Bayesian Hierarchical Models: With Applications

Using R Peter D. Congdon download

https://textbookfull.com/product/bayesian-hierarchical-models-
with-applications-using-r-peter-d-congdon/

Download more ebook from https://textbookfull.com

We believe these products will be a great fit for you. Click
the link to download now, or visit textbookfull.com
to discover even more!

Data analysis using hierarchical generalized linear

models with R 1st Edition Youngjo Lee

https://textbookfull.com/product/data-analysis-using-
hierarchical-generalized-linear-models-with-r-1st-edition-
youngjo-lee/

Bayesian Disease Mapping : Hierarchical Modeling in

Spatial Epidemiology, Third Edition Lawson

https://textbookfull.com/product/bayesian-disease-mapping-
hierarchical-modeling-in-spatial-epidemiology-third-edition-
lawson/

Reasoning with Data An Introduction to Traditional and

Bayesian Statistics Using R 1st Edition Jeffrey M.
Stanton

https://textbookfull.com/product/reasoning-with-data-an-
introduction-to-traditional-and-bayesian-statistics-using-r-1st-
edition-jeffrey-m-stanton/

Habitat Suitability and Distribution Models With

Applications in R 1st Edition Antoine Guisan

https://textbookfull.com/product/habitat-suitability-and-
distribution-models-with-applications-in-r-1st-edition-antoine-
guisan/
Building Intelligent Cloud Applications Develop
Scalable Models Using Serverless Architectures with
Azure 1st Edition Vicente Herrera García

https://textbookfull.com/product/building-intelligent-cloud-
applications-develop-scalable-models-using-serverless-
architectures-with-azure-1st-edition-vicente-herrera-garcia/

Introduction to Deep Learning Using R: A Step-by-Step

Guide to Learning and Implementing Deep Learning Models
Using R Taweh Beysolow Ii

https://textbookfull.com/product/introduction-to-deep-learning-
using-r-a-step-by-step-guide-to-learning-and-implementing-deep-
learning-models-using-r-taweh-beysolow-ii/

Applied Hierarchical Modeling in Ecology. Analysis of

distribution, abundance and species richness in R and
BUGS: Volume 1: Prelude and Static Models 1st Edition
Kéry
https://textbookfull.com/product/applied-hierarchical-modeling-
in-ecology-analysis-of-distribution-abundance-and-species-
richness-in-r-and-bugs-volume-1-prelude-and-static-models-1st-
edition-kery/

Generalized Additive Models: An Introduction With R

(Second Edition) Simon N. Wood

https://textbookfull.com/product/generalized-additive-models-an-
introduction-with-r-second-edition-simon-n-wood/

Statistics and Probability with Applications for

Engineers and Scientists Using MINITAB, R and JMP 2nd
Edition Bhisham C. Gupta

https://textbookfull.com/product/statistics-and-probability-with-
applications-for-engineers-and-scientists-using-minitab-r-and-
jmp-2nd-edition-bhisham-c-gupta/
Bayesian Hierarchical Models
With Applications Using R
Second Edition
Bayesian Hierarchical Models
With Applications Using R
Second Edition

By
Peter D. Congdon
University of London, England
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2020 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-8575-4 (Hardback)

This book contains information obtained from authentic and highly regarded sources. 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 copy-
right 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 pho-
tocopying, 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, please access www.copyright.com (http://
www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA
01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users.
For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been
arranged.

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

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com
Contents

Preface...............................................................................................................................................xi

1. Bayesian Methods for Complex Data: Estimation and Inference.................................. 1

1.1 Introduction.................................................................................................................... 1
1.2 Posterior Inference from Bayes Formula.................................................................... 2
1.3 MCMC Sampling in Relation to Monte Carlo Methods; Obtaining
Posterior Inferences.......................................................................................................3
1.4 Hierarchical Bayes Applications..................................................................................5
1.5 Metropolis Sampling..................................................................................................... 8
1.6 Choice of Proposal Density..........................................................................................9
1.7 Obtaining Full Conditional Densities....................................................................... 10
1.8 Metropolis–Hastings Sampling................................................................................. 14
1.9 Gibbs Sampling............................................................................................................ 17
1.10 Hamiltonian Monte Carlo........................................................................................... 18
1.11 Latent Gaussian Models.............................................................................................. 19
1.12 Assessing Efficiency and Convergence; Ways of Improving Convergence......... 20
1.12.1 Hierarchical Model Parameterisation to Improve Convergence.............22
1.12.2 Multiple Chain Methods................................................................................ 24
1.13 Choice of Prior Density............................................................................................... 25
1.13.1 Including Evidence......................................................................................... 26
1.13.2 Assessing Posterior Sensitivity; Robust Priors........................................... 27
1.13.3 Problems in Prior Selection in Hierarchical Bayes Models...................... 29
1.14 Computational Notes.................................................................................................. 31
References................................................................................................................................ 37

2. Bayesian Analysis Options in R, and Coding for BUGS, JAGS, and Stan................ 45
2.1 Introduction.................................................................................................................. 45
2.2 Coding in BUGS and for R Libraries Calling on BUGS ......................................... 46
2.3 Coding in JAGS and for R Libraries Calling on JAGS............................................ 47
2.4 Coding for rstan .......................................................................................................... 49
2.4.1 Hamiltonian Monte Carlo............................................................................. 49
2.4.2 Stan Program Syntax...................................................................................... 49
2.4.3 The Target + Representation......................................................................... 51
2.4.4 Custom Distributions through a Functions Block..................................... 53
2.5 Miscellaneous Differences between Generic Packages
(BUGS, JAGS, and Stan)............................................................................................... 55
References................................................................................................................................ 56

3. Model Fit, Comparison, and Checking............................................................................. 59

3.1 Introduction.................................................................................................................. 59
3.2 Formal Model Selection.............................................................................................. 59
3.2.1 Formal Methods: Approximating Marginal Likelihoods......................... 62
3.2.2 Importance and Bridge Sampling Estimates..............................................63
3.2.3 Path Sampling.................................................................................................65

v
vi Contents

3.2.4 Marginal Likelihood for Hierarchical Models........................................... 67

3.3 Effective Model Dimension and Penalised Fit Measures...................................... 71
3.3.1 Deviance Information Criterion (DIC)......................................................... 72
3.3.2 Alternative Complexity Measures................................................................ 73
3.3.3 WAIC and LOO-IC......................................................................................... 75
3.3.4 The WBIC.........................................................................................................77
3.4 Variance Component Choice and Model Averaging..............................................80
3.4.1 Random Effects Selection..............................................................................80
3.5 Predictive Methods for Model Choice and Checking............................................ 87
3.5.1 Predictive Model Checking and Choice...................................................... 87
3.5.2 Posterior Predictive Model Checks.............................................................. 89
3.5.3 Mixed Predictive Checks............................................................................... 91
3.6 Computational Notes.................................................................................................. 95
References................................................................................................................................ 98

4. Borrowing Strength via Hierarchical Estimation......................................................... 103

4.1 Introduction................................................................................................................ 103
4.2 Hierarchical Priors for Borrowing Strength Using Continuous Mixtures........ 105
4.3 The Normal-Normal Hierarchical Model and Its Applications.......................... 106
4.3.1 Meta-Regression............................................................................................ 110
4.4 Prior for Second Stage Variance............................................................................... 111
4.4.1 Non-Conjugate Priors................................................................................... 113
4.5 Multivariate Meta-Analysis...................................................................................... 116
4.6 Heterogeneity in Count Data: Hierarchical Poisson Models............................... 121
4.6.1 Non-Conjugate Poisson Mixing.................................................................. 124
4.7 Binomial and Multinomial Heterogeneity............................................................. 126
4.7.1 Non-Conjugate Priors for Binomial Mixing............................................. 128
4.7.2 Multinomial Mixtures.................................................................................. 130
4.7.3 Ecological Inference Using Mixture Models............................................ 131
4.8 Discrete Mixtures and Semiparametric Smoothing Methods............................ 134
4.8.1 Finite Mixtures of Parametric Densities.................................................... 135
4.8.2 Finite Mixtures of Standard Densities....................................................... 136
4.8.3 Inference in Mixture Models...................................................................... 137
4.8.4 Particular Types of Discrete Mixture Model............................................ 141
4.8.5 The Logistic-Normal Alternative to the Dirichlet Prior.......................... 142
4.9 Semiparametric Modelling via Dirichlet Process and Polya Tree Priors.......... 144
4.9.1 Specifying the Baseline Density................................................................. 146
4.9.2 Truncated Dirichlet Processes and Stick-Breaking Priors...................... 148
4.9.3 Polya Tree Priors........................................................................................... 149
4.10 Computational Notes................................................................................................ 154
References.............................................................................................................................. 156

5. Time Structured Priors....................................................................................................... 165

5.1 Introduction................................................................................................................ 165
5.2 Modelling Temporal Structure: Autoregressive Models...................................... 166
5.2.1 Random Coefficient Autoregressive Models............................................ 168
5.2.2 Low Order Autoregressive Models............................................................ 169
5.2.3 Antedependence Models............................................................................. 170
5.3 State-Space Priors for Metric Data........................................................................... 172
Contents vii

5.3.1 Simple Signal Models................................................................................... 175

5.3.2 Sampling Schemes........................................................................................ 176
5.3.3 Basic Structural Model................................................................................. 178
5.3.4 Identification Questions............................................................................... 179
5.3.5 Nonlinear State-Space Models for Continuous Data............................... 184
5.4 Time Series for Discrete Responses; State-Space Priors and Alternatives......... 186
5.4.1 Other Approaches......................................................................................... 188
5.5 Stochastic Variances.................................................................................................. 193
5.6 Modelling Discontinuities in Time......................................................................... 197
5.7 Computational Notes................................................................................................ 202
References.............................................................................................................................. 206

6. Representing Spatial Dependence................................................................................... 213

6.1 Introduction................................................................................................................ 213
6.2 Spatial Smoothing and Prediction for Area Data.................................................. 214
6.2.1 SAR Schemes................................................................................................. 216
6.3 Conditional Autoregressive Priors.......................................................................... 221
6.3.1 Linking Conditional and Joint Specifications...........................................222
6.3.2 Alternative Conditional Priors.................................................................... 223
6.3.3 ICAR(1) and Convolution Priors................................................................. 226
6.4 Priors on Variances in Conditional Spatial Models.............................................. 227
6.5 Spatial Discontinuity and Robust Smoothing....................................................... 229
6.6 Models for Point Processes.......................................................................................234
6.6.1 Covariance Functions................................................................................... 237
6.6.2 Sparse and Low Rank Approaches............................................................ 238
6.7 Discrete Convolution Models................................................................................... 241
6.8 Computational Notes................................................................................................ 245
References.............................................................................................................................. 246

7. Regression Techniques Using Hierarchical Priors....................................................... 253

7.1 Introduction................................................................................................................ 253
7.2 Predictor Selection..................................................................................................... 253
7.2.1 Predictor Selection........................................................................................254
7.2.2 Shrinkage Priors........................................................................................... 256
7.3 Categorical Predictors and the Analysis of Variance........................................... 259
7.3.1 Testing Variance Components.................................................................... 260
7.4 Regression for Overdispersed Data........................................................................ 264
7.4.1 Overdispersed Poisson Regression............................................................ 264
7.4.2 Overdispersed Binomial and Multinomial Regression.......................... 267
7.5 Latent Scales for Binary and Categorical Data...................................................... 270
7.5.1 Augmentation for Ordinal Responses....................................................... 273
7.6 Heteroscedasticity and Regression Heterogeneity............................................... 276
7.6.1 Nonconstant Error Variances...................................................................... 276
7.6.2 Varying Regression Effects via Discrete Mixtures.................................. 277
7.6.3 Other Applications of Discrete Mixtures.................................................. 278
7.7 Time Series Regression: Correlated Errors and Time-Varying
Regression Effects...................................................................................................... 282
7.7.1 Time-Varying Regression Effects............................................................... 283
7.8 Spatial Regression...................................................................................................... 288
viii Contents

7.8.1 Spatial Lag and Spatial Error Models........................................................ 288

7.8.2 Simultaneous Autoregressive Models....................................................... 288
7.8.3 Conditional Autoregression........................................................................ 290
7.8.4 Spatially Varying Regression Effects: GWR and Bayesian SVC
Models............................................................................................................ 291
7.8.5 Bayesian Spatially Varying Coefficients.................................................... 292
7.8.6 Bayesian Spatial Predictor Selection Models............................................ 293
7.9 Adjusting for Selection Bias and Estimating Causal Effects............................... 296
7.9.1 Propensity Score Adjustment...................................................................... 296
7.9.2 Establishing Causal Effects: Mediation and Marginal Models.............. 299
7.9.3 Causal Path Sequences................................................................................. 299
7.9.4 Marginal Structural Models........................................................................306
References..............................................................................................................................308

8. Bayesian Multilevel Models.............................................................................................. 317

8.1 Introduction................................................................................................................ 317
8.2 The Normal Linear Mixed Model for Hierarchical Data..................................... 318
8.2.1 The Lindley–Smith Model Format............................................................. 320
8.3 Discrete Responses: GLMM, Conjugate, and Augmented Data Models........... 322
8.3.1 Augmented Data Multilevel Models.......................................................... 324
8.3.2 Conjugate Cluster Effects............................................................................. 325
8.4 Crossed and Multiple Membership Random Effects........................................... 328
8.5 Robust Multilevel Models......................................................................................... 331
References.............................................................................................................................. 336

9. Factor Analysis, Structural Equation Models, and Multivariate Priors................... 339

9.1 Introduction................................................................................................................ 339
9.2 Normal Linear Structural Equation and Factor Models......................................340
9.2.1 Forms of Model.............................................................................................342
9.2.2 Model Definition...........................................................................................343
9.2.3 Marginal and Complete Data Likelihoods, and MCMC Sampling.......345
9.3 Identifiability and Priors on Loadings....................................................................346
9.3.1 An Illustration of Identifiability Issues......................................................348
9.4 Multivariate Exponential Family Outcomes and Generalised Linear
Factor Models............................................................................................................. 354
9.4.1 Multivariate Count Data.............................................................................. 355
9.4.2 Multivariate Binary Data and Item Response Models............................ 357
9.4.3 Latent Scale IRT Models............................................................................... 359
9.4.4 Categorical Data............................................................................................ 360
9.5 Robust Density Assumptions in Factor Models.................................................... 370
9.6 Multivariate Spatial Priors for Discrete Area Frameworks................................. 373
9.7 Spatial Factor Models................................................................................................ 379
9.8 Multivariate Time Series........................................................................................... 381
9.8.1 Multivariate Dynamic Linear Models....................................................... 381
9.8.2 Dynamic Factor Analysis............................................................................ 386
9.8.3 Multivariate Stochastic Volatility............................................................... 388
9.9 Computational Notes................................................................................................ 396
References.............................................................................................................................. 397
Contents ix

10. Hierarchical Models for Longitudinal Data.................................................................. 405

10.1 Introduction................................................................................................................ 405
10.2 General Linear Mixed Models for Longitudinal Data......................................... 406
10.2.1 Centred or Non-Centred Priors..................................................................408
10.2.2 Priors on Unit Level Random Effects......................................................... 409
10.2.3 Priors for Random Covariance Matrix and
Random Effect Selection.............................................................................. 411
10.2.4 Priors for Multiple Sources of Error Variation.......................................... 415
10.3 Temporal Correlation and Autocorrelated Residuals........................................... 418
10.3.1 Explicit Temporal Schemes for Errors....................................................... 419
10.4 Longitudinal Categorical Choice Data....................................................................423
10.5 Observation Driven Autocorrelation: Dynamic Longitudinal Models............. 427
10.5.1 Dynamic Models for Discrete Data............................................................ 429
10.6 Robust Longitudinal Models: Heteroscedasticity, Generalised Error
Densities, and Discrete Mixtures............................................................................ 433
10.6.1 Robust Longitudinal Data Models: Discrete Mixture Models............... 436
10.7 Multilevel, Multivariate, and Multiple Time Scale Longitudinal Data..............443
10.7.1 Latent Trait Longitudinal Models..............................................................445
10.7.2 Multiple Scale Longitudinal Data..............................................................446
10.8 Missing Data in Longitudinal Models.................................................................... 452
10.8.1 Forms of Missingness Regression (Selection Approach)........................454
10.8.2 Common Factor Models............................................................................... 455
10.8.3 Missing Predictor Data................................................................................ 457
10.8.4 Pattern Mixture Models............................................................................... 459
References.............................................................................................................................. 462

11. Survival and Event History Models................................................................................ 471

11.1 Introduction................................................................................................................ 471
11.2 Survival Analysis in Continuous Time.................................................................. 472
11.2.1 Counting Process Functions....................................................................... 474
11.2.2 Parametric Hazards...................................................................................... 475
11.2.3 Accelerated Hazards.................................................................................... 478
11.3 Semiparametric Hazards.......................................................................................... 481
11.3.1 Piecewise Exponential Priors...................................................................... 482
11.3.2 Cumulative Hazard Specifications.............................................................484
11.4 Including Frailty........................................................................................................ 488
11.4.1 Cure Rate Models.......................................................................................... 490
11.5 Discrete Time Hazard Models................................................................................. 494
11.5.1 Life Tables...................................................................................................... 496
11.6 Dependent Survival Times: Multivariate and Nested Survival Times.............. 502
11.7 Competing Risks........................................................................................................ 507
11.7.1 Modelling Frailty.......................................................................................... 509
11.8 Computational Notes................................................................................................ 514
References.............................................................................................................................. 519

12. Hierarchical Methods for Nonlinear and Quantile Regression................................ 525

12.1 Introduction................................................................................................................ 525
12.2 Non-Parametric Basis Function Models for the Regression Mean..................... 526
12.2.1 Mixed Model Splines.................................................................................... 527
x Contents

12.2.2 Basis Functions Other Than Truncated Polynomials.............................. 529

12.2.3 Model Selection............................................................................................. 532
12.3 Multivariate Basis Function Regression................................................................. 536
12.4 Heteroscedasticity via Adaptive Non-Parametric Regression............................ 541
12.5 General Additive Methods.......................................................................................543
12.6 Non-Parametric Regression Methods for Longitudinal Analysis......................546
12.7 Quantile Regression.................................................................................................. 552
12.7.1 Non-Metric Responses................................................................................. 554
12.8 Computational Notes................................................................................................ 560
References.............................................................................................................................. 560

Index.............................................................................................................................................. 565
Preface

My gratitude is due to Taylor & Francis for proposing a revision of Applied Bayesian
Hierarchical Methods, first published in 2010. The revision maintains the goals of present-
ing an overview of modelling techniques from a Bayesian perspective, with a view to
practical data analysis. The new book is distinctive in its computational environment,
which is entirely R focused. Worked examples are based particularly on rjags and jagsUI,
R2OpenBUGS, and rstan. Many thanks are due to the following for comments on chap-
ters or computing advice: Sid Chib, Andrew Finley, Ken Kellner, Casey Youngflesh,
Kaushik Chowdhury, Mahmoud Torabi, Matt Denwood, Nikolaus Umlauf, Marco Geraci,
Howard Seltman, Longhai Li, Paul Buerkner, Guanpeng Dong, Bob Carpenter, Mitzi
Morris, and Benjamin Cowling. Programs for the book can be obtained from my website
at https://www.qmul.ac.uk/geog/staff/congdonp.html or from https://www.crcpress.com/
Bayesian-Hierarchical-Models-With-Applications-Using-R-Second-Edition/Congdon/p/
book/9781498785754. Please send comments or questions to me at [email protected].

QMUL, London

xi
1
Bayesian Methods for Complex Data:
Estimation and Inference

1.1 Introduction
The Bayesian approach to inference focuses on updating knowledge about unknown
parameters θ in a statistical model on the basis of observations y, with revised knowledge
expressed in the posterior density p(θ|y). The sample of observations y being analysed
provides new information about the unknowns, while the prior density p(θ) represents
accumulated knowledge about them before observing or analysing the data. There is
considerable flexibility with which prior evidence about parameters can be incorporated
into an analysis, and use of informative priors can reduce the possibility of confounding
and provides a natural basis for evidence synthesis (Shoemaker et al., 1999; Dunson, 2001;
Vanpaemel, 2011; Klement et al., 2018). The Bayes approach provides uncertainty intervals
on parameters that are consonant with everyday interpretations (Willink and Lira, 2005;
Wetzels et al., 2014; Krypotos et al., 2017), and has no problem comparing the fit of non-
nested models, such as a nonlinear model and its linearised version.
Furthermore, Bayesian estimation and inference have a number of advantages in terms
of its relevance to the types of data and problems tackled by modern scientific research
which are a primary focus later in the book. Bayesian estimation via repeated sampling
from posterior densities facilitates modelling of complex data, with random effects treated
as unknowns and not integrated out as is sometimes done in frequentist approaches
(Davidian and Giltinan, 2003). For example, much of the data in social and health research
has a complex structure, involving hierarchical nesting of subjects (e.g. pupils within
schools), crossed classifications (e.g. patients classified by clinic and by homeplace),
spatially configured data, or repeated measures on subjects (MacNab et al., 2004). The
Bayesian approach naturally adapts to such hierarchically or spatio-temporally correlated
effects via conditionally specified hierarchical priors under a three-stage scheme (Lindley
and Smith, 1972; Clark and Gelfand, 2006; Gustafson et al., 2006; Cressie et al., 2009), with
the first stage specifying the likelihood of the data, given unknown random individual or
cluster effects; the second stage specifying the density of the random effects; and the third
stage providing priors on parameters underlying the random effects density or densities.
The increased application of Bayesian methods has owed much to the development of
Markov chain Monte Carlo (MCMC) algorithms for estimation (Gelfand and Smith, 1990;
Gilks et al., 1996; Neal, 2011), which draw repeated parameter samples from the posterior
distributions of statistical models, including complex models (e.g. models with multiple
or nested random effects). Sampling based parameter estimation via MCMC provides
a full posterior density of a parameter so that any clear non-normality is apparent, and

1
2 Bayesian Hierarchical Models

hypotheses about parameters or interval estimates can be assessed from the MCMC sam-
ples without the assumptions of asymptotic normality underlying many frequentist tests.
However, MCMC methods may in practice show slow convergence, and implementation of
some MCMC methods (such as Hamiltonian Monte Carlo) with advantageous estimation
features, including faster convergence, has been improved through package development
(rstan) in R.
As mentioned in the Preface, a substantial emphasis in the book is placed on implemen-
tation and data analysis for tutorial purposes, via illustrative data analysis and attention
to statistical computing. Accordingly, worked examples in R code in the rest of the chap-
ter illustrate MCMC sampling and Bayesian posterior inference from first principles. In
subsequent chapters R based packages, such as jagsUI, rjags, R2OpenBUGS, and rstan are
used for computation.
As just mentioned, Bayesian modelling of hierarchical and random effect models via
MCMC techniques has extended the scope for modern data analysis. Despite this, applica-
tion of Bayesian techniques also raises particular issues, although these have been allevi-
ated by developments such as integrated nested Laplace approximation (Rue et al., 2009)
and practical implementation of Hamiltonian Monte Carlo (Carpenter et al., 2017). These
include:

a) Propriety and identifiability issues when diffuse priors are applied to variance or
dispersion parameters for random effects (Hobert and Casella, 1996; Palmer and
Pettit, 1996; Hadjicostas and Berry, 1999; Yue et al., 2012);
b) Selecting the most suitable form of prior for variance parameters (Gelman, 2006)
or the most suitable prior for covariance modelling (Lewandowski et al., 2009);
c) Appropriate priors for models with random effects, to avoid potential overfitting
(Simpson et al., 2017; Fuglstad et al., 2018) or oversmoothing in the presence of
genuine outliers in spatial applications (Conlon and Louis, 1999);
d) The scope for specification bias in hierarchical models for complex data structures
where a range of plausible model structures are possible (Chiang et al., 1999).

1.2 Posterior Inference from Bayes Formula

Statistical analysis uses probability models to summarise univariate or multivariate
observations y = ( y1 , … , y n ) by a collection of unknown parameters of dimension (say
d), q = (q1 ,… ,q d ) . Consider the joint density p( y ,q ) = p( y|q )p(q ), where p(y|θ) is the sam-
pling model or likelihood, and p(θ) defines existing knowledge, or expresses assumptions
regarding the unknowns that can be justified by the nature of the application (e.g. that
random effects are spatially distributed in an area application). A Bayesian analysis seeks
to update knowledge about the unknowns θ using the data y, and so interest focuses on
the posterior density p(θ|y) of the unknowns. Since p(y,θ) also equals p(y)p(θ|y) where p(y)
is the unconditional density of the data (also known as the marginal likelihood), one may
obtain

p( y ,q ) = p( y|q )p(q ) = p( y )p(q |y ). (1.1)

Bayesian Methods for Complex Data 3

This can be rearranged to provide the required posterior density as

p( y|q )p(q )
p(q |y ) = . (1.2)
p( y )
The marginal likelihood p(y) may be obtained by integrating the numerator on the right
side of (1.2) over the support for θ, namely

ò
p( y ) = p( y|q )p(q )dq .

From (1.2), the term p(y) therefore acts as a normalising constant necessary to ensure p(θ|y)
integrates to 1, and so one may write

p(q |y ) = kp( y|q )p(q ), (1.3)

where k = 1/p(y) is an unknown constant. Alternatively stated, the posterior density

(updated evidence) is proportional to the likelihood (data evidence) times the prior (his-
toric evidence or elicited model assumptions). Taking logs in (1.3), one has

log éë p(q |y )ùû = log(k ) + log éë p( y|q )ùû + log éë p(q )ùû

and log[ p( y|q )] + log[ p(q )] is generally referred to as the log posterior, which some R pro-
grams (e.g. rstan) allow to be directly specified as the estimation target.
In some cases, when the prior on θ is conjugate with the posterior on θ (i.e. has the same
density form), the posterior density and marginal likelihood can be obtained analytically.
When θ is low-dimensional, numerical integration is an alternative, and approximations to
the required integrals can be used, such as the Laplace approximation (Raftery, 1996; Chen
and Wang, 2011). In more complex applications, such approximations are not feasible, and
integration to obtain p(y) is intractable, so that direct sampling from p(θ|y) is not feasible.
In such situations, MCMC methods provide a way to sample from p(θ|y) without it having
a specific analytic form. They create a Markov chain of sampled values q (1) ,… ,q (T ) with
transition kernel K(q cand |q curr ) (investigating transitions from current to candidate values
for parameters) that have p(θ|y) as their limiting distribution. Using large samples from
the posterior distribution obtained by MCMC, one can estimate posterior quantities of
interest such as posterior means, medians, and highest density regions (Hyndman, 1996;
Chen and Shao, 1998).

1.3 MCMC Sampling in Relation to Monte Carlo

Methods; Obtaining Posterior Inferences
Markov chain Monte Carlo (MCMC) methods are iterative sampling methods that can be
encompassed within the broad class of Monte Carlo methods. However, MCMC methods
must be distinguished from conventional Monte Carlo methods that generate independent
simulations {u(1) , u(2) … , u(T ) } from a target density π(u). From such simulations, the expecta-
tion of a function g(u) under π(u), namely
4 Bayesian Hierarchical Models

∫
Ep [ g(u)] = g(u)p(u)du,

is estimated as

g= ∑ g (u
t =1
(t )
)

and, under independent sampling from π(u), g tends to Ep [ g(u)] as T → ∞. However, such
independent sampling from the posterior density p(θ|y) is not usually feasible.
When suitably implemented, MCMC methods offer an effective alternative way to gen-
erate samples from the joint posterior distribution, p(θ|y), but differ from conventional
Monte Carlo methods in that successive sampled parameters are dependent or autocorre-
lated. The target density for MCMC samples is therefore the posterior density π(θ) = p(θ|y)
and MCMC sampling is especially relevant when the posterior cannot be stated exactly
in analytic form e.g. when the prior density assumed for θ is not conjugate with the like-
lihood p(y|θ). The fact that successive sampled values are dependent means that larger
samples are needed for equivalent precision, and the effective number of samples is less
than the nominal number.
For the parameter sampling case, assume a preset initial parameter value θ(0). Then
MCMC methods involve repeated iterations to generate a correlated sequence of sampled
values θ(t) (t = 1, 2, 3, …), where updated values θ(t) are drawn from a transition distribution

K (q (t ) |q (0) ,… ,q (t -1) ) = K (q (t ) |q (t -1) )

that is Markovian in the sense of depending only on θ(t−1). The transition distribution
K (q (t ) |q (t -1) ) is chosen to satisfy additional conditions ensuring that the sequence has
the joint posterior density p(θ|y) as its stationary distribution. These conditions typically
reduce to requirements on the proposal and acceptance procedure used to generate can-
didate parameter samples. The proposal density and acceptance rule must be specified in
a way that guarantees irreducibility and positive recurrence; see, for example, Andrieu
and Moulines (2006). Under such conditions, the sampled parameters θ(t) {t = B, B + 1, … , T },
beyond a certain burn-in or warm-up phase in the sampling (of B iterations), can be viewed
as a random sample from p(θ|y) (Roberts and Rosenthal, 2004).
In practice, MCMC methods are applied separately to individual parameters or blocks of
more than one parameter (Roberts and Sahu, 1997). So, assuming θ contains more than one
parameter and consists of C components or blocks {q1 , … , qC } , different updating methods
may be used for each component, including block updates.
There is no limit to the number of samples T of θ which may be taken from the poste-
rior density p(θ|y). Estimates of the marginal posterior densities for each parameter can
be made from the MCMC samples, including estimates of location (e.g. posterior means,
modes, or medians), together with the estimated certainty or precision of these parameters
in terms of posterior standard deviations, credible intervals, or highest posterior density
intervals. For example, the 95% credible interval for θh may be estimated using the 0.025
and 0.975 quantiles of the sampled output {q h(t ) , t = B + 1,… , T } . To reduce irregularities in
the histogram of sampled values for a particular parameter, a smooth form of the posterior
density can be approximated by applying kernel density methods to the sampled values.
Monte Carlo posterior summaries typically include estimated posterior means and vari-
ances of the parameters, obtainable as moment estimates from the MCMC output, namely
Bayesian Methods for Complex Data 5

T
Ê(q h ) = q h = åq
t =B + 1
(t )
h /(T - B)

T
V̂ (q h ) = å (q
t=B+1
(t )
h - q h )2 /(T - B).

This is equivalent to estimating the integrals

ò
E(q h |y ) = q h p(q |y )dq ,

ò
V (q h |y ) = q h2 p(q |y )dq - [E(q h |y )]2

= E(q h2 |y ) - [E(q h |y )]2 .

One may also use the MCMC output to derive obtain posterior means, variances, and
credible intervals for functions Δ = Δ(θ) of the parameters (van Dyk, 2003). These are esti-
mates of the integrals

ò
E[D(q )|y] = D(q )p(q |y )dq ,

∫
V[∆(q )| y] = ∆ 2 p(q | y )dq − [E( ∆ | y )]2

= E( ∆ 2 | y ) − [E( ∆ | y )]2 .

For Δ(θ), its posterior mean is obtained by calculating Δ(t) at every MCMC iteration from
the sampled values θ(t). The theoretical justification for such estimates is provided by the
MCMC version of the law of large numbers (Tierney, 1994), namely that

T
D[q (t ) ]
å T - B ® E [D(q )],
t =B + 1
p

provided that the expectation of Δ(θ) under p (q ) = p(q |y ), denoted Eπ[Δ(θ)], exists. MCMC
methods also allow inferences on parameter comparisons (e.g. ranks of parameters or con-
trasts between them) (Marshall and Spiegelhalter, 1998).

1.4 Hierarchical Bayes Applications

The paradigm in Section 1.2 is appropriate to many problems, where uncertainty is limited
to a few fundamental parameters, the number of which is independent of the sample size
n – this is the case, for example, in a normal linear regression when the independent vari-
ables are known without error and the units are not hierarchically structured. However,
6 Bayesian Hierarchical Models

in more complex data sets or with more complex forms of model or response, a more gen-
eral perspective than that implied by (1.1)–(1.3) is available, and also implementable, using
MCMC methods.
Thus, a class of hierarchical Bayesian models are defined by latent data (Paap, 2002;
Clark and Gelfand, 2006) intermediate between the observed data and the underlying
parameters (hyperparameters) driving the process. A terminology useful for relating hier-
archical models to substantive issues is proposed by Wikle (2003) in which y defines the
data stage, latent effects b define the process stage, and ξ defines the hyperparameter stage.
For example, the observations i = 1,…,n may be arranged in clusters j = 1, …, J, so that the
observations can no longer be regarded as independent. Rather, subjects from the same
cluster will tend to be more alike than individuals from different clusters, reflecting latent
variables that induce dependence within clusters.
Let the parameters θ = [θL,θb] consist of parameter subsets relevant to the likelihood and
to the latent data density respectively. The data are generally taken as independent of θb
given b, so modelling intermediate latent effects involves a three-stage hierarchical Bayes
(HB) prior set-up

p( y , b ,q ) = p( y|b ,q L )p(b|q b )p(q L , q b ), (1.4)

with a first stage likelihood p( y|b ,q L ) and a second stage density p(b|θb) for the latent data,
with conditioning on higher stage parameters θ. The first stage density p(y|b,θL) in (1.4) is
a conditional likelihood, conditioning on b, and sometimes called the complete data or
augmented data likelihood. The application of Bayes’ theorem now specifies

p( y|b ,q L )p(b|q b )p(q )

p(q , b|y ) = ,
p( y )
and the marginal posterior for θ may now be represented as

p(q |y ) = =
ò
p(q )p( y|q ) p(q ) p( y|b ,q L )p(b|q b )db
,
p( y ) p( y )
where

ò ò
p( y|q ) = p( y , b|q )db = p( y|b ,q L )p(b|q b )db ,

is the observed data likelihood, namely the complete data likelihood with b integrated out,
sometimes also known as the integrated likelihood.
Often the latent data exist for every observation, or they may exist for each cluster in
which the observations are structured (e.g. a school specific effect bj for multilevel data yij
on pupils i nested in schools j). The latent variables b can be seen as a population of values
from an underlying density (e.g. varying log odds of disease) and the θb are then popula-
tion hyperparameters (e.g. mean and variance of the log odds) (Dunson, 2001). As exam-
ples, Paap (2002) mentions unobserved states describing the business cycle and Johannes
and Polson (2006) mention unobserved volatilities in stochastic volatility models, while
Albert and Chib (1993) consider the missing or latent continuous data {b1, …, bn} which
underlie binary observations {y1, …, yn}. The subject specific latent traits in psychometric or
educational item analysis can also be considered this way (Fox, 2010), as can the variance
Bayesian Methods for Complex Data 7

scaling factors in the robust Student t errors version of linear regression (Geweke, 1993) or
subject specific slopes in a growth curve analysis of panel data on a collection of subjects
(Oravecz and Muth, 2018).
Typically, the integrated likelihood p(y|θ) cannot be stated in closed form and classical
likelihood estimation relies on numerical integration or simulation (Paap, 2002, p.15). By
contrast, MCMC methods can be used to generate random samples indirectly from the
posterior distribution p(θ,b|y) of parameters and latent data given the observations. This
requires only that the augmented data likelihood be known in closed form, without need-
ing to obtain the integrated likelihood p(y|θ). To see why, note that the marginal posterior
of the parameter set θ may alternatively be derived as

ò ò
p(q |y ) = p(q , b|y )db = p(q |y , b)p(b|y )db ,

with marginal densities for component parameters θh of the form (Paap, 2002, p.5)

p(q h |y ) =
ò ò p(q , b|y)dbdq
q [ h] b
[ h] ,

µ
ò p(q |y)p(q )dq
q [ h]
[ h] =
ò ò p(q )p(y|b,q )p(b|q )dbdq
q [ h] b
[ h] ,

where θ[h] consists of all parameters in θ with the exception of θh. The derivation of suitable
MCMC algorithms to sample from p(θ,b|y) is based on Clifford–Hammersley theorem,
namely that any joint distribution can be fully characterised by its complete conditional
distributions. In the hierarchical Bayes context, this implies that the conditionals p(b|θ,y)
and p(θ|b,y) characterise the joint distribution p(θ,b|y) from which samples are sought, and
so MCMC sampling can alternate between updates p(b(t ) |q (t -1) , y ) and p(q (t ) |b(t ) , y ) on con-
ditional densities, which are usually of simpler form than p(θ,b|y). The imputation of latent
data in this way is sometimes known as data augmentation (van Dyk, 2003).
To illustrate the application of MCMC methods to parameter comparisons and hypoth-
esis tests in an HB setting, Shen and Louis (1998) consider hierarchical models with unit
or cluster specific parameters bj, and show that if such parameters are the focus of interest,
their posterior means are the optimal estimates. Suppose instead that the ranks of the unit
or cluster parameters, namely

Rj = rank(b j ) = ∑ I(b ≥ b ),
k≠i
j k

(where I(A) is an indicator function which equals 1 when A is true, 0 otherwise) are
required for deriving “league tables”. Then the conditional expected ranks are optimal,
and obtained by ranking the bj at each MCMC iteration, and taking the means of these
ranks over all samples. By contrast, ranking posterior means of the bj themselves can
perform poorly (Laird and Louis, 1989; Goldstein and Spiegelhalter, 1996). Similarly,
when the empirical distribution function of the unit parameters (e.g. to be used to obtain
the fraction of parameters above a threshold) is required, the conditional expected EDF
is optimal.
8 Bayesian Hierarchical Models

A posterior probability estimate that a particular bj exceeds a threshold τ, namely of the

∞
integral Pr(b j > t| y ) =
∫ p(b |y)db , is provided by the proportion of iterations where b
t
j j
(t )
j

exceeds τ, namely
T
 ( b j > t| y ) =
Pr ∑ I (b
t =B + 1
(t )
j > t)/(T − B).

Thus, one might, in an epidemiological application, wish to obtain the posterior probabil-
ity that an area’s smoothed relative mortality risk bj exceeds unity, and so count iterations
where this condition holds. If this probability exceeds a threshold such as 0.9, then a sig-
nificant excess risk is indicated, whereas a low exceedance probability (the sampled rela-
tive risk rarely exceeded 1) would indicate a significantly low mortality level in the area.
In fact, the significance of individual random effects is one aspect of assessing the gain of
a random effects model over a model involving only fixed effects, or of assessing whether
a more complex random effects model offers a benefit over a simpler one (Knorr-Held and
Rainer, 2001, p.116). Since the variance can be defined in terms of differences between ele-
ments of the vector (b1 ,..., bJ ), as opposed to deviations from a central value, one may also
consider which contrasts between pairs of b values are significant. Thus, Deely and Smith
(1998) suggest evaluating probabilities Pr(b j ≤ tbk |k ≠ j , y ) where 0 < t ≤ 1, namely, the pos-
terior probability that any one hierarchical effect is smaller by a factor τ than all the others.

1.5 Metropolis Sampling
A range of MCMC techniques is available. The Metropolis sampling algorithm is still a
widely applied MCMC algorithm and is a special case of Metropolis–Hastings consid-
ered in Section 1.8. Let p(y|θ) denote a likelihood, and p(θ) denote the prior density for
θ, or more specifically the prior densities p(q1 ),… p(qC ) of the components of θ. Then the
Metropolis algorithm involves a symmetric proposal density (e.g. a Normal, Student t, or
uniform density) q(q cand |q (t ) ) for generating candidate parameter values θcand, with accep-
tance probability for potential candidate values obtained as

æ p (q cand ) ö æ p(q cand |y ) ö æ p( y|q cand )p(q cand ) ö

a (t ) = min ç 1, (t ) ÷
= min ç 1, ÷ = min ç 1, ÷. (1.5)
è p (q ) ø è p(q |y ) ø p( y|q (t ) )p(q (t ) ) ø
(t )
è
So one compares the (likelihood * prior), namely, p( y|q )p(q ), for the candidate and exist-
ing parameter values. If the (likelihood * prior) is higher for the candidate value, it is auto-
matically accepted, and q (t+1) = q cand. However, even if the (likelihood * prior) is lower for
the candidate value, such that α(t) is less than 1, the candidate value may still be accepted.
This is decided by random sampling from a uniform density, U(t) and the candidate value
is accepted if a(t ) ≥ U (t ) . In practice, comparisons involve the log posteriors for existing and
candidate parameter values.
The third equality in (1.5) follows because the marginal likelihood p(y) = 1/k in the
Bayesian formula

p(q |y ) = p( y|q )p(q )/ p( y ) = kp( y|q )p(q ),

Bayesian Methods for Complex Data 9

cancels out, as it is a constant. Stated more completely, to sample parameters under the
Metropolis algorithm, it is not necessary to know the normalised target distribution,
namely, the posterior density, π(θ|y); it is enough to know it up to a constant factor.
So, for updating parameter subsets, the Metropolis algorithm can be implemented by
using the full posterior distribution

p (q ) = p(q |y ) = kp( y|q )p(q ),

as the target distribution – which in practice involves comparisons of the unnormalised

posterior p(y|θ)p(θ). However, for updating values on a particular parameter θh, it is not just
p(y) that cancels out in the ratio

p( y|q cand )p(q cand )

p (q cand )/p (q (t ) ) = ,
p( y|q (t ) )p(q (t ) )
but any parts of the likelihood or prior not involving θh (these parts are constants when θh
is being updated).
When those parts of the likelihood or prior not relevant to θh are abstracted out, the
remaining part of p(q |y ) = kp( y|q )p(q ), the part relevant to updating θh, is known as the
full conditional density for θh (Gilks, 1996). One may denote the full conditional density
for θh as

p h (q h |q[ h] ) µ p( y|q h )p(q h ),

where θh] denotes the parameter set excluding θh. So, the probability for updating θh can be
obtained either by comparing the full posterior (known up to a constant k), namely

æ p (q h ,cand ,q[(ht]) ) ö æ p( y|q h ,cand ,q[(ht]) )p(q h ,cand ,q[(ht]) ) ö

a = min çç 1, ÷
÷ = min çç 1, ÷÷ ,
è p (q (t ) ) ø è p( y|q (t ) )p(q (t ) ) ø
or by using the full conditional for the hth parameter, namely

æ p h (q h ,cand |q[(ht]) ) ö
a = min çç 1, ÷.
è p h (q h(t ) |q[(ht]) ) ÷ø
Then one sets q h(t +1) = q h ,cand with probability α, and q h(t +1) = q h(t ) otherwise.

1.6 Choice of Proposal Density

There is some flexibility in the choice of proposal density q for generating candidate values
in the Metropolis and other MCMC algorithms, but the chosen density and the parameters
incorporated in it are relevant to successful MCMC updating and convergence (Altaleb
and Chauveau, 2002; Robert, 2015). A standard recommendation is that the proposal den-
sity for a particular parameter θh should approximate the posterior density p(θh|y) of that
parameter. In some cases, one may have an idea (e.g. from a classical analysis) of what
the posterior density is, or what its main defining parameters are. A normal proposal is
10 Bayesian Hierarchical Models

often justified, as many posterior densities do approximate normality. For example, Albert
(2007) applies a Laplace approximation technique to estimate the posterior mode, and uses
the mean and variance parameters to define the proposal densities used in a subsequent
stage of Metropolis–Hastings sampling.
The rate at which a proposal generated by q is accepted (the acceptance rate) depends on
how close θcand is to θ(t), and this in turn depends on the variance sq2 of the proposal density.
A higher acceptance rate would typically follow from reducing sq2 , but with the risk that
the posterior density will take longer to explore. If the acceptance rate is too high, then
autocorrelation in sampled values will be excessive (since the chain tends to move in a
restricted space), while a too low acceptance rate leads to the same problem, since the chain
then gets locked at particular values.
One possibility is to use a variance or dispersion estimate, sm2 or Σm, from a maximum
likelihood or other mode-finding analysis (which approximates the posterior variance)
and then scale this by a constant c > 1, so that the proposal density variance is sq2 = csm2 .
Values of c in the range 2–10 are typical. For θh of dimension dh with covariance Σm, a pro-
posal density dispersion 2.382Σm/dh is shown as optimal in random walk schemes (Roberts
et al., 1997). Working rules are for an acceptance rate of 0.4 when a parameter is updated
singly (e.g. by separate univariate normal proposals), and 0.2 when a group of parameters
are updated simultaneously as a block (e.g. by a multivariate normal proposal). Geyer and
Thompson (1995) suggest acceptance rates should be between 0.2 and 0.4, and optimal
acceptance rates have been proposed (Roberts et al., 1997; Bedard, 2008).
Typical Metropolis updating schemes use variables Wt with known scale, for example,
uniform, standard Normal, or standard Student t. A Normal proposal density q(q cand |q (t ) )
then involves samples Wt ~ N(0,1), with candidate values

q cand = q (t ) + s qWt ,

where σq determines the size of the jump from the current value (and the acceptance
rate). A uniform random walk samples Wt  Unif( −1,1) and scales this to form a proposal
q cand = q (t ) + k Wt , with the value of κ determining the acceptance rate. As noted above, it is
desirable that the proposal density approximately matches the shape of the target density
p(θ|y). The Langevin random walk scheme is an example of a scheme including informa-
tion about the shape of p(θ|y) in the proposal, namely q cand = q (t ) + s q [Wt + 0.5Ñ log( p(q (t ) |y )]
where ∇ denotes the gradient function (Roberts and Tweedie, 1996).
Sometimes candidate parameter values are sampled using a transformed version of a
parameter, for example, normal sampling of a log variance rather than sampling of a vari-
ance (which has to be restricted to positive values). In this case, an appropriate Jacobean
adjustment must be included in the likelihood. Example 1.2 below illustrates this.

1.7 Obtaining Full Conditional Densities

As noted above, Metropolis sampling may be based on the full conditional density when
a particular parameter θh is being updated. These full conditionals are particularly central
in Gibbs sampling (see below). The full conditional densities may be obtained from the
joint density p(q , y ) = p( y|q )p(q ) and in many cases reduce to standard densities (Normal,
Bayesian Methods for Complex Data 11

exponential, gamma, etc.) from which direct sampling is straightforward. Full conditional
densities are derived by abstracting out from the joint model density p(y|θ)p(θ) (likelihood
times prior) only those elements including θh and treating other components as constants
(George et al., 1993; Gilks, 1996).
Consider a conjugate model for Poisson count data yi with means μi that are themselves
gamma-distributed; this is a model appropriate for overdispersed count data with actual
variability var(y) exceeding that under the Poisson model (Molenberghs et al., 2007).
Suppose the second stage prior is μi ~ Ga(α,β), namely,

p( mi |a , b ) = mia -1e - bmi b a /G(a ),

and further that α ~ E(A) (namely, α is exponential with parameter A), and β ~ Ga(B,C)
where A, B, and C are preset constants. So the posterior density p(θ|y) of q = ( m1 ,..mn , a , b )
, given y, is proportional to

∏e ∏m
n
e − Aa b B −1e − C b  − mi
miyi   b a /Γ(a)  a − 1 − bmi
i e 
 (1.6)
 i   i 
where all constants (such as the denominator yi! in the Poisson likelihood, as well as the
inverse marginal likelihood k) are combined in a proportionality constant.
It is apparent from inspecting (1.6) that the full conditional densities of μi and β are also
gamma, namely,

mi ∼ Ga( yi + a , b + 1),

and

 
b ~ Ga  B + na , C +

∑ i
mi  ,

respectively. The full conditional density of α, also obtained from inspecting (1.6), is

∏m
n
p(a| y , b , m) ∝ e − Aa  b a /Γ(a)  i
a −1 
.
 i 
This density is non-standard and cannot be sampled directly (as can the gamma densities
for μi and β). Hence, a Metropolis or Metropolis–Hastings step can be used for updating it.

Example 1.1 Estimating Normal Density Parameters via Metropolis

To illustrate Metropolis sampling in practice using symmetric proposal densities,
consider n = 1000 values yi generated randomly from a N(3,25) distribution, namely a
Normal with mean μ = 3 and variance σ2 = 25. Note that, for the particular set.seed used,
the average sampled yi is 2.87 with variance 24.87. Using the generated y, we seek to
estimate the mean and variance, now treating them as unknowns. Setting θ = (μ,σ2), the
likelihood is

n
 ( y i − m)2 
∏
1
p( y|q ) = exp  − .
i =1
s 2p  2s 2 
12 Bayesian Hierarchical Models

Assume a flat prior for μ, and a prior p(s ) ∝ 1/s on σ; this is a form of noninformative
prior (see Albert, 2007, p.109). Then one has posterior density

n
 ( y i − m)2 
∏ exp  −
1
p(q|y ) ∝ .
s n+1
i =1
2s 2 

with the marginal likelihood and other constants incorporated in the proportionality
sign.
Parameter sampling via the Metropolis algorithm involves σ rather than σ2, and uni-
form proposals. Thus, assume uniform U(−κ,κ) proposal densities around the current
parameter values μ(t) and σ(t), with κ = 0.5 for both parameters. The absolute value of
s (t ) + U( − k , k) is used to generate σcand. Note that varying the lower and upper limit of
the uniform sampling (e.g. taking κ = 1 or κ = 0.25) may considerably affect the accep-
tance rates.
An R code for κ = 0.5 is in the Computational Notes [1] in Section 1.14, and uses the
full posterior density (rather than the full conditional for each parameter) as the tar-
get density for assessing candidate values. In the acceptance step, the log of the ratio
p( y|q cand )p(q cand )
is compared to the log of a random uniform value to avoid computer
p( y|q (t ) )p(q (t ) )
over/underflow. With T = 10000 and B = 1000 warmup iterations, acceptance rates for
the proposals of μ and σ are 48% and 35% respectively, with posterior means 2.87 and
4.99. Other posterior summary tools (e.g. univariate and bivariate kernel density plots,
effective sample sizes) are included in the R code (see Figure 1.1 for a plot of the pos-
terior bivariate density). Also included is a posterior probability calculation to assess
Pr(μ < 3|y), with result 0.80, and a command for a plot of the changing posterior expec-
tation for μ over the iterations. The code uses the full normal likelihood, via the dnorm
function in R.

5.3 10

5.2
8

5.1
6
sigma

5.0

4
4.9

2
4.8

4.7 0
2.6 2.8 3.0 3.2 3.4
mu

FIGURE 1.1
Bivariate density plot, normal density parameters.
Bayesian Methods for Complex Data 13

Example 1.2 Extended Logistic with Metropolis Sampling

Following Carlin and Gelfand (1991), consider an extended logistic model for beetle
mortality data, involving death rates πi at exposure dose wi. Thus, for deaths yi at six
dose points, one has

y i ∼ Bin( ni , p(wi )),

p(wi ) = [exp( zi ) /(1 + exp( zi )]m1 ,

zi = (wi − m)/s ,

where m1 and σ are both positive. To simplify notation, one may write V = σ2.
Consider Metropolis sampling involving log transforms of m1 and V, and separate
univariate normal proposals in a Metropolis scheme. Jacobian adjustments are needed
in the posterior density to account for the two transformed parameters. The full poste-
rior p( m, m1 , V |y ) is proportional to

p(m1 )p( m)p(V ) ∏[p(w )]

i
i
yi
(1 − p(wi )]ni − yi

where p(μ), p(m1) and p(V) are priors for μ, m1 and V. Suppose the priors p(m1) and p(μ)
are as follows:

m1 ∼ Ga( a0 , b0 ),

m ∼ N(c0 , d02 ),

where the gamma has the form

b a a -1 - b x
Ga( x|a , b ) = x e .
G(a )
Also, for p(V) assume

V ∼ IG(e0 , f 0 ),

where the inverse gamma has the form

b a -(a +1) - b /x
IG( x|a , b ) = x e .
G(a )

The parameters ( a0 , b0 , c0 , d0 , e0 , f 0 ) are preset. The posterior is then proportional to

  m − c0   −( e0 + 1) − f0 /V
2

(m1a0 − 1e − b0m1 ) exp  −0.5 

  d0  
 V e ∏[p(w )]
i
i
yi
(1 − p(wi )]ni − yi .

Suppose the likelihood is re-specified in terms of parameters q1 = m, q2 = log(m1 ) and

θ3 = log(V). Then the full posterior in terms of the transformed parameters is propor-
tional to

 ∂m1   ∂V 
 ∂q   ∂q  p( m)p(m1 )p(V )
2 3
∏[p(w )]
i
i
yi
(1 − p(wi )]ni − yi .
14 Bayesian Hierarchical Models

One has (∂m1/∂q2 ) = e q2 = m1 and (∂V/∂q3 ) = e q3 = V . So, taking account of the param-
eterisation (θ1,θ2,θ3), the posterior density is proportional to

  m − c0   − e0 − f0 /V
2

(m1a0 e − b0m1 ) exp  −0.5 

  d0  
 V e ∏[p(w )]
i
i
yi
(1 − p(wi )]ni − yi .

The R code (see Section 1.14 Computational Notes [2]) assumes initial values for μ = θ1
of 1.8, for θ2 = log(m1) of 0, and for θ3 = log(V) of 0. Preset parameters in the prior den-
sities are (a0 = 0.25, b0 = 0.25, c0 = 2, d0 = 10, e0 = 2.000004, f0 = 0.001). Two chains are run
with T = 100000, with inferences based on the last 50,000 iterations. Standard devia-
tions in the respective normal proposal densities are set at 0.01, 0.2, and 0.4. Metropolis
updates involve comparisons of the log posterior and logs of uniform random variables
{U h(t ) , h = 1,… , 3} .
Posterior medians (and 95% intervals) for {μ,m1,V} are obtained as 1.81 (1.78, 1.83), 0.36
(0.20,0.75), 0.00035 (0.00017, 0.00074) with acceptance rates of 0.41, 0.43, and 0.43. The pos-
terior estimates are similar to those of Carlin and Gelfand (1991). Despite satisfactory
convergence according to Gelman–Rubin scale reduction factors, estimation is beset
by high posterior correlations between parameters and low effective sample sizes. The
cross-correlations between the three hyperparameters exceed 0.75 in absolute terms,
effective sample sizes are under 1000, and first lag sampling autocorrelations all exceed
0.90.
It is of interest to apply rstan (and hence HMC) to this dataset (Section 1.10) (see Section
1.14 Computational Notes [3]). Inferences from rstan differ from those from Metropolis
sampling estimation, though are sensitive to priors adopted. In a particular rstan esti-
mation, normal priors are set on the hyperparameters as follows:

m ∼ N(2, 10),

log(m1 ) ∼ N(0, 1),

log(s ) ∼ N(0, 5).

Two chains are applied with 2500 iterations and 250 warm-up. While estimates for μ
are similar to the preceding analysis, the posterior median (95% intervals) for m1 is now
1.21 (0.21, 6.58), with the 95% interval straddling the default unity value. The estimate
for the variance V is lower. As to MCMC diagnostics, effective sample sizes for μ and m1
are larger than from the Metropolis analysis, absolute cross-correlations between the
three hyperparameters in the MCMC sampling are all under 0.40 (see Figure 1.2), and
first lag sampling autocorrelations are all under 0.60.

1.8 Metropolis–Hastings Sampling
The Metropolis–Hastings (M–H) algorithm is the overarching algorithm for MCMC
schemes that simulate a Markov chain θ(t) with p(θ|y) as its stationary distribution.
Following Hastings (1970), the chain is updated from θ(t) to θcand with probability
Bayesian Methods for Complex Data 15

FIGURE 1.2
Posterior densities and MCMC cross-correlations, rstan estimation of beetle mortality data.

æ p(q cand |y )q(q (t ) |q cand ) ö

a (q cand |q (t ) ) = min ç 1, (t ) ÷
,
è p(q |y )q(q cand |q ) ø
(t )

where the proposal density q (Chib and Greenberg, 1995) may be non-symmetric, so
that q(q cand |q (t ) ) does not necessarily equal q(q (t ) |q cand ). q(q cand |q (t ) ) is the probability (or
density ordinate) of θcand for a density centred at θ(t), while q(q (t ) |q cand ) is the probabil-
ity of moving back from θcand to the current value. If the proposal density is symmetric,
with q(q cand |q (t ) ) = q(q (t ) |q cand ) , then the Metropolis–Hastings algorithm reduces to the
Metropolis algorithm discussed above. The M–H transition kernel is

K (q cand |q (t ) ) = a (q cand |q (t ) )q(q cand |q (t ) ),

for q cand ¹ q (t ) , with a nonzero probability of staying in the current state, namely
16 Bayesian Hierarchical Models

ò
K (q (t ) |q (t ) ) = 1 - a (q cand |q (t ) )q(q cand |q (t ) )dq cand .

Conformity of M–H sampling to the requirement that the Markov chain eventually sam-
ples from π(θ) is considered by Mengersen and Tweedie (1996) and Roberts and Rosenthal
(2004).
If the proposed new value θcand is accepted, then θ(t+1) = θcand, while if it is rejected the next
state is the same as the current state, i.e. θ(t+1) = θ(t). As mentioned above, since the target
density p(θ|y) appears in ratio form, it is not necessary to know the normalising constant
k = 1/p(y). If the proposal density has the form

q(q cand |q (t ) ) = q(q (t ) - q cand ),

then a random walk Metropolis scheme is obtained (Albert, 2007, p.105; Sherlock et al.,
2010). Another option is independence sampling, when the density q(θcand) for sampling
candidate values is independent of the current value θ(t).
While it is possible for the target density to relate to the entire parameter set, it is typi-
cally computationally simpler in multi-parameter problems to divide θ into C blocks or
components, and use the full conditional densities in componentwise updating. Consider
the update for the hth parameter or parameter block. At step h of iteration t + 1 the preced-
ing h − 1 parameter blocks are already updated via the M–H algorithm, while qh +1 , … , qC
are still at their iteration t values (Chib and Greenberg, 1995). Let the vector of partially
updated parameters apart from θh be denoted

q[(ht]) = (q1(t +1) ,q 2(t +1) ,… ,q h(t-+11) ,q h(t+)1 ,… ,qC(t ) ),

The candidate value for θh is generated from the hth proposal density, denoted
qh (q h ,cand |q h(t ) ) . Also governing the acceptance of a proposal are full conditional densities
p h (q h(t ) |q[(ht]) ) µ p( y|q h(t ) )p(q h(t ) ) specifying the density of θh conditional on known values of
other parameters θ[h]. The candidate value θh,cand is then accepted with probability

æ p( y|q h ,cand )p(q cand )q(q h(t ) |q cand ) ö

a = min ç 1, ÷. (1.7)
è p( y|q h(t ) )p(q h(t ) )q(q cand |q h(t ) ) ø

Example 1.3 Normal Random Effects in a Hierarchical Binary Regression

To exemplify a hierarchical Bayes model involving a three-stage prior, consider binary
data yi ~ Bern(pi) from Sinharay and Stern (2005) on survival or otherwise of n = 244
newborn turtles arranged in J = 31 clutches, numbered in increasing order of the average
birthweight of the turtles. A known predictor is turtle birthweight xi. Let Ci denote the
clutch that turtle i belongs to. Then to allow for varying clutch effects, one may specify,
for cluster j = Ci, a probit regression with

pi |b j = Φ( b1 + b2 xi + b j ),

where {b j ∼ N(0, 1 / tb ), j = 1,… , J }. It is assumed that bk ∼ N(0, 10) and tb ∼ Ga(1, 0.001).
A Metropolis–Hastings step involving a gamma proposal is used for the random
effects precision τb, and Metropolis updates for other parameters; see Section 1.14
Computational Notes [3]. Trial runs suggest τb is approximately between 5 and 10, and a
Discovering Diverse Content Through
Random Scribd Documents
belong to Christ. For them we shall have, if we be true to our faith,
that mighty love which passeth in excellence both faith and hope.
But the exhortation of St. Peter speaks in this wise: Ye who hold
your brethren in Christ unspeakably dear, do not allow that love to
suffice, to swallow up all regard for other men. They also need your
thoughts, your help. The heathen, the unbelievers—these have the
strongest possible claim, even their great need. And so with the
other pair of precepts. Ye who fear God, which is your foremost
duty, do not let that fear lessen your willingness to do honour to
your earthly rulers. The feelings toward God and the king differ in
character and in degree, but both have their place in proper share in
the heart of the true servant of Christ.
 VIII
CHRISTIAN SERVICE
"Servants, be in subjection to your masters, with all fear; not
only to the good and gentle, but also to the froward. For this is
acceptable, if for conscience toward God a man endureth griefs,
suffering wrongfully. For what glory is it, if, when ye sin, and are
buffeted for it, ye shall take it patiently? but if, when ye do well
and suffer for it, ye shall take it patiently, this is acceptable with
God. For hereunto were ye called: because Christ also suffered
for you, leaving you an example, that ye should follow His
steps: who did no sin, neither was guile found in His mouth:
who, when He was reviled, reviled not again; when He suffered,
threatened not; but committed Himself to Him that judgeth
righteously; who His own self bare our sins in His body upon the
tree, that we, having died unto sins, might live unto
righteousness; by whose stripes ye were healed. For ye were
going astray like sheep, but are now returned unto the
Shepherd and Bishop of your souls."—1 Peter ii. 18-25.

The Gospel history shows very clearly that during our Lord's lifetime
His followers were drawn largely from the ranks of the poor. It was
fitting that He who had been proclaimed in prophecy as "the Servant
of the Lord" should enter the world in humble estate; and, from the
lowly position of the virgin-mother and her husband, the life of Jesus
for thirty years must have been spent in comparative poverty and
amid poor surroundings. The major part of His chosen disciples were
fisherfolk and such-like. And though we read of the wife of Herod's
steward among the women who ministered unto Him and of the
richer Joseph of Arimathæa as a secret disciple, these are marked
exceptions. To the poor His Gospel was preached, and among the
poor it first made its way. The question of the chief priests, "Hath
any of the rulers believed on Him, or the Pharisees?" (John vii. 48),
tells its own tale, as does also the significant record, "The common
people heard Him gladly" (Mark xii. 37).
It need not therefore much surprise us if St. Peter, now that he
begins to classify his counsels, addresses himself first to "household
servants": Servants, be in subjection to your masters, with all fear.
We have, however, to bear in mind, as we consider the Apostle's
exhortation, that most of those whom he addresses were slaves.
They had no power of withdrawing themselves, though their service
should prove burdensome and grievous. St. Paul, in writing to the
same class, nearly always employs the word which means
"bondservants." Yet his counsel agrees with St. Peter's. Thus he
exhorts that their service be "with fear and trembling" (Eph. vi. 5);
in Col. iii. 22, "Obey in all things them that are your masters." And to
Timothy and Titus it is given as a part of their charge to "exhort
servants to be in subjection to their own masters and to be well-
pleasing to them in all things" (1 Tim. vi. 1; Titus ii. 9).
When St. Peter and St. Paul wrote, this slave population was
everywhere very numerous. Gibbon calculates that in the reign of
Claudius the slaves were at least equal in number to the free
inhabitants of the Roman world; Robertson places the estimate
much higher. These formed, then, a very large share of the public to
which the first preachers had to appeal, and we can understand the
importance to the Christian cause of the behaviour of these humble,
but doubtless most numerous, members of the society. Their lives
would be a daily sermon in the houses of their masters. Hence the
very earnest exhortations addressed to them that by their conduct
they should adorn the doctrine of God our Saviour in all things; that
they should count their masters worthy of all honour; that the name
of God and of the doctrine be not blasphemed; that they should be
in subjection with all fear. Everything in the New Testament
concerning slaves goes to show that they were a most important
factor in the early Christian societies.
Men wonder nowadays that there is so little said by any of the
Apostles about freeing slaves from their bondage. The best men in
those times and long before appear to have regarded slavery as one
of the institutions with which they were bound to rest content. It
flourished everywhere; it was countenanced in the Scriptures of the
older dispensation. Eleazar was Abraham's slave, and the Law in
many passages contemplates the possession by Israelites of persons
who were bought with their money. Hence we find no remonstrance
against slave-holding in the New Testament writings, only advice to
those who were in such bondage to cultivate a spirit which would
render it less galling and to strive that by their behaviour the cause
of Christ might be advanced. St. Paul represents the ideas of his age
when, writing to the Corinthians, he says, "Wast thou called being a
bondservant? Care not for it; but if thou canst be made free, use it
rather" (1 Cor. vii. 21). Freedom was worth having, but any heroic
effort to get rid of the yoke is not encouraged in the Epistles. Yet it
must have been a lot which called for the exercise of much moral
strength to make it bearable. Even from the house of the Christian
Philemon the slave Onesimus found cause to run away. But St. Paul
in his letter admits no right on the slave's part to take this course.
With the Apostle there is no question that the first duty is to go back
to his master. All that he urges is that the common profession of
Christianity by slave and master ought to, and doubtless would,
alleviate the conditions of servitude. There were in Christianity, as
time has shown, germs which would fructify, a spirit which some day
would strike on the chains of slaves. But the vision of such a time
had not dawned either for St. Paul or St. Peter. Christ has overcome
the world in many other matters beside slavery. It is only that
Christians are so tardy in awaking to the fulness of His lessons.
So in apostolic days the rights and claims of slave-masters were
looked upon as indisputable. Be subject, not only to the good and
gentle, but also to the froward. There is to be no resistance, no
lapse in duty. About service rendered to good masters there might
be little apprehension, but even here St. Paul finds occasion for
warning. "They that have believing masters," he says, "let them not
despise them because they are brethren" (1 Tim. vi. 2). Christian
freedom was not without its dangers in many forms, especially to
minds wherein liberty was a strange idea. But froward masters are
to be faithfully served likewise, and care is to be taken withal to
remove every occasion for their frowardness. The apostolic lesson is
to make suffering endurable, noble, acceptable to God, by seeing
that it be always undeserved. How strange a doctrine this in the
eyes of the world! The rule of purely human conduct would be just
the opposite. If wrong be undeserved, rebel at once. Christianity
supplies a motive for the contrary course: conscience toward God.
The world's spirit is not His spirit, and to have praise with Him
should be the Christian's single aim. Men can at times be patient
when rebuke is deserved, but the world sees that that deserves no
credit. "What thank have ye?" they cry. But they give no praise for
the bearing of unmerited rebuke.
The world counts such conduct weakness, and is still far from
comprehending the Divineness of the virtue of yielding patiently to
wrong. God has long been teaching the lesson, but it has been
slowly learnt. He chose the milder, timid Jacob rather than the fiery
Esau. Both had faults in multitude. With the world Esau is oft the
favourite. At a later day He stamps with approval the noble mercy of
David in sparing Saul, while round Daniel and his companions in
Babylon there gathers something of a halo of New Testament
sanctity by reason of the noble confession which they made under
persecution. These are chapters in the Divine lesson-book. Such
lives marked stages in the preparation for the Servant of the Lord.
Men, if they would have hearkened, were being trained to estimate
such a character at God's value. Now Christ's example is before us,
and we are bidden to follow it.
For hereunto were ye called. Strange invitation to be dictated by
love, a call to suffering! And yet the Master at first promises nothing
else to His followers: "If any man would come after Me, let him deny
himself, and take up his cross, and follow Me" (Matt. xvi. 24). And
what can a Christian wish for but to be like Christ? And the very
reason given ought to make us love the cross. We are called unto
suffering because Jesus suffered for us, leaving us an example that
we should follow His steps. He has trodden the hard road, the
winepress of the wrath of God, alone and for men. At this point the
Apostle begins to apply to Christ Isaiah's description of the suffering
"Servant of the Lord," "who did no sin, neither was guile found in His
mouth" (Isa. liii.). But soon the memory of the scenes he had
witnessed is present with him; and his words, though holding to the
spirit of Isaiah's picture, become a description of what he himself
had seen and heard when Jesus was taken and crucified: Who,
when He was reviled, reviled not again; when He suffered,
threatened not, but committed Himself to Him that judgeth
righteously. How the brief words sum up and recall the dark history
—Caiaphas, Pilate, and Herod; the mockery, the scourging, the
railing crowd, the dying Jesus, and the parting prayer, "Father, into
Thy hands I commend My spirit."
So far the Apostle speaks of the example of Christ, which, though far
above and beyond us, we are exhorted and called on to follow. And
there are many who will go with him thus far who value our Lord's
work only for its lofty example. Indeed, it is characteristic of those
who deny the mediatorial office of Christ to be loudest in magnifying
the grandeur of His character. To His good works, His love for men,
His spotless life, His noble lessons, they accord untiring praise, as
though thereby they would atone for denying Him that office which
is more glorious still. But St. Peter stops at no such half-way house.
He knows in whom he has believed, knows Him for the Son of the
living God, a Teacher with whom were the words of eternal life. So
in pregnant words he sets forth the doctrine of the Atonement as
the end of Christ's suffering: Who His own self bare our sins in His
own body upon the tree, that we, having died unto sins, might live
unto righteousness. He bare our sins. The words tell of something
beyond our powers to comprehend; but some light is shed on them
by a kindred passage (Matt. viii. 17), where the Evangelist applies to
the work of Jesus those other words from Isa. liii., "Himself took our
infirmities and bare our sicknesses." The narrative in the Gospel has
just recorded how Jesus wrought many miracles. First a leper was
healed, then the centurion's servant, next Simon's wife's mother, and
afterwards many sick and demoniacs beside. There is no record here
of the effect produced on Jesus Himself by these exhibitions of
miraculous power, but from other passages in the Gospels we do
find that He was conscious in Himself of a demand on His power
when such cures were wrought. Thus we are told, at the cure of the
woman with the issue, that Jesus perceived in Himself that the
power proceeding from Him had gone forth (Mark v. 30); and again
when many were cured, that "power came forth from Him and
healed them all" (Luke vi. 19). Of the woman Jesus says expressly,
"Thy faith hath made thee whole"; and the manifestation of
eagerness to touch Jesus is a sign of the faith of the others whom
the Divine power blessed with health.
The Bible recognises everywhere the analogy between sin and
sickness. May we not trace some analogy between the Lord's works
of healing and that mightier deliverance from sin won by Christ upon
the cross, an analogy which may help, if but a little, to give meaning
to the bearing by Christ of human sins? A power went forth when
the sick were healed; and through that imparted power they were
restored to health, faith being the pathway which brought the Divine
virtue to their aid. Thus Jesus bore their diseases and took them
away. Look through this figure on the work of our redemption. Christ
has borne the burden of sin. He has died for sin that men may die
from sin, that sin may be slain in us, the fell disease healed by the
power of His suffering. We cannot comprehend what was done for
the sick when Christ was on earth, nor what is wrought for sinners
by His grace in heaven. Those alone who reap the blessing know its
certainty; and they can but say, as the blind man whose sight was
restored, "One thing I know: that, whereas I was blind, now I see"
(John ix. 25).
To this teaching, that Christ's suffering wrought man's rescue, St.
Peter adds emphasis by another quotation from that chapter of
Isaiah which he has so much in mind: by whose stripes ye were
healed. Christ was stricken, and God grants to His sufferings a power
to heal the souls of those whom He loves because they strive to love
Him. Healing through wounds! Soundness through that which speaks
only of injury! Mysterious dispensation! But long ago it had been
foreshadowed, and shown also how little connexion there was to be,
except through faith, between the remedy and the disease. Those
who were bitten of the serpents in the wilderness gazed on the
brazen serpent, and were healed. In the dead brass was no virtue,
but God was pleased to make of it a speaking sacrament; so has it
pleased Him to give healing of sins to those who by faith appropriate
the sacrifice on Calvary. Christ has claimed the type for Himself: "I, if
I be lifted up from the earth, will draw all men unto Myself" (John
xii. 32).
And now, as is so often his wont, St. Peter varies the figure. The
wounded sinner finding cure becomes the wandering sheep that has
been brought back into the fold: For ye were going astray like sheep,
but are now returned unto the Shepherd and Bishop of your souls.
But the message, the teaching, the love, is all the same. He who
before was the great Exemplar, whose footsteps we should follow, is
now the Shepherd, the Good Shepherd, who goes before His sheep.
This Shepherd has been a Sufferer, too. He has given Himself up as
a prey to the wolves that His flock might be saved. Now, with a
voice of love, He calls His sheep by name; and hearing, they follow
Him.
But He is more than this. Brought within the fold, the sheep still
need His care; and it is freely given. He is the Bishop, the Overseer,
the Watchman for His people's safety, who, having gathered them
within the fold, tends them with constant watchfulness. The figure
passes over thus into the reality in the Apostle's closing words. The
cure which the great Healer desires to accomplish is in the souls of
men. For them His care is bestowed, first to bring them safe out of
the way of evil, then for ever to keep them under the sheltering care
of His abundant love.
 IX
CHRISTIAN WIVES AND HUSBANDS
"In like manner, ye wives, be in subjection to your own
husbands; that, even if any obey not the word, they may
without the word be gained by the behaviour of their wives;
beholding your chaste behaviour coupled with fear. Whose
adorning let it not be the outward adorning of plaiting the hair,
and of wearing jewels of gold, or of putting on apparel; but let
it be the hidden man of the heart, in the incorruptible apparel of
a meek and quiet spirit, which is in the sight of God of great
price. For after this manner aforetime the holy women also, who
hoped in God, adorned themselves, being in subjection to their
own husbands: as Sarah obeyed Abraham, calling him lord:
whose children ye now are, if ye do well, and are not put in fear
by any terror.
"Ye husbands, in like manner, dwell with your wives according to
knowledge, giving honour unto the woman, as unto the weaker
vessel, as being also joint heirs of the grace of life; to the end
that your prayers be not hindered."—1 Peter iii. 1-7.

The Apostle gave at first (ii. 13) the rule of Christian submission
generally; then proceeded to apply it to the cases of citizens and of
servants. In the same way he now gives injunctions concerning the
behaviour of wives and husbands. The precept with which he began
holds good for them also. In like manner, ye wives, be in subjection
to your own husbands. The life and teaching of Jesus had wrought a
great change in the position of women, a change which can be
observed from the earliest days of Christianity. We can gather in
what estimation women were generally held among the Jews at that
time from the expression used in the account of our Lord's interview
with the woman of Samaria. There it is said (John iv. 27) that the
disciples marvelled that Jesus was talking with a woman. Such a
feeling must afterwards have been entirely dispelled, for all through
the earthly life of Christ we find Him attended by women who
ministered unto Him; we read of His close friendship with Mary and
Martha, and are told, at the time of His death (Matt. xxvii. 55), that
many women beheld the Crucifixion afar off, having followed Him
from Galilee. Women were the earliest visitors to the tomb on the
great Easter morning, and to them, among the first (Luke xxiv. 22),
was the Lord's resurrection made known.
We are not surprised, therefore, in the history of the infant Church,
to read (Acts i. 14) that women were present among the disciples
who waited at Jerusalem for the promise of the Father, nor to learn
how the daughters of Philip the evangelist (Acts xxi. 9) took a share
in the labours of their father for the cause of Christ, or that Priscilla
(Acts xviii. 26), equally with her husband, was active in Christian
good offices. Other examples occur in the Acts of the Apostles:
Dorcas, Lydia, and the mother of Timothy; and the constant mention
of women which we find in the salutations with which St. Paul
concludes his letters makes it clear how large a part they played in
the early propagation of the faith. "Fellow-workers," "servants of the
Church," "labourers in the Lord," are among the terms which the
Apostle applies to them; and we know from the Pastoral Epistles
what help the primitive Church derived from the labours of its
deaconesses and widows.
To be occupied in such duties was sure to give to women an
influence which they had never possessed before; and the women
converts, in countries such as these Asiatic provinces, were exposed
to the same sort of danger which beset the slave population at their
acceptance of the Christian faith. They might begin to think meanly
of others, even of their own husbands, if they were still content to
abide in heathenism. Such women might incline at times to take
counsel for their life's guidance with Christian men among the
various congregations to which they belonged and to set a value on
their advice above any which they could obtain from their own
husbands. They might come to entertain doubts also whether they
ought to maintain the relations of married life with their heathen
partners. With the knowledge that such cases might occur, St. Peter
gives his lesson. And as in the case of slaves, so here, he gives no
countenance to the idea that to become a Christian breaks off
previous relations. Wives, though they have accepted the faith, have
wifely duties still. Like Christian citizens living in a heathen
commonwealth, they are not by religion released from their
previously contracted obligations; they are to abide in their estate,
and use it, if it may be done, for the furtherance of the cause of
Christ. Be in subjection to your own husbands; they have still their
claim on your duty.
There is much gentleness in the Apostle's next words. He knows that
there may arise cases where believing wives have husbands who are
heathen. But he speaks hopefully, as thinking they would not be of
frequent occurrence: even if any obey not the word. Wives,
especially if they be of such a character as the Apostle would have
them be, could not have been won to the faith of Christ without
much converse with their husbands on so deep a subject; and the
word which was working effectually in the one would often have its
influence with the other. It might not always be so. But husbands,
though not obeying the word as yet, are not to be despaired of.
And here we may turn aside to dwell on the tone of hope in which
St. Peter speaks of these husbands who obey not. For the word
ἀπειθοῦντες, by which they are described, is the same that is used in
ii. 18 of those who stumble at the word, being disobedient. The
lesson here given to Christian wives, not to despair of winning their
husbands for Christ, gives warrant for what was said on the former
passage: that the disobedience which causes men to stumble need
not last for ever, nor imply final obduracy and rejection from God's
grace. But this by the way.
The Apostle adds the strongest motive to confirm wives in holding to
their married state: That the husbands may without the word be
gained by the behaviour of their wives: beholding your chaste
behaviour coupled with fear. "Without the word" here means that
there is to be no discussion. They are so to live as to make their
lives a sermon without words, to work conviction without debate;
then, when the victory is won, there will remain no trace of combat:
all will tell of gain, and nothing of loss.
And once again St. Peter uses his special word (ἐποπτέυειν) as he
describes how the husbands shall be affected by the behaviour of
their wives. They shall gaze on it as a mystery, the key to which they
do not possess. The wives in heathen homes must have been
obliged to hear and see many things which were grievous and
distasteful. The husbands could hardly fail to know that it was so. If,
then, they still found wifely regard and respect, wifely submission,
with no assertion of a law of their own, no comparison of the lives of
Christian men with those of their own husbands, if a silent,
consistent walk were all the protest which the Christian wives
offered against their heathen environments, such a life could hardly
fail of its effect. There must be a powerful motive, a mighty,
strengthening power, that enabled women to abide uncomplainingly
in their estate. For this the husbands would surely search, and in
their search would learn secrets to which they were strangers, would
learn how the tongue was restrained where remonstrance might
seem more natural, how pure life was maintained in spite of
temptations to laxity, and the marriage bond exalted with religious
observance even when reverence for the husband was meeting with
no equal return. Such lives would be more powerful than oratory,
have a charm beyond resistance, would win the husbands first to
wonder, then to praise, and in the end to imitation.
And from describing the grace of such a life the Apostle turns to
contrast it with other adornments of which the world thinks highly.
Whose adorning, he says, let it not be the outward adorning of
plaiting the hair, and of wearing jewels of gold, and of putting on
apparel. We can see from the catalogue in Isaiah (iii. 18-23) that the
daughters of Zion in old days had gone to great lengths in this
outside bravery, and provoked the Lord to smite them. These had
forgotten the simplicity of Sarah. But that in the house of Abraham
there were found no such ornaments is hardly to be believed. The
patriarch, who sent (Gen. xxiv. 53) to Rebekah jewels of silver and
jewels of gold, did not leave his own wife unadorned. Nor does the
language of St. Peter condemn Rebekah's bracelets, if they be worn
with Rebekah's modesty. The New Testament does not teach us to
neglect or despise the body. A misrendering in the Authorised
Version, "Who shall change our vile body" (Phil. iii. 21), has long
seemed to lend countenance to such a notion. It is one of the gains
of the Revised Version that we now read in that place, "Who shall
fashion anew the body of our humiliation." Sin has robbed the body
of its primal dignity, but it is to be restored and made like unto the
body of Christ's glory. And He did not despise the body when He
deigned to wear it that He might draw nearer unto us. If these
things be present to our thoughts, we shall seek to bestow on the
body whatever may make it comely. The mischief arises when the
adornment of the outer brings neglect of the inner man, when fine
apparel has for its companions the haughtiness, the stretched-forth
necks, and wanton eyes which Isaiah rebukes. Then it is that it
rightly comes under condemnation. When the jewel is (as Rebekah's
was) the gift of some dear one—a parent, a husband, a near
kinsman—it rouses grateful reminiscences, and may fitly be prized,
and holily worn, and ranked near to the rings of betrothal and of
marriage.
Let these be the feelings which regulate womanly adornment, and it
may be made a part of the culture of the heart, the inner man,
which St. Peter urges the Christian wives to be careful to adorn: Let
your adorning be the hidden man of the heart, in the incorruptible
apparel of a meek and quiet spirit, which is in the sight of God of
great price. All Scripture regards man as of twofold nature, the
outward and the inward, of which the latter is the more precious. He
is a Jew who is one inwardly (Rom. ii. 29); the inward man
delighteth in the law of God (Rom. vii. 22); while the outward man
perishes the inward man may be renewed day by day (2 Cor. iv. 16),
being strengthened with power through God's Spirit. This hidden
man is the centre from which all the strength of Christian life comes.
Let this be rightly adorned, and the outward life will need no strict
rules; there will be no fear of excess, least of all when the inner life
is cared for because it is precious before God. Its pure array passeth
gold and gems, be they ever so beautiful. This is a grace which
never fades, but will flourish through eternity.
The Apostle proceeds to commend it by a noble example. The Old
Testament Scriptures do not dwell largely on the lives of women, but
a study of what is said will oftentimes reveal deeper meaning in the
record and put force into a solitary word. The writer of the Epistle to
the Hebrews couples Sarah with Abraham in the list of heroes and
heroines of faith, and St. Peter from a single word finds a text to
extol the submission which she showed to her husband. He probably
refers to Gen. xviii. 12, where she gives the title of "lord" to
Abraham, as Rachel in another place (Gen. xxxi. 35) does to her
father Laban: For after this manner aforetime the holy women also,
who hoped in God, adorned themselves, being in subjection to their
own husbands: as Sarah obeyed Abraham, calling him lord. A
Scripture example which has more in common with the experience
of the Asian women is the life of Hannah. Her lot, for a time at least,
was as full of grief and disappointment as theirs could be; but her
trust in God was unshaken. Her patience under provocation was
exemplary, while the picture of her home life is one full of touching
affection on the part of both husband and wife; and the mother's
gratitude, when her prayer was granted, is set forth in her noble
hymn of thanksgiving and in the devotion of her child to the service
of the God who had bestowed him. Ruth is another of those holy
women who must have been in St. Peter's thoughts, who, though
not of the house of Israel, manifested virtues in her life which made
her fit to be the ancestress of King David. The Apostle, however,
seems to have had a purpose in his special mention of Sarah. As the
sons of Israel looked back to Abraham and to the covenant sealed
with him, yea, not seldom prided themselves on being his children,
so the daughters of Israel counted themselves as Sarah's daughters
after the flesh. St. Peter now gives them another ground for that
claim. God's promises to Abraham have been fulfilled in Christ, and
so Christian Jewesses are more truly than ever daughters of Sarah.
Whose children ye now are. But to the heathen converts the same
door was opened. They by their faith were now made partakers of
the ancient covenant. They too were become Sarah's daughters. Let
them, one and all, continue in the well-doing which has been
commended; let it be seen in the daily round (ἀναστροφὴ) of their
lives, led in quietness and humility. The excessive love of adornment
against which they are warned marks a condition of boldness and
unrest. But unrest may enter into the other actions of their life. Their
behaviour is to be coupled with fear and reverence, but it should
eschew everything which partakes of flighty irregularity. It should be
steady and consistent, running into no extremes either of humiliation
or the contrary. Do well, and be not put in fear by any terror.
The Apostle now addresses Christian husbands. In his counsel to
subjects and slaves he has not dwelt on the duties of rulers and
masters. Perhaps he judged it unlikely that his letter would come to
the hands of many such, or it may be he thought the lessons which
he had to give were more needed by the subject people, if Christ's
cause were to be furthered. But with husbands and wives life has of
necessity a great deal in common, and the one partner can hardly
receive counsel which is not of interest to the other. To the wives the
Apostle spake as though examples of unbelieving husbands might be
rare. Christian husbands with unbelieving wives he hardly seems to
contemplate. We know from St. Paul (1 Cor. vii. 16) that there were
such. But doubtless heathen wives hearkened to Christian husbands
more readily than heathen husbands to their Christian wives. The
husbands are to use their position as heads of their wives with
judgement and discretion: Dwell with your wives according to
knowledge. The knowledge of which St. Peter speaks is not religious,
godly, Christian knowledge, but that foresight and thoughtfulness
which the responsibility of the husband calls for. He will understand
what things for his wife's sake he should do or leave undone. This
knowledge, which results in considerate conduct towards her, will
manifest itself in Christian chivalry. The woman is physically the
feebler of the two. No burden beyond her powers will be laid upon
her; and by reason of her weaker nature regard and honour will be
felt to be her due. For the woman is the glory of the man (1 Cor. xi.
7). Such observance will not degenerate into undue adulation nor
foolish fondness, apt to foster pride and conceit, but will be inspired
by the sense that in God's creation neither is the man without the
woman, nor the woman without the man.
But beyond and above these daily graces of domestic and social
intercourse, the Apostle would have husband and wife knit together
by a higher bond. They are joint heirs of the grace of life. Both are
meant to be partakers of the heavenly inheritance, and such
participation makes their chief duty here to be preparation for the
life to come. Those who are bound together not by wedlock only, but
by the hope of a common salvation, will find a motive in that
thought to help each other in life's pilgrimage, each to shun all that
might cause the other to stumble: That your prayers be not
hindered. They are fellow-travellers with the same needs. Together
they can bring their requests before God, and where the two join in
heart and soul Christ has promised to be present as the Third. And
in praying they will know one another's necessities. This is the
grandest knowledge the husband can attain to for the honouring of
his wife; and using it, he will speed their united supplications to the
throne of grace, and the union of hearts will not fail of its blessing.
 X
THEY WHO BLESS ARE BLESSED
"Finally, be ye all like-minded, compassionate, loving as
brethren, tender-hearted, humble-minded: not rendering evil for
evil, or reviling for reviling; but contrariwise blessing; for
hereunto were ye called, that ye should inherit a blessing. For
he that would love life, and see good days, let him refrain his
tongue from evil, and his lips that they speak no guile: and let
him turn away from evil, and do good; let him seek peace, and
pursue it. For the eyes of the Lord are upon the righteous, and
His ears unto their supplication: but the face of the Lord is upon
them that do evil. And who is he that will harm you, if ye be
zealous of that which is good? But and if ye should suffer for
righteousness' sake, blessed are ye: and fear not their fear,
neither be troubled; but sanctify in your hearts Christ as Lord:
being ready always to give answer to every man that asketh you
a reason concerning the hope that is in you, yet with meekness
and fear: having a good conscience; that, wherein ye are
spoken against, they may be put to shame who revile your good
manner of life in Christ."—1 Peter iii. 8-16.

The Apostle now ceases from his special admonitions, and enforces
generally such qualities and conduct as must mark all who fear the
Lord. Finally, he says—and the word may indicate the close of his
counsels; but the virtues which he inculcates are of so important a
character that he may very well intend them as the apex and crown
of all his previous advice—be ye all like-minded, compassionate,
loving as brethren, tender-hearted, humble-minded. St. Peter has
here grouped together a number of epithets of which all but one are
only used in the New Testament by himself, and they are of that
graphic character which is so conspicuous in all the Apostle's
language. Like-minded. If the word be not there, the spirit is largely
exemplified in the early history of the Church. How often we hear
the phrase "with one accord" in the opening chapters of the Acts.
Thus the disciples continued in prayer (i. 14); thus they went daily to
the Temple (ii. 46); thus they lifted up their voices to God (iv. 24),
for all they that believed were of one heart and one soul (iv. 32).
Such lives exhibit harmony of thought, the same aim and purpose.
The men may not, will not, always use the same means or follow the
same methods, but they will all be seeking one result. Such unity is
worth more than uniformity. Compassionate. This feeling St. Paul
describes (Rom. xii. 15) as rejoicing with them that do rejoice and
weeping with them that weep. For the παθήματα of this life are not
always sorrowful, though the best of them are not worthy to be
compared with the glory that shall be revealed (Rom. viii. 18).Loving
as brethren. The sense of the brotherhood of Christians is strongly
marked in all the New Testament Scriptures. It is the name by which
our Lord claims fellowship with men, being not ashamed to call them
brethren. It is the designation of the Christian body from the first
(Matt. xxiii. 8), is constantly found in the Acts and the Epistles (Acts
vi. 3, ix. 30, xi. 29), and has been used of the Church in every age,
marking how as one family we dwell in Him. Next comes the word
which is not St. Peter's alone: Tender-hearted. St. Paul has it (Eph.
iv. 32), but it is no Greek notion. It was a Jewish idea that deep
feeling was closely connected with some of the organs of the body;
and in the Old Testament, as in the story of Joseph (Gen. xliii. 30)
and elsewhere (1 Kings iii. 26), we come upon such phrases as "His
bowels did yearn upon his brother." This Hebrew notion the LXX. has
conveyed into Greek by the word which St. Peter here uses, and
which those translators had used and consecrated long before. For
them so exalted was the thought contained in it that they employ it
in the prayer of Manasses (ver. 7) to express the tenderness of God
towards the penitent, the yearning love of the Father, who sees the
prodigal afar off, and has compassion. Humble-minded. This word
and those akin to it are almost a New Testament creation. The
heathen had no admiration for the temper it expresses, and where
they do use the word it is in a bad sense as signifying "cowardly"
and "mean-spirited." Before Christ none had taught, "He that is
greatest among you shall be your servant" (Matt. xxiii. 11).
It is manifest that if such harmony, kind feeling, attachment,
affection, and humility flourished among believers, these virtues
would put discord to the rout, and leave no occasion for rending the
oneness of the Christian body. They would also be proof against evil
from without, both in deed and speech, neither tempted to render
evil for evil in their actions nor reviling for reviling in their words.
They have a duty to the world, and cannot thus belie their Christian
profession. They are called to adorn the doctrine of their Saviour,
and the Master's sermon has among its prominent precepts "Bless
them that curse you." This is the spirit of St. Peter's exhortation, But
contrariwise blessing; that is, Be ye of those who bless. For there is
a law of recompense with God in good things as in evil; the blessers
shall be blessed: For hereunto were ye called, that ye should inherit
a blessing. It is as though he urged them thus: Ye were aforetime
enemies of God; but ye have been made partakers of His heavenly
calling (Heb. iii. 1), that ye may come to blessing. This should move
you to bless your enemies. And more than this, the servant of God
may receive no blessing from the world, may get curses for his
blessing; but yet he knows where to flee for consolation. He can
pray with the Psalmist, "Let them curse, but bless Thou" (Psalm cix.
28), conscious that the Lord will stand at the right hand of the
needy.
The psalmists knew much of such trials, and it is from the words of
one of them (Psalm xxxiv. 12-16) that St. Peter enforces his own
lesson. It is a psalm full of the knowledge of the trials of God's
servants: "Many are the afflictions of the righteous"; but it is rich
also in plenitude of comfort: "The Lord delivereth him out of them
all." The father of long ago teaches thus to his children the fear of
the Lord: He that would love life, and see good days, let him refrain
his tongue from evil, and his lips that they speak no guile: and let
him turn away from evil, and do good; let him seek peace, and
pursue it. For the eyes of the Lord are upon the righteous, and His
ears unto their supplication: but the face of the Lord is upon them
that do evil. A glance at the Psalm will show that the Apostle has not
quoted precisely; and though he has much in common with the
Greek of the LXX., he does not adhere closely to that. But he gives
to the full the spirit both of the Hebrew and the Greek. The life of
which the Psalmist speaks is life in this world. The original explains
this by making the latter clause of the verse, "and loveth many days,
that he may see good." And the love is to be a noble feeling, a
desire to make his worth living. Such a life must exhibit watchfulness
over words and actions. The precepts begin at the beginning, with
control of the tongue. Control that, and you are master of the rest.
"It is a little member, but boasteth great things." "The world of
iniquity among our members is the tongue, which defileth the whole
body" (James iii. 5, 6). It needs to be kept as with a bridle, and not
only when the ungodly are in sight, but constantly. But the words of
the Psalm contemplate a further danger. Men may give good words
with the lips while the heart is full of bitterness. Then the lips are
lying, and this is an evil as great as the former, and more perilous to
him who commits it, because the sin does not come to the light that
it may be reproved, but contrives to wear the mask of virtue.
And the actions need watchfulness also. They must not only possess
the negative quality of abstinence from evil, but the positive stamp
of good deeds done. "By their fruits ye shall know them." And the
work will be no light one. Peace is to be sought, and the Apostle
uses a word which implies that a chase is needful to obtain it. St.
Paul has a passage very much in the spirit of St. Peter's teaching
here, and the words of which picture distinctly the difficulties which
the Christian will have to labour against: "Giving diligence to keep
the unity of the Spirit in the bond of peace" (Eph. iv. 3). This tells us
why our Apostle urges the pursuit of peace. It is the clasp which
binds the Christian communion together. From all sorts of causes
men are prone to fall apart, to break the oneness; and peace is able
to hold them fast. Hence the diligence in seeking it, the earnestness
of the pursuit that it may not elude us.
But when all is done, when men have not been sitting with folded
hands waiting and dreaming that peace would come without pursuit,
but have laboured for it, they do not always attain to it. "I am for
peace," says the Psalmist, "but when I speak, they are for war"
(Psalm cxx. 7). And so the disappointed struggler is directed to the
sure source of consolation amid discomfiture. The Lord marks his
efforts, knows their earnest purpose in spite of their ill-success. He
beholds also those who have withstood them, but with far other
regard. St. Peter has not quoted what the Psalmist says of their fate:
"God will root out the remembrance of them from the earth." God's
righteous pilgrim is not forgotten. His prayer is heard, and will be
answered for good. No shadow has come between him and God,
though his lot seem very dark. Neither can the wrong-doer raise a
shadow to screen himself from the all-seeing eyes. All things are
naked and open before the eyes of Him with whom we have to do.
Thus far St. Peter has used the language of the Psalmist, and among
the converts the Jews would be sure to supply from the context
those other words, "O fear the Lord, all ye His saints; for they that
fear Him lack nothing." The Apostle clothes that same thought in his
own words: And who is he that will harm you, if ye be zealous of
that which is good? He has repeatedly dwelt on the power of
goodness to win unbelievers to its side (ii. 12, 15; iii. 1), and the
same idea shapes his words now. In those days the Zealots were
well known, and their unbounded enthusiasm for their evil cause.
Josephus lays the destruction of Jerusalem at their door. The Apostle
would have Christ's disciples "zealots" for Him. Let there be nothing
half-hearted in their service, and its power will be irresistible. It will
avail either to silence and confound the adversaries, or to strengthen
the faithful so that the smell of the furnace of persecution shall not
pass upon them. They shall be enabled to break the chains with
which their foes would bind them as easily as Samson his green
withes. But and if ye should suffer for righteousness' sake, blessed
are ye. If ye endure chastening, God is dealing with you as with
sons. He has called Himself your Father; Christ has claimed you for
brethren. He, the righteous, suffered; shall we not reckon it for a
blessing to be worthy to bear the cross? Only let us be of good
courage. He that endureth to the end shall find salvation. And fear
not their fear, neither be troubled. Again St. Peter applies the
promises of the ancient Scriptures. In the days of Isaiah all Judah
was in terror, king and people alike, before the gathering armies of
Syria and Israel. In their dread comes the prophetic message, and
says to the confederates, "Gird yourselves, and ye shall be broken in
pieces," and to the tiny power of Judah, "Let the Lord of hosts be
your fear, and let Him be your dread, and He shall be for a
sanctuary" (Isa. viii. 12, 13). The condition of these Asian converts
was one of heaviness through manifold temptations. While the
believer lives here he always has his assailants, and in those early
days the rulers of the earth were not seldom among the adversaries
of the Christians. Hence the Apostle's exhortation is most apposite:
Fear not their fear—the things which they would dread, and with
which they will threaten you. For what are they? They may take
away your property. Be not troubled; you would soon have had to
leave it. The loss a few years sooner is no terrible affliction. They
may drive you from one land to another. To strangers and sojourners
what can that signify? If they cast you into prison, the Lord who
shut the lions' mouths for Daniel is your Lord also; and I, Peter,
know how angel-hands have removed chains and opened prison
doors. And should they scourge and torture you, do you shrink from
thus being made like unto your Master? Sanctify in your hearts
Christ as Lord.
Isaiah's message to disheartened Judah was, "The Lord of hosts,
Him shall ye sanctify." On His word shall ye rely, assured that He, the
holy God, will fail neither in wisdom nor power. To think otherwise is
not to sanctify Him. The Lord knoweth how to deliver out of
temptation. St. Peter, who knew Christ as the Son of the living God,
applies to the Son the words first spoken of the Father. The Son is
one with the Father. Hence he bids the afflicted converts, suffering
for righteousness' sake, not to be afraid of the world's terror, but to
sanctify Christ in their hearts as Lord. He is the Emmanuel, whom
Isaiah was sent to promise. God has dwelt among men, and will be
the God and the Deliverer of all His faithful ones. This sense of "God
with us" they know, and with the knowledge comes a power not
their own, and they fear no more the fear of their adversaries.
It is against foes of another sort that the Christian has now to hold
fast his faith, and sanctify Christ as his Lord. There are those who
deny Him all that is supernatural, all that speaks of the Divine in His
history; who treat the resurrection and ascension of the Lord as
groundless legends, due to the ignorance of His followers; and who
leave to the Jesus of the Gospels only the qualities of a better
fellow-man. These are the enemies of the cross of Christ.
And of such dangerous teaching it would seem as if St. Peter had
been thinking in the words that follow: Being ready always to give
answer to every man that asketh you a reason concerning the hope
that is in you. The believer rests on Christ in faith. But though in his
belief there must be much which he cannot fathom, yet it is a belief
for men. His service is a reasonable service; he can point to
abundance of evidence as ground for his faith; he believes because
he has experienced the power of the Spirit, and fears not to trust
the Christ whom he has sanctified in his heart as Lord; he knows in
whom he has believed. But beside this, he can study the Old
Testament; and there he learns how the coming incarnation
dominates every portion of the volume, how from the first
redemption through the seed of the woman was made known; and
he follows the revelation step by step till in the evangel of Isaiah he
has predictions almost as vivid and plain as the narrative of the
Gospels. Those four narratives are another warrant for his faith, their
wondrous agreement amid multitudinous divergences, divergences
so marked that none could have ventured to put them forth as
history except while the knowledge of those who had seen the Lord
and been witnesses of His actions was available to vouch for and
stamp as true these varicoloured pictures of the life of Jesus. He has
further vouchers in the lives and letters of those who knew and
followed the Lord, followed Him, most of them, on the road that led
through persecution unto death. And beside all this, there stands
and grows the Church built upon this history, strong with the power
of this faith and in her holy worship sanctifying Christ as her Lord.
These are things to which the Christian appeals. They are not the
only reasons for belief, but they are those of which he can make
other men cognisant, and to which the world cannot continue always
blind; and they have a force against which the gates of hell have not
yet been, nor ever will be able, to prevail.
These reasons he gives with meekness and fear—with meekness,
because in that spirit all the victories of the Lord are to be won; with
fear, lest by feeble advocacy the cause of Christ may suffer. And he
does not bring words alone with him to the struggle, but the power
of a godly life; he is prepared for the conflict by the possession of a
good conscience before God and men; he bears in mind the
prophetic exhortation, "Be ye clean, ye that bear the vessels of the
Lord" (Isa. lii. 11). That injunction was given to those who were in
their day strangers and pilgrims. But with the good conscience,
pureness of heart in the service of the Lord, there need be no haste,
no flight. The Lord will go before them; the God of Israel will be
their rearward. And the good conscience has lost none of its
efficacy: Wherein ye are spoken against, they may be put to shame
who revile your good manner of life in Christ. Of the Christian's faith
and hope his revilers know nothing, but his good life and his reasons
for it men can see and hear. And these shall gain the victory. But
they must go hand in hand. The deeds must bear out the words.
When he testifies that his hope is placed where neither persecutions
nor revilings avail against it, his life must show him fearless of what
the world can do. His position toward it must be that which St. Peter
himself took: "Whether it be right in the sight of God to hearken
unto you more than unto God, judge ye" (Acts iv. 19). Men may
marvel at what they see in him, but they will take knowledge that he
has been with Jesus. He is created, new-created, in Christ Jesus
unto good works (Eph. ii. 10). His revilers use him despitefully; but,
according to Christ's lesson, he prays for them, and their shafts
glance pointless off. Well does St. Paul close his catalogue of the
Christian armour "with all prayer and supplication praying at all
seasons in the Spirit" (Eph. vi. 18). Thus does the believer wield his
weapons effectually. His revilers have no reason for their words; he
is careful that they shall have none. As with Peter and John the
council could say nothing against their good deed and let them go,
finding nothing how they might punish them, so shall it be with
others of the faithful; and, for very shame at the futility of their
accusations and assaults, the revilers shall be put to silence.
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

textbookfull.com