeira-Tapajos moist forests 1695.456 -225.6561
Marajo varzea 1617.388 -225.6561
Maranhao Babassu forests 1666.274 -225.6561
Mato Grosso tropical dry forests 1686.788 -225.6561
Monte Alegre varzea 1655.775 -225.6561
Purus-Madeira moist forests 1658.655 -225.6561
Tapajos-Xingu moist forests 1696.736 -225.6561
Tocantins/Pindare moist forests 1659.035 -225.6561
Uatuma-Trombetas moist forests 1684.141 -225.6561
Xingu-Tocantins-Araguaia moist forests 1679.111 -225.6561
278 14 Sampling for estimating parameters of domains

The fitted model can now be used to predict the means of the ecoregions as
follows. As a first step, a data frame must be defined, with the size and the
population mean of the covariate lnSWIR2 per domain. This data frame is
passed to function eblup.mse.f.wrap with argument domain.data. This function
computes the model-based prediction, as well as the regression estimator
(Equation (14.7)) and the synthetic estimator (Equation (14.13)) and their
variances. The model-based predictor is the variable EBLUP in the output data
frame. For the model-based predictor, two standard errors are computed, see
Breidenbach and Astrup (2012) for details.

N_eco <- tapply(grdAmazonia$AGB, INDEX = grdAmazonia$Ecoregion, FUN = length)

df_eco <- data.frame(Ecoregion = ecoregions, N = N_eco, lnSWIR2 = mx_eco_pop)
res <- eblup.mse.f.wrap(domain.data = df_eco, lme.obj = lmm_AGB)
df <- data.frame(Ecoregion = res$domain.ID, mb = res$EBLUP,
se.1 = res$EBLUP.se.1, se.2 = res$EBLUP.se.2)

Table 14.7 shows the model-based predictions and the estimated standard
errors of the mean AGB of the ecoregions, obtained with the random intercept
model.

TABLE 14.7: Model-based predictions of the mean AGB (109 kg ha-1 ) of

ecoregions in Eastern Amazonia, for simple random sample without replacement
of size 200, obtained with the random intercept model and lnSWIR2 as a
predictor.

Ecoregion AGB se.1 se.2

Cerrado 101.9 11.4 11.5
Guianan highland moist forests 271.1 25.8 26.4
Guianan lowland moist forests 269.1 25.8 26.4
Gurupa varzea 155.3 36.0 31.8
Madeira-Tapajos moist forests 292.8 9.3 9.3
Marajo varzea 169.3 13.6 13.2
Maranhao Babassu forests 113.1 14.1 14.4
Mato Grosso tropical dry forests 129.6 22.9 23.1
Monte Alegre varzea 218.9 22.2 22.7
Purus-Madeira moist forests 229.0 22.2 22.7
Tapajos-Xingu moist forests 277.2 6.7 6.7
Tocantins/Pindare moist forests 159.6 7.4 7.5
Uatuma-Trombetas moist forests 270.2 6.0 6.0
Xingu-Tocantins-Araguaia moist forests 222.9 7.5 7.6
se.1 and se.2 are standard errors, for explanation see text.
14.3 Model-based prediction 279

Note that with this model no predictions of the mean AGB are obtained for
the unsampled ecoregions. This is because the random intercept 𝑣u� cannot be
predicted in the absence of data, see Equations (14.16) and (14.17).

14.3.2 Geostatistical model

In a geostatistical model, there is only one random variable, the residual of the
model-mean, not two random variables as in the random intercept model. See
Equation (21.2) for a geostatistical model with a constant mean and Equation
(21.16) for a model with a mean that is a linear combination of covariates.
In a geostatistical model, the covariance of the residuals of the mean at two
locations is modelled as a function of the distance (and direction) of the points.
Instead of the covariance, often the semivariance is modelled, i.e., half the
variance of the difference of the residuals at two locations, see Chapter 21 for
details.
The simple random sample of size 200 selected before is used to estimate the
regression coefficients for the mean, an intercept, and a slope coefficient for
lnSWIR2, and besides the parameters of a spherical semivariogram model
for the residuals of the mean. The two regression coefficients and the three
semivariogram parameters are estimated by restricted maximum likelihood
(REML), see Subsection 21.5.2. This estimation procedure is also used in
function lme to fit the random intercept model. Here, function likfit of package
geoR (Ribeiro Jr et al., 2020) is used to estimate the model parameters. First,
a geoR object must be generated with function as.geodata.

library(geoR)
dGeoR <- as.geodata(mysample, header = TRUE,
coords.col = c("x1", "x2"), data.col = "AGB", covar.col = "lnSWIR2")
vgm_REML <- likfit(geodata = dGeoR, trend = ~ lnSWIR2,
cov.model = "spherical",
ini.cov.pars = c(600, 600), nugget = 1500,
lik.method = "REML", messages = FALSE)

The estimated intercept and slope are 1,744 and -236.5, respectively. The
estimated semivariogram parameters are 1,623 (109 kg ha-1 )2 , 700 (109 kg
ha-1 )2 , and 652 km for the nugget, partial sill, and range, respectively. These
model parameters are used to predict AGB for all units in the population,
using function krige of package gstat (Pebesma, 2004). The REML estimates
of the semivariogram parameters are passed to function vgm with arguments
nugget, psill, and range. The coordinates of the sample are shifted to a random
point within a 1 km × 1 km grid cell. This is done to avoid the coincidence
of a sampling point and a prediction point, which leads to an error message
when predicting AGB at the nodes of the grid.
280 14 Sampling for estimating parameters of domains

library(gstat)
mysample$x1 <- jitter(mysample$x1, amount = 0.5)
mysample$x2 <- jitter(mysample$x2, amount = 0.5)
coordinates(mysample) <- ~ x1 + x2
vgm_REML_gstat <- vgm(model = "Sph",
nugget = vgm_REML$nugget, psill = vgm_REML$sigmasq, range = vgm_REML$phi)
coordinates(grdAmazonia) <- ~ x1 + x2
predictions <- krige(
formula = AGB ~ lnSWIR2,
locations = mysample,
newdata = grdAmazonia,
model = vgm_REML_gstat,
debug.level = 0) %>% as("data.frame")

The first six rows of predictions are shown below.

x1 x2 var1.pred var1.var
1 -6628.193 188.4642 295.8800 1987.472
2 -6627.193 188.4642 294.8188 1986.125
3 -6626.193 188.4642 287.2958 1984.268
4 -6625.193 188.4642 280.1586 1982.756
5 -6624.193 188.4642 290.3905 1982.185
6 -6623.193 188.4642 299.5091 1982.027

Besides a prediction (variable var1.pred), for every population unit the variance
of the prediction error is computed (var1.var). The unitwise predictions can be
averaged across all units of an ecoregion to obtain a model-based prediction of
the mean of that ecoregion.

AGBpred_unit <- predictions$var1.pred

mz_eco_mb <- tapply(AGBpred_unit, INDEX = grdAmazonia$Ecoregion,
FUN = mean) %>% round(1)

A difficulty is the computation of the standard error of these model-based

predictions of the ecoregion mean. We cannot simply sum the unitwise variances
and divide the sum by the squared number of units, because the prediction
errors of units with a mutual distance smaller than the estimated range of
the spherical semivariogram are correlated. A straightforward approach to
obtain the standard error of the predicted mean is geostatistical simulation.
A large number of maps are simulated, conditional on the selected sample.
For an infinite number of maps, the “average map”, i.e., the map obtained by
averaging for each unit all simulated values of that unit, is equal to the map
with predicted AGB. For each simulated map, the average of the simulated
values across all units of an ecoregion is computed. This results in as many
averages as we have simulated maps. The variance of the averages of an
14.3 Model-based prediction 281

ecoregion is an estimate of the variance of the predicted mean of that ecoregion.

To reduce computing time, a 5 km × 5 km subgrid of grdAmazonia is used in
the geostatistical simulation.

grdAmazonia <- read_rds(file = "results/grdAmazonia_5km.rds")

grdAmazonia <- grdAmazonia %>%
mutate(lnSWIR2 = log(SWIR2))
nsim <- 1000
coordinates(grdAmazonia) <- ~ x1 + x2
simulations <- krige(
formula = AGB ~ lnSWIR2,
locations = mysample,
newdata = grdAmazonia,
model = vgm_REML_gstat,
nmax = 100, nsim = nsim,
debug.level = 0) %>% as("data.frame")
grdAmazonia <- as_tibble(grdAmazonia)
AGBsim_eco <- matrix(nrow = length(ecoregions), ncol = nsim)
for (i in 1:nsim) {
AGBsim_eco[, i] <- tapply(simulations[, i + 2],
INDEX = grdAmazonia$Ecoregion, FUN = mean)
}

Similar to the synthetic estimator, for all ecoregions an estimate of the mean
AGB is obtained, also for the unsampled ecoregions (Table 14.8). The model-
based prediction is strongly correlated with the synthetic estimate (Figure
14.2).
The most striking difference is the standard error. The standard errors of the
synthetic estimator range from 3.7 to 7.1 (Table 14.6), whereas the standard
errors of the geostatistical predictions range from 6.2 to 28.1. However, these
two standard errors are fundamentally different and should not be compared.
The standard error of the synthetic estimator is a sampling standard error, i.e.,
it quantifies the variation of the estimated mean of an ecoregion over repeated
random sampling with the sampling design, in this case simple random sampling
of 200 units. The model-based standard error is not a sampling standard error
but a model standard error, which expresses our uncertainty about the means
of the domains due to our imperfect knowledge of the spatial variation of
AGB. Given the observations of AGB at the selected sample, the map with the
covariate lnSWIR2, and the estimated semivariogram model parameters, we
are uncertain about the exact value of AGB at unsampled units. No samples are
considered other than the one actually selected. For the fundamental difference
between design-based, model-assisted, and model-based estimates of means,
refer to Section 1.2 and Chapter 26.
282 14 Sampling for estimating parameters of domains

TABLE 14.8: Model-based predictions of the mean AGB (109 kg ha-1 )

of ecoregions in Eastern Amazonia, using a simple random sample without
replacement of size 200, obtained with the geostatistical model and lnSWIR2
as a predictor for the mean.

Ecoregion AGB se
Amazon-Orinoco-Southern Caribbean mangroves 191.1 22.4
Cerrado 96.8 12.1
Guianan highland moist forests 299.1 28.1
Guianan lowland moist forests 279.3 16.4
Guianan savanna 166.1 10.2
Gurupa varzea 201.2 15.6
Madeira-Tapajos moist forests 287.0 8.6
Marajo varzea 194.6 11.3
Maranhao Babassu forests 121.6 12.0
Mato Grosso tropical dry forests 121.4 11.7
Monte Alegre varzea 241.4 10.8
Purus-Madeira moist forests 229.0 24.3
Tapajos-Xingu moist forests 273.8 8.2
Tocantins/Pindare moist forests 163.0 6.8
Uatuma-Trombetas moist forests 265.9 6.4
Xingu-Tocantins-Araguaia moist forests 220.0 6.2
se: standard error of predicted mean.

It makes more sense to compare the two model-based predictions, the random
intercept model predictions and the geostatistical predictions, and their stan-
dard errors. Figure 14.3 shows that the two model-based predictions are very
similar.
For four ecoregions, the standard errors of the geostatistical model predictions
are much smaller than those of the random intercept model predictions (Figure
14.4). These are ecoregions with small sample sizes.

If a different semivariogram model were used, both the predicted means

per ecoregion and the standard errors would be different. Especially the
variance is sensitive to the semivariogram. For this reason, the model-based
predictions are also referred to as model-dependent predictions, see Chapter
26.
14.4 Supplemental probability sampling of small domains 283

FIGURE 14.2: Scatter plot of the model-based prediction and the synthetic
estimate of the mean AGB (109 kg ha-1 ) of ecoregions in Eastern Amazonia.
The solid line is the 1:1 line.

14.4 Supplemental probability sampling of small

domains
The sample size in small domains of interest can be so small that no reliable
statistical estimate of the mean or total of these domains can be obtained. In
this case, we may decide to collect a supplemental sample from these domains.
It is convenient to use these domains as strata in supplemental probability
sampling, so that we can control the sample sizes in the strata. If we can safely
assume that the study variable at the units of the first sample are not changed,
there is no need to revisit these units; otherwise, we must revisit them to
observe the current values.
There are two approaches for using the two probability samples to estimate
the population mean or total of a small domain (Grafström et al., 2019). In
the first approach, the two samples are combined, and then the merged sample
284 14 Sampling for estimating parameters of domains

FIGURE 14.3: Scatter plot of model-based predictions of the mean AGB

(109 kg ha-1 ) of ecoregions in Eastern Amazonia, obtained with the random
intercept model and the geostatistical model. The solid line is the 1:1 line.

is used to estimate the population mean or total. In the second approach, the
samples are not combined, but the two estimates from the separate samples.
In this section only the first approach is illustrated with a simple situation in
which the two samples are easily combined. Refer to Grafström et al. (2019) for
a more general approach of how multiple probability samples can be combined.
Suppose that the original sample is a simple random sample from the entire
study area. A supplemental sample is selected from small domains, i.e., domains
that have few selected units only. For a given small domain, the first sample
is supplemented by selecting a simple random sample from the units not yet
selected in the first sample. The size of the supplemental sample of a domain
depends on the number of units of that domain in the first sample. The first
sample is supplemented so that the total sample size of that domain is fixed.
In this case, the combined sample of a domain is a simple random sample from
that domain, so that the usual estimators for simple random sampling can be
used to estimate the domain mean or total and its standard error.
This sampling strategy is illustrated with Eastern Amazonia. A simple random
sample without replacement of 400 units is selected.
14.4 Supplemental probability sampling of small domains 285

FIGURE 14.4: Scatter plot of the estimated standard error of model-based

predictions of the mean AGB (109 kg ha-1 ) of ecoregions obtained with the
random intercept model and the geostatistical model, using a simple random
sample without replacement of size 200 from Eastern Amazonia. The numbers
refer to the number of sampling units in an ecoregion. The solid line is the 1:1
line.

grdAmazonia$Biome <- as.factor(grdAmazonia$Biome)

biomes <- c("Mangrove", "Forest_dry", "Grassland", "Forest_moist")
levels(grdAmazonia$Biome) <- biomes
n1 <- 400
set.seed(123)
units_1 <- sample(nrow(grdAmazonia), size = n1, replace = FALSE)
mysample_1 <- grdAmazonia[units_1, c("AGB", "Biome")]
print(n1_biome <- table(mysample_1$Biome))

Mangrove Forest_dry Grassland Forest_moist

2 9 26 363

The selected units are removed from the sampling frame. For each of the
three small biomes, Mangrove, Forest_dry, and Grassland, the size of the
286 14 Sampling for estimating parameters of domains

supplemental sample is computed so that the total sample size becomes 40.
The supplemental sample is selected by stratified simple random sampling
without replacement, using the small biomes as strata (Chapter 4).

units_notselected <- grdAmazonia[-units_1, ]

Biomes_NFM <- units_notselected[units_notselected$Biome != "Forest_moist", ]
n_biome <- 40
n2_biome <- rep(n_biome, 3) - n1_biome[-4]
ord <- unique(Biomes_NFM$Biome)
units_2 <- sampling::strata(Biomes_NFM, stratanames = "Biome",
size = n2_biome[ord], method = "srswor")
mysample_2 <- getdata(Biomes_NFM, units_2)
mysample_2 <- mysample_2[c("AGB", "Biome")]

The two samples are merged, and the means of the domains are estimated by
the sample means.

mysample <- rbind(mysample_1, mysample_2)

print(mz_biome <- tapply(mysample$AGB, INDEX = mysample$Biome, FUN = mean))

Mangrove Forest_dry Grassland Forest_moist

112.8000 122.1750 123.6250 233.6749

Finally, the standard error is estimated, accounting for sampling without

replacement from a finite population (Equation (3.15)).

N_biome <- table(grdAmazonia$Biome)

fpc <- (1 - n_biome / N_biome)
S2z_biome <- tapply(mysample$AGB, INDEX = mysample$Biome, FUN = var)
print(se_mz_biome <- sqrt(fpc * (S2z_biome / n_biome)))

Mangrove Forest_dry Grassland Forest_moist

5.227434 8.213407 11.709257 13.937597

This sampling approach and estimation are repeated 10,000 times, i.e., a
simple random sample without replacement of size 400 is selected 10,000 times
from Eastern Amazonia, and the samples from the three small domains are
supplemented so that the total sample sizes in these domains become 40. In
two out of the 10,000 samples, the size of the first sample in one of the domains
exceeded 40 units. These two samples are discarded. Ideally, these samples are
not discarded, but their sizes in the small domains are reduced to 40 units,
which are then used to estimate the means of the domains.
For all three small domains, the average of the 10,000 estimated means of AGB
is about equal to the true mean (Table 14.9). Also the mean of the 10,000
estimated standard errors is very close to the standard deviation of the 10,000
14.4 Supplemental probability sampling of small domains 287

TABLE 14.9: Summary statistics of 10,000 estimated means of AGB (109 kg

ha-1 ) of small domains (biomes) in Eastern Amazonia, estimated by combining
a simple random sample without replacement of size 400 from all units and a
supplemental stratified simple random sample without replacement from the
units in the small domains not included in the simple random sample. The
total sample size per small domain is 40.

Mangrove Dry forest Grassland

Average of estimated means 122.476 113.983 114.462
True means 122.494 113.931 114.390
Standard deviation of estimated means 6.186 7.563 10.986
Average of estimated standard errors 6.169 7.450 10.946
Coverage rate 95% 0.950 0.937 0.945
Coverage rate 90% 0.903 0.888 0.896
Coverage rate 80% 0.801 0.792 0.799

estimated means. The coverage rates of 95, 90, and 80% confidence intervals
are about equal to the nominal coverage rates.
This simple approach is feasible because at the domain level the two merged
samples are a simple random sample. This approach is also applicable when the
first sample is a stratified simple random sample from the entire population,
and the supplemental sample is a stratified simple random sample from a small
domain using as strata the intersections of the strata used in the first phase
and that domain.
15
Repeated sample surveys for monitoring
population parameters

The previous chapters are all about sampling to estimate population parameters
at a given time. The survey is done in a relatively short period of time, so
that we can safely assume that the study variable has not changed during
that period. This chapter is about repeating the sample survey two or more
times, to estimate, for instance, a temporal change in a population parameter.
Sampling locations are selected by probability sampling, by any design type.
In most cases, sampling times are not selected randomly, but purposively. For
instance, to monitor the carbon stock in the soil of a country, we may decide
to repeat the survey after five years, in the same season of the year as the first
survey.

15.1 Space-time designs

An overview of space-time designs is presented by de Gruijter et al. (2006).
Five of these designs are schematically shown in Figure 15.1. With repeated
sampling in two-dimensional space, there are three dimensions: two spatial
dimensions and one time dimension. Sampling locations shown in Figure 15.1
are selected by simple random sampling, but this is not essential for the
space-time designs.
In the static-synchronous (SS) design, referred to as a pure panel by Fuller
(1999), all sampling locations selected in the first survey are revisited in all
subsequent surveys. On the contrary, in an independent synchronous (IS)
design, a probability sample is selected in each survey independently from the
samples selected in the previous surveys. The serially alternating (SA) design
is a compromise between an SS and an IS design. The sample selected in the
first survey is revisited in the third survey. The sample of the second survey is
selected independently from the sample of the first survey, and these locations
are revisited in the fourth survey. In this case, the period of revisits is two,
i.e., two sampling intervals between consecutive surveys, but this can also be
increased. For instance, with a period of three, three samples are selected,

289
290 15 Repeated sample surveys for monitoring population parameters

FIGURE 15.1: Space-time designs for monitoring population parameters.

The sampling locations in 2D are plotted in one dimension, along the horizontal
axis. A selected unit along this axis actually represents a sampling location in
2D. Twenty sampling locations are selected by simple random sampling. SS:
static-synchronous design; IS: independent synchronous design; SA: serially
alternating design; SP: supplemented panel design; RP: rotating panel design.
15.2 Space-time population parameters 291

independently from each other, for the first three surveys, and these samples
are revisited in subsequent surveys.
Two other compromise designs are a supplemented panel (SP) design and a
rotating panel (RP) design. In an SP design only a subset of the sampling
locations of the first survey is revisited in the subsequent surveys. These are
the permanent sampling locations observed in all subsequent surveys. The
permanent sampling locations are supplemented by samples that are selected
independently from the samples in the previous surveys. In Figure 15.1, half
of the sampling locations (ten locations) is permanent (panel a), i.e., revisited
in all surveys, but the proportion of permanent sampling locations can be
smaller or larger and, if prior information on the variation in space and time
is available, even can be optimised for estimating the current mean. Also in
an RP design, sampling units of the previous survey are partially replaced by
new units. The difference with an SP design is that there are no permanent
sampling units, i.e., no units observed in all surveys. All sampling units are
sequentially rotated out and in again at the subsequent sampling times.
In Figure 15.1 the shape and colour of the symbols represent a panel. A panel
is a group of sampling locations that is observed in the same surveys. In the SS
design, there is only one panel. All locations are observed in all surveys, so all
locations are in the same panel. In the IS design, there are as many panels as
there are surveys. In the SA design with a period of two, the number of panels
equals the number of surveys divided by two. In these three space-time designs
(SS, IS, and SA), all sampling locations of a given survey are in the same panel.
This is not the case in the SP and RP designs. In Figure 15.1 in each survey,
two panels are observed. In the SP sample, there is one panel of permanent
sampling locations (pure panel part of sample) and another panel of swarming
sampling locations observed in one survey only. In the RP sample of Figure
15.1, the sampling locations are observed in two consecutive surveys; however,
this number can be increased. For instance, in an ‘in-for-three’ rotational
sample (McLaren and Steel, 2001) the sampling locations stay in the sample
for three consecutive surveys. The number of panels per sampling time is then
three. Also, similar to an SA design, in an IS, SP, and RP design we may decide
after several surveys to stop selecting new sampling locations and to revisit
existing locations. The concept of panels is needed hereafter in estimating
space-time population parameters.

15.2 Space-time population parameters

The data of repeated surveys can be used to estimate various parameters of the
space-time universe. I will focus on the mean as a parameter, but estimation
292 15 Repeated sample surveys for monitoring population parameters

of a total or proportion is straightforward, think for instance of the total

amount of greenhouse gas emissions from the soil in a given study area during
some period. This chapter shows how to estimate the current mean, i.e., the
population mean (spatial mean) in the last survey, the change of the mean
between two surveys, the temporal trend of the mean, and the space-time mean.
The current mean need not be defined here as only one survey (one sampling
time) is involved in this parameter, so that the definition in Subsection 1.1.1
is also relevant here.
The change of the mean is defined as the spatial mean at a given survey minus
this mean at an earlier survey. For finite populations, the definition is

u� u�
1 1 u�
̄ =
𝑑u�u� ( ∑ 𝑧u� (𝑡u� ) − ∑ 𝑧u� (𝑡u� )) = ∑𝑑 , (15.1)
𝑁 u�=1 u�=1
𝑁 u�=1 u�u�u�

with 𝑑u�u�u� the change of the study variable in the period between time 𝑡u� and
𝑡u� for unit 𝑘. For infinite populations, the sums are replaced by integrals:

1 1
̄ =
𝑑u�u� (∫ 𝑧(𝐬, 𝑡u� ) d𝐬 − ∫ 𝑧(𝐬, 𝑡u� ) d𝐬) = ∫ 𝑑u�u� (𝐬) , (15.2)
𝐴 u�∈u� u�∈u�
𝐴 u�∈u�

with 𝑑u�u� (𝐬) the change of the study variable in the period between time 𝑡u�
and 𝑡u� at location 𝐬.
With more than two surveys, an interesting population parameter is the average
change per time unit of the mean, referred to as the temporal trend of the
spatial mean. It is defined as a linear combination of the spatial means at the
sampling times (Breidt and Fuller, 1999):

u�
𝑏 = ∑ 𝑤u� 𝑧u�̄ , (15.3)
u�=1

with 𝑅 the number of sampling times, 𝑧u�̄ the spatial mean at time 𝑡u� , and
weights 𝑤u� equal to

𝑡u� − 𝑡̄
𝑤u� = u�
, (15.4)
∑u�=1 (𝑡u� − 𝑡)̄ 2

with 𝑡 ̄ the mean of the sampling times.

The temporal trend is defined as a parameter of a space-time population,

not as a parameter of a time-series model.
15.3 Design-based generalised least squares estimation of spatial means 293

A space-time mean can be defined as the average of the spatial means at

the sampling times. In this definition, the temporal universe is discrete and
restricted to the sampling times. The target universe consists of a finite set of
spatial populations:

1 u�
𝑧u�̄ = ∑𝑧 ̄ . (15.5)
𝑅 u�=1 u�

Alternatively, a space-time mean for a continuous temporal universe 𝒯 is

defined as

1
𝑧u�̄ = ∫ 𝑧̄ , (15.6)
𝑇 u�∈u� u�

with 𝑇 the length of the monitoring period and 𝑧u�̄ the spatial mean at time 𝑡.

15.3 Design-based generalised least squares estimation

of spatial means
Rotational sampling has a long tradition in agronomy, forestry, and in social
studies. Early papers on how sample data from previous times can be used
to increase the precision of estimates of the current mean are Jessen (1942),
Patterson (1950), Ware and Cunia (1962), and Woodruff (1963). Gurney and
Daly (1965) developed general theory for these estimators, which I will present
now.
In overlapping samples such as an SP and an RP sample, we may define one
estimate of a spatial mean per ‘panel’. These panel-specific estimates of the
mean at a given sampling time, based on observations at that time only, are
referred to as elementary estimates.
The essence of the estimation method described in this section is to estimate
the spatial mean at a given time point as a weighted average of the elementary
estimates of all time points, with the weights determined by the variances and
covariances of the elementary estimates. Collecting all elementary estimates of
the spatial means of the different sampling times in vector 𝐳,̂ we can write

𝐳̂ = 𝐗𝐳 + 𝐞 , (15.7)

with 𝐳 the vector of true spatial means 𝑧(𝑡 ̄ u� ) at the 𝑅 sampling times,
̄ 1 ), … , 𝑧(𝑡
𝐗 the (𝑃 × 𝑅) design matrix with zeroes and ones that selects the appropriate
294 15 Repeated sample surveys for monitoring population parameters

elements from 𝐳 (𝑃 is the total number of elementary estimates), and 𝐞 the

𝑃-vector of sampling errors with variance-covariance matrix 𝐂. With unbiased
elementary estimators, the expectation of 𝐞 is a vector with zeroes.
With an SS, IS, and SA design, the design matrix 𝐗 is the identity matrix of
size 𝑅, i.e., an 𝑅 × 𝑅 square matrix with ones on the diagonal and zeroes in all
off-diagonal entries. For the SP design of Figure 15.1, the design matrix 𝐗 is

1 0 0 0
⎡0 1 0 0⎤
⎢ ⎥
⎢0 0 1 0⎥
⎢0 0 0 1⎥
𝐗=⎢ . (15.8)
1 0 0 0⎥
⎢ ⎥
⎢0 1 0 0⎥
⎢0 0 1 0⎥
⎣0 0 0 1⎦

The first four rows of this matrix are associated with the elementary estimates
of the spatial means at the four sampling times, estimated from panel a, the
panel with permanent sampling locations. Hereafter, this panel is referred
to as the static-synchronous subsample. The remaining rows correspond to
the elementary estimates from the other four panels, the swarming locations,
hereafter referred to as the independent-synchronous subsamples. For the RP
design of Figure 15.1, the design matrix equals

1 0 0 0
⎡1 0 0 0⎤
⎢ ⎥
⎢0 1 0 0⎥
⎢0 1 0 0⎥
𝐗=⎢ . (15.9)
0 0 1 0⎥
⎢ ⎥
⎢0 0 1 0⎥
⎢0 0 0 1⎥
⎣0 0 0 1⎦

The first two rows correspond to the two elementary estimates of the mean
at time 𝑡1 from panels a and b, the third and fourth rows correspond to the
elementary estimate at time 𝑡2 from panels b and c, respectively, etc.
The minimum variance linear unbiased estimator (MVLUE) of the spatial
means at the different times is the design-based generalised least squares (GLS)
estimator (Binder and Hidiroglou, 1988):

𝐳GLS
̂ = (𝐗T 𝐂−1 𝐗)−1 𝐗T 𝐂−1 𝐳̂ . (15.10)

To define matrix 𝐂 for the SP design in Figure 15.1, let 𝑧u�u� ̄̂ denote the estimated
mean at time 𝑡u� , 𝑗 = 1, 2, 3, 4 in subsample 𝑝, 𝑝 ∈ (𝑎, 𝑏, 𝑐, 𝑑, 𝑒), with panel 𝑎 the
15.3 Design-based generalised least squares estimation of spatial means 295

permanent sampling locations (SS subsample) and panels 𝑏, 𝑐, 𝑑, 𝑒 the swarming

sampling locations (IS subsamples). If the eight elementary estimates in 𝐳̂
are ordered as (𝑧1u�
̄̂ , 𝑧2u�
̄̂ , 𝑧3u�
̄̂ , 𝑧4u�
̄̂ , 𝑧1u�
̄̂ , 𝑧2u�
̄̂ , 𝑧3u� ̄̂ ), the variance-covariance
̄̂ , 𝑧4u�
matrix 𝐂 equals

𝑉1u� 𝐶1,2 𝐶1,3 𝐶1,4 0 0 0 0

⎡𝐶 𝑉2u� 𝐶2,3 𝐶2,4 0 0 0 0 ⎤
⎢ 2,1 ⎥
⎢𝐶3,1 𝐶3,2 𝑉3u� 𝐶3,4 0 0 0 0 ⎥
⎢𝐶 𝐶4,2 𝐶4,3 𝑉4u� 0 0 0 0 ⎥
𝐂 = ⎢ 4,1 . (15.11)
0 0 0 0 𝑉1u� 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 𝑉2u� 0 0 ⎥
⎢ 0 0 0 0 0 0 𝑉3u� 0 ⎥
⎣ 0 0 0 0 0 0 0 𝑉4u� ⎦

The covariances of the elementary estimates of the IS subsamples are zero

because these are estimated from independently selected samples (off-diagonal
elements in lower right (4 × 4) submatrix). Also the covariances of the elemen-
tary estimates from the SS subsample and an IS subsample are zero for the
same reason (upper right submatrix and lower left submatrix). The covariances
of the four elementary estimates from the SS subsample are not zero because
these are estimated from the same set of sampling locations.
Ordering the elementary estimates of the RP sample by the sampling times,
the variance-covariance matrix equals

𝑉1u� 0 0 0 0 0 0 0
⎡ 0 𝑉1u� 𝐶1,2 0 0 0 0 0 ⎤
⎢ ⎥
⎢ 0 𝐶2,1 𝑉2u� 0 0 0 0 0 ⎥
⎢ 0 0 0 𝑉2u� 𝐶2,3 0 0 0 ⎥
𝐂=⎢ . (15.12)
0 0 0 𝐶3,2 𝑉3u� 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 0 𝑉3u� 𝐶3,4 0 ⎥
⎢ 0 0 0 0 0 𝐶4,3 𝑉4u� 0 ⎥
⎣ 0 0 0 0 0 0 0 𝑉4u� ⎦

Only the elementary estimates of the same panel are correlated, for instance
the elementary estimates of the spatial means at times 𝑡1 and 𝑡2 , estimated
from panel b.
For the SA design of Figure 15.1 the variance-covariance matrix equals (there
is only one estimate per time)

𝑉1 0 𝐶1,3 0
⎡ 0 𝑉2 0 𝐶2,4 ⎤
𝐂=⎢ ⎥ . (15.13)
⎢𝐶3,1 0 𝑉3 0 ⎥
⎣ 0 𝐶4,2 0 𝑉4 ⎦
296 15 Repeated sample surveys for monitoring population parameters

In practice, matrix 𝐂 in Equation (15.10) is unknown and is replaced by a

matrix with design-based estimates of the variances and covariances of the
elementary estimators. With simple random sampling with replacement from
finite populations and simple random sampling of infinite populations, the
variance of an elementary estimator of the spatial mean at a given time can
be estimated by Equation (3.14). The covariance of the elementary estimators
of the spatial means at two sampling times, using the data of the same panel,
can be estimated by

̂
𝑆2
𝐶̂u�u� = u�u�
, (15.14)
𝑚

with 𝑚 the number of sampling locations in the panel and 𝑆 ̂ 2

u�u� the estimated
covariance of the study variable at times 𝑡u� and 𝑡u� , estimated by

u�
1
̂
𝑆2
u�u� = ̄̂ )(𝑧u�u�u� − 𝑧u�u�
∑ (𝑧u�u�u� − 𝑧u�u� ̄̂ ) , (15.15)
𝑚 − 1 u�=1

with 𝑧u�u�u� the study variable of unit 𝑘 in panel 𝑝 at time 𝑡u� and 𝑧u�u�
̄̂ the spatial
mean at time 𝑡u� as estimated from panel 𝑝.
The variances and covariances of the GLS estimators of the spatial means at
the 𝑅 sampling times can be estimated by

Cov(𝐳GLS
̂ ) = (𝐗T 𝐂
̂−1 𝐗)−1 . (15.16)

Given the design-based GLS estimates of the spatial means at the different
times, it is an easy job to compute the estimated change of the mean between
two surveys, the estimated temporal trend of the mean, and the estimated
space-time mean.

15.3.1 Current mean

As explained above, with the SS, IS, and SA designs, the design matrix 𝐗 is
the identity matrix of size 𝑅. From this it follows that for these space-time
designs 𝐳GLS
̂ = 𝐳,̂ see Equation (15.10), and Cov(𝐳GLS
̂ ) = Cov(𝐳)̂ = 𝐂. In
words, the GLS estimator of the spatial mean at a given time equals the usual
𝜋 estimator of the mean at that time, and the variance-covariance matrix of
the GLS estimators of the spatial means equals the variance-covariance matrix
of the 𝜋 estimators.
With SP and RP sampling in space-time, there is partial overlap between
the samples at the different times, and so the samples at the previous times
can be used to increase the precision of the estimated mean at the last time
15.3 Design-based generalised least squares estimation of spatial means 297

(current mean). The estimated current mean is simply the last element in 𝐳GLS
̂ .
The estimated variance of the estimator of the current mean is the element
in the final row and final column of the variance-covariance matrix of the
estimated spatial means (Equation (15.16)). These two space-time designs with
partial replacement of sampling locations may yield a more precise estimate of
the current mean than the other three space-time designs.

15.3.2 Change of the spatial mean

The change of the spatial mean between two sampling times can simply be
estimated by subtracting the estimated spatial means at these two times. With
the SS, IS, and SA designs, these means are estimated by the 𝜋 estimators.
With the SP and RP designs, the two spatial means are estimated by the GLS
estimators. The variance of the estimator of the change can be estimated by
the sum of the estimated variances of the spatial mean estimators at the two
sampling times, minus two times the estimated covariance of the two estimators.
The covariance is maximal when all sampling locations are revisited, leading
to the most precise estimate of the change of the spatial mean. With SP and
RP sampling, the covariance of the two spatial mean estimators is smaller, and
so the variance of the change estimator is larger.

15.3.3 Temporal trend of the spatial mean

The temporal trend of the mean is estimated by the weighted average of the
(GLS) estimated means at 𝑡1 , … , 𝑡u� , with weights equal to Equation (15.4):

u�
𝑏̂ = ∑ 𝑤u� 𝑧GLS,u�
̄̂ . (15.17)
u�=1

The variance of the estimator of the temporal trend can be estimated by

̂ 𝐳GLS
𝑉(̂ 𝑏)̂ = 𝐰′ 𝐂( ̂ )𝐰 . (15.18)

Brus and de Gruijter (2011) compared the space-time designs for estimating
the temporal trend of the spatial means, under a first order autoregressive
time-series model of the spatial means. The SS design performed best when
the correlation is strong, say > 0.8. What is the best design depends amongst
others on the strength of the correlation and the number of sampling times. A
safe choice is an SA design. With strong positive correlation, say > 0.9, the
SS design can be a good choice, but remarkably for weak positive correlation
this design performed relatively poorly.
298 15 Repeated sample surveys for monitoring population parameters

For an application of an SP design to estimate the temporal trend of the areal

fractions of several vegetation types, see Brus et al. (2014). Sampling locations
are selected by a design that spreads the locations evenly over the study area
(systematic unaligned sampling, which is a modified version of systematic
sampling).

15.3.4 Space-time mean

The space-time mean, as defined in Equation (15.5), can be estimated by

the average of the (GLS) estimated spatial means at the 𝑅 sampling times.
The variance of this estimator can be estimated by Equation (15.18) using
constant weights, equal to 1/𝑅, for the 𝑅 estimators of the mean. An IS design
yields the most precise estimate of the space-time mean compared to the other
space-time designs when the correlation of collocated measurements of the
study variable is positive (Cochran, 1977). With the other space-time designs,
there is redundant information due to the revisits.
For design-based estimation of the space-time mean of a continuous temporal
universe as defined in Equation (15.6), both the sampling locations and the
sampling times must be selected by probability sampling. Figure 15.2 shows
an SS sample and an IS sample of twenty locations and six times, where both
locations and times are selected by simple random sampling.

FIGURE 15.2: Space-time designs for estimating the space-time mean of a

continuous temporal universe. Both locations and times are selected by simple
random sampling. SS: static-synchronous design; IS: independent synchronous
design.

An IS sample in which both locations and times are selected by probability

sampling can be considered as a two-stage cluster random sample from a
15.4 Case study: annual mean daily temperature in Iberia 299

space-time universe (Chapter 7). The primary sampling units (PSUs) are the
spatial sections of that universe (horizontal lines in Figure 15.2), the secondary
sampling units (SSUs) are the sampling locations. The space-time mean can
therefore be estimated by Equation (7.2), and its variance by Equation (7.7).
For simple random sampling, both in space and in time, and a linear costs
model 𝐶 = 𝑐0 + 𝑐1 𝑛 + 𝑐2 𝑛𝑚 Equations (7.9) and (7.10) can be used to optimise
the number of sampling times (PSUs, 𝑛) and the number of sampling locations
(SSUs, 𝑚) per time. The pooled within-unit variance, 𝑆2w , is in this case the
time-averaged spatial variance of the study variable at a given time, and the
between-unit variance, 𝑆2b , is the variance of the spatial means over time.
Estimation of the space-time mean from an SS sample in which both locations
and times are selected by probability sampling is the same as for an IS sample.
However, estimation of the variance of the space-time mean estimator is more
complicated. For the variance of the estimator of the space-time mean with
an SS space-time design and simple random sampling of both locations and
times, see equation (15.8) in de Gruijter et al. (2006). Due to the two-fold
alignment of the sampling units, no unbiased estimator of the variance is
available. The variance estimator of two-stage cluster sampling can be used to
approximate the variance, but this variance estimator does not account for a
possible temporal correlation of the estimated spatial means, resulting in an
underestimation of the variance.
For an application of an IS design, to estimate the space-time mean of nutrients
(nitrogen and phosphorous) in surface waters, see Brus and Knotters (2008) and
Knotters and Brus (2010). In both applications, sampling times are selected
by stratified simple random sampling, with periods of two months as strata. In
Brus and Knotters (2008), sampling locations are selected by stratified simple
random sampling as well.

15.4 Case study: annual mean daily temperature in

Iberia
The space-time designs are illustrated with the annual mean air temperature
at two metres above the earth surface (TAS) in Iberia for 2004, 2009, 2014,
and 2019 (Subsection 1.3.4).
Sampling locations are selected by simple random sampling. The sample size is
100 locations per survey; so with four surveys we have 400 observations in total,
but not with all space-time designs at 400 different sampling locations. In the
next sections, I will show how a space-time sample with a given space-time
design can be selected, and how the population parameters described in Section
15.2 can be estimated.
300 15 Repeated sample surveys for monitoring population parameters

15.4.1 Static-synchronous design

Selecting an SS space-time sample, using simple random sampling as a spatial

design, is straightforward. We can simply select a single simple random sample
of size 𝑛 = 100. The locations of this sample are observed at all times (Figure
15.3).

grd <- grdIberia %>%

mutate(x = x / 1000,
y = y / 1000)
set.seed(314)
n <- 100
units <- sample(nrow(grd), size = n, replace = TRUE)
mysample_SS <- grd[units, ]

FIGURE 15.3: Static-synchronous space-time sample from Iberia. Locations

are selected by simple random sampling.

The spatial means at the four times can be estimated by the sample means
(Equation (3.2)), the variances of the mean estimators by Equation (3.14).
The covariances of the estimators of the means at two different times can
be estimated by Equation (15.14) with 𝑚 equal to the sample size 𝑛. The
15.4 Case study: annual mean daily temperature in Iberia 301

estimated current mean (the spatial mean at the fourth survey) and the es-
timated standard error of the estimator of the current mean can be simply
extracted from the vector with estimated means and from the matrix with
estimated variances and covariances.

mz <- apply(mysample_SS[, -c(1, 2)], MARGIN = 2, FUN = mean)

C <- var(mysample_SS[, -c(1, 2)]) / n
#current mean
mz_cur_SS <- mz[4]
se_mz_cur_SS <- sqrt(C[4, 4])

The change of the spatial mean from the first to the fourth survey can simply
be estimated by subtracting the estimated spatial mean of the first survey from
the estimated mean of the fourth survey. The standard error of this estimator
can be estimated by the sum of the estimated variances of the two spatial
mean estimators, minus two times the estimated covariance of the two mean
estimators, and finally taking the square root.

d_mz_SS <- mz[4] - mz[1]

se_d_mz_SS <- sqrt(C[4, 4] + C[1, 1] - 2 * C[1, 4])

The same estimates are obtained by defining a weight vector with values -1 and
1 for the first and last element, respectively, and 0 for the other two elements.

w <- c(-1, 0, 0, 1)
d_mz_SS <- t(w) %*% mz
se_d_mz_SS <- sqrt(t(w) %*% C %*% w)

The temporal trend of the spatial means can be estimated much in the same
way, but using a different vector with weights (Equation (15.4)).

t <- seq(from = 2004, to = 2019, by = 5)

w <- (t - mean(t)) / sum((t - mean(t))ˆ2)
mz_trend_SS <- t(w) %*% mz
se_mz_trend_SS <- sqrt(t(w) %*% C %*% w)

The estimated temporal trend equals 0.0403∘ C 𝑦−1 , and the estimated standard
error equals 0.0024794∘ C 𝑦−1 . Using t = 1:4 yields the estimated average change
in annual mean temperature per five years.

t <- 1:4
w <- (t - mean(t)) / sum((t - mean(t))ˆ2)
print(mz_trend_SS <- t(w) %*% mz)
302 15 Repeated sample surveys for monitoring population parameters

[,1]
[1,] 0.2016378

Using a constant weight vector with values 1/4 yields the estimated space-time
mean and the standard error of the space-time mean estimator.

w <- rep(1 / 4, 4)
mz_st_SS <- t(w) %*% mz
se_mz_st_SS <- sqrt(t(w) %*% C %*% w)

15.4.2 Independent synchronous design

To select an IS space-time sample with four sampling times, we simply select

4𝑛 = 400 sampling locations. Figure 15.4 shows the selected IS space-time
sample. A variable identifying the panel of the selected sampling units is added
to the data frame.

units <- sample(nrow(grd), size = 4 * n, replace = TRUE)

mysample_IS <- grd[units, ]
mysample_IS$panel <- rep(c("a", "b", "c", "d"), each = n)

The units of the finite population are selected by simple random sampling
with replacement. As a consequence, there can be partial overlap between
the samples at the different times, i.e., some units are observed at multiple
times. However, this overlap is by chance; it is not coordinated as in an SP
and RP design. By selecting the units with replacement, the estimators of
the spatial means are independent (covariance of estimators equals zero). For
infinite populations such as points in an area, there will be no overlap, so
that the covariance of the estimators equals zero.

The spatial means are estimated with the 𝜋 estimator as there is no partial
overlap of the spatial samples. All covariances of the mean estimators are
zero. Estimation of the space-time parameters is done as before with the SS
space-time sample. Four data frames of 𝑛 rows are first made with the data
observed in a specific panel. The variables of these four data frames are joined
into a single data frame. The spatial means at the four times are then estimated
by the sample means, computed with function apply.

panel_a <- filter(mysample_IS, panel == "a")[, c("TAS2004")]

panel_b <- filter(mysample_IS, panel == "b")[, c("TAS2009")]
panel_c <- filter(mysample_IS, panel == "c")[, c("TAS2014")]
panel_d <- filter(mysample_IS, panel == "d")[, c("TAS2019")]
15.4 Case study: annual mean daily temperature in Iberia 303

FIGURE 15.4: Independent synchronous space-time sample from Iberia.

Locations are selected by simple random sampling.

panel_abcd <- cbind(panel_a, panel_b, panel_c, panel_d)

mz <- apply(panel_abcd, MARGIN = 2, FUN = mean)
C <- var(panel_abcd) / n
C[row(C) != col(C)] <- 0
#current mean
mz_cur_IS <- mz[4]
se_mz_cur_IS <- sqrt(C[4, 4])
#change of mean
w <- c(-1, 0, 0, 1)
d_mz_IS <- t(w) %*% mz
se_d_mz_IS <- sqrt(t(w) %*% C %*% w)
#trend of mean
w <- (t - mean(t)) / sum((t - mean(t))ˆ2)
mz_trend_IS <- t(w) %*% mz
se_mz_trend_IS <- sqrt(t(w) %*% C %*% w)
#space-time mean
w <- rep(1 / 4, 4)
304 15 Repeated sample surveys for monitoring population parameters

mz_st_IS <- t(w) %*% mz

se_mz_st_IS <- sqrt(t(w) %*% C %*% w)

15.4.3 Serially alternating design

To select an SA space-time sample with a period of two, 2𝑛 = 200 sampling

locations are selected. A variable is added to mysample_SA indicating the panel of
the sampling locations. Figure 15.5 shows the selected SA space-time sample.

units <- sample(nrow(grd), size = 2 * n, replace = TRUE)

mysample_SA <- grd[units, ]
mysample_SA$panel <- rep(c("a", "b"), each = n)

FIGURE 15.5: Serially alternating space-time sample from Iberia. Locations

are selected by simple random sampling.

In SA sampling there is no partial overlap of the spatial samples at different

times: there is either no coordinated overlap or full overlap. Spatial means
therefore can be estimated by the usual 𝜋 estimator. Two data frames of 𝑛
15.4 Case study: annual mean daily temperature in Iberia 305

rows are first made with the data observed in a specific panel. The variables of
these two data frames are joined into a single data frame. The spatial means
at the four times are then estimated by the sample means, computed with
function apply.

panel_a <- filter(mysample_SA, panel == "a")[, c("TAS2004", "TAS2014")]

panel_b <- filter(mysample_SA, panel == "b")[, c("TAS2009", "TAS2019")]
panel_ab <- cbind(panel_a[1], panel_b[1], panel_a[2], panel_b[2])
mz <- apply(panel_ab, MARGIN = 2, FUN = mean)

To compute the matrix with estimated variances and covariances of the spa-
tial mean estimators, first the full matrix is computed. Then the estimated
covariances of spatial means of consecutive surveys are replaced by zeroes
as the samples of these consecutive surveys are selected independently from
each other, so that the two estimators are independent. Estimation of the
space-time parameters is done as before with the SS space-time sample.

C <- var(panel_ab) / n
odd <- c(1, 3)
C[row(C) %in% odd & !(col(C) %in% odd)] <- 0
C[!(row(C) %in% odd) & col(C) %in% odd] <- 0

#current mean
mz_cur_SA <- mz[4]
se_mz_cur_SA <- sqrt(C[4, 4] / n)
#change of mean from time 1 to time 4
w <- c(-1, 0, 0, 1)
d_mz_SA <- t(w) %*% mz
se_d_mz_SA <- sqrt(t(w) %*% C %*% w)
#trend of mean
w <- (t - mean(t)) / sum((t - mean(t))ˆ2)
mz_trend_SA <- t(w) %*% mz
se_mz_trend_SA <- sqrt(t(w) %*% C %*% w)
#space-time mean
mz_st_SA <- mean(mz)
w <- rep(1 / 4, 4)
se_mz_st_SA <- sqrt(t(w) %*% C %*% w)

15.4.4 Supplemented panel design

With SP sampling and four sampling times we have five panels: one panel
with fixed sampling locations and four panels with swarming locations (Figure
15.1). Each panel consists of 𝑛/2 sampling locations, so in total 5𝑛/2 = 250
306 15 Repeated sample surveys for monitoring population parameters

locations are selected. A variable indicating the panel is added to the data
frame with the selected sampling locations. Figure 15.6 shows the selected SP
sample.

units <- sample(nrow(grd), size = 5 * n / 2, replace = TRUE)

mysample_SP <- grd[units, ]
mysample_SP$panel <- rep(c("a", "b", "c", "d", "e"), each = n / 2)

FIGURE 15.6: Supplemented panel space-time sample from Iberia. Locations

are selected by simple random sampling. The permanent sampling locations
are indicated by the filled dots, the swarming locations by the other symbols.

With SP sampling, the spatial means are estimated by the design-based GLS
estimator (Equation (15.10)). As a first step, the eight elementary estimates
(two per sampling time) are computed and collected in a vector. Note the
order of the elementary estimates: first the estimated spatial means of 2004,
2009, 2014, and 2019, estimated from the panel with fixed sampling locations,
then the spatial means estimated from the panels with swarming locations.
The design matrix 𝐗 corresponding to this order of elementary estimates is
constructed.
15.4 Case study: annual mean daily temperature in Iberia 307

Ordering the elementary estimates by sampling time is also fine, but then
the design matrix 𝐗 should be adapted to this order.

mz_1 <- tapply(

mysample_SP$TAS2004, INDEX = mysample_SP$panel, FUN = mean)[c(1, 2)]
mz_2 <- tapply(
mysample_SP$TAS2009, INDEX = mysample_SP$panel, FUN = mean)[c(1, 3)]
mz_3 <- tapply(
mysample_SP$TAS2014, INDEX = mysample_SP$panel, FUN = mean)[c(1, 4)]
mz_4 <- tapply(
mysample_SP$TAS2019, INDEX = mysample_SP$panel, FUN = mean)[c(1, 5)]
mz_el <- c(mz_1["a"], mz_2["a"], mz_3["a"], mz_4["a"],
mz_1["b"], mz_2["c"], mz_3["d"], mz_4["e"])
X <- rbind(diag(4), diag(4))

The variances and covariances of the elementary estimates of the panel with
fixed locations (SS subsample) are estimated, as well as the variances of the
elementary estimates of the panels with swarming locations (IS subsample).
The two variance-covariance matrices and a 4 × 4 submatrix with all zeroes
are combined into a single matrix, see matrix (15.11).

mysubsample_SS <- mysample_SP[mysample_SP$panel == "a", ]

C_SS <- var(
mysubsample_SS[, c("TAS2004", "TAS2009", "TAS2014", "TAS2019")]) / (n / 2)
mysubsample_IS <- cbind(mysample_SP$TAS2004[mysample_SP$panel == "b"],
mysample_SP$TAS2009[mysample_SP$panel == "c"],
mysample_SP$TAS2014[mysample_SP$panel == "d"],
mysample_SP$TAS2019[mysample_SP$panel == "e"])
C_IS <- var(mysubsample_IS) / (n / 2)
C_IS[row(C_IS) != col(C_IS)] <- 0
zeroes <- matrix(0, nrow = 4, ncol = 4)
C <- rbind(cbind(C_SS, zeroes), cbind(zeroes, C_IS))

The design-based GLS estimates of the spatial means can be computed as

follows, see Equation (15.10).

XCXinv <- solve(crossprod(X, solve(C, X)))

XCz <- crossprod(X, solve(C, mz_el))
mz_GLS <- XCXinv %*% XCz

Computing the estimated space-time parameters from the design-based GLS

estimated spatial means is straightforward.
308 15 Repeated sample surveys for monitoring population parameters

#current mean
mz_cur_SP <- mz_GLS[4]
se_mz_cur_SP <- sqrt(XCXinv[4, 4])
#change of mean
w <- c(-1, 0, 0, 1)
d_mz_SP <- t(w) %*% mz_GLS
se_d_mz_SP <- sqrt(t(w) %*% XCXinv %*% w)
#trend of mean
w <- (t - mean(t)) / sum((t - mean(t))ˆ2)
mz_trend_SP <- t(w) %*% mz_GLS
se_mz_trend_SP <- sqrt(t(w) %*% XCXinv %*% w)
#space-time mean
w <- rep(1 / 4, 4)
mz_st_SP <- t(w) %*% mz_GLS
se_mz_st_SP <- sqrt(t(w) %*% XCXinv %*% w)

15.4.5 Rotating panel design

Similar to the SP design, with an in-for-two RP design and four sampling times
we have five panels, each consisting of 𝑛/2 sampling locations, so that in total
5𝑛/2 = 250 locations are selected. Figure 15.7 shows the selected space-time
sample.

units <- sample(nrow(grd), size = 5 * n / 2, replace = TRUE)

mysample_RP <- grd[units, ]
mysample_RP$panel <- rep(c("a", "b", "c", "d", "e"), each = n / 2)

Two elementary estimates per sampling time are computed and collected in a
vector. Note that now the elementary estimates are ordered by sampling time.
The design matrix corresponding to this order is constructed.

mz_1 <- tapply(

mysample_RP$TAS2004, INDEX = mysample_RP$panel, FUN = mean)[c(1, 2)]
mz_2 <- tapply(
mysample_RP$TAS2009, INDEX = mysample_RP$panel, FUN = mean)[c(2, 3)]
mz_3 <- tapply(
mysample_RP$TAS2014, INDEX = mysample_RP$panel, FUN = mean)[c(3, 4)]
mz_4 <- tapply(
mysample_RP$TAS2019, INDEX = mysample_RP$panel, FUN = mean)[c(4, 5)]
mz_el <- c(mz_1, mz_2, mz_3, mz_4)
X <- matrix(c(rep(c(1, 0, 0, 0), 2),
rep(c(0, 1, 0, 0), 2),
15.4 Case study: annual mean daily temperature in Iberia 309

FIGURE 15.7: In-for-two rotating panel space-time sample from Iberia.

Locations are selected by simple random sampling. Half of the sampling
locations observed in a given year are also observed in the consecutive survey.

rep(c(0, 0, 1, 0), 2),

rep(c(0, 0, 0, 1), 2)), nrow = 8, ncol = 4, byrow = TRUE)

To construct the variance-covariance matrix of the eight elementary estimates,

first the 2 × 2 matrices with estimated variances and covariances are computed
for panels b, c, and d. These three panels are used to estimate the spatial
means of two consecutive years: the locations of panel b are used to estimate
the spatial means for 2004 and 2009, panel c for 2009 and 2014, and panel d
for 2014 and 2019 (Figure 15.1). These three matrices are copied into the 8 ×
8 matrix, see matrix (15.12). Finally, the estimated variances of the estimators
of the spatial means in 2004 and 2019 are computed and copied to the upper
left and lower right entry of the variance-covariance matrix, respectively.

mysubsample <- mysample_RP[mysample_RP$panel == "b", c("TAS2004", "TAS2009")]

C_b <- var(mysubsample) / (n / 2)
mysubsample <- mysample_RP[mysample_RP$panel == "c", c("TAS2009", "TAS2014")]
310 15 Repeated sample surveys for monitoring population parameters

C_c <- var(mysubsample) / (n / 2)

mysubsample <- mysample_RP[mysample_RP$panel == "d", c("TAS2014", "TAS2019")]
C_d <- var(mysubsample) / (n / 2)
C <- matrix(data = 0, ncol = 8, nrow = 8)
C[2:3, 2:3] <- C_b
C[4:5, 4:5] <- C_c
C[6:7, 6:7] <- C_d
C[1, 1] <- var(mysample_RP[mysample_RP$panel == "a", "TAS2004"]) / (n / 2)
C[8, 8] <- var(mysample_RP[mysample_RP$panel == "e", "TAS2019"]) / (n / 2)

The design-based GLS estimates of the spatial means for the four sampling
times are computed as before, followed by computing the estimated space-time
parameters.

XCXinv <- solve(crossprod(X, solve(C, X)))

XCz <- crossprod(X, solve(C, mz_el))
mz_GLS <- XCXinv %*% XCz
#current mean
mz_cur_RP <- mz_GLS[4]
se_mz_cur_RP <- sqrt(XCXinv[4, 4])
#change of mean
w <- c(-1, 0, 0, 1)
d_mz_RP <- t(w) %*% mz_GLS
se_d_mz_RP <- sqrt(t(w) %*% XCXinv %*% w)
#trend of mean
w <- (t - mean(t)) / sum((t - mean(t))ˆ2)
mz_trend_RP <- t(w) %*% mz_GLS
se_mz_trend_RP <- sqrt(t(w) %*% XCXinv %*% w)
#space-time mean
w <- rep(1 / 4, 4)
mz_st_RP <- t(w) %*% mz_GLS
se_mz_st_RP <- sqrt(t(w) %*% XCXinv %*% w)

15.4.6 Sampling experiment

The random sampling with the five space-time designs and estimation of the
four space-time parameters is repeated 10,000 times. The standard deviations
of the 10,000 estimates of a space-time parameter are shown in Table 15.1.
Note that to determine the standard errors of the estimators of the space-
time parameters for the SS, IS, and SA designs, a sampling experiment is
not really needed. These can be computed without error because we have
exhaustive knowledge of the study variable at all sampling times. However, for
15.4 Case study: annual mean daily temperature in Iberia 311

TABLE 15.1: Standard deviations of 10,000 estimates of space-time param-

eters of annual mean temperature in Iberia, with static-synchronous (SS),
independent synchronous (IS), serially alternating (SA), supplemented panel
(SP), and rotating panel (RP) space-time sampling, and simple random sam-
pling with replacement of 100 locations per sampling time.

Current mean Change of mean Trend of mean Space-time mean

SS 0.2709 0.0343 0.0123 0.2741
IS 0.2653 0.3873 0.1221 0.1354
SA 0.2676 0.3894 0.0787 0.1910
SP 0.1755 0.0481 0.0171 0.1770
RP 0.1769 0.0840 0.0299 0.1744

the SP and RP designs, a sampling experiment is needed to approximate the

design-expectation of the standard error of the estimators of the space-time
parameters. This is because the space-time parameters are derived from the
GLS estimator of the spatial means, and the GLS estimator is a function
of the estimated covariances of the elementary estimates (Equation (15.10)).
Using the true covariance of the elementary estimates in the GLS estimator,
instead of the estimated covariance, leads to an underestimation of the standard
error. Computing the true standard errors of the estimators of the space-time
parameters for the various space-time designs is left as an exercise.
As a reference for the standard deviations in Table 15.1: the current mean
(spatial mean air temperature in 2019) equals 14.59∘ C, the change of the mean
from 2004 to 2019 equals 0.611∘ C, the temporal trend equals 0.198∘ C per five
years, and the space-time mean equals 14.44∘ C.
Estimates of the current mean with the SP and RP designs are much more
precise than with the SS, IS, and SA designs. This is because with designs
SP and RP there is partial overlap with the samples of the other years. If the
study variable at different years is correlated, we can profit from the data of
previous years. In our case, this correlation is very strong:

cor(grd[, -c(1, 2)])

TAS2004 TAS2009 TAS2014 TAS2019

TAS2004 1.0000000 0.9950227 0.9878301 0.9927014
TAS2009 0.9950227 1.0000000 0.9875005 0.9960541
TAS2014 0.9878301 0.9875005 1.0000000 0.9938060
TAS2019 0.9927014 0.9960541 0.9938060 1.0000000

The differences in standard deviations of the estimated current means among

the SS, IS, and SA designs are due to random variation: the true standard
312 15 Repeated sample surveys for monitoring population parameters

errors of the current mean estimator are equal for these three space-time
designs.
The strong correlation also explains that the estimated change of the spatial
mean from 2004 to 2019 with the SS design is much more precise than with
the IS and SA designs. With the IS and SA designs, the sample of 2019 is
selected independently from the sample of 2004, so that the covariance of
the two spatial mean estimators is zero. With the SS design this covariance
is subtracted two times from the sum of the variances of the spatial mean
estimators. The standard error of the estimator of the change of the spatial
mean from 2009 to 2019 with the SA design (not shown in Table 15.1) is
much smaller than that of the change from 2004 to 2019, because the spatial
means at these two times are estimated from the same sample, so that we
profit from the strong positive correlation. The standard error of the change
estimator with the SP design is slightly larger than the standard error with
the SS design, because with this design there is only partial overlap, so that
we profit less from the correlation. The standard error with the RP design
is larger than that of the SP design, because with the RP design there is no
overlap of the samples of 2004 and 2019 (Figure 15.1). Despite the absence of
overlap, the standard error is still considerably smaller than those with the IS
and SA designs because the spatial means of 2004 and 2019 are estimated by
the GLS estimator that uses the data of all years, so that we still profit from
the correlation.
Estimation of the temporal trend of the spatial mean is most precise with the
SS design, closely followed by the SP design, and least precise with the IS
design. This is in agreement with the results of Brus and de Gruijter (2011)
and Brus and de Gruijter (2013). On the contrary, estimation of the space-time
mean is most precise with the IS design and least precise with the SS design.
With strong persistence of the spatial patterns, as in our case, it is not efficient
to observe the same sampling locations at all times when interest is in the
space-time mean. In our case with very strong correlation, the larger the total
number of sampling locations over all sampling times, the smaller the standard
error of the space-time mean estimator. The total number of sampling locations
in this case study are 𝑛 with SS, 4𝑛 with IS, 2𝑛 with SA, and 5𝑛/2 with SP
and RP.

15.5 Space-time sampling with stratified random

sampling in space
In real-world applications, one will often use more efficient sampling designs
than simple random sampling for selecting spatial sampling units. For instance,
15.5 Space-time sampling with stratified random sampling in space 313

in the case study of the previous section, stratified simple random sampling
using climate zones is most likely more efficient than simple random sampling.
To select a space-time sample with stratified simple random sampling as a
spatial design, the selection procedures described above for the five basic types
of space-time design are applied at the level of the strata. Estimation of the
space-time parameters goes along the same lines, using the 𝜋 estimator of the
spatial mean and the estimator of the standard error presented in Chapter 4.
With the SP and RP designs, the covariance of the elementary estimators of
the spatial means at two sampling times using the data of the same panel, can
be estimated by

u� ̂
𝑆2
𝐶̂u�u� = ∑ 𝑤2ℎ u�u�ℎ , (15.19)
ℎ=1
𝑚ℎ

with 𝑆 ̂2
u�u�ℎ the estimated covariance of the study variable at times 𝑡u� and
𝑡u� in stratum ℎ, and 𝑚ℎ the number of sampling locations in the panel in
stratum ℎ.
Interesting new developments are presented by Wang and Zhu (2019) and Zhao
and Grafström (2020).

Exercises

1. Compute for the annual mean temperature data of Iberia the true
standard error of the estimator of (i) the spatial mean in 2019; (ii)
the change of the spatial mean from 2004 to 2019; (iii) the temporal
trend of the spatial mean (average change per five years in the period
from 2004 to 2019); and (iv) the space-time mean, for all five space-
time designs and simple random sampling with replacement of 100
units per time. Use for the designs SP and RP the true covariances
of the elementary estimates in the GLS estimators of the spatial
means.
• Compare the standard errors with the standard deviations in
Table 15.1. Explain why for the designs SP and RP the true
standard errors of the estimators of all space-time parameters
are slightly smaller than the standard deviations in Table 15.1.

2. With an SP and an RP design, the change of the mean can also be

estimated by the difference of the 𝜋 estimates of the spatial means
at the two times, instead of the difference of the two GLS estimates.
The variance of the difference of the 𝜋 estimators of two simple
random samples at different times with partial overlap equals

𝑆2u� 𝑆2 𝑚 𝑆2u�u�
̄ )=
𝑉(𝑑u�u� + u� − 2 , (15.20)
𝑛 𝑛 𝑛2
314 15 Repeated sample surveys for monitoring population parameters

with 𝑆2u� and 𝑆2u� the spatial variance at time 𝑡u� and 𝑡u� , respectively,
𝑆2u�u� the spatial covariance of the study variable at times 𝑡u� and
𝑡u� , 𝑛 the size of the simple random samples, and 𝑚 the number of
units observed at both times. Compute for Iberia the standard error
of the estimator of the change of the mean with Equation (15.20),
using the true spatial variances and covariances for 2004 and 2019,
for simple random samples of size 100 (𝑛 = 100) and a matching
proportion of 0.5 (𝑚 = 50). Compare with the true standard error
of the estimator of the change of the mean using the GLS estimators
computed in the previous exercise. Explain that the standard error
of the 𝜋 estimators is larger than the standard error with the GLS
estimators.
 Part II

Sampling for mapping

16
Introduction to sampling for mapping

16.1 When is probability sampling not required?

This second part of the book deals with sampling for mapping, i.e., for predict-
ing the study variable at the nodes of a fine discretisation grid. For mapping,
a model-based sampling approach is the most natural option. When a statisti-
cal model, i.e., a model containing an error term modelled by a probability
distribution, is used to map the study variable from the sample data, selection
of the sampling units by probability sampling is not strictly needed anymore
in order to make statistical statements about the population, i.e., statements
with quantified uncertainty, see Section 1.2. As a consequence, there is room
for optimising the sampling units by searching for those units that lead to the
most accurate map, for instance, the map with the smallest squared prediction
error averaged over all locations in the mapped study area, see Chapter 25.
As an illustration, consider the following statistical model to be used for
mapping: a simple linear regression model for the study variable to be mapped:

𝑍u� = 𝛽0 + 𝛽1 𝑥u� + 𝜖u� , (16.1)

with 𝑍u� the study variable of unit 𝑘, 𝛽0 and 𝛽1 regression coefficients, 𝑥u� a
covariate for unit 𝑘 used as a predictor, and 𝜖u� the error (residual) at unit 𝑘,
normally distributed with mean zero and a constant variance 𝜎2 . The errors
are assumed independent, so that Cov(𝜖u� , 𝜖u� ) = 0 for all 𝑘 ≠ 𝑗. Figure 16.1
shows a simple random sample without replacement and the sample optimised
for mapping with a simple linear regression model. Both samples are plotted
on a map of the covariate 𝑥.
The optimal sample for mapping with a simple linear regression model contains
the units with the smallest and the largest values of the covariate 𝑥. The
optimal sample shows strong spatial clustering. Spatial clustering is not avoided
because in a simple linear regression model we assume that the residuals are
not spatially correlated. In Chapter 23 I will show that when the residuals
are spatially correlated, spatial clustering of sampling units is avoided. The
standard errors of both regression coefficients are considerably smaller for the

317
318 16 Introduction to sampling for mapping

FIGURE 16.1: Simple random sample and optimal sample for mapping with
a simple linear regression model, plotted on a map of the covariate.

TABLE 16.1: Standard errors and determinant of the variance-covariance

matrix of estimators of the regression coefficients for the simple random sample
(SI) and the optimal sample.

Sampling design se intercept se slope Determinant

SI 1.70 0.118 0.003876
Optimal 1.06 0.050 0.000940

optimal sample (Table 16.1). The joint uncertainty about the two regression
coefficients, quantified by the determinant of the variance-covariance matrix
of the regression coefficient estimators, is also much smaller for the optimal
sample. When we are less uncertain about the regression coefficients, we are also
less uncertain about the regression model predictions of the study variable 𝑧
at points where we have observations of the covariate 𝑥 only. We can conclude
that for mapping with a simple linear regression model, in this example simple
random sampling is not a good option.
Of course, this simple example would only be applicable if we have evidence of
a linear relation between study variable 𝑧 and covariate 𝑥, and in addition if
we are willing to rely on the assumption that the residuals are not spatially
correlated.
16.2 Sampling for simultaneously mapping and estimating means 319

16.2 Sampling for simultaneously mapping and

estimating means
Although probability sampling is not strictly needed for mapping with a
statistical model, in some situations, when feasible, it can still be advantageous
to select a probability sample. If the aim of the survey is to map the study
variable, as well as to estimate the mean or total for the entire study area
or for several subareas, probability sampling can be a good option. Think,
for instance, of sampling for the dual objectives of mapping and at the same
time estimating soil carbon stocks. Although the statistical model used for
mapping can also be used for model-based prediction of the total carbon stocks
in the study area and subareas (Section 14.3), we may prefer to estimate
these totals by design-based or model-assisted inference. The advantage of
design-based and model-assisted estimation of these totals is their validity.
Validity means that an objective assessment of the uncertainty of the estimated
mean or total is warranted, and that the coverage of confidence intervals
is (almost) correct, provided that the sample is large enough to assume an
approximately normal distribution of the estimator and design-unbiasedness
of the variance estimator, see Chapter 26. In design-based estimation, no
model of the spatial variation is used. Therefore, discussions are avoided about
how realistic modelling assumptions are. In model-assisted estimation, these
discussions are irrelevant as well, because we do not rely on these assumptions.
A poor model results in large variances of the estimated mean or total and, as a
consequence, a wide confidence interval, so that the coverage of the confidence
interval is in agreement with the nominal coverage, see Section 26.6 for more
details.
The question then is: what is a suitable probability sampling design for both
aims? First, I would recommend a sampling design with equal inclusion proba-
bilities. This is because in standard model-based inference unequal inclusion
probabilities are not accounted for, which may lead to systematic prediction
errors when small or large values are overrepresented in the sample (Section
26.3).
Second, in case we have subareas of which we would like to estimate the
mean or total (domains of interest), using these subareas as strata in stratified
random sampling makes sense, unless there are too many. This requires that
of all population units (nodes of discretisation grid) we must know to which
subarea it belongs, so that this information can be added to the sampling
frame.
Third, a sampling design spreading the sampling units in geographical space
is attractive as well, for instance through compact geographical stratification
(Section 4.6) or sampling with the local pivotal method (Subsection 9.2.1).
320 16 Introduction to sampling for mapping

We may profit from this geographical spreading if some version of kriging is

used for mapping (Chapter 21). In addition, the geographical spreading may
enhance the coverage of the space spanned by covariates related to the study
variable. Spreading in covariate space can also be explicitly accounted for using
the covariates as spreading variables in the local pivotal method.
As an illustration, I selected a single sample of 500 units from Eastern Amazonia
with the dual aim of mapping aboveground biomass (AGB) as well as estimating
the means of AGB for four biomes. The biomes are used as strata in stratified
random sampling. First, the stratum sample sizes are computed for proportional
allocation, so that the inclusion probabilities are approximately equal for all
population units.

grdAmazonia$Biome <- as.factor(grdAmazonia$Biome)

biomes <- c("Mangrove", "Forest_dry", "Grassland", "Forest_moist")
levels(grdAmazonia$Biome) <- biomes
N_h <- table(grdAmazonia$Biome)
n <- 500
n_h <- round(n * N_h / sum(N_h))
n_h[3] <- n_h[3] + 1
print(n_h)

Mangrove Forest_dry Grassland Forest_moist

2 11 28 459

Biome Forest_moist is by far the largest biome with a sample size of 459
points.
In the next code chunk, a balanced sample is selected with equal inclusion
probabilities, using both the categorical variable biome and the continuous
variable lnSWIR2 as balancing variables (Subsection 9.1.3). The geographical
coordinates are used as spreading variables.

library(BalancedSampling)
grdAmazonia$lnSWIR2 <- log(grdAmazonia$SWIR2)
pi <- n_h / N_h
stratalabels <- levels(grdAmazonia$Biome)
lut <- data.frame(Biome = stratalabels, pi = as.numeric(pi))
grdAmazonia <- merge(x = grdAmazonia, y = lut)
Xbal <- model.matrix(~ Biome - 1, data = grdAmazonia) %>%
cbind(grdAmazonia$lnSWIR2)
Xspread <- cbind(grdAmazonia$x1, grdAmazonia$x2)
set.seed(314)
units <- lcube(Xbal = Xbal, Xspread = Xspread, prob = grdAmazonia$pi)
mysample <- grdAmazonia[units, ]
16.3 Broad overview of sampling designs for mapping 321

Figure 16.2 shows the selected sample.

FIGURE 16.2: Balanced sample of size 500 from Eastern Amazonia, bal-
anced on biome and lnSWIR2, with geographical spreading. Equal inclusion
probabilities are used.

I think this is a suitable sample, both for mapping AGB across the entire
study area, for instance by kriging with an external drift (Section 21.3), and
for estimating the mean AGB of the four biomes. For biome Forest_moist,
the population mean can be estimated from the data of this biome only, using
the 𝜋 estimator, as the sample size of this biome is very large (Section 9.3).
For the other three biomes, we may prefer model-assisted estimation for small
domains as described in Section 14.2.
In this example I used one quantitative covariate, lnSWIR2, for balancing
the sample. If we have a legacy sample that can be used to fit a linear or
non-linear model, for instance a random forest using multiple covariates and
factors as predictors (Chapter 10), then this model can be used to predict the
study variable for all population units, so that we can use the predictions of
the study variable to balance the sample, see Section 10.3.

16.3 Broad overview of sampling designs for mapping

The non-probability sampling designs for mapping described in the following
chapters can be grouped into three categories (Brus, 2019):

1. geometric sampling designs (Chapters 17 and 18);

2. adapted experimental designs (Chapters 19 and 20); and
3. model-based sampling designs (Chapters 22 and 23).
322 16 Introduction to sampling for mapping

Square and triangular grids are examples of geometric sampling designs; the
sampling units show a regular, geometric spatial pattern. In other geometric
sampling designs the spatial pattern is not perfectly regular. Yet these are
classified as geometric sampling designs when the samples are obtained by min-
imising some geometric criterion, i.e., a criterion defined in terms of distances
between the sampling units and the nodes of a fine prediction grid discretising
the study area (Section 17.2 and Chapter 18).
In model-based sampling designs, the samples are obtained by minimising a
criterion that is defined in terms of variances of prediction errors. An example
is the mean kriging variance criterion, i.e., the average of the kriging variances
over all nodes of the prediction grid. Model-based sampling therefore requires
prior knowledge of the model of spatial variation. Such a model must be
specified and justified. Once this model is given, the sample can be optimised.
In Chapter 22 I will show how a spatial model can be used to optimise the
spacing of a square grid given a requirement on the accuracy of the map. The
grid spacing determines the number of sampling units, so this optimisation
boils down to determining the required sample size. In Chapter 23 I will show
how a sample of a given size can be further optimised through optimisation of
the spatial coordinates of the sampling units.
In Chapter 1 the design-based and model-based approaches for sampling and
statistical inference were introduced. Note that a model-based approach does
not necessarily imply model-based sampling. The adjective ‘model-based’ refers
to the model-based inference, not to the selection of the units. In a model-based
approach sampling units can be, but need not be, selected by model-based
sampling. If they are, then both in selecting the units and in mapping a
statistical model is used. In most cases, the two models differ: once the sample
data are collected, these are used to update the postulated model used for
designing the sample. The updated model is then used in mapping.
Besides geometric and model-based sampling designs for a spatial survey, a
third category can be distinguished: sampling designs that are adaptations
of experimental designs. An adaptation is necessary because in contrast to
experiments, in observational studies one is not free to choose combinations
of levels of different factors. For instance, when two covariates are strongly
positively correlated, it may happen that there are no units with a relatively
large value for one covariate and a relatively small value for the other covariate.
In a full factorial design, all combinations of factor levels are observed. For
instance, suppose we have only two covariates, e.g., application rates for N
and P in an agricultural experiment, and four levels for each covariate. To
account for possible non-linear effects, a good option is to have multiple plots
for all 4 × 4 combinations. This is referred to as a full factorial design. With
𝑘 factors and 𝑙 levels per factor the total number of observations is 𝑙u� . With
numerous factors and/or numerous levels per factor, this becomes unfeasible in
16.3 Broad overview of sampling designs for mapping 323

practice. Alternative designs have been developed that need fewer observations
but still provide detailed information about how the study variable responds
to changes in the factor levels. Examples are Latin hypercube samples and
response surface designs. The survey sampling analogues of these experimental
designs are described in Chapters 19 and 20.
17
Regular grid and spatial coverage sampling

This chapter describes and illustrates two sampling designs by which the
sampling locations are evenly spread throughout the study area: regular grid
sampling and spatial coverage sampling. In a final section, the spatial coverage
sampling design is used to fill in the empty spaces of an existing sample.

17.1 Regular grid sampling

Sampling on a regular grid is an attractive option for mapping because of its
simplicity. The data collected on the grid nodes are not used for design-based
estimation of the population mean or total. For this reason, the grid need
not be placed randomly on the study area as in systematic random sampling
(Chapter 5). The grid can be located such that the grid nodes optimally cover
the study area in the sense of the average distance of the nodes of a fine
discretisation grid to the nearest node of the sampling grid. Commonly used
grid configurations are square and triangular. If the grid data are used in
kriging (Chapter 21), the optimal configuration depends, among others, on the
semivariogram model. If the study variable shows moderate to strong spatial
autocorrelation (see Section 21.1), triangular grids outperform square grids.
Besides the shape of the sampling grid cells, we must decide on the grid spacing.
The grid spacing determines the number of sampling units in the study area,
i.e., the sample size. There are two options to decide on this spacing, either
starting from the available budget or from a requirement on the quality of
the map. The latter is explained in Chapter 22, as this requires a model of
the spatial variation, and as a consequence this is an example of model-based
sampling. Starting from the available budget and an estimate of the costs
per sampling unit, we first compute the affordable sample size. Then we may
derive from this number the grid spacing. For square grids, the grid spacing in
meters is calculated as √𝐴/𝑛, where 𝐴 is the area in m2 and 𝑛 is the number
of sampling units (sample size).
Grids can be selected with function spsample of package sp (Pebesma and
Bivand, 2005). Argument offset is used to select a grid non-randomly. Either

325
326 17 Regular grid and spatial coverage sampling

a sample size can be specified, using argument n, or a grid spacing, using

argument cellsize. In the next code chunk, a square grid is selected with a
spacing of 200 m.

library(sp)
gridded(grdVoorst) <- ~ s1 + s2
mysample <- spsample(
x = grdVoorst, type = "regular", cellsize = c(200, 200),
offset = c(0.5, 0.5)) %>% as("data.frame")

Figure 17.1 shows the selected square grid.

FIGURE 17.1: Non-random square grid sample with a grid spacing of 200
m from Voorst.

The number of grid points in this example equals 115. Nodes of the square grid
in parts of the area not belonging to the population of interest, such as built-up
areas and roads, are discarded by spsample (these nodes are not included in the
sampling frame file grdVoorst). As a consequence, there are some undersampled
areas, for instance in the middle of the study area where two roads cross. If
we use the square grid in spatial interpolation, e.g., by ordinary kriging, we
are more uncertain about the predictions in these undersampled areas than in
areas where the grid is complete. The next section will show how this local
undersampling can be avoided.

Exercises

1. Write an R script to select a square grid of size 100 from West-

Amhara in Ethiopia. Use grdAmhara of package sswr as a sampling
frame. Use a fixed starting point of the grid, i.e., do not select the
grid randomly.
• Compute the number of selected grid points. How comes it is
not exactly equal to 100?
• Select a square grid with a spacing of 10.2 km, and compute
the sample size.
• Write a for-loop to select 200 times a square grid of, on average,
100 points with random starting point. Set a seed so that
the result can be reproduced. Determine for each randomly
17.2 Spatial coverage sampling 327

selected grid the number of selected grid points, and save this
in a numeric. Compute summary statistics of the sample size,
and plot a histogram.
• Select a square grid of exactly 100 points.

17.2 Spatial coverage sampling

Local undersampling with regular grids can be avoided by relaxing the con-
straint that the sampling units are restricted to the nodes of a regular grid.
This is what is done in spatial coverage sampling or, in case of a sample that
is added to an existing sample, in spatial infill sampling. Spatial coverage and
infill samples cover the area or fill in the empty space as uniformly as possible.
The sampling units are obtained by minimising a criterion that is defined in
terms of the geographic distances between the nodes of a fine discretisation grid
and the sampling units. Brus et al. (2007) proposed to minimise the mean of
the squared distances of the grid nodes to their nearest sampling unit (mean
squared shortest distance, MSSD):

1 u�
𝑀𝑆𝑆𝐷 = ∑ min (𝐷2u�u� ) , (17.1)
𝑁 u�=1 u�

where 𝑁 is the total number of nodes of the discretisation grid and 𝐷u�u� is the
distance between the 𝑘th grid node and the 𝑗th sampling point. This distance
measure can be minimised by the k-means algorithm, which is a numerical,
iterative procedure. Figure 17.2 illustrates the selection of a spatial coverage
sample of four points from a square. In this simple example the optimal spatial
coverage sample is known, being the centres of the four subsquares of equal
size. A simple random sample of four points serves as the initial solution. Each
raster cell is then assigned to the closest sampling point. This is the initial
clustering. In the next iteration, the centres of the initial clusters are computed.
Next, the raster cells are reassigned to the closest new centres. This continues
until there is no change anymore. In this case only nine iterations are needed,
where an iteration consists of computing the clusters by assigning the raster
cells to the nearest centre (sampling unit), followed by computing the centres
of these clusters. Figure 17.2 shows the first, second, and ninth iterations.
The same algorithm was used in Chapter 4 to construct compact geographical
strata (briefly referred to as geostrata) for stratified random sampling. The
clusters serve as strata. In stratified random sampling, one or more sampling
units are selected randomly from each geostratum. However, for mapping
purposes probability sampling is not required, so the random selection of a unit
328 17 Regular grid and spatial coverage sampling

FIGURE 17.2: First, second, and ninth iterations of the k-means algorithm
to select a spatial coverage sample of four points from a square. Iterations
are in rows from top to bottom. In the left column of subfigures, the clusters
are computed by assigning the raster cells to the nearest centre. In the right
column of subfigures, the centres of the clusters are computed.
17.2 Spatial coverage sampling 329

within each stratum is not needed. With random selection, the spatial coverage
is suboptimal. Here, the centres of the final clusters (geostrata) are used as
sampling points. This improves the spatial coverage compared to stratified
random sampling.
In probability sampling, we may want to have strata of equal area (clusters of
equal size) so that the sampling design becomes self-weighting. For mapping,
equally sized clusters are not recommended, as this may lead to samples with
suboptimal spatial coverage.

In Figure 17.2 the clusters are of equal size, but this is an artefact. Equally
sized clusters are not guaranteed by the illustrated k-means algorithm. Clus-
tering the raster cells of a square into four clusters is a very special case. In
other cases, the clusters computed with the k-means algorithm described
above might have unequal sizes. In package spcosa also a different k-means
algorithm is implemented, using swops, enforcing compact clusters of equal
size.

Spatial coverage samples can be computed with package spcosa (Walvoort

et al., 2010), using functions stratify and spsample, see code chunk below.
Argument nTry of function stratify specifies the number of initial stratifications
in k-means clustering. Note that function spsample of package spcosa without
optional argument n selects non-randomly one point in each cluster, being the
centre. Figure 17.3 shows a spatial coverage sample of the same size as the
regular grid in study area Voorst (Figure 17.1). Note that the undersampled
area in the centre of the study area is now covered by a sampling point.

library(spcosa)
n <- 115
set.seed(314)
gridded(grdVoorst) <- ~ s1 + s2
mystrata <- spcosa::stratify(
grdVoorst, nStrata = n, equalArea = FALSE, nTry = 10)
mysample <- spsample(mystrata) %>% as("data.frame")

FIGURE 17.3: Spatial coverage sample from Voorst.

330 17 Regular grid and spatial coverage sampling

If the clusters need not be of equal size, we may also use function kmeans of
the stats package, using the spatial coordinates as clustering variables. This
requires less computing time, especially with large data sets.

grdVoorst <- as_tibble(grdVoorst)

mystrata_kmeans <- kmeans(
grdVoorst[, c("s1", "s2")], centers = n, iter.max = 10000, nstart = 10)
mysample_kmeans <- mystrata_kmeans$centers %>% data.frame()

When function kmeans is used to compute the spatial coverage sample, there
is no guarantee that the computed centres of the clusters, used as sampling
points, are inside the study area. In Figure 17.4 there are eight such centres.

FIGURE 17.4: Centres of spatial clusters computed with function kmeans.

This problem can easily be solved by selecting points inside the study area
closest to the centres that are outside the study area. Function rdist of package
fields is used to compute a matrix with distances between the centres outside
the study area and the nodes of the discretisation grid. Then function apply is
used with argument FUN = which.min to compute the discretisation nodes closest
to the centres outside the study area. A similar procedure is implemented
in function spsample of package spcosa when the centres of the clusters are
selected as sampling points (so, when argument n of function spsample is not
used).

library(fields)
gridded(grdVoorst) <- ~ s1 + s2
coordinates(mysample_kmeans) <- ~ s1 + s2
res <- over(mysample_kmeans, grdVoorst)
inside <- as.factor(!is.na(res$z))
units_out <- which(inside == FALSE)
grdVoorst <- as_tibble(grdVoorst)
mysample_kmeans <- as_tibble(mysample_kmeans)
D <- fields::rdist(x1 = mysample_kmeans[units_out, ],
x2 = grdVoorst[, c("s1", "s2")])
units_close <- apply(D, MARGIN = 1, FUN = which.min)
mysample_kmeans[units_out, ] <- grdVoorst[units_close, c("s1", "s2")]
17.3 Spatial infill sampling 331

Exercises

2. In forestry and vegetation surveys, square and circular plots are

often used as sampling units, for instance, 2 m squares or circles
with a diameter of 2 m. To study the relation between the vegetation
and the soil, soil samples must be collected from the vegetation
plots. Suppose we want to collect four soil samples from a square
plot. Where would you locate the four sampling points, so that they
optimally cover the plot?

3. Suppose we are also interested in the accuracy of the estimated plot

means of the soil properties, not just the means. In that case, the
soil samples should not be bulked into a composite sample, but
analysed separately. How would you select the sampling points in
this case?

4. For circular vegetation plots, it is less clear where the sampling

points with smallest MSSD are (Equation (17.1)). Write an R script
to compute a spatial coverage sample of five points from a circular
plot discretised by the nodes of a fine square grid. Use argument
equalArea = FALSE. Check the size (number of raster cells) of the
strata. Repeat this for six sampling points.

5. Consider the case of six strata. The strata are not of equal size. If the
soil samples are bulked into a composite sample, the measurement
on this single sample is a biased estimator of the plot mean. How
can this bias be avoided?

17.3 Spatial infill sampling

If georeferenced data are available that can be used for mapping the study
variable, but we need more data for mapping, it is attractive to account for
these existing sampling units when selecting the additional units. The aim
now is to fill in the empty spaces, i.e., the parts of the study area not covered
by the existing sampling units. This is referred to as spatial infill sampling.
Existing sampling units can easily be accommodated in the k-means algorithm,
using them as fixed cluster centres.
Figure 17.5 shows a spatial infill sample for West-Amhara. A large set of
legacy data on soil organic matter (SOM) in mass percentage (dag kg-1 ) is
available, but these data come from strongly spatially clustered units along
roads (prior points in Figure 17.5). This is a nice example of a convenience
332 17 Regular grid and spatial coverage sampling

FIGURE 17.5: Spatial infill sample of 100 points from West-Amhara.

sample. The legacy data are not ideal for mapping the SOM concentration
throughout West-Amhara. Clearly, it is desirable to collect additional data in
the off-road parts of the study area, with the exception of the northeastern part
where we have already quite a few data that are not near the main roads. The
legacy data are passed to function stratify of package spcosa with argument
priorPoints. The object assigned to this argument must be of class SpatialPoints
or SpatialPointsDataFrame. This optional argument fixes these points as cluster
centres. A spatial infill sample of 100 points is selected, taking into account
these fixed points.

gridded(grdAmhara) <- ~ s1 + s2
n <- 100
ntot <- n + nrow(sampleAmhara)
coordinates(sampleAmhara) <- ~ s1 + s2
proj4string(sampleAmhara) <- NA_character_
set.seed(314)
mystrata <- spcosa::stratify(grdAmhara, nStrata = ntot,
priorPoints = sampleAmhara, nTry = 10)
17.3 Spatial infill sampling 333

mysample <- spsample(mystrata)

plot(mystrata, mysample)

In the output object of spsample, both the prior and the new sampling points
are included. The new points can be obtained as follows:

units <- which(mysample@isPriorPoint == FALSE)

mysample <- as(mysample, "data.frame")
mysample_new <- mysample[units, ]

Exercises

6. Write an R script to select a spatial infill sample of size 100 from

study area Xuancheng in China. Use the iPSM sample in tibble
sampleXuancheng of package sswr as a legacy sample. To map the
SOM concentration, we want to measure the SOM concentration at
100 more sampling points.
• Read the file data/Elevation_Xuancheng.rds with function rast of
package terra, and use this file as a discretisation of the study
area.
• For computational reasons, there are far too many raster cells.
That many cells are not needed to select a spatial infill sample.
Subsample the raster file by selecting a square grid with a
spacing of 900 m × 900 m. First, convert the SpatRaster object to
a data.frame, and then change it to a SpatialPixelsDataFrame using
function gridded. Then use function spsample with argument type
= "regular".
• Select a spatial infill sample using functions stratify and sample
of package spcosa.
18
Covariate space coverage sampling

Regular grid sampling and spatial coverage sampling are pure spatial sampling
designs. Covariates possibly related to the study variable are not accounted for
in selecting sampling units. This can be suboptimal when the study variable is
related to covariates of which maps are available, think for instance of remote
sensing imagery or digital elevation models related to soil properties. Maps of
these covariates can be used in mapping the study variable by, for instance, a
multiple linear regression model or a random forest. This chapter describes
a simple, straightforward method for selecting sampling units on the basis of
the covariate values of the raster cells.
The simplest option for covariate space coverage (CSC) sampling is to cluster
the raster cells by the k-means clustering algorithm in covariate space. Similar
to spatial coverage sampling (Section 17.2) the mean squared shortest distance
(MSSD) is minimised, but now the distance is not measured in geographical
space but in a 𝑝-dimensional space spanned by the 𝑝 covariates. Think of this
space as a multidimensional scatter plot with the covariates along the axes.
The covariates are centred and scaled so that their means become zero and
standard deviations become one. This is needed because, contrary to the spatial
coordinates used as clustering variables in spatial coverage sampling, the ranges
of the covariates in the population can differ greatly. In the clustering of the
raster cells, the mean squared shortest scaled distance (MSSSD) is minimised.
The name ‘scaled distance’ can be confusing. The distances are not scaled, but
rather they are computed in a space spanned by the scaled covariates.
In the next code chunk, a CSC sample of 20 units is selected from Eastern
Amazonia. All five quantitative covariates, SWIR2, Terra_PP, Prec_dm,
Elevation, and Clay, are used as covariates. To select 20 units, 20 clusters are
constructed using function kmeans of the stats package (R Core Team, 2021).
The number of clusters is passed to function kmeans with argument centers.
Note that the number of clusters is not based, as would be usual in cluster
analysis, on the assumed number of subregions with a high density of units
in the multivariate distribution, but rather on the number of sampling units.
The k-means clustering algorithm is a deterministic algorithm, i.e., the final
optimised clustering is fully determined by the initial clustering. This final
clustering can be suboptimal, i.e., the minimised MSSSD value is somewhat
larger than the global minimum. Therefore, the clustering should be repeated
many times, every time starting with a different random initial clustering.

335
336 18 Covariate space coverage sampling

The number of repeats is specified with argument nstart. The best solution is
automatically kept. To speed up the computations, a 5 km × 5 km subgrid of
grdAmazonia is used.

covs <- c("SWIR2", "Terra_PP", "Prec_dm", "Elevation", "Clay")

n <- 20
set.seed(314)
myclusters <- kmeans(
scale(grdAmazonia[, covs]), centers = n, iter.max = 10000, nstart = 100)
grdAmazonia$cluster <- myclusters$cluster

Raster cells with the shortest scaled Euclidean distance in covariate space
to the centres of the clusters are selected as the sampling units. To this end,
first a matrix with the distances of all the raster cells to the cluster centres
is computed with function rdist of package fields (Nychka et al., 2021). The
raster cells closest to the centres are computed with function apply, using
argument FUN = which.min.

library(fields)
covs_s <- scale(grdAmazonia[, covs])
D <- rdist(x1 = myclusters$centers, x2 = covs_s)
units <- apply(D, MARGIN = 1, FUN = which.min)
myCSCsample <- grdAmazonia[units, ]

Figure 18.1 shows the clustering of the raster cells and the raster cells closest
in covariate space to the centres that are used as the selected sample. In Figure
18.2 the selected sample is plotted in biplots of some pairs of covariates. In the
biplots, some sampling units are clearly clustered. However, this is misleading,
as actually we must look in five-dimensional space to see whether the units
are clustered. Two units with a large separation distance in a five-dimensional
space can look quite close when projected on a two-dimensional plane.
The next code chunk shows how the MSSSD of the selected sample can be
computed.

D <- rdist(x1 = scale(

myCSCsample[, covs], center = attr(covs_s, "scaled:center"),
scale = attr(covs_s, "scaled:scale")), x2 = covs_s)
dmin <- apply(D, MARGIN = 2, min)
MSSSD <- mean(dminˆ2)

Note that to centre and scale the covariate values in the CSC sample, the
population means and the population standard deviations are used, as passed
to function scale with arguments center and scale. If these means and stan-
dard deviations are unspecified, the sample means and the sample standard
18.0 Covariate space coverage sampling 337

deviations are used, resulting in an incorrect value of the minimised MSSSD

value. The MSSSD of the selected sample equals 1.004.

FIGURE 18.1: Covariate space coverage sample of 20 units from Eastern

Amazonia, obtained with k-means clustering using five covariates, plotted on
a map of the clusters.

FIGURE 18.2: Covariate space coverage sample of Figure 18.1 plotted in

biplots of covariates, coloured by cluster.

Instead of function kmeans we may use function kmeanspp of package LICORS

(Goerg, 2013). This function is an implementation of the k-means++ algorithm
(Arthur and Vassilvitskii, 2007). This algorithm consists of two parts, namely
338 18 Covariate space coverage sampling

the selection of an optimised initial sample, followed by the standard k-means.

The algorithm is as follows:

1. Select one unit (raster cell) at random.

2. For each unsampled unit 𝑗, compute 𝑑u�u� , i.e., the distance in

standardised covariate space between 𝑗 and the nearest unit 𝑘 that
has already been selected.

3. Choose one new raster cell at random as a new sampling unit with
probabilities proportional to 𝑑2u�u� and add the selected raster cell to
the set of selected cells.

4. Repeat steps 2 and 3 until 𝑛 centres have been selected.

5. Now that the initial centres have been selected, proceed using stan-
dard k-means.

library(LICORS)
myclusters <- kmeanspp(
scale(grdAmazonia[, covs]), k = n, iter.max = 10000, nstart = 30)

Due to the improved initial centres, the risk of ending in a local minimum is
reduced. The k-means++ algorithm is of interest for small sample sizes. For
large sample sizes, the extra time needed for computing the initial centres can
become substantial and may not outweigh the larger number of starts that can
be afforded with the usual k-means algorithm for the same computing time.

18.1 Covariate space infill sampling

If we have legacy data that can be used to fit a model for mapping, it is more
efficient to select an infill sample, similar to spatial infill sampling explained in
Section 17.3. The only difference with spatial infill sampling is that the legacy
data are now plotted in the space spanned by the covariates. The empty regions
we would like to fill in are now the undersampled regions in this covariate
space. The legacy sample units serve as fixed cluster centres; they cannot move
through the covariate space during the optimisation of the infill sample. In the
next code chunk, a function is defined for covariate space infill sampling.
18.1 Covariate space infill sampling 339

CSIS <- function(fixed, nsup, nstarts, mygrd) {

n_fix <- nrow(fixed)
p <- ncol(mygrd)
units <- fixed$units
mygrd_minfx <- mygrd[-units, ]
MSSSD_cur <- NA
for (s in 1:nstarts) {
units <- sample(nrow(mygrd_minfx), nsup)
centers_sup <- mygrd_minfx[units, ]
centers <- rbind(fixed[, names(mygrd)], centers_sup)
repeat {
D <- rdist(x1 = centers, x2 = mygrd)
clusters <- apply(X = D, MARGIN = 2, FUN = which.min) %>% as.factor(.)
centers_cur <- centers
for (i in 1:p) {
centers[, i] <- tapply(mygrd[, i], INDEX = clusters, FUN = mean)
}
#restore fixed centers
centers[1:n_fix, ] <- centers_cur[1:n_fix, ]
#check convergence
sumd <- diag(rdist(x1 = centers, x2 = centers_cur)) %>% sum(.)
if (sumd < 1E-12) {
D <- rdist(x1 = centers, x2 = mygrd)
Dmin <- apply(X = D, MARGIN = 2, FUN = min)
MSSSD <- mean(Dminˆ2)
if (s == 1 | MSSSD < MSSSD_cur) {
centers_best <- centers
clusters_best <- clusters
MSSSD_cur <- MSSSD
}
break
}
}
}
list(centers = centers_best, clusters = clusters_best)
}

The function is used to select an infill sample of 15 units from Eastern Amazonia.
A legacy sample of five units is randomly selected.

set.seed(314)
units <- sample(nrow(grdAmazonia), 5)
fixed <- data.frame(units, scale(grdAmazonia[, covs])[units, ])
340 18 Covariate space coverage sampling

mygrd <- data.frame(scale(grdAmazonia[, covs]))

res <- CSIS(fixed = fixed, nsup = 15, nstarts = 10, mygrd = mygrd)

Figures 18.3 and 18.4 show the selected sample plotted on a map of the clusters
and in biplots of covariates, respectively.

FIGURE 18.3: Covariate space infill sample of 15 units from Eastern Ama-
zonia, obtained with k-means clustering and five fixed cluster centres, plotted
on a map of the clusters. The dots represent the fixed centres (legacy sample),
the triangles the infill sample.

18.2 Performance of covariate space coverage sampling

in random forest prediction
CSC sampling can be a good candidate for a sampling design if we have
multiple maps of covariates, and in addition if we do not want to rely on a
linear relation between the study variable and the covariates. In this situation,
we may consider mapping with machine learning algorithms such as neural
networks and random forests (RF).
I used the Eastern Amazonia data set to evaluate CSC sampling for mapping
the aboveground biomass (AGB). The five covariates are used as predictors in
RF modelling. The calibrated models are used to predict AGB at the units of
18.2 Performance of covariate space 341

FIGURE 18.4: Covariate space infill sample of Figure 18.3 plotted in biplots
of covariates, coloured by cluster. The dots represent the fixed centres (legacy
sample), the triangles the infill sample.

a validation sample of size 25,000 selected by simple random sampling without

replacement from the 1 km × 1 km grid, excluding the cells of the 10 km × 10
km grid from which the calibration samples are selected. The predicted AGB
values at the validation units are compared with the true AGB values, and the
prediction errors are computed. The sample mean of the squared prediction
error is a design-unbiased estimator of the population mean squared error, i.e.,
the mean of the squared errors at all population units (excluding the units of
the 10 km × 10 km grid), see Chapter 25.
Three sample sizes are used, 𝑛 = 25, 50, 100. Of each sample size, 500 CSC
samples are selected using the k-means algorithm, leading to 1,500 CSC samples
in total. The numbers of starts are 500, 350, and 200 for 𝑛 = 25, 50, and 100,
respectively. With these numbers of starts, the computing time was about
equal to conditioned Latin hypercube sampling, see next chapter. Each sample
is used to calibrate a RF model. Simple random sampling (SI) is used as a
reference sampling design that ignores the covariates.
In Figure 18.5 the root mean squared error (RMSE) of the RF predictions
of AGB is plotted against the minimised MSSSD, both for the 3 × 500 CSC
samples and for the 3 × 500 simple random samples. It is no surprise that
for all three sample sizes the minimised MSSSD values of the CSC samples
are substantially smaller than those of the SI samples. However, despite the
substantial smaller MSSSD values, the RMSE values for the CSC samples are
only a bit smaller than those of the SI samples. Only for 𝑛 = 50 a moderately
strong positive correlation can be seen: 𝑟 = 0.513. For 𝑛 = 25 the correlation
is 0.264 only, and for 𝑛 = 100 it is even negative: 𝑟 = −0.183. On average in
this case study, CSC and SI perform about equally (Table 18.1). However,
342 18 Covariate space coverage sampling

TABLE 18.1: Mean RMSE of RF predictions of AGB in Eastern Amazonia

of 500 covariate space coverage (CSC) samples and 500 simple random (SI)
samples, for three sample sizes.

Sample size CSC SI

25 49.30 50.58
50 40.85 43.16
100 39.33 38.83

especially with 𝑛 = 25 and 50 the sampling distribution of RMSE with SI has

a long right tail. This implies that with SI there is a serious risk that a sample
will be selected resulting in poor RF predictions of AGB. In Chapter 19 CSC
sampling is compared with conditioned Latin hypercube sampling.

FIGURE 18.5: Scatter plot of the minimisation criterion MSSSD and the root
mean squared error (RMSE) of RF predictions of AGB in Eastern Amazonia
for covariate space coverage (CSC) sampling and simple random (SI) sampling,
and three sample sizes.
18.2 Performance of covariate space 343

Exercises

1. Write an R script to select a covariate space coverage sample of size

20 from Hunter Valley (grdHunterValley of package sswr). Use the
covariates cti (compound topographic index, which is the same as
topographic wetness index), ndvi (normalised difference vegetation
index), and elevation_m in k-means clustering of the raster cells.
Plot the clusters and the sample on a map of cti and in a biplot of
cti against ndvi.
19
Conditioned Latin hypercube sampling

This chapter and Chapter 20 on response surface sampling are about exper-
imental designs that have been adapted for spatial surveys. Adaptation is
necessary because, in contrast to experiments, in observational studies one is
not free to choose any possible combination of levels of different factors. When
two covariates are strongly positively correlated, it may happen that there
are no population units with a relatively large value for one covariate and a
relatively small value for the other covariate. By contrast, in experimental
research it is possible to select any combination of factor levels.
In a full factorial design, all combinations of factor levels are observed. With 𝑘
factors and 𝑙 levels per factor, the total number of observations is 𝑙u� . With
numerous factors and/or numerous levels per factor, observing 𝑙u� experimental
units becomes unfeasible in practice. Alternative experimental designs have
been developed that need fewer observations but still provide detailed infor-
mation about how the study variable responds to changes in the factor levels.
This chapter will describe and illustrate the survey sampling analogue of Latin
hypercube sampling. Response surface sampling follows in the next chapter.
Latin hypercube sampling is used in designing industrial processes, agricultural
experiments, and computer experiments, with numerous covariates and/or
factors of which we want to study the effect on the output (McKay et al.,
1979). A much cheaper alternative to a full factorial design is an experiment
with, for all covariates, exactly one observation per level. So, in the agricultural
experiment described in Chapter 16 with the application rates of N and P as
factors and four levels for each factor, this would entail four observations only,
distributed in a square in such way that we have in all rows and in all columns
one observation, see Figure 19.1. This is referred to as a Latin square. The
generalisation of a Latin square to a higher number of dimensions is a Latin
hypercube (LH).
Minasny and McBratney (2006) adapted LH sampling for observational studies;
this adaptation is referred to as conditioned Latin hypercube (cLH) sampling.
For each covariate, a series of intervals (marginal strata) is defined. The number
of marginal strata per covariate is equal to the sample size, so that the total
number of marginal strata equals 𝑝u� , with 𝑝 the number of covariates and 𝑛
the sample size. The bounds of the marginal strata are chosen such that the
numbers of raster cells in these marginal strata are equal. This is achieved by

345
346 19 Conditioned Latin hypercube sampling

FIGURE 19.1: Latin square for agricultural experiment with four application
rates of N and P.

using the quantiles corresponding to evenly spaced cumulative probabilities as

stratum bounds. For instance, for five marginal strata we use the quantiles
corresponding to the cumulative probabilities 0.2, 0.4, 0.6, and 0.8.
The minimisation criterion proposed by Minasny and McBratney (2006) is a
weighted sum of three components:

1. O1: the sum over all marginal strata of the absolute deviations of the
marginal stratum sample size from the targeted sample size (equal
to 1);
2. O2: the sum over all classes of categorical covariates of the abso-
lute deviations of the sample proportion of a given class from the
population proportion of that class; and
3. O3: the sum over all entries of the correlation matrix of the absolute
deviation of the correlation in the sample from the correlation in
the population.

With cLH sampling, the marginal distributions of the covariates in the sample
are close to these distributions in the population. This can be advantageous for
mapping methods that do not rely on linear relations, for instance in machine
learning techniques like classification and regression trees (CART), and random
forests (RF). In addition, criterion O3 ensures that the correlations between
predictors are respected in the sample set.
19.0 Conditioned Latin hypercube sampling 347

cLH samples can be selected with function clhs of package clhs (Roudier,
2021). With this package, the criterion is minimised by simulated annealing,
see Section 23.1 for an explanation of this optimisation method. Arguments
iter, temp, tdecrease, and length.cycle of function clhs are control parameters of
the simulated annealing algorithm. In the next code chunk, I use default values
for these arguments. With argument weights, the weights of the components of
the minimisation criterion can be set. The default weights are equal to 1.
Argument cost is for cost-constrained cLH sampling (Roudier et al., 2012), and
argument eta can be used to control the sampling intensities of the marginal
strata (Minasny and McBratney, 2010). This argument is of interest if we
would like to oversample the marginal strata near the edge of the multivariate
distribution.
cLH sampling is illustrated with the five covariates of Eastern Amazonia that
were used before in covariate space coverage sampling (Chapter 18).

library(clhs)
covs <- c("SWIR2", "Terra_PP", "Prec_dm", "Elevation", "Clay")
set.seed(314)
res <- clhs(
grdAmazonia[, covs], size = 20, iter = 50000, temp = 1, tdecrease = 0.95,
length.cycle = 10, progress = FALSE, simple = FALSE)
mysample_CLH <- grdAmazonia[res$index_samples, ]

Figure 19.2 shows the selected sample in a map of SWIR2. In Figure 19.3
the sample is plotted in a biplot of Prec_dm against SWIR2. Each black
dot in the biplot represents one grid cell in the population. The vertical and
horizontal lines in the biplot are at the bounds of the marginal strata of SWIR2
and Prec_dm, respectively. The number of grid cells between two consecutive
vertical lines is constant, as well as the number of grid cells between two
consecutive horizontal lines, i.e., the marginal strata have equal sizes. The
intervals are the narrowest where the density of grid cells in the plot is highest.
Ideally, in each column and row, there is exactly one sampling unit (red dot).
Figure 19.4 shows the sample sizes for all 100 marginal strata. The next code
chunk shows how the marginal stratum sample sizes are computed.

probs <- seq(from = 0, to = 1, length.out = nrow(mysample_CLH) + 1)

bounds <- apply(grdAmazonia[, covs], MARGIN = 2,
FUN = function(x) quantile(x, probs = probs))
mysample_CLH <- as.data.frame(mysample_CLH)
counts <- lapply(1:5, function(i)
hist(mysample_CLH[, i + 3], bounds[, i], plot = FALSE)$counts)
348 19 Conditioned Latin hypercube sampling

FIGURE 19.2: Conditioned Latin hypercube sample from Eastern Amazonia

in a map of SWIR2.

For all marginal strata with one sampling unit, the contribution to component
O1 of the minimisation criterion is 0. For marginal strata with zero or two
sampling units, the contribution is 1, for marginal strata with three sampling
units the contribution equals 2, etc. In Figure 19.4 there are four marginal
strata with zero units and four marginal strata with two units. Component O1
therefore equals 8 in this case.
Figure 19.5 shows the trace of the objective function, i.e., the values of the
minimisation criterion during the optimisation. The trace plot indicates that
50,000 iterations are sufficient. I do not expect that the criterion can be
reduced anymore. The final value of the minimisation criterion is extracted
with function tail using argument n = 1.

trace <- res$obj

tail(trace, n = 1)

[1] 9.51994

In the next code chunk, the minimised value of the criterion is computed “by
hand”.

O1 <- sum(abs(countslf$counts - 1))

rho <- cor(grdAmazonia[, covs])
r <- cor(mysample_CLH[, covs])
O3 <- sum(abs(rho - r))
print(O1 + O3)

[1] 9.51994
19.0 Conditioned Latin hypercube sampling 349

FIGURE 19.3: Conditioned Latin hypercube sample plotted in a biplot of

precipitation in the dryest month against SWIR2. The vertical and horizontal
lines are at the bounds of the marginal strata of the covariates SWIR2 and
precipitation dryest month, respectively.

Exercises

1. Use the data of Hunter Valley (grdHunterValley of package sswr)

to select a cLH sample of size 50, using elevation_m, slope_deg,
cos_aspect, cti, and ndvi as covariates. Plot the selected sample in
a map of covariate cti, and plot the selected sample in a biplot of
cti against elevation_m. In which part of the biplot are most units
selected, and which part is undersampled?

2. Load the simulated data of Figure 16.1 (results/SimulatedSquare.rda),

and select a cLH sample of size 16, using the covariate 𝑥 and the
350 19 Conditioned Latin hypercube sampling

FIGURE 19.4: Sample sizes of marginal strata for the conditioned Latin
hypercube sample of size 20 from Eastern Amazonia.

FIGURE 19.5: Trace of minimisation criterion during optimisation of condi-

tioned Latin hypercube sampling from Eastern Amazonia.

spatial coordinates as stratification variables. Plot the selected sam-

ple in the square with simulated covariate values.
• What do you think of the geographical spreading of the sam-
pling units? Is it optimal?
• Compute the number of sampling units in the marginal strata
of 𝑠1, 𝑠2, and the covariate 𝑥. First, compute the bounds of
these marginal strata. Are all marginal strata of 𝑠1 and 𝑠2
sampled? Suppose that all marginal strata of 𝑠1 and 𝑠2 are
sampled (contain one sampling point), does this guarantee good
spatial coverage?
• Plot the trace of the minimisation criterion, and retrieve the
minimised value. Is this minimised value in agreement with the
marginal stratum sample sizes?
19.2 Performance of conditioned Latin hypercube sampling 351

19.1 Conditioned Latin hypercube infill sampling

Package clhs can also be used for selecting a conditioned Latin hypercube
sample in addition to existing sampling units (legacy sample), as in spatial
infill sampling (Section 17.2). The units of the legacy sample are assigned to
argument must.include. Argument size must be set to the total sample size,
i.e., the number of mandatory units (legacy sample units) plus the number of
additional infill units.
To illustrate conditioned Latin hypercube infill sampling (cLHIS), in the next
code chunk I select a simple random sample of ten units from Eastern Amazonia
to serve as the legacy sample. Twenty new units are selected by cLHIS. The
ten mandatory units (i.e., units which are already sampled and thus must be
in the sample set computed by cLHIS) are at the end of the vector with the
index of the selected raster cells.

set.seed(314)
units <- sample(nrow(grdAmazonia), 10, replace = FALSE)
res <- clhs(grdAmazonia[, covs], size = 30, must.include = units,
tdecrease = 0.95, iter = 50000, progress = FALSE, simple = FALSE)
mysample_CLHI <- grdAmazonia[res$index_samples, ]
mysample_CLHI$free <- as.factor(rep(c(1, 0), c(20, 10)))

Figure 19.6 shows the selected cLHI sample in a map of SWIR2. In Figure 19.7
the sample is plotted in a biplot of SWIR2 against Prec_dm. The marginal
strata already covered by the legacy sample are mostly avoided by the additional
sample.

19.2 Performance of conditioned Latin hypercube

sampling in random forest prediction
The performance of cLH sampling is studied in the same experiment as covariate
space coverage sampling of the previous chapter. In total, 500 cLH samples
of size 25 are selected and an equal number of samples of sizes 50 and 100.
Each sample is used to calibrate a RF model for the aboveground biomass
(AGB) using five covariates as predictors. The calibrated models are used
to predict AGB at the 25,000 validation units selected by simple random
sampling without replacement. Simple random (SI) sampling is added as a
reference sampling design that ignores the covariates. The prediction errors
352 19 Conditioned Latin hypercube sampling

FIGURE 19.6: Conditioned Latin hypercube infill sample from Eastern

Amazonia in a map of SWIR2. Legacy units have free-value 0; infill units have
free-value 1.

FIGURE 19.7: Conditioned Latin hypercube infill sample plotted in a biplot

of SWIR2 against precipitation in the dryest month. Legacy units have free-
value 0; infill units have free-value 1.
19.2 Performance of conditioned Latin hypercube sampling 353

are used to estimate three map quality indices, the population mean error
(ME), the population root mean squared error (RMSE), and the population
Nash-Sutcliffe model efficiency coefficient (MEC), see Chapter 25.
Figure 19.8 shows the results as boxplots, each based on 500 estimates. For
𝑛 = 25 and 100, cLH sampling performs best in terms of RMSE and MEC,
whereas for 𝑛 = 50 CSC sampling performs best. For 𝑛 = 25 and 50, the
boxplots of cLH and SI show quite a few outliers with large values of RMSE,
resulting in small values of MEC. For CSC, these map quality indices are more
stable. Remarkably, for 𝑛 = 100 SI sampling performs about equally to CSC
and cLH sampling.

FIGURE 19.8: Boxplots of ME, RMSE, and MEC of predictions with RF

models calibrated on conditioned Latin hypercube (cLH), covariate space
coverage (CSC), and simple random (SI) samples from Eastern Amazonia, for
sample sizes of 25, 50, and 100 units.

In Figure 19.9 the RMSE is plotted against the minimised criterion (O1 + O3)
for the cLH and the SI samples. For all three sample sizes, there is a weak
positive correlation of the minimisation criterion and the RMSE: for 𝑛 = 25,
50, and 100 this correlation is 0.369, 0.290, and 0.140, respectively. On average,
cLH performs slightly better than SI for 𝑛 = 25 (Table 19.1). The gain in
accuracy decreases with the sample size. For 𝑛 = 100, the two designs perform
about equally. Especially for 𝑛 = 25 and 50, the distribution of RMSE with SI
has a long right tail. For these small sample sizes, the risk of selecting an SI
354 19 Conditioned Latin hypercube sampling

TABLE 19.1: Mean RMSE of RF predictions of AGB in Eastern Amazonia

of 500 conditioned Latin hypercube (cLH) samples and 500 simple random
(SI) samples, and three sample sizes.

Sample size cLH SI

25 47.64 50.58
50 41.81 43.16
100 38.45 38.83

sample leading to a poor map with large RMSE is much larger than with cLH
sampling.

FIGURE 19.9: Biplot of the minimisation criterion (O1 + O3) and the
RMSE of RF predictions of AGB in Eastern Amazonia for conditioned Latin
hypercube (cLH) sampling and simple random (SI) sampling, and three sample
sizes.

These results are somewhat different from the results of Wadoux et al. (2019)
and Ma et al. (2020). In these case studies, cLH sampling appeared to be
an inefficient design for selecting a calibration sample that is subsequently
used for mapping. Wadoux et al. (2019) compared cLH, CSC, spatial coverage
sampling (Section 17.2), and SI for mapping soil organic carbon in France with
a RF model. The latter two sampling designs do not exploit the covariates
19.2 Performance of conditioned Latin hypercube sampling 355

in selecting the calibration units. Sample sizes were 100, 200, 500, and 1,000.
cLH performed worse (larger RMSE) than CSC and not significantly better
than SI for all sample sizes.
Ma et al. (2020) compared cLH, CSC, and SI for mapping soil classes by
various models, among which a RF model, in a study area in Germany. Sample
sizes were 20, 30, 40, 50, 75, and 100 points. They found no relation between
the minimisation criterion of cLH and the overall accuracy of the map with
predicted soil classes. Models calibrated on CSC samples performed better
on average, i.e., on average the overall accuracy of the maps obtained by
calibrating the models on these CSC samples was higher. cLH was hardly
better than SI.
20
Spatial response surface sampling

As with conditioned Latin hypercube sampling, spatial response surface sam-

pling is an experimental design adapted for spatial surveys. Experimental
response surface designs aim at finding an optimum of the response within
specified ranges of the factors. There are many types of response surface designs,
see Myers et al. (2009). With response surface sampling one assumes that some
type of low order regression model can be used to accurately approximate the
relationship between the study variable and the covariates. A commonly used
design is the central composite design. The data obtained with this design are
used to fit a multiple linear regression model with quadratic terms, yielding a
curved, quadratic surface of the response.
The response surface sampling approach is an adaptation of an experimental
design, but at the same time it is an example of a model-based sampling
design. Sampling units are selected to implicitly optimise the estimation of
the quadratic regression model. However, this optimisation is done under
one or more spatial constraints. Unconstrained optimisation of the sampling
design under the linear regression model will not prevent the units from
spatial clustering, see the optimal sample in Figure 16.1. The assumption of
independent data might be violated when the sampling units are spatially
clustered. For this reason, the response surface sampling design selects samples
with good spatial coverage, so that the design becomes robust against violation
of the independence assumption.
Lesch et al. (1995) adapted the response surface methodology for observational
studies. Several problems needed to be tackled. First, when multiple covariates
are used, the covariates must be decorrelated. Second, when sampling units are
spatially clustered, the assumption in linear regression modelling of spatially
uncorrelated model residuals can be violated. To address these two problems,
Lesch et al. (1995) proposed the following procedure; see also Lesch (2005):

1. Transform the covariate matrix into a scaled, centred, decorrelated

matrix by principal components analysis (PCA).

2. Choose the response surface design type.

357
358 20 Spatial response surface sampling

3. Select candidate sampling units based on the distance from the

design points in the space spanned by the principal components.
Select multiple sampling units per design point.

4. Select the combination of candidate sampling units with the highest

value for a criterion that quantifies how uniform the sample is spread
across the study area.

This design has been applied, among others, for mapping soil salinity (ECe),
using electromagnetic (EM) induction measurements and surface array con-
ductivity measurements as predictors in multiple linear regression models.
For applications, see Corwin and Lesch (2005), Lesch (2005), Fitzgerald et al.
(2006), Corwin et al. (2010), and Fitzgerald (2010).
Spatial response surface sampling is illustrated with the EM measurements
(mS m-1 ) of the apparent electrical conductivity on the 80 ha Cotton Research
Farm in Uzbekistan. The EM measurements in vertical dipole mode, with
transmitter at 1 m and 0.5 m from the receiver, are on transects covering the
Cotton Research Farm (Figure 20.1). As a first step, the natural log of the two
EM measurements, denoted by lnEM, are interpolated by ordinary kriging to
a fine grid (Figure 20.2). These ordinary kriging predictions of lnEM are used
as covariates in response surface sampling. The two covariates are strongly
correlated, 𝑟 = 0.73, as expected since they are interpolations of measurements
of the same variable but of different overlapping layers.

FIGURE 20.1: Natural log of EM measurements on the Cotton Research

Farm (with transmitter at 1 m and 0.5 m from receiver).

Function prcomp of the stats package (R Core Team, 2021) is used to compute
the principal component scores for all units in the population (grid cells). The
two covariates are centred and scaled, i.e., standardised principal components
are computed.

pc <- grdCRF %>%

dplyr::select(lnEM100cm, lnEM50cm) %>%
prcomp(center = TRUE, scale = TRUE)
20.0 Spatial response surface sampling 359

FIGURE 20.2: Interpolated surfaces of natural log of EM measurements

on the Cotton Research Farm, used as covariates in spatial response surface
sampling.

The means of the two principal component scores are 0; however, their standard
deviations are not zero but 1.330 and 0.480. Therefore, the principal component
scores are divided by these standard deviations. They then will have the same
weight in the following steps.

grdCRF <- grdCRF %>%

mutate(
PC1 = pc$x[, 1] / pc$sdev[1],
PC2 = pc$x[, 2] / pc$sdev[2])

Function ccd of package rsm (Lenth, 2009) is now used to generate a central
composite response surface design (CCRSD). Argument basis specifies the
number of factors, which is two in our case. Argument n0 is the number of
centre points, and argument alpha determines the position of the star points
(explained hereafter).

library(rsm)
set.seed(314)
print(ccdesign <- ccd(basis = 2, n0 = 1, alpha = "rotatable"))

run.order std.order x1.as.is x2.as.is Block

1 1 4 1.000000 1.000000 1
2 2 5 0.000000 0.000000 1
3 3 2 1.000000 -1.000000 1
4 4 1 -1.000000 -1.000000 1
5 5 3 -1.000000 1.000000 1
6 1 5 0.000000 0.000000 2
7 2 3 0.000000 -1.414214 2
8 3 1 -1.414214 0.000000 2
9 4 4 0.000000 1.414214 2
10 5 2 1.414214 0.000000 2
360 20 Spatial response surface sampling

Data are stored in coded form using these coding formulas ...
x1 ~ x1.as.is
x2 ~ x2.as.is

The experiment consists of two blocks, each of five experimental units. Block 1,
the so-called cube block, consists of one centre point and four cube points. In
the experimental unit represented by the centre point, both factors have levels
in the centre of the experimental range. In the experimental units represented
by the cube points, the levels of both factors is either -1 or +1 unit in the
design space. Block 2, referred to as the star block, consists of one centre point
and four star points. With alpha = "rotatable" the star points are on the circle
circumscribing the square (Figure 20.3).

FIGURE 20.3: Rotatable central composite design for two factors.

To adapt this design for an observational study, we drop one of the centre
points (0,0).

ccd_df <- data.frame(x1 = ccdesign$x1, x2 = ccdesign$x2)

ccd_df <- ccd_df[-6, ]

The coordinates of the CCRSD points are multiplied by a factor so that a large
proportion 𝑝 of the bivariate standardised principal component scores of the
population units is covered by the circle that passes through the design points
(Figure 20.3). The factor is computed as a sample quantile of the empirical
20.0 Spatial response surface sampling 361

distribution of the distances of the points in the scatter to the centre. For 𝑝, I
chose 0.7.

d <- sqrt(grdCRF$PC1ˆ2 + grdCRF$PC2ˆ2)

fct <- quantile(d, p = 0.7)
print(fct)

70%
1.472547

ccd_df <- ccd_df %>%

mutate(x1 = x1 * fct, x2 = x2 * fct)

The next step is to select for each design point several candidate sampling
points. For each of the nine design points, eight points are selected that are
closest to that design point. This results in 9 × 8 candidate sampling points.

candi_all <- NULL

for (i in seq_len(nrow(ccd_df))) {
d2dpnt <- sqrt((grdCRF$PC1 - ccd_df$x1[i])ˆ2 +
(grdCRF$PC2 - ccd_df$x2[i])ˆ2)
grdCRF <- grdCRF[order(d2dpnt), ]
candi <- grdCRF[c(1:8), c("point_id", "x", "y", "PC1", "PC2")]
candi$dpnt <- i
candi_all <- rbind(candi_all, candi)
}

Figure 20.4 shows the nine clusters of candidate sampling points around
the design points. Note that the location of the candidate sampling points
associated with the design points with coordinates (0,-2.13), (1.51,-1.51), and
(2.13,0) are all far inside the circle that passes through the design points. So,
for the optimised sample, there will be three points with principal component
scores that considerably differ from the ideal values according to the CCRSD
design.
Figure 20.5 shows that in geographical space for most design points there are
multiple spatial clusters of candidate units. For instance, for design point nine,
there are three clusters of candidate sampling units. Therefore, there is scope
to optimise the sample computationally.
As a first step, an initial subsample from the candidate sampling units is
selected by stratified simple random sampling, using the levels of factor dpnt
as strata. Function strata of package sampling is used for stratified random
sampling (Tillé and Matei, 2021).
362 20 Spatial response surface sampling

FIGURE 20.4: Clusters of points (red points) around the design points
(triangles) of a CCRSD (two covariates), serving as candidate sampling points.

FIGURE 20.5: Candidate sampling points plotted on a map of the first

standardised principal component (PC1).
20.0 Spatial response surface sampling 363

library(sampling)
set.seed(314)
units_stsi <- sampling::strata(
candi_all, stratanames = "dpnt", size = rep(1, 9))
mysample0 <- getdata(candi_all, units_stsi) %>%
dplyr::select(-(ID_unit:Stratum))

The locations of the nine sampling units are now optimised by minimising a
criterion that is a function of the distance between the nine sampling points.
Two minimisation criteria are implemented, a geometric criterion and a model-
based criterion.
In the geometric criterion (as proposed by Lesch (2005)) for each sampling
point the log of the shortest distance to the other points is computed. The
minimisation criterion is the negative of the sample mean of these distances.
The model-based minimisation criterion is the average correlation of the
sampling points. This criterion requires as input the parameters of a residual
correlogram (see Section 21.3). I assume an exponential correlogram without
nugget, so that the only parameter to be chosen is the distance parameter 𝜙
(Equation (21.13)). Three times 𝜙 is referred to as the effective range of the
exponential correlogram. The correlation of the random variables at two points
separated by this distance is 0.05.
A penalty term is added to the geometric or the model-based minimisation
criterion, equal to the average distance of the sampling points to the associated
design points, multiplied by a weight. With weights > 0, sampling points close
to the design points are preferred over more distant points.
In the next code chunk, a function is defined for computing the minimisation
criterion. Given a chosen value for 𝜙, the 9 × 9 distance matrix of the sampling
points can be converted into a correlation matrix, using function variogramLine
of package gstat (Pebesma, 2004). Argument weight is an optional argument
with default value 0.

getCriterion <- function(mysample, dpnt, weight = 0, phi = NULL) {

D2dpnt <- sqrt((mysample$PC1 - dpnt$x1)ˆ2 + (mysample$PC2 - dpnt$x2)ˆ2)
D <- as.matrix(dist(mysample[, c("x", "y")]))
if (is.null(phi)) {
diag(D) <- NA
logdmin <- apply(D, MARGIN = 1, FUN = min, na.rm = TRUE) %>% log
criterion_cur <- mean(-logdmin) + mean(D2dpnt) * weight
} else {
vgmodel <- vgm(model = "Exp", psill = 1, range = phi)
C <- variogramLine(vgmodel, dist_vector = D, covariance = TRUE)
criterion_cur <- mean(C) + mean(D2dpnt) * weight
364 20 Spatial response surface sampling

}
return(criterion_cur)
}

Function getCriterion is used to compute the geometric criterion for the initial
sample.

criterion_geo <- getCriterion(mysample = mysample0, dpnt = ccd_df)

The initial value of the geometric criterion is -4.829. In the next code chunk,
the initial value for the model-based criterion is computed for an effective
range of 150 m.

It does not make sense to make the effective range smaller than the size of the
grid cells, which is 25 m in our case. For smaller ranges, the correlation matrix
is for any sample a matrix with zeroes. If the effective range is smaller than
the smallest distance between two points in a cluster, the mean correlation
is equal for all samples.

phi <- 50
criterion_mb <- getCriterion(mysample = mysample0, dpnt = ccd_df, phi = phi)

The initial value of the model-based criterion is 0.134.

The objective function defining the minimisation criterion is minimised with
simulated annealing (Kirkpatrick et al. (1983), Aarts and Korst (1989)). One
sampling point is randomly selected and replaced by another candidate sam-
pling point from the same cluster. If the criterion of the new sample is smaller
than that of the current sample, the new sample is accepted. If it is larger, it
is accepted with a probability that is a function of the change in the criterion
(the larger the increase, the smaller the acceptance probability) and of an
annealing parameter named the temperature (the higher the temperature, the
larger the probability of accepting a new, poorer sample, given an increase of
the criterion). See Section 23.1 for a more detailed introduction to simulated
annealing.
The sampling pattern can be optimised with function anneal of package sswr.
The arguments of this function will be clear from the description of the sampling
procedure above.

set.seed(314)
mySRSsample <- anneal(
mysample = mysample0, candidates = candi_all, dpnt = ccd_df, phi = 50,
20.1 Increasing the sample size 365

T_ini = 1, coolingRate = 0.9, maxPermuted = 25 * nrow(mysample0),

maxNoChange = 20, verbose = TRUE)

Figure 20.6 shows the optimised CCRSD samples plotted in the space spanned
by the two principal components, obtained with the geometric and the model-
based criterion, plotted together with the design points. The two optimised
samples are very similar.

FIGURE 20.6: Principal component scores of the spatial CCRSD sample

(triangles), optimised with the geometric and the model-based criterion. Dots:
design points of CCRSD.

Figure 20.7 shows the two optimised CCRSD samples plotted in geographical
space on the first standardised principal component scores.

20.1 Increasing the sample size

Nine points are rather few for fitting a polynomial regression model, especially
for a second-order polynomial with interaction. Therefore, in experiments often
multiple observations are done for each design point. Increasing the sample size
of a response surface sample in observational studies is not straightforward. The
challenge is to avoid spatial clustering of sampling points. A simple solution is
to select multiple points from each subset of candidate sampling units. The
success of this solution depends on how strong candidate sampling units are
spatially clustered. For the Cotton Research Farm for most design points the
366 20 Spatial response surface sampling

FIGURE 20.7: CCRSD sample from the Cotton Research Farm, optimised
with the geometric and the model-based criterion, plotted on a map of the
first standardised principal component (PC1).

candidate sampling units are not in one spatial cluster; so in this case, the
solution may work properly. I increased the number of candidate sampling
units per design point to 16, so that there is a larger choice in the optimisation
of the sampling pattern.

candi_all <- NULL

for (i in seq_len(nrow(ccd_df))) {
d2dpnt <- sqrt((grdCRF$PC1 - ccd_df$x1[i])ˆ2 +
(grdCRF$PC2 - ccd_df$x2[i])ˆ2)
grdCRF <- grdCRF[order(d2dpnt), ]
candi <- grdCRF[c(1:16), c("point_id", "x", "y", "PC1", "PC2")]
candi$dpnt <- i
20.2 Stratified spatial response surface sampling 367

candi_all <- rbind(candi_all, candi)

}

A stratified simple random subsample of two points per stratum is selected,

which serves as an initial sample.

set.seed(314)
units_stsi <- sampling::strata(
candi_all, stratanames = "dpnt", size = rep(2, 9))
mysample0 <- getdata(candi_all, units_stsi) %>%
dplyr::select(-(ID_unit:Stratum))

The data frame with the design points must be doubled. Note that the order
of the design points must be equal to the order in the stratified subsample.

tmp <- data.frame(ccd_df, dpnt = 1:9)

ccd_df2 <- rbind(tmp, tmp)
ccd_df2 <- ccd_df2[order(ccd_df2$dpnt), ]

Figures 20.8 and 20.9 show the optimised CCRSD sample of 18 points in
geographical and principal component space, respectively, obtained with the
model-based criterion, an effective range of 150 m, and zero weight for the
penalty term. Sampling points are not spatially clustered, so I do not expect
violation of the assumption of independent residuals. In principal component
space, all points are pretty close to the design points, except for the four design
points in the lower right corner, where no candidate units near these design
points are available.

20.2 Stratified spatial response surface sampling

The sample size can also be increased by stratified spatial response surface
sampling. The strata are subareas of the study area. When the subsets of
candidate sampling units for some design points are strongly spatially clustered,
the final optimised sample obtained with the method of the previous section
may also show strong spatial clustering. An alternative is then to split the
study area into two or more subareas (strata) and to select from each stratum
candidate sampling units. This guarantees that for each design point we have
at least as many spatial clusters of candidate units as we have strata.
368 20 Spatial response surface sampling

FIGURE 20.8: CCRSD sample with two points per design point, from the
Cotton Research Farm, plotted on a map of the first standardised principal
component (PC1).

FIGURE 20.9: CCRSD sample (triangles) with two points per design point
(dots), optimised with model-based criterion, plotted in the space spanned by
the two standardised principal components.
20.2 Stratified spatial response surface sampling 369

The spatial strata are not used for fitting separate regression models. All
data are used to fit one (second-order) polynomial regression model.

Figure 20.10 shows two subareas used as strata in stratified response surface
sampling of the Cotton Research Farm.

FIGURE 20.10: Two subareas of the Cotton Research Farm used as strata
in stratified CCRSD sampling.

The candidate sampling units are selected in a double for-loop. The outer loop
is over the strata, the inner loop over the design points. Note that variable dpnt
continues to increase by 1 after the inner loop over the nine design points in
subarea 1 is completed, so that variable dpnt (used as a stratification variable
in subsampling the sample of candidate sampling points) now has values
1, 2, … , 18. An equal number of candidate sampling points per design point
in both strata (eight points) is selected by sorting the points of a stratum by
the distance to a design point using function order. Figure 20.11 shows the
candidate sampling points for stratified CCRSD sampling.

candi_all <- NULL

for (h in c(1, 2)) {
data_stratum <- grdCRF %>%
filter(subarea == h)
candi_stratum <- NULL
for (i in seq_len(nrow(ccd_df))) {
d2dpnt <- sqrt((data_stratum$PC1 - ccd_df$x1[i])ˆ2 +
(data_stratum$PC2 - ccd_df$x2[i])ˆ2)
data_stratum <- data_stratum[order(d2dpnt), ]
candi <- data_stratum[c(1:8),
c("point_id", "x", "y", "PC1", "PC2", "subarea")]
370 20 Spatial response surface sampling

candi$dpnt <- i + (h - 1) * nrow(ccd_df)

candi_stratum <- rbind(candi_stratum, candi)
}
candi_all <- rbind(candi_all, candi_stratum)
}

FIGURE 20.11: Candidate sampling points for stratified CCRSD sampling,

plotted on a map of the first principal component (PC1).

As before, dpnt is used as a stratum identifier to subsample the candidate

sampling units. Finally, the number of rows in data.frame ccd_df with the design
points is doubled.

set.seed(314)
units_stsi <- sampling::strata(
candi_all, stratanames = "dpnt", size = rep(1, 18))
mysample0 <- getdata(candi_all, units_stsi) %>%
dplyr::select(-(ID_unit:Stratum))
ccd_df2 <- rbind(ccd_df, ccd_df)

Figures 20.12 and 20.13 show the optimised sample of 18 points in geographical
and principal component space, obtained with the model-based criterion with
an effective range of 150 m. The pattern in the principal component space is
worse compared to the pattern in Figure 20.9. In stratum 1, the distance to
the star point at the top and the upper left and upper right cube points is very
large. In this stratum no population units are present that are close to these
20.3 Mapping 371

design points. By adding a penalty term to the minimisation criterion that is

proportional to the distance to the design points, the distance is somewhat
decreased, but not really for the three design points mentioned above (Figure
20.9). Also note the spatial cluster of three sampling units in Figure 20.12
obtained with a weight equal to 5.

FIGURE 20.12: Stratified CCRSD samples from the Cotton Research Farm,
optimised with the model-based criterion, obtained without (weight = 0) and
with penalty (weight = 5) for a large average distance to design points.

20.3 Mapping
Once the data are collected, the study variable is mapped by fitting a multiple
linear regression model using the two covariates, in our case the two EM
372 20 Spatial response surface sampling

FIGURE 20.13: Principal component scores of the stratified CCRSD sample,

optimised with the model-based criterion, obtained without (weight = 0) and
with penalty (weight = 5) for a large average distance to design points (dots).

measurements, as predictors. The fitted model is then used to predict the

study variable for all unsampled population units.
The value of the study variable at an unsampled prediction location 𝐬0 is
predicted by
̂ 0 ) = 𝐱0 𝛽 ̂
𝑍(𝐬 OLS
, (20.1)
with 𝐱0 the (𝑝 + 1)-vector with covariate values at prediction location 𝐬0 and 1
in the first entry (𝑝 is the number of covariates) and 𝛽̂ the vector with ordinary
least squares estimates of the regression coefficients:

𝛽̂OLS = (𝐗T 𝐗)−1 (𝐗T 𝐳) , (20.2)

with 𝐗 the (𝑛 × (𝑝 + 1)) matrix with covariate values and ones in the first
column (𝑛 is the sample size, and 𝑝 is the number of covariates) and 𝐳 the
𝑛-vector with observations of the study variable.

Although the principal component scores are used to select the sampling
locations, there is no need to use these scores as predictors in the linear
regression model. When all principal components derived from the covariates
are used as predictors, the predicted values and standard errors obtained
with the model using the principal components as predictors are equal to
those obtained with the model using the covariates as predictors.
20.3 Mapping 373

The variance of the prediction error can be estimated by

𝑉(̂ 𝑍(𝐬
̂ 0 )) = 𝜎̂2u� (1 + 𝐱T T
0 (𝐗 𝐗)
−1
𝐱0 ) , (20.3)

with 𝜎̂2u� the estimated variance of the residuals.

In R the model can be calibrated with function lm, and the predictions can be
obtained with function predict. The standard errors of the estimated means
can be obtained with argument se.fit = TRUE. The variances of Equation (20.3)
can be computed by squaring these standard errors and adding the squared
value of the estimated residual variance, which can be extracted with sigma().

mdl <- lm(lnECe ~ lnEM100cm + lnEM50cm, data = mysample)

zpred <- predict(mdl, newdata = grdCRF, se.fit = TRUE)
v_zpred <- zpred$se.fitˆ2+sigma(mdl)ˆ2

The assumption underlying Equations (20.2) and (20.3) is that the model
residuals are independent. We assume that all the spatial structure of the
study variable is explained by the covariates. Even the residuals at two locations
close to each other are assumed to be uncorrelated. A drawback of the spatial
response surface design is that it is hard or even impossible to check this
assumption, as the sampling locations are spread throughout the study area.
If the residuals are not independent, the covariance of the residuals can
be accounted for by generalised least squares estimation of the regression
coefficients (Equation (21.24)). The study variable can then be mapped by
kriging with an external drift (Section 21.3). However, this requires an estimate
of the semivariogram of the residuals (Section 21.5).
21
Introduction to kriging

In the following chapters a geostatistical model, i.e., a statistical model of

the spatial variation of the study variable, is used to optimise the sample
size and/or spatial pattern of the sampling locations. This chapter is a short
introduction to geostatistical modelling.
In Chapter 13 we have already seen how a geostatistical model can be used to
optimise probability sampling designs for estimating the population mean or
total. In the following chapters, the focus is on mapping. A map of the study
variable is obtained by predicting the study variable at the nodes of a fine
discretisation grid. Spatial prediction using a geostatistical model is referred
to as kriging (Webster and Oliver, 2007).
With this prediction method, besides a map of the kriging predictions, a map
of the variance of the prediction error is obtained. I will show hereafter that the
prediction error variance is not influenced by the values of the study variable at
the sampling locations. For this reason, it is possible to search, before the start
of the survey, for the sampling locations that lead to the minimum prediction
error variance averaged over all nodes of a fine prediction grid, provided that
the geostatistical model is known up to some extent.

21.1 Ordinary kriging

The kriging predictions and prediction error variances are derived from a
statistical model of the spatial variation of the study variable. There are
several versions of kriging, but most of them are special cases of the following
generic model:

𝑍(𝐬) = 𝜇(𝐬) + 𝜖(𝐬)

𝜖(𝐬) ∼ 𝒩(0, 𝜎2 ) (21.1)
Cov(𝜖(𝐬), 𝜖(𝐬′ )) = 𝐶(𝐡) ,

375
376 21 Introduction to kriging

with 𝑍(𝐬) the study variable at location 𝐬, 𝜇(𝐬) the mean at location 𝐬, 𝜖(𝐬)
the residual (difference between study variable 𝑧 and mean 𝜇(𝐬)) at location 𝐬,
and 𝐶(𝐡) the covariance of the residuals at two locations separated by vector
𝐡 = 𝐬 − 𝐬′ .
In ordinary kriging (OK) it is assumed that the mean of the study variable is
constant, i.e., the same everywhere (Webster and Oliver, 2007):

𝑍(𝐬) = 𝜇 + 𝜖(𝐬) , (21.2)

with 𝜇 the constant mean, independent of the location 𝐬. Stated otherwise, in

OK we assume stationarity in the mean over the area to be mapped.
𝐶(⋅) is the covariance function, also referred to as the covariogram, and a scaled
version of it, obtained by dividing 𝐶(⋅) by the variance 𝐶(0), is the correlation
function or correlogram.
If the data set is of substantial size, it is possible to define a neighbourhood:
not all sample data are used to predict the study variable at a prediction
location, but only the sample data in this neighbourhood. This implies that
the stationarity assumption is relaxed to the often more realistic assumption
of a constant mean within neighbourhoods.
In OK, the study variable at a prediction location 𝐬0 , 𝑍(𝐬 ̂ 0 ), is predicted as
a weighted average of the observations at the sampling locations (within the
neighbourhood):

u�
̂OK (𝐬0 ) = ∑ 𝜆u� 𝑍(𝐬u� ) ,
𝑍 (21.3)
u�=1

where 𝑍(𝐬u� ) is the study variable at the 𝑖th sampling location and 𝜆u� is
the weight attached to this location. The weights should be related to the
correlation of the study variable at the sampling location and the prediction
location. Note that as the mean is assumed constant (Equation (21.2)), the
correlation of the study variable 𝑍 is equal to the correlation of the residual 𝜖.
Roughly speaking, the stronger this correlation, the larger the weight must
be. If we have a model for this correlation, then we can use this model to
find the optimal weights. Further, if two sampling locations are very close, the
weight attached to these two locations should not be twice the weight attached
to a single, isolated sampling location at the same distance of the prediction
location. This explains that in computing the kriging weights, besides the
covariances of the 𝑛 pairs of prediction location and sampling location, also
the covariances of the 𝑛(𝑛 − 1)/2 pairs that can be formed with the 𝑛 sampling
units are used, see Isaaks and Srivastava (1989) for a nice intuitive explanation.
For OK, the optimal weights, i.e., the weights that lead to the model-unbiased1
1 Model-unbiasedness is explained in Chapter 26.
21.1 Ordinary kriging 377

predictor with minimum error variance (best linear unbiased predictor), can
be found by solving the following 𝑛 + 1 equations:
u�
∑ 𝜆u� 𝐶(𝐬1 , 𝐬u� ) + 𝜈 = 𝐶(𝐬1 , 𝐬0 )
u�=1
u�
∑ 𝜆u� 𝐶(𝐬2 , 𝐬u� ) + 𝜈 = 𝐶(𝐬2 , 𝐬0 )
u�=1
⋮ , (21.4)
u�
∑ 𝜆u� 𝐶(𝐬u� , 𝐬u� ) + 𝜈 = 𝐶(𝐬u� , 𝐬0 )
u�=1
u�
∑ 𝜆u� = 1
u�=1

where 𝐶(𝐬u� , 𝐬u� ) is the covariance of the 𝑖th and 𝑗th sampling location, 𝐶(𝐬u� , 𝐬0 )
is the covariance of the 𝑖th sampling location and the prediction location 𝑠0 ,
and 𝜈 is an extra parameter to be estimated, referred to as the Lagrange
multiplier. This Lagrange multiplier must be included in the set of equations
because the error variance is minimised under the constraint that the kriging
weights sum to 1, see the final line in Equation (21.4). This constraint ensures
that the OK-predictor is model-unbiased. It is convenient to write this system
of equations in matrix form:

𝐶11 𝐶12 … 𝐶1u� 1 𝜆1 𝐶10

⎡ 𝐶 𝐶22 … 𝐶2u� 1 ⎤⎡ 𝜆 ⎤ ⎡ 𝐶 ⎤
⎢ 21 ⎥⎢ 2 ⎥ ⎢ 20 ⎥
⎢ ⋮ ⋮ … ⋮ ⋮ ⎥⎢ ⋮ ⎥=⎢ ⋮ ⎥ . (21.5)
⎢ 𝐶u�1 𝐶u�2 … 𝐶u�u� 1 ⎥ ⎢ 𝜆u� ⎥ ⎢ 𝐶u�0 ⎥
⎣ 1 1 … 1 0 ⎦⎣ 𝜈 ⎦ ⎣ 1 ⎦

Replacing submatrices by single symbols results in the shorthand matrix

equation:

𝐂 𝟏 𝜆 𝐜
[ ][ ]=[ 0 ] . (21.6)
𝟏T 0 𝜈 1

The kriging weights 𝜆 and the Lagrange multiplier 𝜈 can then be computed
by premultiplying both sides of Equation (21.6) with the inverse of the first
matrix of this equation:
−1
𝜆 𝐂 𝟏 𝐜0
[ ]=[ T ] [ ] . (21.7)
𝜈 𝟏 0 1

The variance of the prediction error (ordinary kriging variance, OK variance)

at a prediction location equals
̂ 0 )) = 𝜎2 − 𝜆T 𝐜0 − 𝜈 ,
𝑉OK (𝑍(𝐬 (21.8)
378 21 Introduction to kriging

with 𝜎2 the a priori variance, see Equation (21.2). This equation shows that
the OK variance is not a function of the data at the sampling locations. Given
a covariance function, it is fully determined by the spatial pattern of the
sampling locations and the prediction location. It is this property of kriging
that makes it possible to optimise the grid spacing (Chapter 22) and, as we will
see in Chapter 23, to optimise the spatial pattern of the sampling locations,
given a requirement on the kriging variance. If the kriging variance were a
function of the data at the sampling locations, optimisation would be much
more complicated.
In general practice, the covariance function is not used in kriging, rather
a semivariogram. A semivariogram 𝛾(𝐡) is a model of the dissimilarity of
the study variable at two locations, as a function of the vector 𝐡 separating
the two locations. The dissimilarity is quantified by half the variance of the
difference of the study variable 𝑍 at two locations. Under the assumption that
the expectation of 𝑍 is constant throughout the study area (stationarity in
the mean), half the variance of the difference is equal to half the expectation
of the squared difference:

𝛾(𝐡) = 0.5𝑉[𝑍(𝐬) − 𝑍(𝐬 + 𝐡)] = 0.5𝐸[{𝑍(𝐬) − 𝑍(𝐬 + 𝐡)}2 ] . (21.9)

A covariance function and semivariogram are related by (see Figure 21.1)

𝛾(𝐡) = 𝜎2 − 𝐶(𝐡) . (21.10)

FIGURE 21.1: Spherical covariance function (red line and dot) and semivar-
iogram (black line and dot).
21.1 Ordinary kriging 379

Expressed in terms of the semivariances between the sampling locations and a

prediction location, the OK variance is

̂ 0 )) = 𝜆T𝛾 0 + 𝜈 ,
𝑉OK (𝑍(𝐬 (21.11)

with 𝛾 0 the vector with semivariances between the sampling locations and a
prediction location.
Computing the kriging predictor requires a model for the covariance (or
semivariance) as a function of the vector separating two locations. Often, the
covariance is modelled as a function of the length of the separation vector
only, so as a function of the Euclidian distance between two locations. We
then assume isotropy: given a separation distance between two locations,
the covariance is the same in all directions. Only authorised functions are
allowed for modelling the semivariance, ensuring that the variance of any
linear combination of random variables, like the kriging predictor, is positive.
Commonly used functions are an exponential and a spherical model.
The spherical semivariogram model has three parameters:

1. nugget (𝑐0 ): where the semivariogram touches the y-axis (in Figure
21.1: 25);
2. partial sill (𝑐1 ): the difference between the maximum semivariance
and the nugget (in Figure 21.1: 75); and
3. range (𝜙): the distance at which the semivariance reaches its maxi-
mum (in Figure 21.1: 250 m).

The formula for the spherical semivariogram is

⎧ 0 if ℎ = 0
{ ℎ 3
𝛾(ℎ) = ⎨ 𝑐0 + 𝑐1 [1 − 3
2
ℎ
( u� ) + 1
2 ( u� ) ] if 0 < ℎ ≤ 𝜙 . (21.12)
{ 𝑐 +𝑐 if ℎ > 𝜙
⎩ 0 1

The sum of the nugget and the partial sill is referred to as the sill (or sill
variance or a priori variance).
An exponential semivariogram model also has three parameters. Its formula is

0 if ℎ = 0
𝛾(ℎ) = { . (21.13)
𝑐0 + 𝑐1 exp(−ℎ/𝜙) if ℎ > 0

In an exponential semivariogram, the semivariance goes asymptotically to a

maximum; it never reaches it. In an exponential semivariogram, the range
parameter is replaced by the distance parameter. In an exponential semivari-
ogram without nugget, the semivariance at three times the distance parameter
380 21 Introduction to kriging

is at 95% of the sill. Three times the distance parameter is referred to as the
effective or practical range.
In following chapters, I also use a correlogram, which is a scaled covariance
function, such that the sill of the correlogram equals 1:

𝐶(𝐡)
𝜌(𝐡) = . (21.14)
𝜎2
To illustrate that the OK variance is independent of the values of the study
variable at the sampling locations, I simulated a spatial population of 50 ×
50 units. For each unit a value of the study variable is simulated, using the
semivariogram of Figure 21.1. This is repeated ten times, resulting in ten maps
of 2,500 units. Figure 21.2 shows two of the ten simulated maps. Note that
the two maps clearly show spatial structure, i.e., there are patches of similar
values.

FIGURE 21.2: Two maps simulated with the spherical semivariogram of

Figure 21.1, the centred square grid of sampling units, and the prediction unit
(red cell with coordinates (590,670)).

The simulated maps are sampled on a centred square grid with a spacing of 100
distance units, resulting in a sample of 100 units. Each sample is used one-by-
one to predict the study variable at one prediction location (see Figure 21.2),
using again the semivariogram of Figure 21.1. The semivariogram is passed to
function vgm of package gstat (Pebesma, 2004). Usually, this semivariogram
is estimated from a sample, see Chapter 24, but here we assume that it is
known. Function krige of package gstat is used for kriging. Argument formula
specifies the dependent (study variable) and independent variables (covariates).
The formula z ~ 1 means that we do not have covariates (we assume that the
model-mean is a constant) and that predictions are done by OK (or simple
21.2 Block-kriging 381

kriging, see Section 21.3). Argument locations is a SpatialPointsDataFrame with

the spatial coordinates and observations. Argument newdata is a SpatialPoints
object with the locations we want to predict. Argument nmax can be used to
specify the neighbourhood in terms of the number of nearest observations to be
used in kriging (not used in the code chunk below, so that all 100 observations
are used).

library(sp)
library(gstat)
vgmodel <- vgm(model = "Sph", nugget = 25, psill = 75, range = 250)
gridded(mypop) <- ~ s1 + s2
mysample <- spsample(x = mypop, type = "regular", cellsize = c(100, 100),
offset = c(0.5, 0.5))
zsim_sample <- over(mysample, mypop)
coordinates(s_0) <- ~ s1 + s2
zpred_OK <- v_zpred_OK <- NULL
for (i in seq_len(ncol(Z))) {
mysample$z <- zsim_sample[, i]
predictions <- krige(
formula = z ~ 1,
locations = mysample,
newdata = s_0,
model = vgmodel,
debug.level = 0)
zpred_OK[i] <- predictions$var1.pred
v_zpred_OK[i] <- predictions$var1.var
}

As can be seen in Table 21.1, unlike the predicted value, the OK variance
produced from the different simulations is constant.

21.2 Block-kriging
In the previous section, the support of the prediction units is equal to that of
the sampling units. So, if the observations are done at points (point support),
the support of the predictions are also points, and if means of small blocks are
observed, the predictions are predicted means of blocks of the same size and
shape. There is no change of support. In some cases we may prefer predictions
at a larger support than that of the observations. For instance, we may prefer
predictions of the average concentration of some soil property of blocks of 5 m
× 5 m, instead of predictions at points, simply because of practical relevance.
382 21 Introduction to kriging

TABLE 21.1: Ordinary kriging predictions and kriging variance at a fixed

prediction location for ten data sets with simulated values at a square sampling
grid.

Kriging prediction Kriging variance

51.44286 55.30602
57.29272 55.30602
52.44407 55.30602
51.40719 55.30602
63.09248 55.30602
40.57517 55.30602
54.48507 55.30602
47.42828 55.30602
43.65313 55.30602
44.52740 55.30602

If the observations are at points, there is a change of support, from points

to blocks. Kriging with a prediction support that is larger than the support
of the sample data is referred to as block-kriging. Kriging without change of
support so that sample support and prediction support are equal, is referred
to as point-kriging. Note that point-kriging does not necessarily imply that
the support is a point; it can be, for instance, a small block.
In block-kriging the mean of a prediction block ℬ0 is predicted as a weighted
average of the observations at the sampling units. The kriging weights are
derived much in the same way as in point-kriging (Equations (21.4) to (21.7)).
In block-kriging the covariance between a sampling point 𝑖 and a prediction
point, 𝐶(𝐬u� , 𝐬0 ), is replaced by the mean covariance between the sampling point
and a prediction block 𝐶(𝐬u� , ℬ0 ) (Equation (21.4)). This mean covariance can
be approximated by discretising the prediction block by a fine grid, computing
the covariance between a sampling point 𝑖 and each of the discretisation points,
and averaging.
The variance of the prediction error of the block-mean (block-kriging variance)
equals

̂
𝑉OBK (𝑍(ℬ T ̄
(21.15)
0 )) = 𝜆 𝛾 (ℬ0 ) + 𝜈 − 𝛾(ℬ
̄ 0 , ℬ0 ) ,

with 𝛾 (ℬ
̄ 0 ) the vector with mean semivariances between the sampling points
and a prediction block and 𝛾(ℬ ̄ 0 , ℬ0 ) the mean semivariance within the
prediction block. Comparing this with Equation (21.11) shows that the block-
kriging variance is smaller than the point-kriging variance by an amount
approximately equal to the mean semivariance within a prediction block. Recall
from Chapter 13 that the mean semivariance within a block is a model-based
prediction of the variance within a block (Equation (13.3)).
21.3 Kriging with an external drift 383

21.3 Kriging with an external drift

In kriging with an external drift (KED), the spatial variation of the study
variable is modelled as the sum of a linear combination of covariates and a
spatially correlated residual:

u�
𝑍(𝐬) = ∑ 𝛽u� 𝑥u� (𝐬) + 𝜖(𝐬)
u�=0
(21.16)
𝜖(𝐬) ∼ 𝒩(0, 𝜎2 )
Cov(𝜖(𝐬), 𝜖(𝐬′ )) = 𝐶(𝐡) ,

with 𝑥u� (𝐬) the value of the 𝑘th covariate at location 𝐬 (𝑥0 = 1 for all locations),
𝑝 the number of covariates, and 𝐶(𝐡) the covariance of the residuals at two
locations separated by vector 𝐡 = 𝐬 − 𝐬′ . The constant mean 𝜇 in Equation
(21.2) is replaced by a linear combination of covariates and, as a consequence,
the mean is not constant anymore but varies in space.
With KED, the study variable at a prediction location 𝐬0 is predicted by

u� u� u�
̂
𝑍KED (𝐬0 ) = ∑ 𝛽u�̂ 𝑥u� (𝐬0 ) + ∑ 𝜆u� {𝑍(𝐬u� ) − ∑ 𝛽u�̂ 𝑥u� (𝐬u� )} , (21.17)
u�=0 u�=1 u�=0

with 𝛽u�̂ the estimated regression coefficient associated with covariate 𝑥u� .
The first component of this predictor is the estimated model-mean at the
new location based on the covariate values at this location and the estimated
regression coefficients. The second component is a weighted sum of the residuals
at the sampling locations.
The optimal kriging weights 𝜆u� , 𝑖 = 1, … , 𝑛 are obtained in a similar way
as in OK. The difference is that additional constraints on the weights are
needed, to ensure unbiased predictions. Not only the weights must sum to 1,
but also for all 𝑝 covariates the weighted sum of the covariate values at the
sampling locations must equal the covariate value at the prediction location:
∑u�=1 𝜆u� 𝑥u� (𝐬u� ) = 𝑥u� (𝐬0 ) for all 𝑘 = 1, … , 𝑝. This leads to a system of 𝑛 + 𝑝 + 1
u�

simultaneous equations that must be solved. In matrix notation, this system is

𝐂 𝐗 𝜆 𝐜
[ ][ ]=[ 0 ] , (21.18)
𝐗T 𝟎 𝜈 𝐱0
384 21 Introduction to kriging

with
1 𝑥11 𝑥12 … 𝑥1u�
⎡ 1 𝑥21 𝑥22 … 𝑥2u� ⎤
𝐗=⎢ ⎥ . (21.19)
⎢ ⋮ ⋮ ⋮ … ⋮ ⎥
⎣ 1 𝑥u�1 𝑥u�2 … 𝑥u�u� ⎦

The submatrix 𝟎 is a ((𝑝 + 1) × (𝑝 + 1)) matrix with zeroes, 𝜈 a (𝑝 + 1) vector

with Lagrange multipliers, and 𝐱0 a (𝑝 + 1) vector with covariate values at the
prediction location (including a 1 as the first entry).
The kriging variance with KED equals

̂ 0 )) = 𝜎2 − 𝜆T 𝐜0 − 𝜈T 𝐱0 .
𝑉KED (𝑍(𝐬 (21.20)

The prediction error variance with KED can also be written as the sum of
the variance of the predictor of the mean and the variance of the error in the
interpolated residuals (Christensen, 1991):

̂ 0 )) = 𝜎2 − 𝐜T
𝑉KED (𝑍(𝐬 0𝐂
−1
𝐜0 +
(21.21)
(𝐱0 − 𝐗 𝐂 𝐜0 )T (𝐗T 𝐂−1 𝐗)−1 (𝐱0 − 𝐗T 𝐂−1 𝐜0 ) .
T −1

The first two terms constitute the interpolation error variance, the third term
the variance of the predictor of the mean.
To illustrate that the kriging variance with KED depends on the values of the
covariate at the sampling locations and the prediction location, values of a
covariate 𝑥 and of a correlated study variable 𝑧 are simulated for the 50 × 50
units of a spatial population (Figure 21.3). First, a field with covariate values
is simulated with a model-mean of 10. Next, a field with residuals is simulated.
The field of the study variable is then obtained by multiplying the simulated
field with covariate values by two (𝛽1 = 2), adding a constant of 10 (𝛽0 = 10),
and finally adding the simulated field with residuals.

#simulate covariate values

vgm_x <- vgm(model = "Sph", psill = 10, range = 200, nugget = 0)
C <- variogramLine(vgm_x, dist_vector = H, covariance = TRUE)
Upper <- chol(C)
set.seed(314)
N <- rnorm(n = nrow(mypop), 0, 1)
mypop$x <- crossprod(Upper, N) + 10
#simulate values for residuals
vgm_resi <- vgm(model = "Sph", psill = 5, range = 100, nugget = 0)
C <- variogramLine(vgm_resi, dist_vector = H, covariance = TRUE)
21.3 Kriging with an external drift 385

Upper <- chol(C)

set.seed(314)
N <- rnorm(n = nrow(mypop), 0, 1)
e <- crossprod(Upper, N)
#compute mean of study variable
betas <- c(10, 2)
mu <- betas[1] + betas[2] * mypop$x
#compute study variable z
mypop$z <- mu + e

As before, a centred square grid with a spacing of 100 distance units is selected.
The simulated values of the study variable 𝑧 and covariate 𝑥 are used to predict
𝑧 at a prediction location 𝐬0 by kriging with an external drift (red cell in
Figure 21.3). Although at the prediction location we have only one simulated
value of covariate 𝑥, a series of covariate values is used to predict 𝑧 at that
location: 𝑥0 = 0, 2, 4, … , 20. In practice, we have of course only one value of
the covariate at a fixed location, but this is for illustration purposes only. Note
that we have only one data set with ‘observations’ of 𝑥 and 𝑧 at the sampling
locations (square grid).

FIGURE 21.3: Maps with simulated values of covariate 𝑥 and study variable
𝑧, the centred square grid of sampling units, and the prediction unit (red cell
with coordinates (590,670)).

zxsim_sample <- over(mysample, mypop)

mysample$z <- zxsim_sample$z
mysample$x <- zxsim_sample$x
x0 <- seq(from = 0, to = 20, by = 2)
v_zpred_KED <- NULL
for (i in seq_len(length(x0))) {
s_0$x <- x0[i]
predictions <- krige(
386 21 Introduction to kriging

formula = z ~ x,
locations = mysample,
newdata = s_0,
model = vgm_resi,
debug.level = 0)
v_zpred_KED[i] <- predictions$var1.var
}

Note the formula z ~ x in the code chunk above, indicating that there is now
an independent variable (covariate). The covariate values are attached to the
file with the prediction location one-by-one in a for-loop. Also note that for
KED we need the semivariogram of the residuals, not of the study variable
itself. The residual semivariogram used in prediction is the same as the one
used in simulating the fields: a spherical model without nugget, with a sill of
5, and a range of 100 distance units.
To assess the contribution of the uncertainty about the model-mean 𝜇(𝐬), I also
predict the values assuming that the model-mean is known. In other words, I
assume that the two regression coefficients 𝛽0 (intercept) and 𝛽1 (slope) are
known. This type of kriging is referred to as simple kriging (SK). With SK,
the constraints explained above are removed, so that there are no Lagrange
multipliers involved. Argument beta is used to specify the known regression
coefficients. I use the same values used in simulation.

v_zpred_SK <- NULL

for (i in seq_len(length(x0))) {
s_0$x <- x0[i]
prediction <- krige(
formula = z ~ x,
locations = mysample,
newdata = s_0,
model = vgm_resi,
beta = betas,
debug.level = 0)
v_zpred_SK[i] <- prediction$var1.var
}

Figure 21.4 shows that, contrary to the SK variance, the kriging variance with
KED is not constant but depends on the covariate value at the prediction
location. It is smallest near the mean of the covariate values at the sampling
locations, which is 10.0. The more extreme the covariate value at the prediction
location, the larger the kriging variance with KED. This is analogous to the
variance of predictions with a linear regression model.
21.4 Estimating the semivariogram 387

The variance with SK is constant. This is because with SK we assume that

the regression coefficients are known, so that we know the model-mean at
a prediction location. What remains is the error in the interpolation of the
residuals (first two terms in Equation (21.21)). This interpolation error is
independent of the value of 𝑥 at the prediction location. In Figure 21.4 the
difference between the variances with KED and SK is the variance of the pre-
dictor of the model-mean, due to uncertainty about the regression coefficients.
In real-world applications these regression coefficients are unknown and must
be estimated from the sample data. This variance is smallest for a covariate
value about equal to the sample mean of the covariate, and increases with the
absolute difference of the covariate value and this sample mean.

FIGURE 21.4: Variance of the prediction error as a function of the covariate

value at a fixed prediction location, obtained with kriging with an external
drift (KED) and simple kriging (SK).

21.4 Estimating the semivariogram

Kriging requires a semivariogram or covariance function as input for computing
the covariance matrix of the study variable at the sampling locations, and the
vector of covariance of the study variable at the sampling locations and the
prediction location. In most cases the semivariogram model is unknown and
must be estimated from sample data. The estimated parameters of the semivar-
iogram model are plugged into the kriging equations. There are two different
388 21 Introduction to kriging

approaches for estimating the semivariogram parameters from sample data:

the method-of-moments and maximum likelihood estimation (Lark, 2000).

21.4.1 Method-of-moments

With the method-of-moments (MoM) approach, the semivariogram is estimated

in two steps. In the first step, a sample semivariogram, also referred to as an
experimental semivariogram, is estimated. This is done by choosing a series
of distance intervals. If a semivariogram in different directions is required, we
must also choose direction intervals.
For each distance interval, all pairs of points with a separation distance in that
interval are identified. For each pair, half the squared difference of the study
variable is computed, and these differences are averaged over all point-pairs
of that interval. This average is the estimated semivariance of that distance
interval. The estimated semivariances for all distance intervals are plotted
against the average separation distances per distance interval. In the second
step, a permissible model is fitted to the sample semivariogram, using some
form of weighted least squares. For details, see Webster and Oliver (2007).
The next code chunk shows how this can be done in R. A simple random
sample of 150 points is selected from the first simulated field shown in Figure
21.2. Function variogram of package gstat is then used to compute the sample
semivariogram. Note that distance intervals need to be passed to function
variogram. This can be done with argument width or argument boundaries. In
their absence, the default value for argument width is equal to the maximum
separation distance divided by 15, so that there are 15 points in the sample
semivariogram. The maximum separation distance can be set with argument
cutoff. The default value for this argument is equal to one-third of the longest
diagonal of the bounding box of the point set. The output is a data frame,
with the number of point-pairs, average separation distance, and estimated
semivariance in the first three columns.

set.seed(123)
units <- sample(nrow(mypop), size = 150)
mysample <- mypop[units, ]
coordinates(mysample) <- ~ s1 + s2
vg <- variogram(z ~ 1, data = mysample)
head(vg[, c(1, 2, 3)])

np dist gamma
1 35 24.02379 38.73456
2 74 49.02384 42.57199
3 179 78.62401 62.13025
4 213 110.05088 72.41387
21.4 Estimating the semivariogram 389

5 239 139.20163 82.21432

6 323 170.55712 88.38168

The next step is to fit a model. This can be done with function fit.variogram of
the gstat package. Many models can be fitted with this function (type vgm()
to see all models). I chose a spherical model. Function fit.variogram requires
initial values of the semivariogram parameters. From the sample semivariogram,
my eyeball estimates are 25 for the nugget, 250 for the range, and 75 for the
partial sill. These are passed to function vgm. Figure 21.5 shows the sample
semivariogram along with the fitted spherical model2 .

FIGURE 21.5: Sample semivariogram and fitted spherical model estimated

from a simple random sample of 150 units selected from the first simulated
field shown in Figure 21.2. Numbers refer to point-pairs used in computing
semivariances.

print(vgm_MoM)

model psill range

1 Nug 26.32048 0.0000
2 Sph 68.36364 227.7999

Function fit.variogram has several options for weighted least squares optimisa-
tion, see ?fit.variogram for details. Also note that this non-linear fit may not
converge to a solution, especially if the starting values passed to vgm are not
near their optimal values.
2 The figure is plotted with package ggplot2. The sample semivariogram and the fitted

model can also be plotted with plot(vg, vgm_MoM, plot.numbers = TRUE).

390 21 Introduction to kriging

Further, this method depends on the choice of cutoff and distance intervals.
We hope that modifying these does not change the fitted model too much, but
this is not always the case, especially with smaller data sets.

21.4.2 Maximum likelihood

In contrast to the MoM, with the maximum likelihood (ML) method the data
are not paired into couples and binned into a sample semivariogram. Instead,
the semivariogram model is estimated in one step. To apply this method, one
typically assumes that (possibly after transformation) the 𝑛 sample data come
from a multivariate normal distribution. If we have one observation from a
normal distribution, the probability density of that observation is given by

1 1 𝑧−𝜇 2
𝑓(𝑧|𝜇, 𝜎2 ) = √ exp {− ( ) } , (21.22)
𝜎 2𝜋 2 𝜎

with 𝜇 the mean and 𝜎2 the variance. With multiple independent observations,
each of them coming from a normal distribution, the joint probability density
is given by the product of the probability densities per observation. However,
if the data are not independent, we must account for the covariances and the
joint probability density can be computed by

1
(21.23)
u� 1
𝜇, 𝜃 ) = (2𝜋)− 2 |𝐂|− 2 exp {− (𝐳 − 𝜇 )T 𝐂−1 (𝐳 − 𝜇 )} ,
𝑓(𝐳|𝜇
2

where 𝐳 is the vector with the 𝑛 sample data, 𝜇 is the vector with means, 𝜃
is the vector with parameters of the covariance function, and 𝐂 is the 𝑛 × 𝑛
matrix with variances and covariances of the sample data. If the probability
density of Equation (21.23) is regarded as a function of 𝜇 and 𝜃 with the data
𝐳 fixed, this equation defines the likelihood.
ML estimates of the semivariogram can be obtained with function likfit of
package geoR (Ribeiro Jr et al., 2020). First, a geoR object must be made
specifying which columns of the data frame contain the spatial coordinates
and the study variable.

library(geoR)
mysample <- as(mysample, "data.frame")
dGeoR <- as.geodata(obj = mysample, header = TRUE,
coords.col = c("s1", "s2"), data.col = "z")

The model parameters can then be estimated with function likfit. Argument
trend = "cte" means that we assume that the mean is constant throughout the
study area.
21.5 Estimating the residual semivariogram 391

TABLE 21.2: Estimated parameters of a spherical semivariogram obtained

with method-of-moments (MoM) and maximum likelihood (ML) estimation.

Parameter MoM ML
nugget 26.3 19.5
partial sill 68.4 83.8
range 227.8 217.8

vgm_ML <- likfit(geodata = dGeoR, trend = "cte",

cov.model = "spherical", ini.cov.pars = c(80, 200),
nugget = 20, lik.method = "ML", messages = FALSE)

Table 21.2 shows the ML estimates together with the MoM estimates. As can
be seen, the estimates are substantially different, especially the division of
the a priori variance (sill) into partial sill and nugget. In general, I prefer the
ML estimates because the arbitrary choice of distance intervals to compute a
sample semivariogram is avoided. Also ML estimates of the parameters are
more precise, given a sample size. On the other hand, in ML estimation we
need to assume that the data are normally distributed.

21.5 Estimating the residual semivariogram

For KED, estimates of the regression coefficients and of the parameters of
the residual semivariogram are needed. Estimation of these model param-
eters is not a trivial problem, as the estimated regression coefficients and
the estimated residual semivariogram parameters are not independent. The
residuals depend on the estimated regression coefficients and, as a consequence,
also the parameters of the residual semivariogram depend on the estimated
coefficients. Conversely, the estimated regression coefficients depend on the
spatial correlation of the residuals, and so on the estimated residual semivari-
ogram parameters. This is a classic “which came first, the chicken or the egg?”
problem.
Estimation of the model parameters is illustrated with the simulated field of
Figure 21.3. A simple random sample of size 150 is selected to estimate the
model parameters.
392 21 Introduction to kriging

21.5.1 Iterative method-of-moments

A simple option is iterative estimation of the regression coefficients followed by

MoM estimation of the sample semivariogram and fitting of a semivariogram
model. In the first iteration, the regression coefficients are estimated by ordinary
least squares (OLS). This implies that the data are assumed independent, i.e.,
we assume a pure nugget residual semivariogram. The sample semivariogram of
the OLS residuals is then computed by the MoM, followed by fitting a model
to this sample semivariogram. With package gstat this can be done with one
line of R code, using a formula specifying the study variable and the predictors
as the first argument of function variogram.

set.seed(314)
units <- sample(nrow(mypop), size = 150)
mysample <- mypop[units, ]
vg_resi <- variogram(z ~ x, data = mysample)
model_eye <- vgm(model = "Sph", psill = 10, range = 150, nugget = 0)
vgmresi_MoM <- fit.variogram(vg_resi, model = model_eye)

Given these estimates of the semivariogram parameters, the regression coeffi-

cients are reestimated by accounting for spatial dependency of the residuals.
This can be done by generalised least squares (GLS):

̂
𝛽 GLS = (𝐗T 𝐂−1 𝐗)−1 (𝐗T 𝐂−1 𝐳) . (21.24)

The next code chunk shows how the GLS estimates of the regression coefficients
can be computed. Function spDists of package sp is used to compute the matrix
with distances between the sampling locations, and function variogramLine of
package gstat is used to transform the distance matrix into a covariance
matrix.

X <- matrix(data = 1, nrow = nrow(as(mysample, "data.frame")), ncol = 2)

X[, 2] <- mysample$x
z <- mysample$z
D <- spDists(mysample)
C <- variogramLine(vgmresi_MoM, dist_vector = D, covariance = TRUE)
Cinv <- solve(C)
XCXinv <- solve(crossprod(X, Cinv) %*% X)
XCz <- crossprod(X, Cinv) %*% z
betaGLS <- XCXinv %*% XCz

The inversion of the covariance matrix 𝐂 can be avoided as follows:

21.5 Estimating the residual semivariogram 393

XCX <- crossprod(X, solve(C, X))

XCz <- crossprod(X, solve(C, z))
betaGLS <- solve(XCX, XCz)

This coding is to be preferred as the inversion of a matrix can be numerically

unstable.
The GLS estimates of the regression coefficients can then be used to re-compute
the residuals of the mean, and so on, until the changes in the model parameters
are negligible.

repeat {
betaGLS.cur <- betaGLS
mu <- X %*% betaGLS
mysample$e <- z - mu
vg_resi <- variogram(e ~ 1, data = mysample)
vgmresi_MoM <- fit.variogram(vg_resi, model = model_eye)
C <- variogramLine(vgmresi_MoM, dist_vector = D, covariance = TRUE)
XCX <- crossprod(X, solve(C, X))
XCz <- crossprod(X, solve(C, z))
betaGLS <- solve(XCX, XCz)
if (sum(abs(betaGLS - betaGLS.cur)) < 0.0001) {
break
}
}

Table 21.3 shows the estimates of the residual semivariogram parameters

and regression coefficients, together with the estimates obtained by restricted
maximum likelihood, which is explained in the next subsection.

TABLE 21.3: Estimated parameters of a spherical semivariogram for the

residuals and estimated regression coefficients, obtained with iterative method-
of-moments (iMoM) and restricted maximum likelihood (REML) estimation.

Parameter iMoM REML

nugget 0.000 0.000
partial sill 2.137 1.909
range 129.147 141.491
intercept 3.255 3.083
x 2.712 2.732
394 21 Introduction to kriging

21.5.2 Restricted maximum likelihood

The estimates of the residual semivariogram parameters obtained by the

iterative MoM procedure are not unbiased. When the mean is not constant but
is a linear combination of one or more covariates, also ML estimation results in
biased estimates of the residual semivariogram parameters. Unbiased estimates
of the regression coefficients and residual semivariogram parameters can be
obtained by restricted maximum likelihood (REML), also referred to as residual
maximum likelihood. In REML the vector with the data is premultiplied by a
so-called projection matrix 𝐏 (Kitanidis, 1983). This projection matrix has
the property that a vector with zeroes is obtained when the matrix 𝐗 with
the covariate values at the sampling locations (and ones in the first column) is
premultiplied with 𝐏:
𝐏𝐗 = 𝟎 . (21.25)

Premultiplying both sides of the KED model (Equation (21.16)) with 𝐏 gives
(Webster and Oliver, 2007)

𝛽 + 𝐏𝜖𝜖(𝐬) = 𝐏𝜖𝜖(𝐬) .
𝐏𝐳(𝐬) = 𝐲(𝐬) = 𝐏𝐗𝛽 (21.26)

In words, by premultiplying variable 𝐳 with matrix 𝐏, a new variable 𝐲 is

obtained that has a constant mean. So, the trend is filtered out, whatever the
regression coefficients are. The semivariogram parameters of this new variable
can be estimated by ML. The projection matrix 𝐏 can be computed by

𝐏 = 𝐈 − 𝐗(𝐗T 𝐗)−1 𝐗T , (21.27)

with 𝐈 the 𝑛 × 𝑛 identity matrix (matrix with ones on the diagonal and zeroes
in all off-diagonal elements). The natural log of the residual likelihood can be
computed by (Lark and Webster, 2006)

ℓ(𝜃𝜃|𝐳) = constant − 0.5(ln|𝐂| + |𝐗T 𝐂−1 𝐗| + 𝐲u� 𝐂−1 (𝐈 − 𝐐)𝐳)) , (21.28)

with 𝐐 = 𝐗(𝐗T 𝐂−1 𝐗)−1 𝐗T 𝐂−1

REML estimates of the semivariogram can be obtained with function likfit
of package geoR, used above in ML estimation of the variogram (Ribeiro Jr
et al., 2020).

vgm_REML <- likfit(geodata = dGeoR, trend = ~ x,

cov.model = "spherical", ini.cov.pars = c(2, 100),
nugget = 0, lik.method = "REML", messages = FALSE)
21.5 Estimating the residual semivariogram 395

Table 21.3 shows that REML yields a smaller estimated (partial) sill and a
larger estimated range than iterative MoM. Of the two regression coefficients,
especially the estimated intercept differs considerably among the two estimation
methods.
Realising that this is a rather short introduction to kriging, refer to Isaaks and
Srivastava (1989) for an introduction to geostatistics, to Goovaerts (1997) for
an exposé of the many versions of kriging, and to Webster and Oliver (2007)
for an elaborate explanation of kriging. A nice educational tool for getting a
feeling for ordinary kriging is E{Z}-Kriging3 .

3 https://wiki.52north.org/AI_GEOSTATS/SWEZKriging
22
Model-based optimisation of the grid spacing

This is the first chapter on model-based sampling1 . In Section 17.2 and Chapter
18 a geometric criterion is minimised, i.e., a criterion defined in terms of
distances, either in geographic space (Section 17.2) or in covariate space
(Chapter 18). In model-based sampling, the minimisation criterion is a function
of the variance of the prediction errors.
This chapter on model-based sampling is about optimisation of the spacing of
a square grid, i.e., the distance between neighbouring points in the grid. The
grid spacing is derived from a requirement on the accuracy of the map. Here
and in following chapters, I assume that the map is constructed by kriging;
see Chapter 21 for an introduction. As we have seen in Chapter 21, a kriging
prediction of the study variable at an unobserved location is accompanied by
a variance of the prediction error, referred to as the kriging variance. The map
accuracy requirement is a population parameter of this kriging variance, e.g.,
the population mean of the kriging variance.

22.1 Optimal grid spacing for ordinary kriging

Suppose that we require the population mean of the kriging variance not to
exceed a given threshold. The question then is what the tolerable or maximum
possible grid spacing is, given this requirement. For finding the tolerable grid
spacing we must have prior knowledge of the spatial variation. I first consider
the situation in which it is reasonable to assume that the model-mean of
the study variable is constant throughout the study area, but is unknown.
When the model-mean is unknown, ordinary kriging (OK) is used for mapping.
Furthermore, we need a semivariogram of the study variable. In practice, we
often do not have a reliable estimate of the semivariogram. In the best case
scenario, we have some existing data, of sufficient quantity and suitable spatial
distribution, that can be used to estimate the semivariogram. In other cases,
1 Spatial response surface sampling can also be considered as model-based sampling,

especially when a model-based criterion is used, see Chapter 20.

397
398 22 Model-based optimisation of the grid spacing

such data are lacking and a best guess of the semivariogram must be made,
for instance using data for the same study variable from other, similar areas.
There is no simple equation that relates the grid spacing to the kriging variance.
What can be done is calculate the mean OK variance for a range of grid spacings,
plot the mean ordinary kriging variances against the grid spacings, and use
this plot inversely to determine the tolerable grid spacing, given a constraint
on the mean OK variance.
In the next code chunks, this procedure is used to compute the tolerable spacing
of a square grid for mapping soil organic matter (SOM) in West-Amhara. The
legacy data of the SOM concentration (dag kg-1 ), used before to design a
spatial infill sample (Section 17.3), are used here to estimate a semivariogram.
A sample semivariogram is estimated by the method-of-moments (MoM), and a
spherical model is fitted using functions of package gstat (Pebesma, 2004). The
values for the partial sill, range, and nugget, passed to function fit.variogram
with argument model, are guesses from an eyeball examination of the sample
semivariogram obtained with function variogram, see Figure 22.1. The ultimate
estimates of the semivariogram parameters differ from these eyeball estimates.
First, the projected coordinates of the sampling points are changed from m
into km using function mutate2 .

library(gstat)
grdAmhara <- grdAmhara %>%
mutate(s1 = s1 / 1000, s2 = s2 / 1000)
sampleAmhara <- sampleAmhara %>%
mutate(s1 = s1 / 1000, s2 = s2 / 1000)
coordinates(sampleAmhara) <- ~ s1 + s2
vg <- variogram(SOM ~ 1, data = sampleAmhara)
model_eye <- vgm(model = "Sph", psill = 0.6, range = 40, nugget = 0.6)
vgm_MoM <- fit.variogram(vg, model = model_eye)

The semivariogram of SOM can also be estimated by maximum likelihood (ML)

using function likfit of package geoR (Ribeiro Jr et al., 2020), see Section
21.4.

library(geoR)
sampleAmhara <- as_tibble(sampleAmhara)
dGeoR <- as.geodata(
obj = sampleAmhara, header = TRUE, coords.col = c("s1", "s2"),
data.col = "SOM")
vgm_ML <- likfit(geodata = dGeoR, trend = "cte",

2 This is mainly done to avoid problems in (restricted) maximum likelihood estimation of

the (residual) semivariogram with function likfit of package geoR.

22.1 Optimal grid spacing for ordinary kriging 399

FIGURE 22.1: Sample semivariogram and fitted spherical model of the SOM
concentration in West-Amhara, estimated from the legacy data.

TABLE 22.1: Method-of-moments (MoM) and maximum likelihood (ML)

estimates of the parameters of a spherical semivariogram of the SOM concen-
tration in West-Amhara.

Parameter MoM ML
nugget 0.62 0.56
partial sill 0.56 0.68
range 45.40 36.90

cov.model = "spherical", ini.cov.pars = c(0.6, 40),

nugget = 0.6, lik.method = "ML", messages = FALSE)

Table 22.1 shows the ML estimates of the parameters of the spherical semivari-
ogram, together with the MoM estimates. Either could be used in the following
steps.
400 22 Model-based optimisation of the grid spacing

22.2 Controlling the mean or a quantile of the ordinary

kriging variance
To decide on the grid spacing, we may require the population mean kriging
variance (MKV) not to exceed a given threshold. Instead of the population
mean, we may use the population median or any other quantile of the cu-
mulative distribution function of the kriging variance, for instance the 0.90
quantile (P90), as a quality criterion. Hereafter, the ML semivariogram is used
to optimise the grid spacing given a requirement on the mean, median, and
P90 of the kriging variance.
As a first step, a series of spacings of the square grid with observations is
specified. Only spacings are considered which would result in expected sample
sizes that are reasonable for kriging. With a spacing of 5 km, the expected
sample size is 434 points; with a spacing of 12 km, these are 75 points.

spacing <- 5:12

The next step is to select a simple random sample of evaluation points. It

is important to select a large sample, so that the precision of the estimated
population mean or quantile of the kriging variance will be high.

To check whether the size of the simple random sample of evaluation points
is sufficiently large, we may estimate the standard error of the estimator of
the MKV, see Chapter 3, substituting the kriging variances at the evaluation
points for the study variable values.

set.seed(314)
mysample <- grdAmhara %>%
slice_sample(n = 5000, replace = TRUE) %>%
mutate(s1 = s1 %>% jitter(amount = 0.5),
s2 = s2 %>% jitter(amount = 0.5))

The R code below shows the next steps. Given a spacing, a square grid with a
fixed starting point is selected with function spsample, using argument offset.
A dummy variable is added to the data frame, having value 1 at all grid points,
but any other value is also fine. The predicted value at all evaluation points
equals 1. However, we are not interested in the predicted value but in the
kriging variance only, and we have seen in Chapter 21 that the kriging variance
is independent of the observations of the study variable. The ML estimates of
the semivariogram are used in function vgm to define a semivariogram model
of class variogramModel that can be handled by function krige. For each grid
22.2 Controlling the mean or a quantile of the ordinary kriging variance 401

spacing the population mean, median, and P90 of the kriging variance are
estimated from the evaluation sample. The estimated median and P90 can be
computed with function quantile.

coordinates(mysample) <- ~ s1 + s2
gridded(grdAmhara) <- ~ s1 + s2
MKV_OK <- P50KV_OK <- P90KV_OK <- samplesize <-
numeric(length = length(spacing))
vgm_ML_gstat <- vgm(model = "Sph", nugget = vgm_ML$nugget,
psill = vgm_ML$sigmasq, range = vgm_ML$phi)
for (i in seq_len(length(spacing))) {
mygrid <- spsample(x = grdAmhara, cellsize = spacing[i],
type = "regular", offset = c(0.5, 0.5))
mygrid$dummy <- rep(1, length(mygrid))
samplesize[i] <- nrow(mygrid)
predictions <- krige(
formula = dummy ~ 1,
locations = mygrid,
newdata = mysample,
model = vgm_ML_gstat,
nmax = 100,
debug.level = 0)
MKV_OK[i] <- mean(predictions$var1.var)
P50KV_OK[i] <- quantile(predictions$var1.var, probs = 0.5)
P90KV_OK[i] <- quantile(predictions$var1.var, probs = 0.9)
}
dfKV_OK <- data.frame(spacing, samplesize, MKV_OK, P50KV_OK, P90KV_OK)

The estimated mean and quantiles of the kriging variance are plotted against
the grid spacing (Figure 22.2).
The tolerable grid spacing for the three quality indices can be computed with
function approx of the base package, as shown below for the median kriging
variance.

spacing_tol_P50 <- approx(x = dfKV_OK$P50, y = dfKV_OK$spacing, xout = 0.8)$y

For a mean kriging variance of 0.8 (dag kg-1 )2 the tolerable grid spacing is 8.6
km. For the median kriging variance this is 9.2 km, which is somewhat larger
leading to a smaller sample size. The smaller grid spacing for the mean can
be explained by the right-skewed distribution of the kriging variance, so that
the mean kriging variance is larger than the median kriging variance. For the
P90 of the kriging variance, the tolerable grid spacing is much smaller, 6.8 km,
leading to a much larger sample size.
402 22 Model-based optimisation of the grid spacing

FIGURE 22.2: Mean, median (P50), and 0.90 quantile (P90) of the ordinary
kriging variance of predictions of the SOM concentration in West-Amhara, as
a function of the spacing of a square grid.

Exercises

1. Write an R script to determine the tolerable grid spacing so

that the 0.50, 0.80, and 0.95 quantiles of the variance of OK
predictions of SOM in West-Amhara do not exceed 0.85. Estimate
the semivariogram by MoM.

2. In practice, we are uncertain about the semivariogram. For this

reason, it can be wise to explore the sensitivity of the tolerable grid
spacing for the semivariogram parameters.
• Increase the nugget parameter of the MoM semivariogram by
5%, and change the partial sill parameter so that the sill (nugget
+ partial sill) is unchanged. Compute the tolerable grid spacing
and the corresponding required sample size for a mean kriging
variance of 0.85 (dag kg-1 )2 . Explain the difference.
• Reduce the range of the MoM semivariogram by 5%. Reset the
nugget and the partial sill to their original values. Compute the
tolerable grid spacing and the corresponding required sample
size for a mean kriging variance of 0.85 (dag kg-1 )2 . Explain
the difference.
22.3 Optimal grid spacing for block-kriging 403

22.3 Optimal grid spacing for block-kriging

In the previous section, the tolerable grid spacing is derived from a constraint on
the mean or quantile of the prediction error variances at points. The alternative
is to put a constraint on the mean or quantile of the error variances of the
predicted means of blocks. These means can be predicted with block-kriging
(Section 21.2). Block-kriging predictions can be obtained with function krige
of package gstat using argument block. In the code chunk below, the means
of 100 m × 100 m blocks are predicted by ordinary block-kriging.

MKV_OBK <- P50KV_OBK <- P90KV_OBK <- numeric(length = length(spacing))

for (i in seq_len(length(spacing))) {
mygrid <- spsample(x = grdAmhara, cellsize = spacing[i],
type = "regular", offset = c(0.5, 0.5))
mygrid$dummy <- rep(1, length(mygrid))
samplesize[i] <- nrow(mygrid)
predictions <- krige(
formula = dummy ~ 1,
locations = mygrid,
newdata = mysample,
model = vgm_ML_gstat,
block = c(0.1, 0.1),
nmax = 100,
debug.level = 0)
MKV_OBK[i] <- mean(predictions$var1.var)
P50KV_OBK[i] <- quantile(predictions$var1.var, probs = 0.5)
P90KV_OBK[i] <- quantile(predictions$var1.var, probs = 0.9)
}
dfKV_OBK <- data.frame(spacing, MKV_OBK, P50KV_OBK, P90KV_OBK)

Figure 22.3 shows that the mean, P50, and P90 of the block-kriging predic-
tions are substantially smaller than those of the point-kriging predictions
(Figure 22.2). This can be explained by the large nugget of the semivari-
ogram (Table 22.1). The side length of a prediction block (100 m) is much
smaller than the range of the semivariogram (36.9 km), so that in this case
the mean semivariance within a prediction block is about equal to the nugget.
Roughly speaking, for a given grid spacing the mean point-kriging variance
is reduced by an amount about equal to this mean semivariance to yield the
mean block-kriging variance for this spacing (Section 21.2). Recall that the
mean semivariance within a block is a model-based prediction of the variance
within a block (Subsection 13.1.1, Equation (13.3)).
404 22 Model-based optimisation of the grid spacing

FIGURE 22.3: Mean, median (P50), and 0.90 quantile (P90) of the ordinary
block-kriging variance of predictions of the mean SOM concentration of blocks
of 100 m × 100 m, in West-Amhara, as a function of the spacing of a square
grid.

22.4 Optimal grid spacing for kriging with an external

drift
In the previous sections I assumed a constant model-mean for the study variable.
I now consider the case where covariates that are related to the study variable
are available. A model is calibrated that is the sum of a linear combination of
the covariates (spatial trend) and a spatially structured residual, see Equation
(21.16). Predictions at the nodes of a fine grid are obtained by kriging with an
external drift (KED).
The SOM concentration data of West-Amhara are used to estimate the pa-
rameters (regression coefficients and residual semivariogram parameters) of
the model by restricted maximum likelihood (REML), see Subsection 21.5.2.

library(geoR)
dGeoR <- as.geodata(obj = sampleAmhara, header = TRUE,
coords.col = c("s1", "s2"), data.col = "SOM",
covar.col = c("dem", "rfl_NIR", "rfl_red", "lst"))
vgm_REML <- likfit(geodata = dGeoR, trend = ~ dem + rfl_NIR + rfl_red + lst,
22.4 Optimal grid spacing for kriging with an external drift 405

TABLE 22.2: Maximum likelihood (ML) estimates of the parameters of a

spherical semivariogram for the SOM concentration and restricted maximum
likelihood (REML) estimates of the parameters of a spherical semivariogram
for the residuals of a multiple linear regression model, for West-Amhara.

Parameter ML REML
nugget 0.56 0.36
partial sill 0.68 0.44
range (km) 36.91 5.24

cov.model = "spherical", ini.cov.pars = c(1, 5),

nugget = 0.2, lik.method = "REML", messages = FALSE)

The total sill (partial sill + nugget) of the residual semivariogram, estimated
by REML, equals 0.80, which is considerably smaller than that of the ML
semivariogram of SOM (Table 22.2). A considerable part of the variance of
SOM is explained by the covariates. Besides, note the much smaller range
of the residual semivariogram. The smaller sill and range of the residual
semivariogram show that the spatial structure of SOM is largely captured by
the covariates. The residuals of the model-mean, which is a linear combination
of the covariates, no longer show much spatial structure.
The mean kriging variance as obtained with KED is used as the evaluation
criterion. With KED, the kriging variance is also a function of the values of
the covariates at the sampling locations and the prediction location (Section
21.3). Compared with the procedure above for OK, in the code chunk below,
a slightly different procedure is used. The square grid of a given spacing is
randomly placed on the area (option offset in function spsample is not used),
and this is repeated ten times.

R <- 10
MKV_KED <- matrix(nrow = length(spacing), ncol = R)
vgm_REML_gstat <- vgm(model = "Sph", nugget = vgm_REML$nugget,
psill = vgm_REML$sigmasq, range = vgm_REML$phi)
set.seed(314)
for (i in seq_len(length(spacing))) {
for (j in 1:R) {
mygrid <- spsample(x = grdAmhara, cellsize = spacing[i], type = "regular")
mygrid$dummy <- rep(1, length(mygrid))
mygrd <- data.frame(over(mygrid, grdAmhara), mygrid)
coordinates(mygrd) <- ~ x1 + x2
predictions <- krige(
formula = dummy ~ dem + rfl_NIR + rfl_red + lst,
406 22 Model-based optimisation of the grid spacing

locations = mygrd,
newdata = mysample,
model = vgm_REML_gstat,
nmax = 100,
debug.level = 0)
MKV_KED[i, j] <- mean(predictions$var1.var)
}
}
dfKV_KED <- data.frame(spacing, MKV_KED)

FIGURE 22.4: Mean kriging variance of OK and KED predictions of the

SOM concentration in West-Amhara, as a function of the spacing of a square
grid. With KED for each spacing, ten MKV values are shown obtained by
selecting ten randomly placed grids of that spacing.

Figure 22.4 shows the mean kriging variances, obtained with OK and KED,
as a function of the grid spacing. Interestingly, for grid spacings smaller than
about 9 km, the mean kriging variance with KED is larger than with OK. In
this case only for larger grid spacings KED outperforms OK in terms of the
mean kriging variance. Only for mean kriging variances larger than about 0.82
(dag kg-1 )2 we can afford with KED a larger grid spacing (smaller sample size)
than with OK. Only with large spacings (small sample sizes) we profit from
modelling the mean as a linear function of covariates.

MMKV_KED <- apply(dfKV_KED[, -1], MARGIN = 1, FUN = mean)

spacing_tol_KED <- approx(x = MMKV_KED, y = dfKV_KED$spacing, xout = 0.8)$y
22.5 Bayesian approach 407

The tolerable grid spacing for a mean kriging variance of 0.8 (dag kg-1 )2 , using
KED, equals 7.9 km.

Exercises

3. Given a grid spacing, the mean kriging variance varies among

randomly selected grids, especially for large spacings. Explain why.

4. Write an R script to compute the tolerable grid spacing for KED

of natural logs of the electrical conductivity of the soil across the
Cotton Research Farm of Uzbekistan, using natural logs of the
electromagnetic induction (EM) measurements (lnEM100cm) as a
covariate. Use a nugget of 0.126, a partial sill of 0.083, and a range
of 230 m for an exponential semivariogram of the residuals (Table
23.1). Select a simple random sample of size 1,000 of evaluation
points from the discretisation grid with interpolated lnEM100cm
values to compute the mean kriging variance. Do this by selecting
1,000 grid cells by simple random sampling with replacement and
jittering the centres of the selected grid cells by an amount equal
to half the size of the grid cell. Use as grid spacings 70, 75, … , 100
m. With a spacing of 100 m the number of grid points is about 100
(the farm has an area of about 97 ha). What is the tolerable grid
spacing for a mean kriging variance of 0.165?

22.5 Bayesian approach

In practice, we do not know the semivariogram. In the best case we have prior
data that can be used to estimate the semivariogram. However, even in this
case we are uncertain about the semivariogram model (spherical, exponential,
etc.) and the semivariogram parameters. Lark et al. (2017) showed how in a
Bayesian approach we can account for uncertainty about the semivariogram
parameters when we must decide on the grid spacing. In this approach a prior
distribution of the semivariogram parameters is updated with the sample data
to a posterior distribution (Gelman et al., 2013):

𝑓(𝜃𝜃)𝑓(𝐳|𝜃𝜃)
𝑓(𝜃𝜃|𝐳) = , (22.1)
𝑓(𝐳)

with 𝑓(𝜃𝜃|𝐳) the posterior distribution function, i.e., the probability density
function of the semivariogram parameters given the sample data, 𝑓(𝜃𝜃) our prior
belief in the parameters specified by a probability density function, 𝑓(𝐳|𝜃𝜃) the
408 22 Model-based optimisation of the grid spacing

likelihood of the data, and 𝑓(𝐳) the probability density function of the data.
This probability density function 𝑓(𝐳) is hard to obtain.
Problems with analytical derivation of the posterior distribution are avoided
by selecting a large sample of units (vectors with semivariogram parameters)
from the posterior distribution through Markov chain Monte Carlo (MCMC)
sampling, see Subsection 13.1.3.
In a Bayesian approach, we must define the likelihood function of the data, see
Subsection 13.1.3. I assume that the SOM concentration data in West-Amhara
have a multivariate normal distribution, and that the spatial covariance of the
data can be modelled by a spherical model, see Subsection 21.4.2. The likelihood
is a function of the semivariogram parameters. Given a vector of semivariogram
parameters, the variance-covariance matrix of the data is computed from the
matrix with geographic distances between the sampling points. Inputs of the
loglikelihood function ll are the matrix with distances between the sampling
points, the design matrix X, and the vector with observations of the study
variable z, see Subsection 13.1.3.

D <- as.matrix(dist(sampleAmhara[,c("s1","s2")]))
X <- matrix(1, nrow(sampleAmhara), 1)
z <- sampleAmhara$SOM

Besides the likelihood function, in a Bayesian approach we must define prior

distributions for the semivariogram parameters. Here, we combine the partial
sill and nugget into the ratio of spatial dependence, i.e., the proportion of the
sill attributable to the partial sill. For the ratio of spatial dependence 𝜉 and
the distance parameter 𝜙, I use uniform distributions as priors, with a lower
bound of 0 and an upper bound of 1 for the ratio of spatial dependence, and
a lower bound of 10−6 km and an upper bound of 100 km for the range. A
uniform distribution for the sill is not recommended (Gelman et al., 2013).
Instead, I assume a uniform distribution for the inverse of the sill, with a lower
bound of 10−6 and an upper bound of 2.
These priors can be defined by function createUniformPrior of package
BayesianTools (Hartig et al., 2019). There are also functions to define a
beta density function (commonly used as a prior for proportions) and a trun-
cated normal distribution as a prior. Function createBayesianSetup is then used
to define the setup of the MCMC sampling, specifying the likelihood function,
the prior, and the vector with best prior estimates of the model parameters,
specified with argument best. The ML estimates computed in Section 22.1 are
used as starting values for the inverse of the sill parameter, the ratio of spatial
dependence, and the range.

library(BayesianTools)
priors <- createUniformPrior(
22.5 Bayesian approach 409

lower = c(1E-6, 0, 1E-6), upper = c(2, 1, 100))

sill_ML <- vgm_ML$nugget + vgm_ML$sigmasq
thetas_ML <- c(1 / sill_ML, vgm_ML$sigmasq / sill_ML, vgm_ML$phi)
model <- "Sph"
setup <- createBayesianSetup(likelihood = ll, prior = priors,
best = thetas_ML, names = c("lambda", "xi", "range"))

A sample from the posterior distribution of the semivariogram parameters is

then obtained with function runMCMC. Various sampling algorithms are imple-
mented in package BayesianTools. I used the default sampler DEzs, which
is based on the differential evolution Markov chain (ter Braak and Vrugt,
2008). This algorithm is passed to function runMCMC with argument sampler. It is
common not to use all sampled units, but to discard the units of the burn-in
period that are possibly influenced by the initial arbitrary settings, and to
thin the series of units after this period. The extraction of the ultimate sample
is done with function getSample. Argument start specifies the unit where the
extraction starts, and argument numSamples specifies how many units are selected
through systematic sampling of the full MCMC sample. The alternative is to
use argument thin which defines the thinning interval.

set.seed(314)
res <- runMCMC(setup, sampler = "DEzs")
mcmcsample <- getSample(res, start = 1000, numSamples = 1000) %>%
data.frame()

Table 22.3 shows the first ten units of the MCMC sample from the posterior
distribution of the semivariogram parameters.
The units of the MCMC sample (vectors with semivariogram parameters) are
used one-by-one to compute the average of the kriging variances at the simple
random sample of evaluation points.
For each unit in the MCMC sample, the tolerable grid spacing is computed for
a target MKV of 0.8. Figure 22.5 shows that for most sampled semivariograms
(MCMC sample units) the tolerable grid spacing equals 8 km, which roughly
corresponds with the tolerable grid spacing derived above for OK. For 165
sampled semivariograms, the tolerable grid spacing exceeds 12 km. However,
this grid spacing leads to a sample size that is too small for estimating the
semivariogram and kriging.
Finally, for each grid spacing, the proportion of MCMC samples with a MKV
smaller than or equal to the target MKV of 0.8 is computed. Figure 22.6 shows,
for instance, that if the MKV is required not to exceed a target MKV of 0.8
with a probability of 80%, the tolerable grid spacing is 6.6 km. With a grid
spacing of 8.6 km, as determined before, the probability that the MKV exceeds
0.8 is 54%.
410 22 Model-based optimisation of the grid spacing

TABLE 22.3: First ten units of a MCMC sample from the posterior distribu-
tion of the parameters of a spherical semivariogram for the SOM concentration
in West-Amhara.

Inverse of sill Ratio of spatial dependence Range (km)

0.624 0.655 55.0
0.828 0.588 33.3
0.769 0.480 51.9
0.624 0.655 55.0
0.872 0.538 47.2
0.559 0.643 60.6
0.605 0.690 66.8
0.764 0.585 42.7
0.793 0.608 44.8
0.480 0.741 87.1

FIGURE 22.5: Frequency distribution of tolerable grid spacings for a target

MKV of 0.8.
22.5 Bayesian approach 411

FIGURE 22.6: Proportion of sampled semivariograms with a MKV smaller

than or equal to a target MKV of 0.8.
23
Model-based optimisation of the sampling
pattern

In Chapter 22 a model of the spatial variation is used to optimise the spacing

of a regular grid. The grid spacing determines the number of grid points within
the study area, so optimisation of the grid spacing is equivalent to optimisation
of the sample size of a square grid.
This chapter is about optimisation of the spatial coordinates of the sampling
units given the sample size. So, we are searching for the optimal spatial
sampling pattern of a fixed number of sampling units. The constraint of
sampling on a regular grid is dropped. In general, the optimal spatial sampling
pattern is irregular. Similar to spatial coverage sampling (Section 17.2), we
search for the optimal sampling pattern through minimisation of an explicit
criterion. In spatial coverage sampling, the minimisation criterion is the mean
squared shortest distance (MSSD) which is minimised by k-means. In this
chapter the minimisation criterion is the mean kriging variance (MKV) or
a quantile of the kriging variance. Algorithm k-means cannot be used for
minimising this criterion, as it uses (standardised) distances between cluster
centres (the sampling locations) and the nodes of a discretisation grid, and the
kriging variance is not a simple linear function of these distances. A different
optimisation algorithm is needed. Here, spatial simulated annealing is used
which is explained in the next subsection. Non-spatial simulated annealing
was used before in conditioned Latin hypercube sampling using package clhs
(Chapter 19).

23.1 Spatial simulated annealing

Inspired by the potentials of optimisation through simulated annealing (Kirk-
patrick et al., 1983), van Groenigen and Stein (1998) proposed to optimise the
sampling pattern by spatial simulated annealing (SSA), see also van Groenigen
et al. (1999) and van Groenigen et al. (2000). This is an iterative, random search
procedure, in which a sequence of samples is generated. A newly proposed
sample is obtained by slightly modifying the current sample. One sampling

413
414 23 Model-based optimisation of the sampling pattern

location of the current sample is randomly selected, and this location is shifted
to a random location within the neighbourhood of the selected location.
The minimisation criterion is computed for the proposed sample and compared
with that of the current sample. If the criterion of the proposed sample is
smaller, the sample is accepted. If the criterion is larger, the sample is accepted
with a probability equal to

(23.1)
−Δ
𝑃=𝑒 u� ,

with Δ the increase of the criterion and 𝑇 the “temperature”.

The name of this parameter shows the link with annealing in metallurgy.
Annealing is a heat treatment of a material above its recrystallisation tem-
perature. Simulated annealing mimics the gradual cooling of metal alloys,
resulting in an optimum or near-optimum structure of the atoms in the alloy.

The larger the value of 𝑇, the larger the probability that a proposed sample
with a given increase of the criterion is accepted (Figure 23.1). The temperature
𝑇 is stepwise decreased during the optimisation: 𝑇u�+1 = 𝛼𝑇u� . In Figure 23.1
𝛼 equals 0.9. The effect of decreasing the temperature is that the acceptance
probability of worse samples decreases during the optimisation and approaches
0 towards the end of the optimisation. Note that the temperature remains
constant during a number of iterations, referred to as the chain length. In
Figure 23.1 this chain length equals 100 iterations. Finally, a stopping criterion
is required. Various stopping criteria are possible; one option is to set the
maximum numbers of chains with no improvement. 𝑇, 𝛼, the chain length, and

FIGURE 23.1: Acceptance probability as a function of the change in the

mean kriging variance (MKV) used as a minimisation criterion, and cooling
schedule in spatial simulated annealing. For negative changes (MKV of proposed
sample smaller than of current sample) the acceptance probability equals 1.
23.2 Optimising the sampling pattern for ordinary kriging 415

the stopping criterion are annealing schedule parameters that must be chosen
by the user.

23.2 Optimising the sampling pattern for ordinary

kriging
In ordinary kriging (OK), we assume a constant model-mean. No covariates are
available that are related to the study variable. Optimisation of the sampling
pattern for OK is illustrated with the Cotton Research Farm in Uzbekistan
which was used before to illustrate spatial response surface sampling (Chapter
20). The spatial coordinates of 50 sampling locations are optimised for mapping
of the soil salinity (as measured by the electrical conductivity, ECe, of the soil)
by OK. In this section, the coordinates of the sampling points are optimised
for OK. In Section 23.3 this is done for kriging with an external drift. In that
section, a map of interpolated electromagnetic (EM) induction measurements
is used to further optimise the coordinates of the sampling points.
Model-based optimisation of the sampling pattern for OK requires as input a
semivariogram of the study variable. For the Cotton Research Farm, I used the
ECe (dS m-1 ) data collected in eight surveys in the period from 2008 to 2011
at 142 points to estimate this semivariogram (Akramkhanov et al., 2014). The
ECe data are natural-log transformed. The sample semivariogram is shown in
Figure 23.2. The R code below shows how I fitted the semivariogram model
with function nls (“non-linear least squares”) of the stat package. I did not
use function fit.variogram of the gstat package (Pebesma, 2004), because this
function requires the output of function variogram as input, whereas the sample
semivariogram is here computed in a different way.

The sample semivariogram is computed by first estimating sample semi-

variograms for each of the eight surveys separately, followed by computing
weighted averages of semivariances and distances per lag, using the numbers
of pairs as weights (R code not shown).

The semivariogram parameters as estimated by nls are then used to define a

semivariogram model of class variogramModel of package gstat, using function
vgm. This is done because function optimMKV requires a semivariogram model of
this class, see hereafter. As already mentioned in Chapter 22, in practice we
often do not have legacy data from which we can estimate the semivariogram,
and a best guess of the semivariogram then must be made.
416 23 Model-based optimisation of the sampling pattern

library(gstat)
res_nls <- nls(semivar ~ nugget + psill * (1 - exp(-h / range)),
start = list(nugget = 0.1, psill = 0.4, range = 200), weights = somnp)
vgm_lnECe <- vgm(model = "Exp", nugget = coef(res_nls)[1],
psill = coef(res_nls)[2], range = coef(res_nls)[3])

FIGURE 23.2: Sample semivariogram and fitted exponential model of lnECe

at the Cotton Research Farm.

The estimated semivariogram parameters are shown in Table 23.1. The nugget-
to-sill ratio is about 1/4, and the effective range is about 575 m (three times
the distance parameter of an exponential model).
The coordinates of the sampling points are optimised with function optimMKV of
package spsann (Samuel-Rosa, 2019)1 . First, the candidate sampling points
are specified by the nodes of a grid discretising the population. As explained
hereafter, this does not necessarily imply that the population is treated as a
finite population. Next, the parameters of the annealing schedule are set. Note
that both the initial acceptance rate and the initial temperature are set, which
may seem weird as the acceptance rate is a function of the temperature, see
Equation (23.1). The optimisation stops when an initial temperature is chosen
leading to an acceptance rate outside the interval specified with argument
initial.acceptance. If the acceptance rate is smaller than the lower bound of the
interval, a larger value for the initial temperature must be chosen; if the rate
is larger than the upper bound, a smaller initial temperature must be chosen.
1 At the moment of writing this book, package spsann is not available on CRAN. You
can install spsann and its dependency pedometrics with remotes::install_github("samuel-
rosa/spsann") and remotes::install_github("samuel-rosa/pedometrics").
23.2 Optimising the sampling pattern for ordinary kriging 417

TABLE 23.1: Estimated parameters of an exponential semivariogram for

lnECe (estimated by method-of-moments) and for the residuals of a linear
regression model for lnECe using lnEM100cm as a predictor (estimated by
REML).

Variable Nugget Partial sill Distance parameter (m)

lnECe 0.116 0.336 192
residuals 0.126 0.083 230

Arguments chain.length and stopping of function scheduleSPSANN are multipliers.

So, for a chain length of five, the number of iterations equals 5𝑛, with 𝑛 the
sample size.
During the optimisation, a sample is perturbed by replacing one randomly
selected point of the current sample by a new point. This selection of the new
point is done in two steps. In the first step, one node of the discretisation grid
(specified with argument candi) is randomly selected. Only the nodes within a
neighbourhood defined by x.min, x.max, y.min, and y.max can be selected. The
nodes within this neighbourhood have equal probability of being selected. In
the second step, one point is selected within a grid cell with the selected node
at its centre and a side length specified with argument cellsize. So, it is natural
to set cellsize to the spacing of the discretisation grid. With cellsize = 0, the
sampling points are restricted to the nodes of the discretisation grid.

library(spsann)
candi <- grdCRF[, c("x", "y")]
schedule <- scheduleSPSANN(
initial.acceptance = c(0.8,0.95),
initial.temperature = 0.004, temperature.decrease = 0.95,
chains = 500, chain.length = 2, stopping = 10, cellsize = 25)

The R code for optimising the sampling pattern is as follows.

set.seed(314)
res <- optimMKV(
points = 50, candi = candi,
vgm = vgm_lnECe, eqn = z ~ 1,
schedule = schedule, nmax = 20,
plotit = FALSE, track = TRUE)
mysample <- res$points
trace <- res$objective$energy

The spatial pattern of the sample in Figure 23.3 and the trace of the MKV in
Figure 23.4 suggest that we are close to the global optimum.
418 23 Model-based optimisation of the sampling pattern

FIGURE 23.3: Optimised sampling pattern for the mean variance of OK

predictions of lnECe (model-based sample) and spatial coverage sample of the
Cotton Research Farm.

FIGURE 23.4: Trace of the mean ordinary kriging variance (MOKV).

For comparison I also computed a spatial coverage sample of the same size.
The spatial patterns of the two samples are quite similar (Figure 23.3). The
MKV of the spatial coverage sample equals 0.2633 (dS m-1 )2 , whereas for
the model-based sample the MKV equals 0.2642 (dS m-1 )2 . So, no gain in
precision is achieved by the model-based optimisation of the sampling pattern
compared to spatial coverage sampling. With cellsize = 0 the minimised MKV
is slightly smaller: 0.2578 (dS m-1 )2 . This outcome is in agreement with the
results reported by Brus et al. (2007).
Instead of the mean OK variance (MOKV), we may prefer to use some quantile
of the cumulative distribution function of the OK variance as a minimisation
criterion. For instance, if we use the 0.90 quantile as a criterion, we are
searching for the sampling locations so that the 90th percentile (P90) of the
OK variance is minimal. This can be done with function optimUSER of package
23.2 Optimising the sampling pattern for ordinary kriging 419

spsann. The objective function to be minimised can be passed to this function

with argument fun. In this case the objective function is as follows.

QOKV <- function(points, esample, model, nmax, prob) {

points <- as.data.frame(points)
coordinates(points) <- ~ x + y
points$dum <- 1
res <- krige(
formula = dum ~ 1,
locations = points,
newdata = esample,
model = model,
nmax = nmax,
debug.level = 0)
quantile(res$var1.var, probs = prob)
}

The next code chunk shows how this objective function can be minimised.

mysample_eval <- candi

coordinates(mysample_eval) <- ~ x + y
set.seed(314)
res <- optimUSER(
points = 50, candi = candi,
fun = QOKV,
esample = mysample_eval,
model = vgm_lnECe,
nmax = 20, prob = 0.9,
schedule = schedule)

Argument esample specifies a SpatialPoints object with the evaluation points,

i.e., the points at which the kriging variance is computed. Above I used
all candidate sampling points as evaluation points. Computing time can be
reduced by selecting a coarser square grid with evaluation points. The number
of points used in kriging is specified with argument nmax, and the probability
of the cumulative distribution function of the kriging variance is specified with
argument prob. Optimisation of the sampling pattern for a quantile of the OK
variance is left as an exercise.

Exercises

1. Write an R script to optimise the spatial coordinates of 16 points

in a square for OK. First, create a discretisation grid of 20 × 20
nodes. Use an exponential semivariogram without nugget, with a
420 23 Model-based optimisation of the sampling pattern

sill of 2, and a distance parameter of four times the spacing of the

discretisation grid. Optimise the sampling pattern with SSA (using
functions scheduleSPSANN and optimMKV of package spsann).
• Check whether the optimisation has converged by plotting the
trace of the optimisation criterion MKV.
• Based on the coordinates of the sampling points, do you think
the sample is the global optimum, i.e., the sample with the
smallest possible MKV?

2. Write an R script to optimise the sampling pattern of 50 points, using

the P90 of the variance of OK predictions of lnECe on the Cotton
Research Farm as a minimisation criterion. Use the semivariogram
parameters of Table 23.1. Compare the optimised sample with the
sample optimised with the mean OK variance (shown in Figure
23.3).

23.3 Optimising the sampling pattern for kriging with

an external drift
If we have one or more covariates that are linearly related to the study variable,
the study variable can be mapped by kriging with an external drift (KED).
A requirement is that we have maps of the covariates so that, once we have
estimated the parameters of the model for KED from the data collected at the
optimised sample, these covariate maps can be used to map the study variable
(see Equation (21.21)).
Optimisation of the sampling pattern for KED requires as input the semi-
variogram of the residuals. Besides, we must decide on the covariates for the
model-mean. Note that we do not need estimates of the regression coeffi-
cients associated with the covariates as input, but just which combination of
covariates we want to use for modelling the model-mean of the study variable.
Optimisation of the sampling pattern for KED is illustrated with the Cotton
Research Farm. The interpolated natural log of the EM data (with transmitter
at 1 m) is used as a covariate, see Figure 20.1. The data for fitting the model
are in data file sampleCRF. The parameters of the residual semivariogram are
estimated by restricted maximum likelihood (REML), see Subsection 21.5.2.
At several points, multiple pairs of observations of the study variable ECe and
the covariate EM have been made. These calibration data have exactly the
same spatial coordinates. This leads to problems with REML estimation. The
covariance matrix is not positive definite, so that it cannot be inverted. To
23.3 Optimising the sampling pattern for kriging with an external drift 421

solve this problem, I jittered the coordinates of the sampling points by a small
amount.

library(geoR)
sampleCRF$lnEM100 <- log(sampleCRF$EMv1m)
sampleCRF$x <- jitter(sampleCRF$x, amount = 0.001)
sampleCRF$y <- jitter(sampleCRF$y, amount = 0.001)
dGeoR <- as.geodata(obj = sampleCRF, header = TRUE,
coords.col = c("x", "y"), data.col = "lnECe", covar.col = "lnEM100")
vgm_REML <- likfit(geodata = dGeoR, trend = ~ lnEM100,
cov.model = "exponential", ini.cov.pars = c(0.1, 200),
nugget = 0.1, lik.method = "REML", messages = FALSE)

The REML estimates of the parameters of the residual semivariogram are

shown in Table 23.1. The estimated sill (sum of nugget and partial sill) of the
residual semivariogram is substantially smaller than that of lnECe, showing
that the linear model for the model-mean explains a considerable part of the
spatial variation of lnECe.
To optimise the sampling pattern for KED, using the mean KED variance as
a minimisation criterion, a data frame with the covariates at the candidate
sampling points must be specified with argument covars. The formula for the
model-mean is specified with argument eqn.

set.seed(314)
res <- optimMKV(
points = 50, candi = candi, covars = grdCRF,
vgm = vgm_REML_gstat, eqn = z ~ lnEM100cm,
schedule = schedule, nmax = 20,
plotit = FALSE, track = FALSE)

Figure 23.5 shows the optimised locations of a sample of 50 points. This

clearly shows the irregular spatial pattern of the sampling points induced by
the covariate lnEM100cm. Computing time was substantial: 35.06 minutes
(processor AMD Ryzen 5, 16 GB RAM).
Comparing the population and sample histograms of the covariate clearly shows
that locations with small and large values for the covariate are preferentially
selected (Figure 23.6). The optimised sampling pattern is a compromise between
spreading in geographic space and covariate space, see also Heuvelink et al.
(2007) and Brus and Heuvelink (2007). More precisely, locations are selected by
spreading them throughout the study area, while accounting for the values of
the covariates at the selected locations, in a way that locations with covariate
values near the minimum and maximum are preferred. This can be explained by
422 23 Model-based optimisation of the sampling pattern

FIGURE 23.5: Optimised sampling pattern for KED of lnECe at the Cotton
Research Farm, using lnEM100cm as a covariate.

noting that the variance of the KED prediction error can be decomposed into
two components: the variance of the interpolated residuals and the variance
of the estimator of the model-mean, see Section 21.3. The contribution of the
first variance component is minimised through geographical spreading, that of
the second component by selecting locations with covariate values near the
minimum and maximum.

FIGURE 23.6: Sample frequency distribution and population frequency

distribution of lnEM100cm used as covariate in model-based optimisation of
the sampling pattern for mapping with KED.
23.3 Optimising the sampling pattern for kriging with an external drift 423

A sample with covariate values close to the minimum and maximum only
is not desirable if we do not want to rely on the assumption of a linear
relation between the study variable and the covariates. To identify a non-
linear relation, locations with intermediate covariate values are needed.
Optimisation using a semivariogram with clear spatial structure leads to
geographical spreading of the sampling units, so that most likely also locations
with intermediate covariate values are selected.

When one or more covariates are used in optimisation of the sampling pattern
but not used in KED once the data are collected, the sample is suboptimal for
the model used in prediction. Inversely, ignoring a covariate in optimisation
of the sampling pattern while using this covariate as a predictor also leads to
suboptimal samples. The selection of covariates to be used in sampling design
therefore should be done with care. Besides, as we will see in the next exercise,
the nugget of the residual semivariogram has a strong effect on the optimised
sampling pattern, stressing the importance of a reliable prior estimate of this
semivariogram parameter.

Exercises

3. Write an R script to optimise the sampling pattern of 16 points in

a square for KED. Use the 𝑥-coordinate as a covariate. First, create
a discretisation grid of 20 × 20 nodes. Use an exponential residual
semivariogram without nugget, with a sill of 2, and a distance pa-
rameter of four times the spacing of the discretisation grid. Optimise
the sampling pattern with SSA (using functions scheduleSPSANN and
optimMKV of package spsann).
• What do you think of the spatial coverage of the optimised
sample? Compare the sample with the optimised sample for
OK, see exercise of Section 23.2.
• Repeat the optimisation using a residual semivariogram with a
nugget of 1.5 and a partial sill of 0.5. Note that the sill is again
2, as before.
• Compare the optimised sample with the previous sample. What
is the most striking difference?
• How will the optimised sample look with a pure nugget semi-
variogram? Check your assumption using such semivariogram
in SSA.
424 23 Model-based optimisation of the sampling pattern

23.4 Model-based infill sampling for ordinary kriging

Similar to spatial infill sampling using MSSD as a minimisation criterion
(Section 17.3), we may design a model-based infill sample. Package spsann
can be used for this, using argument points of function optimMKV.
In Section 22.1 the legacy data of West-Amhara were used to estimate the
parameters of a spherical semivariogram for the SOM concentration. The
estimated parameters are shown in Table 22.1. The maximum likelihood
estimates are used in this section to optimise the spatial coordinates of the
infill sample.
In the next code chunk, a list is created containing a data frame with the
coordinates of the fixed points (specified with subargument fixed) and an integer
of the number of additional points to be selected (specified with subargument
free). The list is passed to function optimMKV with argument points. For kriging,
I reduced the number of legacy points by keeping one point only per grid cell
of 1 km × 1 km. This is done with function remove.duplicates of package sp.

library(sp)
coordinates(sampleAmhara) <- ~ s1 + s2
legacy <- remove.duplicates(sampleAmhara, zero = 1, remove.second = TRUE)
pnts <- list(fixed = coordinates(legacy), free = 100)
candi <- grdAmhara[, c("s1", "s2")]
names(candi) <- c("x", "y")

The number of points used in kriging can be passed to function optimMKV with
argument nmax.

set.seed(314)
vgm_ML_gstat <- vgm(model = "Sph", psill = vgm_ML$sigmasq,
range = vgm_ML$phi, nugget = vgm_ML$nugget)
res <- optimMKV(
points = pnts, candi = candi,
vgm = vgm_ML_gstat, eqn = z ~ 1,
nmax = 20, schedule = schedule, track = FALSE)
infillSample <- res$points %>%
filter(free == 1)

Figure 23.7 shows a model-based infill sample of 100 points for OK of the
soil organic matter (SOM) concentration (dag kg-1 ) throughout West-Amhara.
Comparison of the model-based infill sample with the spatial infill sample of
Figure 17.5 shows that in a wider zone on both sides of the roads no new
23.5 Model-based infill sampling for kriging with an external drift 425

sampling points are selected. This can be explained by the large range, 36.9
km, of the semivariogram.

FIGURE 23.7: Model-based infill sample for OK of the SOM concentration

throughout West-Amhara. Legacy units have free-value 0; infill units have
free-value 1.

23.5 Model-based infill sampling for kriging with an

external drift
For West-Amhara, maps of covariates are available that can be used in KED,
see Section 22.4. The prediction error variance with KED is partly determined
by the covariate values (see Section 23.3), and therefore, when filling in the
undersampled areas, locations with extreme values for the covariates are
426 23 Model-based optimisation of the sampling pattern

preferably selected. In Section 22.4 the legacy data were used to estimate the
residual semivariogram by REML, see Table 22.2. In the next code chunk,
the estimated parameters of the residual semivariogram are used to optimise
the spatial pattern of an infill sample of 100 points for mapping the SOM
concentration throughout West-Amhara by KED, using elevation (dem), NIR-
reflectance (rfl_NIR), red-reflectance (rfl_red), and land surface temperature
(lst) as predictors for the model-mean.

covars <- grdAmhara[, c("dem", "rfl_NIR", "rfl_red", "lst")]

vgm_REML_gstat <- vgm(model = "Sph", psill = vgm_REML$sigmasq,
range = vgm_REML$phi, nugget = vgm_REML$nugget)
set.seed(314)
res <- optimMKV(
points = pnts, candi = candi, covars = covars,
vgm = vgm_REML_gstat,
eqn = z ~ dem + rfl_NIR + rfl_red + lst,
nmax = 20, schedule = schedule, track = TRUE)

Figure 23.8 shows the optimised sample. Again the legacy points are avoided,
but the infill sampling of the undersampled areas is less uniform compared to
Figure 23.7. Spreading in geographical space is less important than with OK
because the residual semivariogram has a much smaller range (Table 22.2).
Spreading in covariate space does not play any role with OK, whereas with
KED selecting locations with extreme values for the covariates is important to
minimise the uncertainty about the estimated model-mean.
The MKV of the optimised sample equals 0.878 (dag kg-1 )2 , which is somewhat
larger than the sill (sum of nugget and partial sill) of the residual semivariogram
(Table 22.2). This can be explained by the very small range of the semivariogram,
so that ignoring the uncertainty about the model-mean, the kriging variance at
nearly all locations in the study area equals the sill. Besides, we are uncertain
about the model-mean, explaining that the MKV can be larger than the sill.
23.5 Model-based infill sampling for kriging with an external drift 427

FIGURE 23.8: Model-based infill sample for KED of the SOM concentration
throughout West-Amhara, plotted on a map of one of the covariates. Legacy
units have free-value 0; infill units have free-value 1.
24
Sampling for estimating the semivariogram

For model-based sampling as described in Chapters 22 and 23, we must specify

the (residual) semivariogram. In case we do not have a reasonable prior estimate
of the semivariogram, we may decide to first collect data with the specific aim
of estimating the semivariogram. This semivariogram is subsequently used to
design a model-based sample for mapping. This chapter is about how to design
a reconnaissance sample survey for estimating the semivariogram.
The first question is how many observations we need for this. Webster and
Oliver (1992) gave as a rule of thumb that 150 to 225 points are needed to
obtain a reliable semivariogram when estimated by the method-of-moments
(MoM). Lark (2000) showed that with maximum likelihood (ML) estimation
two-thirds to only half of the observations are needed to achieve equal precision
of the estimated semivariogram parameters.
Once we have decided on the sample size, we must select the locations of the
sampling units. Two random sampling designs for semivariogram estimation are
described, nested sampling (Section 24.1) and independent sampling of pairs
of points (Section 24.2). Section 24.3 is devoted to model-based optimisation
of the sampling pattern for semivariogram estimation. Section 24.4 is about
how to design a single sample that can be used both for estimation of the
semivariogram and prediction (mapping). In a final section, a practical solution
is described for the problem of how to design a sample for semivariogram
estimation and prediction.

24.1 Nested sampling

Nested sampling can be used to estimate the semivariance at a limited number
of separation distances, see Oliver and Webster (1986) and Webster et al.
(2006). First, we must decide on these separation distances. We need point-
pairs at various separation distances, especially for small separation distances,
so that we get reliable estimates of this part of the semivariogram which has a
strong effect on the kriging weights. Usually, separation distances are chosen in
a geometric progression, for instance 2, 8, 32, 128, and 512 m. The multiplier,

429
430 24 Sampling for estimating the semivariogram

which is four in this example, should not be too small; as a rule of thumb, use
three or larger.
There are two versions of nested sampling. In the first stage of the first version,
several main stations are selected in a way that they cover the study area
well, for instance by spatial coverage sampling. In the second stage, each of
the main stations is used as a starting point to select one point at a distance
equal to the largest chosen separation distance (512 m in the example) in a
random direction from the main station. This doubles the sample size. In the
third stage, all points selected in the previous stages (main stations of stage
1 plus the points of stage 2) are used as starting points to select one point
at a distance equal to the second largest separation distance (128 m), and so
on. All points selected in the various stages are included in the nested sample.
The code chunk below shows the function for random selection of one point
at distance ℎ from a starting point. Note the while loop which continues until
a point is found that is inside the area. This is checked with function over of
package sp.

SelectPoint <- function(start, h, area) {

dxy <- numeric(length = 2)
inArea <- NA
while (is.na(inArea)) {
angle <- runif(n = 1, min = 0, max = 2 * pi)
dxy[1] <- h * sin(angle); dxy[2] <- h * cos(angle)
xypnt <- start + dxy
coordinates(xypnt) <- ~ s1 + s2
inArea <- as.numeric(over(x = xypnt, y = area))[1]
}
xypoint <- as.data.frame(xypnt)
xypoint
}

The first stage of the second version is equal to that of the first version. However,
in the second stage each of the main stations serves as a starting point for
randomly selecting a pair of points with a separation distance equal to the
largest chosen separation distance. The main station is halfway the selected
pair of points. In the third stage, each of the substations is used to select in
the same way a pair of points separated by the second largest chosen distance,
and so on. Only the points selected in the final stage are used as sampling
points. The R code below shows the function for random selection of two points
separated by ℎ distance units with a starting point halfway the pair of points.
The while loop continues until both points of a pair are inside the area.

SelectPair <- function(start, h, area) {

dxy <- numeric(length = 2)
24.1 Nested sampling 431

xypoints <- NULL

inArea1 <- inArea2 <- NA
while (is.na(inArea1) | is.na(inArea2)) {
angle <- runif(n = 1, min = 0, max = 2 * pi)
dxy[1] <- h * sin(angle) / 2; dxy[2] <- h * cos(angle) / 2
xypnt1 <- start + dxy
coordinates(xypnt1) <- ~ s1 + s2
inArea1 <- as.numeric(over(x = xypnt1, y = area))[1]
dxy[1] <- -dxy[1]; dxy[2] <- -dxy[2]
xypnt2 <- start + dxy
coordinates(xypnt2) <- ~ s1 + s2
inArea2 <- as.numeric(over(x = xypnt2, y = area))[1]
}
xypoints <- rbind(as.data.frame(xypnt1), as.data.frame(xypnt2))
xypoints
}

The R code below shows the selection of a nested sample from Hunter Valley
using both versions. Only one main station is selected. In total 16 points are
selected in four stages. The separation distances are 2,000, 1,000, 500, and
250 m. Sixteen points is not enough for estimating the semivariogram and a
multiplier of two is rather small, but this example is for illustrative purposes
only.
Note that the separation distances are in descending order. The largest sepa-
ration distance should not be chosen too large, because then, when the main
station is somewhere in the middle of the study area, it may happen that
using the first version, no pair can be found with that separation distance. A
similar problem may occur with the second version when in subsequent stages
a station is selected near the border of the study area. A copy of grdHunterValley
is made because both the original data frame is needed, as well as a gridded
version of this data frame.

library(sp)
grid <- grdHunterValley
gridded(grdHunterValley) <- ~ s1 + s2
lags <- c(2000, 1000, 500, 250)

The next code chunk is an implementation of the first version of nested

sampling.

set.seed(614)
unit <- sample(nrow(grid), 1)
mainstation <- grid[unit, c("s1", "s2")]
432 24 Sampling for estimating the semivariogram

newpnt <- SelectPoint(start = mainstation, h = lags[1], area = grdHunterValley)

mysample_nested <- rbind(mainstation, newpnt)
for (j in 2:length(lags)) {
newpnts <- NULL
for (i in seq_len(nrow(mysample_nested))) {
pnts <- SelectPoint(
start = mysample_nested[i, ], h = lags[j], area = grdHunterValley)
newpnts <- rbind(newpnts, pnts)
}
mysample_nested <- rbind(mysample_nested, newpnts)
}

The R code for the second version is presented in the next code chunk.

unit <- sample(nrow(grid), 1)

mainstation <- grid[unit, c("s1", "s2")]
pnt <- SelectPoint(start = mainstation, h = lags[1], area = grdHunterValley)
stations <- rbind(mainstation, pnt)
allstations <- rbind(mainstation, pnt)
for (j in 2:length(lags)) {
newstations <- NULL
for (i in seq_len(nrow(stations))) {
pnts <- SelectPair(
start = stations[i, ], h = lags[j], area = grdHunterValley)
newstations <- rbind(newstations, pnts)
allstations <- rbind(allstations, pnts)
}
stations <- newstations
}
mysample_nested_2 <- as_tibble(stations)

Figure 24.1 shows the two selected nested samples. For the sample selected
with the second version, also the stations that served as starting points for the
selection of the point-pairs are plotted.
The samples of Figure 24.1 are examples of balanced nested samples. The
number of point-pairs separated by a given distance doubles with every stage.
As a consequence, the estimated semivariances for the smallest separation
distance are much more precise than for the largest distance. We are most
uncertain about the estimated semivariances for the largest separation distances.
If in the first stage only one point-pair separated by the largest distance is
selected, then we have only one degree of freedom for estimating the variance
component associated with this stage. It is more efficient to select more than
one main station, say about ten, and to select fewer points in the final stages.
For instance, with the second version we may decide to select a point-pair at
24.1 Nested sampling 433

FIGURE 24.1: Balanced nested samples from Hunter Valley, selected with
the two versions of nested sampling. In the subfigure of the second version
the selected sampling points (symbol x) are plotted together with the selected
stations (halfway the two points of a pair).
434 24 Sampling for estimating the semivariogram

only half the number of stations selected in the one-but-last stage. The nested
sample then becomes unbalanced.
The model for nested sampling with four stages is a hierarchical analysis of
variance (ANOVA) model with random effects:

𝑍u�u�u�u� = 𝜇 + 𝐴u� + 𝐵u�u� + 𝐶u�u�u� + 𝜖u�u�u�u� , (24.1)

with 𝜇 the mean, 𝐴u� the effect of the 𝑖th first stage station, 𝐵u�u� the effect
of the 𝑗th second stage station within the 𝑖th first stage station, and so on.
𝐴u� , 𝐵u�u� , 𝐶u�u�u� , and 𝜖u�u�u�u� are random quantities (random effects) all with zero
mean and variances 𝜎21 , 𝜎22 , 𝜎23 , and 𝜎24 , respectively.
For balanced designs, the variance components can be estimated by the MoM
from a hierarchical ANOVA. The first step is to assign factors to the sampling
points that indicate the grouping of the sampling points in the various stages.
The number of factors needed is the number of stages minus 1. All factors have
two levels. Figures 24.2 and 24.3 show the levels of the three factors. The levels
of the first factor show the strongest spatial clustering, those of the second
factor the one-but-strongest, and so on.

mysample_nested$factor1 <- rep(rep(1:2), times = 8)

mysample_nested$factor2 <- rep(rep(1:2, each = 2), times = 4)
mysample_nested$factor3 <- rep(rep(1:2, each = 4), times = 2)

FIGURE 24.2: The levels of the three factors assigned to the sampling points
of the balanced nested sample selected with the first version.

The R code below shows the construction of the three factors for the second
version of nested sampling.
24.1 Nested sampling 435

mysample_nested_2$factor1 <- rep(1:2, each = 8)

mysample_nested_2$factor2 <- rep(rep(1:2, each = 4), times = 2)
mysample_nested_2$factor3 <- rep(rep(1:2, each = 2), times = 4)

FIGURE 24.3: The levels of the three factors assigned to the sampling points
of the balanced nested sample selected with the second version.

For unbalanced nested designs, the variance components can be estimated

by restricted maximum likelihood (REML) (Webster et al., 2006). REML
estimation is also recommended if in Equation (24.1) instead of a constant
mean 𝜇 the mean is a linear combination of one or more covariates (fixed
effects). The semivariances at the chosen separation distances are obtained by
cumulating the estimated variance components.
The R code below shows how the variance components and the semivariances
can be estimated with function lme of the package nlme (Pinheiro et al., 2021),
once the data are collected and added to the data frame. This function fits
linear mixed-effects models and allows for nested random effects. It can be used
both for balanced and unbalanced nested samples, and for a constant mean or
a mean that is a linear combination of covariates. Argument fixed is a formula
describing the fixed effects with the response variable on the left-hand side of
the ~ operator and the covariates for the mean on the right-hand side. If a
constant mean is assumed, as in our example, this is indicated by the 1 on the
right-hand side of the ~ operator. Argument random is a one-sided formula (no
response variable is on the left-hand side of the ~ operator). On the right-hand
side of the | separator, the nested structure of the data is specified using the
factors of Figure 24.2. The 1 on the left-hand side of the | separator means
that we assume that all regression coefficients associated with the covariates
are fixed (non-random) quantities.

library(nlme)
lmodel <- lme(
436 24 Sampling for estimating the semivariogram

fixed = z ~ 1, data = mysample_nested,

random = ~ 1 | factor1 / factor2 / factor3)
res <- as.matrix(VarCorr(lmodel))
sigmas <- as.numeric(res[c(2, 4, 6, 7), 1])
sigma <- rev(sigmas)
semivar <- cumsum(sigmas)

Random sampling of the points is not strictly needed, because a model-based

approach is followed here. The model of Equation (24.1) is a superpopulation
model, i.e., we assume that the population is generated by this model (see
Chapter 26). Papritz et al. (2011), for instance, selected the points (using the
second version) non-randomly to improve the control of the nested subareas
and the average separation distances.
Lark (2011) describes a method for optimisation of a nested design, given the
total number of points and the chosen separation distances.

Exercises

1. Write an R script to select with the first version a balanced nested

sample from Hunter Valley. Use as separation distances 1,000, 500,
200, 100, and 50 m.
• Add the factors that are needed for estimating the variance
components to the data frame with the selected sampling points.
• Overlay the sampling points with the SpatialPixelsDataFrame,
and estimate the semivariances for the attribute compound
topographic index (cti).

24.2 Independent sampling of pairs of points

With the nested design, the estimated semivariances for the different separation
distances are not independent. Independent estimated semivariances can be
obtained by independent random selection of pairs of points (IPP sampling).
Independence here means design-independence, see Section 26.2. Similar to a
regression model, a semivariogram can be defined as a superpopulation model
or as a population model. Only in the current section a semivariogram is defined
at the population level. Such a semivariogram is referred to as a non-ergodic
semivariogram or local semivariogram (Brus and de Gruijter, 1994).
IPP sampling is straightforward for simple random sampling. For each sepa-
ration distance a point-pair is selected by first selecting fully randomly one
24.2 Independent sampling of pairs of points 437

point from the study area. Then the second point is randomly selected from
the circle with the first point at its centre and a radius equal to the chosen
separation distance. If this second point is outside the study area, both points
are discarded. This is repeated until we have the required point-pairs for this
separation distance. The next code chunk is an implementation of this selection
procedure.

SIpairs <- function(h, n, area) {

topo <- as(getGridTopology(area), "data.frame")
cell_size <- topo$cellsize[1]
xy <- coordinates(area)
dxy <- numeric(length = 2)
xypnts1 <- xypnts2 <- NULL
i <- 1
while (i <= n) {
unit1 <- sample(length(area), size = 1)
xypnt1 <- xy[unit1, ]
xypnt1[1] <- jitter(xypnt1[1], amount = cell_size / 2)
xypnt1[2] <- jitter(xypnt1[2], amount = cell_size / 2)
angle <- runif(n = 1, min = 0, max = 2 * pi)
dxy[1] <- h * sin(angle); dxy[2] <- h * cos(angle)
xypnt2 <- as.data.frame(t(xypnt1 + dxy))
coordinates(xypnt2) <- ~ s1 + s2
inArea <- as.numeric(over(x = xypnt2, y = area))[1]
if (!is.na(inArea)) {
xypnts1 <- rbind(xypnts1, xypnt1)
xypnts2 <- rbind(xypnts2, as.data.frame(xypnt2))
i <- i + 1
}
rm(xypnt1, xypnt2)
}
cbind(xypnts1, xypnts2)
}

IPP sampling is illustrated with the compound topographic index (cti, which is
the same as topographic wetness index) data of Hunter Valley. Five separation
distances are chosen, collected in numeric h, and for each distance 𝑛 = 100
point-pairs are selected by simple random sampling.

library(sp)
h <- c(50, 100, 200, 500, 1000)
n <- 100
set.seed(123)
allpairs <- NULL
438 24 Sampling for estimating the semivariogram

for (i in seq_len(length(h))) {
pairs <- SIpairs(h = h[i], n = n, area = grdHunterValley)
allpairs <- rbind(allpairs, pairs, make.row.names = FALSE)
}

The data.frame allpairs has four variables: the spatial coordinates of the first
and of the second point of a pair. An overlay is made of the selected points
with the SpatialPixelsDataFrame, and the cti values are extracted.

pnt1 <- allpairs[, c(1, 2)]

coordinates(pnt1) <- ~ s1 + s2
z1 <- over(x = pnt1, y = grdHunterValley)["cti"]
pnt2 <- allpairs[, c(3, 4)]
coordinates(pnt2) <- ~ s1 + s2
z2 <- over(x = pnt2, y = grdHunterValley)["cti"]
mysample <- data.frame(h = rep(h, each = n), z1, z2)
names(mysample)[c(2, 3)] <- c("z1", "z2")

The semivariances for the chosen separation distances are estimated as well as
the variance of these estimated semivariances.

gammah <- vgammah <- numeric(length = length(h))

for (i in seq_len(length(h))) {
units <- which(mysample$h == h[i])
pairsh <- mysample[units, ]
gammah[i] <- mean((pairsh$z1 - pairsh$z2)ˆ2, na.rm = TRUE) / 2
vgammah[i] <- var((pairsh$z1 - pairsh$z2)ˆ2, na.rm = TRUE) / (n * 4)
}

A spherical model with nugget is fitted to the sample semivariogram, using

function nls, with weights equal to the reciprocal of the estimated variances of
the estimated semivariances.

sample_vg <- data.frame(h, gammah, vgammah)

SphNug <- function(h, range, psill, nugget) {
h <- h / range
nugget + psill * ifelse(h < 1, (1.5 * h - 0.5 * hˆ3), 1)
}
fit.var <- nls(gammah ~ SphNug(h, range, psill, nugget),
data = sample_vg, start = list(psill = 4, range = 200, nugget = 1),
weights = 1 / vgammah, algorithm = "port", lower = c(0, 0, 0))
print(pars <- signif(coef(fit.var), 3))
24.2 Independent sampling of pairs of points 439

psill range nugget

3.29 188.00 1.21

Figure 24.4 shows the sample semivariogram and the fitted model.

FIGURE 24.4: Sample semivariogram, obtained by independent sampling

of pairs of points, and fitted spherical model of compount topographic index
in Hunter Valley.

The covariances of the estimated semivariances at different separation distances

are zero, as the point-pairs are selected independently. This keeps estimation
of the variances and covariances of the estimated semivariogram parameters
simple. In the next code chunk, this is done by bootstrapping.
In bootstrapping, for each separation distance a simple random sample with
replacement of point-pairs is selected from the original sample of point-pairs.
A point-pair can be selected more than once. The sample size (number of
draws) is equal to the total number of point-pairs per separation distance in
the original sample.
Every run of the bootstrap results in as many bootstrap samples as there are
separation distances. The bootstrap samples are used to fit a semivariogram
model. The whole procedure is repeated 500 times, resulting in 500 vectors
with model parameters. These vectors can be used to estimate the variances
and covariances of the estimators of the three semivariogram parameters.

allpars <- NULL

R <- 500
for (j in 1:R) {
gammah <- vgammah <- numeric(length = length(h))
440 24 Sampling for estimating the semivariogram

for (i in seq_len(length(h))) {
units <- which(mysample$h == h[i])
mysam_btsp <- mysample[units, ] %>%
slice_sample(n = n, replace = TRUE)
gammah[i] <- mean((mysam_btsp$z1 - mysam_btsp$z2)ˆ2, na.rm = TRUE) / 2
vgammah[i] <- var((mysam_btsp$z1 - mysam_btsp$z2)ˆ2, na.rm = TRUE) / (n * 4)
}
sample_vg <- data.frame(h, gammah, vgammah)
tryCatch({
fittedvariogram <- nls(gammah ~ SphNug(h, range, psill, nugget),
data = sample_vg, start = list(psill = 4, range = 200, nugget = 1),
weights = 1 / vgammah, algorithm = "port", lower = c(0, 0, 0))
pars <- coef(fittedvariogram)
allpars <- rbind(allpars, pars)}, error = function(e) {})
}
#compute variance-covariance matrix
signif(var(allpars), 3)

psill range nugget

psill 1.160 -44.6 -0.803
range -44.600 66700.0 172.000
nugget -0.803 172.0 1.070

Note the large variance for the range parameter (the standard deviation is
258 m) as well as the negative covariance of the nugget and the partial sill
parameter (the Pearson correlation coefficient is -0.72). Histograms of the three
estimated semivariogram parameters are shown in Figure 24.5.

Marcelli et al. (2019) show how a probability sample of points (instead of pairs
of points) can be used in design-based estimation of the semivariogram. From
the 𝑛 randomly selected points all 𝑛(𝑛−1)/2 point-pairs are constructed. The
second-order inclusion probabilities of these point-pairs are used to estimate
the mean semivariance for separation distance classes. This sampling strategy
makes better use of the data and is therefore potentially more efficient than
IPP sampling.

Exercises

2. Write an R script to select simple random samples of pairs of points

for estimating the semivariogram of cti in Hunter Valley. Use as
separation distances 25, 50, 100, 200, and 400 m. Note that these
separation distances are smaller than those used above. Select 100
pairs per separation distance.
24.3 Optimisation of sampling pattern for semivariogram estimation 441

FIGURE 24.5: Frequency distributions of estimated parameters of a spherical

semivariogram of compound topographic index in Hunter Valley.

• Compute the sample semivariogram, and estimate a spherical

model with nugget using function nls.
• Compare the estimated semivariogram parameters with the
estimates obtained with the larger separation distances.
• Estimate the variance-covariance matrix of the estimated semi-
variogram parameters by bootstrapping.
• Compare the variances of the estimated semivariogram param-
eters with the variances obtained with the larger separation
distances. Which variance is changed most?

24.3 Optimisation of sampling pattern for

semivariogram estimation
There is rich literature on model-based optimisation of the sampling locations
for semivariogram estimation. Several design criteria (minimisation criteria)
have been proposed for optimising the sampling pattern. Various authors have
proposed a measure of the uncertainty about the semivariogram parameters as
a minimisation criterion. These criteria are described and illustrated in the
next subsection. The kriging variance is sensitive to errors in the estimated
semivariogram. Therefore, Lark (2002) proposed to use a measure of the
442 24 Sampling for estimating the semivariogram

uncertainty about the kriging variance as a minimisation criterion (Subsection

24.3.2).

24.3.1 Uncertainty about semivariogram parameters

Müller and Zimmerman (1999) as well as Bogaert and Russo (1999) pro-
posed the determinant of the variance-covariance matrix of semivariogram
parameters, estimated by generalised least squares to fit the MoM sample
semivariogram. For instance, if we have two semivariogram parameters, 𝜃1
and 𝜃2 , the determinant of the 2 × 2 variance-covariance matrix equals the
sum of the variances of the two estimated parameters minus two times the
covariance of the two estimated parameters. If the two estimated parameters
are positively correlated, the determinant of the matrix is smaller than if
they are uncorrelated, and the covariance term is zero. The determinant is a
measure of our joint uncertainty about the semivariogram parameters.
Zhu and Stein (2005) proposed as a minimisation criterion the log of the
determinant of the inverse Fisher information matrix in ML estimation of
the semivariogram, hereafter shortly denoted by logdet. The Fisher informa-
tion about a semivariogram parameter is a function of the likelihood of the
semivariogram parameter; the likelihood of a semivariogram parameter is the
probability of the data as a function of the semivariogram parameter. The log
of this likelihood can be plotted against values of the parameter. The flatter the
log-likelihood surface, the less information is in the data about the parameter.
The flatness of the surface can be measured by the first derivative of the
log-likelihood to the semivariogram parameter. Strong negative or positive
derivative values indicate a steep surface. The Fisher information for a model
parameter is defined as the expectation of the square of the first derivative of
the log-likelihood to that semivariogram parameter, see Ly et al. (2017) for a
nice tutorial on this subject. The more information we have about a semivari-
ogram, the less uncertain we are about that parameter. This explains why the
inverse of the Fisher information can be used as a measure of uncertainty. The
inverse Fisher information matrix contains the variances and covariances of
the estimated semivariogram parameters.
The code chunks hereafter show how logdet can be computed. It makes use
of the result of Kitanidis (1987) who showed that each element of the Fisher
information matrix 𝐈(𝜃) can be obtained with (see also Lark (2002))

1 𝜕𝐀 −1 𝜕𝐀
[𝐈(𝜃)]u�u� = Tr [𝐀−1 𝐀 ] , (24.2)
2 𝜕𝜃u� 𝜕𝜃u�

with 𝐀 the correlation matrix of the sampling points, u�u�

u�u�
the partial derivative
u�
of the correlation matrix to the 𝑖th semivariogram parameter, and Tr[⋅] the
trace of a matrix.
24.3 Optimisation of sampling pattern for semivariogram estimation 443

As an illustration, I selected a simple random sample of 50 points from Hunter

Valley. A matrix with distances between the points of a sample is computed.
Preliminary values for the semivariogram parameters 𝜉 (ratio of spatial depen-
dence) and 𝜙 (distance parameter) are obtained by visual inspection of the
sample semivariogram, and the gstat (Pebesma, 2004) function variogramLine
is used to compute the correlation matrix.

library(sp)
library(gstat)
set.seed(314)
mysample0 <- grdHunterValley %>%
slice_sample(n = 50)
coordinates(mysample0) <- ~ s1 + s2
D <- spDists(mysample0)
xi <- 0.8; phi <- 200
thetas <- c(xi, phi)
vgmodel <- vgm(model = "Exp", psill = thetas[1],
range = thetas[2], nugget = 1 - thetas[1])
A <- variogramLine(vgmodel, dist_vector = D, covariance = TRUE)

In the next step, the semivariogram parameters are slightly changed one-by-
one. The changes, referred to as perturbations, are a small fraction of the
preliminary semivariogram parameter values. The perturbed semivariogram
parameters are used to compute the perturbed correlation matrices (pA) and
the partial derivatives of the correlation matrix (dA) for each perturbation.

perturbation <- 0.01

pA <- dA <- list()
for (i in seq_len(length(thetas))) {
thetas_pert <- thetas
thetas_pert[i] <- (1 + perturbation) * thetas[i]
vgmodel_pert <- vgm(model = "Exp", psill = thetas_pert[1],
range = thetas_pert[2], nugget = 1 - thetas_pert[1])
pA[[i]] <- variogramLine(vgmodel_pert, dist_vector = D, covariance = TRUE)
dA[[i]] <- (pA[[i]] - A) / (thetas[i] * perturbation)
}

Finally, the Fisher information matrix is computed using Equation (24.2).

We do not need to compute the inverse of the Fisher information matrix,
because the determinant of the inverse of a matrix is equal to the inverse of
the determinant of the (not inverted) matrix. The determinant is computed
with function determinant.
444 24 Sampling for estimating the semivariogram

I <- matrix(0, length(thetas), length(thetas))

for (i in seq_len(length(thetas))) {
m_i <- solve(A, dA[[i]])
for (j in i:length(thetas)) {
m_j <- solve(A, dA[[j]])
I[i, j] <- I[j, i] <- 0.5 * sum(diag(m_i %*% m_j))
}
}
logdet0 <- -determinant(I, logarithm = TRUE)$modulus

The joint uncertainty about the semivariogram parameters, as quantified by

the log of the determinant of the inverse of the information matrix, equals
8.198. Hereafter, we will see how much this joint uncertainty can be reduced by
optimising the spatial pattern of the sample used for semivariogram estimation,
compared to the simple random sample used in the above calculation. Note
that a preliminary semivariogram is needed to compute an optimised sampling
pattern for semivariogram estimation.
Function optimUSER of package spsann can be used to search for the sampling
locations with the minimum value of logdet. This function has been used before
in Section 23.2. Package spsann cannot deal with the 22,124 candidate grid
nodes of Hunter Valley; these are too many. I therefore selected a subgrid of
50 m × 50 m.
The size of the sample for estimation of the semivariogram is passed to function
optimUSER with argument points. The objective function to be minimised is
passed to function optimUSER with argument fun. The objective function logdet
is defined in package sswr. Argument model specifies the model type, using the
characters for model types of package gstat. Argument thetas specifies the
preliminary semivariogram parameter values. Argument perturbation specifies
how much the semivariogram parameters are changed to compute the perturbed
correlation matrices (pA) and the partial derivatives of the correlation matrix
(dA).

gridded(grdHunterValley) <- ~ s1 + s2
candi <- spsample(grdHunterValley, type = "regular",
cellsize = c(50, 50), offset = c(0.5, 0.5))
candi <- as.data.frame(candi)
names(candi) <- c("x", "y")
schedule <- scheduleSPSANN(
initial.acceptance = c(0.8, 0.95),
initial.temperature = 0.15, temperature.decrease = 0.9,
chains = 300, chain.length = 10, stopping = 10,
x.min = 0, y.min = 0, cellsize = 50)
set.seed(314)
24.3 Optimisation of sampling pattern for semivariogram estimation 445

res <- optimUSER(

points = 50, candi = candi,
fun = logdet,
model = "Exp", thetas = thetas, perturbation = 0.01,
schedule = schedule, track = TRUE)

FIGURE 24.6: Optimised sampling pattern of 100 points for semivariogram

estimation, using the log of the determinant of the inverse Fisher information
matrix of the semivariogram parameters (logdet) and the mean estimation
variance of the kriging variance (MVKV) as a minimisation criterion.

Figure 24.6 shows the optimised sampling pattern of 50 points. The logdet of
the optimised sample equals 3.548, which is 43% of the value of the simple
random sample used above to illustrate the computations. The optimised
sample consists of two clusters. There are quite a few point-pairs with nearly
coinciding points.

24.3.2 Uncertainty about the kriging variance

Lark (2002) proposed as a minimisation criterion the estimation variance of

the kriging variance (VKV) due to uncertainty in the ML estimates of the
semivariogram parameters. This variance is approximated by a first order
Taylor series, requiring the partial derivatives of the kriging variance with
respect to the semivariogram parameters:
u� u�
𝜕𝑉OK (𝐬0 ) 𝜕𝑉OK (𝐬0 )
𝑉𝐾𝑉(𝐬0 ) = ∑ ∑ Cov(𝜃u� , 𝜃u� ) , (24.3)
u�=1 u�=1
𝜕𝜃u� 𝜕𝜃u�
446 24 Sampling for estimating the semivariogram

with 𝑝 the number of semivariogram parameters, Cov(𝜃u� , 𝜃u� ) the covariances

of the semivariogram parameters 𝜃u� and 𝜃u� (elements of the inverse of the
Fisher information matrix 𝐈−1 (𝜃𝜃), Equation (24.2)), and u�u�u�u�
OK (u�0 )
the partial
u�
derivative of the kriging variance to the 𝑖th semivariogram parameter at
prediction location 𝐬0 .
The first step in designing a sample for semivariogram estimation using a
population parameter of VKV as a minimisation criterion is to select a sample
for the second sampling round. In the code chunk below, a spatial coverage
sample of 100 points is selected, using function stratify of package spcosa, see
Section 17.2. Once the observations of this sample are collected in the second
sampling round, these data are used for mapping by ordinary kriging.
To optimise the sampling pattern of the first sampling round for variogram
estimation, the population mean of VKV (MVKV) is used as a minimisation
criterion. This population mean is estimated from a centred square grid of 200
points, the evaluation sample.

library(spcosa)
gridded(grdHunterValley) <- ~ s1 + s2
set.seed(314)
mystrata <- stratify(grdHunterValley, nStrata = 100, equalArea = FALSE, nTry = 10)
mysample_SC <- as(spsample(mystrata), "SpatialPoints")
mysample_eval <- spsample(
x = grdHunterValley, n = 200, type = "regular", offset = c(0.5, 0.5))

The following code chunks show how VKV at the evaluation point is computed.
First, the correlation matrix of the spatial coverage sample (A) is computed as
well as the correlation matrix of the spatial coverage sample and the evaluation
points (A0). Correlation matrix A is extended with a column and a row with
ones, see Equation (21.5).

D <- spDists(mysample_SC)
vgmodel <- vgm(model = "Exp", psill = thetas[1], range = thetas[2],
nugget = 1 - thetas[1])
A <- variogramLine(vgmodel, dist_vector = D, covariance = TRUE)
nobs <- length(mysample_SC)
B <- matrix(data = 1, nrow = nobs + 1, ncol = nobs + 1)
B[1:nobs, 1:nobs] <- A
B[nobs + 1, nobs + 1] <- 0
D0 <- spDists(x = mysample_eval, y = mysample_SC)
A0 <- variogramLine(vgmodel, dist_vector = D0, covariance = TRUE)
b <- cbind(A0, 1)
24.3 Optimisation of sampling pattern for semivariogram estimation 447

Next, the semivariogram parameters are perturbed one-by-one, and the per-
turbed correlation matrices pA and pA0 are computed.

pA <- pA0 <- list()

for (i in seq_len(length(thetas))) {
thetas_pert <- thetas
thetas_pert[i] <- (1 + perturbation) * thetas[i]
vgmodel_pert <- vgm(model = "Exp", psill = thetas_pert[1],
range = thetas_pert[2], nugget = 1 - thetas_pert[1])
pA[[i]] <- variogramLine(vgmodel_pert, dist_vector = D, covariance = TRUE)
pA0[[i]] <- variogramLine(vgmodel_pert, dist_vector = D0, covariance = TRUE)
}
pB <- pb <- list()
for (i in seq_len(length(thetas))) {
pB[[i]] <- B
pB[[i]][1:nobs, 1:nobs] <- pA[[i]]
pb[[i]] <- cbind(pA0[[i]], 1)
}

Next, the kriging variance and the perturbed kriging variances are computed,
and the partial derivatives of the kriging variance with respect to the semivari-
ogram parameters are approximated. See Equations (21.7) and (21.8) for how
the kriging weights l and the kriging variance var are computed.

var <- numeric(length = length(mysample_eval))

pvar <- matrix(nrow = length(mysample_eval), ncol = length(thetas))
for (i in seq_len(length(mysample_eval))) {
#compute kriging weights and Lagrange multiplier
l <- solve(B, b[i,])
var[i] <- 1 - l[1:nobs] %*% A0[i, ] - l[nobs + 1]
for (j in seq_len(length(thetas))) {
pl <- solve(pB[[j]], pb[[j]][i, ])
pvar[i, j] <- 1 - pl[1:nobs] %*% pA0[[j]][i, ] - pl[nobs + 1]
}
}
dvar <- list()
for (i in seq_len(length(thetas))) {
dvar[[i]] <- (pvar[, i] - var) / (thetas[i] * perturbation)
}

Finally, the partial derivatives of the kriging variance are used to approximate
VKV at the 200 evaluation points (Equation (24.3)). For this, the variances and
covariances of the estimated semivariogram parameters are needed, estimated
by the inverse of the Fisher information matrix. The Fisher information matrix
448 24 Sampling for estimating the semivariogram

computed in Subsection 24.3.1 for the simple random sample of 50 points is

also used here.
Note that the variance-covariance matrix of estimated semivariogram param-
eters is computed from the sample for semivariogram estimation only. The
spatial coverage sample of the second sampling round is not used for estimating
the semivariogram, but for prediction only. Section 24.4 is about designing one
sample, instead of two samples, that is used both for estimation of the model
parameters and for prediction.

invI <- solve(I)

VKV <- numeric(length = length(var))
for (i in seq_len(length(thetas))) {
for (j in seq_len(length(thetas))) {
VKVij <- invI[i, j] * dvar[[i]] * dvar[[j]]
VKV <- VKV + VKVij
}
}
MVKV0 <- mean(VKV)

For the simple random sample, the square root of MVKV equals 0.223. The
mean kriging variance (MKV) at these points equals 0.787, so the uncer-
tainty about the kriging variance is substantial. Hereafter, we will see how
much MVKV can be reduced by optimising the sampling pattern with spatial
simulated annealing.
As for logdet, the sample with minimum value for MVKV can be searched for
using spsann function optimUSER. The objective function MVKV is defined in pack-
age sswr. Argument points specifies the size of the sample for semivariogram
estimation. Argument psample is to specify the sample used for prediction at
the evaluation points (after the second round of sampling). Argument esample
is to specify the sample with evaluation points for estimating MVKV. The
optimisation requires substantial computing time. With 200 evaluation points
and the annealing schedule specified below the computing time was 46.25
minutes (processor AMD Ryzen 5, 16 GB RAM).

schedule <- scheduleSPSANN(

initial.acceptance = c(0.8, 0.95),
initial.temperature = 0.002, temperature.decrease = 0.8,
chains = 300, chain.length = 2, stopping = 10,
x.min = 0, y.min = 0, cellsize = 50)
set.seed(314)
res <- optimUSER(
points = 50, candi = candi,
fun = MVKV,
24.3 Optimisation of sampling pattern for semivariogram estimation 449

psample = mysample_SC, esample = mysample_eval,

model = "Exp", thetas = thetas, perturbation = 0.01,
schedule = schedule, track = TRUE)

Figure 24.6 shows the optimised sample. The minimised value of MVKV is 29%
of the value of the simple random sample used to illustrate the computations.
The optimised sample points are clustered in an ellipse.

Both minimisation criteria, logdet and MVKV, are a function of the semivari-
ogram parameters 𝜃 , showing that the problem is circular. Using a preliminary
estimate of the semivariogram parameters, 𝜃 ,̂ leads to a locally optimal design
at 𝜃 .̂ For this reason, Bogaert and Russo (1999) and Zhu and Stein (2005)
proposed a Bayesian approach, in which a multivariate prior distribution for
the semivariogram parameters is postulated. The expected value over this
distribution of the criterion is minimised. Lark (2002) computed the average
of VKV over a number of semivariograms.

Both methods for sample optimisation rely, amongst others, on the assumption
that the mean and the variance are constant throughout the area. Under this
assumption, it is no problem that the sampling units are spatially clustered.
So, we assume that the semivariogram estimated from the data collected in a
small portion of the study area is representative for the whole study area. If
we do not feel comfortable with this assumption, spreading the sampling units
throughout the study area by the sampling methods described in the next two
sections can be a good option.

Exercises

3. Write an R script to design a model-based sample of 50 points for

Hunter Valley, to estimate the semivariogram for a study variable.
Use logdet as a minimisation criterion. Use as a prior estimate of the
semivariogram an exponential model with a distance parameter of
200 m and a ratio of spatial dependence of 0.5. Compare the sample
with the optimised sample in Figure 24.6, which was obtained with
the same value for the distance parameter and a spatial dependence
ratio of 0.8.

4. Repeat this for MVKV as a minimisation criterion. Use a spatial

coverage sample of 100 points for prediction and a square grid of
200 points for evaluation.
450 24 Sampling for estimating the semivariogram

24.4 Optimisation of sampling pattern for

semivariogram estimation and mapping
In practice, a reconnaissance survey for semivariogram estimation often is not
feasible. A single sample must be designed that is suitable both for estimating
the semivariogram parameters and mapping, i.e., prediction with the estimated
semivariogram parameters at the nodes of a fine discretisation grid. Another
reason is that in a reconnaissance survey we can seldom afford a sample size
large enough to obtain reliable estimates of the semivariogram parameters.
Papritz et al. (2011) found that for a sample size of 192 points the estimated
variance components with balanced and unbalanced nested designs were highly
uncertain. For this reason, it is attractive to use also the sampling points
designed for spatial prediction (mapping) for estimating the semivariogram.
Designing two samples, one for estimation of the semivariogram and one for
spatial prediction, is suboptimal. Designing one sample that can be used both
for estimation of the semivariogram parameters and for prediction potentially
is more efficient.
Finally, with nested sampling and IPP sampling, we aim at estimating the
semivariogram of the residuals of a constant mean (see Equation (24.1)). In
other words, with these designs we aim at estimating the parameters of a
semivariogram model used in ordinary kriging. In situations where we have
covariates that can partly explain the spatial variation of the study variable,
kriging with an external drift is more appropriate. In these situations, the
reconnaissance survey should be tailored to estimating both the regression
coefficients associated with the covariates and the parameters of the residual
semivariogram.
Model-based methods for designing a single sample for estimation of the
model parameters and for prediction with the estimated model parameters are
proposed, amongst others, by Zimmerman (2006), Zhu and Stein (2006), Zhu
and Zhang (2006), and Marchant and Lark (2007). The methods use a different
minimisation criterion. Zimmerman (2006) proposed to minimise the kriging
variance (at the centre of a square grid cell) that is augmented by an amount
that accounts for the additional uncertainty in the kriging predictions due to
uncertainty in the estimated semivariogram parameters, hereafter referred to
as the augmented kriging variance (AKV):

𝐴𝐾𝑉(𝐬0 ) = 𝑉OK (𝐬0 ) + E[𝜏2 (𝐬0 )] , (24.4)

with 𝑉OK (𝐬0 ) the ordinary kriging variance, see Equation (21.8), and E[𝜏2 (𝐬0 )]
the expectation of the additional variance component due to uncertainty about
the semivariogram parameters estimated by ML. The additional variance
24.4 Optimisation of sampling pattern 451

component is approximated by a first order Taylor series:

u� u�
𝜕𝜆T 𝜕𝜆
E[𝜏2 (𝐬0 )] = ∑ ∑ Cov(𝜃u� , 𝜃u� ) 𝐀 , (24.5)
u�=1 u�=1
𝜕𝜃u� 𝜕𝜃u�

with u�u�
u�u�
the vector of partial derivatives of the kriging weights with respect
u�
to the 𝑗th semivariogram parameter. Comparing Equations (24.5) and (24.3)
shows that the two variances differ. VKV quantifies our uncertainty about the
estimated kriging variance, whereas E[𝜏2 ] quantifies our uncertainty about the
kriging prediction due to uncertainty about the semivariogram parameters.
I use the mean of the AKV over the nodes of a prediction grid (evaluation
grid) as a minimisation criterion (MAKV). The same criterion can also be
used in situations where we have maps of covariates that we want to use in
prediction. In that case, the aim is to design a single sample that is used both
for estimation of the residual semivariogram and for prediction by kriging with
an external drift. The ordinary kriging variance 𝑉OK (𝐬0 ) in Equation (24.4)
is then replaced by the prediction error variance with kriging with an external
drift 𝑉KED (𝐬0 ), see Equation (21.20).
Zhu and Stein (2006) proposed as a minimisation criterion a linear combination
of AKV (Equation (24.4)) and VKV (Equation (24.3)), referred to as the
estimation adjusted criterion (EAC):

1
𝐸𝐴𝐶(𝐬0 ) = 𝐴𝐾𝑉(𝐬0 ) + 𝑉𝐾𝑉(𝐬0 ) . (24.6)
2𝑉OK (𝐬0 )

Again, the mean of the EAC values (MEAC) over the nodes of a prediction
grid (evaluation) is used as a minimisation criterion.
Computing time for optimisation of the coordinates of a large sample, say, > 50
points, can become prohibitively long. To reduce computing time, Zhu and Stein
(2006) proposed a two-step approach. In the first step, for a fixed proportion
𝑝 ∈ (0, 1) the locations of (1 − 𝑝) 𝑛 points are optimised for prediction with
given parameters, for instance by minimising MKV. This ‘prediction sample’
is supplemented with 𝑝 𝑛 points, so that the two combined samples of size 𝑛
minimise logdet or MVKV. This is repeated for different values of 𝑝. In the
second step, MEAC is computed for the combined samples of size 𝑛, and the
proportion and the associated sample with minimum MEAC are selected.
A simplification of this two-step approach is to select in the first step a square
grid or a spatial coverage sample (Section 17.2), and to supplement this
sample by a fixed number of points whose coordinates are optimised by spatial
simulated annealing (SSA), using either MAKV or MEAC computed from both
samples (grid sample or spatial coverage sample plus supplemental sample)
as a minimisation criterion. In SSA the grid or spatial coverage sample is
fixed, i.e., the locations are not further optimised. Lark and Marchant (2018)
452 24 Sampling for estimating the semivariogram

recommended as a rule of thumb to add about 10% of the fixed sample as

short distance points.
The following code chunks show how the AKV and EAC can be computed.
First, a spatial coverage sample of 90 points is selected using function stratify
of package spcosa, see Section 17.2. In addition, a simple random sample of
10 points is selected. This sample is the initial supplemental sample, whose
locations are optimised. As before, a square grid of 200 points is used as an
evaluation sample.

library(spcosa)
set.seed(314)
mystrata <- stratify(grdHunterValley, nStrata = 90, equalArea = FALSE, nTry = 10)
mysample_SC <- as(spsample(mystrata), "SpatialPoints")
nsup <- 10
units <- sample(nrow(grdHunterValley), nsup)
mysample_sup0 <- as(grdHunterValley[units, ], "SpatialPoints")
mysample_eval <- spsample(
x = grdHunterValley, n = 200, type = "regular", offset = c(0.5, 0.5))

The next step is to compute the inverse of the Fisher information matrix, given
a preliminary semivariogram model, which is used as the variance-covariance
matrix of the estimated semivariogram parameters. Contrary to Section 24.3
now all sampling locations are used to compute this matrix. The locations of
the spatial coverage sample and the supplemental sample are merged into one
SpatialPoints object.

mysample <- rbind(mysample_sup0, mysample_SC)

To learn how the Fisher information matrix is computed, refer to the code
chunks in Section 24.3. The inverse of this matrix can be computed with
function solve.
In the next code chunk, for each evaluation point the kriging weights (L), the
kriging variance (var), the perturbed kriging weights (pL), and the perturbed
kriging variances (pvar) are computed. In the final lines, the partial derivatives
of the kriging weights (dL) and the kriging variances (dvar) with respect to
the semivariogram parameters are computed. The partial derivatives of the
kriging variances with respect to the semivariogram parameters are needed for
computing VKV, see Equation (24.3), which in turn is needed for computing
criterion EAC, see Equation (24.6).

L <- matrix(nrow = length(mysample_eval), ncol = nobs)

pL <- array(
dim = c(length(mysample_eval), length(mysample), length(thetas)))
24.4 Optimisation of sampling pattern 453

var <- numeric(length = length(mysample_eval))

pvar <- matrix(nrow = length(mysample_eval), ncol = length(thetas))
for (i in seq_len(length(mysample_eval))) {
b <- c(A0[i, ], 1)
l <- solve(B, b)
L[i, ] <- l[1:nobs]
var[i] <- 1 - l[1:nobs] %*% A0[i, ] - l[-(1:nobs)]
for (j in seq_len(length(thetas))) {
pl <- solve(pB[[j]], pb[[j]][i, ])
pL[i, , j] <- pl[1:nobs]
pvar[i, j] <- 1 - pl[1:nobs] %*% pA0[[j]][i, ] - pl[-(1:nobs)]
}
}
dL <- dvar <- list()
for (i in seq_len(length(thetas))) {
dL[[i]] <- (pL[, , i] - L) / (thetas[i] * perturbation)
dvar[[i]] <- (pvar[, i] - var) / (thetas[i] * perturbation)
}

In the next code chunk, the expected variance due to uncertainty about the
semivariogram parameters (Equation (24.5)) is computed.

tausq <- numeric(length = length(mysample_eval))

tausqk <- 0
for (k in seq_len(length(mysample_eval))) {
for (i in seq_len(length(dL))) {
for (j in seq_len(length(dL))) {
tausqijk <- invI[i, j] * t(dL[[i]][k, ]) %*% A %*% dL[[j]][k, ]
tausqk <- tausqk + tausqijk
}
}
tausq[k] <- tausqk
tausqk <- 0
}

The AKVs are computed by adding the kriging variances and the extra variances
due to semivariogram uncertainty (Equation (24.4)). The VKV values and
the EAC values are computed. Both the AKV and the EAC differ among the
evaluation points. As a summary, the mean of the two variables is computed.

augmentedvar <- var + tausq

MAKV0 <- mean(augmentedvar)
VKV <- numeric(length = length(var))
454 24 Sampling for estimating the semivariogram

for (i in seq_len(length(dvar))) {
for (j in seq_len(length(dvar))) {
VKVij <- invI[i, j] * dvar[[i]] * dvar[[j]]
VKV <- VKV + VKVij
}
}
EAC <- augmentedvar + (VKV / (2 * var))
MEAC0 <- mean(EAC)

For the spatial coverage sample of 90 points supplemented by a simple random

sample of 10 points, MAKV equals 0.849 and MEAC equals 0.890.
The sample can be optimised with spsann function optimUSER. Argument
points is a list containing a data frame, or matrix, with the coordinates of the
fixed points (assigned to subargument fixed) and an integer of the number
of supplemental points of which the locations are optimised (assigned to sub-
argument free). As already stressed above, an important difference with Section
24.3 is that the free and the fixed sample are merged and used together both
for estimation of the semivariogram and for prediction. The objective functions
MAKV and MEAC are defined in package sswr. Computing time for minimisation
of MAKV was 28.16 minutes and of MEAC 28.88 minutes.

candi <- spsample(grdHunterValley, type = "regular",

cellsize = c(50, 50), offset = c(0.5, 0.5))
candi <- as.data.frame(candi)
names(candi) <- c("x", "y")
schedule <- scheduleSPSANN(
initial.acceptance = c(0.8, 0.95),
initial.temperature = 0.008, temperature.decrease = 0.8,
chains = 300, chain.length = 10, stopping = 10,
x.min = 0, y.min = 0, cellsize = 50)
fixed <- coordinates(mysample_SC)
names(fixed) <- c("x", "y")
pnts <- list(fixed = fixed, free = nsup)
set.seed(314)
res <- optimUSER(
points = pnts, candi = candi,
fun = MAKV,
esample = mysample_eval,
model = "Exp", thetas = thetas, perturbation = 0.01,
schedule = schedule, track = TRUE)

Figure 24.7 shows for Hunter Valley a spatial coverage sample of 90 points,
supplemented by 10 points optimised by SSA, using MAKV and MEAC as a
minimisation criterion.
24.4 Optimisation of sampling pattern 455

FIGURE 24.7: Optimised sampling pattern of 10 points supplemented to

spatial coverage sample of 90 points, for semivariogram estimation and pre-
diction, using the mean augmented kriging variance (MAKV) and the mean
estimation adjusted criterion (MEAC) as a minimisation criterion. The prior
semivariogram used in optimising the sampling pattern of the supplemental
sample is an exponential semivariogram with a range of 200 m and a ratio of
spatial dependence of 0.5.

The frequency distribution of the shortest distance to the spatial coverage sam-
ple is shown in Figure 24.8. With both criteria there are several supplemental
points at very short distance of a point of the spatial coverage sample. The
remaining points are at large distances of spatial coverage sample points. The
average distance between neighbouring spatial coverage sampling points equals
381 m.
MAKV of the optimised sample equals 0.795 which is 94% of MAKV of the
initial sample. MEAC of the optimised sample equals 0.808 which is 91% of
MEAC of the initial sample. The reduction of these two criteria through the
optimisation is much smaller than for logdet and MVKV in Section 24.3. This
can be explained by the small number of sampling units that is optimised:
only the locations of 10 points are optimised, 90 are fixed. In Section 24.3 all
100 locations were optimised.

Exercises

5. Write an R script to select from Hunter Valley a spatial coverage

sample of 80 points supplemented by 20 points. Use MEAC as a
minimisation criterion, an exponential semivariogram with a distance
parameter of 200 m and a ratio of spatial dependence of 0.8. Compare
456 24 Sampling for estimating the semivariogram

FIGURE 24.8: Frequency distributions of the shortest distance to the spatial

coverage sample of the supplemental sample, optimised with the mean aug-
mented kriging variance (MAKV) and the mean estimation adjusted criterion
(MEAC).

the minimised MEAC with MEAC reported above, obtained by

supplementing a spatial coverage sample of 90 points by 10 points.

24.5 A practical solution

Based on the optimised samples shown above, a straightforward, simple sam-
pling design for estimation of the model parameters and for prediction is a
spatial coverage sample supplemented with randomly selected points between
the points of the spatial coverage sample at some chosen fixed distances. Figure
24.9 shows an example. A simple random subsample without replacement of
size 10 is selected from the 90 points of the spatial coverage sample. These
points are used as starting points to select a point at a distance of 20 m in a
random direction.

h <- 20
m <- 10
set.seed(314)
units <- sample(nrow(mysample_SC), m, replace = FALSE)
mySCsubsample <- mysample_SC[units, ]
dxy <- matrix(nrow = m, ncol = 2)
24.5 A practical solution 457

angle <- runif(n = m, min = 0, max = 2 * pi)

dxy[, 1] <- h * sin(angle); dxy[, 2] <- h * cos(angle)
mysupsample <- mySCsubsample + dxy

FIGURE 24.9: Spatial coverage sample of 90 points supplemented by 10

points at short distance (20 m) from randomly selected spatial coverage points.

MAKV of this sample equals 0.808, and MEAC equals 0.816. For MAKV, 25%
of the maximal reduction is realised by this practical solution; for MEAC, this
is 10%.
25
Sampling for validation of maps

In the previous chapters of Part II, various methods are described for selecting
sampling units with the aim to map the study variable. Once the map has been
made, we would like to know how good it is. It should come as no surprise
that the value of the study variable at a randomly selected location as shown
on the map differs from the value at that location in reality. This difference is
a prediction error. The question is how large this error is on average, and how
variable it is. This chapter describes and illustrates with a real-world case study
how to select sampling units at which we will confront the predictions with
the true values, and how to estimate map quality indices from the prediction
errors of these sampling units.
If the map has been made with a statistical model, then the predictors are
typically model-unbiased and the variance of the prediction errors can be
computed from the model. Think, for instance, of kriging which also yields
a map of the kriging variance. In Chapters 22 and 23 I showed how this
kriging variance can be used to optimise the grid spacing (sample size) and
the sampling pattern for mapping, respectively. So, if we have a map of these
variances, why do we still need to collect new data for estimating the map
quality?
The problem is that the kriging variances rely on the validity of the assumptions
made in modelling the spatial variation of the study variable. Do we assume
a constant mean, or a mean that is a linear combination of some covariates?
In the latter case, which covariates are assumed to be related to the study
variable? Or should we model the mean with a non-linear function as in a
random forest model? How certain are we about the semivariogram model
type (spherical, exponential, etc.), and how good are our estimates of the
semivariogram parameters? If one or more of the modelling assumptions are
violated, the variances of the prediction errors as computed with the model
may become biased. For this reason, the quality of the map is preferably
determined through independent validation, i.e., by comparing predictions
with observations not used in mapping, followed by design-based estimation of
the map quality indices. This process is often referred to as validation, perhaps
better statistical validation, a subset of the more comprehensive term map
quality evaluation, which includes the concept of fitness-for-use.

459
460 25 Sampling for validation of maps

Statistical validation of maps is often done through data splitting or cross-

validation. In data splitting the data are split into two subsets, one for cali-
brating the model and mapping and one for validation. In cross-validation the
data set is split into a number of disjoint subsets of equal size. Each subset
is used one-by-one for calibration and prediction. The remaining subsets are
used for validation. Leave-one-out cross-validation (LOOCV) is a special case
of this, in which each sampling unit is left out one-by-one, and all other units
are used for calibration and prediction of the study variable of the unit that
is left out. The problem with data splitting and cross-validation is that the
data used for mapping typically are from non-probability samples. This makes
design-based estimation of the map quality indices unfeasible (Brus et al.,
2011). Designing a sampling scheme starts with a comprehensive description
of the aim of the sampling project (de Gruijter et al., 2006). Mapping and
validation are different aims which ask for different sampling approaches. For
validation probability sampling is the best option because then a statistical
model of the spatial variation of the prediction errors is not needed. Map
quality indices, defined as population parameters, can be estimated model-free,
by design-based inference (see also Section 1.2).

In statistical learning using large data sets a common approach is to randomly

partition the data set in three subsets: a training subset, a validation subset,
and a test subset (Hastie et al. (2009), chapter 7). The training subset is used
for fitting the models, the validation subset is used to estimate prediction
error for model selection and hyperparameter tuning, while the test subset
is used for assessing the accuracy of the final model. The term validation as
used in this chapter is therefore the equivalent of testing as used in statistical
learning.

All probability sampling designs described in Part I are in principle appropriate

for validation. Stehman (1999) evaluated five basic probability sampling designs
and concluded that in general stratified random sampling is a good choice. For
validation of categorical maps, natural strata are the map units, i.e., the groups
of polygons or grid cells assigned to each class. Systematic random sampling is
less suitable, as no unbiased estimator of the sampling variance of the estimator
of a population mean exists for this design (see Chapter 5). For validation
of maps of extensive areas, think of whole continents, travel time between
sampling locations can become substantial. In this case, sampling designs
that lead to spatial clustering of validation locations can become efficient, for
instance cluster random sampling (Chapter 6) or two-stage cluster random
sampling (Chapter 7).
25.1 Map quality indices 461

25.1 Map quality indices

In validation we want to assess the accuracy of the map as a whole. We are not
interested in the accuracy at a sample of population units only. For instance,
we would like to know the population mean of the prediction error, i.e., the
average of the errors over all population units, and not merely the average
prediction error at a sample of units. Map quality indices are therefore defined
as population parameters. We cannot afford to determine the prediction error
for each unit of the mapping area to calculate the population means. If we could
do that, there would be no need for a mapping model. Therefore, we have to
take a sample of units at which the predictions of the study variable are con-
fronted with the observations. This sample is then used to estimate population
parameters of the prediction error and our uncertainty about these population
parameters, as quantified, for instance, by their standard errors or confidence
interval.
For quantitative maps, i.e., maps depicting a quantitative study variable,
popular map quality indices are (i) the population mean error (ME); (ii) the
population mean absolute error (MAE); and (iii) the population mean squared
error (MSE), defined as

1 u�
𝑀𝐸 = ∑ (𝑧 ̂ − 𝑧u� ) (25.1)
𝑁 u�=1 u�
1 u�
𝑀𝐴𝐸 = ∑ (|𝑧 ̂ − 𝑧u� |) (25.2)
𝑁 u�=1 u�
1 u�
𝑀𝑆𝐸 = ∑ (𝑧 ̂ − 𝑧u� )2 , (25.3)
𝑁 u�=1 u�

with 𝑁 the total number of units (e.g., raster cells) in the population, 𝑧u�̂
the predicted value for unit 𝑘, 𝑧u� the true value of that unit, and | ⋅ | the
absolute value operator. For infinite populations, the sum must be replaced by
an integral over all locations in the mapped area and divided by the size of the
area. The ME quantifies the systematic error and ideally equals 0. It can be
positive (in case of overprediction) and negative (in case of underprediction).
Positive and negative errors cancel out and, as a consequence, the ME does
not quantify the magnitude of the prediction errors. The MAE and MSE do
quantify the magnitude of the errors, they are non-negative. Often, the square
root of MSE is taken, denoted by RMSE, which is in the same units as the
study variable and is therefore more intelligible. The RMSE is strongly affected
by outliers, i.e., large prediction errors, due to the squaring of the errors, and
for this reason I recommend estimating both MAE and RMSE.
462 25 Sampling for validation of maps

Two other important map quality indices are the population coefficient of
determination (𝑅2 ) and the Nash-Sutcliffe model efficiency coefficient (MEC).
𝑅2 is defined as the square of the Pearson correlation coefficient 𝑟 of the study
variable and the predictions of the study variable, given by

∑u� ̄ 𝑧u�̂ − 𝑧)̄̂

(𝑧 − 𝑧)(
u�=1 u� 𝑆2 (𝑧, 𝑧)̂
𝑟= = , (25.4)
√∑u� (𝑧u� − 𝑧)̄ 2 √(𝑧u�̂ − 𝑧)̄̂ 2 𝑆(𝑧)𝑆(𝑧)̂
u�=1

with 𝑧 ̄ the population mean of the study variable, 𝑧 ̄̂ the population mean of the
predictions, 𝑆2 (𝑧, 𝑧)̂ the population covariance of the study variable and the
predictions of 𝑧, 𝑆(𝑧) the population standard deviation of the study variable,
and 𝑆(𝑧)̂ the population standard deviation of the predictions. Note that 𝑅2
is unaffected by bias and therefore should not be used in isolation, but should
always be accompanied by ME.
MEC is defined as (Janssen and Heuberger, 1995)

∑u� (𝑧 ̂ − 𝑧u� )2
u�=1 u� 𝑀𝑆𝐸
𝑀𝐸𝐶 = 1 − =1− , (25.5)
∑u� (𝑧
u�=1 u�
− 𝑧)̄ 2 𝑆2 (𝑧)

with 𝑆2 (𝑧) the population variance of the study variable. MEC quantifies the
improvement made by the model over using the mean of the observations as a
predictor. An MEC value of 1 indicates a perfect match between the observed
and the predicted values of the study variable, whereas a value of 0 indicates
that the mean of the observations is as good a predictor as the model. A
negative value occurs when the mean of the observations is a better predictor
than the model, i.e., when the residual variance is larger than the variance of
the measurements.
For categorical maps, a commonly used map quality index is the overall purity,
which is defined as the proportion of units that is correctly classified (mapped):

1 u�
𝑃= ∑𝑦 , (25.6)
𝑁 u�=1 u�

with 𝑦u� an indicator for unit 𝑘, having value 1 if the predicted class equals the
true class, and 0 otherwise:

1 if 𝑐u�̂ = 𝑐u�
𝑦u� = { (25.7)
0 otherwise ,

with 𝑐u� and 𝑐u�̂ the true and the predicted class of unit 𝑘, respectively. For
infinite populations the purity is the fraction of the area that is correctly
classified (mapped).
25.1 Map quality indices 463

The population ME, MSE, 𝑅2 , MEC, and purity can also be defined for
subpopulations. For categorical maps, natural subpopulations are the classes
depicted in the map, the map units. In that case, for infinite populations the
purity of map unit 𝑢 is defined as the fraction of the area of map unit 𝑢 that
is correctly mapped as 𝑢.
A different subpopulation is the part of the population that is in reality class
𝑢 (but possibly not mapped as 𝑢). We are interested in the fraction of the area
covered by this subpopulation that is correctly mapped as 𝑢. This is referred
to as the class representation of class 𝑢, for which I use hereafter the symbol
𝑅u� .

25.1.1 Estimation of map quality indices

The map quality indices are defined as population or subpopulation means.

To estimate these (sub)population means, a design-based sampling approach
is the most appropriate. Sampling units are selected by probability sampling,
and the map quality indices are estimated by design-based inference. For
instance, the ME of a finite population can be estimated by the 𝜋 estimator
(see Equation (2.4)):

̂ = 1 ∑ 1 𝑒u� ,
𝑀𝐸 (25.8)
𝑁 u�∈u� 𝜋u�

with 𝑒u� = 𝑧u�̂ − 𝑧u� the prediction error for unit 𝑘. By taking the absolute
value of the prediction errors 𝑒u� in Equation (25.8) or by squaring them, the 𝜋
estimators for the MAE and MSE are obtained, respectively. By replacing 𝑒u�
by the indicator 𝑦u� of Equation (25.7), the 𝜋 estimator for the overall purity
is obtained.
With simple random sampling, the square of the sample correlation coefficient,
i.e., the correlation of the study variable and the predictions of the study
variable in the sample, is an unbiased estimator of 𝑅2 . See Särndal et al. (1992)
(p. 486 − 491) for how to estimate 𝑅2 for other sampling designs.
The population MEC can be estimated by

̂
̂ = 1 − 𝑀𝑆𝐸 .
𝑀𝐸𝐶 (25.9)
̂
𝑆2 (𝑧)

For simple random sampling the sample variance, i.e., the variance of the
observations of 𝑧 in the sample, is an unbiased estimator of the population
variance 𝑆2 (𝑧). For other sampling designs, this population variance can be
estimated by Equation (4.9).
464 25 Sampling for validation of maps

Estimation of the class representations is slightly more difficult, because the

sizes of the classes (number of raster cells or area where in reality class 𝑢 is
present) are unknown and must therefore also be estimated. This leads to the
ratio estimator:

u�u�
∑u�∈u� u�u�
𝑅̂ u� = u�u� , (25.10)
∑u�∈u� u�
u�

where 𝑦u� denotes an indicator defined as

1 if 𝑐u�̂ = 𝑐u� = 𝑢
𝑦u� = { (25.11)
0 otherwise ,

and 𝑥u� denotes an indicator defined as

1 if 𝑐u� = 𝑢
𝑥u� = { (25.12)
0 otherwise .

This estimator is also recommended for estimating other map quality indices
from a sample with a sample size that is not fixed but varies among samples
selected with the sampling design. This is the case, for instance, when estimating
the mean (absolute or squared) error or the purity of a given map unit from a
simple random sample. The number of selected sampling units within the map
unit is uncontrolled and varies among the simple random samples. In this case,
we can estimate the mean error or the purity of a map unit 𝑢 by dividing the
estimated population total by either the known size (number of raster cells,
area) of map unit 𝑢 or by the estimated size. Interestingly, in general using the
estimated size in the denominator, instead of the known size, yields a more
precise estimator (Särndal et al., 1992). See also Section 14.1.

25.2 Real-world case study

As an illustration, two soil maps of three northern counties of Xuancheng
(China), both depicting soil organic matter (SOM) concentration (g kg-1 ) in
the topsoil, are evaluated. In Section 13.2 the data of three samples, including
the stratified random sample, were merged to estimate the parameters of a
spatial model for the natural log of the SOM concentration. Here, only the
data of the two non-random samples, the grid sample and the iPSM sample,
are used to map the SOM concentration. The stratified simple random sample
is used for validation.
25.2 Real-world case study 465

Two methods are used in mapping, kriging with an external drift (KED) and
random forest prediction (RF). For mapping with RF, seven covariates are
used: planar curvature, profile curvature, slope, temperature, precipitation,
topographic wetness index, and elevation. For mapping with KED only the
two most important covariates in the RF model are used: precipitation and
elevation.
The two maps that are to be validated are shown in Figure 25.1. Note that
non-soil areas (built-up, water, roads) are not predicted. The maps are quite
similar. The most striking difference between the maps is the smaller range of
the RF predictions: they range from 9.8 to 61.5, whereas the KED predictions
range from 5.3 to 90.5.

FIGURE 25.1: Map of the SOM concentration (g kg-1 ) in the topsoil of

Xuancheng, obtained by kriging with an external drift (KED) and random
forest (RF).

The two maps are evaluated by statistical validation with a stratified simple
random sample of 62 units (points). The strata are the eight units of a geological
map (Figure 25.2).

25.2.1 Estimation of the population mean error and mean

squared error

To estimate the population MSE of the two maps, first the squared prediction
errors are computed. The name of the measured study variable at the validation
sample in data.frame sample_test is SOM_A_hori. Four new variables are added to
sample_test using function mutate, by computing the prediction errors for KED
and RF and squaring these errors.

sample_test <- read.csv(file = "results/STSI_Xuancheng_SOMpred.csv")

sample_test <- sample_test %>%
466 25 Sampling for validation of maps

FIGURE 25.2: Stratified simple random sample for validation of the two
maps of the SOM concentration in Xuancheng.

mutate(
eKED = SOM_A_hori - SOM_KED,
eRF = SOM_A_hori - SOM_RF,
e2KED = (SOM_A_hori - SOM_KED)ˆ2,
e2RF = (SOM_A_hori - SOM_RF)ˆ2)

These four new variables now are our study variables of which we would like
to estimate the population means. The population means can be estimated
as explained in Chapter 4. First, the stratum sizes and stratum weights are
computed, i.e., the number and relative number of raster cells per stratum
(Figure 25.2).

rmap <- rast(x = system.file("extdata/Geo_Xuancheng.tif", package = "sswr"))

strata_Xuancheng <- as.data.frame(rmap, xy = TRUE, na.rm = TRUE) %>%
rename(stratum = Geo_Xuancheng) %>%
filter(stratum != 99) %>%
group_by(stratum) %>%
summarise(N_h = n()) %>%
mutate(w_h = N_h / sum(N_h))

Next, the stratum means of the prediction errors, obtained with KED and RF,
are estimated by the sample means, and the population mean of the errors are
estimated by the weighted mean of the estimated stratum means.
25.2 Real-world case study 467

me <- sample_test %>%

group_by(stratum) %>%
summarise(
meKED_h = mean(eKED),
meRF_h = mean(eRF)) %>%
left_join(strata_Xuancheng, by = "stratum") %>%
summarise(
meKED = sum(w_h * meKED_h),
meRF = sum(w_h * meRF_h))

This is repeated for the squared prediction errors.

mse <- sample_test %>%

group_by(stratum) %>%
summarise(
mseKED_h = mean(e2KED),
mseRF_h = mean(e2RF)) %>%
left_join(strata_Xuancheng, by = "stratum") %>%
summarise(
mseKED = sum(w_h * mseKED_h),
mseRF = sum(w_h * mseRF_h))

The estimated MSE of the KED map equals 89.3 (g kg-1 )2 , that of the RF
map 93.8 (g kg-1 )2 .

Exercises

1. Are you certain that the population MSE of the KED map is smaller
than the population MSE of the RF map?

25.2.2 Estimation of the standard error of the estimator of

the population mean error and mean squared error

We are uncertain about both population MSEs, as we measured the squared

errors at 62 sampling points only. So, we would like to know how uncertain we
are. This uncertainty is quantified by the standard error of the estimator of
the population MSE. A problem is that in the second stratum we have only
one sampling point. So, for this stratum we cannot compute the variance of
the squared errors. To compute the variance, we need at least two sampling
points.
468 25 Sampling for validation of maps

n_strata <- sample_test %>%

group_by(stratum) %>%
summarise(n = n())
n_strata

# A tibble: 8 x 2
stratum n
<int> <int>
1 1 5
2 2 1
3 3 8
4 4 10
5 5 2
6 6 23
7 7 9
8 8 4

A solution is to merge stratum 2 with stratum 1, which is a similar geological

map unit (we know this from the domain expert). This is referred to as
collapsing the strata. An identifier for the collapsed strata is added to n_strata.
This table is subsequently used to add the collapsed stratum identifiers to
sample_test and strata_Xuancheng.

n_strata <- n_strata %>%

mutate(stratum_clp = c(1, 1:7))
sample_test <- sample_test %>%
left_join(n_strata, by = "stratum")
strata_Xuancheng <- strata_Xuancheng %>%
left_join(n_strata, by = "stratum")

The collapsed strata can be used to estimate the standard errors of the
estimators of the population MSEs. As a first step, the weights and the sample
sizes of the collapsed strata are computed.

strata_clp_Xuancheng <- strata_Xuancheng %>%

group_by(stratum_clp) %>%
summarise(N_hc = sum(N_h)) %>%
mutate(w_hc = N_hc / sum(N_hc)) %>%
left_join(
sample_test %>%
group_by(stratum_clp) %>%
summarise(n_hc = n()),
by = "stratum_clp")
25.2 Real-world case study 469

TABLE 25.1: Estimated population mean error (ME) and population mean
squared error (MSE) of KED and RF map, and their standard errors.

KED seKED RF seRF

ME 0.83 1.2 0.4 1.29
MSE 89.30 25.5 93.8 25.80

The sampling variance of the estimator of the mean of the (squared) prediction
error can be estimated by Equation (4.4). The estimated ME and MSE and
their estimated standard errors are shown in Table 25.1.

se <- sample_test %>%

group_by(stratum_clp) %>%
summarise(
s2e_KED_hc = var(eKED),
s2e2_KED_hc = var(e2KED),
s2e_RF_hc = var(eRF),
s2e2_RF_hc = var(e2RF)) %>%
left_join(strata_clp_Xuancheng, by = "stratum_clp") %>%
summarise(
se_me_KED = sqrt(sum(w_hcˆ2 * s2e_KED_hc / n_hc)),
se_mse_KED = sqrt(sum(w_hcˆ2 * s2e2_KED_hc / n_hc)),
se_me_RF = sqrt(sum(w_hcˆ2 * s2e_RF_hc / n_hc)),
se_mse_RF = sqrt(sum(w_hcˆ2 * s2e2_RF_hc / n_hc)))

Exercises

2. Do you think there is a systematic error in the KED and the RF

predictions?

3. Do you think the difference between the two estimated population

MSEs is statistically significant?

25.2.3 Estimation of model efficiency coefficient

To estimate the MEC, we must first estimate the population variance of the
study variable from the stratified simple random sample (the denominator in
Equation (25.9)). First, the sizes and the sample sizes of the collapsed strata
must be added to sample_test. Then the population variance is estimated with
function s2 of package surveyplanning (Subsection 4.1.2).
470 25 Sampling for validation of maps

library(surveyplanning)
s2z <- sample_test %>%
left_join(strata_clp_Xuancheng, by = "stratum_clp") %>%
summarise(s2z = s2(SOM_A_hori, w = N_hc / n_hc)) %>%
flatten_dbl

Now the MECs for KED and RF can be estimated.

mec <- 1 - mse / s2z

The estimated MEC for KED equals 0.016 and for RF -0.034, showing that
the two models used in mapping are no better than the estimated mean SOM
concentration used as a predictor. This is quite a disappointing result.

25.2.4 Statistical testing of hypothesis about population ME

and MSE

The hypothesis that the population ME equals 0 can be tested by a one-

sample t-test. The alternative hypothesis is that ME is unequal to 0 (two-
sided alternative). The number of degrees of freedom of the t distribution is
approximated by the total sample size minus the number of strata (Section
4.2). Note that we have a two-sided alternative hypothesis, so we must compute
a two-sided p-value.

t_KED <- me$meKED / se$se_me_KED

df <- nrow(sample_test) - length(unique(sample_test$stratum_clp))
p_KED <- 2 * pt(t_KED, df = df, lower.tail = t_KED < 0)

The outcomes of the test statistics are 0.690 and 0.309 for KED and RF,
respectively, with p-values 0.493 and 0.759. So, we clearly have not enough
evidence for systematic errors, neither with KED nor with RF mapping.
Now we test whether the two population MSEs differ significantly. This can
be done by a paired t-test. The first step in a paired t-test is to compute
pairwise differences of squared prediction errors, and then we can proceed as
in a one-sample t-test.

m_de2 <- sample_test %>%

mutate(de2 = e2KED - e2RF) %>%
group_by(stratum) %>%
summarise(m_de2_h = mean(de2)) %>%
left_join(strata_Xuancheng, by = "stratum") %>%
summarise(m_de2 = sum(w_h * m_de2_h)) %>%
25.2 Real-world case study 471

flatten_dbl

se_m_de2 <- sample_test %>%

mutate(de2 = e2KED - e2RF) %>%
group_by(stratum_clp) %>%
summarise(s2_de2_hc = var(de2)) %>%
left_join(strata_clp_Xuancheng, by = "stratum_clp") %>%
summarise(se_m_de2 = sqrt(sum(w_hcˆ2 * s2_de2_hc / n_hc))) %>%
flatten_dbl

t <- m_de2 / se_m_de2

p <- 2 * pt(t, df = df, lower.tail = t < 0)

The outcome of the test statistic is -0.438, with a p-value of 0.663, so we clearly
do not have enough evidence that the population MSEs obtained with the two
mapping methods are different.
26
Design-based, model-based, and
model-assisted approach for sampling and
inference

Section 1.2 already mentioned the design-based and the model-based approach
for sampling and statistical inference. In this chapter, the fundamental differ-
ences between these two approaches are explained in more detail. Several
misconceptions about the design-based approach for sampling and statistical
inference, based on classical sampling theory, seem to be quite persistent.
These misconceptions are the result of confusion about basic statistical con-
cepts such as independence, expectation, and bias and variance of estimators
or predictors. These concepts have a different meaning in the design-based
and the model-based approach. Besides, a population mean is still often con-
fused with a model-mean, and a population variance with a model-variance,
leading to invalid formulas for the sampling variance of an estimator of the
population mean. The fundamental differences between these two approaches
are illustrated with simulations, so that hopefully a better understanding
of this subject is obtained. Besides, the difference between model-dependent
inference (as used in the model-based approach) and model-assisted inference
is explained. This chapter has been published as part of a journal paper, see
Brus (2021).

26.1 Two sources of randomness

In my classes about spatial sampling, I ask the participants the following
question. Suppose we have measurements of a soil property, for instance soil
organic carbon content, at two locations separated by 20 cm. Do you think
these two measurements are correlated? I ask them to vote for one of three
answers:

1. yes, they are (>80% confident);

2. no, they are not (>80% confident); or
3. I do not know.

473
474 26 Design-based approach for sampling and inference

Most students vote for answer 1, the other students vote for answer 3, nearly
no one votes for answer 2. Then I explain that you cannot say which answer
is correct, simply because for correlation we need two series of data, not just
two numbers. The question then is how to generate two series of data. We
need some random process for this. This random process differs between the
design-based and the model-based approach.
In the design-based approach, the random process is the random selection of
sampling units, whereas in the model-based approach randomness is introduced
via the statistical model of the spatial variation (Table 1.1). So, the design-
based approach requires probability sampling, i.e., random sampling, using
a random number generator, in such way that all population units have a
positive probability of being included in the sample and that these inclusion
probabilities are known for at least the selected population units (Särndal
et al., 1992). A probability sampling design can be used to generate an infinite
number of samples in theory, although in practical applications only one is
selected.
The spatial variation model used in the model-based approach contains two
terms, one for the mean (deterministic part) and one for the error with a
specified probability distribution. For instance, Equation (21.2) in Chapter 21
describes the model used in ordinary kriging. This model can be used to simulate
an infinite number of spatial populations. All these populations together are
referred to as a superpopulation (Särndal et al. (1992), Lohr (1999)). Depending
on the model of spatial variation, the simulated populations may show spatial
structure because the mean is a function of covariates, as in kriging with
an external drift, and/or because the errors are spatially autocorrelated. A
superpopulation is a construct, the populations do not exist in the real world.
The populations are similar, but not identical. For instance, the mean differs
among the populations. The expectation of the population mean, i.e., the
average over all possible simulated populations, equals the superpopulation
mean, commonly referred to as the model-mean, parameter 𝜇 in Equation
(21.2). The variance also differs among the populations. Contrary to the mean,
the average of the population variance over all populations generally is not
equal to the model-variance, parameter 𝜎2 in Equation (21.2), but smaller.
I will come back to this later. The differences between the simulated spatial
populations illustrate our uncertainty about the spatial variation of the study
variable in the population that is sampled or will be sampled.
In the design-based approach, only one population is considered, the one sam-
pled, but the statistical inference is based on all samples that can be generated
by a probability sampling. The top row of Figure 26.1 shows five simple random
samples of size ten. The population is the same in all plots. Proponents of
the design-based approach do not like to consider other populations than the
one sampled. Their challenge is to characterise this one population from a
probability sample.
26.1 Two sources of randomness 475

On the contrary, in the model-based approach only one sample is considered,

but the statistical inference is based on all populations that can be generated
with the spatial variation model. Proponents of the model-based approach do
not like to consider other samples than the one selected. Their challenge is to
get the most out of the sample that is selected. The bottom row of Figure 26.1
shows a spatial coverage sample, superimposed on five different populations
simulated with an ordinary kriging model, using a spherical semivariogram
with a nugget of 0.1, partial sill of 0.6, and a range of 75 m. Note that in
the model-based approach there is no need to select a probability sample (see
Table 1.1); there are no requirements on how the units are selected.

FIGURE 26.1: Random process considered in the design-based (top row) and
the model-based approach (bottom row). The design-based approach considers
only the sampled population, but all samples that can be generated by the
sampling design. The model-based approach considers only the selected sample,
but all populations that can be generated by the model.

As stressed by de Gruijter and ter Braak (1990) and Brus and de Gruijter (1997),
both approaches have their strengths and weaknesses. Broadly speaking, the
design-based approach is the most appropriate if interest is in the population
mean (total, proportion) or the population means (totals, proportions) of a
restricted number of subpopulations (subareas). The model-based approach
is the most appropriate if our aim is to map the study variable. Further, the
strength of the design-based approach is the strict validity of the estimates.
Validity means that an objective assessment of the uncertainty of the estimator
is warranted and that the coverage of confidence intervals is (almost) correct,
provided that the sample is large enough to assume an approximately normal
distribution of the estimator and design-unbiasedness of the variance estimator
476 26 Design-based approach for sampling and inference

(Särndal et al., 1992). The strength of the model-based approach is efficiency,

i.e., more precise estimates of the (sub)population mean given the sample
size, provided that a reasonably good model is used. So, if validity is more
important than efficiency, the design-based approach is the best choice; in the
reverse case, the model-based approach is preferable. For further reading, I
recommend Cassel et al. (1977) and Hansen et al. (1983).

26.2 Identically and independently distributed

In a review paper on spatial sampling by Wang et al. (2012), there is a
section with the caption ‘Sampling of i.i.d. populations’. Here, i.i.d. stands
for “identically and independently distributed”. In this section of Wang et al.
(2012) we can read: “In SRS (simple random sampling) it is assumed that
the population is independent and identically distributed”. This is one of the
old misconceptions revitalised by this review paper. I will make clear that in
statistics i.i.d. is not a characteristic of populations, so the concept of i.i.d.
populations does not make sense. The same misconception can be found in
Plant (2012): “There is considerable literature on sample size estimation, much
of which is discussed by Cochran (1977, chapter 4). This literature, however,
is valid for samples of independent data but may not retain its validity for
spatial data”. Also according to Wang et al. (2010), the classical formula
for the variance of the estimator of the mean with simple random sampling,
𝑉 = 𝜎2 /𝑛, only holds when data are independent. They say: “However in
the case of spatial data, although members of the sample are independent by
construction, data values that are near to one another in space, are unlikely to
be independent because of a fundamental property of attributes in space, which
is that they show spatial structure or continuity (spatial autocorrelation)”.
According to Wang et al. (2010), the variance should be approximated by

𝜎2 − Cov(𝑧u� , 𝑧u� )
𝑉(𝑧)̄̂ = , (26.1)
𝑛

with 𝑉(𝑧)̄̂ the variance of the estimator of the regional mean (mean of spatial
population), 𝜎2 the population variance, 𝑛 the sample size, and Cov(𝑧u� , 𝑧u� ) the
average autocovariance between all pairs of individuals (𝑖, 𝑗) in the population
(sampled and unsampled). So, according to this formula, ignoring the mean
covariance within the population leads to an overestimation of the variance of
the estimator of the mean. In Section 26.4 I will make clear that this formula
is incorrect and that the classical formula is still valid, also for populations
showing spatial structure or continuity.
26.2 Identically and independently distributed 477

Remarkably, in other publications we can read that the classical formula for the
variance of the estimator of the population mean with simple random sampling
underestimates the true variance for populations showing spatial structure, see
for instance Griffith (2005) and Plant (2012). The reasoning is that due to
the spatial structure, there is less information in the sample data about the
population mean. In Section 26.4 I explain that this is also a misconception.
Do not get confused by these publications and stick to the classical formulas
which you can find in standard textbooks on sampling theory, such as Cochran
(1977) and Lohr (1999), as well as in Chapter 3 of this book.
The concept of independence of random variables is illustrated with a simulation.
The top row of Figure 26.2 shows five simple random samples of size two. The
two points are repeatedly selected from the same population (showing clear
spatial structure), so this top row represents the design-based approach. The
bottom row shows two points, not selected randomly and independently, but at
a fixed distance of 10 m. These two points are placed on different populations
generated by the model described above, so the bottom row represents the
model-based approach.

FIGURE 26.2: Illustration of independence in design-based and model-based

approach. The top row shows five samples of two points selected randomly and
independently from each other from one population (design-based approach).
The bottom row shows two points not selected randomly, at a distance of 10
m from each other, from five model realisations (model-based approach).

The values measured at the two points are plotted against each other in a
scatter plot, not for just five simple random samples or five populations, but
for 1,000 samples and 1,000 populations (Figure 26.3). As we can see there is
478 26 Design-based approach for sampling and inference

FIGURE 26.3: Scatter plots of the values of a study variable 𝑧 at two

randomly and independently selected points, 1,000 times selected from one
population (design-based approach), and at two fixed points with a separation
distance of 10 m, selected non-randomly from 1,000 model realisations (model-
based approach).

no correlation between the two variables generated by the repeated random

selection of the two points (design-based), whereas the two variables generated
by the repeated simulation of populations (model-based) are correlated.
Instead of two points, we may select two series of probability samples inde-
pendently from each other, for instance two series of simple random samples
(SI) of size 10, or two series of systematic random samples with random origin
(SY) with an average size of 10, see Figure 26.4.
Again, if we plot the sample means of pairs of simple random samples and pairs
of systematic random samples against each other, we see that the two averages
are not correlated (Figure 26.5). Note that the variation of the averages of the
systematic random samples is considerably smaller than that of the simple
random samples. The sampled population shows spatial structure. By spreading
the sampling units over the spatial population, the precision of the estimated
population mean is increased, see Chapter 5.
This sampling experiment shows that independence is not a characteristic
of a population, as stated by Wang et al. (2012), but of random variables
generated by a random process (in the experiment the values at points or the
sample means). As the random process differs between the design-based and
the model-based approach, independence has a different meaning in these two
approaches. For this reason, it is imperative to be more specific when using
26.2 Identically and independently distributed 479

FIGURE 26.4: Two series (a and b) of simple random samples of ten points
(top) and two series (a and b) of systematic random samples of, on average,
ten points (bottom). The samples of series a and b are selected independently
from each other.
480 26 Design-based approach for sampling and inference

FIGURE 26.5: Scatter plots of averages of 1,000 pairs of simple random

samples of ten points (SI) and of averages of 1,000 pairs of systematic random
samples of ten points on average (SY).

the term independence, by saying that data are design-independent or that

you assume that the data are model-independent.

26.3 Bias and variance

Bias and variance are commonly used statistics to quantify the quality of an
estimator. Bias quantifies the systematic error, variance the random error of
the estimator. Both are defined as expectations. But are these expectations
over realisations of a probability sampling design (samples) or realisations
of a statistical model (populations)? Like independence, it is important to
distinguish design-bias from model-bias and design-variance (commonly referred
to as sampling variance) from model-variance.
The concept of model-unbiasedness deserves more attention. Figure 26.6 shows
a preferential sample from a population simulated by sequential Gaussian
simulation with a constant mean of 10 and an exponential semivariogram
without nugget, a sill of 5, and a distance parameter of 20. The points are
selected by sampling with draw-by-draw selection probabilities proportional to
size (pps sampling, Chapter 8), using the square of the simulated values as
a size variable. We may have a similar sample that is collected for delineating
soil contamination or detecting hot spots of soil bacteria, etc. Many samples
26.3 Bias and variance 481

are selected at locations with a large value, few points at locations with
a small value. The sample data are used in ordinary kriging (Figure 26.6).
The prediction errors are computed by subtracting the kriged map from the
simulated population.

FIGURE 26.6: Preferential sample (size of open dots is proportional to value

of study variable) from a simulated field (z) and map of ordinary kriging
predictions (zpred).

error <- preds$zpred - sim$z

Figure 26.7 shows a histogram of the prediction errors. The population mean
error equals 0.482, not 0. You may have expected a positive systematic error
because of the overrepresentation of locations with large values, but on the other
hand, kriging predictions are best linear unbiased predictions (BLUP), so from
that point of view, this systematic error might be unexpected. BLUP means that
at individual locations the ordinary kriging predictions are unbiased. However,
apparently this does not guarantee that the average of the prediction errors,
averaged over all population units, equals 0. The reason is that unbiasedness
is defined here over all realisations (populations) of the statistical model of
spatial variation. So, the U in BLUP stands for model-unbiasedness. For other
model realisations, sampled at the same points, we may have much smaller
values, leading to a negative mean error of that population. On average, over
all populations, the error at any point will be 0 and consequently also the
average over all populations of the mean error.
This experiment shows that model-unbiasedness does not protect us against
selection bias, i.e., bias due to preferential sampling.
482 26 Design-based approach for sampling and inference

FIGURE 26.7: Frequency distribution of the errors of ordinary kriging

predictions from a preferential sample.

26.4 Effective sample size

Another persistent misconception is that when estimating the variance of the
estimator of the mean of a spatial population or the correlation of two variables
of a population, we must account for autocorrelation of the sample data. This
misconception occurs, for instance, in Griffith (2005) and in various sections
(for instance, sections 3.5, 10.1, and 11.2) of Plant (2012). The reasoning is that,
due to the spatial autocorrelation in the sample data, there is less information
in the data about the parameter of interest, and so the effective sample size is
smaller than the actual sample size. An early example of this misconception is
Barnes’ publication on the required sample size for estimating nonparametric
tolerance intervals (Barnes, 1988). de Gruijter and ter Braak (1992) showed
that a basic probability sampling design like simple random sampling requires
fewer sampling points than the model-based sampling design proposed by
Barnes.
26.4 Effective sample size 483

The misconception is caused by confusing population parameters with model

parameters. Recall that the population mean and the model-mean are not the
same; the model-mean 𝜇 of Equation (21.2) is the expectation of the population
means over all populations that can be simulated with the model. The same
holds for the variance of a variable as well as for the covariance and the Pearson
correlation coefficient of two variables. All these parameters can be defined as
parameters of a finite or infinite population or of random variables generated
by a superpopulation model. Using an effective sample size to quantify the
variance of an estimator is perfectly correct for model parameters, but not so
for population parameters. For instance, when the correlation coefficient is
defined as a population parameter and sampling units are selected by simple
random sampling, there is no need to apply the method proposed by Clifford
et al. (1989) to correct the p-value in a significance test for the presence of
spatial autocorrelation.
I elaborate on this for the mean as the parameter of interest. Suppose a sample
is selected in some way (need not be random) and the sample mean is used as
an estimator of the model-mean. Note that for a model with a constant mean
as in Equation (21.2), the sample mean is a model-unbiased estimator of the
model-mean, but in general not the best linear unbiased estimator (BLUE) of
the model-mean. If the random variables are model-independent, the variance
of the sample mean, used as an estimator of the model-mean, can be computed
by

𝜎2
𝑉(𝜇)̂ = , (26.2)
𝑛
with 𝜎2 the model-variance of the random variable (see Equation (21.2)). The
variance presented in Equation (26.2) necessarily is a model-variance as it
quantifies our uncertainty about the model-mean, which only exists in the
model-based approach. If the random variables are not model-independent,
the model-variance of the sample mean can be computed by (de Gruijter et al.,
2006)

𝜎2
𝑉(𝜇)̂ = {1 + (𝑛 − 1)𝜌}̄ , (26.3)
𝑛
with 𝜌 ̄ the mean correlation within the sample (the average of the correlation
of all pairs of sampling points). The term inside the curly brackets is larger
than 1, unless 𝜌 ̄ equals 0. So, the variance of the estimator of the model-mean
with dependent data is larger than when data are independent. The number of
independent observations that is equivalent to a spatially autocorrelated data
set’s sample size 𝑛, referred to as the effective sample size, can be computed
by (de Gruijter et al., 2006)
𝑛
𝑛eff = . (26.4)
{1 + (𝑛 − 1)𝜌}̄
484 26 Design-based approach for sampling and inference

So, if we substitute 𝑛eff for 𝑛 in Equation (26.2), we obtain the variance

presented in Equation (26.3). Equation (26.4) is equivalent to equation (2)
in Griffith (2005). Figure 26.8 shows that the effective sample size decreases
sharply with the mean correlation. With a mean correlation of 0, the effective
sample size equals the actual sample size; with a mean correlation of 1, the
effective sample size equals 1.

FIGURE 26.8: Effective sample sizes as a function of the mean correlation

within the sample, for samples of size 25 and 100.

To illustrate the difference between the model-variance and the design-variance

of a sample mean, I simulated a finite population of 100 units, located at the
nodes of a square grid, with a model-mean of 10, an exponential semivariogram
without nugget, an effective range of three times the distance between adjacent
population units, and a sill of 1 (Figure 26.9). The model-variance of the
average of a simple random sample without replacement of size 𝑛 is computed
using Equation (26.3), and the design-variance of the sample mean, used as
an estimate of the population mean, is computed by (see Equation (3.15))

𝑛 𝑆2
𝑉(𝑧)̄̂ = (1 − ) , (26.5)
𝑁 𝑛

with 𝑁 the total number of population units (𝑁 = 100). This is done for a
range of sample sizes: 𝑛 = 10, 11, … , 100. Note that for 𝑛 < 100 the model-
variance of the sample mean for a given 𝑛, differs between samples. For samples
showing strong spatial clustering, the mean correlation is relatively large, and
consequently the model-variance is relatively large (see Equation (26.3)). There
is less information in these samples about the model-mean than in samples
without spatial clustering of the points. Therefore, to estimate the expectation
26.4 Effective sample size 485

of the model-variance over repeated simple random sampling for a given 𝑛,

I selected 200 simple random samples of that size 𝑛, and I averaged the 200
model-variances. Figure 26.10 shows the result. Both the model-variance and
the design-variance of the sample mean decrease with the sample size. For
all sample sizes, the model-variance is larger than the design-variance. The
design-variance goes to 0 for 𝑛 = 100 (see Equation (26.5)), whereas the model-
variance for 𝑛 = 100 equals 0.0509. This can be explained as follows. Although
with 𝑛 = 100 we know the population mean without error, this population
mean is only an estimate of the model-mean. Recall that the model-mean is
the expectation of the population mean over all realisations of the model.

FIGURE 26.9: Simple random sample without replacement of ten points

from a finite population simulated with a model with a model-mean of 10, a
model-variance of 1, and an exponential semivariogram (without nugget) with
a distance parameter equal to the distance between neighbours (effective range
is three times this distance). The mean correlation within the sample equals
0.135, and the model-variance of the estimator of the model-mean equals 0.222.

In Figure 26.11 we can see that the population mean shows considerable
variation. The variance of 10,000 simulated population means equals 0.0513,
which is nearly equal to the value of 0.0509 for the model-variance computed
with Equation (26.3).
In observational research, I cannot think of situations in which interest is
in estimation of the mean of a superpopulation model. This in contrast to
experimental research. In experimental research, we are interested in the effects
486 26 Design-based approach for sampling and inference

FIGURE 26.10: Model-variance (MB) and design-variance (DB) of the

average of a simple random sample without replacement as a function of the
sample size.

of treatments; think for instance of the effects of different types of soil tillage
on the soil carbon stock. These treatment effects are quantified by different
model-means. Also, in time-series analysis of data collected in observational
studies, we might be more interested in the model-mean than in the mean over
a bounded period of time.

FIGURE 26.11: Frequency distributions of means and variances of 10,000

simulated populations.
26.5 Exploiting spatial structure in design-based approach 487

Now let us return to Equation (26.1). What is wrong with this variance
estimator? Where Griffith (2005) confused the population mean and the model-
mean, Wang et al. (2010) confused the population variance with the sill (a
priori variance) of the random process that has generated the population
(Webster and Oliver, 2007). The parameter 𝜎2 in their formula is defined as
the population variance. In doing so, the variance estimator is clearly wrong.
However, if we define 𝜎2 in this formula as the sill, the formula makes more
sense, but even then, the equation is not fully correct. The variance computed
with this equation is not the design-variance of the average of a simple random
sample selected from the sampled population, but the expectation of this design-
variance over all realisations of the model. So, it is a model-based prediction of
the design-variance of the estimator of the population mean, estimated from a
simple random sample, see Chapter 13. For the population actually sampled,
the design-variance is either smaller or larger than this expectation. Figure
26.11 shows that there is considerable variation in the population variance
among the 10,000 populations simulated with the model. Consequently, for
an individual population, the variance of the estimator of the population
mean, estimated from a simple random sample, can largely differ from the
model-expectation of this variance. Do not use Equation (26.1) for estimating
the design-variance of the estimator of the population mean, but simply use
Equation (26.5) (for simple random sampling with replacement and simple
random sampling of infinite populations the term (1 − 𝑛/𝑁) can be dropped).
Equation (26.1) is only relevant for comparing simple random sampling under
a variety of models of spatial variation (Ripley (1981), Domburg et al. (1994)).

26.5 Exploiting spatial structure in design-based

approach
A further misconception is that, in the design-based approach, the possibilities
of exploiting our knowledge about the spatial structure of the study variable
are limited, because the sampling units are selected randomly. This would
indeed be a very serious drawback, but happily enough, this is not true. There
are various ways of utilising this knowledge. Our knowledge about the spatial
structure can be used either at the stage of designing the sample and/or at
the stage of the statistical inference once the data are collected (Table 26.1).
I distinguish the situation in which maps of covariates are available from the
situation in which such maps are lacking. In the first situation, the covariate
maps can be used, for instance, to stratify the population (Chapter 4). With
a quantitative covariate, optimal stratification methods are available. Other
options are, for instance, pps sampling (Chapter 8), and balanced sampling
and well-spread sampling in covariate space with the local pivotal method
488 26 Design-based approach for sampling and inference

TABLE 26.1: Strategies in the design-based approach for exploiting

knowledge about the spatial structure of the study variable.

Stage Covariates available No covariates

Sampling Stratified random sampling Systematic random sampling
pps sampling Compact geographical stratification
Balanced sampling Geographical spreading with LPM
Covariate space spreading with LPM GRTS sampling
Inference Model-assisted: regression model Model-assisted: spatial regression model
LPM: local pivotal method.

(Chapter 9). At the inference stage, the covariate maps can be used in a model-
assisted approach, using, for instance, a linear regression model to increase the
precision of the design-based estimator (Chapter 10, Section 26.6).
If no covariate maps are available, we may anticipate the presence of spatial
structure by spreading the sampling units throughout the study area. This
spreading can be done in many ways, for instance by systematic random sam-
pling (Chapter 5), compact geographical stratification (Section 4.6), well-spread
sampling in geographical space with the local pivotal method (LPM) (Subsec-
tion 9.2.1), and generalised random-tessellation stratified (GRTS) sampling
(Subsection 9.2.2). At the inference stage, again a model-assisted approach can
be advantageous, using the spatial coordinates in a regression model.

26.6 Model-assisted vs. model-dependent

In this section the difference between the model-assisted approach and the
model-based approach is explained. The model-assisted approach is a hybrid
approach in between the design-based and the model-based approach. It tries
to build the strength of the model-based approach, a potential increase of
the accuracy of estimates, into the design-based approach. As in the design-
based approach, sampling units are selected by probability sampling, and
consequently bias and variance are defined as design-bias and design-variance
(Table 26.2). As in the model-based approach, a superpopulation model is used.
However, the role of this model in the two approaches is fundamentally different.
In both approaches we assume that the population of interest is a realisation of
the superpopulation model. However, as explained above, in the model-based
approach, the statistical properties of the estimators (predictors), such as bias
and variance, are defined over all possible realisations of the model (Table
26.2). So, unbiasedness and minimum variance of an estimator (predictor)
means model-unbiasedness and minimum model-variance. On the contrary, in
26.6 Model-assisted vs. model-dependent 489

TABLE 26.2: Three statistical approaches for sampling and inference.

Approach Sampling Inference Regression coefficients Quality criteria

Design-based Prob. sampling Design-based No model Design-bias,

Design-variance
Model-assisted Prob. sampling Model-assisted Population par. Design-bias,
Design-variance
Model-based No requirement Model-depend. Superpop. par. Model-bias,
Model-variance

the model-assisted approach, the model is used to derive an efficient estimator

(Chapter 10). To stress its different role in the model-assisted approach, the
model is referred to as a working model.
An important property of model-assisted estimators is that, if a poor working
model is used (our assumptions about how our population is generated are
incorrect), then for moderate sample sizes the results are still valid, i.e., the
empirical coverage rate of a model-assisted estimate of the confidence interval of
the population mean still is approximately equal to the nominal coverage rate.
This is because the mismatch of the superpopulation model and the applied
model-assisted estimator results in a large design-variance of the estimator of
the population mean. This is illustrated with a simulation study, in which I
compare the effect of using a correct vs. an incorrect model in estimation.
A population is simulated with a simple linear regression model with an
intercept of 15 (𝛽0 = 15), a slope coefficient of 0.5 (𝛽1 = 0.5), and a constant
residual standard deviation of 2 (𝜎u� = 2). This is done by first simulating a
population with covariate values with a model-mean of 20 (𝜇(𝑥) = 20), using
an exponential semivariogram without nugget, a sill variance of 25, and a
distance parameter of 20 distance units. This field with covariate values is then
linearly transformed using the above-mentioned regression coefficients. Finally,
‘white noise’ is added by drawing independently for each population unit a
random number from a normal distribution with zero mean and a standard
deviation of 2 (Figure 26.12).
The population mean of the study variable equals 25.052, which is pretty close
to the known model-mean 𝜇(𝑧): 𝛽0 + 𝛽1 𝜇(𝑥) = 15 + 0.5 ⋅ 20.
Figure 26.13 shows a scatter plot for all population units. The Pearson correla-
tion coefficient equals 0.745. Two models are fitted to the exhaustive scatter
plot: a simple linear regression model and a ratio model. The ratio model
assumes that the intercept 𝛽0 equals 0 and that the residual variance is propor-
tional to the covariate values: 𝜎2u� ∝ 𝑥. The population fits of the coefficients of
the simple linear regression model are 14.9989 and 0.4997, which are very close
to the model regression coefficients. The fitted ratio model is clearly very poor.
The residual standard deviation of the population fit of the ratio model equals
3.887, which is much larger than 2.001 of the simple linear regression model.
490 26 Design-based approach for sampling and inference

FIGURE 26.12: Realisation of simple linear regression model. x is the

covariate, z is the study variable.

FIGURE 26.13: Exhaustive scatter plot of the simulated population, with

population fit of a simple linear regression model (green line), and of a ratio
model fitted with weights inversely proportional to the covariate (red line).
26.6 Model-assisted vs. model-dependent 491

The population mean of the study variable is estimated by selecting 5,000

times a simple random sample of 25 units. Each sample is used to estimate
the population mean by two model-assisted estimators: the simple regression
estimator and the ratio estimator (Chapter 10). The first estimator correctly
assumes that the population is a realisation of a simple linear regression model,
whereas the latter incorrectly assumes that it is a realisation of a ratio model.
For each sample, the standard error of the two estimators are estimated as
well, which is used to compute a 95% confidence interval of the population
mean. Then the empirical coverage rate is computed, i.e., the proportion of
samples for which the population mean is inside the 95% confidence interval.
Ideally, this empirical coverage rate is equal to the nominal coverage rate of
0.95.
The coverage rates of the simple regression estimator and the ratio estimator
equal 0.931 and 0.948, respectively. Both coverage rates are very close to the
nominal coverage rate of 0.95. So, despite the fact that the ratio estimator
assumes an improper superpopulation model, the estimated confidence interval
is still valid. The price we pay for the invalid model assumption is not an
overestimated coverage rate of a confidence interval, but an increased standard
error of the estimated population mean. The average over the 5,000 samples of
the estimated standard error of the regression estimator equals 0.387, whereas
that of the ratio estimator equals 0.772. The larger standard error of the ratio
estimator leads to wider confidence intervals, which explains that the coverage
rate is still correct.
This sampling experiment is now repeated for samples sizes 𝑛 = 10, 25, 50, 100
and for confidence levels 1 − 𝛼 = 0.01, 0.02, … , 0.99.
Figures 26.14 and 26.15 show that the empirical coverage rates are close to
the nominal coverage rate, for all four sample sizes, both estimators, and
all confidence levels. For the regression estimator and 𝑛 = 10, the empirical
coverage rate is somewhat too small. This is because the standard error
of the regression estimator is slightly underestimated at this sample size.
The average of the estimated standard errors (square root of estimated variance
of regression estimator) equals 0.609, which is somewhat smaller than the
standard deviation of the 5,000 regression estimates of 0.678. For all sample
sizes, the standard deviation of the 5,000 ratio estimates is considerably
larger than that of the 5,000 regression estimates (Table 26.3). For 𝑛 = 10
also the standard error of the ratio estimator is underestimated (the average
of the 5,000 estimated standard errors is smaller than the standard deviation of
the 5,000 ratio estimates), but as a percentage of the standard deviation of the
5,000 ratio estimates, this underestimation is smaller than for the regression
estimator.
492 26 Design-based approach for sampling and inference

FIGURE 26.14: Empirical vs. nominal coverage rates of confidence intervals

for the population mean, estimated by the simple regression estimator, for
sample sizes 10, 25, 50, and 100.

The relative bias, computed by

5000
1
∑u�=1 (𝑧u�̄̂ − 𝑧)̄
(26.6)
5000
𝑏𝑖𝑎𝑠 = ,
𝑧̄

is about 0 for both estimators and all four sample sizes.

Contrarily, if in the model-based approach a poor superpopulation model is
used, the predictions and the prediction error variances still are model-unbiased.
However, for the sampled population, we may have serious systematic error in
the estimated population mean and the variance of local predictions may be
26.6 Model-assisted vs. model-dependent 493

FIGURE 26.15: Empirical versus nominal coverage rates of confidence in-

tervals for the population mean, estimated by the ratio estimator, for sample
sizes 10, 25, 50, and 100.

seriously over- or underestimated. For this reason, model-based inference is

also referred to as model-dependent inference, stressing that we fully rely on
the model and that the validity of the estimates and predictions depends on
the quality of the model (Hansen et al., 1983).
494 Design-based approach for sampling and inference

TABLE 26.3: Estimated relative bias of the regression estimator and ratio
estimator, standard deviation of 5,000 regression/ratio estimates, and average
of 5,000 estimated standard errors of the regression/ratio estimator.

Estimator n Bias (%) Standard deviation Average standard error

Regression 10 0.0037 0.678 0.609
Regression 25 0.0086 0.411 0.397
Regression 50 0.0172 0.282 0.282
Regression 100 0.0058 0.202 0.200
Ratio 10 -0.3191 1.273 1.206
Ratio 25 -0.1318 0.776 0.769
Ratio 50 -0.0611 0.548 0.549
Ratio 100 0.0175 0.390 0.388
n: sample size.
A
Answers to exercises

R scripts of the answers to the exercises are available in the Exercises folder
at the github repository of this book.

Introduction to probability sampling

1. No, this is not a probability sample because, with this implementa-
tion, the probabilities of selection of the units are unknown.

2. For simple random sampling without replacement, the inclusion

probability is 0.5 (𝜋u� = 𝑛/𝑁 = 2/4). For simple random sampling
with replacement, the inclusion probability is 0.4375 (𝜋u� = 1 − (1 −
1/𝑁)u� = 1 − 0.752 ).

Simple random sampling

1. The most remarkable difference is the much smaller range of values
in the sampling distribution of the estimator of the population
mean (Figure 3.2). This can be explained by the smaller variance of
the average of 𝑛 randomly selected values compared to the variance
of an individual randomly selected value. A second difference is that
the sampling distribution is more symmetric, less skewed to the
right. This is an illustration of the central limit theorem.

2. The variance (and so the standard deviation) becomes smaller.

3. Then the difference between the average of the estimated population

means and the true population mean will be very close to 0, showing
that the estimator is unbiased.

495
496 A Answers to exercises

4. For simple random sampling without replacement (from a finite

population), the sampling variance will be smaller. When units are
selected with replacement, a unit can be selected more than once.
This is inefficient, as there is no extra information in the unit that
has been selected before.

5. The larger the population size 𝑁, the smaller the difference between
the sampling variances of the estimator of the mean for simple
random sampling with replacement and simple random sampling
without replacement (given a sample size 𝑛).

6. The true sampling variance of the estimator of the mean for simple
random sampling from an infinite population can be computed with
the population variance divided by the sample size: 𝑉(𝑧)̄̂ = 𝑆2 (𝑧)/𝑛.

7. In reality, we cannot compute the true sampling variance, because

we do not know the values of 𝑧 for all units in the population, so
that we do not know the population variance 𝑆2 (𝑧).

8. See SI.R1 . The 90% confidence interval is less wide than the 95%
interval, because a larger proportion of samples is allowed not to
cover the population mean. The estimated standard error of the
estimated total underestimates the true standard error, because a
constant bulk density is used. In reality this bulk density also varies.

Stratified simple random sampling

2
1. See STSI1.R .

2. Strata EA and PA can be merged without losing much precision:

their means are about equal.

3. See STSI2.R3 . The true sampling variance of the 𝜋 estimator of the

mean SOM obtained by collapsing strata EA and PA equals 42.89,
whereas the sampling variance with the original stratification equals
42.53. So, the new stratification with four strata is only slightly worse.

1 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SI.R
2 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/STSI1.R
3 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/STSI2.R
A.0 Answers to exercises 497

4. The proof is as follows: ∑u� 𝜋u� = ∑u� ∑u� 𝜋ℎu� =

ℎ
∑u� ∑u� 𝑛ℎ /𝑁ℎ = ∑u� 𝑛ℎ = 𝑛.
ℎ

5. See STSIcumrootf.R4 . The default allocation is Neyman allocation, see

help of function strata.cumrootf. The true sampling variance of the
estimator of the mean equals 20.0. The stratification effect equals
4.26.

6. With at least two points per geostratum, the variance of the

estimator of the stratum mean can be estimated without bias by
the estimated stratum variance divided by the number of points in
that stratum.

7. On average, the sampling variance of the estimator of the mean

with 100 × 1 point is smaller than with 50 × 2 points, because
the geographical spreading will be somewhat better (less spatial
clustering).

8. With geostrata of equal size and equal number of sampling points

per geostratum, the sampling intensity is equal for all strata, so that
the sample mean is an unbiased estimator of the population mean. In
u� u�
formula: 𝑧 ̄̂ = ∑ 𝑤ℎ 𝑧u�ℎ
̄ = 1
u� ̄ = 𝑧u�̄ , with 𝑧u�ℎ
∑ 𝑧u�ℎ ̄ the average of
ℎ=1 ℎ=1
the sample from stratum ℎ and 𝑧u�̄ the average of all sampling points.

9. See STSIgeostrata.R5 .
• Collapsing the geostrata on the basis of the measurements of
the study variable is not a proper way, as it will lead to a biased
estimator of the sampling variance of the estimator of the mean.
The estimated stratum variances 𝑆2̂ (𝑧) will be small, and so
the estimated sampling variance will underestimate the true
sampling variance.
• I propose to group neighbouring geostrata, i.e., geostrata that
are close to each other.
• The sampling variance estimator is not unbiased. The sampling
variance is slightly overestimated, because we assume that the
two (or three) points within a collapsed stratum are selected
by simple random sampling, whereas they are selected by
stratified random sampling (a collapsed stratum consists of
two or three geostrata), and so there is less spatial clustering
compared to simple random sampling.

4 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/STSIcumrootf.R
5 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/STSIgeostrata.R
498 A Answers to exercises

10. See STSIgeostrata_composite.R6 .

• No, with bulking within strata the sampling variance cannot
be estimated, because then we cannot estimate the sampling
variances of the estimated stratum means, which are needed
for estimating the sampling variance of the estimator of the
population mean.
• If all aliquots are analysed separately, the estimated population
mean is more precise than with composite sampling (variance of
the estimator of the mean is smaller), because the contribution
of the measurement error to the total variance of the estimator
of the mean is smaller.
• This combination of arguments of function stratify does not
work, because with geostrata of unequal area the mean of a
composite sample is a biased estimator of the population mean.
All aliquots bulked into a composite get equal weight, but they
should get different weights, because they do not represent
equal fractions of the population.

Systematic random sampling

7
1. See SY.R .

2. As can be seen in the plot, the spatial coverage of the study area by
the two systematic random samples can be quite poor. So, I expect
that the variance of the estimator of the mean using the data of two
systematic random samples of half the expected size is larger than
the variance of the estimator of the mean based on the data of a
single systematic random sample.

Cluster random sampling

8
1. See Cluster.R .

6 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/STSIgeostrata_composite.

R
7 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SY.R
8 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/Cluster.R
A.0 Answers to exercises 499

2. I expect that the sampling variance with three transects is larger

than with six transects of half the length, as the sampling points
are more spatially clustered.

3. With two independently selected clusters per stratum, the sampling

variance of the estimator of the mean can be estimated without bias,
as the variance of cluster means within the strata can be estimated
from the two cluster means.

Two-stage cluster random sampling

9
1. See TwoStage.R .

2. With ten PSU draws and four SSUs per PSU draw (10 × 4), the
expected standard error of the estimator of the population mean
is smaller than with four PSU draws and ten SSUs per PSU draw
(4 × 10), because spatial clustering of the sampling points is less
strong.

10
3. See TwoStage.R .

11
4. See TwoStage.R .

12
5. See TwoStage.R .

6. For the first variance component:

2 2
1 u� 𝑡u� (𝑧) 1 u� 𝑡u� (𝑧)
∑ 𝑝u� ( − 𝑡(𝑧)) = ∑ 𝑝u� (𝑀 − 𝑀𝑧)
̄
𝑛 u�=1 𝑝u� 𝑛 u�=1 𝑀u�
(A.1)
1 u� 2 𝑀2 u� 2
= ∑ 𝑝u� (𝑀 (𝑧u�̄ − 𝑧))
̄ = ∑ 𝑝 (𝑧 ̄ − 𝑧)̄ .
𝑛 u�=1 𝑛 u�=1 u� u�

9 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/TwoStage.R
10 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/TwoStage.R
11 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/TwoStage.R
12 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/TwoStage.R
500 A Answers to exercises

For the second variance component:

1 u� 𝑀2u� 𝑆2u� 1 u� 𝑀2u� 𝑆2u� 1 u�

∑ = ∑ = ∑ 𝑀𝑀u� 𝑆2u�
𝑛 u�=1 𝑚u� 𝑝u� 𝑛𝑚 u�=1 𝑝u� 𝑛𝑚 u�=1
(A.2)
1 u� 2 𝑀u� 2 𝑀2 u�
= ∑𝑀 𝑆u� = ∑ 𝑝 𝑆2 .
𝑛𝑚 u�=1 𝑀 𝑛𝑚 u�=1 u� u�

Division of both variance components by 𝑀2 yields the variance

of the estimator of the population mean, see Equations (7.3), (7.4),
and (7.5).

Sampling with probabilities proportional to size

13
1. See pps.R .

2. No, this field should not be included in the poppy area of that
sampling unit, because it is located outside the target area.

3. Yes, this field must be included in the poppy area of that sampling
unit, as it is located inside the target area. The target area is the
territory of Kandahar, regardless of how an area inside this territory
is depicted on the map, as agricultural land or otherwise.

Balanced and well-spread sampling

14
1. See Balanced.R .

2. Spatial clustering of sampling units with balanced sampling may

lead to a less precise estimate of the population mean. This will be
the case when the residuals of the regression model are spatially
correlated (show spatial structure). The residuals will be correlated
when the spatial variation of the study variable is also determined
by covariates or factors that are not used in balancing the sample.
If the residuals are not spatially correlated (white noise), spatial
clustering does no harm.
13 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/pps.R
14 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/Balanced.R
A.0 Answers to exercises 501

3. One advantage is that unequal inclusion probabilities can be used

in the LPM design. If the sampling units have unequal size (as in
the poppy survey of Kandahar), or if a covariate is available that
is linearly related to the study variable (as in the AGB survey of
Eastern Amazonia), the sampling efficiency can be increased by
sampling with (inclusion) probabilities proportional to size. The
only option for random sampling from geostrata is then to select
the units within geostrata by pps sampling.

Model-assisted estimation
1. See RegressionEstimator.R15 . The approximate standard error esti-
mator that uses the 𝑔-weights (computed with functions calibrate
and svymean of package survey) has a larger mean (7.194) than
the approximated standard error (7.130) computed with Equation
(10.13).

2. See VarianceRegressionEstimator.R16 . In reality, we do not have a pop-

ulation fit of the regression coefficients, but these coefficients must
be estimated from a sample. The estimated coefficients vary among
the samples, which explains that the experimental variance, i.e.,
the variance of the 10,000 regression estimates obtained by esti-
mating the coefficients from the sample (Sample in Figure A.1), is
larger than the variance as computed with the population fit of the
regression coefficients (Exhaust in Figure A.1).
The difference between the experimental variance (variance of re-
gression estimator with sample fit of coefficients) and the variance
obtained with the population fit, as a proportion of the experimental
variance, decreases with the sample size. The same holds for the
difference between the approximated variance and the experimental
variance as a proportion of the experimental variance. Both findings
can be explained by the smaller contribution of the variance of the
estimated regression coefficients to the variance of the regression
estimator with the large sample size. The approximated variance
does not account for the uncertainty about the regression coefficients,
so that for all three sample sizes this approximated variance is about
15 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/RegressionEstimator.R
16 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/

VarianceRegressionEstimator.R
502 A Answers to exercises

equal to the variance of the regression estimator as computed with

the population fit of the regression coefficients.

FIGURE A.1: Variance of the regression estimator of the mean AGB in

Eastern Amazonia with population fit of regression coefficients (Exhaust), with
sample fit of regression coefficients (Sample), and approximated variance of
regression estimator (Approx).

3. See RatioEstimator.R17 . The population fit of the slope coefficient of

the homoscedastic model differs from the ratio of the population
total poppy area to the population total agricultural area. For the
heteroscedastic model, these are equal.

Two-phase random sampling

1. See RegressionEstimator_Twophase.R18 . Figure A.2 shows the approxi-
mated sampling distribution of the simple regression estimator of
the mean AGB in Eastern Amazonia when lnSWIR2 is observed
for all sampling units (one-phase), and when AGB is observed for
17 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/RatioEstimator.R
18 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/RegressionEstimator_

Twophase.R
A.0 Answers to exercises 503

the subsample only (two-phase). The variance of the regression

estimator with two-phase sampling is considerably larger. Without
subsampling, the regression estimator exploits our knowledge of
the population mean of the covariate lnSWIR2, whereas in two-
phase sampling, this population mean must be estimated from the
first-phase sample, introducing additional uncertainty.
The average of the 10,000 approximated variances equals 40.5 (109
kg ha-1 )2 , which is considerably smaller than the variance of the
10,000 regression estimates for two-phase sampling, which is equal
to 51.2 (109 kg ha-1 )2 .

FIGURE A.2: Approximated sampling distribution of the simple regression

estimator of the mean AGB (109 kg ha-1 ) in Eastern Amazonia in the case
that the covariate is observed for all sampling units (one-phase) and for the
subsample only (two-phase).
504 A Answers to exercises

Computing the required sample size

1. See RequiredSampleSize_CIprop.R19 . Figure A.3 shows that the required
sample size decreases sharply with the length of the confidence
interval and increases with the prior (anticipated) proportion.
A prior for the proportion is needed because the standard error of
the estimated proportion is a function of the estimated proportion 𝑝̂
itself: 𝑠𝑒(𝑝)̂ = √u� , so that the length of the confidence interval,
√u�(1−
̂ u�)
̂

computed with the normal approximation, is also a function of 𝑝,̂

see Equation (12.11).
For a prior proportion 𝑝∗ of 0.5 the standard deviation √𝑝∗ (1 − 𝑝∗ )
is maximum. The closer the prior proportion to zero or one, the
smaller the standard error of the estimated proportion, the smaller
the required sample size.

FIGURE A.3: Required sample size as a function of the half-length of a 95%

confidence interval of the population proportion, for a prior proportion of 0.1
(left subfigure), and as a function of the prior proportion for a half-length of a
95% confidence interval of 0.2 (right subfigure).

2. See RequiredSampleSize_CIprop.R20 . There is no need to compute the

required sample size for prior proportions > 0.5, as this required
19 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/RequiredSampleSize_

CIprop.R
20 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/RequiredSampleSize_

CIprop.R
A.0 Answers to exercises 505

sample size is symmetric. For instance, the required sample size for
𝑝∗ = 0.7 is equal to the required sample size for 𝑝∗ = 0.3.

Model-based optimisation of probability sampling designs

1. See MBSamplingVarSI_VariogramwithNugget.R21 . The predicted sampling
variance is slightly larger compared to the predicted sampling
variance obtained with the semivariogram without nugget (and
the same sill and range), because 50% of the spatial varia-
tion is not spatially structured, so that the model-expectation of
the population variance (the predicted population variance) is larger.

22
2. See first part of MBRequiredSampleSize_SIandSY.R .

3. See second part of MBRequiredSampleSize_SIandSY.R. The model-based

prediction of the required sample size for simple random sampling
is 34 and for systematic random sampling 13. The design effect at a
sample size of 34 equals 0.185. The design effect decreases with the
sample size, i.e., the ratio of the variance with systematic random
sampling to the variance with simple random sampling becomes
smaller. This is because the larger the sample size, the more we
profit from the spatial correlation.

Repeated sample surveys for monitoring population

parameters
1. See SE_STparameters.R23 . For the designs SP and RP, the true
standard errors of all space-time parameters are slightly smaller
than the standard deviations in Table 15.1, because in the sampling
experiment the estimated covariances of the elementary estimates
are used in the GLS estimator of the spatial means, whereas in this
exercise the true covariances are used. The estimated covariances
21 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSamplingVarSI_

VariogramwithNugget.R
22 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBRequiredSampleSize_

SIandSY.R
23 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SE_STparameters.R
506 A Answers to exercises

vary among the space-time samples. This variation propagates to

the GLS estimates of the spatial means and so to the estimated
space-time parameters.

2. See SE_ChangeofMean_HT.R24 . The standard error of the change with

the GLS estimators of the two spatial means is much smaller than
the standard error of the change with the 𝜋 estimators, because the
GLS estimators use the data of all four years to estimate the spatial
means of 2004 and 2019, whereas with the 𝜋 estimators only the
data of 2004 and 2019 are used.

Regular grid and spatial coverage sampling

1. See SquareGrid.R25 . The number of grid points specified with
argument n is the expected number of grid points over repeated
selection of square grids with a random start. With a fixed start
(using argument offset), the number of grid points can differ from
the expected sample size.

2. The optimal spatial coverage sample, optimal in terms of MSSD,

consists of the four points in the centre of the four subsquares of
equal size.

3. If we are also interested in the accuracy of the estimated plot means,

the sampling units can best be selected by probability sampling,
for instance by simple random sampling, from the subsquares, or
strata. Preferably at least two points should then be selected from
the strata, see Section 4.6.

26
4. See SpatialCoverageCircularPlot.R . See Figure A.4.

5. Bias can be avoided by constructing strata of equal size. Note that

in this case we cannot use function spsample to select the centres of
these geostrata. These centres must be computed by hand.

27
6. See SpatialInfill.R .

24 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SE_ChangeofMean_HT.R
25 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SquareGrid.R
26 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/

SpatialCoverageCircularPlot.R
27 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SpatialInfill.R
A.0 Answers to exercises 507

FIGURE A.4: Spatial coverage samples of five and six points in a circular
plot.

Covariate space coverage sampling

28
1. See CovariateSpaceCoverageSample.R . See Figure A.5.

FIGURE A.5: Covariate space coverage sample from Hunter Valley, using
cti, ndvi, and elevation as clustering variables, plotted on a map of cti.

28 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/

CovariateSpaceCoverageSample.R
508 A Answers to exercises

Conditioned Latin hypercube sampling

1. See cLHS.R29 . Most units are selected in the part of the diagram with
the highest density of raster cells. Raster cells with a large cti value
and low elevation and raster cells with high elevation and small cti
value are (nearly) absent in the sample. In the population, not many
raster cells are present with these combinations of covariate values.

2. See cLHS_Square.R30 .
• Spatial coverage is improved by using the spatial coordinates
as covariates, but it is not optimal in terms of MSSD.
• It may happen that not all marginal strata of 𝑠1 and 𝑠2 are
sampled. Even when all these marginal strata are sampled, this
does not guarantee a perfect spatial coverage.
• With set.seed(314) and default values for the arguments of
function clhs, there is one unsampled marginal stratum and one
marginal stratum with two sampling locations. So, component
O1 equals 2. The minimised value (2.62) is slightly larger due
to the contribution of O3 to the criterion.

Model-based optimisation of the grid spacing

1. See MBGridspacing_QOKV.Rmd31 . For P50 not to exceed 0.85 the tolerable
grid spacing is about 11.7 km, for P80 it is 9.4 km, and for P95 it is
7.1 km (Figure A.6).

2. See MBGridspacing_Sensitivity.Rmd32 . Increasing the nugget by 5% and

decreasing the range by 5% yields a tolerable grid spacing that is
smaller than that with the original semivariogram (Figure A.7). The
tolerable grid spacings for a mean kriging variance of 0.85 are 10.6,
8.9, and 7.4 km for the original semivariogram, the semivariogram
with increased nugget, and the semivariogram with the smaller range,
respectively, leading to a required expected sample size of 97, 137,
and 200 points.
29 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/cLHS.R
30 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/cLHS_Square.R
31 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBGridspacing_QOKV.Rmd
32 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBGridspacing_

Sensitivity.Rmd
A.0 Answers to exercises 509

FIGURE A.6: Three quantiles of the ordinary kriging variance of predicted

SOM concentrations in West-Amhara, as a function of the grid spacing.

FIGURE A.7: Mean ordinary kriging variance of predicted SOM concentra-

tions in West-Amhara, as a function of grid spacing for three semivariograms.

3. The variation in MKV for a given grid spacing can be explained

by the random sample size: for a given spacing, the number
of points of a randomly selected grid inside the study area is
not fixed but varies. Besides, the covariate values at the grid
510 A Answers to exercises

points vary, so that also the variance of the estimator of the mean,
which contributes to the kriging variance, differs among grid samples.

4. See MBGridspacing_MKEDV.Rmd33 . The tolerable grid spacing for a mean

kriging variance of 0.165 is 79 m.

Model-based optimisation of the sampling pattern

1. See MBSampleSquare_OK.Rmd34 . The optimised sample (Figure A.8) is
most likely not the global optimum. The spatial pattern is somewhat
irregular. I expect the optimal sampling locations to be close to the
centres of the subsquares.

FIGURE A.8: Optimised sampling pattern of 16 points in a square for

ordinary kriging.
33 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBGridspacing_MKEDV.Rmd
34 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSampleSquare_OK.Rmd
A.0 Answers to exercises 511

2. See MBSample_QOKV.Rmd35 . Figure A.9 shows the optimised sampling

pattern. Compared with the optimised sampling pattern using the
mean ordinary kriging variance (MOKV) as a minimisation criterion
(Figure 23.3), the sampling locations are pushed more to the border
of the study area. This is because with a sample optimised for
MOKV (and a spatial coverage sample) near the border the kriging
variances are the largest. By pushing sampling locations towards
the border, the kriging variances in this border zone are strongly
reduced.

FIGURE A.9: Optimised sampling pattern of 50 points on the Cotton

Research Farm, using the P90 of ordinary kriging predictions of lnECe as a
minimisation criterion.

3. See MBSampleSquare_KED.Rmd36 . Figure A.10 shows the optimised sam-

pling patterns with the three semivariograms.
• With zero nugget and a (partial) sill of 2, the sampling points
are well spread throughout the area (subfigure on the left).
• With a nugget of 1.5 and a partial sill of 0.5, the sampling points
are pushed towards the left and right side of the square. With
this residual semivariogram, the contribution of the variance of
the predictor of the mean (as a proportion) to the total kriging
variance is larger than with the previous semivariogram. By
shifting the sampling points towards the left and right side of
the square this contribution becomes smaller. At the same time
the variance of the interpolation error increases as the spatial
coverage becomes worse. The optimised sample is the right
35 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSample_QOKV.Rmd
36 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSampleSquare_KED.Rmd
512 A Answers to exercises

balance of these two variance components (subfigure in the

middle).
• With a pure nugget semivariogram, all sampling points are at
the left and right side of the square. This is because with a pure
nugget semivariogram, the variance of the interpolation error
is independent of the locations (the variance equals the nugget
variance everywhere), while the variance of the predictor of the
mean is minimal for this sample (subfigure on the right).

FIGURE A.10: Effect of the nugget (no nugget, large nugget, pure nugget)
on the optimised sampling pattern of 16 points for KED, using Easting as a
covariate for the mean.

Sampling for estimating the semivariogram

37
1. See NestedSampling_v1.R .

2. See SI_PointPairs.R38 . With the seed I used (314), the variance of

the estimator of the range parameter with the smaller separation
distances is much smaller compared to that obtained with the larger
separation distances (the estimated standard error is 115 m).

3. See MBSample_SSA_logdet.R39 . Figure A.11 shows the optimised sam-

pling pattern. The smaller ratio of spatial dependence of 0.5 (larger
nugget) results in one cluster of sampling points and many point-
pairs with nearly coinciding points.
37 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/NestedSampling_v1.R
38 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/SI_PointPairs.R
39 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSample_SSA_logdet.R
A.0 Answers to exercises 513

4. See MBSample_SSA_MVKV.R40 . Figure A.11 shows the optimised sampling

pattern. The circular cluster of sampling points covers a larger area
than the cluster obtained with a ratio of spatial dependence of 0.8.

FIGURE A.11: Model-based sample for estimating the semivariogram, using

the log of the determinant of the inverse Fisher information matrix (logdet) and
the mean variance of the kriging variance (MVKV) as a minimisation criterion.
The sampling pattern is optimised with an exponential semivariogram with a
range of 200 m and a ratio of spatial dependence of 0.5.

5. See MBSample_SSA_MEAC.R41 . Figure A.12 shows the optimised sampling

pattern of the 20 sampling points together with the 80 spatial
coverage sampling points. The minimised MEAC value equals 0.801,
which is slightly smaller than that for the spatial coverage sample
of 90 points supplemented by 10 points (0.808).

Sampling for validation of maps

1. I am not certain about that, because the computed MSEs are
estimates of the population MSEs only and I am uncertain about
both population MSEs.

40 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSample_SSA_MVKV.R
41 https://github.com/DickBrus/SpatialSamplingwithR/tree/master/Exercises/MBSample_SSA_MEAC.R
514 A Answers to exercises

FIGURE A.12: Optimised sample of 20 points supplemented to a spatial

coverage sample of 80 points, using MEAC as a minimisation criterion. The
sampling pattern of the supplemental sample is optimised with an exponential
semivariogram with a range of 200 m and a ratio of spatial dependence of 0.8.

2. The standard errors of the estimated MEs are large when related
to the estimated MEs, so my guess is that we do not have enough
evidence against the hypothesis that there is no systematic error.

3. Both standard errors are large compared to the difference in MSEs,

so maybe there is no significant difference. However, we must be
careful, because the variance of the difference in MSEs cannot be
computed as the sum of the variances of estimated MSEs. This is
because the two prediction errors at the same location are correlated,
so the covariance must be subtracted from the sum of the variances
to obtain the variance of the estimator of the difference in MSEs.
Bibliography

Aarts, E. and Korst, J. (1989). Simulated Annealing and Boltzmann Machines:

A Stochastic Approach to Combinatorial Optimization and Neural Computing.
John Wiley & Sons, Chichester.
Adcock, C. J. (1988). A Bayesian approach to calculating sample sizes. Statis-
tician, 37:433–439.
Akramkhanov, A., Brus, D. J., and Walvoort, D. J. J. (2014). Geostatistical
monitoring of soil salinity in Uzbekistan by repeated EMI surveys. Geoderma,
213:600–607.
Anonymous (2014). Afghanistan opium survey 2014. Technical report, United
Nations Office on Drugs and Crime, Islamic Republic of Afghanistan, Min-
istry of Counter Narcotics.
Arthur, D. and Vassilvitskii, S. (2007). k-means++: The advantages of careful
seeding. Proceedings of the Annual ACM-SIAM Symposium on Discrete
Algorithms, 07-09-January-2007:1027–1035.
Baccini, A., Goetz, S. J., Walker, W. S., Laporte, N. T., Sun, M., Sulla-Menashe,
D., Hackler, J., Beck, P. S. A., Dubayah, R., Friedl, M. A., Samanta, S., and
Houghton, R. A. (2012). Estimated carbon dioxide emissions from tropical
deforestation improved by carbon-density maps. Nature Climate Change,
2(3):182–185.
Bache, S. M. and Wickham, H. (2020). magrittr: A Forward-Pipe Operator for
R. R package version 2.0.1.
Baillargeon, S. and Rivest, L.-P. (2011). The construction of stratified designs
in R with the package stratification. Survey Methodology, 37(1):53–65.
Ballabio, C., Lugato, E., Fernández-Ugalde, O., Orgiazzi, A., Jones, A., Borrelli,
P., Montanarella, L., and Panagos, P. (2019). Mapping LUCAS topsoil
chemical properties at European scale using Gaussian process regression.
Geoderma, 355:113912.
Ballin, M., Barcaroli, G., Masselli, M., and Scarnó, M. (2018). Redesign sample
for Land Use/Cover Area frame Survey (LUCAS) 2018. Technical report,
Eurostat.
Barcaroli, G. (2014). SamplingStrata: An R Package for the Optimization of
Stratified Sampling. Journal of Statistical Software, 61(4):1–24.

515
516 Bibliography

Barcaroli, G., Ballin, M., Odendaal, H., Pagliuca, D., Willighagen, E., and
Zardetto, D. (2020). SamplingStrata: Optimal Stratification of Sampling
Frames for Multipurpose Sampling Surveys. R package version 1.5-1.
Barnes, R. J. (1988). Bounding the required sample size for geologic site
characterization. Mathematical Geology, 20:477–490.
Berger, Y. G. (2004). A simple variance estimator for unequal probability
sampling without replacement. Canadian Journal of Applied Statistics,
31(3):305–315.
Bethel, J. (1989). Sample allocation in multivariate surveys. Survey Methodol-
ogy, 15(1):47–57.
Binder, D. A. and Hidiroglou, M. A. (1988). Sampling in time. In Krishnaiah,
P. R. and Rao, C. R., editors, Handbook of Statistics, volume 6, pages
187–211. North-Holland, Amsterdam.
Bivand, R. S., Pebesma, E., and Gómez-Rubio, V. (2013). Applied Spatial
Data Analysis with R. Springer, New York, second edition.
Bogaert, P. and Russo, D. (1999). Optimal spatial sampling design for the
estimation of the variogram based on a least squares approach. Water
Resources Research, 35(4):1275–1289.
Breidaks, J., Liberts, M., and Jukams, J. (2020). surveyplanning: Survey
planning tools. R package version 4.0.
Breidenbach, J. (2018). JoSAE: Unit-Level and Area-Level Small Area Esti-
mation. R package version 0.3.0.
Breidenbach, J. and Astrup, R. (2012). Small area estimation of forest attributes
in the Norwegian National Forest Inventory. European Journal of Forest
Research, 131:1255–1267.
Breidt, F. J. and Fuller, W. A. (1999). Design of supplemented panel surveys
with application to the National Resources Inventory. Journal of Agricultural,
Biological, and Environmental Statistics, 4(4):391–403.
Breidt, F. J. and Opsomer, J. D. (2017). Model-assisted survey estimation
with modern prediction techniques. Statistical Science, 32(2):190–205.
Brown, L. D., Cai, T. T., and DasGupta, A. (2001). Interval estimation
for a binomial proportion - Comment - Rejoinder. Statistical Science,
16(2):101–133.
Brus, D. J. (2019). Sampling for digital soil mapping: A tutorial supported by
R scripts. Geoderma, 338:464–480.
Brus, D. J. (2021). Statistical approaches for spatial sample survey: Persistent
misconceptions and new developments. European Journal of Soil Science,
72(2):686–703.
Bibliography 517

Brus, D. J. and de Gruijter, J. J. (1994). Estimation of non-ergodic vari-

ograms and their sampling variance by design-based sampling strategies.
Mathematical Geology, 26:437–454.
Brus, D. J. and de Gruijter, J. J. (1997). Random sampling or geostatisti-
cal modelling? Choosing between design-based and model-based sampling
strategies for soil (with Discussion). Geoderma, 80(1–2):1–59.
Brus, D. J. and de Gruijter, J. J. (2003). A method to combine non-probability
sample data with probability sample data in estimating spatial means
of environmental variables. Environmental Monitoring and Assessment,
83(3):303–317.
Brus, D. J. and de Gruijter, J. J. (2011). Design-based Generalized Least
Squares estimation of status and trend of soil properties from monitoring
data. Geoderma, 164(3–4):172–180.
Brus, D. J. and de Gruijter, J. J. (2013). Effects of spatial pattern persistence on
the performance of sampling designs for regional trend monitoring analyzed
by simulation of space-time fields. Computers & Geosciences, 61:175–183.
Brus, D. J., de Gruijter, J. J., and van Groenigen, J. W. (2007). Designing
Spatial Coverage Samples Using the k-means Clustering Algorithm. In La-
gacherie, P., McBratney, A. B., and Voltz, M., editors, Digital Soil Mapping.
An Introductory Perspective, pages 183–192. Elsevier.
Brus, D. J. and Heuvelink, G. B. M. (2007). Optimization of sample patterns
for universal kriging of environmental variables. Geoderma, 138(1–2):86–95.
Brus, D. J., Kempen, B., and Heuvelink, G. B. M. (2011). Sampling for valida-
tion of digital soil maps. European Journal of Soil Science, 62(3):394–407.
Brus, D. J. and Knotters, M. (2008). Sampling design for compliance monitoring
of surface water quality: A case study in a Polder area. Water Resources
Research, 44:W11410.
Brus, D. J. and Saby, N. P. A. (2016). Approximating the variance of estimated
means for systematic random sampling, illustrated with data of the French
Soil Monitoring Network. Geoderma, 279:77–86.
Brus, D. J., Slim, P. A., Heidema, A. H., and van Dobben, H. F. (2014). Trend
monitoring of the areal extent of habitats in a subsiding coastal area by
spatial probability sampling. Ecological Indicators, 45:313–319.
Brus, D. J., Spätjens, L. E. E. M., and de Gruijter, J. J. (1999). A sampling
scheme for estimating the mean extractable phosphorus concentration of
fields for environmental regulation. Geoderma, 89(1–2):129–148.
Cassel, C. M., Särndal, C. E., and Wretman, J. H. (1977). Foundations of
Inference in Survey Sampling. Wiley, New York.
Champely, S. (2020). pwr: Basic Functions for Power Analysis. R package
version 1.3-0.
518 Bibliography

Chaudhuri, A. (1994). Small domain statistics: a review. Statistica Neerlandica,

48(3):215–236.
Christensen, R. (1991). Linear Models for Multivariate, Time Series, and
Spatial Data. Springer, New York.
Clifford, P., Richardson, S., and Hémon, D. (1989). Assessing the significance
of the correlation between two spatial processes. Biometrics, 45:123–134.
Cochran, W. G. (1977). Sampling Techniques. Wiley, New York.
Corwin, D. L. and Lesch, S. M. (2005). Characterizing soil spatial variability
with apparent soil electrical conductivity: Part II. Case study. Computers
and Electronics in Agriculture, 46(1–3):135–152.
Corwin, D. L., Lesch, S. M., Segal, E., Skaggs, T. H., and Bradford, S. A. (2010).
Comparison of sampling strategies for characterizing spatial variability with
apparent soil electrical conductivity directed soil sampling. Journal of
Environmental and Engineering Geophysics, 15(3):147–162.
Csárdi, G., Hester, J., Wickham, H., Chang, W., Morgan, M., and Tenenbaum,
D. (2021). remotes: R Package Installation from Remote Repositories,
Including ’GitHub’. R package version 2.4.2.
Cullman, A. D. (2021). maSAE: Mandallaz’ Model-Assisted Small Area Esti-
mators. R package version 2.0.3.
Dalenius, T. and Hodges, J. L. (1959). Minimum variance stratification. Journal
of the American Statistical Association, 54(285):88–101.
de Gruijter, J. J., Brus, D. J., Bierkens, M. F. P., and Knotters, M. (2006).
Sampling for Natural Resource Monitoring. Springer, Berlin.
de Gruijter, J. J., Minasny, B., and McBratney, A. B. (2015). Optimizing
stratification and allocation for design-based estimation of spatial means
using predictions with error. Journal of Survey Statistics and Methodology,
3(1):19–42.
de Gruijter, J. J. and ter Braak, C. J. F. (1990). Model-free estimation from
spatial samples: a reappraisal of classical sampling theory. Mathematical
Geology, 22(4):407–415.
de Gruijter, J. J. and ter Braak, C. J. F. (1992). Design-based versus model-
based sampling strategies: Comment on R.J. Barnes’ ‘Bounding the re-
quired sample size for geologic site characterization’. Mathematical Geology,
24(7):859–864.
De Vries, P. G. (1986). Sampling Theory for Forest Inventory: A Teach-Yourself
Course. Springer-Verlag, New York.
Deville, J. C. and Tillé, Y. (1998). Unequal probability sampling without
replacement through a splitting method. Biometrika, 85(1):89–101.
Bibliography 519

Deville, J. C. and Tillé, Y. (2004). Efficient balanced sampling: The cube

method. Biometrika, 91(4):893–912.
Deville, J. C. and Tillé, Y. (2005). Variance approximation under balanced
sampling. Journal of Statistical Planning and Inference, 128(2):569–591.
Domburg, P., de Gruijter, J. J., and Brus, D. J. (1994). A structured approach
to designing soil survey schemes with prediction of sampling error from
variograms. Geoderma, 62:151–164.
Dumelle, M., Kincaid, T. M., Olsen, A. R., and Weber, M. H. (2021). spsurvey:
Spatial Sampling Design and Analysis. R package version 5.0.1.
Escobar, E. L., Zamudio, E. B., and Munoz Rosas, J. F. (2019). samplingVarEst:
Sampling Variance Estimation. R package version 1.4.
Falorsi, P. D. and Righi, P. (2008). A balanced sampling approach for multi-
way stratification designs for small area estimation. Survey Methodology,
34(2):223–234.
Fitzgerald, G. J. (2010). Response surface sampling of remotely sensed imagery
for precision agriculture. In Viscarra Rossel, R. A., McBratney, A. B.,
and Minasny, B., editors, Proximal Soil Sensing, pages 121–129. Springer,
Dordrecht.
Fitzgerald, G. J., Lesch, S. M., Barnes, E. M., and Luckett, W. E. (2006).
Directed sampling using remote sensing with a response surface sampling
design for site-specific agriculture. Computers and Electronics in Agriculture,
53(2):98–112.
Fuller, W. A. (1999). Environmental surveys over time. Journal of Agricultural,
Biological, and Environmental Statistics, 4(4):331–345.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin,
D. B. (2013). Bayesian Data Analysis. Chapman & Hall/CRC, Boca Raton,
third edition.
Gershenfeld, N. (1999). The Nature of Mathematical Modeling. Cambridge
University Press, Cambridge.
Goerg, G. M. (2013). LICORS: Light Cone Reconstruction of States - Predictive
State Estimation From Spatio-Temporal Data. R package version 0.2.0.
Goovaerts, P. (1997). Geostatistics for Natural Resources Evaluation. Oxford
University Press, New York.
Grafström, A. (2012). Spatially correlated Poisson sampling. Journal of
Statistical Planning and Inference, 142(1):139–147.
Grafström, A., Ekström, M., Jonsson, B. G., Esseen, P.-A., and Ståhl, G.
(2019). On combining independent probability samples. Survey Methodology,
45(2):349–364.
520 Bibliography

Grafström, A. and Lisic, J. (2019). BalancedSampling: Balanced and Spatially

Balanced Sampling. R package version 1.5.5.
Grafström, A., Lundström, N. L. P., and Schelin, L. (2012). Spatially balanced
sampling through the pivotal method. Biometrics, 68(2):514–520.
Grafström, A. and Schelin, L. (2014). How to select representative samples.
Scandinavian Journal of Statistics, 41(2):277–290.
Grafström, A. and Tillé, Y. (2013). Doubly balanced spatial sampling with
spreading and restitution of auxiliary totals. Environmetrics, 24(2):120–131.
Griffith, D. A. (2005). Effective geographic sample size in the presence of
spatial autocorrelation. Annals of the Association of American Geographers,
95(4):740–760.
Gurney, M. and Daly, J. F. (1965). A multivariate approach to estimation
in periodic sample surveys. In Proceedings of the American Statistical
Association, Section on Social Statistics, pages 242–257.
Hankin, D. G., Mohr, M. S., and Newman, K. B. (2019). Sampling Theory:
For the Ecological and Natural Resource Sciences. Oxford University Press,
United Kingdom.
Hansen, M. H., Madow, W. G., and Tepping, B. J. (1983). An evaluation
of model-dependent and probability sampling inferences in sample surveys.
Journal of the American Statistical Association, 78(384):776–793.
Hartig, F., Minunno, F., and Paul, S. (2019). BayesianTools: General-Purpose
MCMC and SMC Samplers and Tools for Bayesian Statistics. R package
version 0.1.7.
Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical
Learning. Data Mining, Inference, and Prediction. Springer, New York,
second edition.
Heuvelink, G. B. M., Brus, D. J., and de Gruijter, J. J. (2007). Optimization
of sample configurations for digital mapping of soil properties with universal
kriging. In Lagacherie, P., McBratney, A. B., and Voltz, M., editors, Digital
Soil Mapping. An Introductory Perspective, pages 137–151. Elsevier.
Hill, A., Massey, A., and Mandallaz, D. (2021). The R package forestinven-
tory: Design-based global and small area estimations for multiphase forest
inventories. Journal of Statistical Software, 97(4):1–40.
Hofman, S. C. K. and Brus, D. J. (2021). How many sampling points are
needed to estimate the mean nitrate-N content of agricultural fields? A
geostatistical simulation approach with uncertain variograms. Geoderma,
385:114816.
Höhle, M. (2017). binomSamSize: Confidence Intervals and Sample Size
Determination for a Binomial Proportion under Simple Random Sampling
and Pooled Sampling. R package version 0.1-5.
Bibliography 521

Huang, Z. (1998). Extensions to the k-means algorithm for clustering large

data sets with categorical values. Data Mining and Knowledge Discovery,
2:283–304.
Hutson, A. D. (1999). Calculating nonparametric confidence intervals for
quantiles using fractional order statistics. Journal of Applied Statistics,
26(3):343–353.
Hutson, N., Hutson, A., and Yan, L. (2019). QuantileNPCI: Nonparametric
Confidence Intervals for Quantiles. R package version 0.9.0.
Hyndman, R. J. and Fan, Y. (1996). Sample quantiles in statistical packages.
American Statistician, 50(4):361–365.
Isaaks, E. H. and Srivastava, R. M. (1989). An Introduction to Applied
Geostatistics. Oxford University Press, New York.
Janssen, P. H. M. and Heuberger, P. S. C. (1995). Calibration of process-
oriented models. Ecological Modelling, 83(1–2):55–66.
Jessen, R. J. (1942). Statistical investigation of a sample survey for obtaining
farm facts. Research Bulletin (Iowa Agriculture and Home Economics
Experiment Station), 26(304):54–59.
Joseph, L. and Bélisle, P. (1997). Bayesian sample size determination for
normal means and differences between normal means. Journal of the Royal
Statistical Society. Series D (The Statistician), 46(2):209–226.
Joseph, L. and Bélisle, P. (2012). SampleSizeMeans: Sample size calculations
for normal means. R package version 1.1.
Joseph, L., Wolfson, D. B., and du Berger, R. (1995). Sample size calculations
for binomial proportions via highest posterior density intervals. Journal of
the Royal Statistical Society. Series D (The Statistician), 44(2):143–154.
Joseph, L., Wolfson, D. B., du Berger, R., and Bélisle, P. (2018). SampleSizeBi-
nomial: Bayesian sample size determination for a single binomial proportion.
R package version 1.1.
Karger, D. N., Conrad, O., Böhner, J., Kawohl, T., Kreft, H., Soria-Auza, R. W.,
Zimmermann, N. E., Linder, H. P., and Kessler, M. (2017). Climatologies at
high resolution for the earth’s land surface areas. Scientific Data, 4:170122.
Kim, J. and Wang, Z. (2019). Sampling techniques for big data analysis.
International Statistical Review, 87(S1):S177–S191.
Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. (1983). Optimization by
simulated annealing. Science, 220(4598):671–680.
Kitanidis, P. K. (1983). Statistical estimation of polynomial generalized
covariance functions and hydrologic applications. Water Resources Research,
19(4):909–921.
522 Bibliography

Kitanidis, P. K. (1987). Parametric estimation of covariances of regionalized

variables. Water Resources Bulletin, 23(4):557–567.
Knotters, M. and Brus, D. J. (2010). Estimating space-time mean concen-
trations of nutrients in surface waters of variable depth. Water Resources
Research, 46:W08502.
Lark, R. M. (2000). Estimating variograms of soil properties by the method-
of-moments and maximum likelihood. European Journal of Soil Science,
51(4):717–728.
Lark, R. M. (2002). Optimized spatial sampling of soil for estimation of the
variogram by maximum likelihood. Geoderma, 105(1):49–80.
Lark, R. M. (2011). Spatially nested sampling schemes for spatial vari-
ance components: Scope for their optimization. Computers & Geosciences,
37(10):1633–1641.
Lark, R. M., Hamilton, E. M., Kaninga, B., Maseka, K. K., Mutondo, M.,
Sakala, G. M., and Watts, M. J. (2017). Planning spatial sampling of the
soil from an uncertain reconnaissance variogram. SOIL, 3(4):235–244.
Lark, R. M. and Marchant, B. P. (2018). How should a spatial-coverage sample
design for a geostatistical soil survey be supplemented to support estimation
of spatial covariance parameters? Geoderma, 319:89–99.
Lark, R. M. and Webster, R. (2006). Geostatistical mapping of geomorphic
variables in the presence of trend. Earth Surface Processes and Landforms,
31(7):862–874.
Lenth, R. V. (2009). Response-Surface Methods in R, Using rsm. Journal of
Statistical Software, 32(7):1–17.
Lesch, S. M. (2005). Sensor-directed response surface sampling designs for
characterizing spatial variation in soil properties. Computers and Electronics
in Agriculture, 46(1–3):153–179.
Lesch, S. M., Strauss, D. J., and Rhoades, J. D. (1995). Spatial prediction
of soil salinity using electromagnetic induction techniques, 2. An efficient
spatial sampling algorithm suitable for multiple linear regression model
identification and estimation. Water Resources Research, 31(2):387–398.
Lisic, J. and Grafström, A. (2018). SamplingBigData: Sampling Methods for
Big Data. R package version 1.0.0.
Lohr, S. L. (1999). Sampling: Design and Analysis. Duxbury Press, Pacific
Grove.
Lumley, T. (2020). leaps: Regression Subset Selection. R package version 3.1.
Lumley, T. (2021). survey: Analysis of Complex Survey Samples. R package
version 4.1-1.
Bibliography 523

Ly, A., Marsman, M., Verhagen, J., Grasman, R. P. P. P., and Wagenmakers,
E. J. (2017). A tutorial on Fisher information. Journal of Mathematical
Psychology, 80:40–55.
Ma, T., Brus, D. J., Zhu, A.-X., Zhang, L., and Scholten, T. (2020). Compar-
ison of conditioned Latin hypercube and feature space coverage sampling
for predicting soil classes using simulation from soil maps. Geoderma,
370:114366.
Mandallaz, D. (2007). Sampling Techniques for Forest Inventories. Chapman
& Hall/CRC, Boca Raton.
Mandallaz, D., Breschan, J., and Hill, A. (2013). New regression estimators in
forest inventories with two-phase sampling and partially exhaustive informa-
tion: A design-based Monte Carlo approach with applications to small-area
estimation. Canadian Journal of Forest Research, 43(11):1023–1031.
Marcelli, A., Corona, P., and Fattorini, L. (2019). Design-based estimation of
mark variograms in forest ecosystem surveys. Spatial Statistics, 30:27–38.
Marchant, B. P. and Lark, R. M. (2007). Optimized sample schemes for
geostatistical surveys. Mathematical Geology, 39(1):113–134.
Matérn, B. (1947). Methods of estimating the accuracy of line and sample
plot surveys. Meddelanden från Statens Skogsforskningsinstitut, 36(1).
Matérn, B. (1986). Spatial Variation, volume 36 of Lecture Notes in Statistics.
Springer-Verlag, Berlin, second edition.
McConville, K., Tang, B., Zhu, G., Li, S., Chueng, S., and Toth, D. (2021).
mase: Model-Assisted Survey Estimation. R package version 0.1.3.
McConville, K. S., Moisen, G. G., and Frescino, T. S. (2020). A tutorial
on model-assisted estimation with application to forest inventory. Forests,
11(2):244.
McKay, M. D., Beckman, R. J., and Conover, W. J. (1979). A comparison
of three methods for selecting values of input variables in the analysis of
output from a computer code. Technometrics, 21(2):239–245.
McLaren, C. H. and Steel, D. G. (2001). Rotation patterns and trend estimation
for repeated surveys using rotation group estimates. Statistica Neerlandica,
55(2):221–238.
Minasny, B. and McBratney, A. B. (2006). A conditioned Latin hypercube
method for sampling in the presence of ancillary information. Computers &
Geosciences, 32(9):1378–1388.
Minasny, B. and McBratney, A. B. (2010). Conditioned Latin hypercube
sampling for calibrating soil sensor data to soil properties. In Viscarra
Rossel, R. A., McBratney, A. B., and Minasny, B., editors, Proximal Soil
Sensing, pages 111–119. Springer, Dordrecht.
524 Bibliography

Müller, W. G. and Zimmerman, D. L. (1999). Optimal designs for variogram

estimation. Environmetrics, 10(1):23–37.
Myers, R. H., Montgomery, D. C., and Anderson-Cook, C. M. (2009). Response
Surface Methodology: Process and Product Optimization Using Designed
Experiments. John Wiley & Sons, Hoboken, third edition.
Müller, K. and Wickham, H. (2021). tibble: Simple Data Frames. R package
version 3.1.6.
Nanthakumar, A. and Selvavel, K. (2004). Estimation of proportion of success
from a stratified population: a comparative study. Communications in
Statistics, 33(9):2245–2257.
Nychka, D., Furrer, R., Paige, J., Sain, S., Gerber, F., and Iverson, M. (2021).
fields: Tools for Spatial Data. University Corporation for Atmospheric
Research, Boulder, CO, USA. R package version 13.3.
Oliver, M. A. and Webster, R. (1986). Combining nested and linear sampling for
determining the scale and form of spatial variation of regionalized variables.
Geographical Analysis, 18(3):227–242.
Ott, R. L. and Longnecker, M. (2015). An Introduction to Statistical Methods
and Data Analysis. Cengage Learning, Boston, seventh edition.
Papritz, A. J., Dümig, A., Zimmermann, C., Gerke, H. H., Felderer, B., Kögel-
Knabner, I., Schaaf, W., and Schulin, R. (2011). Uncertainty of variance
component estimates in nested sampling: a case study on the field-scale
spatial variability of a restored soil. European Journal of Soil Science,
62(3):479–495.
Patterson, H. D. (1950). Sampling on successive occasions with partial re-
placement of units. Journal of the Royal Statistical Society. Series B
(Methodological), 12(2):241–255.
Pebesma, E. (2018). Simple features for R: Standardized support for spatial
vector data. The R Journal, 10(1):439–446.
Pebesma, E. J. (2004). Multivariable geostatistics in S: the gstat package.
Computers & Geosciences, 30(7):683–691.
Pebesma, E. J. and Bivand, R. S. (2005). Classes and methods for spatial data
in R. R News, 5(2):9–13.
Pinheiro, J., Bates, D., DebRoy, S., Sarkar, D., and R Core Team (2021). nlme:
Linear and Nonlinear Mixed Effects Models. R package version 3.1-152.
Plant, R. E. (2012). Spatial Data Analysis in Ecology and Agriculture Using R.
CRC Press, Boca Raton.
R Core Team (2021). R: A language and environment for statistical computing.
R Foundation for Statistical Computing, Vienna, Austria.
Bibliography 525

Rao, J. N. K. (2003). Small Area Estimation. John Wiley & Sons, Hoboken.
Ribeiro Jr, P. J., Diggle, P. J., Schlather, M., Bivand, R., and Ripley, B. (2020).
geoR: Analysis of Geostatistical Data. R package version 1.8-1.
Ripley, B. D. (1981). Spatial Statistics. John Wiley & Sons, New York.
Robertson, B. L., Brown, J. A., McDonald, T., and Jaksons, P. (2013).
BAS: Balanced Acceptance Sampling of Natural Resources. Biometrics,
69(3):776–784.
Rosén, B. (1997). On sampling with probability proportional to size. Journal
of Statistical Planning and Inference, 62(2):159–191.
Roudier, P. (2021). clhs: a R package for conditioned Latin hypercube sampling.
R package version 0.9.0.
Roudier, P., Hewitt, A. E., and Beaudette, D. E. (2012). A conditioned Latin
hypercube sampling algorithm incorporating operational constraints. In
Minasny, B., Malone, B. P., and McBratney, A. B., editors, Digital Soil
Assessments and Beyond. Proceedings of the 5th Global Workshop on Digital
Soil Mapping, pages 227–232. CRC Press/Balkema, Leiden.
Samuel-Rosa, A. (2019). spsann: Optimization of Sample Configurations using
Spatial Simulated Annealing. R package version 2.2.0.
Särndal, C. E., Swensson, B., and Wretman, J. (1992). Model Assisted Survey
Sampling. Springer, New York.
Schloerke, B., Cook, D., Larmarange, J., Briatte, F., Marbach, M., Thoen,
E., Elberg, A., and Crowley, J. (2021). GGally: Extension to ’ggplot2’. R
package version 2.1.2.
Schoch, T. (2014). rsae: Robust Small Area Estimation. R package version
0.1-5.
Signorell, A. (2021). DescTools: Tools for Descriptive Statistics. R package
version 0.99.43.
Stehman, S. V. (1999). Basic probability sampling designs for thematic
map accuracy assessment. International Journal of Remote Sensing,
20(12):2423–2441.
Stehman, S. V., Fonte, C. C., Foody, G. M., and See, L. (2018). Using volun-
teered geographic information (VGI) in design-based statistical inference for
area estimation and accuracy assessment of land cover. Remote Sensing of
Environment, 212:47–59.
Stevens, D. L. and Olson, A. R. (2004). Spatially balanced sampling of natural
resources. Journal of the American Statistical Association, 99(465):262–278.
Szepannek, G. (2018). clustMixType: User-friendly clustering of mixed-type
data in R. The R Journal, 10(2):200–208.
526 Bibliography

ter Braak, C. J. F. and Vrugt, J. A. (2008). Differential Evolution Markov

Chain with snooker updater and fewer chains. Statistics and Computing,
18(4):435–446.
Tillé, Y. (2006). Sampling Algorithms. Springer, New York.
Tillé, Y. and Matei, A. (2021). sampling: Survey Sampling. R package version
2.9.
Toth, D. (2021). rpms: Recursive Partitioning for Modeling Survey Data. R
package version 0.5.1.
Valliant, R., Dever, J. A., and Kreuter, F. (2018). Practical Tools for Designing
and Weighting Survey Samples. Springer, New York, second edition.
Valliant, R., Dever, J. A., and Kreuter, F. (2021). PracTools: Tools for
Designing and Weighting Survey Samples. R package version 1.2.3.
van Groenigen, J. W., Pieters, G., and Stein, A. (2000). Optimizing spatial
sampling for multivariate contamination in urban areas. Environmetrics,
11(2):227–244.
van Groenigen, J. W., Siderius, W., and Stein, A. (1999). Constrained optimi-
sation of soil sampling for minimisation of the kriging variance. Geoderma,
87(3):239–259.
van Groenigen, J. W. and Stein, A. (1998). Constrained optimization of spatial
sampling using continuous simulated annealing. Journal of Environmental
Quality, 27(5):1078–1086.
Venables, W. N. and Ripley, B. D. (2002). Modern Applied Statistics with S.
Springer, New York, fourth edition.
Wadoux, A. M. J.-C., Brus, D. J., and Heuvelink, G. B. M. (2019). Sam-
pling design optimization for soil mapping with random forest. Geoderma,
355:113913.
Walvoort, D., Brus, D., and de Gruijter, J. (2020). Spatial Coverage Sampling
and Random Sampling from Compact Geographical Strata. R package version
0.3-9.
Walvoort, D. J. J., Brus, D. J., and de Gruijter, J. J. (2010). An R package for
spatial coverage sampling and random sampling from compact geographical
strata by k-means. Computers & Geosciences, 36(10):1261–1267.
Wang, J., Haining, R., and Cao, Z. (2010). Sample surveying to estimate the
mean of a heterogeneous surface: reducing the error variance through zoning.
International Journal of Geographical Information Science, 24(4):523–543.
Wang, J., Stein, A., Gao, B., and Ge, Y. (2012). A review of spatial sampling.
Spatial Statistics, 2(1):1–14.
Bibliography 527

Wang, Z. and Zhu, Z. (2019). Spatiotemporal balanced sampling design for

longitudinal area surveys. Journal of Agricultural, Biological, and Environ-
mental Statistics, 24(2):245–263.
Ware, K. D. W. and Cunia, T. (1962). Continuous forest inventory with partial
replacement of samples. Forest Science Monographs, 3:1–40.
Webster, R. and Oliver, M. A. (1992). Sample adequately to estimate vari-
ograms of soil properties. Journal of Soil Science, 43(1):177–192.
Webster, R. and Oliver, M. A. (2007). Geostatistics for Environmental Scien-
tists. Wiley, Chichester, second edition.
Webster, R., Welham, S. J., Potts, J. M., and Oliver, M. A. (2006). Estimating
the spatial scales of regionalized variables by nested sampling, hierarchi-
cal analysis of variance and residual maximum likelihood. Computers &
Geosciences, 32(9):1320–1333.
Wickham, H. (2019). stringr: Simple, Consistent Wrappers for Common String
Operations. R package version 1.4.0.
Wickham, H. (2021). forcats: Tools for Working with Categorical Variables
(Factors). R package version 0.5.1.
Wickham, H., Averick, M., Bryan, J., Chang, W., McGowan, L. D., François,
R., Grolemund, G., Hayes, A., Henry, L., Hester, J., Kuhn, M., Pedersen,
T. L., Miller, E., Bache, S. M., Müller, K., Ooms, J., Robinson, D., Seidel,
D. P., Spinu, V., Takahashi, K., Vaughan, D., Wilke, C., Woo, K., and
Yutani, H. (2019). Welcome to the tidyverse. Journal of Open Source
Software, 4(43):1686.
Wickham, H., François, R., Henry, L., and Müller, K. (2021). dplyr: A Grammar
of Data Manipulation. R package version 1.0.7.
Wickham, H. and Grolemund, G. (2017). R for Data Science. Import, Tidy,
Transform, Visualize, and Model Data. O’Reilly Media, Sebastopol.
Woodruff, R. S. (1963). The use of rotating samples in the Census Bu-
reau’s monthly surveys. Journal of the American Statistical Association,
58(302):454–467.
Wright, M. N. and Ziegler, A. (2017). ranger: A Fast Implementation of
Random Forests for High Dimensional Data in C++ and R. Journal of
Statistical Software, 77(1):1–17.
Wu, C. (2003). Optimal calibration estimators in survey sampling. Biometrika,
90(4):937–951.
Wu, C. and Sitter, R. R. (2001). A model-calibration approach to using
complete auxiliary information from survey data. Journal of the American
Statistical Association, 96(453):185–193.
528 Bibliography

Xie, Y., Dervieux, C., and Riederer, E. (2020). R Markdown Cookbook. Chap-
man & Hall/CRC, Boca Raton, first edition.
Zhao, X. and Grafström, A. (2020). A sample coordination method to monitor
totals of environmental variables. Environmetrics, 31(6).
Zhu, Z. and Stein, M. L. (2005). Spatial sampling design for parameter
estimation of the covariance function. Journal of Statistical Planning and
Inference, 134(2):583–603.
Zhu, Z. and Stein, M. L. (2006). Spatial sampling design for prediction with
estimated parameters. Journal of Agricultural, Biological, and Environmental
Statistics, 11(1):24–44.
Zhu, Z. and Zhang, H. (2006). Spatial sampling under the infill asymptotic
framework. Environmetrics, 17:323–337.
Zimmerman, D. L. (2006). Optimal network design for spatial prediction,
covariance parameter estimation, and empirical prediction. Environmetrics,
17(6):635–652.
Index

𝜋 estimator, 22 to prediction of sampling

𝜋-expanded value, 22 variance, 245
𝜋∗ estimator, 202 to sample size determination, 219
Cum-root-f stratification, 64 Best linear unbiased estimator, 483
kth order statistic, 35 Best linear unbiased predictor, 377,
p-value of a test, 246, 483 481
two-sided p-value, 470, 471 Beta distribution, 226
t-test Bias, 33
one-sample t-test, 470 design-bias, 480
paired t-test, 470 model-bias, 480
relative bias, 197
A priori variance, 378, 379 selection bias, 481
Accuracy, 33 Big data, 199
Adapted experimental design, 321 Binomial distribution, 43
Allocation Binomial proportion, 43
Bethel allocation, 62, 73 Binomial test, 216
Neyman allocation, 61 Bivariate normal-gamma distribution,
optimal allocation, 61 223
proportional allocation, 60 Block-kriging variance, 382
ANCOVA model, 176 Bootstrap sample, 195, 439
ANOVA model Brewer’s variance estimator, 102, 118,
hierarchical ANOVA model, 434 133
one-way ANOVA model, 187 Burn-in period, 248
Anticipated variance, 233
Arbitrary sampling, 7, 21 Calibrated weights, 169, 181
Areal sampling unit, 2, 124 Categorical map, 462
Augmented kriging variance, 450 Census, 1
Autocorrelation, 482 Central composite design, 357
Average coverage criterion, 221 Central composite response surface
Average length criterion, 221 design, 359
Central limit theorem, 41
Back-up list of sampling units, 29
Certainty stratum approach, 199
Balanced acceptance sampling, 152
Chain length, 414
Balanced sampling, 137
Change of spatial mean (total), 292
Bayes’ rule, 220
Change of support, 381
Bayesian approach
Circular sampling plot, 44
to grid spacing determination,
fixed, 44
407

529
530 Index

floating, 45 Design weight, 22

Class representation, 463 Design-based approach, 7, 473
Classical sampling theory, 42, 473 Design-based generalised least
Clopper-Pearson method, 43 squares estimator, 294
Cluster random sampling, 93 design-variance, 480
Cluster size, 94 Determinant of variance-covariance
Coefficient of variation, 73 matrix, 318, 442
Cohen’s ℎ, 217 Differential evolution algorithm, 248
Collapsed strata, 72, 468 Direct estimator, 264
Compact geographical stratum, 69, Discretisation error, 5
236 Distance parameter, 258, 443
Composite sampling, 72, 238 Domain of interest, 263
Conditioned Latin hypercube Double sampling, 201
sampling, 345 Doubly balanced sampling, 158
Confidence interval, 40 Draw-by-draw selection probability
Confidence level, 41 of a cluster, 97
Conjugate prior, 226 of a primary sampling unit, 113
Conservative variance estimator, 154
Consistent estimator, 57 Effective range, 363, 416
Convenience sample, 240 Effective sample size, 482
Correlation function, 376 Elementary estimate, 293
Correlogram, 376, 380 Elementary sampling unit, 2
Covariance function, 376 Empirical coverage rate, 489
Covariate space coverage sampling, Estimation adjusted criterion, 451
335 Estimator, 22
Covariate space infill sampling, 338 Evaluation point, 419
Covariogram, 376 Exhaustive data set, 9
Coverage probability, 221 Expectation of estimator, 32
Credible interval, 220 Expected sample size, 79, 100
Cross-classification stratum, 73, 148 Experimental semivariogram, see
Cross-validation, 460 Sample semivariogram
leave-one-out cross-validation,
460 Factor level, 345
Cube algorithm for balanced Finite population correction, 37
sampling, 140 First order Taylor series, 445
Cube point, 360 First-phase sample, 202
Cumulative distribution function, 5 Fisher information matrix, 442
Fixed cluster centre, 331
Data snooping, 190 Fixed effect, 276
Data splitting, 460 Full factorial design, 322, 345
Degrees of freedom, 41
Gamma distribution, 224
Design effect, 57, 99, 115, 218
Generalised difference estimator, 162
Design matrix, 145
Design parameter, 212, 219 Generalised least squares, 392
Design point, 361 Generalised random-tessellation
stratified sampling, 154
Index 531

Genetic algorithm, 252 Likelihood function, 220

Geographical stratification, 69 Linear costs model, 61
Geometric sampling design, 321 Linear mixed model, 274
Geostatistical model, 240, 375 Linear regression model
Geostatistical simulation, 239, 280 heteroscedastic, 162
conditional, 241 homoscedastic, 164
unconditional, 241 Local pivotal method, 152
Geostratum, see Compact Local semivariogram, 436
geographical stratum Log-likelihood surface, 442
Grid spacing, 79, 325
Machine learning technique, 190, 340
Hájek estimator, 132 Map quality evaluation, 459
Hájek-Rosén variance estimator, 133 Map quality index, 353, 461
Hansen-Hurwitz estimator, see pwr Margin of error, 41, 213
estimator Marginal stratum, 148, 348
Haphazard sampling, 21 Markov chain Monte Carlo sampling,
Hartley-Rao’s variance estimator, 103, 247, 408
133 Matérn’s variance approximation
Highest posterior density interval, 221 method, 88
Horvitz-Thompson estimator, see 𝜋 Maximum likelihood estimation, 246,
estimator 388, 442
Mean semivariance
Identically and independently within a prediction block, 382
distributed, 476 within a sample, 233
Inclusion probability density, 22 within a study area, 233
Inclusion probability of a unit, 19 Mean squared error, 33
Independence Mean squared shortest distance, 327
design-independence, 480 Measurement error, 238
model-independence, 480 Median kriging variance, 401
Isotropy, 379 Method-of-moments, 388
Iterative reallocation, 252 iterative method-of-moments,
394
k-means clustering, 68, 85 Mixed Bayesian-likelihood approach,
k-means++ algorithm, 338 223
Kriging Model efficiency coefficient, 353, 462
block-kriging, 382, 403 Model regression coefficient, 163
kriging with an external drift, Model-assisted approach, 25, 161, 267,
383, 420, 465 488
ordinary kriging, 376 Model-based approach, 7, 473
point-kriging, 382 Model-based sampling, 322
simple kriging, 386 Model-calibration estimator, 191
Kriging weight, 377 Model-dependent predictor, 282, 493
Model-mean, 161, 474
Lagrange multiplier, 377
Model-variance, 480
Latin hypercube sample, 323, 345
Multiple linear regression model, 165
Likelihood, 390, 408
532 Index

Multivariate normal distribution, 390 Population quantile, 6, 35

Multivariate stratification, 73 Population regression coefficient, 141,
Multiway stratification, 148 163
Population root mean squared error,
Nested sampling, 430 353
balanced nested sampling, 432 Population total, 5, 22
unbalanced nested sampling, 435 Population unit, 2
Non-ergodic semivariogram, 436 Population variance, 33, 37, 462
Non-probability sampling, 7, 199 Posterior distribution, 220
Nugget, 258, 379 Poststratified estimator, 187
Nugget-to-sill ratio, 416 Power of a test, 215
pps sampling
Objective function, 419 with replacement (ppswr), 97,
Optimal sample size in two-stage 109, 125
cluster random sampling, without replacement (ppswor),
114 102, 118, 128
Optimal spatial stratification, 252 Precision, 33
Ordered sample, 24 Precision parameter, 223
Ordinary kriging variance, 377, 380 Predictive distribution, 221
Ordinary least squares, 163, 372 Preferential sample, 480
Out-of-bag predictions, 196 Preposterior distribution, 221
Overall purity, 462 Primary sampling unit, 109
Pairwise inclusion probability, 133 Principal component score, 358
Partial sill, 258, 379 Prior belief, 407
Prior distribution, 220
Penalised least squares, 172
Prior estimate
Perturbation, 443
Pivotal method for ppswor sampling, of population coefficient of
130 variation, 212
Point support, 381 of population proportion, 212
Point-kriging variance, 382 of population standard deviation,
212
Population
Probability of a type I error, 215
continuous population, 2
discrete population, 2 Probability of a type II error, 215
finite population, 22 Probability sampling, 7, 474
infinite population, 22 Pure panel, 289
Population coefficient of pwr estimator, 23
determination, 462 Q-Q plot, 246, 255
Population fit of model parameters, Quantitative map, 461
162 Quartile, 35
Population mean, 5, 22
Population mean absolute error, 461 Random effect, 277
Population mean error, 353, 461 Random error, 33
Population mean squared error, 461 Random forest, 192, 340, 465
Population parameter, 5, 8 Random intercept model, 274
Population proportion, 5, 34
Index 533

Random number generator, 474 Sampling design, 19

Random process, 474 Sampling distribution, 9, 32
Range, 379 Sampling error, 9
Ratio estimator, 31, 84, 104, 182, 264 Sampling experiment, 38
combined ratio estimator, 185 Sampling fraction, 125
separate ratio estimator, 185 Sampling frame, 3
Ratio of spatial dependence, 246, 408, Sampling intensity, 22
443 Sampling pattern, 423
Reconnaissance sample survey, 429 Sampling strategy, 33
Rectangular grid, 80 Sampling variance, 33
Regression estimator, 140 Sampling with replacement, 23
combined regression estimator, Sattherwaite’s method, 60
178 Second-phase sample, 202
elastic net regression estimator, Secondary sampling unit, 109
173 Selection probability of a sample, 19
for small domains, 267 Self-weighting sampling design, 112
for two-phase sampling, 207 Semivariance, 379
generalised regression estimator, Semivariogram, 378
162, 163 Semivariogram model, 235
multiple regression estimator, exponential, 379
165 spherical, 379
separate regression estimator, Separation distance, 388
174 Sequential Gaussian simulation, 241
simple regression estimator, 164 Set-sample, 24
Regression tree, 192 Shapiro-Wilk test, 246
Regular grid, 79, 325 Significance level, 218, 246
Relative efficiency, 197 Sill, 379, 487
Repeated surveys for monitoring, 289 Simple linear regression model, 164
Required sample size, 211 Simple random sampling, 27
Residual semivariogram, 421 Simple random sampling of
Response design, 2 point-pairs, 436–440
Response surface design, 323 Simulated annealing, 364, 413
Restricted maximum likelihood Simulation, 9
estimation, 240, 258, 394, Small domain, 267
404, 435 Smallest relevant difference, 215
Soil aliquot, 72, 238
Sample mean, 30 Space-time design, 289
Sample quantile, 35 independent-synchronous design,
Sample selection scheme, 19 289
Sample semivariogram, 388 rotating panel design, 291
Sample size, 20 serially alternating design, 289
Sample support, 2 static-synchronous design, 289
Sample survey, 1 supplemented panel design, 291
Sample variance, 37, 463 Spatial cluster, 93
Sampling algorithm, 19 Spatial coverage, 69, 80
534 Index

Spatial coverage sampling, 327 Transect, 93

Spatial infill sampling, 327 Triangular grid, 81
Spatial response surface sampling, Two-phase random sampling, 201
357 Two-phase sampling
Spatial structure, 474 for regression, 206
Spatial variance, 37 for stratification, 203
Spatial variation model, 474 Two-sided alternative hypothesis, 470
Spatially correlated Poisson sampling, Two-stage cluster random sampling,
152 109
Square grid, 79
Standard error, 34 Unbiasedness
Standard normal distribution, 41 design-unbiasedness, 32
Star point, 360 model-unbiasedness, 376, 483
Stationarity in the mean, 376 Undersampled area, 326
Stratification effect, 56, 237 Uninformative prior, 223
Stratified random sampling, 49 Univariate stratification, 62
stratified cluster random Upper tail probability, 43
sampling, 106
stratified simple random Validation, 459
sampling, 49 Validation sample, 465
stratified two-stage cluster Validity, 319, 459, 475, 493
random sampling, 121 Variance component, 434
Stratified spatial response surface Variance to mean squared error ratio,
sampling, 367 197
Stratum, 49 Volunteer data, 199
Stratum bound, 65, 252
Wald interval, 44, 214
Stratum weight, 51
Weighted least squares, 180
Student’s t distribution, 41
Well-spread sample, 137
Subpopulation, 8
White noise, 489
Superpopulation, 474
Within-stratum variance, 52
Superpopulation mean, 474
Working model, 161, 182, 489
Superpopulation model, 161
Worst outcome criterion, 221
Supplemental sample, 283
Survey, 1 Yates-Grundy variance estimator, 133
Synthetic estimator, 267, 271
Systematic error, 29, 461
Systematic ppswor sampling, 128
Systematic random sampling, 79

Taylor linearisation of regression

estimator, 165
Temporal trend of spatial mean, 292
Test value, 215
Tolerable grid spacing, 397
Trans-Gaussian variable, 240