An Introduction to Spatial

Data Science with GeoDa

Volume 1 – Exploring Spatial Data
This book is the first in a two-volume series that introduces the field of spatial data science.
It offers an accessible overview of the methodology of exploratory spatial data analysis. It also
constitutes the definitive user’s guide for the widely adopted GeoDa open-source software for
spatial analysis. Leveraging a large number of real-world empirical illustrations, readers will gain
an understanding of the main concepts and techniques, using dynamic graphics for thematic
mapping, statistical graphing, and, most centrally, the analysis of spatial autocorrelation. Key to
this analysis is the concept of local indicators of spatial association, pioneered by the author and
recently extended to the analysis of multivariate data.
The focus of the book is on intuitive methods to discover interesting patterns in spatial data.
It offers a progression from basic data manipulation through description and exploration to
the identification of clusters and outliers by means of local spatial autocorrelation analysis. A
distinctive approach is to spatialize intrinsically non-spatial methods by means of linking and
brushing with a range of map representations, including several that are unique to the GeoDa
software. The book also represents the most in-depth treatment of local spatial autocorrelation
and its visualization and interpretation by means of GeoDa.
The book is intended for readers interested in going beyond simple mapping of geographical
data to gain insight into interesting patterns. Some basic familiarity with statistical concepts is
assumed, but no previous knowledge of GIS or mapping is required.

Key Features:
• Includes spatial perspectives on cluster analysis
• Focuses on exploring spatial data
•  Supplemented by extensive support with sample data sets and examples on the GeoDaCenter
website
This book is both useful as a reference for the software and as a text for students and researchers
of spatial data science.

Luc Anselin is the Founding Director of the Center for Spatial Data Science at the University of Chi-
cago, where he is also the Stein-Freiler Distinguished Service Professor of Sociology and the College,
as well as a member of the Committee on Data Science. He is the creator of the GeoDa software and
an active contributor to the PySAL Python open-source software library for spatial analysis. He has
written widely on topics dealing with the methodology of spatial data analysis, including his clas-
sic 1988 text on Spatial Econometrics. His work has been recognized by many awards, such as his
election to the U.S. National Academy of Science and the American Academy of Arts and Science.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
An Introduction to Spatial
Data Science with GeoDa
Volume 1 – Exploring Spatial Data

Luc Anselin
Designed cover image: Luc Anselin
First edition published 2024
by CRC Press
2385 NW Executive Center Drive, Suite 320, Boca Raton FL 33431

and by CRC Press

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

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

© 2024 Luc Anselin

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

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or
utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including 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, access www.copyright.com or contact the
Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. For works that are
not available on CCC, please contact [email protected]

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

Library of Congress Cataloging-in-Publication Data

Names: Anselin, Luc, 1953- author.

Title: An introduction to spatial data science with GeoDa / Luc Anselin.
Description: First edition. | Boca Raton, FL : CRC Press, 2024. | Includes
bibliographical references and index. | Contents: Volume 1. Exploring
spatial data -- Volume 2. Clustering spatial data. | Summary: “This book
is the first in a two-volume series that introduces the field of spatial
data science. It offers an accessible overview of the methodology of
exploratory spatial data analysis. It also constitutes the definitive
user’s guide for the widely adopted GeoDa open source software for
spatial analysis. Leveraging a large number of real-world empirical
illustrations, readers will gain an understanding of the main concepts
and techniques, using dynamic graphics for thematic mapping, statistical
graphing, and, most centrally, the analysis of spatial autocorrelation.
Key to this analysis is the concept of local indicators of spatial
association, pioneered by the author and recently extended to the
analysis of multivariate data.”--Provided by publisher.
Identifiers: LCCN 2023048617 | ISBN 9781032229188 (volume 1 ; hardback) |
ISBN 9781032229621 (volume 1 ; paperback) | ISBN 9781003274919
(volume 1 ; ebook) | ISBN 9781032713021 (volume 2 ; hardback)
| ISBN 9781032713168 (volume 2 ; paperback) | ISBN 9781032713175
(volume 2 ; ebook)
Subjects: LCSH: Spatial analysis (Statistics) | Spatial analysis
(Statistics)--Data processing. | GeoDa (Computer file)
Classification: LCC QA278.2 .A565 2024 | DDC 519.5/3--dc23/eng/20240201
LC record available at https://lccn.loc.gov/2023048617

ISBN: 978-1-032-22918-8 (hbk)

ISBN: 978-1-032-22962-1 (pbk)
ISBN: 978-1-003-27491-9 (ebk)
DOI: 10.1201/9781003274919
Typeset in Latin Modern font
by KnowledgeWorks Global Ltd.
Publisher’s note: This book has been prepared from camera-ready copy provided by the authors.
To Emily
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
Contents

List of Figures xv

Preface xxiii

Acknowledgments xxv

About the Author xxvii

1 Introduction 1
1.1 Overview of Volume 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 A Quick Tour of GeoDa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Data entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Data manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 GIS operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.4 Weights manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.5 Mapping and geovisualization . . . . . . . . . . . . . . . . . . . . . . 5
1.2.6 Exploratory data analysis . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.7 Space-time analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.8 Spatial autocorrelation analysis . . . . . . . . . . . . . . . . . . . . . 6
1.2.9 Cluster analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Sample Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

I Spatial Data Wrangling 9

2 Basic Data Operations 11
2.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Spatial Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 GIS files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.2 Tabular files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.3 Other spatial data input . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Creating Spatial Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1 Point layers from coordinates . . . . . . . . . . . . . . . . . . . . . . 19
2.3.2 Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Table Manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Variable properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4.2 Calculator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.3 Merging tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5.1 Selection Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.2 Indicator variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.3 Save selected observations . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.4 Spatial selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

vii
viii Contents

3 GIS Operations 37
3.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.1 Coordinate reference system . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.2 Selecting a projection . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.3 Reprojection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Converting Between Points and Polygons . . . . . . . . . . . . . . . . . . . . 41
3.3.1 Mean centers and centroids . . . . . . . . . . . . . . . . . . . . . . . 42
3.3.2 Tessellations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Minimum Spanning Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4.1 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4.2 Minimum Spanning Tree options . . . . . . . . . . . . . . . . . . . . 49
3.5 Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.1 Dissolve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.5.2 Aggregation in table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.6 Multi-Layer Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6.1 Loading multiple layers . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6.2 Automatic reprojection . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.6.3 Selection in multiple layers . . . . . . . . . . . . . . . . . . . . . . . 54
3.7 Spatial Join . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.7.1 Spatial assign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.7.2 Spatial count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.8 Linked Multi-Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.8.1 Specifying an inter-layer linkage . . . . . . . . . . . . . . . . . . . . 59
3.8.2 Visualizing linked selections . . . . . . . . . . . . . . . . . . . . . . . 59

II EDA and ESDA 61

4 Geovisualization 63
4.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 From EDA to ESDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.1 Exploratory data analysis . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.2 Mapping as exploration . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.3 Exploratory spatial data analysis . . . . . . . . . . . . . . . . . . . . 66
4.2.4 Linking and brushing . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Thematic Maps – Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.1 Choropleth map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.2 Map classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.3 Legend and color . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4 Common Map Classifications . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.1 Quantile map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.2 Equal intervals map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4.3 Natural breaks map . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.5 Map options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5.1 The map legend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.5.2 The map toolbar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.5.3 Map base layer (base map) . . . . . . . . . . . . . . . . . . . . . . . 75
4.5.4 Saving the classification as a categorical variable . . . . . . . . . . . 78
4.5.5 Saving the map as an image . . . . . . . . . . . . . . . . . . . . . . . 79
Contents ix

4.5.6 Other map options . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

4.6 Custom Classifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.6.1 Category editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.6.2 Applications of custom categories . . . . . . . . . . . . . . . . . . . . 83
4.6.3 Saving the custom categories – the Project File . . . . . . . . . . . . 83

5 Statistical Maps 87
5.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2 Extreme Value Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.1 Percentile map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.2 Box map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.3 Standard deviation map . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3 Mapping Categorical Variables . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3.1 Unique values map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.3.2 Co-location map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.4 Cartogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4.1 Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4.2 Creating a cartogram . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.5 Map Animation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

6 Maps for Rates 101

6.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2 Choropleth Maps for Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.2.1 Spatially extensive and spatially intensive variables . . . . . . . . . . 102
6.2.2 Raw rate map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Excess Risk – SMR – LQ . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.1 Relative risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.2 Excess risk map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.4 Rate Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.4.1 Variance instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.4.2 Borrowing strength . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.4.3 Empirical Bayes smoothed rate map . . . . . . . . . . . . . . . . . . 112

7 Univariate and Bivariate Data Exploration 115

7.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.2 Analyzing the Distribution of a Single Variable . . . . . . . . . . . . . . . . 116
7.2.1 Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2.2 Box plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.3 Bivariate Analysis – The Scatter Plot . . . . . . . . . . . . . . . . . . . . . 123
7.3.1 Scatter plot basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.3.2 Smoother option – local regression . . . . . . . . . . . . . . . . . . . 128
7.4 Scatter Plot Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.4.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.5 Spatial Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.5.1 Averages chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.5.2 Brushing the scatter plot . . . . . . . . . . . . . . . . . . . . . . . . 138

8 Multivariate Data Exploration 141

8.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.2 The Curse of Dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . 142
8.2.1 The empty space phenomenon . . . . . . . . . . . . . . . . . . . . . 143
x Contents

8.3 Three Variables: Bubble Chart and 3-D Scatter Plot . . . . . . . . . . . . . 144
8.3.1 Bubble chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.3.2 3-D scatter plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.4 Conditional Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.4.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.4.2 Conditional statistical graphs . . . . . . . . . . . . . . . . . . . . . . 153
8.4.3 Conditional maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.5 Parallel Coordinate Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.5.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.5.2 Clusters and outliers in PCP . . . . . . . . . . . . . . . . . . . . . . 159

9 Space-Time Exploration 161

9.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.2 Time Editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
9.2.1 Creating grouped variables . . . . . . . . . . . . . . . . . . . . . . . 162
9.3 Time Player – Comparative Statics . . . . . . . . . . . . . . . . . . . . . . . 166
9.3.1 Box plot over time . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.3.2 Scatter plot with a time-lagged variable . . . . . . . . . . . . . . . . 167
9.3.3 Thematic map over time . . . . . . . . . . . . . . . . . . . . . . . . . 169
9.4 Treatment Effect Analysis – Averages Chart . . . . . . . . . . . . . . . . . . . 171
9.4.1 Difference in means . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
9.4.2 Difference in difference . . . . . . . . . . . . . . . . . . . . . . . . . . 172

III Spatial Weights 177

10 Contiguity-Based Spatial Weights 179
10.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
10.2 The Concept of Spatial Weights . . . . . . . . . . . . . . . . . . . . . . . . . 180
10.2.1 Spatial weights matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
10.2.2 Types of contiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
10.2.3 Higher-order contiguity . . . . . . . . . . . . . . . . . . . . . . . . . 185
10.2.4 Practical considerations . . . . . . . . . . . . . . . . . . . . . . . . . 187
10.3 Creating Contiguity Weigths . . . . . . . . . . . . . . . . . . . . . . . . . . 188
10.3.1 Weights manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
10.3.2 Rook contiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
10.3.3 Queen contiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.3.4 Higher-order contiguity . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.3.5 Project file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.3.6 Using existing weights . . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.3.7 Space-time weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
10.4 Weights Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.4.1 Summary characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.4.2 Connectivity histogram . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.4.3 Connectivity map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
10.4.4 Connectivity graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
10.4.5 Connectivity option . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

11 Distance-Based Spatial Weights 203

11.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
11.2 Distance Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Contents xi

11.2.1 Distance in a Cartesian coordinate system . . . . . . . . . . . . . . . 204

11.2.2 Great circle distance . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
11.2.3 General distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
11.3 Distance-Based Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
11.3.1 Distance-band weights . . . . . . . . . . . . . . . . . . . . . . . . . . 206
11.3.2 K-nearest neighbor weights . . . . . . . . . . . . . . . . . . . . . . . 207
11.3.3 General distance weights . . . . . . . . . . . . . . . . . . . . . . . . . 208
11.4 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.4.1 Distance-band weights . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
11.4.2 Isolates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
11.4.3 K-nearest neighbor weights . . . . . . . . . . . . . . . . . . . . . . . 215
11.4.4 General distance matrix . . . . . . . . . . . . . . . . . . . . . . . . . 216
11.5 Broadening the Concept of Contiguity . . . . . . . . . . . . . . . . . . . . . 217
11.5.1 Contiguity-based weights for points . . . . . . . . . . . . . . . . . . . 218
11.5.2 Distance-based weights for polygons . . . . . . . . . . . . . . . . . . 220
11.5.3 Block weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
11.5.4 Space-time weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
11.6 Manipulating Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
11.6.1 Set operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
11.6.2 Make symmetric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

12 Special Weights Operations 225

12.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
12.2 Spatial Weights as Distance Functions . . . . . . . . . . . . . . . . . . . . . 226
12.2.1 Inverse distance weights . . . . . . . . . . . . . . . . . . . . . . . . . 226
12.2.2 Kernel weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
12.3 Spatial Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
12.3.1 Spatially lagged variables . . . . . . . . . . . . . . . . . . . . . . . . 233
12.3.2 Using inverse distance weights – potential measure . . . . . . . . . . 236
12.3.3 Using kernel weights . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
12.4 Spatial Weights in Rate Smoothing . . . . . . . . . . . . . . . . . . . . . . . 239
12.4.1 Spatial rate smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . 240
12.4.2 Spatial Empirical Bayes smoothing . . . . . . . . . . . . . . . . . . . 242

IV Global Spatial Autocorrelation 245

13 Spatial Autocorrelation 247
13.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
13.2 Spatial Randomness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
13.2.1 Parametric approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
13.2.2 Randomization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
13.2.3 Simulating spatial randomness . . . . . . . . . . . . . . . . . . . . . 250
13.3 Positive and negative spatial autocorrelation . . . . . . . . . . . . . . . . . . 250
13.3.1 Positive spatial autocorrelation . . . . . . . . . . . . . . . . . . . . . 250
13.3.2 Negative Spatial Autocorrelation . . . . . . . . . . . . . . . . . . . . . 251
13.4 Spatial Autocorrelation Statistic . . . . . . . . . . . . . . . . . . . . . . . . 252
13.4.1 Attribute similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
13.4.2 Locational similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
13.4.3 Examples of spatial autocorrelation statistics . . . . . . . . . . . . . 253
13.5 Moran’s I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
xii Contents

13.5.1 Statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

13.5.2 Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
13.5.3 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
13.5.4 Moran scatter plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
13.6 Visualizing spatial autocorrelation with the Moran scatter plot . . . . . . . . 261
13.6.1 Creating a Moran scatter plot . . . . . . . . . . . . . . . . . . . . . . . 261
13.6.2 Moran scatter plot options . . . . . . . . . . . . . . . . . . . . . . . 262
13.6.3 Brushing the Moran scatter plot . . . . . . . . . . . . . . . . . . . . 264

14 Advanced Global Spatial Autocorrelation 267

14.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
14.2 Specialized Moran Scatter Plots . . . . . . . . . . . . . . . . . . . . . . . . . 267
14.2.1 Differential Moran scatter plot . . . . . . . . . . . . . . . . . . . . . 268
14.2.2 Moran scatter plot for EB rates . . . . . . . . . . . . . . . . . . . . . . 271
14.3 Bivariate Spatial Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 273
14.3.1 Bivariate Moran scatter plot . . . . . . . . . . . . . . . . . . . . . . 273
14.4 The Many Faces of Space-Time Correlation . . . . . . . . . . . . . . . . . . 276
14.4.1 Serial (temporal) correlation . . . . . . . . . . . . . . . . . . . . . . 276
14.4.2 Serial (temporal) correlation between spatial lags . . . . . . . . . . . 276
14.4.3 Space-time regression . . . . . . . . . . . . . . . . . . . . . . . . . . 277
14.4.4 Serial and space-time regression . . . . . . . . . . . . . . . . . . . . . 279

15 Nonparametric Spatial Autocorrelation 281

15.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
15.2 Non-Parametric Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
15.3 Spatial Correlogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
15.3.1 Creating a spatial correlogram . . . . . . . . . . . . . . . . . . . . . 283
15.3.2 Spatial correlogram options . . . . . . . . . . . . . . . . . . . . . . . 285
15.4 Smoothed Distance Scatter Plot . . . . . . . . . . . . . . . . . . . . . . . . 288
15.4.1 Multivariate extension . . . . . . . . . . . . . . . . . . . . . . . . . . 290
15.4.2 Creating a smoothed distance scatter plot . . . . . . . . . . . . . . . 290
15.4.3 Smoothed distance scatter plot options . . . . . . . . . . . . . . . . . . 291

V Local Spatial Autocorrelation 295

16 LISA and Local Moran 297
16.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
16.2 LISA Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
16.3 Local Moran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
16.3.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
16.3.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
16.4 Clusters and Spatial Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . 302
16.4.1 Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
16.4.2 Spatial outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
16.5 Significance and Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . 305
16.5.1 Multiple comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . 306
16.5.2 Interpretation of significance . . . . . . . . . . . . . . . . . . . . . . 309
16.5.3 Interpretation of clusters . . . . . . . . . . . . . . . . . . . . . . . . . 310
16.6 Conditional Local Cluster Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Contents xiii

17 Other Local Spatial Autocorrelation Statistics 313

17.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
17.2 Extensions of the Local Moran . . . . . . . . . . . . . . . . . . . . . . . . . 314
17.2.1 Median Local Moran . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
17.2.2 Differential Local Moran . . . . . . . . . . . . . . . . . . . . . . . . . 318
17.2.3 Local Moran with EB Rate . . . . . . . . . . . . . . . . . . . . . . . 320
17.3 Local Geary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
17.3.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
17.3.2 Clusters and spatial outliers . . . . . . . . . . . . . . . . . . . . . . . 325
17.3.3 Comparing Local Geary and Local Moran . . . . . . . . . . . . . . . 326
17.4 Getis-Ord Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
17.4.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
17.4.2 Clusters and outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
17.4.3 Comparing G statistics and Local Moran . . . . . . . . . . . . . . . . 331
17.5 Which Local Statistic to Use? . . . . . . . . . . . . . . . . . . . . . . . . . . 333

18 Multivariate Local Spatial Autocorrelation 335

18.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
18.2 The Multivariate Spatial Autocorrelation Problem . . . . . . . . . . . . . . 335
18.3 Bivariate Local Moran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
18.3.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
18.3.2 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
18.4 Multivariate Local Geary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
18.4.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
18.4.2 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
18.5 Local Neighbor Match Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
18.5.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
18.5.2 Local neighbor match and Multivariate Local Geary . . . . . . . . . 349

19 LISA for Discrete Variables 351

19.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
19.2 Univariate Local Join Count Statistic . . . . . . . . . . . . . . . . . . . . . . 351
19.2.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
19.3 Bivariate Local Join Count Statistic . . . . . . . . . . . . . . . . . . . . . . 355
19.3.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
19.4 Co-Location Local Join Count Statistic . . . . . . . . . . . . . . . . . . . . 357
19.4.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
19.5 Quantile LISA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
19.5.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
19.5.2 Univariate Quantile LISA . . . . . . . . . . . . . . . . . . . . . . . . 360
19.5.3 Bivariate and Multivariate Quantile LISA . . . . . . . . . . . . . . . 360

20 Density-Based Clustering Methods 363

20.1 Topics Covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
20.2 Heat Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
20.2.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
20.3 DBSCAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
20.3.1 Important concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
20.3.2 DBSCAN algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
20.3.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
20.4 DBSCAN* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
xiv Contents

20.4.1 Mutual reachability distance . . . . . . . . . . . . . . . . . . . . . . 373

20.4.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
20.5 HDBSCAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
20.5.1 Important concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
20.5.2 HDBSCAN algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 379
20.5.3 Cluster membership . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
20.5.4 Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
20.5.5 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

VI Epilogue 389
21 Postscript – The Limits of Exploration 391

A Appendix A – GeoDa Preference Settings 395

B Appendix B – Menu Structure 399

C Appendix C – Scripting with GeoDa via the geodalib Library 403

Bibliography 405

Index 417
List of Figures

1.1 GeoDa toolbar icons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Data entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Data manipulation/table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 GIS operations/tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Weights manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.6 Mapping and geovisualization . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.7 Exploratory data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.8 Space-time analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.9 Spatial autocorrelation analysis . . . . . . . . . . . . . . . . . . . . . . . . 6
1.10 Cluster analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1 Open | Close | Save | Table | Tools . . . . . . . . . . . . . . . . . . . . . . 12

2.2 Connect to Data Source dialog . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Supported spatial file formats . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Themeless polygon map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Example GeoJSON file contents . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Example GML file contents . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.7 Themeless point map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.8 CSV format file input dialog . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.9 Coordinate variables in data table . . . . . . . . . . . . . . . . . . . . . . . 19
2.10 Specifying the point coordinates . . . . . . . . . . . . . . . . . . . . . . . . 19
2.11 Point layer from coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.12 Blank CRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.13 CRS with projection information . . . . . . . . . . . . . . . . . . . . . . . . 21
2.14 Grid creation dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.15 Grid layer over Chicago community areas . . . . . . . . . . . . . . . . . . . 22
2.16 Table options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.17 Edit variable properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.18 Create a date/time variable . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.19 Date/time format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.20 Customizing the date/time format . . . . . . . . . . . . . . . . . . . . . . . 26
2.21 Calculator interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.22 Merge dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.23 Merged tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.24 Selection Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.25 Selected observations in themeless map . . . . . . . . . . . . . . . . . . . . 32
2.26 Selected observations in new map . . . . . . . . . . . . . . . . . . . . . . . 32
2.27 Car thefts by month . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.28 Selection on the map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.29 Selection on a map category . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.1 UTM zones for North America (source: GISGeography) . . . . . . . . . . . 40

xv
xvi List of Figures

3.2 Load CRS from another layer . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.3 New CRS from another layer . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Reprojected Chicago community areas layer . . . . . . . . . . . . . . . . . 42
3.5 Display mean centers on map . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.6 Add mean centers to table . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.7 Chicago community area mean center layer . . . . . . . . . . . . . . . . . . 44
3.8 Display Thiessen polygons . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.9 Connectivity graph of mean centers . . . . . . . . . . . . . . . . . . . . . . 46
3.10 Mininum spanning tree for connectivity graph of mean centers . . . . . . . 47
3.11 Toy example point coordinates . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.12 Inter-point distance matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.13 Connectivity graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.14 Initial steps in Prim’s algorithm . . . . . . . . . . . . . . . . . . . . . . . . 48
3.15 Minimum spanning tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.16 Dissolve dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.17 Dissolved districts map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.18 Dissolved districts table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.19 Map layer settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.20 Polygon layer on top with opacity=0 . . . . . . . . . . . . . . . . . . . . . 53
3.21 Points reprojected to lat-lon . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.22 Select on current map layer (community areas) . . . . . . . . . . . . . . . 55
3.23 Spatial join dialog – areal unit identifier for point . . . . . . . . . . . . . . 56
3.24 Community area identifier (SA) in point table . . . . . . . . . . . . . . . . 56
3.25 Point in polygon mismatch . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.26 Spatial join dialog – aggregation options . . . . . . . . . . . . . . . . . . . 57
3.27 Spatial count in polygon area table . . . . . . . . . . . . . . . . . . . . . . 58
3.28 Set association dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.29 Linked selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.1 Maps and Rates | Cartogram | Map Movie | Category Editor . . . . . . . . 64

4.2 Mapping options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3 Quintile map for housing index, Ceará . . . . . . . . . . . . . . . . . . . . 69
4.4 Tied values in a quantile map . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Equal intervals map for housing index, Ceará . . . . . . . . . . . . . . . . . 71
4.6 Equal intervals map and histogram . . . . . . . . . . . . . . . . . . . . . . 72
4.7 Natural breaks map for housing index, Ceará . . . . . . . . . . . . . . . . . 73
4.8 Map options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.9 Legend options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.10 Map toolbar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.11 Map base layer options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.12 Stamen TonerLight base map, default map transparency . . . . . . . . . . 76
4.13 Stamen TonerLight base map, 0.40 map transparency . . . . . . . . . . . . 77
4.14 ESRI WorldStreetMap base map, 0.40 map transparency . . . . . . . . . . 78
4.15 Saved Image As dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.16 Map without outlines visible . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.17 Category editor interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.18 Quintile maps for sanitation and infrastructure . . . . . . . . . . . . . . . 84
4.19 Thematic maps for sanitation and infrastructure using custom classification 84
4.20 Custom category definition in project file . . . . . . . . . . . . . . . . . . . 85

5.1 Percentile map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

List of Figures xvii

5.2 Box map, hinge = 1.5, with box plot . . . . . . . . . . . . . . . . . . . . . 89

5.3 Outliers in box plot and box map . . . . . . . . . . . . . . . . . . . . . . . 90
5.4 Standard deviation map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.5 Unique values map, Zika indicator . . . . . . . . . . . . . . . . . . . . . . . 92
5.6 Co-location map variable selection, Zika and Microcephaly indicators . . . 93
5.7 Co-location map, Zika and Microcephaly incidence . . . . . . . . . . . . . 94
5.8 Selection of co-located observations . . . . . . . . . . . . . . . . . . . . . . 95
5.9 Cartogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.10 Linked Cartogram and map . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.11 Animation control panel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.12 Linked animation sanitation-infrastructure . . . . . . . . . . . . . . . . . . 99

6.1 Population and population density . . . . . . . . . . . . . . . . . . . . . . 103

6.2 GDP and GDP per capita . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3 Rate map interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.4 GDP per capita as a rate map . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.5 Rate calculation in table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.6 GDP per capita in table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.7 GDP per capita excess risk . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.8 GDP per capita excess risk in table . . . . . . . . . . . . . . . . . . . . . . 108
6.9 Crude and EB rate map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.10 Effect of smoothing – small populations . . . . . . . . . . . . . . . . . . . . 113
6.11 Effect of smoothing – large populations . . . . . . . . . . . . . . . . . . . . 113

7.1 Histogram | Box Plot | Scatter Plot | Scatter Plot Matrix . . . . . . . . . . 116
7.2 Histogram for percent poverty 2020 . . . . . . . . . . . . . . . . . . . . . . 117
7.3 Histogram for percent poverty 2020 – 12 bins . . . . . . . . . . . . . . . . 117
7.4 Bar chart for settlement categories . . . . . . . . . . . . . . . . . . . . . . 118
7.5 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.6 Linking between histogram and map . . . . . . . . . . . . . . . . . . . . . 120
7.7 Linking between map and histogram . . . . . . . . . . . . . . . . . . . . . 120
7.8 Box plot for population change 2020–2010 . . . . . . . . . . . . . . . . . . . 121
7.9 Box map for population change 2020–2010 (upper outliers selected) . . . . 122
7.10 Scatter plot of food insecurity on poverty . . . . . . . . . . . . . . . . . . . 124
7.11 Regime regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.12 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.13 Default LOWESS local regression fit . . . . . . . . . . . . . . . . . . . . . 127
7.14 Default LOWESS local regression fit with bandwidth 0.6 . . . . . . . . . . 128
7.15 Default LOWESS local regression fit with bandwidth 0.05 . . . . . . . . . 129
7.16 Scatter plot matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7.17 Scatter plot correlation matrix . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.18 Scatter plot matrix with LOWESS fit . . . . . . . . . . . . . . . . . . . . . 133
7.19 Food insecurity 2020 in Valles Centrales . . . . . . . . . . . . . . . . . . . 134
7.20 Averages Chart – Valles Centrales . . . . . . . . . . . . . . . . . . . . . . . 135
7.21 Map brushing and the averages chart – 1 . . . . . . . . . . . . . . . . . . . 136
7.22 Map brushing and the averages chart – 2 . . . . . . . . . . . . . . . . . . . 136
7.23 Map brushing and the averages chart – 3 . . . . . . . . . . . . . . . . . . . 137
7.24 Map brushing and the scatter plot – 1 . . . . . . . . . . . . . . . . . . . . 138
7.25 Map brushing and the scatter plot – 2 . . . . . . . . . . . . . . . . . . . . 138
7.26 Map brushing and the scatter plot – 3 . . . . . . . . . . . . . . . . . . . . 139
xviii List of Figures

8.1 Bubble Chart | 3D Scatter Plot | Parallel Coordinate Plot | Conditional Plot 142
8.2 One dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.3 Two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.4 Discrete evaluation points in three variable dimensions . . . . . . . . . . . 143
8.5 Two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.6 Three dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.7 Bubble chart – default settings: education, basic services, extreme poverty 145
8.8 Bubble chart – bubble size adjusted: education, basic services, extreme
poverty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8.9 Bubble chart – no color: education, basic services, extreme poverty . . . . 147
8.10 Bubble chart – categories: education, basic services, region . . . . . . . . . 147
8.11 3D scatter plot: education, basic services, extreme poverty . . . . . . . . . 148
8.12 Interacting with 3D scatter plot . . . . . . . . . . . . . . . . . . . . . . . . 149
8.13 Selection in the 3D scatter plot . . . . . . . . . . . . . . . . . . . . . . . . 150
8.14 Brushing and linking with the 3D scatter plot . . . . . . . . . . . . . . . . . 151
8.15 Conditional scatter plot – 3 by 3 . . . . . . . . . . . . . . . . . . . . . . . . 152
8.16 Conditional scatter plot – unique values . . . . . . . . . . . . . . . . . . . 153
8.17 Conditional box plot – lack of health access by region . . . . . . . . . . . . 154
8.18 Conditional box map – 2 by 2 . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.19 Parallel coordinate plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.20 Parallel coordinate plot – standardized variables . . . . . . . . . . . . . . . 157
8.21 Brushing the PCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.22 Brushing map and PCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.23 Clusters of observations in PCP . . . . . . . . . . . . . . . . . . . . . . . . 159
8.24 Outlier observation in PCP . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

9.1 Time | Averages Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

9.2 Time Editor interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
9.3 Grouped variables in data table . . . . . . . . . . . . . . . . . . . . . . . . 163
9.4 Grouped variables in project file . . . . . . . . . . . . . . . . . . . . . . . . 165
9.5 Time Player dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.6 Box plot over time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
9.7 Box plot for two time periods . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.8 Time lagged scatter plot variables . . . . . . . . . . . . . . . . . . . . . . . 169
9.9 Scatter plots with time lagged variables . . . . . . . . . . . . . . . . . . . . 169
9.10 Thematic maps for car ownership in 2000 . . . . . . . . . . . . . . . . . . . 170
9.11 Thematic maps for car ownership in 2010 . . . . . . . . . . . . . . . . . . . 170
9.12 Thematic maps for car ownership in 2020 . . . . . . . . . . . . . . . . . . . 170
9.13 Selected municipalities with altitude in two upper quantiles . . . . . . . . . 171
9.14 Difference in means, selected and unselected, static, in 2000 and 2020 . . . 172
9.15 Difference in difference setup, 2000 to 2020 . . . . . . . . . . . . . . . . . . 173
9.16 Difference in difference results . . . . . . . . . . . . . . . . . . . . . . . . . 174
9.17 Difference in means, unselected, dynamic, 2000 to 2020 . . . . . . . . . . . 174

10.1 Weights Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

10.2 Example spatial layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
10.3 Contiguity as a network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
10.4 Regular grid layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.5 Rook contiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.6 Queen contiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.7 Second-order contiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
List of Figures xix

10.8 Higher order contiguity as multiple steps in the graph . . . . . . . . . . . . 186

10.9 Weights Manager interface . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
10.10 Rook contiguity in the Weights File Creation interface . . . . . . . . . . . 189
10.11 Rook weights summary interface . . . . . . . . . . . . . . . . . . . . . . . . 190
10.12 GAL file contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
10.13 Summary properties of first-order queen contiguity . . . . . . . . . . . . . . 191
10.14 Summary properties of second-order queen contiguity . . . . . . . . . . . . 193
10.15 Summary properties of inclusive second-order queen contiguity . . . . . . . 193
10.16 Weights description in Project File . . . . . . . . . . . . . . . . . . . . . . 194
10.17 Second-order contiguity weights loaded from file . . . . . . . . . . . . . . . 195
10.18 Saved space-time table as csv file . . . . . . . . . . . . . . . . . . . . . . . 195
10.19 Space-time GAL contiguity file . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.20 Rook contiguity histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
10.21 Rook and queen contiguity histograms . . . . . . . . . . . . . . . . . . . . 197
10.22 Linked contiguity histogram and map . . . . . . . . . . . . . . . . . . . . . 198
10.23 Connectivity map for queen contiguity . . . . . . . . . . . . . . . . . . . . 199
10.24 Connectivity graph for queen contiguity . . . . . . . . . . . . . . . . . . . 200
10.25 Connectivity graph for queen contiguity without map . . . . . . . . . . . . 200

11.1 Nearest neighbor properties . . . . . . . . . . . . . . . . . . . . . . . . . . 207

11.2 Italian community bank point layer . . . . . . . . . . . . . . . . . . . . . . 209
11.3 Distance-based weights in the Weights File Creation interface . . . . . . . 210
11.4 Default distance-based weights summary properties . . . . . . . . . . . . . 210
11.5 GWT weights file format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
11.6 Connectivity histogram for default distance weights . . . . . . . . . . . . . 212
11.7 Connectivity graph for default distance weights . . . . . . . . . . . . . . . 213
11.8 Weights summary properties for distance 73 km . . . . . . . . . . . . . . . 214
11.9 Connectivity graph for distance band 73 km . . . . . . . . . . . . . . . . . 215
11.10 Weights summary properties for 6 k-nearest neighbors . . . . . . . . . . . 216
11.11 Connectivity graph for 6 k-nearest neighbors . . . . . . . . . . . . . . . . . 217
11.12 Contiguity for points from Thiessen polygons . . . . . . . . . . . . . . . . 218
11.13 Weights summary properties for point contiguity (Thiessen polygons) . . . 218
11.14 Connectivity histogram for point contiguity (Thiessen polygons) . . . . . . 219
11.15 Connectivity graph for point contiguity (Thiessen polygons) . . . . . . . . 220
11.16 Block weights in the Weights File Creation interface . . . . . . . . . . . . . . 221
11.17 Connectivity graph for block weights by region . . . . . . . . . . . . . . . . . 221
11.18 Space-time weights GAL file . . . . . . . . . . . . . . . . . . . . . . . . . . 222
11.19 Connectivity graph for intersection distance 73 km and knn 6 . . . . . . . 224

12.1 Summary properties of inverse distance weights . . . . . . . . . . . . . . . 228

12.2 GWT files with distance for original and rescaled coordinates . . . . . . . 229
12.3 Kernel weights in Weights File Creation interface . . . . . . . . . . . . . . . 231
12.4 Properties of kernel weights . . . . . . . . . . . . . . . . . . . . . . . . . . 232
12.5 KWT file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
12.6 Spatial lag in calculator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
12.7 Spatially lagged variables added to table . . . . . . . . . . . . . . . . . . . 235
12.8 Variables computed with inverse distance weights added to table . . . . . . 237
12.9 Disability rate: raw rate and EB smoothed rate . . . . . . . . . . . . . . . 239
12.10 Spatial rate map: first- and second-order queen contiguity . . . . . . . . . . 241
12.11 Spatial rate map: k-nearest neighbor and inverse distance weights . . . . . 242
12.12 Spatial rate map: inverse distance weights from Calculator . . . . . . . . . 242
xx List of Figures

12.13 Spatial Empirical Bayes smoothed rate map using region block weights . . 243

13.1 Moran Scatter Plot | Spatial Correlogram | Cluster Maps . . . . . . . . . . 248

13.2 Spatially random observations – i.i.d . . . . . . . . . . . . . . . . . . . . . 249
13.3 Spatially random observations – true and randomized per capita income . 249
13.4 True and randomized per capita income – histograms . . . . . . . . . . . . 250
13.5 Positive spatial autocorrelation . . . . . . . . . . . . . . . . . . . . . . . . . 251
13.6 Negative spatial autocorrelation . . . . . . . . . . . . . . . . . . . . . . . . . 251
13.7 Reference distribution, Moran’s I for per capita income . . . . . . . . . . . 257
13.8 Reference distribution, Moran’s I for randomized per capita income . . . . 258
13.9 Moran scatter plot, community area per capita income . . . . . . . . . . . 259
13.10 Positive spatial autocorrelation, high-high . . . . . . . . . . . . . . . . . . 260
13.11 Negative spatial autocorrelation, low-high . . . . . . . . . . . . . . . . . . . 261
13.12 Moran scatter plot LOWESS smoother . . . . . . . . . . . . . . . . . . . . 263
13.13 Brushing the Moran scatter plot . . . . . . . . . . . . . . . . . . . . . . . . 264
13.14 Brushing and linking the Moran scatter plot . . . . . . . . . . . . . . . . . 265
13.15 Static spatial weights in Moran scatter plot . . . . . . . . . . . . . . . . . 265

14.1 Access to health care in 2010 and 2020 . . . . . . . . . . . . . . . . . . . . 269

14.2 Change in access to health care between 2010 and 2020 . . . . . . . . . . . 269
14.3 Moran scatter plot for access to health care in 2010 and 2020 . . . . . . . 270
14.4 Differential Moran scatter plot for access to health care 2020–2010 . . . . . . 271
14.5 Moran scatter plot for raw rate and EB Moran scatter plot . . . . . . . . . 273
14.6 Bivariate Moran scatter plot for access to health care 2020 and its spatial
lag in 2010 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
14.7 Correlation between access to health care in 2020 and 2010 . . . . . . . . . 277
14.8 Correlation between the spatial lag of access to health care in 2020 and the
spatial lag in 2010 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
14.9 Correlation between access to health care in 2020 and its spatial lag in 2010 278
14.10 OLS regression of pPHA 2020 on pPHA 2010 and W PHA 2010 . . . . . . 279

15.1 Moran Scatter Plot | Spatial Correlogram | Cluster Maps . . . . . . . . . . 282

15.2 Spatial distribution of loan loss provisions – Italian banks 2016 . . . . . . 283
15.3 Correlogram parameter selection . . . . . . . . . . . . . . . . . . . . . . . . 284
15.4 Default spatial correlogram – LLP (2016) . . . . . . . . . . . . . . . . . . . 285
15.5 Spatial correlogram – LLP, 150 km cut off, 10 bins . . . . . . . . . . . . . 286
15.6 Spatial correlogram – LLP, 150 km cut off, 15 bins . . . . . . . . . . . . . 287
15.7 Smoothed distance scatter plot . . . . . . . . . . . . . . . . . . . . . . . . 289
15.8 Smoothed distance scatter plot with 150 km distance cut-off . . . . . . . . . 291
15.9 Smoothed distance scatter plot with 150 km distance cut-off and customized
Y-Axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
15.10 Smoothed distance scatter plot with 150 km distance cut-off and span 0.5 293
15.11 Smoothed distance scatter plot with 150 km distance cut-off, span 0.5, linear
fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
15.12 Bivariate smoothed distance scatter plot . . . . . . . . . . . . . . . . . . . 294

16.1 Moran Scatter Plot | Spatial Correlogram | Cluster Maps . . . . . . . . . . 298

16.2 Significance map and cluster map . . . . . . . . . . . . . . . . . . . . . . . 300
16.3 Significance map and cluster map – 99999 permutations . . . . . . . . . . . 301
16.4 High-High observations in Moran scatter plot . . . . . . . . . . . . . . . . 303
16.5 Significant High-High observations in Moran scatter plot . . . . . . . . . . 303
List of Figures xxi

16.6 High-High cluster location . . . . . . . . . . . . . . . . . . . . . . . . . . . 304

16.7 Significant High-Low observations in Moran scatter plot . . . . . . . . . . 305
16.8 High-Low spatial outlier location . . . . . . . . . . . . . . . . . . . . . . . 305
16.9 Significance map – p < 0.01 . . . . . . . . . . . . . . . . . . . . . . . . . . 306
16.10 Cluster map – p < 0.01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
16.11 Cluster map – Bonferroni bound . . . . . . . . . . . . . . . . . . . . . . . . 308
16.12 False Discovery Rate calculation . . . . . . . . . . . . . . . . . . . . . . . . 309
16.13 Cluster map – FDR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
16.14 Cores and neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
16.15 Conditional cluster map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

17.1 Significance maps conventional and Median Local Moran . . . . . . . . . . 315

17.2 High-High clusters in Median Local Moran . . . . . . . . . . . . . . . . . . 316
17.3 Low-Low clusters in Median Local Moran . . . . . . . . . . . . . . . . . . . 316
17.4 Low-High spatial outliers in Median Local Moran . . . . . . . . . . . . . . 317
17.5 High-Low spatial outliers in Median Local Moran . . . . . . . . . . . . . . 317
17.6 Cluster maps conventional and Median Local Moran . . . . . . . . . . . . 317
17.7 Cluster maps for Local Moran in 2010 and 2020 . . . . . . . . . . . . . . . 319
17.8 Differential Local Moran significance and cluster maps . . . . . . . . . . . 319
17.9 Significance maps – raw rate and EB rate . . . . . . . . . . . . . . . . . . . 321
17.10 Cluster maps – raw rate and EB rate . . . . . . . . . . . . . . . . . . . . . . 321
17.11 Local Geary significance and cluster maps . . . . . . . . . . . . . . . . . . 323
17.12 Local Geary – High-High clusters . . . . . . . . . . . . . . . . . . . . . . . 324
17.13 Local Geary – Low-Low clusters . . . . . . . . . . . . . . . . . . . . . . . . 324
17.14 Local Geary – Other clusters . . . . . . . . . . . . . . . . . . . . . . . . . . 324
17.15 Local Geary – Spatial outliers . . . . . . . . . . . . . . . . . . . . . . . . . 325
17.16 Local Geary and Local Moran cluster maps . . . . . . . . . . . . . . . . . 326
17.17 Local Geary and Local Moran – High-High clusters . . . . . . . . . . . . . 327
17.18 Local Geary and Local Moran – Low-Low clusters . . . . . . . . . . . . . . 327
17.19 Local Geary and Local Moran – Other positive clusters . . . . . . . . . . . 327
17.20 Local Geary and Local Moran – Spatial outliers . . . . . . . . . . . . . . . 328
17.21 Local Moran and Local Geary – Spatial outliers . . . . . . . . . . . . . . . 328
17.22 Gi* signficance map and cluster map . . . . . . . . . . . . . . . . . . . . . 330
17.23 Gi* High-High clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
17.24 Gi* Low-Low clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
17.25 Gi* and Local Moran High-High clusters . . . . . . . . . . . . . . . . . . . 332
17.26 Gi* and Local Moran Low-Low clusters . . . . . . . . . . . . . . . . . . . . 332
17.27 Gi* and Local Moran Low-High spatial outliers . . . . . . . . . . . . . . . 333
17.28 Gi* and Local Moran High-Low spatial outliers . . . . . . . . . . . . . . . 333

18.1 Box maps – Child Poverty and Crowded Housing . . . . . . . . . . . . . . 338

18.2 Co-location of Child Poverty and Crowded Housing . . . . . . . . . . . . . 338
18.3 Local Moran cluster map – Child Poverty and Crowded Housing . . . . . . 339
18.4 Co-location of Local Moran for Child Poverty and Crowded Housing . . . 339
18.5 Bivariate Local Moran cluster map – Child Poverty and Crowded Housing 340
18.6 Box map and Local Geary cluster map – Uninsured . . . . . . . . . . . . . 342
18.7 Local Geary cluster maps – Child Poverty, Crowded Housing . . . . . . . . 343
18.8 Co-location of Local Geary . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
18.9 Multivariate Local Geary cluster map – Child Poverty, Crowded Housing,
Uninsured . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
18.10 Multivariate Local Geary cluster map and Univariate Local Geary co-location344
xxii List of Figures

18.11 Multivariate Local Geary cluster in PCP . . . . . . . . . . . . . . . . . . . 345

18.12 K nearest neighbor connectivity graphs . . . . . . . . . . . . . . . . . . . . 346
18.13 Intersection of K nearest neighbor connectivity graphs . . . . . . . . . . . 347
18.14 Local neighbor match cardinality map . . . . . . . . . . . . . . . . . . . . 348
18.15 Local neighbor match significance map . . . . . . . . . . . . . . . . . . . . 348
18.16 Local neighbor significant matches . . . . . . . . . . . . . . . . . . . . . . . 349
18.17 Local neighbor match and Multivariate Local Geary . . . . . . . . . . . . . 349

19.1 Unique values Black-Hispanic tracts . . . . . . . . . . . . . . . . . . . . . . 353

19.2 Local join count significance map . . . . . . . . . . . . . . . . . . . . . . . 354
19.3 Bivariate local join count significance map . . . . . . . . . . . . . . . . . . 356
19.4 Local join count spatial outlier . . . . . . . . . . . . . . . . . . . . . . . . . 356
19.5 Co-location local join count . . . . . . . . . . . . . . . . . . . . . . . . . . 358
19.6 Univariate Quantile LISA . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
19.7 Spatial outliers with Quantile LISA . . . . . . . . . . . . . . . . . . . . . . . 361
19.8 Multivariate Quantile LISA and Multivariate Local Geary . . . . . . . . . 362

20.1 Clusters > DBSCAN | HDBSCAN . . . . . . . . . . . . . . . . . . . . . . 364

20.2 Default heat map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
20.3 Heat map with bandwidth 73 km . . . . . . . . . . . . . . . . . . . . . . . 366
20.4 DBSCAN Core, Border and Noise points . . . . . . . . . . . . . . . . . . . 367
20.5 Connectivity graph for Eps = 20 . . . . . . . . . . . . . . . . . . . . . . . . 368
20.6 DBSCAN clusters for Eps = 20 and MinPts = 4 . . . . . . . . . . . . . . . 368
20.7 DBSCAN clustering settings . . . . . . . . . . . . . . . . . . . . . . . . . . 370
20.8 Default DBSCAN cluster map . . . . . . . . . . . . . . . . . . . . . . . . . . 371
20.9 DBSCAN cluster map, Eps=50, Min Pts=10 . . . . . . . . . . . . . . . . . 372
20.10 DBSCAN-Star dendrogram, Eps=50, Min Pts=10 . . . . . . . . . . . . . . 374
20.11 DBSCAN-Star cluster map, Eps=50, Min Pts=10 . . . . . . . . . . . . . . 375
20.12 Exploring the dendrogram . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
20.13 Level set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
20.14 Core distance and lambda for each point . . . . . . . . . . . . . . . . . . . 379
20.15 Minimum spanning tree connectivity graph . . . . . . . . . . . . . . . . . . 380
20.16 Pruning the MST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
20.17 Cluster stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
20.18 HDBSCAN cluster map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
20.19 HDBSCAN dendrogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
20.20 HDBSCAN condensed tree . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
20.21 HDBSCAN cluster membership . . . . . . . . . . . . . . . . . . . . . . . . 385
20.22 HDBSCAN core distance heat map . . . . . . . . . . . . . . . . . . . . . . 386

A.1 Preferences settings – System . . . . . . . . . . . . . . . . . . . . . . . . . 396

A.2 Preferences settings – Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 397

C.1 libgeoda architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404

Preface

This two-volume set is the long overdue successor to the GeoDa Workbook that I wrote
almost twenty years ago (Anselin, 2005a). It was intended to facilitate instruction in spatial
analysis and spatial regression by means of the GeoDa software (Anselin et al., 2006b). In
spite of its age, the workbook is still widely used and much cited, but it is due for a major
update.
The update is two-fold. On the one hand, many new methods have been developed or original
measures refined. This pertains not only to the spatial autocorrelation indices covered in the
original Workbook but also to a collection of newer methods that have become to define
spatial data science. Secondly, the GeoDa software has seen substantial changes to become an
open-source and cross-platform ecosystem that encompasses a much wider range of methods
than its legacy predecessor.
The two volumes outline my vision for an Introduction to Spatial Data Science. They include
a collection of methods that I view as the core of what is special about spatial data science,
as distinct from applying data science to spatial data. They are not intended to be a
comprehensive overview but constitute my personal selection of materials that I see as
central to promoting spatial thinking through teaching spatial data science.
The level in the current volume is introductory, aimed at my typical audience, which is
largely composed of researchers and students (both undergraduate and graduate) who have
not been exposed to any geographic or spatial concepts or have only limited familiarity with
the subject. So, by design, some of the treatment is rudimentary, covering basic concepts in
GIS and spatial data manipulation, as well as elementary statistical graphs. I have included
this material to keep the books accessible to a larger audience. Readers already familiar
with these topics can easily skip to the core techniques.
I believe the two volumes offer a unique perspective, in that they approach the identification
of spatial patterns from a number of different standpoints. The first volume includes an
in-depth treatment of local indicators of spatial association, whereas Volume 2 focuses on
spatial clustering techniques. The main objective is to indicate where a spatial perspective
contributes to the broader field of data science and what is unique about it. In addition,
the aim is to create an intuition for the type of method that should be applied in different
empirical situations. In that sense, the volumes serve both as the complete user guide to
the GeoDa software and as a primer on spatial data science. However, in contrast with the
original Workbook, spatial regression methods are not included. Those are covered in Anselin
and Rey (2014) and not discussed here.
Most methods contained in the two volumes are treated in more technical detail in the
various references provided. With respect to my own work, these include Anselin(1994; 1995;
1996; 1998; 1999; 2005b), Anselin et al. (2002), Anselin et al. (2004), Anselin et al. (2006b),
and, more recently, Anselin (2019a; 2019b; 2020), Anselin and Li (2019; 2020) and Anselin
et al. (2022). However, a few methods are new and have not been reported elsewhere or are

xxiii
xxiv Preface

discussed here in greater depth than previously appeared. In this volume, these include the
co-location map and the local neighbor match test.
The methods are illustrated with a completely new collection of seven sample data sets that
deal with topics ranging from crime, socio-economic determinants of health, and disease
spread, to poverty, food insecurity and bank performance. The data pertain not only to the
U.S. (Chicago) but also include municipalities in Brazil (the State of Ceará) and in Mexico
(the State of Oaxaca), and community banks in Italy. Many of these data sets were used in
previous empirical analyses. They are included as built-in Sample Data in the latest version
of the GeoDa software.
The empirical illustrations are based on Version 1.22 of the software, available in Summer
2023. Later versions may include slight changes as well as additional features, but the
treatment provided here should remain valid. The software is free, cross-platform and
open-source and can be downloaded from https://geodacenter.github.io/download.html.
Acknowledgments

This work would not have existed without the tremendous efforts by the people behind
the development of the GeoDa software over the past twenty-some years. This started in
the early 2000s in the Spatial Analysis Laboratory at the University of Illinois, with major
contributions by Ibnu Syabri, supported by the NSF-funded Center for Spatially Integrated
Social Science (CSISS). Later, the software development was continued at the GeoDa Center
of Arizona State University with Marc McCann as the main software engineer. The last
ten years, Xun Li has served as the lead developer of the software. In addition, he has also
been a close collaborator on several methodological refinements of the LISA approach. Julia
Koschinsky has been on the team for some twenty years, as a constant inspiration and
collaborator, starting at UIUC and most recently at the Center for Spatial Data Science of
the University of Chicago. Xun and Julia have been instrumental in the migration of GeoDa
from a closed-source Windows-based desktop software to an open-source and cross-platform
ecosystem for exploring spatial data. Julia in particular has been at the forefront of refining
the role of ESDA within a scientific reasoning framework, which I have tried to represent in
the book.
In addition, I would like to thank Lara Spieker from Taylor & Francis Group for her expert
guidance in this project.
Over the years, the research behind the methods covered in this book and the accompanying
software development has been funded by grants from the U.S. National Science Foundation,
the National Institutes for Health and the Centers for Disease Control, as well as by
institutional support by the University of Chicago to the Center for Spatial Data Science.
Finally, Emily has been patiently living with my GeoDa obsession for many years. This book
is dedicated to her.
Shelby, MI, Summer 2023

xxv
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
About the Author

Luc Anselin is the Founding Director of the Center for Spatial Data Science at the University
of Chicago, where he is also the Stein-Freiler Distinguished Service Professor of Sociology
and the College. He previously held faculty appointments at Arizona State University, the
University of Illinois at Urbana-Champaign, the University of Texas at Dallas, the Regional
Research Institute at West Virginia University, the University of California, Santa Barbara,
and The Ohio State University. He also was a visiting professor at Brown University and
MIT. He holds a PhD in Regional Science from Cornell University.
Over the past four decades, he has developed new methods for exploratory spatial data
analysis and spatial econometrics, including the widely used local indicators of spatial
autocorrelation. His 1988 Spatial Econometrics text has been cited some 17,000 times. He
has implemented these methods into software, including the original SpaceStat software, as
well as GeoDa, and as part of the Python PySAL library for spatial analysis.
His work has been recognized by several awards, including election to the U.S. National
Academy of Sciences and the American Academy of Arts and Sciences.

xxvii
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
1
Introduction

Spatial data are special in that the location of the observations, the where, plays a critical
role in the methodology required for their analysis. Two aspects in particular distinguish
spatial data from the standard independent and identically distributed paradigm (i.i.d.)
in statistics and data analysis, i.e., spatial dependence and spatial heterogeneity (Anselin,
1988; 1990). Spatial dependence refers to the similarity of values observed at neighboring
locations, or “everything is related to everything else, but closer places more so,” known
as Tobler’s first law of geography (Tobler, 1970). Spatial heterogeneity is a particular form
of structural change associated with spatial subregions of the data, i.e., showing a clear
break in the spatial distribution of a phenomenon. Both spatial dependence and spatial
heterogeneity require a specialized methodology for data analysis, generically referred to as
spatial analysis.
Spatial data science is an emerging paradigm that extends spatial analysis situated at the
interface between spatial statistics and geocomputation. What the term actually encompasses
is not settled, and the collection of methods and software tools it represents is also sometimes
referred to as geographic data science or geospatial data science (Anselin, 2020; Comber and
Brunsdon, 2021; Singleton and Arribas-Bel, 2021; Rey et al., 2023). The concept is closely
related to, overlaps somewhat with and has many methods and approaches in common
with fields such as geocomputation (Brunsdon and Comber, 2015; Lovelace et al., 2019),
cyberGIScience (Wang, 2010; Wang et al., 2013), and, more recently, GeoAI (Janowicz et al.,
2020; Gao, 2021).
This two-volume collection is intended as an introduction to the field of spatial data science,
emphasizing data exploration and visualization and focusing on the importance of a spatial
perspective. It represents an attempt to promote spatial thinking in the practice of data
science. It is admittedly a selection of methods that reflects my own biases, but it has proven
to be an effective collection over many years of teaching and research. The first volume deals
with the exploration of spatial data, whereas the second volume focuses on spatial clustering
methods.
The methods covered in both volumes work well for so-called small to medium data settings,
but not all of them scale well to big data settings. However, some important principles do
scale well, like local indicators of spatial association. Even though data sets of very large size
have become commonplace and arguably have been the drivers behind a lot of methodological
development in modern data science, this is not always relevant for spatial data analysis. The
point of departure is often big data (e.g., geo-located social media messages), but eventually,
the analysis is carried out at a more spatially aggregate level, where the techniques covered
here remain totally relevant.
The methodological approach outlined in this first volume supports an abductive process
of exploration, a dynamic interaction between the analyst and the data with the goal of
obtaining new insights. The focus is on insights that pertain to spatial patterns in the data,
such as the location of interesting observations (hot spots and cold spots), the presence of

DOI: 10.1201/9781003274919-1 1
2 Introduction

structural breaks in the spatial distribution of the data, and the comparison of such patterns
between different variables and over time.
The identification of the patterns is intended to provide cues about the types of processes
that may have generated them. It is important to appreciate that exploration is not the
same as explanation. In my opinion, exploration nevertheless constitutes an important and
necessary step to obtain effective and falsifiable hypotheses to be used in the next stages
of the analysis. However, in practice, the line between pure exploration and confirmation
(hypothesis testing) is not always that clear, and the process of scientific discovery may
move back and forth between the two. I return to this question in more detail in the closing
chapter.
The two volumes are both an introduction to the methodology of spatial data science and the
definitive guide to the GeoDa software. This software represents the implementation of my
vision for a gradual progression in the exploration of spatial data, from simple description
and mapping to more structured identification of patterns and clusters, culminating with
the estimation of spatial regression models. It came at the end of a series of software
developments that started in the late 1980s (for a historical overview, see Anselin, 2012).
GeoDa is designed to be user-friendly and intuitive, working through a graphical user interface,
and therefore it does not require any programming expertise. Similarly, the emphasis in the
two volumes is on spatial concepts and how they can be implemented through the software,
but it does not deal with geocomputation as such.
A distinctive characteristic of GeoDa is the efficient implementation of dynamically linked
graphs, in the sense that one or more selected observations in a “view” of the data (a graph
or map) are immediately also selected in all the other views, allowing interactive linking and
brushing of the data (Anselin et al., 2006b). Since its initial release in 2003 (through the
NSF-funded Center for Spatially Integrated Science), the software has been adopted widely
for both teaching and research, with close to 600,000 unique downloads at the time of this
writing.
In the remainder of this introduction, I first provide a broad overview of the organization of
this first volume. This is followed by a quick tour of the GeoDa software and a listing of the
sample data sets used to illustrate the methods.

1.1 Overview of Volume 1

The first volume is organized into five main parts and an Epilogue as a sixth, offering
a progression from basic data manipulation, through description and exploration, to the
identification of clusters and outliers by means of spatial autocorrelation analysis. It closes
with some reflections on the limits of exploration and its role in scientific discovery. As
mentioned, spatial clustering methods are covered in Volume 2.
The six parts are:
• Spatial data wrangling
• EDA and ESDA
• Spatial weights
• Global spatial autocorrelation
Overview of Volume 1 3

• Local spatial autocorrelation

• Epilogue
Part I deals with basic data operations for both tabular and spatial data, covered in two
chapters. The material includes a review of the distinctive characteristics of spatial data,
how to create spatial layers inside GeoDa, as well as essential transformations and data
queries. There is also a rudimentary discussion of a range of basic GIS operations, such as
projections, converting between points and polygons, and spatial joins. Even though GeoDa
is not (and not intended to be) a GIS, this functionality has been included over the years in
response to user demand.
Part II covers the principles behind exploratory data analysis (EDA) and its spatial counter-
part, exploratory spatial data analysis (ESDA). This includes six chapters. Three of these are
devoted to map use in various degrees of complexity, starting with basic mapping concepts
and moving to statistical maps and maps for rates. The other three chapters deal with
conventional (non-spatial) EDA, in the form of univariate and bivariate data exploration,
multivariate data exploration and space-time exploration. The core idea here is to leverage
linking and brushing between various graphical representations (views of the data), which is
central to the architecture of GeoDa.
The remaining three main parts deal with the topic of spatial autocorrelation. First, in Part
III, three chapters are devoted to spatial weights, both contiguity-based and distance-based
spatial weights, and various spatial weights operations. These are essential pre-requisites for
the computation of the global and local spatial autocorrelation indices covered in Parts IV
and V.
Part IV contains three chapters on global spatial autocorrelation, centered around the
Moran scatter plot as a visualization device. The basic concepts are covered, as well as more
advanced applications and extensions to a bivariate setting. The third chapter provides an
overview of some non-parametric techniques, such as a spatial correlogram.
Part V includes an in-depth treatment of local spatial autocorrelation, spread over five
chapters. It starts with the introduction of the concept of a LISA and the Local Moran
statistic. The second chapter deals with other local spatial autocorrelation statistics, such as
the Local Geary and the Getis-Ord statistics. The next two chapters outline extensions to
the multivariate domain and to discrete variables. These chapters contain material that was
only fairly recently developed. The last chapter of Part V reviews density-based clustering
methods applied to point locations, such as DBScan and HDBScan.
The Epilogue offers some thoughts on the limits of the exploratory perspective. This includes
an assessment of the role of data exploration in aiding with scientific discovery and scientific
reasoning, the limits of spatial analysis, and reproducibility in the exploratory framework as
implemented in the GeoDa software.
An Appendix includes detailed preference settings for the software and an outline of the
complete menu structure. To close, a brief discussion is offered of the new scripting possibilities
through the geodalib library.
The division of the material in two volumes follows my own teaching practice. The first
volume corresponds to what I cover in an Introduction to Spatial Data Science course,
whereas the second volume matches the content of a Spatial Cluster Analysis course. The
volumes are also designed to constitute a self-study guide. In fact, a previous version was
used as such for remote teaching during the Covid pandemic (in the form of laboratory
workbooks, available at https://geodacenter.github.io/documentation.html).
4 Introduction

In addition to the material covered in the two volumes, the GeoDaCenter Github site
(https://geodacenter.github.io) contains an extensive support infrastructure. This includes
detailed documentation and illustrations, as well as a large collection of sample data sets,
cookbook examples, and links to a YouTube channel containing lectures and tutorials.
Specific software support is provided by means of a list of frequently asked questions and
answers to common technical questions, as well as by the community through the Google
Groups Openspace list.

1.2 A Quick Tour of GeoDa

Before delving into the specifics of particular methods, I provide a broad overview of the
functionality and overall organization of the GeoDa software. The complete toolbar with
icons corresponding to a collection of related operations is shown in Figure 1.1. Each icon
is matched by a menu item, detailed in Appendix B. The menu and user interface can be
customized to several languages (details are in Appendix A). The default is English, but
options are available for Simplified Chinese, Russian, Spanish, Portuguese and French, with
more to come in the future.
With each toolbar icon typically corresponds a drop-down list of specific functions. The
structure of the drop-down list matches the menu sub-items (Appendix B).

Figure 1.1: GeoDa toolbar icons

The organization of the toolbar (and menu) follows the same logic as the layout of the parts
and chapters in the two books. It represents a progression in the exploration, from left to
right, from support functions to queries, description and visualization, and more and more
formal methods, ending up with the estimation of actual spatial models in the regression
module (not covered here).
A brief overview of each of the major parts is given next. This also includes the spatial
clustering functionality, which is discussed more specifically in Volume 2.

1.2.1 Data entry

Figure 1.2: Data entry

The three left-most icons, highlighted in Figure 1.2, deal with data entry and general input-
output. This includes the loading of spatial and non-spatial (e.g., tabular) data layers from
a range of GIS and other file formats (supported through the open-source GDAL library).
In addition, it offers connections to spatial databases, such as PostGIS and Oracle Spatial.
It also supports a Save As function, which allows the software to work as a GIS file format
converter. Further details are provided in Chapter 2.
A Quick Tour of GeoDa 5

1.2.2 Data manipulation

Figure 1.3: Data manipulation/table

Functionality for data manipulation and transformation is provided by the Table icon,
highlighted in Figure 1.3. This allows new variables to be created, observations selected,
queries formulated and includes other data table operations, such as merger and aggregation,
detailed in Chapter 2.

1.2.3 GIS operations

Figure 1.4: GIS operations/tools

Spatial data operations are invoked through the Tools icon, highlighted in Figure 1.4. These
include many GIS-like operations that were added over the years to provide access to spatial
data for users who are not familiar with GIS. For example, point layers can be easily created
from tabular data with X,Y coordinates, point in polygon operations support a spatial join,
an indicator variable can be used to implement a dissolve application, and reprojection can
be readily implemented by means of a Save As operation. Specific illustrations are included
in Chapter 3.

1.2.4 Weights manager

Figure 1.5: Weights manager

The Weights Manager icon, Figure 1.5, contains a final set of functions that are in
support of the analytical capabilities. It gives access to a wide range of weight creation
and manipulation operations, discussed at length in the chapters of Part III. This includes
constructing spatial weights from spatial layers, as well as loading them from external files,
summarizing and visualizing their properties, and operations like union and intersection.

1.2.5 Mapping and geovisualization

Figure 1.6: Mapping and geovisualization

The mapping and geovisualization functionality is represented by four icons, highlighted in

Figure 1.6: the Map icon, Cartogram, Map Movie and Category Editor. The mapping
function supports all the customary types of choropleth maps, as well as some specialized
features, such as extreme value maps, co-location maps and smoothed maps for rates. The
6 Introduction

cartogram is a specialized type of map that replaces the actual outline of spatial units by a
circle, whose area is proportional to a given variable of interest. Animation, in the sense of
moving through the locations of observations in increasing or decreasing order of the value
for a given variable is implemented by means of the map movie icon. Finally, the category
editor provides a way to design custom classifications for use in maps as well as in statistical
graphs, such as a histogram. Details are provided in Chapters 4 through 6.

1.2.6 Exploratory data analysis

Figure 1.7: Exploratory data analysis

The next eight icons, grouped in Figure 1.7, contain the functionality for exploratory data
analysis and statistical graphs. This includes a Histogram, Box Plot, Scatter Plot,
Scatter Plot Matrix, Bubble Chart, 3D Scatter Plot, Parallel Coordinate Plot
and Conditional Plots. These provide an array of methods for univariate, bivariate and
multivariate exploration. All the graphs are connected to any other open window (graph or
map) for instantaneous linking and brushing. This is covered in more detail in Chapters 7
and 8.

1.2.7 Space-time analysis

Figure 1.8: Space-time analysis

The exploration of space-time data, treated in Chapter 9, is invoked by means of the icons
on the right, highlighted in Figure 1.8. This includes a Time Editor, which is required
to transform the cross-sectional observations into a proper (time) sequence. In addition,
the Averages Chart implements a simple form of treatment analysis, with treatment and
controls defined over time and/or across space.

1.2.8 Spatial autocorrelation analysis

Figure 1.9: Spatial autocorrelation analysis

Spatial autocorrelation analysis is invoked through the three icons highlighted in Figure 1.9.
The first two pertain to global spatial autocorrelation. The left-most icon corresponds to
various implementations of the Moran scatter plot (Chapters 13 and 14). The middle icon
invokes nonparametric approaches to visualize global spatial autocorrelation as a spatial
correlogram and distance scatter plot (Chapter 15).
The third icon contains a long list of various implementations of local spatial autocorrelation
statistics, including various forms of the Local Moran’s I, the Local Geary c, the Getis-Ord
statistics and extensions to multivariate settings and discrete variables. The local neighbor
Sample Data Sets 7

match test is a new method based on an explicit assessment of the overlap between locational
and attribute similarity. Details are provided in the chapters of Part V.

1.2.9 Cluster analysis

Figure 1.10: Cluster analysis

Finally, cluster analysis is invoked through the icon highlighted in Figure 1.10. An extensive
drop-down list also includes the density-based cluster methods DBScan and HDBScan, which
are treated in this volume under local spatial autocorrelation (Chapter 20).
The other methods are covered in Volume 2. They include dimension reduction, classic
clustering methods and spatially constrained clustering methods. The last items in the drop-
down list associated with the cluster icon pertain to the quantitative and visual assessment
of cluster validity, including a new cluster match map (see Volume 2).

1.3 Sample Data Sets

As mentioned in the Preface, the methods and software are illustrated by means of empirical
examples that use seven new sample data sets. They are available directly from inside the
GeoDa software through the Sample Data tab of the input/output interface (see Figure
2.2).
The specific data sets are:
• Chicago Carjackings (n = 1,412)
– point locations of carjackings in 2020 (Chicago Open Data Portal)
– see Chapters 2 and 3
• Ceará Zika, municipalities in the State of Ceará, Brazil (n = 184)
– Zika and Microcephaly infections and socio-economic profiles for 2013–2016 (adapted
from Amaral et al., 2019)
– see Chapters 4–6, 10 and Part III of Volume 2 (Spatial Clustering)
• Oaxaca Development, municipalities in the State of Oaxaca, Mexico (n = 570)
– poverty and food insecurity indicators and census variables for 2010 and 2020
(CONEVAL and INEGI) (based on the same original sources as Farah Rivadeneyra,
2017)
– see Chapters 7–9, 12, 14 and 16–17
• Italy Community Banks (n = 261)
– bank performance indicators for 2011–17 (used by Algeri et al., 2022)
– see Chapters 11–12, 15 and 20, as well as in Part I of Volume 2 (Dimension
Reduction)
• Chicago Community Areas, CCA Profiles (n = 77)
– socio-economic snapshot for Chicago Community Areas in 2020 (American Commu-
nity Survey from the Chicago Metropolitan Agency for Planning – CMAP – data
portal)
– see Chapters 13 and Chapter 5 of Volume 2 (Hierarchical Clustering Methods)
8 Introduction

• Chicago SDOH, census Tracts (n = 791)

– socio-economic determinants of health in 2014 (a subset of the data used in Kolak
et al., 2020)
– see Chapters 18–19 and Chapters 6 and 7 of Volume 2 (Partitioning Clustering
Methods and Advanced Clustering Methods)
• Spirals (n = 300)
– canonical data set to test spectral clustering
– only used in Volume 2 (Chapter 8, Spectral Clustering)
In addition, a few auxiliary files are employed to illustrate basic data handling operations in
Chapters 2 and 3, such as a boundary layer for Chicago community areas and input data
files in comma-separated text format. These files are available from the GeoDaCenter sample
data site at https://geodacenter.github.io/data-and-lab/.
Further details are provided in the context of specific methods.
 Part I

Spatial Data Wrangling

Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
2
Basic Data Operations

In this and the following chapter, I introduce the topic of data wrangling, i.e., the process
of getting data from its raw input into a form that is amenable for analysis. This is often
considered to be the most time consuming part of a data science project, taking as much
as 80% of the effort (Dasu and Johnson, 2003). Even though the focus in this book is on
analysis and not on data manipulation per se, I provide a quick overview of the functionality
contained in GeoDa to assist with these operations. Increasingly, data wrangling has evolved
into a field of its own, with a growing number of operations turning into automatic procedures
embedded into software (Rattenbury et al., 2017). A detailed discussion of this topic is
beyond the scope of the book.
The coverage in this chapter is aimed at novices who are not very familiar with spatial
data manipulations. Most of the features illustrated can be readily accomplished by means
of dedicated GIS software or by exploiting the spatial data functionality available in the
R and Python worlds. Readers knowledgeable in such operations may want to just skim
the materials in order to become familiar with the way they are implemented in GeoDa.
Alternatively, these operations can be performed outside GeoDa, with the end result loaded
as a spatial data layer.
In the current chapter, I focus on essential input operations and data manipulations contained
in the Table functionality. In the next chapter, I consider a range of basic GIS operations
pertaining to spatial data wrangling.
To illustrate these features, I will use a data set with point locations of car jackings in
Chicago in 2020. The Chicago Carjackings data layer is available from the Sample Data
tab in the GeoDa file dialog (Figure 2.2).
In addition, in order to replicate the detailed steps used in the illustrations,
three original input files are needed as well. These are available from the GeoDa-
Center sample data site. They include a simple outline of the community areas,
Chicago_community_areas.shp, as well as comma delimited (csv) text files with the socio-
economic characteristics (Chicago_CCA_Profiles.csv), and the coordinates of the car jackings
(Chicago_2020_carjackings.csv). The sample data site also contains the detailed listing of
the variable names.

2.1 Topics Covered

• Load a spatial layer from a range of formats
• Convert between spatial formats
• Create a point layer from coordinates in a table
• Create a grid layer

DOI: 10.1201/9781003274919-2 11
12 Basic Data Operations

• Become familiar with the table options

• Use the Calculator Tool to create new variables
• Variable standardization
• Merging tables
• Use the Selection Tool to select observations in a table
• Use a selection shape to select observations in a map
GeoDa Functions
• File > Open
• File > Save
• File > Save As
• Tools > Shape > Points from Table
• Tools > Shape > Create Grid
• Table > Edit Variable Properties
• Table > Add Variable
• Table > Delete Variable(s)
• Table > Rename Variable
• Table > Encode
• Table > Setup Number Formatting
• Table > Move Selected to Top
• Table > Calculator
• Table > Selection Tool
• File > Save Selected As
• Map > Unique Values Map
• Map > Selection Shape
• Map > Save Selection
Toolbar Icons

Figure 2.1: Open | Close | Save | Table | Tools

2.2 Spatial Data

Spatial data are characterized by the combination of two important aspects. First, there is
information on variables, just as in any other statistical analysis. In the spatial world, this is
referred to as attribute information. Typically, it is contained in a flat (rectangular) table
with observations as rows and variables as columns.
The second aspect of spatial data is special and is referred to as locational information.
It consists of the precise definition of spatial objects, classified as points, lines or areas
(polygons). In essence, the formal characterization of any spatial object boils down to the
description of X-Y coordinates of points in space, as well as of a mechanism that spells out
how these points are combined into spatial entities.
For a single point, the description simply consists of its coordinates. For areal units, such as
census tracts, counties, or states, the associated polygon boundary is defined as a series of
Spatial Data 13

line segments, each characterized by the coordinates of their starting and ending points. In
other words, what may seem like a continuous boundary, is turned into discrete segments.
Traditional data tables have no problem including X and Y coordinates as columns, but
as such cannot deal with the boundary definition of irregular spatial units. Since the
number of line segments defining an areal boundary can easily vary from observation to
observation, there is no efficient way to include this in a fixed number of columns of a
flat table. Consequently, a specialized data structure is required, typically contained in a
geographic information system or GIS.
Several specialized formats have been developed to efficiently combine both the attribute
information and the locational information. Such spatial data can be contained in files with
a special structure, or in spatially enabled relational data base systems.
I first consider common GIS file formats that can serve as input to GeoDa. This is followed
by an illustration of simple tabular input of non-spatial files. Finally, a brief overview is
given of connections to other input formats.

2.2.1 GIS files

Historically, a wide range of different formats have been developed for GIS data, both propri-
etary as well as open-source. In addition, there has been considerable effort at standardization,
led by the Open Geospatial Consortium (OGC).1 GeoDa leverages the open-source GDAL
library2 to support input and output of many of the most popular formats in use today.
While it is impossible to cover all of these specifications in detail, I will illustrate three
specific formats here. First is the use of the proprietary shape file format of the leading
GIS vendor ESRI.3 In addition, the open-source GeoJSON format4 will be covered, as well
as the Geography Markup Language of the OGC, a standard XML grammar for defining
geographical features.5
In GeoDa, one can load both polygon and point GIS data, but in the current implementation,
line files are not supported (e.g., to represent road networks).

2.2.1.1 Spatial file formats

Arguably, the most familiar proprietary spatial data format is the shape file format, developed
by ESRI. The terminology is a bit confusing, since there is no such thing as one shape file,
but there is instead a collection of three (or four) files. One file has the extension .shp, one
.shx, one .dbf and one .prj (with the projection information). The first three are required,
the fourth one is optional, but highly recommended. The files should all be in the same
directory and have the same file name, except for the file extension.
In the open-source world, an increasingly common format is GeoJSON, the geographic
augmentation of the JSON standard, which stands from JavaScript Object Notation. This
format is contained in a text file and is easy for machines to read, due to its highly structured
nature.
Finally, the GML standard, or Geographic Markup Language, is a XML implementation
that prescribes the formal description of geographic features.
1 https://www.ogc.org
2 https://gdal.org
3 https://www.esri.com/library/whitepapers/pdfs/shapefile.pdf
4 https://geojson.org
5 https://www.ogc.org/standards/gml
14 Basic Data Operations

Figure 2.2: Connect to Data Source dialog

Figure 2.3: Supported spatial file formats

A detailed discussion of the individual formats is beyond the current scope. All are well-
documented, with many additional resources available online. Although it is always helpful,
there is no need to know the underlying formats in detail in order to use GeoDa, since the
interaction with the data structures is handled under the hood.
The main file manipulations are invoked from the File item in the menu, or by the three
left-most icons on the toolbar in Figure 2.1.

2.2.1.2 Polygon layers

Since GeoDa is particularly geared to the exploration of areal unit data, the input of a
so-called polygon layer is illustrated first. Any spatial layer present as a file can be loaded
by invoking File > Open File from the menu, or by clicking on the left-most Open icon
on the toolbar in Figure 2.1.
This brings up the Connect to Data Source dialog, shown in Figure 2.2. The left panel
has File as the active input format. Other formats are Database and Web, which are
briefly covered in Section 2.2.3. The right panel shows a series of Sample Data data
that are included with GeoDa. In addition, after some files have been loaded in the current
application, the Recent panel will contain their file names as well. Files listed in either
panel can be loaded by simply clicking on the corresponding icon.
The small folder icon to the right of the Input file box brings up a list of supported file
formats, as in Figure 2.3. In this first example, the top item in the list is selected, ESRI
Shapefile (*.shp).
To illustrate this feature, the four files associated with the Chicago_community_areas
shape file must be available in a working directory (they must be downloaded from the
GeoDaCenter sample data site).
Spatial Data 15

Figure 2.4: Themeless polygon map

Figure 2.5: Example GeoJSON file contents

Using the navigation dialog and conventions appropriate for each operating system, the
shape file can be selected from this directory. This opens a new map window with the
spatial layer represented as a themeless choropleth map, as in Figure 2.4. The number of
observations is shown in parentheses next to the small green rectangle in the upper-left
panel, as well as in the status bar at the bottom (#obs = 77).
The current layer is cleared by clicking on the Close toolbar icon, the second item on the
left in Figure 2.1, or by selecting File > Close from the menu. This removes the base map.
At this point, the Close icon on the toolbar becomes inactive.
A more efficient way to open files is to select the file name in the directory window and
to drag it onto the Drop files here box in the dialog. Even easier is to load a one of the
sample data sets or a recently used one, where a simple click on the associated icon in the
Sample Data or Recent tab suffices.
In contrast to the shape file format, which is binary, a GeoJSON file is simple text and can
easily be read by humans. As shown for the Chicago_community_areas.geojson file from
the sample data site in Figure 2.5 (this file must be downloaded to a working directory),
the locational information is combined with the attributes. After some header information
follows a list of features. Each of these contains properties, of which the first set consists
16 Basic Data Operations

Figure 2.6: Example GML file contents

of the different variable names with the associated values, just as would be the case in
any standard data table. The final item refers to the geometry. This includes the type,
here a MultiPolygon, followed by a list of X-Y coordinates. In this fashion, the spatial
information is integrated with the attribute information.
To view the corresponding map, the Chicago_community_areas.geojson file name can be
selected in its directory and dragged onto the Drop files here box. This brings up the
same base map as in Figure 2.4.

2.2.1.3 File format conversion

The file just loaded was originally specified in the GeoJSON format. It can be easily converted
to a different format by means of the File > Save As functionality. For example, to change
it into a GML format file (e.g., for use in a different program), Geographic Markup
Language (*.gml) can be selected from the drop-down list of available formats, shown in
Figure 2.3.
This will yield a text file in the GML XML format, illustrated in Figure 2.6. It shows the
characteristic < > and </ > delimiters of the markup elements, typical of XML files. In
the file snippet in the figure, the top lines pertain to the geography of the first polygon,
ended by </ogr:geometryProperty>. Next follow the actual observations, with variable
names and associated values, finally closed off with </ogr:featureMember>. After this, a
new observation is listed, delineated by the <ogr:featureMember> tag, followed by the
geographic characteristics. Again, this illustrates how spatial information is combined with
attribute information in an efficient file format.
In sum, the File > Save As feature in GeoDa turns the program into an effective GIS
format converter.

2.2.1.4 Point layers

In the same fashion as for polygon layers, spatial data layers containing point locations
can be loaded for the file formats listed in Figure 2.3. As before, the file is either selected
explicitly from the proper directory, or the file name is dragged directly into the Drop files
here box in the dialog.
Spatial Data 17

Figure 2.7: Themeless point map

The point map in Figure 2.7 shows the locations of the 1,412 carjackings that oc-
curred in the City of Chicago during the year 2020. It is generated by clicking on the
Chicago Car Jackings icon in the Sample Data tab, or by dragging the file name
Chicago_carjack_2020_pts.shp from a working directory that contains the shape file.
The shape of the city portrayed by the outline of the points is slightly different from that in
the polygon map in Figure 2.4. This is due to a difference in projections: the point map is
in the State Plane Illinois East NAD 1983 projection (EPSG 3435), whereas the polygon
map uses decimal degrees latitude and longitude (EPSG 4326). This important aspect of
spatial data is often a source of confusion for non-GIS specialists. GeoDa provides an intuitive
interface to deal with projection issues. I return to this topic in Section 2.3.1.1 below and in
Section 3.2 in the next chapter.

2.2.2 Tabular files

In addition to GIS files, GeoDa can also read regular non-spatial tabular data. While this does
not allow for spatial analysis (unless coordinates are contained in the table, see Section 2.3.1),
all non-spatial operations and graphs are supported. Specifically, all standard techniques
of exploratory data analysis (EDA) can be applied, as covered in Chapters 7 and 8 in this
Volume. This does not necessitate a map layer.
The data for the Community area socio-economic profiles are contained in a comma-separated
file (csv format) on the sample data site (the Chicago CCA Profiles are also available as a
spatial layer in the Sample Data tab). Selecting this file (after downloading it to a working
directory) generates the dialog shown in Figure 2.8.
Since a csv file is pure text, there is no information on the type of the variables included in
the file. GeoDa tries to guess the type and lists the Data Type for each field, as well as a
brief preview of the table. At this point, the type can be changed before the data are moved
into the actual data table (see also Section 2.4.1.1).
Instead of a base map, which is the default opening window for spatial data, the data
table is brought up in a spreadsheet-like format (see also Figure 2.9 for an illustration). In
18 Basic Data Operations

Figure 2.8: CSV format file input dialog

addition to comma-separated files, GeoDa also supports tabular input from dBase database
files, Microsoft Excel and Open Document Spreadsheet (*.ods) formatted files. As is the case
for the various GIS formats, the File > Save As command allows for the ready conversion
from one tabular format to another (e.g., from csv to dBase).
An additional feature of the input dialog in Figure 2.8 is inclusion of optional Longitude/X
and Latitude/Y drop-down lists. With coordinate variables specified, this will create a
point layer. As such, it provides an alternative to the approach outlined in Section 2.3.1.

2.2.3 Other spatial data input

In addition to file input, GeoDa can connect to a number of spatially enabled relational data
bases, such as PostgesQL/PostGIS, Oracle Spatial and MySQL Spatial. This is available
through the Database button in the Connect to Data Source interface, the middle tab
in the dialog shown in Figure 2.2. Each data base system has its own requirements in terms
of specifying the host, port, user name, password and other settings. Once the connection is
established, the data can also be saved to a file format. In the current version of GeoDa, the
data base connection is limited to loading a single table. So far, no other SQL commands
are supported.
A third data source option is referred to as Web in the interface, the right-most tab in
Figure 2.2. It allows data to be loaded directly from a GeoJSON URL, or, alternatively,
from the older web feature server WFS URL.
Creating Spatial Layers 19

Figure 2.9: Coordinate variables in data table

Figure 2.10: Specifying the point coordinates

2.3 Creating Spatial Layers

In addition to loading spatial layers from existing files or data base sources, new spatial
layers can be created as well. This is implemented through the Tools functionality.
It is invoked from the menu as Tools > Shape, or, alternatively, from the Tools icon on
the toolbar, the right-most icon in Figure 2.1. Two different operations are supported: one is
to turn X-Y coordinates from a table into a point layer, Tools > Shape > Points from
Table, the other creates a rectangular shape of grid cells,Tools > Shape > Create Grid.

2.3.1 Point layers from coordinates

Point layers are created from two variables contained in a data table that can serve as X and
Y coordinates. This is illustrated with the file Chicago_2020_carjackings.csv, downloaded
from the GeoDaCenter sample data collection. This file contains the same information as
Chicago Carjacking in the Sample Data tab, but is a simple comma-separated text file.
Once loaded, the contents of the data table are shown in a spreadsheet-like format, as
illustrated in Figure 2.9. In the example, the latitude and longitude of the point locations
(in decimal degrees) are included as the two last columns in the table. Even though these
variables are commonly referred to as lat-lon, latitude is actually the vertical dimension (Y),
while longitude is the horizontal dimension (X).
20 Basic Data Operations

Figure 2.11: Point layer from coordinates

Figure 2.12: Blank CRS

The Tools > Shape > Points from Table command opens up a dialog that lists all
numerical variables contained in the data table. In the example, shown in Figure 2.10, the
variables to be selected are Longitude and Latitude.
After specifying the coordinate variables and clicking on OK, a point map appears, as in
Figure 2.11.
The outline of Chicago suggested by the points is different from the one in Figure 2.7. Again,
this is due to the difference in projections, which is discussed next.
At this stage, the point layer can be saved as a GIS file, such as a shape file, using the
familiar File > Save As command (the file name should differ from any existing carjackings
point shape file, or the latter will be overwritten).

2.3.1.1 A first note about projections

At the bottom of the dialog to save the just created layer is an option to enter CRS (proj4
format). This refers to the projection information, or, more precisely, to the coordinate
reference system. Projections will be covered in more detail in the next chapter, but at this
point it is useful to be aware of the implications.
As mentioned, a careful comparison of the point pattern in Figure 2.11 and in Figure 2.7
reveals that they are not identical. While the relative positions of the points are the same,
they appear more spread out in Figure 2.11. The reason for the discrepancy is the use
of different projections. The point shape file in Figure 2.7 is based on the projected X-Y
coordinates, whereas the just created point layer uses the latitude-longitude decimal degrees.
Creating Spatial Layers 21

Figure 2.13: CRS with projection information

Figure 2.14: Grid creation dialog

As a default, the Coordinate Reference System (CRS) entry in the file save dialog will
be empty, as in Figure 2.12. In other words, since nothing was specified other than the
coordinates, there is no projection information. As a result, if the layer was saved as a shape
file, it would not include a file with a prj file extension.
The lack of projection information severely hampers a range of data manipulations. Specifi-
cally, for computations such as Euclidean (straight line) distance, the decimal degrees are
inappropriate and either special calculations must be used (such as great circle distance), or
the coordinates need to be reprojected.
Even though latitude-longitude decimal degrees are not an actual projection, there is a
corresponding CRS. While in and of itself this information is not that meaningful, including
it when saving the file makes future reprojections easy. The CRS in proj4 format for decimal
degrees is (this is discussed in more detail in the next chapter):
+proj=longlat +datum=WGS84 + no_defs.
Including this entry in the CRS box (see Figure 2.13) when saving the file will ensure that
future reprojections will be possible.
In other words, in the absence of specific projection information, the best practice is to
create a spatial layer using decimal latitude-longitude degrees and recording the CRS from
Figure 2.13 when saving the information as a GIS format file.
This is pursued in greater technical detail in the next chapter.
22 Basic Data Operations

Figure 2.15: Grid layer over Chicago community areas

2.3.2 Grid
The second type of spatial layer that can be created is a set of grid cells. Such cells are
rarely of interest in and of themselves, but are useful to aggregate counts of point locations
in order to calculate point pattern statistics, such as quadrat counts.
The grid creation is one of the few instances where GeoDa functionality can be invoked from
the menu or toolbar without any table or layer open. From the menu, Tools > Shape
> Create Grid brings up a dialog that contains a range of options, shown in Figure
2.14. The most important aspect to enter is the Number of Rows and the Number of
Columns, listed under Grid Size at the bottom of the dialog. This determines the number
of observations. In the example, the entries are 10 rows and 5 columns, for 50 observations.
There are four options to determine the Grid Bounding Box, i.e., the rectangular extent
of the grids. The easiest approach is to use the bounding box of a currently open layer or of
an available shape data source. The Lower-left and Upper-right corners of the bounding
box can also be set manually, or read from an ASCII file.
In the example in Figure 2.14, the Chicago community area boundary layer, contained in
the Chicago_community_areas.shp file is shown as the source for the bounding box.
After the grid layer is saved as a file, it can be loaded into GeoDa. In Figure 2.15, the result
is shown, superimposed onto the community areas layer (how to implement this will be
covered in the next chapter). Clearly, the extent of the grid cells matches the bounding box
for the area layer.

2.4 Table Manipulations

A range of variable transformations is supported through the Table functionality. This
requires that the data table associated with a spatial layer is open. In contrast to the other
Table Manipulations 23

Figure 2.16: Table options

operations in GeoDa, opening a data table can only be accomplished by selecting a toolbar
icon, i.e., the second icon from right on the toolbar shown in Figure 2.1.
With the table open, the options can be accessed from the menu, by selecting the Table
menu item, or by right clicking on the table itself. This brings up the list shown in Figure
2.16.
In addition to a range of variable transformations, the options menu also includes the
Selection Tool, which is the way in which data queries can be carried out. I discuss this in
more detail in Section 2.5. The more commonly used variable manipulations are covered in
what follows.

2.4.1 Variable properties

2.4.1.1 Edit variable properties
One of the most used initial transformations pertains to getting the data into the right format.
Often, observation identifiers, such as the FIPS code used by the U.S. census, are recorded
as character or string variables. In order to use these variables effectively as observation
identifiers, they need to be converted to a numeric format. This is accomplished through the
Edit Variable Properties option.
After this option is invoked by clicking on it, a window appears with a list of all variables,
listing their type, as well as a number of other properties, like precision.
For example, consider the data table associated with the Chicago_community_areas commu-
nity area boundaries, shown in Figure 2.17, after selecting the Edit Variable Properties
option.
In the table, the area identifiers – both area_num_1 and area_numbe (which are
identical) – are left-aligned in their columns. This indicates they are character variables and
not in a numeric format.
By selecting the proper integer format from the drop-down list, the change is immediate.
As mentioned, the most common use of this functionality is to change string variables to a
numeric format.
In order to make the change permanent, the table needs to be saved, either by selecting the
middle icon in the toolbar in Figure 2.1, or by means of File > Save. Note that this will
24 Basic Data Operations

Figure 2.17: Edit variable properties

overwrite the current files. To avoid this behavior, select File > Save As and specify a
new file name.
To facilitate data table operations like Merge, the community area identifier must be
changed to integer.

2.4.1.2 Formatting date/time variables

One special case where the Edit Variable Properties option is extremely useful is when
information on the date and time of an observation is included in the data table. Typically,
this is contained in a string variable, which does not allow for any operations such as searching
or sorting. Instead, a special date/time format must be used, which requires converting from
string to this format.
For example, this is the case for the observations in the Chicago Carjackings points data
set. The third column of the data table in Figure 2.18 shows the Date as a string variable.
Specifically, the first entry is 12/11/2020 11:45:00 PM, which lists the month, day and
year, separated by a forward slash (/), followed by the time in hours, minutes and seconds
(separated by a colon) and AM or PM.
Selecting datetime from the drop-down list will change the format into a standard time
expression, as in Figure 2.19. Note how the order of the information has changed to year-
month-day and the time is in 24 hours, minutes and seconds (no more AM or PM). At this
point, the date information can be manipulated by means of the Date/Time functions in
the calculator (see Section 2.4.2.5).
However, there is a catch. GeoDa recognizes many date-time formats, but not all. The default
ones are listed under the Data tab of the GeoDa Preference Setup (see also Appendix
A). Sometimes, the particular format used is not part of the default. It therefore may have
to be added to the list, as shown in Figure 2.20. For example, the proper entry for the
carjacking format is %m/%d/%Y %I:%M:%S %p, the last item in the Date/Time
formats list. This follows the standard format for date and time information.6
6 Specifically, %m stands for two digit month (including a leading zero), %d for two digit day (again,

including a leading zero), %Y for the year in full four digits, %I for the hour in 12 hour segments (this
requires AM or PM in addition), %M for minutes in two digits, %S for seconds in two digits and %p for AM
or PM.
Table Manipulations 25

Figure 2.18: Create a date/time variable

Figure 2.19: Date/time format

Once a customized format is included into the setup interface, it will be recognized in all
future instances of datetime format conversion.

2.4.1.3 Other variable operations

The other variable operations listed in the drop-down list in Figure 2.16 are mostly self-
explanatory. For example, a new variable can be included through Add Variable (its value
is obtained through the Calculator, see Section 2.4.2), or variables can be removed by
means of Delete Variable(s). Rename Variable operates on a specific column, which
needs to be selected first.
Two more obscure options are Encode and Setup Number Formatting. The default
encoding in GeoDa is Unicode (UTF-8), but a range of other encodings that support
non-western characters are available as well. The default numeric formatting is to use a
period for decimals and to separate thousands by a comma, but the reverse can be specified
as well, which is common in Europe.

2.4.1.4 Operations on the table

The columns, rows and cells of the table can be manipulated directly, similar to spreadsheet
operations. For example, the values in a column can be sorted in increasing or decreasing
order by clicking on the variable name. A > or < sign appears next to the variable name to
26 Basic Data Operations

Figure 2.20: Customizing the date/time format

indicate the sorting order. The original order can be restored by sorting on the row numbers
(the left-most column).
Individual cell values can be edited by selecting the cell and entering a new value. As is
always the case in GeoDa, such changes are only permanent after the table is saved.
Specific observations can be selected manually by clicking on their row number (selection
through queries is covered in Section 2.5). The corresponding entries will be highlighted in
yellow. Additional observations can be added to the selection in the usual way, for example,
by means of shift-click (or another key combination, depending on the operating system).
The selected observations can be located at the top of the table for easy reference by means
of Move Selected to Top.
Finally, the arrangement of the columns in the table can be altered by dragging the variable
name and moving it to a different position. This is often handy to locate related variables
next to each other in the table for easy comparison when the original layout is not convenient.

2.4.2 Calculator
The Calculator functionality provides a rudimentary interface to create new variables,
transform existing variables and carry out a number of algebraic operations. It is limited to
one operation at a time, so it is not suitable for bulk programming.
The Calculator interface contains six tabs, dealing, respectively, with Special functions,
Univariate and Bivariate operations, Spatial Lag, Rates and Date/Time operations.
Spatial Lag and Rates are advanced functions that are discussed separately in Chapters 6
and 12.
The calculator interface is shown in Figure 2.21, illustrated for the Bivariate operations.
Each tab at the top includes a series of functions, either creating a new variable (Special),
operating on a single variable (Univariate), two variables (Bivariate), or values contained
in a special Date/Time format. The specific functions are selected from the Operator
drop-down list.
Each operation begins by identifying a target variable in which the result of the operation
will be stored. Typically, this will be a new variable, which is created by means of the Add
Variable button located to the right of Result.7 This brings up an interface in which the
Name for the new variable is specified, as well as its Type, its precision, and where to
7A new variable can also be created directly from the table options menu as Table > Add Variable.
Table Manipulations 27

Figure 2.21: Calculator interface

insert it into the table. The default is to place a new variable to the left of the first column
in the table, but any other variable can be taken as a reference point (i.e., the new column is
placed to the left of the specified variable). In addition, the new variable can also be placed
at the right most end of the table (the last option in the list).
With the target variable specified, the actual calculations can be carried out by selecting
the operation from the drop-down list and choosing the variables involved. Again, to make
the new variable permanent, the data table needs to be saved.
The various operations are briefly reviewed next.

2.4.2.1 Special functions

Special functions create a new variable, either a random variable, NORMAL or UNI-
FORM RANDOM, or the rank order of the current observations. The latter is generated
by the ENUMERATE function. This yields a new variable with the current order of the
observations. It is especially useful to retain the order of observations after sorting on a
given variable, such as the rank order corresponding to a given statistic.

2.4.2.2 Univariate functions

Operations that pertain to a single variable are included in the Univariate drop-down list.
ASSIGN allows a variable to be set equal to any other variable, or to a constant (the typical
use). The next five operations are straightforward transformations, including NEGATIVE
(change the sign), INVERT, SQUARE ROOT, and base 10 and natural LOG.
SHUFFLE randomly permutes the values for a given variable to different observations.
This is an efficient way to implement spatial randomness, i.e., an allocation of values to
locations, but where the location itself does not matter (any location is equally likely to
receive a given observation value).

2.4.2.3 Variable standardization

The univariate operations also include five types of variable standardization. The most
commonly used is undoubtedly STANDARDIZED (Z). This converts the specified variable
such that its mean is zero and variance one, i.e., it creates a z-value as
(x − x̄)
z= ,
σ(x)

with x̄ as the mean of the original variable x, and σ(x) as its standard deviation.
28 Basic Data Operations

A subset of this standardization is DEVIATION FROM MEAN, which only computes

the numerator of the z-transformation.
An alternative standardization is STANDARDIZED (MAD), which uses the mean
absolute deviation (MAD) as the denominator in the standardization. This is preferred in
some of the clustering literature, since it diminishes the effect of outliers on the standard
deviation (see, for example, the illustration in Kaufman and Rousseeuw, 2005, pp. 8–9). The
mean absolute deviation for a variable x is computed as:

mad = (1/n) |xi − x̄|,
i

i.e., the average of the absolute deviations between an observation and the mean for
that variable. The estimate for mad takes the place of σ(x) in the denominator of the
standardization expression.
Two additional transformations that are based on the range of the observations, i.e., the
difference between the maximum and minimum. These the RANGE ADJUST and the
RANGE STANDARDIZE options.
RANGE ADJUST divides each value by the range of observations:
xi
ra = .
xmax − xmin
While RANGE ADJUST simply re-scales observations in function of their range, RANGE
STANDARDIZE turns them into a value between zero (for the minimum) and one (for
the maximum):
xi − xmin
rs = .
xmax − xmin

Note that any standardization is limited to one variable at a time, which admittedly is not
very efficient. However, most analyses where variable standardization is recommended, such
as in the multivariate clustering techniques covered in Volume 2, include a transformation
option that can be applied to all the variables in the analysis simultaneously. The same five
options as discussed here are available to each analysis.

2.4.2.4 Bivariate operations

The bivariate functionality includes all classic algebraic operations. The target variable is
selected from the drop-down list on the left in Figure 2.21. Then the respective variables (or
a constant) are entered in the dialog with the appropriate operation selected from the list.
This includes ADD, SUBTRACT, MULTIPLY, DIVIDE and POWER.

2.4.2.5 Date and time functions

The calculator also contains limited functionality to operate on date and time fields. In
practice, this can be quite challenging, due to the proliferation of formats used to define
variables that pertain to dates and time, as encountered in Section 2.4.1.2.
The Date/Time operators implement functions to extract specific parts from the data
format and store them as integer variables. These include Get Year, Get Month, Get
Day, Get Hour, Get Minute and Get Second. The resulting integer variable can then
be used to carry out queries (see Section 2.5.1).
Table Manipulations 29

Figure 2.22: Merge dialog

2.4.3 Merging tables

An important operation on tables is the ability to Merge new variables into an existing
data set, or, in other words, to join an external table to the currently active one.
For example, the Chicago_community_areas.shp file only contains information on
the community area boundaries, but no actual data. On the other hand, the
Chicago_CCA_profiles_2020.csv file contains a wealth of socio-economic data, but no
spatial information other than the ID number for the community area. In order to combine
the spatial information with the non-spatial data, the two tables need to be joined.
With the community area boundaries loaded, the Table > Merge option brings up the
dialog shown in Figure 2.22.
The first step is to load the file to be merged (Select datasource). In Figure 2.22 this is
Chicago_CCA_profiles_2020.csv. The default method is Merge, but Stack is supported
as well. The latter operation is used to add observations to an existing data set.
Rather than relying on the order of the observations in both data sets, it is best practice to
carry out a merging operation by selecting a key. This is a variable that contains (numeric)
values that match the observations in both data sets. Using a key is superior to merging
by record order, since there is often no guarantee that the actual ordering of the data
in the two sets is the same, even though it may seem so in the interface (such as the table
view). In the example, area_num_1 is the key in the first data set (assuming it has been
reformatted to integer) and GEOID the key in the second.
The variables to be merged are selected by moving their names from the Exclude box to
the Include box. Using the » button selects all of them.
Upon clicking the Merge button at the bottom of the dialog, the new variables will be
added to the current data table, unless there is a potential conflict of variable names. This
conflict can be two-fold. First, when variables have identical names in both data sets, they
need to be differentiated. If that is the case, alternatives are suggested in a dialog. These
alternatives are not always the most intuitive and can be readily edited in the entry box.
30 Basic Data Operations

Figure 2.23: Merged tables

One exception to this is when you do not want to merge in the duplicate fields. Checking
the box Use existing field name makes sure that only the original columns are kept.
A second potential conflict is when the variable name to be included contains more than 10
characters. In csv input files, there are no constraints on variable name length, but the dbf
format used for the attribute values in a shape file format has a limit of 10 characters for a
variable name. Again, an alternative will be suggested in the dialog, but it can be easily
edited (as long as it stays within the 10 character limit).8
The merged tables are shown in Figure 2.23. By keeping the identifiers for both data sets in
the merged table (respectively area_num_1 – in the first column – and GEOID – in the
8th column), one can easily verify that the observations are lined up correctly.
As before, the merger only becomes permanent after a File > Save operation.
Note that the merge operation in GeoDa implements a left join, in the sense that only
observations that match the ID in the current table are included in the merged table.
Specifically, if the table to be merged does not contain certain observations, the corresponding
entries in the merged table will be missing values. Also, if the table to be merged contains
observations that are not part of the current table, they will be ignored. Other forms of
joins are not currently implemented.

2.5 Queries
While queries of the data are somewhat distinct from data wrangling per se, drilling down
the data is often an important part of selecting the right subset of variables and observations.
In order to select particular observations or rows in the data table, the Selection Tool is
used. This is arguably one of the most important features of the table options.
The Chicago Carjackings data set is used to illustrate the selection functionality. First, a
few adjustments are needed: the Date column must be reformatted to datetime, and new
variables for the Month and the Day must be included (using Table > Calculator as in
Section 2.4.2.5).
8 The
same problem can occur when using File > Save As to convert a csv file to the dBase format.
Here too, a dialog will suggest alternative variable names, which can be edited.
Queries 31

Figure 2.24: Selection Tool

2.5.1 Selection Tool

Queries can be carried out by invoking Table > Selection Tool or right clicking on the
table to select this option. The interface, shown in Figure 2.24, supports quite complex
searches and operations, even though it may seem somewhat rudimentary.

2.5.1.1 New selection

The main panel of the interface deals with the selection criteria. In Figure 2.24, a New
Selection has been started, as indicated by the radio button in the top line. To select the
carjackings that occurred in the month of November, the Selection Variable must be set
to the newly created variable Month. Also, the beginning and end value of the Select All
in Range option must be set to 11 (for November). Clicking on the button will select the
observations that meet the selection criterion.
The selected observations are immediately highlighted in the table. For clarity, they can
also be moved to the top of the table by means of Table > Move Selected to Top.
Simultaneously, the selected observations are also highlighted in the themeless point map.
More precisely, they retain their original shading, while unselected observations become
transparent.9
This is illustrated in Figure 2.25. The status bar lists that 207 observations have been
selected. In addition, the selected observations are also immediately highlighted in any other
open graph or map. This is the implementation of linking, which will be discussed in more
detail in Chapter 4.

2.5.1.2 Other selection options

The Selection Tool contains several more options to construct a more refined query. For
example, a second criterion can be chosen to Select From Current Selection. This could
be used to select a particular day of the month of November (provided the corresponding
integer variable was created). Alternatively, to combine observations from both November
and December, Append To Current Selection is appropriate.
9 This behavior can be changed in the GeoDa Preferences settings, see Appendix A.
32 Basic Data Operations

Figure 2.25: Selected observations in themeless map

Figure 2.26: Selected observations in new map

A useful feature is to use Invert Selection to choose all observations except the selected
ones. For example, this could be used to choose all months but November.
Often, this approach is the most practical way to remove unwanted observations, since there
is no Save Unselected function. First, the observations to be removed are selected, followed
by inverting the selection. At this point, File > Save Selected As can be employed to
create a new data set (see Section 2.5.3).
Queries 33

Figure 2.27: Car thefts by month

The same approach can also be used in combination with Select All Undefined, which
identifies the observations with missing values for a given variable. The inverted selection
can then be saved as a data set without missing values.
The Add Neighbors To Selection option will be discussed in the chapters dealing with
spatial weights, in Part III.

2.5.2 Indicator variable

Once observations are selected in the table, a new indicator variable can be created that
typically holds the value of 1 for the selected observations, and 0 for the others, as shown in the
bottom panel of Figure 2.24. More generally, any combination of values for presence/absence
can be specified in the dialog.
With a Target Variable specified, clicking on Apply will add the 0–1 values to the table.
This then makes the variable available for use as an indicator or conditioning variable in a
range of statistical analyses, including conditional plots (covered in Part II) and analysis of
variance.

2.5.3 Save selected observations

Arguably one of the most useful features of the selection tool in terms of data wrangling
is the ability to save the selected observations as a new data set. This is accomplished by
means of File > Save Selected As. For example, a new file can be specified to save just
the observations for November, shown in Figure 2.26. The point pattern has the same shape
as the selected observations in Figure 2.25, but the number of observations is now listed as
207.
34 Basic Data Operations

Figure 2.28: Selection on the map

2.5.4 Spatial selection

In addition to selecting observations by means of Table > Selection Tool, queries can
also be constructed visually from any open map.
To illustrate the selection feature, a simple map of the points classified by month is shown in
Figure 2.27.10 The number of observations in each category is listed next to the corresponding
legend item, with (207) as the value for November.

2.5.4.1 Selection by shape

The map selection tool – the pointer icon at the left of the map window toolbar – is the
default interaction with a map view. Observations are selected by clicking on them or by
drawing a selection shape around the target area. The default selection shape is a Rectangle,
as in Figure 2.28, but Circle and Line are available as well. The particular shape is chosen
by selecting Selection Shape from the map options menu (right click on the map).
Identical to table selection, all selected observations are highlighted in the table in yellow
and in any other open map or graph (through linking). Also, they can be saved to a new
data set using File > Save Selected As, in the same way as for a table selection.
The selection can be inverted by clicking on the second left-most icon in the map toolbar.
This works in the same way as for table selection.
10 This is implemented by means of the Map > Unique Values Map command, with Month as the

variable for the unique values. A detailed discussion of mapping functionality is deferred until Chapter 4.
Queries 35

Figure 2.29: Selection on a map category

2.5.4.2 Selection on map classification

In addition to using a selection shape on a map, observations that fall into a particular map
classification category can be selected by clicking on the corresponding legend icon. Map
classifications are discussed in more detail in Chapter 4, but Figure 2.29 illustrates how the
observations for the month November are selected by clicking on the small rectangle next to
11. The selected points match the pattern in Figure 2.25.

2.5.4.3 Save selection indicator variable

Finally, as is the case for a table selection, a new indicator variables can be saved to the
table, with, by default, a value of 1 for the selected observations, and 0 for the others. This
is invoked from the map options (right click on the map) by selecting Save Selection.
The default variable name is SELECTED, but this can be easily changed, as can the values
assigned to selected and unselected. After the indicator variable is added to the table, it can
be made permanent through a File > Save command, in the usual fashion.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
3
GIS Operations

By design, GeoDa is primarily focused on analysis and is not intended to be a geographic

information system (GIS). However, over the years and in response to user requests, a
range of spatial data manipulation functions have been incorporated. These include efficient
treatment of projections, conversion between points and polygons, the computation of a
minimum spanning tree, spatial aggregation and spatial join through multi-layer support.
These functions are not intended to replace the range of operations that can be carried out
in a proper GIS, but rather serve as an alternative for those without access to or familiarity
with a GIS.
In this chapter, I present a quick overview of this functionality, again primarily aimed at
novices in the field. Readers familiar with GIS operations can skim over the material to check
on how these elementary spatial data manipulations are handled in GeoDa. The discussion is
limited to the essentials. More extensive treatments are offered in standard GIS texts, such
as Law and Collins (2018), Brunsdon and Comber (2019), Shellito (2020) and Madry (2021),
among many others.
The various operations will be again be illustrated with the Chicago Carjackings data set
from the Sample Data, as well as with the Community Area boundary layer from the
GeoDaCenter sample data site, Chicago_community_areas.shp.

3.1 Topics Covered

• Understand how projections are expressed in a coordinate reference system (CRS)
• Be able to efficiently change from one projection to another
• Create shape centers (mean center and centroid) for polygons
• Construct Thiessen polygons for a point layer
• Compute a minimum spanning tree based on the max-min distance between points
• Operate on more than one layer
• Dissolve areal units and compute aggregate values
• Compute aggregate values based on a common indicator in a table
• Implement a spatial join to carry out point in polygon operations
• Visualize the common link between two layers
GeoDa Functions
• File > Save As
• Map > Shape Centers
• Map > Thiessen Polygons
• Map > Minimum Spanning Tree
• Map > Add Map Layer

DOI: 10.1201/9781003274919-3 37
38 GIS Operations

• Map > Map Layer Settings

• Tools > Dissolve
• Table > Aggregate
• Tools > Spatial Join
• Map Layer Settings > Set Highlight Association
• Map Layer Settings > Change Associate Line Color

3.2 Projections
Spatial observations need to be georeferenced, i.e., associated with specific geometric objects
for which the location is represented in a two-dimensional Cartesian coordinate system
(i.e., on a flat map). Since all observations originate on the surface of the three-dimensional
near-spherical earth, this requires a transformation from three to two dimensions.
The transformation involves two concepts that are often confused by non-geographers,
i.e., the geodetic datum and the projection. The topic is complex and forms the basis for
the discipline of geodesy. A detailed treatment is beyond the current scope, but a good
understanding of the fundamental concepts is important. The classic reference is Snyder
(1993), and a recent overview of a range of technical issues is offered in Kessler and Battersby
(2019).
The basic building blocks are degrees (and minutes and seconds) latitude and longitude
that situate each location with respect to the equator and the Greenwich Meridian (near
London, England). Longitude is the horizontal dimension (x) and is measured in degrees
East (positive) and West (negative) of Greenwich, ranging from 0 to 180 degrees. Since
the U.S. is west of Greenwich, the longitude for U.S. locations is negative. Latitude is the
vertical dimension (y) and is measured in degrees North (positive) and South (negative) of
the equator, ranging from 0 to 90 degrees. Since the U.S. is in the northern hemisphere,
its latitude values will be positive. Latitude and longitude are typically given as decimal
degrees, but, if not, a conversion from degrees, minutes and seconds is straightforward.
In order to turn a geographic description, such as an address, into latitude-longitude degrees,
it is necessary to adhere to a so-called geodetic datum, a three-dimensional coordinate system
or model that represents the shape of the earth. Currently, the most commonly used datum is
the World Geodetic System of 1984, WGS 84, which represents the earth as an ellipsoid (and
not as a perfect sphere). In North America, an alternative is NAD 83, the North American
Datum of 1983. In practice, for U.S. locations, there is not much difference between the two.
Both these standards are about to be replaced by reference frames that take advantage of
Global Navigation Satellite Systems (GNSS).1
The second step in the process of georeferencing consists of converting the latitude-longitude
coordinates to Cartesian x-y coordinates in a planar system, using a cartographic projection.
Hundreds of projections have been developed, each addressing different aspects of the
mathematical problem of converting a three-dimensional object (on a sphere or ellipsoid)
to two dimensions (a flat map). In this conversion, every projection involves distortion
of one or more fundamental properties of geographic objects: angles (sometimes confused
with shape), area, distance and direction. It is important to be aware of this limitation,
since many investigations rely on the computation of variables such as distance, or density
1 For details, see https://geodesy.noaa.gov/datums/newdatums/index.shtml.
Projections 39

(which involves area). The use of an inappropriate projection or distance metric may yield
misleading results, especially for analyses that cover large areas (e.g., continent-wide).
For our purposes, three aspects are most important. One is to recognize whether spatial
coordinates are projected (typically in units of feet or meters), or unprojected (i.e., in decimal
degrees). To confuse matters more, the latter are sometimes referred to as a geographic
projection, even though there is no projection involved. It is important to keep in mind that
latitude and longitude are not expressed in distance units, but are degrees.
For example, a graph showing locations with longitude as the x-axis and latitude as the
y-axis, treated as if they were regular distance units, can be misleading, even though it
is seen quite commonly in publications. It ignores the fundamental property that latitude
and longitude are degrees (angles). Similarly, the calculation of Euclidean distance is only
supported for projected coordinates and should not be performed on longitude-latitude pairs.
In many instances, GeoDa will generate a warning when an attempt to compute distances
with decimal degrees is made, but this cannot be detected in all situations.
A second aspect is to be aware of the characteristics of a particular projection that is being
used. Specifically, it is important to know whether the projection respects properties such as
area (equal area) or distance (equidistant), although such properties typically only pertain
to a subset of the projected map.2
A final important aspect is to make sure that layers are in the same projection when combined.
GeoDa has some functionality to reproject layers on the fly to accomplish this, but it is not
fail-safe.

3.2.1 Coordinate reference system

A coordinate reference system or CRS is a formal representation of location. It typically
contains both a datum and a particular projection, as well as some information on the
location of the coordinate center, units of measurement and related items. A CRS is identified
by a code, such as an EPSG code, referred to as such because it was originally developed by
the European Petroleum Survey Group (EPSG).
Interestingly, the lack of a projection, i.e., coordinates as simple latitude-longitude decimal
degrees, has an EPSG code of 4326. As mentioned, this is sometimes referred to as a
geographic projection, even though strictly speaking no projection is involved. Nevertheless,
identifying latitude and longitude coordinates with a CRS of EPSG 4326 will provide a
valid point of departure, Without it, there would be no information on what the coordinates
represent. This is critical for use in later transformations (see Section 3.2.3).
GeoDa uses the proj4 convention to specify the CRS. This is compatible with the open-source
PROJ library upon which the projection functionality is built.3 Specifically, in order to save
a spatial layer that is expressed as latitude-longitude decimal degrees, the proper entry in
the CRS box is:
+proj=longlat +ellps=WGS84 +datum=WGS84 +no_defs

3.2.2 Selecting a projection

Non-geographers are often at a loss when faced with specifying an appropriate projection.
An excellent resource in this respect is the web site spatialreference.org, which contains
2 For a detailed technical discussion, see, for example Kessler and Battersby (2019).
3 https://proj.org
40 GIS Operations

Figure 3.1: UTM zones for North America (source: GISGeography)

literally hundreds of projection definitions that can be easily searched.4 For each projection,
there is a summary of its properties and a list of its CRS specification in a range of different
formats. For example, this includes the format used by ESRI in the *.prj files (part of the
shape file specification), as well as the proj4 specification used by GeoDa.
While it is perfectly fine to use locations expressed as latitude-longitude decimal degrees, it is
imperative to use the proper mathematical operations when computing properties like area
and distance. For the latter, it is necessary to use great circle distance (arc distance), which
is expressed in terms of the angles represented by latitude and longitude (see Chapter 11).
In addition, the proper conversion of the great circle distance from angular units to distance
unit (e.g., miles or kilometers) needs to differentiate by degree latitude. The implementation
in GeoDa is only approximate and uses the distances at the equator.
In practice, it is preferred to use projected coordinates so that Euclidean distance operations
and area calculations are straightforward to perform. For North America, a useful projection
is universal transverse Mercator, or UTM. The country is divided into parallel zones, as
shown in Figure 3.1. With each zone corresponds a specific projection that is represented by
a CRS (e.g., an ESPG code).5
Let’s say we need a projection for locations in Chicago (indicated by the arrow on the figure).
From the map, we can see that this city is located in UTM zone 16 (north – there is a
southern hemisphere equivalent). Searching the spatialreference.org site for this projection
yields a list of specifications (i.e., combinations of different datums). For example, for a
WGS84 datum we can find an associated EPSG code of 32616. In the proj4 format, the
corresponding CRS is:
+proj=utm +zone=16 +ellps=WGS84 +datum=WGS84 +units=m +no_defs

3.2.2.1 Specifying the CRS

As illustrated in the previous chapter, the File > Save As dialog contains a entry box at
the bottom to specify the CRS for a spatial layer (see Figures 2.12 and 2.13). This will yield
4 https://spatialreference.org
5 For more details on the fundamentals behind the UTM projection, see, for example, the GISGeography

site at https://gisgeography.com/utm-universal-transverse-mercator-projection/.
Converting Between Points and Polygons 41

Figure 3.2: Load CRS from another layer

Figure 3.3: New CRS from another layer

a new layer with the proper projection. An alternative approach that avoids the need to
type in the actual proj4 specification is considered next.

3.2.3 Reprojection
To avoid the manual entry of the correct CRS code, the Save As process in GeoDa provides
an alternative that does not need an explicit specification, but requires the presence of
another layer with the desired projection. The CRS information is then copied from that
other layer and used in the reprojection of a current layer.
For example, consider the 77 Chicago community areas expressed in latitude-longitude
coordinates in Figure 2.4. We saw how the locations of the carjackings in Figure 2.7
suggested a somewhat different shape for the city, because it was expressed in the State
Plane Illinois East NAD 1983 projection (EPSG 3435). To convert the polygon community
areas to the same projection, we could enter the proper proj4 specification in the CRS box.
An alternative is to copy the CRS information from a different layer. For example, after
opening the community area file (Chicago_community_areas.shp) and invoking File >
Save As, the CRS box shows the proj4 specification for latitude-longitude degrees, as in
Figure 3.2. Instead of typing in the new specification, the small globe icon to the right can
be selected to load a CRS specification from another file.
This brings up a file load interface into which the file name for the point shape file with the
carjacking locations can be specified (Chicago_carjack_2020_pts.shp). The easiest way to
accomplish this is by dragging the file name into the Drop file here area. Once the file
name is loaded, the contents of the CRS box change to the new proj4 specification, as in
Figure 3.3.
After saving the new file (e.g., as community_areas_proj.shp), the current project should be
closed. A new project is started by loading the just created projected layer. The corresponding
themeless base map is as in Figure 3.4. The more compressed shape matches the layout of
the car jacking locations in Figure 2.7.

3.3 Converting Between Points and Polygons

So far, the geography of the community areas has been represented by their actual boundaries.
However, this is not the only possibility. Equivalently, a representative point for each area
can be specified, such as a mean center or a centroid. In addition, new polygons can be
42 GIS Operations

Figure 3.4: Reprojected Chicago community areas layer

constructed to represent the community areas as tessellations around those central points,
such as Thiessen polygons. The key factor is that all three representations are connected to
the same cross-sectional data set. As discussed in more detail in later chapters, for some
types of analyses it is advantageous to treat the areal units as points, whereas in other
situations Thiessen polygons form a useful alternative to dealing with an actual point layer.
The center point and Thiessen polygon functionality is invoked through the options menu
associated with the map view (right click on the map to bring up the menu). This is
illustrated with the just created projected community area map (Figure 3.4).

3.3.1 Mean centers and centroids

Two concepts are in common use to represent the center of a polygon. One is the mean center,

which is simply
 the (unweighted) average of the coordinates of the vertices, so (1/k) i xi
and (1/k) i yi (where k is the number of vertices). The notion of a centroid is more complex,
as it represents the center of gravity of the polygon. A classic exercise is to create a cutout
of the polygon and to find the centroid as the location where a pin would hold it up in a
stable equilibrium.
The main difference between the mean center and the centroid is that the latter takes into
account the area of the polygon. It is computed as an area weighted average of the centers
of the set of triangles that represent the polygon. This so-called triangulation of a polygon is
a standard operation in computational geometry (see, for example, Worboys and Duckham,
2004; de Berg et al., 2008). Once the triangulation is obtained, the centers of the triangles
(the average of their coordinates) and their areas can be readily computed to obtain the
centroid.6
The calculation of the mean center and centroid can yield strange results when the polygons
are highly irregular (not convex), have holes, or represent a so-called multi-polygon situation
6 GeoDa uses the centroid implementation from the GEOS library, see https://libgeos.org/doxygen/classg

eos_1_1algorithm_1_1Centroid.html.
Converting Between Points and Polygons 43

Figure 3.5: Display mean centers on map

Figure 3.6: Add mean centers to table

(different polygons associated with the same ID, such as a mainland area and an island
belonging to the same county). In those instances, the centers can end up being located
outside the polygon. Nevertheless, the shape centers are a handy way to convert a polygon
layer to a corresponding point layer with the same underlying geography. For example, as
covered in Chapter 11, they are used under the hood to calculate distance-based spatial
weights for polygons.

3.3.1.1 Shape Center options

In GeoDa, the Shape Center option from the map view has three main functions, applied
to either the Mean Centers or the Centroids. These are: Add to Table, Display and
Save.
The simplest is to Display the center points on the map. This is purely illustrative and
the points are not available for any further manipulation, in contrast with the multi-layer
option discussed in Section 3.6. For example, in Figure 3.5, the mean centers of the Chicago
community areas are displayed as small circles on top of the default themeless map from
Figure 3.4.
44 GIS Operations

Figure 3.7: Chicago community area mean center layer

A second option is to Add the coordinates of the central points to Table. This brings up
a dialog that prompts for the variable names for the coordinates, with COORD_X and
COORD_Y as defaults. For example, the mean center coordinates could be set to CO-
ORD_XM and COORD_YM and the centroids to COORD_XC and COORD_YC.
In Figure 3.6, the associated coordinate values are shown in the table (after some reshuffling
of columns). Close examination of the values reveals only minor differences between the two
types of mean centers in this particular case, in part because the community areas have
mostly fairly convex shapes (the coordinates are in meters).
Finally, the Save option results in the usual prompt for a file name (e.g.,
comm_area_centers.shp) and creates the new layer. In the example, the point layer associated
with the community area mean centers is as in Figure 3.7.

3.3.2 Tessellations
A tessellation is the exhaustive covering of an area with polygons or tiles. Of particular
interest in spatial analysis are so-called Thiessen polygons or Voronoi diagrams (for an
extensive discussion, see Okabe et al., 2000).
Thiessen polygons are a regular tiling of polygons centered around points, such that they
define an area that is closer to the central point than to any other point. In other words, the
polygon contains all possible locations that are nearest neighbors to the central point, rather
than to any other point in the data set.7 The polygons are constructed by considering a
line perpendicular at the midpoint to a line connecting the central point to its neighbors.
The latter is referred to as Delaunay triangulation and is in fact the dual problem to the
tessellation.
In sum, any point layer can be converted to a polygon layer with the same geography by
constructing Thiessen polygons around the points.
7 In economic geography, this corresponds to the notion of a market area, assuming a uniform distribution

of demand.
Minimum Spanning Tree 45

Figure 3.8: Display Thiessen polygons

3.3.2.1 Thiessen Polygons options

The Thiessen Polygons functionality is available as an option in the map view for point
layers. As for the mean centers, there are three options, although only two of those are used
regularly.
The Display option simply draws the polygons around the points, as shown in Figure 3.8
for the mean center points from Figure 3.7. As in the case of the display of mean centers,
the resulting polygons cannot be used for any other operations.
The Save option will create a new layer with the polygons.8 This layer can then be used in
the same way as any other polygon layer, e.g., to create maps or carry out spatial analyses.

3.4 Minimum Spanning Tree

3.4.1 Concept
A central concept in the treatment of spatial autocorrelation is the so-called spatial weights
matrix, discussed in detail in Part III of this Volume. For now, it is sufficient to know that
this matrix shows the neighbor relations between observations. An element of the matrix in
row i and column j, wij , is non-zero if i and j are neighbors (using one of many possible
definitions). Technically, this is the same as an adjacency matrix for a graph or network,
where the observations are nodes and the non-zero weights correspond to the edges that
8 A third option is to Save Duplicate Thiessen Polygons to Table. This is only available when there

is a problem in the calculation of the polygons due to multiple observations with the same coordinates. This
may happen when considering different housing units in a high rise, or when the precision of the coordinates
is insufficient to distinguish between all points. The option adds an additional column to the table, labeled
DUP_IDS, which contains a value of 1 for those points that have multiple observations with the same
coordinates.
46 GIS Operations

Figure 3.9: Connectivity graph of mean centers

connect the nodes. In the treatment of spatial weights, the graph structure will be called
the connectivity graph.
For example, in Figure 3.9, the connectivity graph is shown for the community area mean
centers from Figure 3.7. This graph represents a distance-band weights matrix with a cut-off
distance that corresponds to the largest nearest neighbor distance between points, the
so-called max-min distance criterion. This ensures that each observation (node) has at least
one neighbor.
A Minimum Spanning Tree (MST) is a subset of this graph that constitutes a tree, i.e., “a
connected, undirected network that contains no loops” (Newman, 2018). Importantly, this
means that all nodes are connected (no isolates or islands) and there is exactly one path
between any pair of nodes (no loops). As a result, a tree associated with n nodes contains
exactly n − 1 edges. The MST is a special tree in that the total sum of the edge lengths
(typically a weight) is minimized. It constitutes a simplification of the full network structure
that fulfills a minimum cost objective.
For example, the MST computed from the network structure in Figure 3.9 is shown in Figure
3.10. The complexity of the graph is greatly reduced and now consists of only 76 edges (for
77 observations or nodes).
The MST is similar to, but distinct from the solution to the so-called traveling salesperson
problem, which looks for a route through the network that achieves a minimum cost objective,
but visits each node just once. This implies that the nodes in a solution to the traveling
salesperson problem cannot have a degree (i.e., the number of edges connected to it) larger
than two. In Figure 3.10, the maximum degree is four.
A classic solution to the MST problem is achieved by Prim’s algorithm (Prim, 1957),
illustrated in the following worked example. The MST plays a major role in some of the
spatial clustering procedures discussed in Volume 2.
Minimum Spanning Tree 47

Figure 3.10: Mininum spanning tree for connectivity graph of mean centers

Figure 3.11: Toy example point coordinates

3.4.1.1 Worked example – Prim’s algorithm for MST

A small toy example consisting of 9 points is used to illustrate the mechanics of Prim’s
MST algorithm. The coordinates of the points are listed in Figure 3.11. The corresponding
inter-point distance matrix is given in Figure 3.12. The inter-point distances form the basis to
construct a connectivity graph for a distance band that ensures that each point is connected
to at least one other point. In the example, this distance is 29.15.9 The corresponding graph
representation is given in Figure 3.13, where each point is a node and the edges show the
connections and associated distances.
Prim’s algorithm proceeds by starting at a random node and selecting the shortest edge
to the next node. For example, starting with node 1, there is only a single edge, such that
the first path is from node 1 to 2. At node 2, there are two edges, 2–4 with length 15.81,
and 2–5 with length 18.03. Therefore, the next link is between 2 and 4. At this point, the
shortest path follows the red lines as in Figure 3.14.
9 Connectivity graphs are discussed in more detail in the chapters dealing with spatial weights in Part III.
48 GIS Operations

Figure 3.12: Inter-point distance matrix

Figure 3.13: Connectivity graph

Figure 3.14: Initial steps in Prim’s algorithm

At node 4, the two edges have the same length, so 4 could be connected to either 6 or 5.
The order is immaterial, since the length of 15 is the smallest in the network and will have
priority over any other edge. At this point, the path consists of 1–2, 2–4, 4–6 and 4–5. The
next step connects 5 to 7 (18.03), then 7 to 8 (18.03) and 8 to 9 (15). The only remaining
unconnected node is 3, which is linked to 5 (29.15). The resulting MST follows the red lines
in Figure 3.15. Every node is visited and the total distance of the path is minimized. The
solution consists of eight edges, which is n − 1. The MST solution is not unique since there
can be several trees that achieve the same minimum cost solution.
Aggregation 49

Figure 3.15: Minimum spanning tree

3.4.2 Minimum Spanning Tree options

Minimum Spanning Tree is one of the entries in the options menu associated with a
map view for a point layer (right click on the layer to invoke the menu). There are two
main options: Display Minimum Spanning Tree, which simply shows the edges on the
point map, and Save Minimum Spanning Tree. The latter saves the MST information
in the gwt spatial weights format (see Chapter 11), which makes it available for use in other
analyses.
In addition, there are two options to (slightly) change the format of the MST display: Change
Edge Thickness (with three options: Light, Normal – the default – and Strong), and
Change Edge Color. The display in Figure 3.10 has the Edge Thickness as Strong and
the color changed to blue (from the default black).

3.5 Aggregation
Spatial data sets often contain identifiers of larger encompassing units, such as the state that
a county belongs to, or the census tract that contains individual household data. Such nested
spatial scales readily lend themselves to aggregation to the larger unit. The aggregation
computes observations for the new spatial scale as the sum, average or other operation
applied to the values at the lower scale. Note that this only makes sense if the underlying
variables are spatially extensive, such as counts (total population, total households). Without
proper re-weighting, the aggregation gives misleading results for variables expressed as
medians or percentages (e.g., median household value, median income).
GeoDa supports two different ways to compute aggregates from the smaller units. One
approach also creates a dissolved layer, i.e., a spatial layer that includes both a map for the
larger spatial units and a data table with the corresponding aggregated observations. The
second approach is limited to computing a new table with the aggregate information. It does
not generate a spatial component.
50 GIS Operations

Figure 3.16: Dissolve dialog

3.5.1 Dissolve
The Dissolve functionality is part of the Tools menu (Tools > Dissolve), or can be
accessed by selecting the Dissolve option from the Tools icon on the toolbar, the right-most
icon in Figure 2.1.
To illustrate this feature, the point of departure is the merged data set created in Section
2.4.3 (e.g., saved as Chicago_2020.shp). This data set contains socio-economic data for
the 77 Chicago community areas. In addition, it also has an identifier that associates each
community area with a larger encompassing district, i.e., the variable districtno. The
corresponding layout is given by the themeless base map from Figure 2.4
The dialog, shown in 3.16, requests three important pieces of information. The most critical
is the variable for dissolving, i.e., the key that indicates the larger scale to which the
observations will be aggregated (here, districtno). Next follows the selection of the variables
for which the aggregate values will be calculated. In the example in Figure 3.16, these
are 2000_POP, 2010_POP and TOT_POP (the population in 2020 from ACS). The
variables are selected by double clicking on them, or by selecting them and then using the
> key to move them to the right-hand side column.
Finally, we need the proper aggregation Method must be selected. The available options
are Count, Average, Max, Min and Sum. The Count function doesn’t actually compute
any aggregate values, but provides a count of the number of smaller units included in the
larger unit (e.g., the number of community areas in each district). For example, Sum will
yield the population totals for each district.
After activating the Dissolve button, the usual file creation dialog is produced to specify a
file type and file name, e.g., Chicago_district_2020.shp.
Figure 3.17 shows the outline of the nine districts (in red), together with the original
community area boundaries (in blue). This is an illustration of a multiple layer operation,
since both the community area layer and the district layer are combined on the same map.
The multilayer functionality is further detailed in Section 3.6.1. The associated table (Figure
3.18) lists the population totals in the three census years for the district aggregates. The table
also includes a new variable, AGG_COUNT, which indicates the number of community
areas aggregated into the respective district.
Aggregation 51

Figure 3.17: Dissolved districts map

Figure 3.18: Dissolved districts table

3.5.2 Aggregation in table

The same aggregation function is also available as an option for the table, invoked as Table
> Aggregate from the menu, or, alternatively, from the list of table options, shown in
Figure 2.16. The main difference with the Dissolve function is that the results are only
available as a table, without an associated dissolved map.
The dialog requests the same three features and operates in the same way as shown in Figure
3.16.10 The result is saved as a table, which contains the same information as shown in
Figure 3.18.
10 The order of the items is slightly different, and the aggregation variable is referred to as Select key,

but otherwise the operation is identical to that in Figure 3.16.

52 GIS Operations

Figure 3.19: Map layer settings

3.6 Multi-Layer Support

So far, the focus has been on a single spatial layer, associated with one cross-sectional data
set. Over time, GeoDa has added some limited functionality to handle multiple layers, to
facilitate operations such as a spatial join between two layers (see Section 3.7).
While multiple layers can be displayed in the same map window, interaction is limited to
a single layer, the so-called current layer. This pertains to table manipulations, such as
selection and variable transformation but also to the full range of mapping and analytical
functionality. The current layer is always the layer that was loaded first, irrespective of the
actual order of the layers displayed in the map window.
Before delving deeper into the interaction between layers, the basics of the multiple layer
logic in GeoDa are outlined.

3.6.1 Loading multiple layers

The starting point in any multi-layer operation is the current layer, i.e., the data set first
loaded. For example, to display the carjacking locations together with the boundaries of the
Chicago community areas, one of these layers must be the starting point, such as the points
layer (shown Figure 2.7). An additional layer is loaded by means of the Add Map Layer
option. This is represented by the large + icon, third from left in the map window toolbar.
This brings up the usual data input interface, to which the projected community area
boundaries are loaded (e.g., community_areas_proj.shp, shown in Figure 3.4). The result is
displayed in Figure 3.19, with the point locations portrayed on top of the community area
outlines. The Map Layer Settings option (fourth icon from the left on the map window
toolbar) brings up a small dialog that lists the order of the layers and indicates the current
map.
Multi-Layer Support 53

Figure 3.20: Polygon layer on top with opacity=0

It is important to keep in mind that the visual properties (such as color and opacity) of the
current layer are managed by the map legend (see Chapter 4). The properties of the other
layers are controlled through the Map Layer Settings.

3.6.1.1 Map layer setting options

A right click on the small legend rectangle associated with each layer (other than the current
layer) brings up a list of options that pertain to the visual representation of that layer. The
two most often used functions are Change Fill Color and Change Outline Color. In
addition to setting the color, it may be necessary to adjust the opacity in order to make
sure that the relevant information is visible. Specifically, setting opacity to 0 makes the
areas invisible and only the boundary outlines are shown. The latter can be turned off by
unchecking the Outline Visible option. The Only Map Boundary option results in a
dissolve, in that only the outer outline of the polygon layer is shown. This operation is purely
visual and does not involve any aggregation, in contrast to what was the case in Section
3.5.1.
The interface also includes options to set the association highlight, which is covered in
Section 3.8. The final item in the dialog is to Remove the layer.11

3.6.1.2 Layer position

The position of the layers in the map view can be altered by means of the Map Layer
Settings interface. One grabs the entry for a layer and drags it to a different position. For
example, in Figure 3.20, the community area layer is moved to the top (this requires setting
the opacity to 0 in order for the points to show). However, even though the polygon layer is
now on top, the point layer remains the current map layer, since it was loaded first.
11 In addition, it is possible to change the map window scope by using Zoom to Selected and Layer

Full Extent.
54 GIS Operations

Figure 3.21: Points reprojected to lat-lon

3.6.2 Automatic reprojection

In the previous example in Figure 3.19, the two layers that combined were expressed in the
same projection. More precisely, the projected community area boundaries were superimposed
on the point layer.
However, as long as layers have projection information associated with them, any added
layer will be reprojected to the projection of the current layer. This is for visualization only,
and it does not change the projection information in the files of the reprojected layers.
For example, in Figure 3.21, the order in which the layers are loaded is reversed from Figure
3.19. The community area layer (Chicago_community_areas.shp) is loaded first (i.e., it is
the current layer). This layer is not projected, but expressed in lat-lon coordinates. To this
the point layer Chicago_carjack_2020_pts.shp is added, which is projected.
Initially, the point layer is invisible, since it is positioned under the polygon layer. The
default green coloring of the community areas masks the points underneath. After setting
the opacity of the top layer to zero, the points become visible.
In Figure 3.21, a slightly more spread out pattern can be noticed relative to Figure 3.19.
Under the hood, the point layer has been reprojected to lat-lon decimal degrees. This is only
for visualization in the map view and does not change the CRS associated with the added
layers.

3.6.3 Selection in multiple layers

It is important to keep in mind that any selection only pertains to the current map layer,
which may or may not also be the top layer. For example, continuing with the previous
illustration (Figure 3.21), any selection will pertain to the community area layer, which is
the current map layer. This is irrespective of which layer is shown as the top layer.
In Figure 3.22, the selection status bar indicates that 14 observations were selected. If the
point layer was moved to the top, this would give the same result.
Spatial Join 55

Figure 3.22: Select on current map layer (community areas)

This logic may seem a bit counter-intuitive at first, but it is based on the linking and
brushing architecture implemented in GeoDa (see Chapters 4 and 7 for a more extensive
discussion). Currently, the linking operation works only for a single layer, which is always
the first loaded layer.

3.7 Spatial Join

The multi-layer infrastructure allows for the calculation of variables for one layer, based on
the observations in a different layer. This is an example of a point in polygon GIS operation.
There are two applications of this process. In one, referred to as Spatial Assign, the ID
variable (or any other uniquely identifying variable) of a spatial area is assigned to each
point that is within the area’s boundary. The reverse of this process is Spatial Count, i.e.,
a count or other form of aggregation over the number of points that are within a given area,
similar to the operation of Aggregation and Dissolve (see Section 3.5).
Even though the default application is simply assigning an ID or counting the number of
points, more complex assignments and aggregations are possible as well. For example, rather
than just counting the points, an aggregate over the points can be computed for any variable,
such as the mean, median, standard deviation, or sum across the point observations for that
variable.
It is important to keep in mind the order in which the layers are loaded into the multi-layer
setup. For spatial assign, the point layer is first, since the areal unit identifier is assigned
to each point. As mentioned before, only the table of the first loaded layer is available for
manipulation or analysis. So, in order to add the ID code to the point layer table, it must
be loaded first.
For spatial count, the reverse is the case. The polygon layer needs to be loaded first, since
the number of points (or any other aggregation) is added to the table with observations for
the areal units.
56 GIS Operations

Figure 3.23: Spatial join dialog – areal unit identifier for point

Figure 3.24: Community area identifier (SA) in point table

This functionality is invoked from the menu as Tools > Spatial Join, or by selecting the
Spatial Join option from the Tools icon in Figure 2.1.

3.7.1 Spatial assign

To illustrate the spatial assign application, the layer arrangement from Figure 3.19 is
considered, with the (projected) car jacking locations loaded first, followed by the (projected)
community area boundaries.
Invoking Tools > Spatial Join generates the dialog shown in Figure 3.23. This differs
slightly between the spatial assign and spatial count functionalities (see Figure 3.26 for the
latter).
The dialog shows how the current map layer, Chicago_carjack_2020_pts is joined to
another layer. In the example, there is only one other layer, i.e., community_areas_proj,
but in general, there could be several layers to select from. The ID Variable list shows all
the variables that could serve as unique identifiers for the areal units. In the example, there
are three variables listed, but only the fist two pertain to the community areas in the map
layer (all integer variables are listed by default). In Figure 3.23, area_num_1 is selected
as the ID variable.
The next dialog requests the variable name of the spatial assign variable, which is SA by
default. After specifying the variable name, it is added to the data table for the point layer.
In Figure 3.24, the new SA column is listed next to the original Commu_Area variable
from the point location data base. A quick check confirms that they are the same. The
difference is that the values for the SA variable were derived from a spatial point in polygon
operation, joining the two layers, whereas the original variable was included during data
entry.
In cases where the point in polygon operation results in a mismatch (e.g., due to differences
in precision of the location coordinates in the two layers), a value of −1 is used for the
Spatial Join 57

Figure 3.25: Point in polygon mismatch

Figure 3.26: Spatial join dialog – aggregation options

spatial assign. In the example, there are three such points, all situated at the very edge of
a polygon. This is illustrated in Figure 3.25 for the two mismatches at the southern and
south-western edge of the city (one other mismatch is in the very north).
Depending on the goals of the analysis, one could either eliminate the points from the data
set, or manually edit the values in the table after zooming in on the actual location.
The new spatial assign variable can now be used to aggregate observations, as in Section
3.5. However, the dissolve operation does not work, since there is no boundary information
associated with the points. Their spatial information consists of the coordinates only, which
do not lend themselves to a dissolve operation.
As always, a File > Save or File > Save As operation must be used to make the addition
of the spatial assign variable permanent.

3.7.2 Spatial count

To illustrate the spatial count feature, the order of the two layers from Figure 3.19 is reversed.
The community_areas_proj layer is loaded first (since the counts will be added to the
58 GIS Operations

Figure 3.27: Spatial count in polygon area table

polygon table observations), followed by Chicago_carjack_2020_pts. Note that in order to

make the point layer visible, it needs to be moved to the top.12
The Tools > Spatial Join command generates the dialog shown in Figure 3.26. In the
default setup, the Spatial Count radio button is checked, which does not require any
further input. In the figure, the Spatial Join option is shown, which is essentially the same
as aggregation (Section 3.5). Several variables can be selected, with the aggregation following
Sum, Mean, Median or Standard Deviation.
In the illustration, the default is used, i.e., Spatial Count. The next dialog asks to specify
the variable name for the result, with the default as SC. This yields a new variable in
the table, which shows the number of car jackings occurring in each community area, as
illustrated in Figure 3.27. As before, the addition only becomes permanent after a File >
Save or File > Save As operation.

3.8 Linked Multi-Layers

In the previous example, the connection between two layers is used to compute a new
variable (e.g., the number of points in a polygon). An additional feature of the multi-layer
architecture is that a connection can be established between observations in two layers. A
typical application of this feature is to show the connection between two point layers, e.g.,
where one contains the location of headquarters and the second of branches, connected by
a shared corporate ID. The latter is critical, since the observations in the two layers must
share a common key.
The linkage is established through the Highlight Association options in the Map Layer
Settings dialog. There are three aspects of this: Set Highlight Association, Clear
Highlight Association and Change Associate Line Color.
Once the linkage is established, it can be highlighted for any selection in the current map
layer. Each observation in the current map is connected to the matching observation in the
other layer, either a point location or the centroid of a polygon. It is important to remember
that any selection pertains to the current map layer, irrespective of which layer is shown on
top in the map view.
12 Alternatively, the opacity of the polygon layer can be reduced to zero to make the points visible.
Linked Multi-Layers 59

Figure 3.28: Set association dialog

Figure 3.29: Linked selection

3.8.1 Specifying an inter-layer linkage

The layer arrangement from Figure 3.19 has the (projected) car jacking locations loaded
first (current map), followed by the (projected) community area boundaries. Observations in
both layers share the community area identifier, Commu_Area in the point layer, and
area_num_1 in the polygon layer.
Selecting Set Highlight Association from the options generates the dialog shown in
Figure 3.28. This interface brings all the elements together to connect observations in the
two layers through a series of drop-down lists.
The point of departure is the current layer, shown at the top of the dialog
(Chicago_carjack_2020_pts). The first drop-down list, Select layer is to select the
layer to link with. In the example, there is only one option, community_areas_proj. Next
is the key for that layer (field), area_num_1, and the corresponding key in the current
layer (associated), Commu_Area. Finally there is a check box to activate connection
lines between the two layers.

3.8.2 Visualizing linked selections

The choices in the dialog of Figure 3.28 establish the connection between the two layers, but
this is only activated through an actual selection in the current map (see Section 2.5.4). For
example, in Figure 3.29, point locations have been selected that belong to three different
community areas. A line connects each selected point with the centroid of the linked area.
The default color is red, but this can be readily changed through the Change Associate
Line Color option.
The operation can also be reversed, i.e., by making the polygon layer the current map from
which observations can be selected, showing the linked points.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
 Part II

EDA and ESDA

Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
4
Geovisualization

In this first of three chapters dealing with mapping, I begin to explore the concept of
geovisualization. Before getting into specific methods, I start with a brief discussion of the
larger context of exploratory data analysis (EDA) and especially its spatial counterpart,
exploratory spatial data analysis (ESDA). Central to this and to its implementation in GeoDa
are the concepts of linking and brushing.
The technical discussion begins with an overview of key aspects of thematic map construction,
followed by a review of traditional map classifications, i.e., quantile maps, equal interval
maps and natural breaks maps. More statistically inspired maps are discussed in Chapter 5.
Conditional maps are considered in the treatment of conditional plots in Chapter 8. The
particular problem of mapping rates or proportions is covered in Chapter 6.
The common map types are followed by a discussion of various mapping options in GeoDa
that allow interaction with the map window and the creation of output for use in other media.
The chapter closes with an introduction of the implementation of custom classifications and
the use of the project file.
Even though there is substantial mapping functionality in GeoDa, it is worth noting that it
is not a cartographic software. The main objective is to use mapping as part of an overall
framework that interacts with the data in the process of exploration, through so-called
dynamic graphics (elaborated upon in Section 4.2). By design, maps in GeoDa do not have
some standard cartographic features, such as a directional arrow, or a scale bar, since they
are intended to be part of an interactive framework. However, any map can be saved as an
image file for further manipulation in specialized graphics software.
As in the two previous chapters, the discussion here is aimed at novices, who are less familiar
with basic cartographic principles. Others may just want to skim the material to see how the
functionality is implemented in GeoDa. The treatment focuses on gaining familiarity with
essential concepts, sufficient to be able to carry out the various operations in GeoDa. More
technical details can be found in classic cartography texts, such as Brewer (2016), Kraak
and Ormeling (2020) and Slocum et al. (2023), among others. Also highly recommended for
those not familiar with mapping is Monmonier’s How to lie with maps (Monmonier, 2018),
which provides an easy to read overview of critical aspects of the visualization of spatial
data.
To illustrate the various techniques, a new data set is used. It contains information on Zika
infection and Microcephaly in municipios of the state of Ceará in northeastern Brazil. This
constitutes a subset of the data for the whole of Brazil reported on in Amaral et al. (2019).
In addition to the incidence of the two diseases in 2016, the data set also contains several
socio-economic indicators from the Brazilian Index of Urban Structure (IBEU) for 2013 (see
Amaral et al., 2019, for a detailed definition of each variable). Ceará Zika is included as a
GeoDa sample data set.

DOI: 10.1201/9781003274919-4 63
64 Geovisualization

4.1 Topics Covered

• Situate the role of geovisualization within a larger exploratory context
• Appreciate the importance of linking and brushing in exploratory spatial data analysis
• Create commonly used thematic maps
• Change legend colors
• Manipulate the map by zooming, panning and selection
• Add a background layer (base map) to a thematic map
• Save map classifications as a categorical variable
• Save the map as an image
• Create a custom classification
• Use the project file in GeoDa
GeoDa Functions
• Map > Quantile Map
– select number of categories
• Map > Natural Breaks Map
– select number of categories
• Map > Equal Intervals Map
– select number of categories
• Map toolbar options
• Map selection
– select all observations in a legend category
• Map options
– Change current map type
– Save Categories
– Save the map as an image
– Change the look of a map
– Make map outlines invisible
• Map > Create New Custom
– Category Editor
• File > Save Project
– saving a custom classification in the project file
Toolbar Icons

Figure 4.1: Maps and Rates | Cartogram | Map Movie | Category Editor

4.2 From EDA to ESDA

The mapping functionality outlined in this and the next chapters is part of an overall
framework to guide and facilitate learning from spatial data, which also includes a range of
traditional statistical graphs (see Chapters 7 and 8). GeoDa implements general ideas from
From EDA to ESDA 65

exploratory data analysis (EDA) and its spatial extension, exploratory spatial data analysis
(ESDA). To put these various concepts into context, next, I briefly discuss the evolution of
EDA and visual analytics as well as how geovisualization and ESDA fit into this framework.
I close with a brief introduction of the important concepts of linking and brushing.

4.2.1 Exploratory data analysis

EDA is generally viewed to have originated in the 1960s as a reaction by computationally
oriented statisticians to the primary focus on mathematics and modeling in their discipline.
It was first comprehensively outlined in the classic book by John Tukey (Tukey, 1977). His
presentation stressed the value of investigating the raw data by means of a range of graphic
devices, several of which were developed by Tukey himself. Well-known examples include
the box plot and stem and leaf plot.
The objective of EDA is to create effective tools to guide the analyst in the discovery of
information in the data, specifically, “indications of unexpected phenomena” or to “display
the unanticipated” (Tukey, 1962; Tukey and Wilk, 1966), or even to “discover potentially
explicable patterns” (Good, 1983). In this context, EDA is often contrasted with confirmatory
data analysis, or CDA. This reflects the traditional distinction in knowledge discovery
between an inductive approach (data first, hypothesis later) and a deductive approach
(hypothesis/model first, data later). However, with the emphasis on visual exploration (see
also Tufte, 1997), EDA is really about an abductive approach, where the interaction between
data exploration and human perception leads to the detection of patterns jointly with the
formulation of hypotheses (e.g., Gahegan, 2009).
Early approaches to visualizing data go back to Greek times, although major innovations
did not occur until the work of William Playfair during the late 18th and early 19th century
(Friendly, 2008). However, practical visual exploration of large data sets had to wait for the
development of powerful computer graphics hardware. This allowed direct interaction with
the data shown on a computer screen through so-called dynamic graphics (Becker et al., 1987;
Cleveland and McGill, 1988; Cleveland, 1993). Dynamic graphics allowed the data to be
represented simultaneously by means of different views (Buja et al., 1996), i.e., graphs, tables,
charts and even maps that focus on various aspects of the data distribution. Especially when
dealing with high-dimensional data, insight is gained through careful manipulation of the
views, such as linking, focusing and arranging (for a recent review of the range of techniques,
see Chen et al., 2008).
In the computer science literature, a similar and largely parallel development occurred in the
form of visual analytics, an approach to knowledge discovery based on the use of statistical
graphics and other visual tools to facilitate pattern recognition and data mining (Thomas
and Cook, 2005; Kielman et al., 2009). As in EDA, in visual analytics much attention is paid
to human-computer interaction so as to facilitate analytical reasoning in order to “detect
the expected and discover the unexpected” (Kielman et al., 2009, p.245).
Whereas early on EDA and CDA were mostly seen as opposing strategies to gain knowledge,
more recently attempts are being made to bring the two closer together. Specifically, a
concern that unstructured exploration may lead to spurious results has yielded methods to
augment EDA in order to provide some measure as to how unusual the findings may be.
For example, Buja et al. (2009) and Wickham et al. (2010) suggest a Rorschach test and
a line up to compare what outcomes may be generated under a specific null hypothesis to
assess how unusual the graphs and charts obtained in exploration actually are.1 A Bayesian
1 The permutation approach toward inference in spatial autocorrelation analysis can be viewed as an

implementation of the line up approach. See Chapters 13 and 16.

66 Geovisualization

perspective is taken by Hullman and Gelman (2021), who suggest the use of a graph as a
model check. This topic is the subject of ongoing discussion and debate. I revisit some of
these ideas in the postscript (Chapter 21).

4.2.2 Mapping as exploration

Historically, cartography, the science (and art) of map making, has focused on the map
as a presentation or an expository device (Monmonier, 1993). In this context, the map is
an end product, created to represent findings, but it is not part of the analytical process
itself. In recent years, while this aspect is still important, attention has shifted to making
the map an integral part of knowledge discovery from data. This is variously referred
to as geovisualization, geovisual analytics, or even geospatial visual analytics. This effort
consists of a combination of new analytical tools, visual representations and their software
implementation. It represents a shift away from the map as an end product to the integration
of the map and a spatial focus in an interactive process of data exploration. In other words,
the map becomes one of the views manipulated in a dynamic graphics environment (for a
historical perspective, see Anselin, 2005b; 2012).2

4.2.3 Exploratory spatial data analysis

Exploratory spatial data analysis (ESDA) similarly incorporates the map and spatial
information as an integral part of the data exploration process, but the focus is on spatial
patterns. As such, ESDA can be viewed as “a collection of techniques to describe and
visualize spatial distributions, identify atypical locations or spatial outliers, discover patterns
of spatial association, clusters or hot spots and suggest spatial regimes or other forms of
spatial heterogeneity” (Anselin, 1999).
ESDA techniques, as implemented in GeoDa thus augment the map as a view of the data with
targeted searches for spatial patterns, while leveraging global and local spatial autocorrelation
statistics (spatial dependence). In addition, the discovery of the location of structural breaks
in the spatial distribution (spatial heterogeneity) is facilitated (early overviews of this
literature are contained in Anselin, 1994; 1998; 1999). In Volume 2, spatial constraints are
introduced into multivariate clustering techniques.

4.2.4 Linking and brushing

Two concepts central to the way ESDA is implemented in the architecture of GeoDa are
so-called linking and brushing (Anselin et al., 2006b). In Chapter 2, it was shown how
observations can be selected, either by creating a query in the selection tool (Section 2.5.1),
or by means of a spatial selection in a map view (Section 2.5.4).
When a selection is made in a map, such as in Figure 2.28, the corresponding observations
are also highlighted in the table. The same happens in the other direction as well. When
observations are selected in the table, the matching locations (points or polygons) are
highlighted in any open map view. The connection between the selections in all open
windows is referred to as linking. This works not just for a map and associated table (as
is the case in many GIS) but also simultaneously for any map view and all the statistical
graphs (see Chapters 7 and 8).
2 Early discussions of these concepts are included in, among others, Haslett et al. (1991), Dykes (1997),

MacEachren and Kraak (1997) and MacEachren et al. (1999). Overviews of various techniques can be found
in Dykes et al. (2005), Kraak and MacEachren (2005), Rhyne et al. (2006), Andrienko et al. (2011) and
Andrienko et al. (2018), among others.
Thematic Maps – Overview 67

A dynamic version of this process is brushing, first proposed for scatter plots in the statistical
literature by Stuetzle (1987) and Becker and Cleveland (1987). It was further extended to
choropleth maps by Monmonier (1989).
The idea behind brushing is that the selection tool (e.g., the selection rectangle on a map)
becomes a moving object. As the rectangle moves over the map, the collection of selected
objects is immediately updated. In this way, one can move the selection brush over the map
(or any statistical graph) and assess the effect of the changing selection.
In GeoDa, the concept of brushing is combined with linking in the sense that the updated
selection is instantaneously transmitted to all the open windows through the linking process.
This provides a very powerful visual tool to assess the effect of the changing selection on
various aspects of the spatial and statistical distributions, in both univariate and multivariate
settings. The linking-brushing combination is critical to support ESDA. The map plays a
central role in this process as an interactive visualization tool, discussed in more detail in
the remainder of the chapter.

4.3 Thematic Maps – Overview

4.3.1 Choropleth map
A thematic map is commonly known as a choropleth map, from the Greek word choros,
which stands for area or region. Even though sometimes referred to as chloropleth map, the
latter term is incorrect. A choropleth map is a means to visualize the spatial distribution of
a variable over discrete areal units.
The proper design of a map involves the manipulation of several parameters, including
scale, symbols, legend, intervals and colors (Monmonier, 2018). As mentioned in the chapter
introduction, this topic is much too broad to be covered in-depth here. Therefore, only some
essential elements are touched upon.
The role of datum and projection was introduced in Section 3.2. The scale of a map, i.e., the
correspondence between a distance unit on the map and a matching distance in real-life is
an aspect that is ignored in GeoDa, since any map window can be resized and/or subject to
zooming in and out. This operation changes the scale of the displayed map on the fly. As a
result, it becomes impractical to portray these continuous changes in scale.
The three map design parameters that play an important role in GeoDa are classification,
legend and color. These are briefly reviewed next.

4.3.2 Map classification

A map classification is the process of binning the observations from a continuous distribution
into discrete categories. Each of these then corresponds to a different color (or shading)
on the map. There have been proposals to visually represent the full continuity of the
distribution, e.g., by special cross-hatching (Tobler, 1973) or the use of a full color spectrum
(Brewer, 1997). However, this quickly becomes impractical for larger data sets. Alternatively,
symbols with different sizes and/or colors can be used as well. GeoDa only supports discrete
categories.
68 Geovisualization

Figure 4.2: Mapping options

In terms of designing a classification, important decisions pertain to the number of categories

and how the cut-off or break points are determined. These can be based on external (or
exogenous) considerations, such as thresholds determined by policy (e.g., income categories),
or, more commonly, endogenously. Endogenous classification exploits the properties of the
underlying distribution to come up with meaningful break points. Common methods use the
quantiles of the distribution, in a quantile map, or divide the range of the variable into a
number of equal intervals, similar to what is customary for a histogram. An approach based
on a clustering logic results in so-called natural breaks, i.e., optimal groupings of observations
into categories that are most similar internally and maximize the distinction between the
categories. These traditional methods are discussed in Section 4.4.
Other classifications that emphasize extreme observations are covered in Chapter 5.

4.3.3 Legend and color

The legend is the way in which the map classification is symbolized, typically positioned
next to the map itself. There are two important aspects to the legend. One is the choice of
color, which can have major impacts on the perception of value and pattern in the map. For
example, red colors are typically associated with hot locations, whereas blues are associated
with cold. More importantly, red is often used to represent or imply danger, especially in
political maps. In GeoDa, the selection of default color hues and gradations follows the
ColorBrewer recommendations, which can be found at colorbrewer2.org (Brewer et al., 2003;
Harrower and Brewer, 2003).
The second aspect is the gradation of values represented in the legend bar. Three broad
categories can be distinguished. For maps where the values represented are ordered and
follow a single direction, from low to high, a sequential legend is appropriate. Such a legend
typically uses a single hue and associates higher categories with increasingly darker values.
For example, this is the case for the three traditional map classifications reviewed in Section
4.4.
In contrast, for the extreme value maps in Chapter 5, the focus is on the central tendency
(mean or median) and how observations sort themselves away from the center, either in
downward or upward direction. An appropriate legend for this situation is a diverging legend,
which emphasizes the extremes in either direction. It uses two different hues, one for the
downward direction (typically blue) and one for the upward direction (typically red or
brown).
Finally, for categorical data, no order should be implied (no high or low values) and the
legend should suggest the equivalence of categories. For such discrete categories, a qualitative
legend is appropriate, as in Section 5.3.1.
Common Map Classifications 69

Figure 4.3: Quintile map for housing index, Ceará

4.3.4 Implementation
A thematic map is created in GeoDa from the Map menu item, or by selecting the left-most
toolbar icon in Figure 4.1, Maps and Rates. This yields a list of options, shown in Figure
4.2.
In the current chapter, the focus is on traditional map classifications, i.e., the Quantile
Map, Equal Intervals Map and Natural Breaks Map. The other options are treated
in the next two chapters.

4.4 Common Map Classifications

4.4.1 Quantile map
A quantile map is based on sorted values for a variable that are then grouped into bins.
Each bin has the same number of observations, the so-called quantile. The number of bins
corresponds to the particular quantile, e.g., four bins for a quartile map, or five bins for a
quintile map, two of the most commonly used categories.
This is illustrated in Figure 4.3 with a quintile map for the housing index in the Ceará data
set. The index is scaled between zero and one, with higher values corresponding to better
housing conditions.3 This map is obtained in two steps. First, the classification Quantile
Map > 5 is selected from the options in Figure 4.2. This brings up a dialog to choose the
variable to be mapped, in this example housing.
3 The housing condition index is based on the proportion of people living in shanty towns, number of
bedrooms with a maximum of two people, number of households with a maximum ratio of four people per
restroom, proportion of households whose walls are made of brick or appropriate wood, and proportion of
inadequate households. See Amaral et al. (2019) for further details.
70 Geovisualization

Figure 4.4: Tied values in a quantile map

The legend, to the left of the map, shows the type of classification (quantile) and lists the
name of the variable (housing). The five legend categories correspond with increasingly
darker shades of brown. The range of values included in each category is listed in square
parentheses and the number of observations in round parentheses. The map suggests that
lower values for the housing index tend to be concentrated in the north of the state, with
the highest (darkest) values occupying a band in the southern part.
In a quintile map, each category should contain one fifth of the total number of observations,
or, 184/5 = 36.8, so roughly 37 observations. However, in the legend in Figure 4.3, the
number of observations varies from 36 to 37 and 38. This is examined more closely in Section
4.4.1.1.
Another important characteristic of the quantile map is that the range of values in each
category is not constant. In the example, this varies from 0.1 in the lowest quintile to 0.016
for the fourth. The larger the range, the more heterogeneous the observations are that were
grouped into the same category. In other words, very distinct values may be associated
with the same color on the map, which could easily create misleading impressions about the
characteristics of the spatial distribution.

4.4.1.1 The problem with quantile maps

In a non-spatial analysis, the computation of quantiles is straightforward. One sorts the
data from low to high and looks up the value for the given quantile. In the example, those
would be at observations ranked 37, 74, 111, etc. A quick check in the table yields the values
0.797 for observation ranked 37 (actual observation 58), and 0.823 for observation ranked
74 (actual observation 59). However, the second category in the legend goes from 0.798
(observation ranked 38) to 0.822 (observation ranked 73). The true cut-off value of 0.823 is
moved to the third quintile.
A closer inspection of Figure 4.4 illustrates the problem. The third category is selected from
the legend in the map, with the corresponding observations highlighted in yellow in the
table. The observations for housing have been sorted (as indicated by the > next to the
variable name), with the corresponding rank shown in the next column.
Common Map Classifications 71

Figure 4.5: Equal intervals map for housing index, Ceará

In a non-spatial analysis, the correct value for the quintile is 0.823, without having to
necessarily specify the actual observation that matches this value. However, in a spatial
analysis, such as a thematic map, each location/observations must be allocated to a map
category. In the example, observations ranked 74 (actual 59), 75 (actual 19), 76 (actual 140)
and 77 (actual 18) all have the same value. In a thematic map, one cannot arbitrarily decide
which of these should be in category 2 and which in category 3. This is the problem of ties
in the ranking that forms the basis for the computation of the quantiles.
GeoDa uses a heuristic that assigns tied observations to the next higher map category. As a
result, the second category in Figure 4.3 contains only 36 observations, whereas the third
category contains 38.
Even though it is often used as the default setting in thematic mapping software, a quantile
map should be interpreted with caution. Widely different value ranges for the quantile
categories could mask underlying heterogeneity. In addition, the existence of ties can create
problems in practice. For example, when a large number of observations have the same value,
some quantile categories may turn out to be empty (GeoDa moves tied observations to the
next higher category). This is the case for the Zika and Microcephaly incidence variables
in the Ceará example data set, where many municipios have an incidence of zero (see also
Section 5.3.1).

4.4.2 Equal intervals map

The observations for the housing index go from a minimum of 0.697 to a maximum value of
0.93, resulting in a range of 0.233. In an equal intervals map, this range is divided into a
number of bins of equal size. For example, for a five category map, that would yield intervals
of 0.233/5 = 0.0466.
Figure 4.5 illustrates the result for the housing variable with five categories. This is produced
as Equal Intervals Map > 5, followed by selecting housing (the previously selected
variable will be the default in the variable selection dialog). The overall design of the map
window is the same as for the quantile map.
72 Geovisualization

Figure 4.6: Equal intervals map and histogram

In contrast to the quantile map, each category now contains greatly varying numbers of
observations, ranging from 7 for the highest category to 75 for the middle category. As
intended, the range of each category is exactly 0.0466. Both the lowest and the highest
categories contain much fewer observations than in the quantile map.
The patterns suggested by the equal intervals map are quite distinct from those in the
quantile map. The overall impression of a band of the lowest index value municipios in
the north has been replaced by a scattering of 8 observations, not showing any apparent
systematic pattern. Also, the band of 36 observations with the highest category in the
quantile map has been replaced by a seven locations, not grouped in any particular way.
The equal intervals map follows the same classification logic as the histogram, illustrated
in Figure 4.6. In the Figure, the two graphs are located side by side, with the highest
bar selected in the histogram. Through the process of linking, this results in the seven
observations from the fifth map category being selected in the map. In practice, inspecting
the histogram (or box map) for the variable under consideration is often instructive in
suggesting whether a quantile map or equal intervals map is appropriate for the distribution
in question.

4.4.3 Natural breaks map

A natural breaks map uses a nonlinear algorithm to group observations such that the
within-group homogeneity is maximized, following the path-breaking work of Fisher (1958)
and Jenks (1977). In essence, this is a clustering algorithm in one dimension to determine the
break points that yield groups with the largest internal similarity, i.e., the smallest internal
variance (see also Volume 2). The algorithm to obtain the optimal break points is quite
complex, and for large data sets heuristics based on sampling strategies may be necessary
(Rey et al., 2013, 2017).
An example is shown in Figure 4.7 for the housing variable, using Natural Breaks Map
> 5 to obtain five intervals.
The format of the legend in the natural breaks map differs slightly from the one used in
the previous two map types. Each interval is depicted as half open, with the lower value
included – shown by the left square bracket [ – and the upper value excluded – shown by
the right parenthesis ). Similarly, the bounds of the lowest and highest category are shown,
not the lowest and highest value as before.
Map options 73

Figure 4.7: Natural breaks map for housing index, Ceará

The patterns suggested by the map have several similar features to the equal intervals
map, in that the middle categories contain most observations, and the lowest and highest
categories are small. However, the intervals are far from equal, ranging from 0.031 for the
fifth category to 0.071 for the first.
The natural breaks bring out a northern pattern for the second category, similar to what
was suggested by the first quintile in the quantile map, and again there is a seeming band
of higher values. However, in contrast to what the quintile map suggested, the associated
observations are not the extremes.
In practice, it is important to go beyond using a single map type, and to compare the
similarities and differences between the patterns suggested by the various map classifications.

4.5 Map options

Once a map is created, there are three types of options available. One set pertains directly to
the manipulation within the current map window, such as zooming, panning and selecting.
This is implemented by means of the icons on the map toolbar (see Figure 4.10 and Section
4.5.2). The second set of options is triggered by right clicking on the map window itself.
This invokes the options dialog shown in Figure 4.8, some aspects of which were already
encountered in Chapter 3.
The top item (Change Current Map Type) brings up the same list with map types
as invoked by clicking on the Map icon on the main toolbar (see the list in Figure 4.2).
Selecting a different map type from a current map window precludes the need to choose a
variable, but it also overwrites the current map.
The various features are considered in further detail below, except for the rates functions
(Rates and Save Rates), and operations related to the spatial weights (Connectivity),
74 Geovisualization

Figure 4.8: Map options

Figure 4.9: Legend options

which are covered respectively in Chapters 6 and 10. Also, the options in the fourth group
(Shape Centers, etc.), were introduced earlier in Section 3.3.4
A final type of customization pertains to the map legend.

4.5.1 The map legend

The map legend color shades are generated automatically from a Color Brewer palette.
However, they are also fully customizable, although this is generally not recommended.
A right click on one of the legend rectangles brings up the list of options shown in Figure
4.9.5 Each of the items in the list allows for the customization of an aspect, i.e., Fill Color,
Outline Color, Display Precision, Scientific Notation and Background Color using
the standard interfaces specific to each operating system (color wheels, as well as ways to set
an explicit RGB, or CMYK value). In addition, the Copy Legend to Clipboard option
makes it possible to then paste an image of the legend into any other document.
However, it should be pointed out that for most applications, the shades suggested by Color
Brewer should not be changed.
4 The Heat Map option is covered in Chapter 20.
5A single click selects the observations that belong to that category.
Map options 75

4.5.2 The map toolbar

The map toolbar contains ten icons that facilitate selection and viewing of the map, shown
in Figure 4.10.

Figure 4.10: Map toolbar

From left to right, the icons allow the following actions:

• Select, to select one or more (using shift click) observations, or an observation region
by means of a selection shape (the default is a rectangle, see Section 2.5.4); this is the
default operation and works in the same fashion as selection on any graph in GeoDa
• Invert Select, switches to the complement of the current selection (same functionality
as Table > Invert Selection, see Section 2.5.1.2)
• Add Map Layer, add another layer to the current map window, discussed in Section
3.6.1
• Map Layer Settings, multiple map layer interface (see Section 3.6.1)
• Zoom In, zooms in on the map by drawing a rectangle for the new map extent
• Zoom Out, zooms out of the map by repeatedly clicking on the map
• Pan, implements panning by dragging the map in any given direction with the pointer
• Full Extent, returns the map to its default full extent
• Base Map, allows the map to be superimposed onto a base layer that contains roads
and other realistic geographic features (see Section 4.5.3)
• Refresh the image (in case something went wrong)
These functions are mostly standard and self-explanatory, except for the base map, which is
introduced next.

4.5.3 Map base layer (base map)

The choropleth maps created by GeoDa are abstractions and lack context. To provide a way
to connect the spatial distribution of a variable to the underlying reality on the ground, a
base layer can be added. This shows geographical features such as roads, rivers and lakes, as
well as names of places. The base layer uses the map tile concept, essentially a collection of
scalable background images (it is referred to as a base map, although it is actually a picture).
There are now several sources of such base layers available on the internet. GeoDa supports
a limited number of these tile layer sources.
In order to add a base layer to the current map window, an active internet connection must
be available (without internet, it does not work). The Base Map icon is the second from
the right in the map toolbar (Figure 4.10). This brings up a list of options, shown in Figure
4.11. The default setting is No Basemap.
Currently, GeoDa supports six main sources for base layers: Carto, ESRI, HERE, Open-
StreetMap, Stamen and Other (China).6 Each of these has several sub-options, which
can also be customized (see Section 4.5.3.1).
6 Other (China) contains tile data primarily aimed at users in China.
76 Geovisualization

Figure 4.11: Map base layer options

Figure 4.12: Stamen TonerLight base map, default map transparency

For example, continuing with the Natural Breaks map for the housing variable (Figure 4.7),
a base map that uses the Stamen > TonerLite option (second from bottom in Figure
4.11) with the default map transparency would provide the background illustrated in Figure
4.12. This adds the names of several places as well as some major roads (and the ocean) as
the geographical context for the thematic map.
However, the default map transparency of 0.69 does not do justice to the features of the
thematic map. The Change Map Transparency option (third item in the list in Figure
4.11) provides a slider bar to change the transparency between 1.0 (only the base layer
visible) and 0.0 (only the thematic map visible).7
7 Note that the transparency pertains to the base layer, not to the thematic map, hence for a larger value,

less of the thematic map will be visible.

Map options 77

Figure 4.13: Stamen TonerLight base map, 0.40 map transparency

To illustrate this effect, Figure 4.13 shows the same base layer with the transparency set to
0.40. The features of the thematic map are clearly more visible. In practice, some trial and
error may be needed to find a configuration that finds the proper compromise between the
thematic map and the background layer.
The various base layer sources have quite distinct design features and focus on different
aspects of the geography: some have more detailed street information, others include natural
features. For example, Figure 4.14 shows a base layer from the ESRI > WorldStreetMap
option (with transparency set to 0.40). This provides a visual impression that is quite distinct
from that in the previous maps.

4.5.3.1 Detailed base layer options

Each of the six base layer sources has several sub-options. These are:
• Carto: Light, Dark, Light (No Label), Dark (No Label)
• ESRI: WorldStreetMap, WorldTopoMap, WorldTerrain, Ocean
• HERE: Day, Night, Hybrid, Satellite8
• OpenStreetMap: MapNik
• Stamen: Toner, TonerLite, WaterColor
• Other (China): GaoDe, GaoDe (Satellite)
8 The HERE base map support is provided under a generic account for all GeoDa users. This account has

general usage restrictions, so users may want to set up their own accounts for the HERE platform. This can
be accomplished by means of the Basemap Configuration option. A free HERE account can be obtained
from https://developer.here.com. The basemap can then be configured with the App ID and App Key.
78 Geovisualization

Figure 4.14: ESRI WorldStreetMap base map, 0.40 map transparency

The Basemap Configuration option (at the top of Figure 4.11) provides a way to further
customize the sources for the base layer map tiles. Each entry corresponds with an entry in
the layer options following a format group_name.basemap_name,basemap_url.9 In
a typical application, these entries should not be touched. On the other hand, experts are
free to customize.
Finally, in a pinch, it may be necessary to use the Clean Basemap Cache option when
things go wrong (the second option in 4.11).

4.5.4 Saving the classification as a categorical variable

The classification associated with a given map can be added to the data table as a new
categorical variable by means of the Save Categories option (the second option in Figure
4.8 ). This brings up a variable selection dialog in which the name for the new variable can
be specified (the default is CATEGORIES).
Upon selecting OK, a new field is added to the data table. The categories are labeled from
low to high, starting with the value 1. In the example that uses natural breaks for the
housing variable (Figure 4.7), the range would go from 1 to 5. The resulting categorical
variable can be used as input into a Unique Values Map or a Co-location Map (see
Sections 5.3.1 and 5.3.2).
Note that there is no metadata associated with this new variable. As a result, all information
is lost regarding important properties, such as the variable on which the classification is
based, the type of classification, and the ranges associated with each category. To some
9 For example, the entry for Stamen > TonerLite is: Stamen.TonerLite, https://maps.stamen.com/

#toner/12/37.7706/-122.3782.
Map options 79

Figure 4.15: Saved Image As dialog

extent, in practice, one can compensate for this lack of metadata by creative naming of
variables, although this is limited by the 10 character constraint.

4.5.5 Saving the map as an image

Some other options to save features of the map are provided by the bottom three items in
the menu shown in Figure 4.8. Save Selection creates an indicator variable for the selected
observations, as covered in Section 2.5.4.3. For example, this allows one to create a separate
indicator variables for each category in the map classification, which is often useful for later
statistical analyses.10
Copy Image to Clipboard is straightforward and allows for the image to be pasted
directly into other documents.
The Save Image As option is the closest GeoDa comes to publication quality output. It
allows for limited manipulation of the map view to then be saved in a standard image output
format: png (the default), bmp, or SVG. For example, in Figure 4.15, the legend in the
natural breaks map of Figure 4.7 was moved from the top left to the lower right corner.
Limited customization is possible, such as moving and resizing legend box, changing its
transparency, or even removing it altogether. Also, the image dimensions can be set precisely
and the resolution specified (in dpi – the default is 300).
10 Select the category by clicking on its legend rectangle.
80 Geovisualization

Figure 4.16: Map without outlines visible

The saved image file can be further manipulated in specialized graphics software.

4.5.6 Other map options

The remaining three options in the Figure 4.8 menu pertain to the look of the map view.
Show Status Bar is on by default, which means that the number of observations and the
size of any selection will be displayed on the status bar. The Selection Shape is set by
default to Rectangle, with the other two options as Circle and Line (see Section 2.5.4.1).
Finally, the Color options determine how the map is displayed. By default, the polygon
outlines (i.e., in our example, the boundaries of the municipios in Ceará) are shown, with
Outlines Visible checked,
By unchecking this option, the outlines disappear. This is particularly useful when the map
contains many small areas that will tend to be dominated by their boundary lines. In the
example in Figure 4.16, this results in an impression of larger regions compared to Figure 4.7,
because many adjoining municipios end up in the same category according to the natural
breaks classification.
One further option is to Show Map Boundary, only available with Outlines Visible
turned off. This option is particularly useful when only a few polygons are colored in a map
with many observations, for example, in the local cluster maps (Chapters 16 to 19). Turning
off the outlines in such an instance makes the polygons floating in white space. With the
Show Map Boundary option turned on, the initial outline of the overall map is retained.
Custom Classifications 81

The Background Color determines the background against which the map is drawn. In
most instances, the default of white is the best choice.

4.6 Custom Classifications

So far, the classifications considered for the map legend have been endogenous, i.e., they
were derived from the distribution of the variable under consideration (e.g., quantile, equal
interval, natural breaks). However, such endogenous classification may not be that insightful
when comparing the spatial pattern of different variables (on comparable scales, e.g., indices
between 0 and 100), or when assessing the change in spatial pattern for a single variable
over time. In addition, sometimes substantive concerns dictate the cut points, rather than
data driven criteria. For example, this may be appropriate when specific income categories
are attached to a policy. In those instances, the policy-based values for the classification are
important, rather than what the internal data distribution would suggest.
When assessing the spatial distribution of a variable over time, the endogenous classifications
are relative and re-computed for each time period, based on the variable distribution at
that time. This can yield misleading impressions when the interest is in the evolution of
the actual values. For example, when mapping crime rates over time, in an era of declining
rates, the observations in the upper quartile in a later period may have crime rates that
correspond to a much lower category in an earlier period. This would not be apparent using
the endogenous classifications, but could be visualized by setting the same break points for
each time period, i.e., an exogenous classification.
In GeoDa, this is accomplished through the Category Editor.

4.6.1 Category editor

The Category Editor is a complex tool that allows for the creation of a fully customized
classification scheme. It can be invoked in a number of different ways. One option is to use
the main menu, as Map > Custom Breaks > Create New Custom. A second way is
as the bottom item from map options menu in Figure 4.2. Alternatively, the category editor
can also be invoked directly by selecting the right-most icon on the toolbar in Figure 4.1.

4.6.1.1 Design
The custom categories are designed through the interface in Figure 4.17, shown after all
editing has been completed. The initial settings and editing operations will be discussed
below.
There are slight differences in the behavior of the dialog, depending on whether it is invoked
from the toolbar, or as an option from an existing map. In the latter case, the associated
map will be updated as a new classification is being developed. When invoked from the
toolbar, there is nothing to update, but the initial variable may not be meaningful. It is the
first variable in the data table, which is often just an identifier for each observations.
When the dialog opens (or at any point when the New button is selected), the first query is
for the New Categories Title, where a name for the new classification must be specified
(the default is Custom Breaks). In the example in Figure 4.17, Custom1 is the name of
the new classification.
82 Geovisualization

Figure 4.17: Category editor interface

There are three major functions in the interface. First is the general definition of the
classification, carried out through the items in the Create Custom Breaks panel at the
top. This includes the naming of the categories, here Custom1.
The computations for the classification and their visualization in a histogram are based
on the values for a specific variable, the Assoc Var., here sanitation. When the custom
breaks are invoked from an existing map, the variable is entered automatically, but it is
possible to change it from the drop-down list.
As the breaks are edited, the impact on the associated histogram for the classification variable
is shown instantly on the right hand side of the dialog.
Other important categories in the interface are whether the Breaks are User Defined
(the default), or follow a traditional classification (Quantile, the default, Unique Values,
Natural Breaks, or Equal Intervals). These choices may seem counter intuitive for a
custom break editing operation, but it allows for the creation of custom labels for the
categories (Section 4.6.1.3).
The Color Scheme for the legend is either sequential, diverging (the default), thematic
(i.e., categorical) or custom. This provides an automatic selection of legend colors based
on the ColorBrewer palettes. However, the color for each category can also be specified by
clicking on its box below Edit Custom Breaks by means of the standard color editor (see
Section 4.5.1). Finally, the number of Categories must be specified (the default is 4).11
11 The classification can also be saved as a new variable in the table, using the Save Categories to Table

button.
Custom Classifications 83

In the example, in Figure 4.17, the color scheme has been set to sequential and the number
of Categories to 5. The main effort consists of determining appropriate break points, which
is considered next.

4.6.1.2 Editing break points

The variable sanitation is a constructed index with values between 0 and 1. The observed
range for the municipios in the State of Ceará goes from 0.417 to 0.926. These values
are shown at, respectively, the left-most point and right-most point at the bottom of the
histogram in Figure 4.17. An initial set of break points is suggested, which is almost never
the final result. Instead of the suggested values, some absolute cut-offs must be entered, such
as 0.4, 0.6, 0.8 and 0.9. This will allow for the comparison of the absolute position of each
municipio across different variables, not just their relative position (which is produced by all
endogenous classifications).
As new values for the break points are entered, the histogram is immediately updated. The
values can be typed in, or obtained by moving the slider bar to the right.
The new break points are immediately associated with the specific custom break definition,
without a saving operation. However, they can only be preserved through the use of a project
file (see Section 4.6.3). The corresponding distribution is shown in the histogram in Figure
4.17.

4.6.1.3 Custom category labels

When new break points are defined, the category labels are immediately updated to the
corresponding intervals, as long as Automatic Labels is checked. With that box unchecked,
any new label can be specified. This new label will then be used in any map legend.

4.6.2 Applications of custom categories

To illustrate the application of custom categories, consider the variables sanitation and
infra (infrastructure). Both are obtained from the Brazilian Index of Urban Structure
(IBEU) for 2013. They take on values from 0 to 1. The relative position of the municipios
in Ceará on those two variables can be visualized by means of a quintile map, as in Figure
4.18. The spatial patterns are different, but some high performing municipios overlap.
Contrast this pattern with the maps in Figure 4.19, based on the absolute cut-off points of
0.4, 0.6. 0.8 and 0.9. Clearly, the municipios do not perform as well on the infrastructure
index as they do on the sanitation index. For example, the latter has 0 observations in the
lowest category (< 0.4), whereas for infrastructure there are 16 such observations. Also,
infrastructure has 0 municipios in the top category (> 0.9), whereas sanitation has 2. This
allows for the comparison of not only the relative performance of the municipios but also
how they meet some absolute standards (admittedly arbitrary in this example).

4.6.3 Saving the custom categories – the Project File

When a project is closed, the information on the custom classification is lost, unless it
is first saved in a so-called Project File. This file, associated with a geographic input
layer, contains information on various variable transformations and other operations, most
importantly dealing with spatial weights (see Chapter 10). It also contains the definition
of any custom categories that were created. If this definition is not saved in a project file,
84 Geovisualization

Figure 4.18: Quintile maps for sanitation and infrastructure

Figure 4.19: Thematic maps for sanitation and infrastructure using custom classification

then it will be lost and will need to be recreated from scratch the next time the data set is
analyzed.
The project file is created from the menu as File > Save Project. The file is saved with a
file extension of gda. It is a text file that includes XML encoding. For example, a project
file associated with the Ceará data could be zika_ceara.gda. Closer inspection reveals the
section pertaining to custom_classifications in Figure 4.20.
The custom classification section contains all the aspects needed for the definition of the
custom category.
When an analysis is started with the project file as input (e.g., instead of a shape file) then
all transformations and custom categories become immediately available as an option for
any map or histogram.
Custom Classifications 85

Figure 4.20: Custom category definition in project file

Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
5
Statistical Maps

In this Chapter, I continue the exploration of mapping options with a focus on statistical
maps, in particular maps that are designed to highlight extreme values or outliers. Some of
these classifications were originally introduced in GeoDa and are gradually being adopted by
other exploratory software (e.g., the Python PySAL library). Their layout illustrates the
emphasis on statistical exploration rather than cartographic design. Specifically, this includes
the Percentile Map, Box Map (with two options for the hinge) and the Standard
Deviation Map.
Other topics covered in this Chapter include maps for categorical variables in the form of a
Unique Values Map. Construction of this type of map does not involve a classification
algorithm, since it uses the integer values of a categorical variable itself as the map categories.
The Co-location Map is an extension of this principle to multiple categorical variables.
The Chapter closes with a brief discussion of the Cartogram and map animation (Map
Movie), which move beyond the traditional choropleth framework to visualize a spatial
distribution.
I continue to use the Ceará Zika sample data set to illustrate the various features.

5.1 Topics Covered

• Create extreme value statistical maps
• Identify outliers using the box map and standard deviation map
• Create a map for a categorical variable
• Examine multivariate co-location patterns with a co-location map
• Construct and interpret a cartogram
• Visually explore patterns using map animation
GeoDa Functions
• Map > Percentile Map
• Map > Box Map
– Set hinges as 1.5 or 3.0
• Map > Standard Deviation Map
• Map > Unique Values Map
• Map > Co-location Map
• Map > Cartogram
– improving the cartogram fit
• Map > Map Movie
– setting animation controls

DOI: 10.1201/9781003274919-5 87
88 Statistical Maps

Figure 5.1: Percentile map

5.2 Extreme Value Maps

Extreme value maps are variations of common choropleth maps where the classification is
designed to highlight observations at the extreme lower and upper end of the scale. The
objective is to identify and highlight outliers. These maps were developed in the spirit of
spatializing EDA, i.e., adding spatial features to commonly used approaches in non-spatial
EDA (Anselin, 1994).
GeoDa currently supports three such map types in the Map menu: a Percentile Map,
a Box Map and a Standard Deviation Map. These are briefly described below. Only
their distinctive features are highlighted, since they share all the same options with the
traditional thematic map types (Section 4.5). It should be noted that the extreme value
maps are examples of a diverging legend, whereas the conventional maps have a sequential
legend.
The extreme value maps are the third to sixth items in the map menu of Figure 4.2.

5.2.1 Percentile map

The percentile map is a variant of a quantile map that would start off with 100 categories.
However, rather than having these 100 categories, the map classification is reduced to six
ranges, the lowest 1%, 1–10%, 10–50%, 50–90%, 90–99% and the top 1%. This is shown in
Figure 5.1 for the housing variable from the Ceará sample data set.
This map is created by selecting Map > Percentile Map from the map menu or the map
toolbar icon and specifying the variable.
Compared to the traditional map classifications discussed in Section 4.4, the extreme values
are much better highlighted. Both the bottom and top 1% contain two observations, with a
very small range, respectively 0.697 to 0.715 and 0.927 to 0.930. Interestingly, these extreme
Extreme Value Maps 89

Figure 5.2: Box map, hinge = 1.5, with box plot

values are not closely together in space, countering any suggestion of clustering. However,
at least in one case (the northernmost upper percentile observation), a very high value is
surrounded by much lower ones, suggesting the potential of a spatial outlier.
With the focus on the large range of middle values (i.e., 10–50 percentile and 50–90), two
large bands of adjoining municipios in the same category seem to manifest themselves, with
below median values in the northeast and above median values forming a U-shape in the
center of the state.
As for any quantile map, there are some drawbacks to this approach related to ties and
potential heterogeneity of values within the same category. In addition, a percentile map
only makes sense when there are more than 100 observations, which is the case here. It is
particularly useful to identify the location in space of the truly extreme observations.

5.2.2 Box map

Whereas the category bounds in the percentile map are to some extent arbitrary, in the box
map these are connected to the visualization of the distribution of a variable in a box plot
(see Chapter 7).
The box map (Anselin, 1994) is thus the mapping counterpart of the idea behind a box plot.
The point of departure is again a quantile map, or, more specifically, a quartile map. The
four categories are extended to six bins, to separately identify the lower and upper outliers.
The definition of outliers is a function of a multiple of the inter-quartile range (IQR), the
difference between the values for the 75 and 25 percentile. As is customary, there are two
options for these cut-off values, or hinges in a box plot: 1.5 and 3.0. The box map uses the
same convention.
The box map is created by selecting Map > Box Map (Hinge=1.5) from the map options
and specifying the variable. For the Ceará housing index variable, this yields the map shown
90 Statistical Maps

Figure 5.3: Outliers in box plot and box map

in the left panel of Figure 5.2. For easy reference, the corresponding box plot is contained in
the right panel.
Compared to a standard quartile map, the box map in Figure 5.2 separates out six lower
outliers from the other 40 observations in the first quartile. They are depicted in dark blue.
Similarly, it separates a single upper outlier from the 43 other observations in the upper
quartile. The upper outlier is colored dark red. The main focus of interest in a box map is to
identify the extent to which the outliers show any kind of spatial pattern. In this example,
there does not seem to be a suggestion of clustering, but possibly of the presence of spatial
outliers (this is explored more formally in Chapters 16 to 19). This constitutes the spatial
perspective that the box map adds to the data exploration.
To further illustrate the correspondence between the box plot and the box map, in Figure
5.3, the six lower outliers are selected in the box plot in the right hand panel. Through
linking, these are also highlighted in the map and correspond exactly to the lower outlier
category. Again, there is a fundamental difference between a traditional box plot where
outlying observations are identified, and the box map, where their location is taken into
account as well.
From the current map, the box map can be switched between the hinge criterion of 1.5 and
3.0 by opening the options menu (right click on the map) and selecting Change Current
Map Type > Box Map (Hinge = 3.0). Alternatively, a new map window can be opened
from the main menu or map toolbar icon by selecting Box Map (Hinge=3.0) as the
option.
The box map is arguably the preferred method to quickly and efficiently identify outliers
and broad spatial patterns in a data set, although it does suffer form the same drawbacks as
any other quantile map.
Extreme Value Maps 91

Figure 5.4: Standard deviation map

5.2.3 Standard deviation map

The third type of extreme values map is a standard deviation map. In some way, this is
a parametric counterpart to the box map, in that the standard deviation is used as the
criterion to identify outliers, instead of the inter-quartile range.
In a standard deviation map, the variable under consideration is represented as standard
deviational units (with mean 0 and standard deviation 1). This is equivalent to the z-
standardization (see Section 2.4.2.3).
The number of categories in the classification depends on the range of values, i.e., how many
standard deviational units cover the range from lowest to highest. It is possible that some
categories do not contain any observations, since there may be gaps in the distribution for a
given standard deviational range.1
The relevant option is Map > Standard Deviation Map, which yields the map in Figure
5.4 for the Ceará housing index example.
In the example, there are six categories, each with a range of 0.040. There are eight
observations that are more than two standard deviational units away from the mean in the
lower direction, colored dark blue. As in the other extreme value maps, they do not seem to
show any particular spatial pattern.
In the upward direction, there are four observations more than two standard deviational
units from the mean, which is quite a bit more than identified with the box map. The one
observations in the north seems to be both an outlier in the value distribution as well as a
spatial outlier. Otherwise, there is no indication of any clustering of the extremes.
On the other hand, when focusing on the central values in the distribution, again there is
the suggestion of a northern band of lower values and a U-shape pattern of observations
within one standard deviation above the mean.
1 For example, this would be the case when outlying observations are more than one standard deviational

unit away from the rest of the distribution.

92 Statistical Maps

Figure 5.5: Unique values map, Zika indicator

In practice, it is often useful to compare the outliers identified by both the non-parametric
(box map) and parametric (standard deviation map) approaches. Of particular importance is
whether extreme values show an interesting spatial pattern, which can then be more formally
investigated by means of the local spatial autocorrelation indicators in Chapters 16 to 19.

5.3 Mapping Categorical Variables

In the discussion so far, the variable of interest (housing) was continuous, with a clear
ordering from low to high. However, one often deals with indicator variables (0–1), or, even
more generally, with variables that represent multiple categories. For the latter, each category
is distinguished by a different integer value, but those values are typically not meaningful in
and of themselves. Most importantly, the numerical values do not imply any ordering of the
categories.
For a single variable under consideration, a map of its spatial distribution is created through
the Unique Values Map option, the seventh item in the list in Figure 4.2. The functionality
is illustrated with an indicator variable constructed from the cumulative incidence of Zika
during the first three quarters of 2016 in the state of Ceará (zika_d).
A Co-location Map (eighth item in Figure 4.2) extends this to the comparison of multiple
categorical variables, using the logic of map algebra (Tomlin, 1990). This is illustrated with a
comparison of the cumulative incidence of Zika (zika_d) with the presence of Microcephaly
in the fourth quarter of 2016 (mic_d).
Mapping Categorical Variables 93

Figure 5.6: Co-location map variable selection, Zika and Microcephaly indicators

5.3.1 Unique values map

A map for a categorical variable is created by invoking Map > Unique Values Map and
specifying the variable. Note that only variables that take on integer values are included in
the drop-down list.2
In the example in Figure 5.5, the categorical variable is zika_d. It only takes on the values
of 1 (presence) and 0 (absence), reflected in the two colors in the map legend. The legend
colors are generated from the ColorBrewer categorical map palette.3 As mentioned in Section
4.5.1, they can also be customized.
The map indicates that 32 of the 184 municipios had an incidence of Zika during the period
under consideration. There seems to be a suggestion of a greater presence in the northern
part of the state, with some groupings consisting of adjoining locations. A more formal
assessment of these patterns is investigated in Chapter 19.
All the standard map options apply to the unique values map. One additional option is to
change the order (and color) of the categories. Since the categories are just that, and the
2 If a variable that should be categorical appears to be missing from the list, it may have been formatted

as real. This can be readily changed using Edit Variable Properties in the table, see Section 2.4.1.1.
3 http://colorbrewer2.org/#type=qualitative&scheme=Accent&n=3
94 Statistical Maps

Figure 5.7: Co-location map, Zika and Microcephaly incidence

numerical values associated with them do not imply any ordering, they can be changed.
This is accomplished by grabbing the associated legend rectangle and moving it up or down
in the list.
Such reordering can be handy in a so-called cluster map where the categories correspond to
different cluster classifications. The comparison of two cluster maps can be facilitated by
moving the categories around so that the same colors more or less correspond to the same
locations.4

5.3.2 Co-location map

A co-location map combines the information from multiple categorical variables into a unique
values map that shows those locations where the categories match. This is an example of
map algebra, but applied to irregular spatial units rather than the more customary raster
data.5
In essence, the process boils down to finding those locations where the codes for different
categorical variables match. This is handled slightly differently for the simple case of binary
variables and the more complex situation with multiple categories.

5.3.2.1 Binary categories

When the variables under consideration take on binary values, a co-location map can be
constructed by hand as a unique values map of the product of the respective indicator
4 This is covered in Volume 2.
5 An exhaustive treatment can be found in Tomlin (1990).
Mapping Categorical Variables 95

Figure 5.8: Selection of co-located observations

variables. Only those locations that take on a value of 1 for all variables will be coded as
one in the resulting unique values map.
However, a co-location map is different in that it also provides information on matches of
locations with 0 for all variables, as well as on the locations of mismatches. In sum, rather
than just two categories (match and no match), there are three: match of 1, match of 0 and
mismatch.
The map is invoked in the usual way as Map > Co-location Map, which brings up a
dialog to specify the different variables to be considered, as in Figure 5.6. The interface
is used to select the variables but also to choose the associated legend structure. Several
options are provided, ranging from Unique Values, the suggested default in this case, to
several customized legends associate with different types of maps (such as extreme values
maps) and visualizations of spatial analyses (e.g., LISA Map, see Chapter 16).
In our example, the Unique Values default will do fine.
The two variables selected are indicator variables for the presence of respectively Zika
(zika_d) and Microcephaly (mic_d). The corresponding co-location map is shown in
Figure 5.7.
The legend contains three categories. One pertains to those locations with a common
occurrence of 1 (17 observations), and another to those observations that share a value of 0
(115 observations). The third category (in grey) highlights the locations where there is a
mismatch between the values (52 observations).
The logic of the co-location map is further illustrated in Figure 5.8. It shows the selected
observations in each of the respective unique values maps that correspond to locations with
a value of 1 in the co-location map. There are 17 observations selected in each of the maps.
Clearly, these are the only locations where both maps have a dark blue color.

5.3.2.2 Multiple categories

When the variables pertain to multiple categories, the logic of the co-location map is the
same, but it must be applied with caution. It is based on the equality of the categorical
codes for each variable.
It is up to the user to ensure that the categories across variables are meaningful, since the
co-location is based on the variables having the same code. For example, this is useful when
comparing the extent to which the quartiles across different variables occur at the same
96 Statistical Maps

Figure 5.9: Cartogram

locations. Or, to assess whether significant patterns of local spatial autocorrelation match
across multiple variables. But it is also very easy to generate nonsensical results, for example,
when the labels are not comparable.
One can change the label and color of a category by moving the legend rectangle up or
down in the legend. This again highlights that the category values do not have any intrinsic
numerical value.

5.4 Cartogram
5.4.1 Principle
A cartogram is a map type where the original layout of the areal unit is replaced by a regular
geometric shape (usually a circle, rectangle, or hexagon) that is proportional to the value of
the variable for the location. This is in contrast to a standard choropleth map, where the
size of the polygon corresponds to the area of the location in question. The cartogram has a
long history and many variants have been suggested, some quite creative (see Dorling, 1996;
Tobler, 2004, for an extensive discussion of various aspects of the cartogram). In essence, the
construction of a cartogram is an example of a nonlinear optimization problem, where the
geometric forms have to be located such that they reflect the topology (spatial arrangement)
of the locations as closely as possible.
GeoDa implements a circular cartogram, in which the areal units are represented as circles,
whose size (and color) is proportional to the value observed at that location. The changed
Cartogram 97

Figure 5.10: Linked Cartogram and map

shapes remove the misleading effect that the area of the unit might have on perception of
magnitude.

5.4.2 Creating a cartogram

The cartogram is created from the main menu, as Map > Cartogram, or by selecting the
second item in the toolbar shown in Figure 4.1.
This brings up a dialog to specify the Cartogram Variables, both for the Circle Size and
the Circle Color. In most applications, those should be the same variable. In the example
shown in Figure 5.9, the variable used is again the housing index.
The default is a Box Map classification for the circle colors (with hinge at 1.5 IQR).
The cartogram is most insightful when used in conjunction with a regular choropleth map.
Selecting an observation in the cartogram then immediately links it with the corresponding
area in the choropleth map. In Figure 5.10, this identifies an observation in the upper quartile
as a tiny municipio in the south-east of the state. Note how the location of the observation
in the cartogram is not identical to that in the thematic map.
Except for multi-layer and base map, the cartogram has all the same options as a regular
choropleth map, with one addition. The positioning of the circles in the cartogram is the
result of a non-linear optimization approach. The algorithm attempts to locate the center of
the circle as close as possible to the centroid of the areal unit with which it corresponds, while
respecting the contiguity structure as much as possible. As illustrated in Figure 5.10, this is
not perfect. There is no unique solution to this problem, and it may be useful to experiment
with further iterations that will slightly reposition the circles. This is implemented in the
Improve Cartogram option. A number of different iteration options are listed, together
with the estimated time. The latter is particularly useful for larger data sets. Each additional
set of iterations will result in a slight reshuffling of the circles. In most applications, the
initial layout is sufficient for linking and brushing.
98 Statistical Maps

Figure 5.11: Animation control panel

5.5 Map Animation

Animation is a method to discover patterns in data by moving through observations in a
systematic fashion. A typical example is to consider maps or graphs for the same variable over
time (see Chapter 9). Alternatively, a given map or graph can be explored by moving through
the observations from low to high, or from high to low, either individually (highlighting one
at a time), or cumulatively (taking up increasingly larger parts of the map or graph).
In GeoDa, animation is implemented through the Map Movie functionality (referred to as
such for historical reasons). It is invoked from the menu as Map > Map Movie or by
selecting the third icon in the toolbar shown in Figure 4.1.
This brings up the Animation dialog, shown in Figure 5.11, the control center through
which the various aspects of the animation are manipulated. The first item to specify is the
Variable, selected from a drop-down list, here sanitation.
At the bottom of the dialog are the main controls: the start > button, step-by-step forward
» or backward «, whether the animation loops or stops at the end, an option to Reverse
the progress, the speed of the animation, and whether the order followed is ascending or
descending. The defaults are usually good, with Cumulative checked (i.e., the selection
grows as the animation progresses) and Ascending order.
Once the forward button is activated, each observation is selected in turn, starting with the
lowest value. This selection is not only for the map, but, through linking, for all currently
active windows. The slider in the Animation dialog moves from left to right. Under the
variable name, the currently selected observation and its value are listed.
The animation tool can be paused at any point, reversed, changed from continuous change
to step-by-step, etc., using the controls provided. The main point of the animation is to
visually check for any patterns, such as all the lowest or highest values occurring in one
location, or an increase in value that follows a given spatial trend (e.g., core-periphery, or
East-West). Of course, this visual impression is only that and will need to be confirmed with
the more formal pattern detection methods covered in later chapters.
A particular powerful feature of the animation tool is to assess the spatial relationship
between two (or more) variables. This is carried out by selecting one as the variable in the
animation control, but following the selected observations for a different variable.
In the example in Figure 5.12, the progression is based on the variable sanitation. Two
quintile maps are shown side by side, one for sanitation and one for infra (see also Figure
4.18).
Map Animation 99

Figure 5.12: Linked animation sanitation-infrastructure

As shown in Figure 5.11, the current cut-off value for sanitation is 0.536. This results in
21 of the lowest observations to be highlighted in the map on the left. Through linking,
the corresponding observations in the map on the right, for infra, are highlighted as well.
This allows for the comparison of the cumulative locations of the values from low to high
between the two variables. More precisely, one assesses the extent to which the colors for
the matching locations in the two maps follow the same pattern. In the example, several of
the lowest observations for sanitation belong to a higher quintile in the map for infra.
An extension to multiple maps and graphs is straightforward.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
6
Maps for Rates

In Chapters 4 and 5, the maps pertained to a single variable. In the current chapter, I
deal with some special aspects related to the mapping of rates or proportions. In GeoDa,
such variables can be constructed on the fly, by specifying a numerator and a denominator.
The numerator is typically a type of event, such as the incidence of a disease, and the
denominator is the population at risk. Such data have broad applications in public health
and criminology. However, several of the principles covered can be equally applied to any
ratio, such as an unemployment rate (number of unemployed people in a given region as a
share of the labor force), or any other per capita measure (e.g., gross regional product per
capita).
The chapter contains three main sections. First, the creation of a raw rate or crude rate map
is considered, which is the same as the types of maps considered so far, except that the rate
is calculated on the fly. Next, the focus shifts to so-called excess risk maps, which compute a
measure that compares the value at each location to an overall average, highlighting extreme
observations. Such excess risk is known under different terms in various literatures, such
as a standardized mortality rate (SMR) in demography (Preston et al., 2001), or a location
quotient (LQ) in regional economics (McCann, 2001).
Finally, attention shifts to the important topic of variance instability that pertains to any
rate measure, and the associated concept of rate smoothing. In essence, the precision of
the rate as an estimate for the underlying risk depends on the size of the denominator. In
practice, this means that rates estimated for small populations (e.g., rural areas) may have
large standard errors and provide imprecise estimates for the actual risk. This may lead to
erroneous suggestions of extreme values, such as the presence of outliers. Rate smoothing
techniques use a Bayesian logic to borrow strength and adjust the small area estimates. There
is a large literature in statistics dealing with such techniques (e.g., Lawson et al., 2003).
Here, the discussion will be limited to the basic principle, illustrated by the most common
form of Empirical Bayes smoothing.
I continue to use the Ceará Zika sample data set to illustrate the various features.

6.1 Topics Covered

• Create thematic maps for rates
• Assess extreme rate values by means of an excess risk map
• Understand the principle behind shrinkage estimation or smoothing rates
• Apply the Empirical Bayes smoothing principle to maps for rates
• Compute crude rates and smoothed rates in the table

DOI: 10.1201/9781003274919-6 101

102 Maps for Rates

GeoDa Functions
• Map > Rates-Calculated Map
– Raw Rate
– Excess Risk
– Empirical Bayes
– saving calculated rates to the table
• Table > Calculator > Rates
– Raw Rate
– Excess Risk
– Empirical Bayes

6.2 Choropleth Maps for Rates

The distinguishing characteristic of the implementation of maps for rates or proportions in
GeoDa is the explicit computation of the ratio by specifying variables for the numerator and
the denominator. The map types are the same as those reviewed in Chapters 4 and 5. In
addition, there are some options to save the rates in the table.
Before delving into the specifics, I first discuss the important difference between spatially
extensive and spatially intensive variables. Only the latter should be used in a choropleth
map. Next, I outline the implementation of the rate maps and some important options.

6.2.1 Spatially extensive and spatially intensive variables

Many variables in empirical studies are directly correlated with the size of the observational
unit. Examples include total population, number of housing units, count of crimes, etc. In
essence, everything else being the same, larger observational units should have larger values
for such variables. They are therefore referred to as spatially extensive, being directly related
to the size of the observational unit. Unless all spatial units have the same area, this can
lead to spurious suggestions of importance or outlier values. Consequently, choropleth maps,
which use the area of the spatial unit to represent observations, are not an appropriate
application for spatially extensive variables.
Instead, the variables should be standardized in some fashion, so that they reflect intrinsic
variation, rather than variation in the size of the observational unit. This is readily accom-
plished by dividing by some measure of size. The classic example is to use the population
at risk, such as the population exposed to the risk of a disease. This implies that the rate
is interpreted as an estimate for the underlying risk (see Section 6.4.1). Alternatives are
total area (resulting in density measures), or some other total, such as total population
(resulting in per capita measures), without the implication of measuring an underlying risk.
The resulting spatially intensive variables are appropriate for thematic mapping.
More formally, if Oi is the value for the numerator in area i, and Pi is the corresponding
denominator, then the raw or crude rate or proportion follows as:
Oi
ri = .
Pi

To illustrate this concept, consider the box maps in Figures 6.1 and 6.2, depicting, respectively,
population (pop) and population density (popdens), and GDP (gdp) and GDP per capita
Choropleth Maps for Rates 103

Figure 6.1: Population and population density

Figure 6.2: GDP and GDP per capita

(gdpcap) for the municipios in the state of Ceará. The left panel pertains to the spatially
extensive variable, while the right-hand panel shows the corresponding spatially intensive
variable.
Whereas the population box maps suggests 10 outliers, several of which are spread throughout
the state, the population density map has 13 outliers, mostly concentrated around the urban
area of Fortaleza. Also, as highlighted in the two maps, several of the larger areas that rank
in the upper quartile for total population, drop to the first quartile in terms of population
density, i.e., after the area of the municipios is corrected for (their colors changes from
browns to blues). In addition, most of the outliers in the larger (rural) areas, are no longer
characterized as such in the population density map.
A similar phenomenon occurs in the maps for total GDP and GDP per capita. The number
of outliers for GDP (a highly skewed variable) drops from 25 to 18 in the per capita map.
Again, most of the larger (and rural) areas drop in the ranking, highlighted by their change
in color from browns to blues.
In sum, a proper indication of the variability in the spatial distribution of the variable of
interest is only obtained when the latter is converted to a spatially intensive form.
104 Maps for Rates

Figure 6.3: Rate map interface

Figure 6.4: GDP per capita as a rate map

6.2.2 Raw rate map

A Raw Rate or crude rate map is invoked from the map menu or toolbar icon in the usual
fashion, by selecting the bottom item in the interface from Figure 4.2: Map > Rates-
Calculated Maps > Raw Rate. Alternatively, it can be created from an existing map
window by selecting Rates in the list of map options (Figure 4.8).
In contrast to the previous maps, there is not a query for a single variable, but instead both
the numerator (Event Variable) and denominator (Base Variable) of the ratio need to
be specified, as shown in Figure 6.3. In the example, the respective variables are gdp and
pop. The drop-down list provides all the available map types. Here, a Box Map with hinge
1.5 is selected, resulting in the map shown in Figure 6.4.
Choropleth Maps for Rates 105

Figure 6.5: Rate calculation in table

Figure 6.6: GDP per capita in table

Everything is the same as before, except that the map heading spells out both the numerator
and denominator variables: Raw Rate gdp over pop. Specifically, the map is identical
to the right-hand panel in Figure 6.2, which uses the gdpcap variable from the data table,
rather than computing the ratio explicitly.
A rate map retains all the standard options of a map in GeoDa (see Section 4.5). Two
additional features are Rates and Save Rates. Rates brings up a list of the different rate
map options, since these are not available in the Change Current Map Type function in
the main options menu. The Save Rates option allows for the calculated rates to be added
to the data table. The default variable name for a raw rate is R_RAW_RT). As before,
the new variable only becomes a permanent addition after a Save or Save As operation.

6.2.2.1 Rates in the table

Rates can also be computed directly in the table as part of the Calculator functionality,
by selecting the Rates tab in the interface, shown in Figure 6.5 (see also Section 2.4.2 in
Chapter 2). The particular type of rate is selected from the Method drop-down list. In this
example, the Raw Rate is used. As usual, a new variable must be added (here, r_gdpcap)
and both the Event Variable (gdp) and Base Variable (pop) must be spelled out.
Figure 6.6 illustrates the result. It shows side by side the original per capita GDP variable
(gdpcap), the rate saved from the map window (R_RAW_RT), and the rate computed in
the table (r_gdpcap). Rounded to the third decimal of precision, the values are identical.
106 Maps for Rates

6.3 Excess Risk – SMR – LQ

6.3.1 Relative risk
In Chapter 5, several map types were introduced that focus the attention on extreme values.
Clearly, these same maps can be applied to rates, as in the illustration in Figure 6.4, showing
a box map. However, for rates, there is an additional concept that helps in identifying
extreme observations.
The concept is that of a relative rate or a relative risk, also referred to as excess risk. In
essence, this boils down to the ratio of the observed rate at a location to some reference
rate. In demography and public health analysis, such a ratio is referred to as a standardized
mortality rate (SMR). In regional economics, when applied to the fraction of employment in
a given sector, the corresponding concept is that of a location quotient (LQ).
The idea is to compare the observed rate at a small area to a national (or regional) standard.
More specifically, the observed number of events (Oi ) is compared to the number of events
that would be expected (Ei ), had a reference risk been applied.
In most applications, the reference risk is estimated from the aggregate of all the observations
under consideration, as the ratio of the sum of numerators over the sum of denominators.
Formally, this is expressed as:
i=n
Oi
π̃ = i=1
i=n
,
i=1 Pi
with Oi and Pi as before. Note that this is different from a simple average of the rates. In
fact, it is a population weighted average that properly assigns the contribution of each area
to the overall total.
If the ratio π̃ would apply equally to each observation, the result would be the expected value
for that observation. In demography and public health, this would be an expected number
of deaths or incidence of a disease. In regional economics, it would be the expected share of
employment in a sector (equal to the reference share). In general, this can be expressed as:
Ei = π̃ × Pi .
The relative risk then follows as the ratio of the observed rate over the reference rate, or,
equivalently, of the observed number of events over the expected number of events:
ri Oi /Pi Oi
RRi = = =
π̃i Ei /Pi Ei
.
If an area matches the (regional) reference rate, the corresponding relative risk is one. Values
greater than one suggest an excess, whereas values smaller than one suggest a shortfall. The
interpretation depends on the context. For example, in disease analysis, a relative risk larger
than one would indicate an area where the prevalence of the disease is greater than would be
expected. In regional economics, a location quotient greater than one, suggests employment
in a sector that exceeds the local needs, implying an export sector.

6.3.2 Excess risk map

An Excess Risk map is a special rate map that uses a custom classification to indicate
observations that fall below or exceed the reference relative risk rate of one. The legend is
Excess Risk – SMR – LQ 107

Figure 6.7: GDP per capita excess risk

divergent, using blue hues to indicate values smaller than one and brown hues for values
larger than one. The intervals are 1–2, 2–4 and greater than 4 on the high end, and 0.5–1,
0.25–0.5, and <0.25 on the low end.
The map is invoked as Maps > Rates-Calculated Maps > Excess Risk in the usual
fashion. Again, both the Event Variable and the Base Variable must be specified. In
Figure 6.7, the resulting map is shown for gdp and pop as the numerator and denominator.
The great majority of municipalities do not reach the state GDP per capita, yielding a
blue hue in the map. Only nine observations have a GDP per capita that exceeds the state
average by a multiple: five between 1 and 2, three between 2 and 4, and one observation has
an excess risk rate greater than 4, highlighted in the map.
As it turns out, the identified observation is the small rural municipality of Quixeré, which
specializes in high-tech agriculture. Due to the high capital intensive nature of that production
and its small population, it has an extremely high GDP per capita from that activity (see
also Figure 6.8). However, this extreme value may be an artifact of the particular data set.1
All standard map options also apply to the excess risk map.

6.3.2.1 Saving and calculating excess risk

As shown in Section 6.2.2.1, the rates can be saved to the table, as well as calculated directly
in the table. This works in the same way for the excess risk ratios. The default variable
name when saving the rate is R_EXCESS.
Figure 6.8 illustrates this for the GDP per capita. The original gdpcap is shown, together
with the excess rate. The observations are ranked by the excess rate. The highest rate of 4.34
is obtained by Quixeré, with a GDP per capita of 40. The total GDP for the state of Ceará
1 There may be a problem with the GDP data for that municipality in the Ceará data set. The figure for

GDP is from the 2013 IBGE publication. As it turns out, in later years, this figure was revised downward.
108 Maps for Rates

Figure 6.8: GDP per capita excess risk in table

is 77,865,442 rais (i.e., the sum of the GDP in all municipalities), with a matching total
population of 8,452,380. Consequently, the state GDP per capital is 9.21. Clearly, the GDP
per capita of 40 for Quixeré exceeds this more than four-fold. As pointed out in the review
of the map, this is actually an exception, with most of the municipalities not matching the
state average. Figure 6.8 shows the top nine municipalities with excess rates greater than
one. More importantly, the spatial distribution shown in the map in Figure 6.7 highlights
the importance of spatial heterogeneity. Specifically, assuming that the economies of the
municipalities in the state follow the state GDP per capita is highly misleading, as only a
few such locations meet or exceed that standard.

6.4 Rate Smoothing

In many empirical investigations, the real interest is in the spatial distribution of the
underlying risk of an event. While the concept of risk has many meanings, in this context it
represents the probability of an event occurring. To the extent that the risk is not constant
across space, its spatial distribution is represented by a so-called risk surface. The problem
is that such a surface is not observed and thus must be estimated.
As it turns out, the crude rate is a good estimate for the unknown risk in that it is unbiased
(on average, it is on target). However, it suffers from the problem that its precision (or,
alternatively, its variance) is not constant and depends on the size of the corresponding
population at risk (the denominator). This intrinsic variance instability of rates causes
problems for the interpretation of maps, especially in terms of the identification of outliers.
In addition, there are challenges to create meaningful maps when dealing with many locations
that have no observed events.
In response, a number of procedures have been suggested that trade off some bias in order
to increase overall precision, by borrowing strength.
In this section, I first present more formally some important concepts related to variance
instability, as well as ways to improve the risk estimates by means of smoothing rates. In the
current chapter, the discussion is limited to one particular approach, i.e., Empirical Bayes
smoothing. Spatial smoothing methods are covered in Chapter 12.
Rate Smoothing 109

6.4.1 Variance instability

The point of departure in the consideration of rate smoothing is to view the count of events
that occur in a location i, Oi , as the number of successes in Pi draws from a probability
distribution driven by a risk parameter πi . The textbook analog is to draw n balls from an
urn filled with red and white balls, say with a proportion of p red balls. Given n such draws,
what is the probability that q balls will be red?
The formal mathematical framework for this type of situation is the binomial distribution,
B(n, p). The probability function for this distribution gives the probability of k successes in
n successive draws from a population with risk parameter p:
 
n k
Prob[k] = p (1 − p)n−k ,
k
 
where nk is the binomial coefficient.2 Translated into the context of Oi events out of a
population of Pi , the corresponding probability is:
 
Pi Oi
Prob[Oi ] = π (1 − πi )Pi −Oi ,
Oi i

with πi as the underlying risk parameter. The mean of this binomial distribution is πi Pi ,
and the variance is πi (1 − πi )Pi .
Returning to the crude rate ri = Oi /Pi , it can be readily seen that its mean corresponds to
the underlying risk:3
Oi E[Oi ] πi Pi
E[ri ] = E[ ] = = = πi .
Pi Pi Pi
Consequently, the crude rate is an unbiased estimator for the underlying risk.
However, the variance has some undesirable properties. A little algebra yields the variance
as:
πi (1 − πi )Pi πi (1 − πi )
V ar[ri ] = = .
Pi2 Pi
This result implies that the variance depends on the mean, a non-standard situation and an
additional degree of complexity. More importantly, it implies that the larger the population
of an area (Pi in the denominator), the smaller the variance for the estimator, or, in other
words, the greater the precision.
The flip side of this result is that for areas with sparse populations (small Pi ), the estimate
for the risk will be imprecise (large variance). Moreover, since the population typically
varies across the areas under consideration, the precision of each rate will vary as well. This
variance instability needs to somehow be reflected in the map, or corrected for, to avoid a
spurious representation of the spatial distribution of the underlying risk. This is the main
motivation for smoothing rates, considered next.

6.4.2 Borrowing strength

Approaches to smooth rates, also called shrinkage estimators, improve on the precision
of the crude rate by borrowing strength from the other observations. This idea goes back
to the fundamental contributions of James and Stein (the so-called James-Stein paradox),
 
2 n
k
= n!/k!(n − k)!, or the number of different combinations of k observations out of n.
3 Only the number of observed events Oi is random, the population size Pi is not.
110 Maps for Rates

who showed that in some instances biased estimators may have better precision in a mean
squared error sense (James and Stein, 1961).
Formally, the mean squared error or MSE is the sum of the variance and the square of the
bias. For an unbiased estimator, the latter term is zero, so then MSE and variance are the
same. The idea of borrowing strength is to trade off a (small) increase in bias for a (large)
reduction in the variance component of the MSE. While the resulting estimator is biased, it
is more precise in a MSE sense. In practice, this means that the chance is much smaller to
be far away from the true value of the parameter.
The implementation of this idea is based on principles of Bayesian statistics, which are
briefly reviewed next.

6.4.2.1 Bayes law

The formal logic behind the idea of smoothing is situated in a Bayesian framework, in which
the distribution of a random variable is updated after observing data.4 The principle behind
this is the so-called Bayes Law, which follows from the decomposition of a joint probability
(or density) of A and B into two conditional probabilities:

P [AB] = P [A|B] × P [B] = P [B|A] × P [A],

where A and B are random events, and | stands for the conditional probability of one event,
given a value for the other. The second equality yields the formal expression of Bayes law as:

P [B|A] × P [A]
P [A|B] = .
P [B]

In most instances in practice, the denominator in this expression can be ignored, and the
equality sign is replaced by a proportionality sign:

P [A|B] ∝ P [B|A] × P [A].

In the context of estimation and inference, the A typically stands for a parameter (or a
set of parameters) and B stands for the data. The general strategy is to update what is
known about the parameter A a priori (reflected in the prior distribution P [A]), after
observing the data B, to yield a posterior distribution, P [A|B], i.e., what is known about
the parameter after observing the data. The link between the prior and posterior distribution
is established through the likelihood, P [B|A]. Using a more conventional notation with π
as the parameter and y as the observations, this gives:5

P [π|y] ∝ P [y|π] × P [π].

For each particular estimation problem, a distribution must be specified for both the prior
and the likelihood. This must be carried out in such a way that a proper posterior distribution
results. Of particular interest are so-called conjugate priors, which result in a closed form
expression for the combination of likelihood and prior distribution. In the context of rate
smoothing, there are a few commonly used priors, such as the Gamma and the Gaussian
4 There are several excellent books and articles on Bayesian statistics, with Gelman et al. (2014) as a

classic reference.
5 Note that in a Bayesian approach, the likelihood is expressed as a probability of the data conditional

upon a value (or distribution) of the parameters. In classical statistics, it is the other way around.
Rate Smoothing 111

(normal) distribution.6 A formal mathematical treatment is beyond the current scope, but it
is useful to get a sense of the intuition behind smoothing approaches.
In essence, it means that the estimate from the data (i.e., the crude rate) is adjusted with
some prior information, such as the reference rate for a larger region (e.g., the state or
country). Unreliable small area estimates are then shrunk toward this reference rate. For
example, if a small area is observed with zero occurrences of an event, does that mean that
the risk is zero as well? Typically, the answer will be no, and the smoothed risk will be
computed by borrowing information from the reference rate.

6.4.2.2 The Empirical Bayes approach

The Empirical Bayes approach is based on the so-called Poisson-Gamma model, where the
observed count of events (O) is viewed as the outcome of a Poisson distribution with a
random intensity (mean), for example expressed as πP . The Bayesian aspect comes in the
form of a prior Gamma distribution for the risk parameter π. This results in a particular
expression for the posterior distribution, which is also Gamma, with mean and variance:
O+α
E[π] = ,
P +β

and
O+α
V ar[π] =
(P + β)2
where α and β are the the shape and scale parameters of the prior (Gamma) distribution.7
In the Empirical Bayes approach, values for α and β are estimated from the actual data.
The smoothed rate is then expressed as a weighted average of the crude rate, say r, and
the prior estimate, say θ. The latter is estimated as a reference rate, typically the overall
statewide average or some other standard.
In essence, the EB technique consists of computing a weighted average between the raw
rate for each small area and the reference rate, with weights proportional to the underlying
population at risk. Simply put, small areas (i.e., with a small population at risk) will tend to
have their rates adjusted considerably, whereas for larger areas the rates will barely change.8
More formally, the EB estimate for the risk in location i is:

πiEB = wi ri + (1 − wi )θ.

In this expression, the weights are:

σ2
wi = ,
(σ 2 + μ/Pi )

with Pi as the population at risk in area i, and μ and σ 2 as the mean and variance of the
prior distribution. 9
6 For an extensive discussion, see, for example, the classic papers by Clayton and Kaldor (1987) and

Marshall (1991).
7 A Gamma(α, β) distribution with shape and scale parameters α and β has mean E[π] = α/β and

variance V ar[π] = α/β 2 .

8 For an extensive technical discussion, see also Anselin et al. (2006a).
9 In the Poisson-Gamma model, this turns out to be the same as w = P /(P + β).
i i i
112 Maps for Rates

Figure 6.9: Crude and EB rate map

In the empirical Bayes approach, the mean μ and variance σ 2 of the prior (which determine
the scale and shape parameters of the Gamma distribution) are estimated from the data.
For μ this estimate is simply the reference rate (the same reference used in the computation
i=n i=n
of the SMR), i=1 Oi / i=1 Pi . The estimate of the variance is a bit more complex:
i=n
2 Pi (ri − μ)2 μ
σ = i=1
i=n − i=n .
i=1 Pi i=1 Pi /n

While easy to calculate, the estimate for the variance can yield negative values. In such
instances, the conventional approach is to set σ 2 to zero. As a result, the weight wi becomes
zero, which in essence equates the smoothed rate estimate to the reference rate.

6.4.3 Empirical Bayes smoothed rate map

In Chapter 5, a unique values map was shown for the cumulative occurrence of Zika in the
state of Ceará in 2016. The map in Figure 5.5 indicated that only 32 of the 184 municipios had
an incidence of Zika during the period under consideration, meaning that 152 municipalities
had no cases. Suggesting that these municipalities therefore had zero risk of exposure to Zika
does not seem realistic. Hence the need to borrow some strength from other information.
A crude rate map that uses the cumulative Zika cases (acu_zik_3q) as the Event Variable
and the population at risk (pop_zikv) as the Base Variable is shown in the left panel
of Figure 6.9 as a quantile map with six categories. Clearly, all but one of the 32 non-zero
observations are grouped in the top category, one is the next lower category and everything
else is zero.
An Empirical Bayes smoothed map is invoked as Maps > Rates-Calculated Maps >
Empirical Bayes in the usual fashion. The same Event Variable and Base Variable
are specified and the map type is selected from the same drop-down list as in Figure 6.3. The
resulting six category quantile map is shown in the right panel of Figure 6.9. The contrast
with the crude rate map is striking. Each category has the required 31 or 32 observations,
and the map shows much more spatial variety than before. Areas that had zero observations
are shrunk toward an average rate, either below or above the median. The highlighted
Rate Smoothing 113

Figure 6.10: Effect of smoothing – small populations

Figure 6.11: Effect of smoothing – large populations

municipalities in the map (blue rectangle) shown three small areas whose smoothed rates
pulled the original zero value above the median.

6.4.3.1 The effect of smoothing

In the Ceará example, the total cumulative number of Zika cases was 2,262, for a total
population of 8,904,459. This yields a reference rate of 0.000254, or 25.4 per 100,000. The
effect of Empirical Bayes smoothing is that the original crude rate is moved toward this
overall rate as a function of its population: areas with large populations do not see much
change, whereas areas with small populations can take on very different smoothed value.
In Figure 6.10, the effect on small areas is illustrated. The municipalities are ordered by
population size, with the 10 smallest areas listed. This clearly illustrates how the crude
rate (R_RAW_RT) gets moved toward the overall mean in the Empirical Bayes rate
(R_EBS). For example, for the smallest area, the crude rate of zero becomes 5.5 per 100,000
and most of the other crude rates get pulled up toward the overall mean. For the municipio
of Altaneira, the opposite happens, i.e., the crude rate of 27.2 per 100,000 becomes 27.0 per
100,000, pulling it down toward the overall mean.
For areas with large populations, there is a minimal effect of the smoothing. Figure 6.11 lists
the ten largest municipalities with their crude and smoothed rates. The difference between
the two is marginal.
114 Maps for Rates

6.4.3.2 Calculating EB smoothed rates in the table

Just as for the other rate maps, the calculated Empirical Bayes smoothed rate can be saved
to the table (the default variable name is R_EBS). In addition, the smoothed rates can
be computed directly by means of the Calculator option in the Table, following the same
procedures as outlined in Section 6.2.2.1.
7
Univariate and Bivariate Data
Exploration

In this and the next two chapters, I continue the review of data exploration, but now shift
the focus to traditional non-spatial EDA methods. Through the use of linking and brushing,
the connection with a spatial representation (a map) can always be made explicit. This idea
of spatializing EDA is central to the perspective taken here.
In the current chapter, I focus on techniques to describe the distribution of one variable
at a time (univariate), and on the relationship between two variables (bivariate) through
standard statistical graphs. These include the histogram, box plot, scatter plot and scatter
plot matrix, considered in turn. The chapter closes with a discussion of spatial heterogeneity,
both for a single variable (through the averages chart) and pertaining to the relationship
between two variables (brushing the scatter plot).
To illustrate the methods covered in this and the next two chapters, I introduce a new
sample data set with poverty indicators and census data for 570 municipalities in the state
of Oaxaca in Mexico. The poverty indicators are from CONEVAL (the National Council for
the Evaluation of Social Development Policy) for 2010 and 2020. The census variables cover
2000, 2010 and 2020 and are from INEGI (National Institute of Statistics and Geography).
The data are contained in the Oaxaca Development sample data set.

7.1 Topics Covered

• Computing descriptive statistics and creating visualizations of the distribution of a single
variable (histogram, box plot)
• Interpreting a scatter plot and scatter plot smoothing (LOWESS)
• Linking and brushing maps and statistical graphs
• Analyzing the bivariate relationship for multiple variables in a scatter plot matrix
• Assessing spatial heterogeneity by means of the Averages Chart
• Assessing spatial heterogeneity through the Chow test
GeoDa Functions
• Explore > Histogram
– Choose Intervals option
– Histogram Classification option
– View > Set as Unique Value
– View > Display Statistics option

DOI: 10.1201/9781003274919-7 115

116 Univariate and Bivariate Data Exploration

• Explore > Box Plot

– Hinge option
• Explore > Scatter Plot
– View > Display Precision option
– Data option
– Smoother option
– LOWESS parameters setting
– Regimes Regression option and Chow test
• Explore > Scatter Plot Matrix
– changing the variable order in a scatter plot matrix
– smoothing and brushing the scatter plot matrix
• Explore > Averages Chart
Toolbar Icons

Figure 7.1: Histogram | Box Plot | Scatter Plot | Scatter Plot Matrix

7.2 Analyzing the Distribution of a Single Variable

The first step in the analysis of the distribution of a single variable is a summary of its
characteristics, focusing on central tendency (mean, median), spread (variance, interquartile
range) and shape (skewness, kurtosis). Here, I briefly review two familiar visualizations of a
univariate distribution in the form of the histogram and the box plot.

7.2.1 Histogram
Arguably the most familiar statistical graphic is the histogram, which is a discrete rep-
resentation of the density function of a continuous variable. In essence, the range of the
variable (the difference between maximum and minimum) is divided into a number of equal
intervals (or bins), and the number of observations that fall within each bin is depicted
proportional to the height of a bar. This classification is the same as the principle underlying
the equal intervals map, which was covered in Section 4.4.2. The main challenge in creating
an effective visualization is to find a compromise between too much detail (many bins,
containing few observations) and too much generalization (few bins, containing a broad
range of observations).

7.2.1.1 Histogram basics

The histogram functionality is started by selecting Explore > Histogram from the menu,
or by clicking on the Histogram toolbar icon, the left-most icon in the univariate set (within
the blue rectangle) in Figure 7.1.
This brings up the Variable Settings dialog, which lists all the numeric variables in the
data set (string variables cannot be analyzed). For example, selecting ppov_20 from the
list (percent of population living in poverty in 2020) creates the default histogram with seven
bins, shown in Figure 7.2. The distribution is highly left-skewed, with a long tail toward the
Analyzing the Distribution of a Single Variable 117

Figure 7.2: Histogram for percent poverty 2020

Figure 7.3: Histogram for percent poverty 2020 – 12 bins

lower percentages, highlighting the prevalence of high-poverty municipalities (on the right).
The histogram options are brought up in the usual fashion, by right-clicking on the graph.
The options are grouped into three categories, the first consisting of histogram-specific items:
Choose Intervals, Histogram Classification and View. Next is a Color option to set
the background color (the default of white is usually best). Finally, the options to Save
Selection, Copy Image to Clipboard and Save Image As work in the same way as
for maps (see Section 4.5.5).
Of these options, the Choose Intervals is the most commonly used. It brings up a dialog
to specify the number of bins for the histogram. For example, after changing from the default
of seven to twelve bins, the histogram becomes as in Figure 7.3. In this particular instance,
there is not much gained by the greater detail since the same overall pattern is maintained.
The Histogram Classification option allows for the selection of a custom category speci-
fication. This works in the same way as for map classification, covered in Section 4.6.
118 Univariate and Bivariate Data Exploration

Figure 7.4: Bar chart for settlement categories

The View option contains six items. Four of these pertain to the overall look of the graph
and are self-explanatory: Set Display Precision (for the variable being depicted), Set
Display Precision on Axes, Show Axes and Status Bar. The latter two are checked
by default.

7.2.1.2 Bar charts

The first item under the View options, Set as Unique Value, is useful when the variable
under consideration takes on integer values, e.g., when it is categorical or ordinal. While
the corresponding graph is not technically a histogram (since categories do not imply any
order), the same mechanics can be used to create a bar chart.
Under the default settings, the bins are computed by taking the range of values into
consideration, i.e., the difference between the largest and smallest integer value. This is then
divided by the number of bins, leading to unrepresentative category labels for integer-valued
variables. This issue is remedied by means of the View > Set as Unique Value option.
As a result, the values taken by the variable of interest are recognized as integer and depicted
as such.
For example, the left panel in Figure 7.4 shows the default histogram for the variable CAT-
MAX20, a categorical variable (or, more precisely, an ordinal variable) that corresponds to
the largest settlement category in each municipality (each municipality consists of one of
more localities or settlements).1
For Oaxaca in 2020, this variable ranges from 1 to 9. As a result, the default of seven
histogram bins results in a bin width of 8/7, or roughly 1.14, which does not yield meaningful
intervals.
By contrast, the right-hand panel in the figure shows the result after invoking Set as Unique
Value. Now, each bar corresponds to a distinct discrete value, matching each category. The
distribution is fairly symmetric around a mode of 4 (1,000 to 2,499 inhabitants), albeit
slightly right-skewed .
1 The original variable is tamloc, which ranges from 1, for 1 to 249 inhabitants, to 14, for more than a

million inhabitants. CATMAX20 is the highest category obtained for a locality in the municipality. In this
example, the maximum is 9, for 30,000 to 49,999 inhabitants.
Analyzing the Distribution of a Single Variable 119

Figure 7.5: Descriptive statistics

7.2.1.3 Descriptive statistics

The View > Display Statistics option brings up a number of descriptive statistics, shown
below the histogram, as in Figure 7.5 for the ppov_20 variable.
The overall summary statistics are listed at the bottom: minimum of 16.9, maximum of
99.6, median of 82.2, mean of 77.3, standard deviation of 18.15, for 570 observations. In
addition, each histogram bar is associated with a specific range, the number of observations
it contains, % of total in that bin and the number of standard deviations from the mean.
For the latter, results greater than two in absolute value can be classified as outliers. In the
example, this is the case for the two lowest categories.

7.2.1.4 Linking map and histogram

The concepts of linking and brushing were already briefly introduced in Chapter 4. They
are most powerful when connecting a-spatial EDA graphs with a map as a visualization of
the spatial distribution of a variable. The map in question can pertain to the same variable
(or be themeless), allowing to visually explore the extent to which observations in the same
histogram categories also occur in similar locations (a pre-cursor to the more formal notion
of spatial autocorrelation). Alternatively, the map can be for a different variable, providing
a way to explore the association between observations in histogram bins to the spatial
distribution of a different variable.
For example, in Figure 7.6, the map on the left is a six category quantile map for ALTID, a
representative altitude for the municipality, in meters.2 When selecting the 73 observations
in the lowest two histogram bins on the right (the largest settlement having respectively 1 to
249 or 250 to 499 inhabitants), the process of linking simultaneous selects the corresponding
2 The Basemap is ESRI > WorldTopoMap. As is to be expected, the darker colors are associated

with higher elevations on the topographical map.

120 Univariate and Bivariate Data Exploration

Figure 7.6: Linking between histogram and map

Figure 7.7: Linking between map and histogram

locations in the elevation map. The resulting configuration does not appear to be random
(our prior expectation, or null hypothesis), but the respective municipalities are concentrated
in the top elevation categories. In other words, settlements with smaller populations tend to
be located in higher elevations (not an unexpected result, but still worthy of checking).
The reverse linking between the map and the histogram is illustrated in Figure 7.7. The
selection rectangle in the map results in the corresponding observations to be highlighted
in the histogram. Here, the focus is on spatial heterogeneity (further elaborated upon in
Section 7.5). Everything else being the same, the expectation (null hypothesis) would be
that the distribution of the selected observations largely follows that of the whole. In other
words, in the histogram, the heights of the bars of the selected observations should roughly
be proportional to the corresponding height in the full data set. In the example, this clearly
is not the case, suggesting a difference between the distribution in the spatial subset and the
overall distribution. In other words, this suggests the presence of spatial heterogeneity.
The process can be made dynamic through brushing, i.e., by moving the selection rectangle
over the map. This results in an immediate adjustment of the selected observations in
the histogram, allowing for an assessment of spatial heterogeneity by eye. A more formal
assessment is pursued in Section 7.5.
Analyzing the Distribution of a Single Variable 121

Figure 7.8: Box plot for population change 2020–2010

7.2.2 Box plot

The box plot is an alternative visualization of the distribution of a single variable that
focuses on quartiles (see also Section 5.2.2 for a discussion of the associated box map). The
observations are ranked from lowest to highest and (optionally) represented as dots on a
vertical or horizontal line.3 The box consists of a rectangle, with the lower bound drawn at
the first quartile (25%) and the higher bound at the third quartile (75%). Typically, a line is
also drawn at the location of the median (50%).
The box plot introduces the notion of outliers, i.e., observations that are far from the central
tendency (the median), suggesting they may not belong to the same distribution as the rest.
The key concept in this regard is that of the inter-quartile range or IQR, the difference
between the values at the third and first quartile. This is a measure of the spread of the
distribution around the median, a counterpart to the variance. The IQR is used to compute
a hinge (or fence), i.e., a value that is 1.5 times (or, alternatively, 3 times) the IQR above
the third quartile or below the first quartile. Observations that lie outside the hinges are
designated as outliers.

7.2.2.1 Implementation
The box map is invoked as Explore > Box Plot from the menu, or by selecting the Box
Plot as the second icon from the left in the toolbar in Figure 7.1. Identical to the approach
followed for the histogram, next appears a Variable Settings dialog to select the variable.4
To illustrate this graph, we select the variable c_ptot12, the percentage population change
between 2020 and 2010 (positive values are population growth). The default box plot, with
3 GeoDa takes the vertical approach.
4 In GeoDa, the default is that the variable from any previous analysis is already selected.
122 Univariate and Bivariate Data Exploration

Figure 7.9: Box map for population change 2020–2010 (upper outliers selected)

a hinge of 1.5, is shown in the left-hand panel of Figure 7.8. The descriptive statistics are
listed at the bottom.5
The observations range from −31.2% to 113.3%, with a slightly positive median of 3.0%
(the mean is 4.0%, reflecting the influence of upper outliers). The interquartile range is
13.9. Consequently, the upper hinge is roughly 10.7 (Q3) + 1.5 × 13.9, or 37.9%. Fifteen
observations take on values that are larger than this upper hinge and are designated as
upper outliers. The lower hinge is −3.2 − 1.5 × 13.9, or −21.1%. Ten observations have
population decreases that are even larger, hence these are lower outliers.

7.2.2.2 Box plot and box map

Selections in the box plot can be linked to a map, and vice versa. However, to a large extent
this is already accomplished by means of the box map, discussed in Section 5.2.2.
The interesting research question is the extent to which a-spatial outlying observations, such
as the fifteen outliers identified in the box plot in Figure 7.8 show a spatial pattern that
may suggest some structure, rather than randomness. The box map for c_ptot12 in Figure
7.9 shows the outlying observations highlighted in dark red (selected in the box plot).6 The
map indicates that about ten of the outliers are located in the center of the country (inside
the blue rectangle). The research question is then whether this is purely due to chance, or,
instead, suggests a pattern of clustering. This is pursued more formally in Parts IV and V.

7.2.2.3 Box plot options

Several of the box plot options are shared with the other graphs (and maps). There are
seven categories. The last three are Save Selection, Copy Image To Clipboard and
Save Image As, which work as before. The Show Status Bar is listed as a main option
(checked by default) and the Color option also provides a way to change the color of the
points (in addition to the background color).
5 For the box plot, this is the default. It can be turned off by unchecking View > Display Statistics in

the options.
6 The basemap is Stamen > TonerLite.
Bivariate Analysis – The Scatter Plot 123

The View option contains four items, all self-explanatory: Set Display Precision, Set
Display Precision on Axes, Display Statistics and Show Vertical Axis.
The typical multiplier for the IQR to determine outliers is 1.5 (roughly equivalent to the
practice of using two standard deviations in a parametric setting). However, a value of 3.0
is fairly common as well, which considers only truly extreme observations as outliers. The
multiplier to determine the fence can be changed with the Hinge > 3.0 option (obtained
by right clicking in the plot to select the options menu, and then choosing the hinge value).
This yields the box plot shown in the right-hand panel of Figure 7.8. The new hinge no
longer yields lower outliers, and the upper outliers are reduced to five extreme observations.
The main purpose of the box plot in an exploratory strategy is to identify outlier observations
in an a-spatial sense. Its spatial counterpart is the box map.

7.3 Bivariate Analysis – The Scatter Plot

The most commonly used tool to assess the relationship between two variables is the scatter
plot, a diagram with two orthogonal axes, each corresponding to one of the variables. The
observation (X, Y) coordinate pairs are plotted as points in the diagram.
Typically, the relationship between the two variables is summarized by means of a linear
regression fit through the points. The regression is:

y = a + bx + ,

where a is the intercept, b is the slope and  is a random error term. The coefficients
are estimated by minimizing the sum of squared residuals, a so-called least squares fit, or
ordinary least squares (OLS).
The intercept a is the average of the dependent variable (y) when the explanatory variable
(x) is zero. The slope shows how much the dependent variable changes on average (Δy) for
a one unit change in the explanatory variable (Δx). It is important to keep in mind that
the regression applied to the scatter plot pertains to a linear relationship. It may not be
appropriate when the variables are non-linearly related (e.g., a U shape).
When y and x are standardized (mean zero and variance one), then the slope of the regression
line is the same as the correlation between the two variables (the intercept is zero). Note
that while correlation is a symmetric relationship, regression is not, and the slope will be
different when the explanatory variable takes on the role of dependent variable, or vice
versa. Just as a linear regression fit may not be appropriate when the variables are related
in a non-linear way, so is the correlation coefficient. It only measures a linear relationship
between two variables, which is clearly demonstrated by its equivalence to the regression
slope between standardized variables.

7.3.1 Scatter plot basics

A scatter plot is created by selecting Explore > Scatter Plot from the menu, or by
clicking on the matching toolbar icon, the third from the left in Figure 7.1. This brings up
the Scatter Plot Variables dialog, where two variables need to be selected: one for the X
axis (Independent Var X) and one for the Y axis (Dependent Var Y).
124 Univariate and Bivariate Data Exploration

Figure 7.10: Scatter plot of food insecurity on poverty

To illustrate this graph, the X variable is ppov_20 (percent population living in poverty in
2020) and the Y variable pfood_20 (percent population with food insecurity in 2020). The
resulting scatter plot, with all the default settings, is as in Figure 7.10.
The top of the graph spells out the X and Y variables, which are also listed along the axes.
The linear regression fit is drawn over the points, with summary statistics listed below the
graph. These included the fit (R2 ) and the estimates for constant and slope, with estimated
standard error, t-statistic and associated p-value. In the example, the fit is a low 0.091,
which is not surprising, given the cone shape of the scatter plot, rather than the ideal cigar
shape. Nevertheless, both intercept and slope coefficients are highly significant (rejecting the
null hypothesis that they are zero) with p-values of 0.002 and 0.000 respectively.
The positive relationship suggests that as poverty goes up by one percent, the share in food
insecurity increases by 0.24 percent.
In the current setup, no observations are selected, so that the second line in the statistical
summary in Figure 7.10 (all red zeros) has no values. This line pertains to the selected
observations. The blue line at the bottom relates to the unselected observation. The sum
of the number of observations in each of the two subsets always equals the total number
of observations, listed on the top line. The three lines are included because of the default
View setting of Regimes Regression, even though there is currently no active selection
(see Section 7.3.1.1).
The scatter plot has several options, invoked in the customary fashion by right clicking on
the graph. They are grouped into eight categories:
Bivariate Analysis – The Scatter Plot 125

Figure 7.11: Regime regression

• Selection Shape
• Data
• Smoother
• View
• Color
• Save Selection
• Copy Image to Clipboard
• Save Image As
Selection Shape works as in the other graphs and maps (see Section 2.5.4), and so do
the three last options. The Color options allow for the Regression Line Color, Point
Color and Background Color to be set. The remaining options are discussed in more
detail below.

7.3.1.1 View options

The View option manages seven settings, all but the first two checked by default:
• Set Display Precision
• Set Display Precision on Axes
• Fixed Aspect Ratio Mode
• Regimes Regression
• Statistics
• Axes Through Origin
• Status Bar
126 Univariate and Bivariate Data Exploration

Figure 7.12: Correlation

These options are mostly self-explanatory, or work in the same way as before (such as
Statistics), except for the Regimes Regression.
The latter allows for a dynamic assessment of structural stability by comparing the regression
line for a selected subset of the data to its complement (unselected). For example, in Figure
7.11, 206 observations have been selected (inside the selection rectangle). This yields three
regression lines: a purple one for the full data set (same results as in Figure 7.10), a red one
for the selected observations and a blue one for its complement (384 observations). With
each regression line corresponds a line in the statistics summary. In Figure 7.11, the selected
observations show no significant linear relationship between the two variables. In other words,
for a range of poverty roughly between 30% and 70%, there is no linear relationship with
food insecurity (p-value of 0.542), whereas for the whole and for the complement, there is.
This is an indication of structural instability.
A formal assessment of structural stability is based on the Chow test (Chow, 1960). This
statistic is computed from a comparison of the fit in the overall regression to the combination
of the fits of the separate regressions, while taking into account the number of regressors (k).
In our simple example, there is only the intercept and the slope, so k = 2. The residual sum
of squares can be computed for the full regression (RSS) and for the two subsets, say RSS1
and RSS2 . Then, the Chow test follows as:
(RSS − (RSS1 + RSS2 ))/k
C= ,
(RSS1 + RSS2 )/(n − 2k)
distributed as an F statistic with k and n − 2k degrees of freedom. Alternatively, the statistic
can also be expressed as having a chi-squared distribution, which is more appropriate when
Bivariate Analysis – The Scatter Plot 127

Figure 7.13: Default LOWESS local regression fit

multiple breaks and several coefficients are considered.7 In our example, the p-value of the
Chow test is 0.039, which is only weakly significant (in part this is due to the fact that the
overall regression coefficient of 0.243 is not that different from zero itself). The assessment
of structural stability is considered further in the context of spatial heterogeneity in Section
7.5.2.
The Regimes Regression option is on by default. Turning it off disables the effect of any
selection and reduces the Statistics to a single line.

7.3.1.2 Data options

The default setting for the Data option is to View Original Data. In other words, the
points in the scatter plot correspond to the respective entries for the variables in the
data table. With this option turned to View Standardized Data, each variable is used in
standardized form. As a result, the axes can be interpreted as expressing standard deviational
units. Values larger than two in absolute value indicate potential outliers.
The slope of the regression line applied to the standardized variables corresponds to the
correlation coefficient. The intercept will be zero.
In Figure 7.12, some outliers can be detected at the lower end for ppov_20, but not at
the upper end. For, pfood_20, the reverse is the case, with several upper outliers, but no
lower outliers. The correlation coefficient is 0.301.
7 Technical details can be found in in Anselin and Rey (2014), pp. 287–289.
128 Univariate and Bivariate Data Exploration

Figure 7.14: Default LOWESS local regression fit with bandwidth 0.6

7.3.2 Smoother option – local regression

The default fit to the scatter plot points is a global linear regression. This can be turned off
by means of the Smoother > Show Linear Smoother option.
When the relationship between the two variables in the scatter plot is not linear, or shows
complex sub-patterns in the data, a linear fit may not be appropriate. An alternative
consists of a so-called local regression fit. A local regression fit is a nonlinear procedure
that computes the slope from a subset of observations on a small interval on the X-axis.
In addition, values further removed from the center of the interval may be weighted down
(similar to a kernel smoother). As the interval (the bandwidth) moves over the full range of
observations, the different slopes are combined into a smooth curve (for a detailed overview
of the methodological issues, see, e.g., Cleveland, 1979; and Loader, 1999; Loader, 2004).
A local regression fit reveals potential nonlinearities in the bivariate relationship or may
suggest the presence of structural breaks. Two common implementations are LOWESS,
locally weighted scatter plot smoother, and LOESS, a local polynomial regression. The two
are often confused, but implement different fitting algorithms.
The option Smoother > Show LOWESS Smoother calculates a local regression fit
based on the LOWESS algorithm. Continuing with the same two variables, the result is as
in Figure 7.13, using the default settings.
The smoothness of the local fit depends on a few parameters, considered next. Note that the
Regimes Regression setting has no effect on the LOWESS fit. Also, in Figure 7.13, the
Bivariate Analysis – The Scatter Plot 129

Figure 7.15: Default LOWESS local regression fit with bandwidth 0.05

linear fit has been turned off. However, in some instances, it may be insightful to have both
fits showing in the scatter plot.
In the example, the curve shows a fairly regular, near linear increase up to a poverty level of
about 60%, but a much more irregular pattern at higher poverty rates, with a very steep
positive relationship at the very high end. In the middle range (the same range of ppov_20
as selected in Figure 7.11), the curve is almost flat, in rough agreement with what was found
in the regime regression example.

7.3.2.1 LOWESS parameters

The nonlinear fit is driven by a number of parameters, the most important of which is the
bandwidth, i.e., the range of values on the X-axis on which the local slope estimate is based.
The smoothing parameters can be changed in the options by selecting Edit LOWESS
Parameters in the Smoother option. There are three parameters: Bandwidth (default
setting 0.20), Iterations and Delta Factor.
The bandwidth determines the smoothness of the curve and is given as a fraction of the
total range in X values. In other words, the default bandwidth of 0.20 implies that for each
local fit (centered on a value for X), about one fifth of the range of X-values is taken into
account. The other options are technical and are best left to their default settings.8
8 The LOWESS algorithm is complex and uses a weighted local polynomial fit. The Iterations setting

determines how many times the fit is adjusted by refining the weights. A smaller value for this option will
speed up computation, but result in a less robust fit. The Delta Factor drops points from the calculation of
the local fit if they are too close (within Delta) to speed up the computations. Technical details are covered
in Cleveland (1979).
130 Univariate and Bivariate Data Exploration

A higher value of the bandwidth results in a smoother curve. For example, with the bandwidth
set to 0.6, the curve in Figure 7.14 results. This more clearly suggests different patterns
for three subsets of poverty rates: a slowly increasing slope at the lower end, a near flat
slope in the middle and a much steeper slope at the upper end. In a data exploration, one
would be interested in finding out whether these a-spatial subsets have an interesting spatial
counterpart, for example, through linking with a map.
The opposite effect is obtained when the bandwidth is made smaller. For example, with a
value of 0.05, the resulting curve in Figure 7.15 is much more jagged and less informative.
The literature contains many discussions of the notion of an optimal bandwidth, but in
practice a trial and error approach is often more effective. In any case, a value for the
bandwidth that follows one of these rules of thumb can be entered in the dialog. Currently,
GeoDa does not compute optimal bandwidth values.
Finally, while one might expect the LOWESS fit and a linear fit to coincide with a bandwidth
of 1.0, this is not the case. The LOWESS fit will be near-linear, but slightly different from a
standard least squares result due to the locally weighted nature of the algorithm.

7.4 Scatter Plot Matrix

A scatter plot matrix visualizes the bivariate relationships among several pairs of variables.
The individual scatter plots are stacked such that each variable is in turn on the X-axis and
on the Y-axis.
When applied to standardized variables (with mean zero and variance one), it is the visual
counterpart of a correlation matrix. In general, however, while the scatter plot matrix shows
the linear association between variables, the graphs are not symmetric. In addition, in
some instances a variable may not be appropriate to be both a dependent (Y-axis) and
an explanatory variable (X-axis). For example, a variable such as elevation may affect
socio-economic outcomes, but it is hard to imagine that socio-economic factors could in turn
affect elevation.
The main interest is in the magnitude and sign of the slope in each of the scatter plots,
and the extent to which this points to a significant bivariate relationship, similar to the
insight provided by a correlation matrix. In GeoDa, the diagonal elements also contain
a histogram for the variable in the corresponding row/column to provide a sense of the
univariate distribution.
Even though this method deals with multiple variables, it is not really a multivariate
exploration. Rather, it consists of a collection of bivariate explorations for multiple variables.

7.4.1 Implementation
The scatter plot matrix operation is started by selecting the fourth icon on the EDA toolbar
(Figure 7.1), or by choosing Explore > Scatter Plot Matrix from the menu.
This brings up a Scatter Plot Matrix Variables Add/Remove dialog, through which
the variables are selected. The design of the interface is such that one selects a variable from
the Variables list on the left and clicks on the right arrow > to include it in the Include
Scatter Plot Matrix 131

Figure 7.16: Scatter plot matrix

list on the right. Alternatively, one can double click on the variable name to move it to the
right-hand column. The left arrow < is there to remove a variable from the Include list.
As soon as two variables are selected, the scatter plot matrix is rendered in the background.
As new variables are added to the Include list, the matrix in the background is updated
with the additional scatter plots.
Figure 7.16 shows an example with four variables that were already used in the illustrations
above: ppov_20, pfood_20, c_ptot12 and ALTID. Note that the latter really should
not be portrayed on the Y-axis, although by construction this always happens for every
variable in the scatter plot matrix.
With the default setting, each scatter plot shows the slope coefficient at the top, together with
a designation of significance (** is p < 0.01, * is p < 0.05). Ignoring the bottom row (with
ALTID on the Y-axis), significant coefficients are obtained in all instances, except between
pfood_20 and c_ptot12. In other words, food insecurity and population change do not
seem to be related. On the other hand, there is a strong and negative relationship between
ppov_20 and c_ptot12. Altitude is significantly related with all three socio-economic
outcomes, positively with ppov_20 (greater poverty at higher altitudes), but negatively
with the other two. A substantive interpretation is beyond the scope.

7.4.1.1 Scatter plot matrix options

The scatter plot matrix options consist of seven items:
• Add/Remove Variables
• Selection Shape
132 Univariate and Bivariate Data Exploration

Figure 7.17: Scatter plot correlation matrix

• Data
• Smoother
• View
• Color
• Save Image As
The first option invokes the Scatter Plot Matrix Variables Add/Remove dialog to
allow for changes in the variable selection. Selection Shape and Save Image As work in
the usual fashion.
View has options to Set Display Precision on Axes, Regimes Regression and Dis-
play Slope Values. The latter is checked by default. In contrast to the standalone scatter
plot, Regimes Regression is turned off by default in the scatter plot matrix. However,
it can readily be invoked, allowing for all features of linking and brushing, e.g., to explore
spatial heterogeneity across all variables (see Section 7.5.2).
The Color option provides a way to change the Regression Line Color and Point Color.
Data and Smoother are separately considered next.

7.4.1.2 Correlation matrix

The Data option contains the same two items as for the standalone scatter plot. With Data
> View Standardized Data, the scatter plot matrix turns into a correlation matrix. As
pointed out, in contrast to the regression slope, the correlation coefficient is fully symmetric
and the results above the diagonal are the mirror image of the results below the diagonal.
On the graph, they may look slightly different due to the use of different scales on the axes,
but the coefficients are identical.
Scatter Plot Matrix 133

Figure 7.18: Scatter plot matrix with LOWESS fit

The result for the four variables considered before is as in Figure 7.17.

7.4.1.3 LOWESS scatter plot matrix

Finally, the Smoother option allows for local regression, as in the standalone scatter plot,
but now applied to all bivariate relationships. This provides a quick overview of the extent to
which these relationships are linear, or instead show local instability and nonlinear patterns.
The LOWESS parameters can be edited in the same way as before.
The result of a LOWESS fit for a bandwidth of 0.6 is illustrated in Figure 7.18. In order to
make sure that the regression coefficient estimates are no longer shown, the option Display
Slope Values must be turned off (it remains on by default). Also, in the figure, the linear
fits are not shown.
The result shows some interesting patterns, especially for the relationship between altitude
and respectively poverty and food insecurity. Instead of the strong overall positive and
negative slopes, the evidence is much more nuanced. In fact, especially for the altitude-food
insecurity graph, the negative slope for lower altitudes turns into a positive slope for the
higher altitudes. Such patterns cannot be distinguished in a purely linear approach.
134 Univariate and Bivariate Data Exploration

Figure 7.19: Food insecurity 2020 in Valles Centrales

7.5 Spatial Heterogeneity

Together with spatial dependence, spatial heterogeneity is the other critical aspect in any
spatial data analysis. In essence, its presence suggests that more than one underlying
distribution may be responsible for generating the observed sample. This is evidenced in the
form of structural breaks, such as different mean or median values in different subsets of
the data, or different slopes in a bivariate regression over observations in those subsets (see
Anselin, 1988, for a more formal treatment).
The spatial aspect in the heterogeneity comes from the nature of the subsets in the data,
which are spatially defined. Examples are the difference between center and periphery, east
and west and north and south. In an exploratory data analysis, spatial heterogeneity can be
assessed by selection in a map linked to a statistical graph. An application of this idea was
illustrated in the context of a linked map and histogram (Section 7.2.1.4). Next, two further
implementations of this approach are considered. One is the analysis of differences in means
through the Averages Chart, the other the investigation of structural stability in a scatter
plot by means of brushing a map linked to the scatter plot.

7.5.1 Averages chart

The Averages Chart is an implementation of a simple test on the difference in means
between selected and unselected observations. Its most meaningful use is in the context of
observations at different points in time (see Chapter 9), but it is equally applicable in a
cross-sectional setting, illustrated here.
The core functionality of this chart is to illustrate and quantify the difference between the
mean of a variable for selected observations and unselected observations (the complement).
In GeoDa, this is not implemented as a traditional t-test, but rather as an F-statistic for a
regression that includes an indicator variable for the selection (i.e., value = 1 for selected,
and zero otherwise). The F-statistic on the significance of the joint slopes in that regression
Spatial Heterogeneity 135

Figure 7.20: Averages Chart – Valles Centrales

is equivalent to a t-test on the coefficient of the indicator variable, since there is only one
slope.
This F-statistic is basically a test on whether there is a significant gain in explanation in the
regression beyond the overall mean (i.e., the constant term). Formally, the statistic uses the
sum of squared residuals in the regression RSS and the sum of squared deviations from the
mean for the dependent variable RSY. The statistic follows as:
RSS − RSY RSY
F = / ,
k−1 n−k
with k as the number of explanatory variables. In our simple dummy variable regression,
k = 2, so that the degrees of freedom for the F-statistic are 1, n − 2 (see also Anselin and
Rey, 2014, pp. 98–99).

7.5.1.1 Selected and unselected observations

The averages chart is invoked from the menu as Explore > Averages Chart. However, its
toolbar icon is not part of the EDA group, but instead is included as the right-most icon in
the time management toolbar in Figure 9.1. Its most effective application is in a space-time
context, but here its use is illustrated in a cross-sectional setup.
The example explores the difference in mean food insecurity in 2020 between the Oaxaca
Central Valleys and the rest of the country. The so-called Valles Centrales are the location of
the original Zapotec civilization and are still characterized by a large indigenous population.
They also contain the state capital of Oaxaca.
The overall spatial distribution of pfood_20 is shown in the box map in Figure 7.19, with
the central valley municipalities highlighted. The selection is obtained by setting region =
8 in the selection tool (see Section 2.5.1).
The selection is used in the averages chart to define selected and unselected as the
Groups. In Figure 7.20, the Variable is specified as pfood_20. In the current context,
136 Univariate and Bivariate Data Exploration

Figure 7.21: Map brushing and the averages chart – 1

Figure 7.22: Map brushing and the averages chart – 2

the time settings can be ignored. The statistics for the two groups are listed in a small table.
Selected has 121 observations with a Mean of 17.65 and S.D. of 7.74. In contrast, the
Unselected group contains 449 observations, with Mean of 29.09 and S.D. of 15.07.
The formal test on equality of means results in an F-statistic of 64.99, which yields a very
small p-value (essentially zero) for 568 degrees of freedom.
In the right-hand panel of the window, the overall mean (black), selected mean (red) and
unselected mean (blue) are represented graphically.
In this case, there is strong evidence that food insecurity is less severe in the Central Valleys
compared to the rest of the state.

7.5.1.2 Map brushing and the averages chart

A more comprehensive exploration of spatial heterogeneity is achieved by brushing a map
linked to the averages chart. As the selection in the map changes, the statistics in the
averages chart are updated. As a result, one can assess the extent to which subregions of
the data have a different mean for the variable under consideration.
For example, Figure 7.21 has a six category quantile map for elevation on the left-hand side.
As the brush is moved from west to east, the effect on the test of means can be assessed in
Spatial Heterogeneity 137

Figure 7.23: Map brushing and the averages chart – 3

the averages chart on the right. Note that the brushing is not based on the altitude variable,
but is purely a move over space. However, by including a different variable, one may be able
to discover what lies behind the spatial structural differences.
The variable under consideration in the averages chart is ppov_20. The first selection
contains 213 observations with a mean of 75.42, compared to a mean of 78.46 for the rest.
The test on the difference between the means is not significant at p = 0.052. Thus this initial
selection of higher elevation locations is not substantially different from the rest of the state.
In Figure 7.22, the selection is moved to the center of the state, with 215 selected observations,
yielding a mean of 73.69. The unselected mean is 79.53. For this selection, the test rejects
the null hypothesis with a p-value of 0.000, suggesting strong spatial heterogeneity.
Finally, in Figure 7.23, the selection is moved even further to the east. The 87 selected
observations have a mean of 79.01, whereas the complement has a mean of 77.02. The test
on difference between the means is not significant at p = 0.348.
By moving the brush over different regions of the map, the extent of spatial structural
instability can be assessed. In addition, by using a map for a different variable (such as
altitude here), one can possibly gain insight into factors that may be behind the spatial
structural instability. However, for this to be meaningful, one has to make sure that sufficient
observations are contained in each selection.

7.5.1.3 Averages chart options

The default setting for the averages chart is to have a Fixed scale over time. This means
that the minimum and maximum tick marks on the left side axis remain the same as the
selection changes. In some instances, the respective means may no longer fall within this
range. As a result, they will not be shown in the graph. To remedy this, the Axis Option
can be set to Enable User Defined Value Range of Y-Axis, through which the range
can be customized.
A final option is to Set Display Precision of Y-Axis.
138 Univariate and Bivariate Data Exploration

Figure 7.24: Map brushing and the scatter plot – 1

Figure 7.25: Map brushing and the scatter plot – 2

7.5.2 Brushing the scatter plot

7.5.2.1 Scatter plot brushing
The concept of scatter plot brushing was initially suggested by Stuetzle (1987), and extended
to the map context by Monmonier (1989). The idea is to dynamically adjust the selection of
observations in the scatter plot within a selection brush (typically a rectangle). As the brush
moves over the plot, observations are added to and dropped from the selection and the slope
of the linear fit is adjusted. As discussed in Section 7.3.1.1, this is implemented in GeoDa
through the Regimes Regression option (see also Figure 7.11).
Since all windows are simultaneously linked in the basic architecture of the GeoDa software
(Anselin et al., 2006b), the dynamic selection through brushing can be initiated in any
open window. This results in an immediate adjustment of the slope of the linear fit in the
scatter plot and an updated computation of the Chow test. As in the averages chart, spatial
heterogeneity can be assessed by initiating a spatial selection through the brushing operation
in a map.
Spatial Heterogeneity 139

Figure 7.26: Map brushing and the scatter plot – 3

7.5.2.2 Map brushing and scatter plot

To illustrate the dynamic map brushing operation, a scatter plot of c_ptot12 on ppov_20
is considered jointly with the elevation map used before. The global linear fit yields a slope
coefficient estimate of −0.272, with an R2 of 0.121. The brush moves from west to east.
In Figure 7.24, the brush is initiated by selecting 93 observations in the higher altitude
region in the western part of the state. The regression slope of the selected observations is
−0.217, compared to −0.303 for the unselected, yielding a significant Chow test result at
p = 0.0012. Note that the coefficient for the selected observations is not significant (p =
0.128).
As the brush is moved to the right, the selection is adjusted. In Figure 7.25, the slope of
the 77 selected observations has changed to −0.390, compared to −0.266 for the unselected.
Even though the estimates are both significant, they cannot be deemed to be different, with
the Chow test yielding a p-value of 0.387.
In Figure 7.26, the brush is moved even further to the east, yielding a selection of 33
observations. Note that since the brush size is fixed, but the spatial extent of the municipalities
varies, the number of observations contained in each spatial selection will not be constant.
The municipalities in this part of the state tend to be larger, resulting in fewer observations
in the selection window. This will affect the precision of the estimates in the subset, and,
indirectly, also the Chow test statistic.
At this point, the selected observations show no significant relationship, which a slightly
positive coefficient of 0.056 (p-value of 0.335). In contrast, the coefficient in the unselected
observations is negative at −0.307, and highly significant.
A careful assessment of the effect of different spatial selections can provide insight into
spatially defined structural instabilities in the relationship between any two variables. The
inclusion of a Chow test provides some guidance, even though its results need to be interpreted
with caution due to the problem of multiple comparisons (many tests carried out with the
same data). However, it allows for a more quantitative measure of the spatial heterogeneity,
instead of relying on a purely visual assessment, which can be misleading. This is in the
spirit of the more recently developed perspectives on EDA, e.g., as discussed in Section 4.2.1.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
8
Multivariate Data Exploration

In this chapter, I switch to a full multivariate exploration of spatial data, where the focus
is on the potential interaction among multiple variables. This is distinct from a univariate
(or even bivariate) analysis of multiple variables, where the properties of a distribution
are considered in isolation. The goal is to discover potential pathways of interaction. For
example, in a bivariate analysis, one may have found a strong correlation between lung
cancer and socio-economic factors, but after controlling for smoking behavior (itself strongly
correlated with SES), this relationship disappears.
The methods considered share the same objective, i.e., how to represent relationships in
higher dimensions on a two-dimensional screen (or piece of paper). Three techniques, the
bubble chart, three-dimensional scatter plot and conditional plots are limited to the analysis
of three to four variables. True multivariate analyses for several variables (more than four)
can be carried out by means of the parallel coordinate plot (PCP). Each is considered in
turn. I continue to use the Oaxaca Development data set for illustrations.
As in the previous chapter, the methods covered are inherently non-spatial, but by means of
linking and brushing with one or more maps, they can be spatialized.
Before focusing on the specific techniques, some special features of multivariate analysis are
outlined in a brief discussion of the curse of dimensionality.

8.1 Topics Covered

• Illustrate the curse of dimensionality
• Interpreting a bubble chart
• Interpreting and manipulating a 3D scatter plot
• Interpreting and manipulating conditional plots
• Interpreting a parallel coordinate plot
GeoDa Functions
• Explore > Bubble Chart
– bubble chart classification schemes
– bubble size
• Explore > 3D Scatter Plot
– rotating and zooming the 3D scatter plot
– projecting onto one axis
– selection in the 3D scatter plot
• Explore > Conditional Plot
– conditional scatter plot

DOI: 10.1201/9781003274919-8 141

142 Multivariate Data Exploration

– conditional histogram
– conditional box plot
– conditional map
– conditional plot option
– changing the condition breakpoints
• Explore > Parallel Coordinate Plot
– changing the classification theme for the PCP
– changing the order of the axes
– brushing the PCP
Toolbar Icons

Figure 8.1: Bubble Chart | 3D Scatter Plot | Parallel Coordinate Plot | Conditional Plot

8.2 The Curse of Dimensionality

Before considering specific techniques, it is useful to focus attention on the so-called curse
of dimensionality, a concept initially conceived by Bellman (1961). In a nutshell, the curse
of dimensionality means that techniques that work well in small dimensions (e.g., k-nearest
neighbor searches), either break down or become too complex or computationally impractical
in higher dimensions.
In the original discussion by Bellman (1961), the problem was situated in the context of
optimization. To illustrate this, consider the simple example of the search for the maximum
of a function in one variable dimension, as in Figure 8.2.
One simple way to try to find the maximum is to evaluate the function at a number of
discrete values for x and find the point(s) where the function value is highest. For example,
in Figure 8.2, there are 10 evaluation points, equally dividing the x-range of 0–10. Here, the
maximum is between 5 and 6. In a naive one-dimensional numerical optimization routine,
the interval between 5 and 6 could then be further divided into ten more sub-intervals and
so on. However, what happens if the optimization was with respect to two variables? At
this stage, the attribute domain becomes two-dimensional. Using the same approach and
dividing the range for each dimension into 10 discrete values now results in 100 evaluation
points, as shown in Figure 8.3.
Similarly, for three dimensions, there would be 103 or 1,000 evaluation points, as in Figure
8.4.
In general, and keeping the principle of 10 discrete values per dimension, the required number
of evaluation points is 10k , with k as the number of variables. This very quickly becomes
unwieldy, e.g., for 10 dimensions, there would need to be 1010 , or 10 billion evaluation
points, and for 20 dimensions 1020 or 100,000,000,000,000,000,000, hence the curse of
dimensionality. Some further illustrations of methods that break down or become impractical
in high-dimensional spaces are given in Chapter 2 of Hastie et al. (2009).
The Curse of Dimensionality 143

Figure 8.2: One dimension Figure 8.3: Two dimensions

Figure 8.4: Discrete evaluation points in three variable dimensions

8.2.1 The empty space phenomenon

A related and strange property of high-dimensional analysis is the so-called empty space
phenomenon. In essence, the same number of data points are much farther apart in higher
dimensional spaces than in lower dimensional ones. This means that in order to carry out
effective analysis in high dimensions, (very) large data sets are needed.
To illustrate this point, consider three variables, say x, y and z, each generated as 10 uniform
random variates in the unit interval (0–1). In one dimension, the observations on these
variables are fairly close together, since their range will be less than one by construction.
However, when represented in two dimensions, say in the form of a scatter plot, large empty
spaces appear in the unit quadrant, as in Figure 8.5 for x and y (similar graphs are obtained
for the other pairs). The sparsity of the attribute space gets even more pronounced in three
dimensions, as in Figure 8.6.
To quantify this phenomenon, consider the nearest neighbor distances between the observation
points in each of the dimensions.1 In the example (the results will depend on the actual
1 In GeoDa, this can be accomplished in the weights manager, using the variables option for distance

weights, and making sure the transformation is set to raw. The nearest neighbor distances are the third
column in a weights file created for k = 1. See Chapter 11 for a detailed discussion.
144 Multivariate Data Exploration

Figure 8.5: Two dimensions

Figure 8.6: Three dimensions

random numbers generated), the smallest nearest neighbor distance in the x dimension is
0.009. In two dimensions, between x and y, the smallest distance is 0.094, and in three
dimensions, it is 0.146. All distances are in the same units, since the observations fall within
the unit interval for each variable.
Two insights follow from this small experiment. One is that the nearest neighbors in a lower
dimension are not necessarily also nearest neighbors in higher dimensions. For example, in
the x dimension, the shortest distance is between points 5 and 8, whereas in the x-y space,
it is between 1 and 4. The other property is that the nearest neighbor distance increases
with the dimensionality. In other words, more of the attribute space needs to be searched
before neighbors are found. This becomes a critical issue for searches in high dimensions,
where most lower dimensional techniques (such as k-nearest neighbors) become impractical.
Further examples of some strange aspects of data in higher dimensional spaces can be found
in Chapter 1 of Lee and Verleysen (2007).

8.3 Three Variables: Bubble Chart and 3-D Scatter Plot

When the number of dimensions is at most three, it remains relatively easy to visualize the
multivariate attribute space. The bubble chart implements this by augmenting a standard
two-dimensional scatter plot with an additional attribute for each point, i.e., the size
of the point or bubble. More directly, observations can also be visualized as points in a
three-dimensional scatter plot or data cube.

8.3.1 Bubble chart

The bubble chart is an extension of the scatter plot to include a third and possibly a fourth
variable into the two-dimensional chart. While the points in the scatter plot still show the
association between the two variables on the axes, the size of the points, the bubble, is used
to introduce a third variable. In addition, the color shading of the points can be used to
consider a fourth variable as well, although this may stretch one’s perceptual abilities.
Three Variables: Bubble Chart and 3-D Scatter Plot 145

Figure 8.7: Bubble chart – default settings: education, basic services, extreme poverty

The introduction of the extra variable allows for the exploration of interaction. The point of
departure (null hypothesis) is that there is no interaction. In the plot, this would be reflected
by a seeming random distribution of the bubble sizes among the scatter plot points. On the
other hand, if larger (or smaller) bubbles tend to be systematically located in particular
subareas of the plot, this may suggest an interaction. Through the use of linking and brushing
with the map, the extent to which systematic variation in the bubbles corresponds with
particular spatial patterns can be readily investigated, in the same way as illustrated in the
previous chapter.
The bubble chart was popularized through the well-known Gapminder web site, where
it is used extensively, especially to show changes over time.2 This aspect is currently not
implemented in GeoDa.

8.3.1.1 Implementation
The bubble chart is invoked from the menu as Explore > Bubble Chart, and from the
toolbar by selecting the left-most icon in the multivariate group of the EDA functionality,
shown in Figure 8.1. This brings up a Bubble Chart Variables dialog to select the
variables for up to four dimensions: X-Axis, Y-Axis, Bubble Size and Bubble Color.
To illustrate this method, three variables from the Oaxaca Development sample data set are
selected: peduc_20 (percent population with an educational gap in 2020) for the X-Axis,
pserv_20 (percent population without access to basic services in the dwelling in 2020) for
the Y-Axis, and pepov_20 (percent population living in extreme poverty in 2020) for
both Bubble Size and Bubble Color. The default setting uses a standard deviational
diverging legend (see also Section 5.2.3) and results in a somewhat unsightly graph, with
circles that typically are too large, as in Figure 8.7.
Before going into the options in more detail, it is useful to know that the size of the circle can
be readily adjusted by means of the Adjust Bubble Size option. With the size significantly
reduced, a more appealing Figure 8.8 is the result.
2 https://www.gapminder.org/tools/#$chart-type=bubbles&url=v1
146 Multivariate Data Exploration

Figure 8.8: Bubble chart – bubble size adjusted: education, basic services, extreme poverty

As mentioned, the null hypothesis is that the size distribution (and matching colors) of the
bubbles should be seemingly random throughout the chart. In the example, this is clearly
not the case, with high values (dark brown colors and large circles) for pepov_20 being
primarily located in the upper right quadrant of the graph, corresponding to low education
and low services (high values correspond with deprivation). Similarly, low values (small
circles and blue colors) tend to be located in the lower left quadrant. In other words, the
three variables under consideration seem to be strongly related.
Given the screen real estate taken up by the circles in the bubble chart, this is a technique
that lends itself well to applications for small to medium sized data sets. For larger size data
sets, this particular graph is less appropriate.

8.3.1.2 Bubble chart options

In addition to Adjust Bubble Size, the bubble chart has nine main options, invoked in
the usual fashion by right clicking on the graph. Several of these are shared with other
graphs, in particular with the scatter plot, such as Save Selection, Copy Image To
Clipboard, Save Image As, View, Show Status Bar and Selection Shape. They
will not be further discussed here (see, e.g., Section 7.3.1). Classification Themes, Save
Categories and Color work the same as for any map (see Section 4.5).
Figure 8.9 provides an illustration of the flexibility that these options provide. The graph
pertains to the same three variables as before, but the Classification Themes option has
been set to Themeless. As such, this results in green circles, but the Color option applied
to the legend (see Section 4.5) allows the Opacity for the Fill Color for Category to be
set to zero, resulting in the empty circles.

8.3.1.3 Bubble chart with categorical variables

One particular useful feature to investigate potential structural change in the data is to
use the Unique Values classification, in combination with setting the Size to a constant
value.3
3 A constant variable can be readily created by means of the Calculator option in the data table.

Specifically, the Univariate tab with ASSIGN allows one to set the a variable equal to a constant.
Three Variables: Bubble Chart and 3-D Scatter Plot 147

Figure 8.9: Bubble chart – no color: education, basic services, extreme poverty

Figure 8.10: Bubble chart – categories: education, basic services, region

With peduc_20 for the X-Axis and ppov_20 (percentage population living in poverty)
as the Y-Axis, Bubble Size is now set to the constant Const, with the categorical variable
Region for Bubble Color.4 First, this yields a rather meaningless graph, based on the
default standard deviation classification. Changing Classification Themes to Unique
Values results in a scatter plot of the two variables of interest, with the bubble color
corresponding to the values of the categorical variable. In the example, this is Region, as
in Figure 8.10.
In this case, there seems to be little systematic variation along the region category, in line
with our starting hypothesis.
4 The regional classifications are: (1) Canada; (2) Costa; (3) Istmo; (4) Mixteca; (5) Papaloapan; (6)

Sierra Norte; (7) Sierra Sur; and (8) Valles Centrales.

148 Multivariate Data Exploration

Figure 8.11: 3D scatter plot: education, basic services, extreme poverty

The use of a bubble chart to address structural change is particularly effective when more
than two categories are involved. In such an instance, the binary selected/unselected logic of
scatter plot brushing no longer works. Using the bubble chart in this fashion allows for an
investigation of structural changes in the bivariate relationship between two variables along
multiple categories, each represented by a different color bubble. This forms an alternative
to the conditional scatter plot, considered in Section 8.4.2.

8.3.2 3-D scatter plot

An explicit visualization of the relationship between three variables is possible in a three-
dimensional scatter plot, the direct extension of principles used in two dimensions to a
three-dimensional data cube. Each of the axes of the cube corresponds to a variable, and
the observations are shown as a point cloud in three dimensions (of course, rendered as a
perspective plot onto the two-dimensional plane of the screen).
The data cube can be manipulated by zooming in and out, in combination with rotation,
to get a better sense of the alignment of the points in three-dimensional space. This takes
some practice and is not necessarily that intuitive to many users. The main challenge is that
points that seem close in the two-dimensional rendering on the screen may in fact be far
apart in the actual data cube. Only by careful interaction can one get a proper sense of the
alignment of the points.

8.3.2.1 Implementation
The three-dimensional scatter plot method is invoked as Explore > 3D Scatter Plot from
the menu, or by selecting the second icon from the left on the toolbar depicted in Figure 8.1.
This brings up a 3D Scatter Plot Variables selection dialog for the variables corresponding
to the X, Y and Z dimensions. Again, peduc_20, pserv_20 and pepov_20 are used.
The corresponding initial default data cube is as in Figure 8.11, with the Y-axis (pserv_20)
as vertical, and the X (peduc_20) and Z-axes (pepov_20) as horizontal. Note that the
Three Variables: Bubble Chart and 3-D Scatter Plot 149

Figure 8.12: Interacting with 3D scatter plot

axis marker (e.g., the X etc.) is at the end of the axis, so that the origin is at the unmarked
side, i.e., the lower left corner where the green, blue and purple axes meet.

8.3.2.2 Interacting with the 3-D scatter plot

The data cube can be re-sized by zooming in and out. It is sometimes a bit ambiguous what
is meant by these terms. For the purposes of this illustration, zooming out refers to making
the cube smaller, and zooming into making the cube larger (moving into the cube, so to
speak).
The zoom functionality is carried out by pressing down on the track pad with two fingers
and moving up (zoom out) or down (zoom in). Alternatively, one can press Control and
press down on the track pad with one finger and move up or down. With a mouse, one
moves the middle button up or down.
In addition, the cube can be rotated by means of the pointer: one can click anywhere in the
window and drag the cube by moving the pointer to a different location. The controls on the
left-hand side of the view allow for the projection of the point cloud onto any two-dimensional
pane, and the construction of a selection box by checking the relevant boxes and/or moving
the slider bar (see Section 8.3.2.3).
For example, in Figure 8.12, the cube has been zoomed out and the axes rotated such that Z
is now vertical. With the Project to X-Y box checked on the left, a two-dimensional scatter
plot is projected onto the X-Y plane, i.e., showing the relationship between peduc_20 and
pserv_20.

8.3.2.3 Selection in the 3-D scatter plot

Selection in the three-dimensional plot (or, rather, its two-dimensional rendering) is a little
tricky and takes some practice. The selection can be done either manually, by pressing down
the command key while moving the pointer, or by using the guides under the Select, hold
CMD for brushing check box.
150 Multivariate Data Exploration

Figure 8.13: Selection in the 3D scatter plot

Checking this box creates a small red selection cube in the graph. The selection cube can be
moved around with the command key pressed, or can be moved and resized by using the
controls to the left, next to X:, Y: and Z:, with the corresponding variables listed.
The first set of controls (to the left) move the box along the matching dimension, e.g., up or
down the X values for larger or smaller values of peduc_20, and the same for the other two
variables. The slider to the right changes the size of the box in the corresponding dimension
(e.g., larger along the X dimension). The combination of these controls moves the box around
to select observation points, with the selected points colored yellow.
The most effective way to approach this is to combine moving around the selection box and
rotating the cube. The reason for this is that the cube is in effect a perspective plot, and one
cannot always judge exactly where the selection box is located in three-dimensional space.
A selection is illustrated in Figure 8.13, where only a few out of seemingly close points in
the data cube are selected (yellow). By further rotating the plot, one can get a better sense
of their location in the point cloud.
As in all other graphs, linking and brushing is implemented for the 3D scatter plot as well.
Figure 8.14 shows an example of brushing in the six quantile map for ALTID and the
associated selection in the point cloud. The assessment of the match between closeness
in geographical space (the selection in the map) and closeness in multivariate attribute
space (the point cloud) is a fundamental notion in the consideration of multivariate spatial
correlation, discussed in Chapter 18.
Similar to the bubble chart, the 3D scatter plot is most useful for small to medium sized
data sets. For larger numbers of observations, the point cloud quickly becomes overwhelming
and is no longer effective for visualization.

8.3.2.4 3-D scatter options

The 3D scatter plot has a few specialized options, available either on the Data pane (which
was considered exclusively so far), or on the OpenGL pane. The latter gives the option
Conditional Plots 151

Figure 8.14: Brushing and linking with the 3D scatter plot

to adjust the Rendering quality of points, the Radius of points, as well as the Line
width/thickness and Line color. These are fairly technical options that basically affect
the quality of the graph and the speed by which it is updated. In most situations, the default
settings are fine.
Finally, under the Data button, there are the Show selection and neighbors and Show
connection line options. These items are relevant when a spatial weights matrix has been
specified. This is covered in Chapter 10.

8.4 Conditional Plots

Conditional plots, also known as small multiples (Tufte, 1983), Trellis graphs (Becker et al.,
1996), or facet graphs (Wickham, 2016), provide a means to assess interactions among up
to four variables. Multiple graphs are constructed for different subsets of the observations,
obtained as a result of conditioning on the value of one or two variables, different from the
variable(s) of interest. The graphs can be any kind of statistical graph, such as a histogram,
box plot, or scatter plot. The same principle can be applied to choropleth maps, resulting in
so-called micromaps (Carr and Pickle, 2010).
With one conditioning variable, the observations are grouped into subsets according to the
values taken by the conditioning variable, organized along the x-axis from low to high, e.g.,
below or above the median. For each of these subsets, a separate graph or map is drawn for
the variable(s) of interest. The same principle is applied when there are two conditioning
variables, resulting in a matrix of graphs, each corresponding to a subset of the observations
that fall in the specified intervals for the conditioning variables on the x and y-axis. Of
course, for this to be meaningful, one has to make sure that each of the subsets contains a
sufficient number of observations.
The point of departure in the conditional plots is that the subsetting of observations should
have no impact on the statistic in question. In other words, all the graphs or maps should
look more or less the same, irrespective of the subset. Systematic variation across subsets
indicates the presence of heterogeneity, either in the form of structural breaks (discrete
152 Multivariate Data Exploration

Figure 8.15: Conditional scatter plot – 3 by 3

categories), or suggesting an interaction effect between the conditioning variable(s) and the
statistic under scrutiny.
In GeoDa, conditional graphs are implemented for the histogram, box plot, scatter plot and
thematic map.

8.4.1 Implementation
A conditional plot is invoked from the menu by selecting Explore > Conditional Plot.
This brings up a list of four options: Map, Histogram, Scatter Plot and Box Plot. The
same list of four options is also created after selecting the Conditional Plot icon, the
right-most in the multivariate EDA subset in the toolbar shown in Figure 8.1. In addition,
the conditional map can also be started as the third item from the bottom in the Map
menu.
Next follows a dialog to select the conditioning variables and the variable(s) of interest. The
conditioning variables are referred to as Horizontal Cells for the x-axis, and Vertical
Cells for the y-axis. They do not need to be chosen both. Conditioning can also be carried
out for a single conditioning variable, on either the horizontal or on the vertical axis alone.
The remaining columns in the dialog are either for a single variable (histogram, box plot,
map), or for both the Independent Var (x-axis) and the Dependent Var (y-axis) in
the scatter plot option.
Conditional Plots 153

Figure 8.16: Conditional scatter plot – unique values

8.4.1.1 Conditional plot options

Two important options, special to the conditional plot, are Vertical Bins Breaks and
Horizontal Bins Breaks. These options provide a way to select the observation subsets for
each conditioning variable. They include all the same options used in map classification (see
Figure 4.2 in Chapter 4). Of special interest are the Unique Values option and the Custom
Breaks options. Unique Values is particularly appropriate when the conditioning variable
takes on discrete values, in which case other classifications may result in meaningless subsets
(see the discussion of Figure 8.15). Custom Breaks is useful when subclasses for the
conditioning variable were determined previously, preferably contained in a project file (see
Section 4.6.3).
Each graph also has its usual range of options available, with the exception of Regimes
Regression for the conditional scatter plot. Selection of observations is implemented, but
it does not result in a re-computation of the linear regression fit. On the other hand, the
conditional scatter plot does include the LOWESS Smoother option.

8.4.2 Conditional statistical graphs

To illustrate these concepts, first, a set of conditional scatter plots is created for the
relationship between peduc_20 (x-axis) and pfood_20 (y-axis). This is conditioned by
three subcategories (the default number) for pch12 in the horizontal dimension (an indicator
variable for whether the population increased, =1, or decreased, =0, between 2010 and 2020).
154 Multivariate Data Exploration

Figure 8.17: Conditional box plot – lack of health access by region

The conditioning variable in the vertical dimension is ALTID. The default classification is
to use Quantile with three categories for both conditioning variables.
This yields the plot shown in Figure 8.15. Something clearly is wrong, since there turns
out to be only one subcategory for the horizontal conditioning variable. This is because a
3-quantile classification of the indicator variable does not provide meaningful categories.
This will be dealt with next.
However, first consider the substantive interpretation of this graph. The three plots show a
strong positive and significant relationship in each case (significance indicated by **). In
other words less access to education is linearly related to less food security, more or less in
line with our prior expectations. The lack of difference between the graphs would suggest
that the relationship is stable across all three ranges for altitude, the conditioning variable.
Figure 8.16 results after changing the Horizontal Bins Breaks option to Unique Values.
At this point, the two categories for the horizontal conditioning variable correspond to the
values of 0 and 1 for the population change indicator variable.
Conditioning on this variable does provide some indication of interaction. For pch12 = 1,
the linear relationship between peduc_20 and pfood_20 is strong and positive, and not
affected by altitude. However, for pch12 = 0, there does not seem to be a significant slope
for any of the subgraphs, suggesting a lack of relationship between the two variables in those
municipalities suffering from population loss. The change in sign of the slope along altitude
is not meaningful since the slopes are not significant.
If only the horizontal classification had been used as a condition, the result would be the
same as a Regimes Regression in a standard scatter plot. However, in order to consider
the simultaneous conditioning on two variable,six separate selection sets would be required
to create six individual scatter plots with regimes, which is not very practical. In addition,
the double conditioning allows for the investigation of more complex interactions among the
variables.
Conditional Plots 155

Figure 8.18: Conditional box map – 2 by 2

An alternative as well as generalization of the consideration of spatial heterogeneity by

means of the Averages Chart (see Section 7.5.1) is the application of a Conditional Box
Plot with a conditioning variable that corresponds to spatial subsets of the data. In Figure
8.17, this is illustrated for pheal_20 (percent population that lacks access to health care),
using the Region classification as the conditioning variable along the horizontal axis. In the
graph, Valles Centrales (8) shows a median value that is higher than the other regions,
with Canada (1) having the best median health outcome (lowest value for lack of access).
While the conditional box plot is not an alternative to a formal difference in means (or
medians) test, it provides a visual overview of the extent to which heterogeneity may be
present.

8.4.3 Conditional maps

A final example is the conditional map or micromap matrix shown in Figure 8.18. Four box
maps are included for ppov_20, conditioned by peduc_20 on the horizontal axis and
pheal_20 on the vertical axis. The graph is obtained after changing the bin breaks to two
quantiles, i.e., below and above the median for each of the conditioning variable.
As before, the interest lies in the extent to which the maps represent similar spatial patterns
in each of the subcategories. The maps suggest an interaction with peduc, where more of
the brown values (more poverty) are found for above median values for lack of education,
and more blue values (less poverty) in the lower median (better education). A potential
interaction with health access is less pronounced.
156 Multivariate Data Exploration

As in any EDA exercise, considerable experimentation may be needed before any meaningful
patterns are found for the right categories of conditioning variables. Of course, this runs the
danger than one ends up finding what one wants to find, an issue which revisited in Chapter
21.

8.5 Parallel Coordinate Plot

So far, the methods discussed were limited to three or four variables. Once more than three
variables are considered, it is no longer possible to graphically represent the data points in a
multi-dimensional data cube. The parallel coordinate plot or PCP provides an alternative
that can be generalized to many dimensions. Data points are replaced by data lines, which
allows a large number of variables to be considered. The only limitation is human perception
and screen real estate.
This approach was originally suggested by Inselberg (1985; see also Inselberg and Dimsdale,
1990), and it has become a main feature in many visual data mining frameworks (e.g.,
Wegman, 1990; Wegman and Dorfman, 2003).
In a PCP, each variable is represented as a parallel axis, and each observation consists of
a line that connects points on the axes. Additional variables are represented by new axes,
so that it is easy to go beyond three variables. The only visual limitation is the number of
observations, which can make the plot of lines very crowded. However, techniques exist to
overcome this hurdle, such as binning, although this is currently not implemented in GeoDa.
A main objective in the application of PCP is the identification of clusters and outliers in
multi-attribute space. Clusters are found as groups of lines (i.e., observations) that follow a
similar path. This is equivalent to points that are close together in multidimensional variable
space. Outliers in a PCP are lines that show a very different pattern from the rest, similar
to outlying points in a multi-dimensional cloud.

8.5.1 Implementation
The PCP functionality is invoked from the menu as Explore > Parallel Coordinate
Plot, or by means of the PCP toolbar icon, the third item from the left in Figure 8.1.
This brings up a Parallel Coordinate Plot variable selection interface that consists of
two columns: Exclude and Include. Variables are chosen by moving them to the Include
column by means of arrow buttons.
Four variables are selected to illustrate the PCP: peduc_20, pserv_20 and pepov_20,
as well as pss_20 (percent of population without social security). With the default settings,
the result is as in Figure 8.19. Each variable is represented by a horizontal axis, with the
observations as lines connecting points on each axis. The variable name is listed at the left,
together with the range of the variable, its mean and standard deviation. Note that the
axes are represented with equal lengths, which corresponds to the range for each individual
variable. This implies that the distance between points on each axis does not necessarily
correspond with the same difference in value for each variable.
Parallel Coordinate Plot 157

Figure 8.19: Parallel coordinate plot

Figure 8.20: Parallel coordinate plot – standardized variables

8.5.1.1 PCP options

The PCP has nine main options, invoked by right clicking on the graph. They are the same
as discussed for other graphs and are not further considered in detail.5 The Data option
5 Specifically, the options are Classification Themes, Save Categories, Data, View, Selection

Shape, Color, Save Selection, Copy Image to Clipboard and Save Image As. Note that Classifi-
cation Themes always pertains to the variable that was originally listed on top, even when the order of
axes is later changed.
158 Multivariate Data Exploration

Figure 8.21: Brushing the PCP

Figure 8.22: Brushing map and PCP

offers a way to make all axes lengths equivalent by selecting View Standardized Data.
As a result, the data points on each axis are given in standard deviational units and become
directly comparable between the different variables, as in Figure 8.20. Observations that are
outside the −2 to +2 range can be considered to be outliers.
This graph also illustrates another feature of the PCP. By grabbing the small circular handle
on the left of an axis, it can be moved up or down in the graph, changing the order of the
variables. For example, in Figure 8.20, the axis for pepov_20 has been moved up to be
above pserv_20. Realigning the axes in this manner can often bring out patterns in the
data. However, in practice, it is not always clear which is the optimal alignment.

8.5.1.2 Brushing the PCP

Selection of observations in the PCP is implemented by moving the selection shape over
one of the axes. As for the other graphs, the default is a rectangle. Figure 8.21 illustrates
this with a small selection rectangle situated over the highest values for pss_20. This
highlights the lines (observations) that are selected. Of particular interest is the extent to
Parallel Coordinate Plot 159

Figure 8.23: Clusters of observations in PCP

Figure 8.24: Outlier observation in PCP

which these lines move together. For example, in the example, several observations follow a
similar path for pss_20, pserv_20 and pepov_20, but not for peduc_20. Insight into
any corresponding spatial pattern can be obtained by linking the selection with a map, as in
the example, using the six quantile map for ALTID.
A more explicit spatial perspective is taken by brushing a map linked to the PCP. In Figure
8.22, this is again implemented with the same six quantile map. The selection in the map
suggests that several locations within close geographical proximity also take on similar values
for a subset of the variables (close lines), but not for all. This illustrates some of the practical
difficulties associated with the concept of multivariate spatial correlation.

8.5.2 Clusters and outliers in PCP

Visual cluster detection (e.g., Wegman and Dorfman, 2003) consists of locating groups of
lines in the PCP that follow a similar path. To illustrate this feature, Figure 8.23 shows
a three-variable PCP side by side with a three-dimensional data cube for the variables
peduc_20, pserv_20 and pepov_20. The selection rectangle is centered on the three
lowest values of peduc_20 in the top axis (i.e., the best educational access). In the PCP,
160 Multivariate Data Exploration

the observation lines closely track for the three variables. In the data cube on the right, the
same three observations are highlighted in yellow (with the red arrow pointing to them), very
close together near the origin of the axes. This demonstrates the equivalence of lines being
close together in the PCP and observation points being close together in multi-attribute
variable space. Clearly, for more than three variables, only the PCP remains a viable method
to assess such closeness.
Figure 8.24 illustrates the concept of an outlier in the PCP. Now, the selection rectangle
is over the highest value for pepov_20 (in the lowest axis), an observation with extreme
poverty. As the graph shows, this also corresponds to an extreme value for pserv_20, but
much less so for peduc_20. In the data cube, the selected observation, highlighted in yellow
(with the red arrow pointing to it), is indeed somewhat removed from the rest of the data
cloud.
The number of variables that can be visualized with a PCP is only constrained by screen real
estate. However, as pointed out, for larger data sets, the graph can become quite cluttered,
making it less practical.
9
Space-Time Exploration

While the primary focus of GeoDa is on the exploration of cross-sectional or static data, it
also includes limited functionality to investigate space-time dynamics through the use of the
Time Editor, Time Player and Averages Chart.
This allows for the evolution of a variable to be shown over time through maps and graphs,
in a form of comparative statics. One limitation of this approach is that it is cross-section
oriented. More precisely, the different time periods are considered as separate cross-sectional
variables (one cross-section for each time period). While this results in quite a bit of flexibility
when it comes to grouping variables, it also means there is no inherent time-awareness, nor
a dedicated panel data structure.
After explaining the approach to space-time exploration through the Time Editor and
Time Player, I revisit the use of the Averages Chart for treatment effect analysis. In
contrast to the discussion in Chapter 7, here a dynamic analysis is possible, in that spatial
structural change can be assessed at two points in time through a difference in difference
approach.
To illustrate these methods, I continue to use the Oaxaca Development sample data set,
particularly the variables from the Mexican census at three different points in time: 2000,
2010 and 2020.

9.1 Topics Covered

• Create grouped variables by means of the Time Editor
• Save group definitions in the Project File
• Assess space-time patterns by means of the Time Player
• Carry out treatment effect analysis with the Averages Chart
GeoDa Functions
• Time > Time Editor
– grouping variables
– ungrouping variables
– editing the Time label of a variable
– removing time periods from a time-enabled variable
• Time > Time Player
– starting and pausing the player
• Box Plot
– changing the number of time-enabled variables displayed
– changing axes and synchronization options

DOI: 10.1201/9781003274919-9 161

162 Space-Time Exploration

• Scatter Plot
– time-wise autoregressive scatter plot
• Map
– moving a choropleth map through time using custom categories
• Averages Chart
– difference in means test
– difference in difference (DID) analysis
– saving the dummy variables
Toolbar Icons

Figure 9.1: Time | Averages Chart

9.2 Time Editor

The Time Editor is the mechanism through which different cross-sectional variables are
grouped into a specific temporal sequence. In general, this could be any set of variables,
but it is most relevant if the variables pertain to successive time periods. This requires
that the observations for each time period are included in the data table as a separate
cross-section. The design is that of a balanced panel, with the same cross-sectional units in
each time period.1 The system is agnostic to the fact that the variables pertain to different
time periods, so care must be taken to avoid counter intuitive situations (e.g., by selecting
the wrong sequence of variables).
The space-time functionality is invoked from the menu as Time > Time Editor. The
menu includes a separate item for each of the Time Editor and the Time Player. In
contrast, when using the toolbar icon, i.e., the left icon in Figure 9.1, both are started.
The main purpose of the Time Editor dialog is to create grouped variables.

9.2.1 Creating grouped variables

The creation of grouped variables is a somewhat tedious process that involves identifying
the variable names for each time period and combining them into a group. The interface,
shown in Figure 9.2, consists of three major panels: Ungrouped Variables, New Group
Details and Grouped Variables. Each has a brief help window to explain the meaning of
these terms, brought up by clicking on the ? symbol.
The left-most panel contains the variable names for all the numeric variables in the data
table. They are selected by double clicking or by pressing on the arrow key (>). This moves
the variable to the middle panel, where they are listed under Name. At the same time,
the variables to include item is updated. The column associated with Time includes a
period identifier, initially time 0, time 1, etc., but this can easily be edited. The name
1 An unbalanced panel would have different cross-sectional units by time period, which cannot be supported

by GeoDa's current data structure.

Time Editor 163

Figure 9.2: Time Editor interface

Figure 9.3: Grouped variables in data table

for the group is initially derived from the variable name entered but also typically requires
some further editing.
Individual variables can be moved up or down in the sequence (the sequence determines
the order of the comparative statics graphs and maps) by selecting the variable and using
the buttons at the bottom of the panel (Move Up and Move Down). In addition, any
variable can be moved out of the grouping, back to the ungrouped variables list, by selecting
it and clicking on the left arrow buttons (<).
When all the variables are included under New Group Details, in the correct order and
with the correct Time label, another arrow key (to the right of the variable list) adds
the group under the Grouped Variables list with the name as specified in the central
panel. At this point, the process can be repeated for additional variables. However, for each
grouped variable, the same number of original variables must be combined. In other words,
the grouping always pertains to a fixed number of time periods, set when the first grouped
variable is created.
The example in Figure 9.2 is for the Oaxaca data set. It shows the situation after the variable
p_P614NS (percentage of the population between 6 and 14 years of age that received
no schooling) has been created by grouping p_P614NS00 (for 2000), p_P614NS10 (for
2010) and p_P614NS20 (for 2020). The Time labels have been set to 2000, 2010 and
2020. A new grouped variable is in the process of being created using p_PJOB00 for
2000 (percentage of the population over 12 years of age with a job). Note how the entry for
name is not very useful at this point, and should be edited to p_PJOB, for example. The
next variable to be selected should be p_PJOB10, for 2010.
164 Space-Time Exploration

To support a range of illustrations, two more groups are created: p_PHA (percentage of
the population with access to health services), and p_CAR (percentage private inhabited
dwellings with a car or van).
The grouped variables are shown in the data table with their new name, followed by the first
time period in parenthesis. For example, in Figure 9.3, the entries are shown for p_PJOB
and p_PHA.
Note that the grouped variables have not been dropped as individual entries from the table,
they are just not displayed as such. The new label in parentheses indicates that they have
been grouped.
Finally, as mentioned, GeoDa is agnostic to what variables are grouped and in which order
they are arranged, so this procedure can also be used to group variables that do not necessarily
correspond to different time periods. For example, this may be useful when constructing a
box plot graph that shows the box plots for multiple variables in one window.

9.2.1.1 Grouped variables in Project File

As discussed in the case of custom map categories (Section 4.6.3), any special definition
such as a grouped variable will be lost if it is not saved in a project file (accomplished
through File > Save Project). Once contained in a project file, the grouped variables can
be re-used in later analyses. There can be several different project files associated with the
same data layer, each focused on a special aspect of time dynamics.
The project file contains the time_ids specified in the time editor, and the new variable
definitions under the group entry. For example, in Figure 9.4, the three years for the Oaxaca
data set are included under the time_ids listing below the start of variable_order.
Further down in the file, the group definitions for the p_PJOB and p_PHA variables
are shown, together with some of the non-grouped variable definitions.
Note that the grouping of variables can only use one set of time periods, due to the
synchronization mechanism that underlies the time player. So, in order to look at different
time periods, such as two periods for one variable and three for another, two separate project
files would need to be created. More precisely, any active grouping has to pertain to the same
time labels for all the grouped variable. For example, in the Oaxaca data set, the poverty
indicators pertain only to 2010 and 2020, whereas the census variables are for three years.
In order to carry out any type of space-time analysis for the poverty indicators, a separate
Time Editor operation would need to be carried out on a file that does not contain any
other grouped variables (more precisely, one would need to start over). The new definitions
should be saved in a different dedicated project file. Consequently, in the Oaxaca situation,
there could need to be separate project files for a two-period analysis (2010 and 2020) and
for a three period analysis (including 2000).
As seen before, instead of loading a spatial layer in the Connect to Data Source of the
file open operation, one should load a project file with the gda extension. This will retain
all grouped variable definitions contained in the project file.

9.2.1.2 Saving the space-time data table

A final option in the lower right hand side of the interface shown in Figure 9.2 is invoked
by the Save Space-Time Table/Weights button. This creates a data file (typically in
csv format) that contains a new space-time identifier, STID, as well as identifiers for the
Time Editor 165

Figure 9.4: Grouped variables in project file

cross-sectional observation, the time period and time period dummy variables.2 The time
grouped variables are stacked as cross-sections for each successive time period.
The space-time data file allows for pooled cross-section and time series analyses, but does
not work with a corresponding map. In other words, the data can be loaded into a table for
analysis (and, if appropriate, spatial weights included in the weights manager), but no map
is available.
In addition, if a spatial weights file is active, a corresponding space-time weights GAL file is
created as well, as detailed in Section 11.5.4.
2 When a spatial weights file is active, its ID variable is used as the cross-sectional identifier. In the

absence of an active spatial weights file, a sequence of integer values is used.

166 Space-Time Exploration

Figure 9.5: Time Player dialog

9.3 Time Player – Comparative Statics

The Time Player is the mechanism through which the space-time dynamics are implemented.
It is automatically started when selecting the Time icon from the toolbar, but not from
the menu, where it is a separate item from Time Editor. From the menu, it is invoked as
Time > Time Player.
The dialog, shown in Figure 9.5, shares several aspects with the design of the Animation
interface (see Section 5.5). The Current time is shown at the top of the interface (2000 in
the example). The default is to Loop through the different time periods, which is started by
pressing the > button (the loop can also be run in Reverse by checking the corresponding
box). Single step forward or backward is implemented through the double arrows, » and «.
Finally, the Speed of the loop can be adjusted using the slider, just as in the animation
tool.
Whenever a grouped (time-enabled) variable is selected for a graph or map, the Time
Player becomes active. The starting point (time) is determined from a dropdown list below
the variable name, which includes all available periods (for an example, see Figure 9.8 for a
scatter plot). After the initial period is selected, one can move through discrete time steps
by means of the player. While such a process does not assess true space-time dynamics,
it allows one to carry out comparative statics, i.e., to compare cross-sectional patterns at
different discrete time periods.
The comparative statics approach is illustrated with three examples that highlight some
additional features of space-time exploration: the evolution of a box plot over time; a scatter
plot with a time lagged variable; and a thematic map over time.

9.3.1 Box plot over time

To assess the evolution of the distribution of a variable over time, it is often useful to plot
the corresponding box plots side-by-side. This is the default behavior when selecting the
box plot functionality for a time-enabled (grouped) variable.
For example, in Figure 9.6, this is implemented for p_P614NS. The three graphs are shown
side-by-side, with summary statistics below. It is clear that the schooling situation improved
considerably between 2000 and 2010, but less so from 2010 to 2020. The median percent
unschooled 6 to 14 year olds decreases from 9.13% to 4.53% in the first decade, and to 4.51%
in the ten years afterward. However, the distribution is slightly more compact in 2020, with
an IQR of 3.85 compared to 4.27 in 2010 (in 2020, the IQR was 6.8). Interestingly, this
resulted in more upper outliers (i.e., municipalities that do much worse than the bulk), 22 in
2020 vs. 13 in 2010. The combination of the descriptive statistics with a visual representation
of the distribution provides an effective way to assess broad changes over time.
Time Player – Comparative Statics 167

Figure 9.6: Box plot over time

9.3.1.1 Dynamic box plot options

The dynamic box plots share all the standard options with the regular box plot, see Section
7.2.2.3. In addition, there are three sets of special options: Scale Options, Number
of Box Plots and Time Variable Options. The latter is by default checked to allow
synchronization with the Time Control. This can be specified for each individual graph.
Turning this check off disables the forward and backward movement of time periods through
the Time Player, i.e., it makes the graph a static graph.
The Scale Options have by default Fixed scale over time. Turning this off adjusts the
scale for each time period, in function of the relevant variable range. This is immaterial
when all three box plots are shown, but matters when this is not the case.
Finally, Number of Box Plots is set by default to All, resulting in the side-by-side plots
as in Figure 9.6. Other options are all values up to the number of periods in the group, in
this example, 1 and 2. Figure 9.7 illustrates the case with two plots, for 2000 and 2010.

9.3.2 Scatter plot with a time-lagged variable

With time-enabled variables, it becomes possible to create a scatter plot of the same variable
at two points in time. The slope of the linear fit then corresponds to the serial autoregressive
168 Space-Time Exploration

Figure 9.7: Box plot for two time periods

coefficient, i.e., how the variable in one period is related to its values in a previous period.
More formally, this is the slope coefficient in a linear regression of yt on yt−1 .
Since the time periods for the variables are available from a drop-down list in the variable
selection dialog, it is easy to change the time selection for both X and Y variables. In Figure
9.8, this is illustrated for p_PJOB. The Independent Var X is set to Time 2000, and
the drop-down list for Dependent Var Y is checked for Time 2010.
This results in the scatter plot shown in the left-hand panel of Figure 9.9, including both
a linear fit and a LOWESS smoother (with the bandwidth changed to 0.4). The slope of
0.288, while highly significant, only results in an overall fit of 10%. The LOWESS smoother
illustrates why this is the case. There is a clear break in the slope, with a positive slope
(steeper than the overall 0.288) for values of p_PJOB in 2000 up to roughly 52%, but
beyond that it turns to a negative slope. So, for municipalities that already had a high
percentages employed people, the relationship between 2000 and 2010 is negative, suggesting
a loss of employment for those municipalities. In an actual application, this structural break
should be further investigated by means of a Chow test. Also, its spatial footprint can be
assessed by means of linking and brushing with a map (see Sections 7.3.1.1 and 7.5.2 in
Chapter 7).
Time Player – Comparative Statics 169

Figure 9.8: Time lagged scatter plot variables

Figure 9.9: Scatter plots with time lagged variables

After adjusting the Time for the X-variable to 2010 and for the Y-variable to 2020, the
scatter plot in the right-hand panel of Figure 9.9 results. Now, the slope is much steeper,
and the fit quite a bit better (R2 of 24.9%). However, there is still considerable evidence of
structural instability, as emphasized by the S-shaped LOWESS fit (again with bandwidth at
0.4). In an actual application, this would need to be investigated further.
Unlike what is the case for univariate graphs (and maps), the Time Player does not
necessarily have the desired effect. It is essentially a univariate tool, so it only changes the
Time period for the dependent variable (Y). The X-variable remains unchanged, rather
than updating both axes to move the time lag forward or backward.

9.3.3 Thematic map over time

Since a thematic map shows the spatial distribution of a single variable, it can be readily
manipulated by means of the Time Player when it pertains to a time enabled variable.
However, since any map classification is by design relative, this may lead to an unsatisfactory
outcome.
For example, the left-hand panel of Figures 9.10–9.12 contains a six quantile map for the
percentage car ownership, p_CAR, in 2000, 2010 and 2020. These maps show an overall
spatial distribution that seems to undergo little change. However, unless one carefully inspects
170 Space-Time Exploration

Figure 9.10: Thematic maps for car ownership in 2000

Figure 9.11: Thematic maps for car ownership in 2010

Figure 9.12: Thematic maps for car ownership in 2020

each map legend, the maps represent very different realities. In 2000, the percentage car
ownership ranged from 0% to 46.8%, whereas in 2020, it varied between 1.1% and 65.8%, a
20% difference for the maximum.
In order to better represent these absolute changes, a set of custom categories must be
created by means of the Category Editor (see Section 4.6). In the right-hand panel of
Figures 9.10–9.12, a six-category equal interval classification is employed, going from 0 to 66
(the map legend lists car_map as the classification method).
The result is very different. The paucity of car ownership in the first period is emphasized
(there are only five observations over 33%) and the maps show a gradual increase over time,
with growing prevalence of the darker colors (from the center outward and to the east).
Treatment Effect Analysis – Averages Chart 171

Figure 9.13: Selected municipalities with altitude in two upper quantiles

9.4 Treatment Effect Analysis – Averages Chart

As pointed out in Section 7.5.1, the real power of the Averages Chart lies in the comparison
over time of the differences in a variable distribution between two subsets of the data. This
is particularly relevant in so-called treatment effect analysis, where one group of observations
(the Selected) is subject to a policy (treatment) between a beginning point and an endpoint.
A target variable (outcome) is observed for both time periods, i.e., before and after the
policy is implemented. The objective is to assess whether the treatment had a significant
effect on the target variable by comparing its change over time to that for a control group
(the Unselected).
In the example used here, the target variable is p_PHA, the percentage population with
access to health care. Consider an imaginary policy experiment, where between 2000 and
2020 a targeted set of policies would have been applied to improve access to health care in
higher elevation municipalities, specifically municipalities in the upper two quantiles of a six
quantile distribution for altitude (ALTID). Note that this example is used only to illustrate
the logic and implementation of a treatment effect analysis. It does not in fact correspond
to an actual policy carried out in Oaxaca.
The treated municipalities (Selected) are depicted in the familiar quantile map for ALTID,
with their selection highlighted in Figure 9.13.
The Averages Charts is invoked in the same manner as in the cross-sectional case, either
from the menu, as Explore > Averages Chart, or by selecting the right icon in the
toolbar shown in Figure 9.1. Unlike what was the case in the cross-sectional context, once
the Averages Chart is applied to a time enabled variable, it becomes possible to specify
Period 1 and Period 2 in the interface.
172 Space-Time Exploration

Figure 9.14: Difference in means, selected and unselected, static, in 2000 and 2020

This allows for two different types of applications. One is the same Difference-in-Means
Test as considered before, applied statically to the variable at different points in time.
A second is a more complete treatment effect analysis, implemented through the Run
Diff-in-Diff Test option. Each is considered in turn.

9.4.1 Difference in means

To carry out a simple static difference in means test, the time-enabled variable p_PHA
must be specified in the Variable drop-down list, with the Groups set as Selected
vs. Unselected. The time-enabled property of the variable is revealed by the beginning
and end period in parentheses, (2000–2020), following the variable name.
As before, the Difference-in-Means Test requires Selected for Group 1, and Unse-
lected for Group 2, but now the Period selection allows a choice among three time periods
for 1 and for 2 (the period for 1 must always be before or equal to that for 2). A static
analysis requires both periods to be the same. In the left-hand panel of Figure 9.14, the date
is set to 2000, whereas in the right-hand panel, it pertains to 2020.
The graph in each panel shows the values for all three time periods. However, the relevant
period is highlighted with a yellow bar and the statistics pertain only to that period.
In the illustration, in 2000, the mean for Selected is 10.45, compared to 14.74 for Unse-
lected, for a difference of −4.29. In 2020, the situation is reversed, with a mean of 79.51 for
Selected, compared to 73.90 for Unselected, now yielding a positive difference of 5.61.
In both cases the static difference in means test is highly significant.
To what extent might the difference of between the mean of 79.51 in 2020 and 10.45 in 2000
be due to our imaginary policy? This is answered by the Diff-in-Diff Test.

9.4.2 Difference in difference

The principle behind treatment effect analysis is to compare the evolution of the target
variable in the treated group before and after the policy implementation (respectively Period
1 and Period 2) to a counterfactual, i.e., what would have happened to that variable had
the treatment not been applied. Rather than a simple before and after difference in means
Treatment Effect Analysis – Averages Chart 173

Figure 9.15: Difference in difference setup, 2000 to 2020

test, the evolution of the target variable for a control group is taken into account as well.
This is the basis for a difference in difference analysis. The problem is that the counterfactual
is not actually observed. Its behavior is inferred from what happens to the control group.
A critical assumption is that in the absence of the policy, the target variable follows parallel
paths over time for the treatment and control group. In other words, the difference between
the treatment and control group at period 1 would be the same in period 2, in the absence
of a treatment effect.
The counterfactual is thus a simple trend extrapolation of the treated group. In the absence
of a treatment effect, the value of the target variable in period 2 should be equal to the
difference between treated and control in period 1 added to the change for the control
between period 1 and period 2, i.e., the trend extrapolation. The difference between that
extrapolated value for the counterfactual and the actual target variable at period 2 is then
the estimate of the treatment effect.
Formally, this can be expressed as a simple linear regression of the target variable stacked
over the two periods, using a dummy variable for the treatment-control dichotomy (the space
dummy, S, equal to 1 for the Selected), a second one for the before and after dichotomy (the
time dummy, T , equal to 1 for Period 2) and a third dummy for the interaction between
the two (i.e., treatment in the second period, S × T ):3

yt = β0 + β1 S + β2 T + β3 (S × T ) + ,

with β as the estimated coefficients, and  as a random error term. The treatment effect is the
coefficient β3 . The coefficient β0 corresponds to the mean in the control group (Unselected)
in 2000 (14.74 in Figure 9.14). β1 is the pure space effect, the difference in means between
Selected and Unselected in 2000 (-4.29 in Figure 9.14). Finally, β2 is the time trend in
3 A full discussion of the econometrics of difference in difference analysis is beyond the scope of this

chapter and can be found in Angrist and Pischke (2015), among others.
174 Space-Time Exploration

Figure 9.16: Difference in difference results

Figure 9.17: Difference in means, unselected, dynamic, 2000 to 2020

the control group, which has not yet been computed in the cross-sectional analysis (but see
Figure 9.17).

9.4.2.1 Implementation
The difference in difference implementation uses the same Averages Chart interface as
before, but with settings that disable the difference in means test. As shown in Figure 9.15,
the two groups are Selected and Unselected, but Period 1 is 2000, with Period 2 as
2020. Even though the two means are listed and the graph is shown in the right hand panel,
the results of the difference in means test are listed as 0.
Treatment Effect Analysis – Averages Chart 175

The analysis is carried out by selecting the Run Diff-in-Diff Test button. This yields the
standard GeoDa regression output, shown in Figure 9.16.4 The overall fit is quite good, with
an R2 of 83%, and all coefficients are highly significant.

9.4.2.2 Interpretation and options of Diff-in-Diff results

The interpretation of the Diff-in-Diff results centers on the regression coefficients in Figure
9.16. The estimate of the treatment effect is the coefficient of INTERACT, 9.90 and
highly significant. This would suggest that our imaginary high altitude policies indeed had a
significant positive effect on health care access.
The other pieces of the puzzle can be found in the remaining coefficients. The CONSTANT
is 14.7386, which is the value shown for the analysis in 2000 in Figure 9.14 as the mean of
the Unselected. The SPACE coefficient estimate of -4.2855 corresponds to the difference
in means between Selected and Unselected in 2000 in Figure 9.14. The trend coefficient
(T2000_2020) is 59.16. In Figure 9.17, the means for Unselected are shown for Period
1 (14.74) and Period 2 (73.90). The difference between the two is 59.16, the same as the
coefficient for the time trend, T2000_2020.
The interface offers two options to Save Dummy and Save Test Results. Selecting those
options brings up a File > Save dialog to specify the file type and file name to save a
panel data listing of the observations and dummy variables. This can be used for more
sophisticated regression analysis in other software (including GeoDa), e.g., to include relevant
control variables, and correct for serial and spatial correlation. This is beyond the current
scope.5
In spite of the limitations of the simple difference in difference analysis, it provides a way to
more formally assess the potential impact of spatial policy initiatives.

4 The regression functionality of GeoDa in not considered in this book. For a detailed treatment, see

Anselin and Rey (2014).

5 For recent reviews of some of the methodological issues associated with incorporating spatial effects in

treatment effects analysis, see, among others, Kolak and Anselin (2020), Reich et al. (2021) and Akbari et al.
(2023).
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
 Part III

Spatial Weights
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
10
Contiguity-Based Spatial Weights

A core concept in the exploration of spatial patterns is spatial autocorrelation. It is considered

in depth in Parts IV and V. At this point, it suffices to recognize that spatial autocorrelation
is essentially a compromise between attribute similarity (the usual notion of correlation) and
spatial similarity.
Spatial similarity in this context is typically interpreted as locational or geographical
similarity, i.e., the presence of a neighbor relationship between a pair of observations.
However, the concept is perfectly general, and pertains to any network structure that
represents observations as nodes, connected by edges. The existence of an edge between a
pair of observations then becomes the neighbor structure.
The matrix representation of this network structure is the so-called spatial weights matrix.
In the current and the next two chapters, this concept is discussed in more detail, although
with a particular focus on its implementation in GeoDa. Consequently, only a limited set of
aspects of this broad ranging topic will be covered. More extensive and technical discussions
as well as further references can be found in Anselin (1988), Bavaud (1998), Getis (2009),
Harris et al. (2011), Anselin and Rey (2014) and Anselin (2021), among others.
This chapter deals with spatial weights derived from the presence of a common border
between two observations (typically polygons), or contiguity-based weights. It begins with
a more formal introduction of the concept of spatial weights, followed by a review of the
creation of so-called rook and queen weights, as well as higher order contiguity weights and
space-time weights. The chapter closes with a discussion of a range of methods that can be
used to analyze and visualize characteristics of the spatial weights.
To illustrate these concepts, I will mostly use the Ceará Zika sample data set, except in the
discussion of space-time weights, where Oaxaca Development is employed. Note that to deal
efficiently with the resulting spatial weights files, it is preferable to save the sample data
sets onto a designated working directory before constructing the weights.1

10.1 Topics Covered

• Understand the concept of spatial weight
• Construct rook and queen contiguity-based spatial weights
• Compute higher order contiguity weights
• Create space-time weights
• Store the weights information in a Project File
1 The sample data sets are contained within a special file attached to the GeoDa software and cannot be

changed.

DOI: 10.1201/9781003274919-10 179

180 Contiguity-Based Spatial Weights

• Assess the characteristics of spatial weights

• Visualize the graph structure of spatial weights
• Identify the neighbors of selected observations
GeoDa Functions
• Weights Manager (Tools > Weights Manager)
– weights file creation interface
– ID variable
– rook and queen contiguity
– precision threshold option
– spatial weights file name
– weights properties in the weights manager
– loading weights from a file
• Time Editor > Save Space-Time Table/Weights
• Structure of a GAL file
• Connectivity histogram
• Connectivity map
• Connectivity graph
• Select neighbors of selected
– Map > Connectivity
– Table > Add Neighbors To Selection
Toolbar Icons

Figure 10.1: Weights Manager

10.2 The Concept of Spatial Weights

Spatial weights are a key component in any cross-sectional analysis of spatial dependence.
As mentioned in the introduction, they are an essential element in the construction of spatial
autocorrelation statistics. In addition, spatial weights provide the means to create spatially
explicit variables, such as spatially lagged variables and spatially smoothed rates (see Chapter
12).
Spatial weights are a necessary evil. In principle, one would like to have the data determine
the strength of the correlation between any pair of observations. However, with only n
observations in a cross-section, it is impossible to estimate all n × (n − 1)/2 pairwise
correlations. In this situation, growing the sample, i.e., so-called asymptotic analysis, does
not help, since the number of correlation parameters grows with n2 , while the sample only
increases linearly. In the literature, this is referred to as the incidental parameter problem.2
It has no solution other than imposing some structure that simplifies the problem.
The spatial weights accomplish such a simplification by constraining the number of possible
neighbor relations. In addition, the same correlation is imposed for all pairs of observations,
2 See Lancaster (2000) for a technical review.
The Concept of Spatial Weights 181

Figure 10.2: Example spatial layout

but this is only relevant in a discussion of global spatial autocorrelation statistics, which is
postponed until Part IV.
There are different criteria that can be used to reduce the number of neighbors for each
observation. In this chapter, approaches based on the notion of contiguity are considered, i.e.,
sharing a common border. In spatial analysis, this is typically based on geographic borders,
although as discussed in Chapter 11, the concept is perfectly general.
In the remainder of this section, the concept is introduced more formally first, followed
by a brief discussion of the two most common forms, i.e., rook and queen contiguity. This
is followed by a treatment of the concept of higher order contiguity and some practical
considerations.

10.2.1 Spatial weights matrix

A spatial weights matrix expresses the neighbor structure between the n observations in a
cross-section as a n × n matrix W. For each pair i − j, where i is the matrix row and j the
matrix column, there is a spatial weight element wij :
⎡ ⎤
w11 w12 . . . w1n
⎢ w21 w22 . . . w2n ⎥
⎢ ⎥
W=⎢ . .. .. .. ⎥ .
⎣ .. . . . ⎦
wn1 wn2 ... wnn

The spatial weights wij are non-zero when i and j are neighbors, and zero otherwise. By
convention, the self-neighbor relation is excluded, so that the diagonal elements of W are
zero, wii = 0.3
To make the concept of a spatial weights matrix more concrete, a spatial layout is used with
six polygons, shown in Figure 10.2. Two spatial units are defined as neighbors when they
share a common border, i.e., when they are contiguous. For example, in the figure, unit 1
shares a border with units 2, 4 and 5, and hence they are neighbors.
3 An exception to this rule are the diagonal elements in kernel-based weights, which are considered in

Chapter 12.
182 Contiguity-Based Spatial Weights

Figure 10.3: Contiguity as a network

The same neighbor structure can also be represented in the form of a network or graph, as
in Figure 10.3. Here, each spatial unit becomes a node in the network, and the existence of
a neighbor relation is represented by an edge or link connecting the respective nodes. Again,
node 1 is connected with a link to nodes 2, 4 and 5. It is very important to keep in mind
the generality of this network representation, since it applies well beyond purely geographic
concepts of contiguity (see Chapter 11).
In its simplest form, the spatial weights matrix expresses the existence of a neighbor relation
in binary form, with weights 1 and 0. For the layout in Figure 10.2, this yields a 6 × 6 matrix:
⎡ ⎤
0 1 0 1 1 0
⎢1 0 0 1 1 0⎥
⎢ ⎥
⎢0 0 0 0 1 1⎥
W=⎢ ⎢ ⎥. (10.1)
⎥
⎢1 1 0 0 1 0⎥
⎣1 1 1 1 0 0⎦
0 0 1 0 0 0

In the first row of this matrix, matching unit 1, the non-zero elements correspond to columns
(units) 2, 4 and 5.
In sum, the polygon boundaries in the map layout, the presence of edges in the network
and the non-zero weights in the matrix are all equivalent representations of the topology or
spatial arrangement of the data.
While naturally associated with areal units represented as polygons, the notion of contiguity
can also be applied in situations where the observations are represented as points, see Section
11.5.1.
Finally, it is important to note that even though it is referred to as a spatial weights matrix,
no such matrix is actually used in actual software operations. Spatial weights are typically
very sparse matrices, and this sparsity is exploited by using specialized data structures (there
is no point in storing lots and lots of zeros).
The Concept of Spatial Weights 183

10.2.1.1 Row-standardizing spatial weights

With a few exceptions, the analyses in GeoDa that employ spatial weights use them in
row-standardized form. Row-standardization takes the given weights wij (e.g., the binary
zero-one weights) and divides them by the row sum:

wij(s) = wij / wij .
j

As a 
result,
 each row sum of the new matrix equals one. Also, the sum of all weights,
S0 = i j wij , equals n, the total number of observations.4
For the layout in Figure 10.2, the matching row-standardized weights matrix is:
⎡ ⎤
0 1/3 0 1/3 1/3 0
⎢1/3 0 0 1/3 1/3 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 1/2 1/2 ⎥
Ws = ⎢ ⎢ ⎥. (10.2)
1/3 1/3 0 0 1/3 0 ⎥
⎢ ⎥
⎣1/4 1/4 1/4 1/4 0 0 ⎦
0 0 1 0 0 0

Importantly, whereas the original binary weights matrix was symmetric, this is no longer the
case with the row-standardized weights. For example, the element in row 1-column 5 equals
1/3, whereas the element in row 5-column 1 equals 1/4. This has important consequences for
some operations in spatial regression analysis, but less so in an exploratory context (but see
Section 12.3.1). Also, the weights are no longer equal. In fact, they are inversely related to
the number of neighbors. With more neighbors (say, k), the weight given to each neighboring
observation (1/k) becomes smaller. More importantly, when there is only one neighbor (as in
row 6, column 3 in the example), the weights remains 1. This may lead to counter intuitive
results when computing spatial transformations, as in Chapter 12.

10.2.2 Types of contiguity

Sharing a common border may seem a straightforward definition of contiguity. However, in
practice, it is not always clear what constitutes a border of non-zero length. For example,
consider the regular three by three grid layout shown in Figure 10.4
In this layout, contiguity can be defined in three ways, each yielding a different spatial
weights matrix. The three criteria are referred to as rook, bishop and queen, in analogy to
the moves allowed by matching pieces on a chess board.
The rook criterion defines neighbors as those cells to the east, west, north and south, yielding
four neighbors for each spatial unit that is not on the boundary of the spatial layout. In the
example, this yields 2, 4, 6 and 8 as the neighbors of cell 5, as shown in Figure 10.5.
The matching binary contiguity weights matrix for the nine spatial units is then the 9 × 9
4 Strictly speaking, this is only correct in the absence of so-called isolates, i.e., observations without

neighbors. With q isolates, the sum S0 = n − q. See Section 11.4.2.

184 Contiguity-Based Spatial Weights

Figure 10.4: Regular grid layout

Figure 10.5: Rook contiguity

Figure 10.6: Queen contiguity

matrix: ⎡ ⎤
0 1 0 1 0 0 0 0 0
⎢1 0 1 0 1 0 0 0 0⎥
⎢ ⎥
⎢0 1 0 0 0 1 0 0 0⎥
⎢ ⎥
⎢1 0 0 0 1 0 1 0 0⎥
⎢ ⎥
W=⎢
⎢0 1 0 1 0 1 0 1 0⎥⎥. (10.3)
⎢0 0 1 0 1 0 0 0 1⎥
⎢ ⎥
⎢0 0 0 1 0 0 0 1 0⎥
⎢ ⎥
⎣0 0 0 0 1 0 1 0 1⎦
0 0 0 0 0 1 0 1 0
The structure of the weights matrix is characterized by a band of ones. In addition, there is
a strong effect of boundary units in this small data set. Internal units have four neighbors,
but corner units only have two, and the other boundary units only have three.
The Concept of Spatial Weights 185

A second criterion to define contiguity, less frequently used, is based on common vertices
(corners) as the convention to identify neighbors. This is referred to as bishop contiguity.
Due to its lack of adoption in practice, it is not further considered here.
Finally, the queen criterion combines rook and bishop. As a result, neighbors share either
a common edge or a common vertex. This yields eight neighbors for the non-boundary
units. For example, as shown in Figure 10.6, the central cell 5 has all the other cells being
neighbors.
The matching 9 × 9 binary contiguity matrix is much denser than before:
⎡ ⎤
0 1 0 1 1 0 0 0 0
⎢1 0 1 1 1 1 0 0 0⎥
⎢ ⎥
⎢0 1 0 0 1 1 0 0 0⎥
⎢ ⎥
⎢1 1 0 0 1 0 1 1 0⎥
⎢ ⎥
W=⎢ ⎢1 1 1 1 0 1 1 1 1⎥ ⎥
⎢0 1 1 0 1 0 0 1 1⎥
⎢ ⎥
⎢0 0 0 1 1 0 0 1 0⎥
⎢ ⎥
⎣0 0 0 1 1 1 1 0 1⎦
0 0 0 0 1 1 0 1 0

Now, the corner cells have three neighbors, and the other boundary cells have five.
The concepts of rook and queen contiguity can easily be extended to other regular tessellations,
such as hexagons.5 In addition, they apply to irregular lattice layouts as well. In this context,
rook neighbors are those spatial units that share a common edge, and queen neighbors
those that share either a common edge or a common vertex. As a result, spatial weights
based on the queen criterion will always have at least as many neighbors as a corresponding
rook-based weights matrix.
In practice, whether a common segment is categorized as a point (vertex) or line segment
(edge) depends on the precision of the geocoding in a GIS. Due to generalization (a process
of simplifying the detail in a spatial layer), some small line segments may be represented
as points in one GIS and not in another. In addition, similar quirks in the representation
of polygon boundaries may result in (very) small empty spaces between them. In a naive
approach to defining common boundaries, such instances would fail to identify neighbors.
Therefore, some tolerance may be required to make the weights creation robust to such
characteristics.

10.2.3 Higher-order contiguity

So far, the focus has been on the notion of a direct neighbor, or, more precisely, a first-order
neighbor. The concept of contiguity can be generalized to allow for higher order neighbors,
similar to the time series context, where a time shift can pertain to multiple periods.
A higher order neighbor is defined in a recursive fashion, as a first-order neighbor to a lower
order neighbor. More formally, j is a neighbor of order k to i if:
• j is a first-order neighbor to h,
• h and i are neighbors of order k − 1 and
• j is not already a lower order neighbor to i.
5 In the literature pertaining to cellular automata, counterparts to the rook and queen contiguity are the

so-called von Neumann and Moore neighborhoods. The main difference is that here, the central cell is not
included in the neighborhood concept, whereas it is in cellular automata.
186 Contiguity-Based Spatial Weights

Figure 10.7: Second-order contiguity

Figure 10.8: Higher order contiguity as multiple steps in the graph

Using this logic, candidates to be a second-order neighbor would be any first-order neighbor
to another observation that is already a first-order neighbor. However, this would be limited
to only those locations that are not already first-order neighbors. Parenthetically, it should
be noted that the concept of higher order contiguity is similar to the notion of reachability
in social networks (e.g., Newman, 2018).
In the layout shown in Figure 10.7, cells 4 and 3 are highlighted. Cell 4 has three first-order
neighbors, immediately surrounding it: 1, 2 and 5. Cell 3 is a first-order neighbor to 5, hence
3 is a second-order neighbor to 4.
In the graph representation of the contiguity structure in Figure 10.8, second-order contiguity
corresponds to of a path of length 2 from node 4 to node 3, illustrated by the dashed lines.
This path requires the traversal of two edges in the graph, [4-5] and [5-3] and is thus of
length 2. As it turns out, the number of edges separating a pair of nodes corresponds to the
order of contiguity.
The Concept of Spatial Weights 187

As such, identifying paths of length 2 between nodes is not sufficient to find the correct
second-order neighbors, since it does not preclude redundant or circular paths. For example,
a closer examination of Figure 10.8 reveals that 1 and 5 can be reached in two steps from 4
as well (e.g., 4-1-2, 4-5-1). In fact, multiple paths of length 2 can exist between two nodes
(e.g., 4-1-2 and 4-5-2).
These paths illustrate the complexity of the problem of finding higher order neighbors. For
example, nodes 2 and 5 cannot be both first-order and second-order neighbors of node 4, so
a strict application of the graph-theoretical approach is not sufficient in the spatial weights
case. Ways must be found to eliminate these redundant paths in order to yield the proper
second-order neighbor, node 3.

10.2.3.1 Circular and redundant paths

The problem of circular and redundant paths was pointed out early on by Blommestein
(1985), in the context of the specification of spatial autoregressive models. For such models,
the inclusion of the same observation as a neighbor at different orders of contiguity (e.g.,
both first- and second-order neighbor) constitutes a so-called identification problem (it
is impossible to tell certain parameters apart). An efficient algorithm to remove such
redundancies is developed by Anselin and Smirnov (1996), based on shortest paths in the
network structure.
However, in an exploratory context, it may be useful to include lower order neighbors in
order to mimic the effect of expanding bands of neighbors, similar to an expanding distance
band (see Section 11.3.1). GeoDa includes both concepts of higher order weights (Section
10.3.4).

10.2.4 Practical considerations

In practice, the construction of the spatial weights from the geometry of the data cannot
be done by visual inspection or manual calculation, except in the most trivial of situations.
Specialized software is needed to compute distances between points from their coordinates
or to derive the presence of common borders from the boundary information for polygons.
To assess whether two polygons are contiguous is quite demanding. This requires the use of
explicit spatial data structures to deal with the location and arrangement of the polygons, a
functionality that is lacking in most econometric and statistical software. The specialized
spatial data representation that differentiates a GIS from a standard data base system, and
especially the use of a spatial index are crucial in this respect.
Many commonly used spatial data structures are in a so-called spaghetti representation.
While this approach is more efficient for rapid rendering of maps, it lacks indexing and
topology. It is popular in many cartographic packages and map-oriented GIS. A well known
example is ESRI’s shape file format.
GeoDa uses its own internal data structure to represent polygons loaded from different sources
(files as well as spatial data bases), so that it can take advantage of spatial indexing to
optimize the generation of the weights information.
It is important to keep in mind that the spatial weights are critically dependent on the
quality of the spatial data source (GIS) from which they are constructed. Problems with
the topology in the GIS (e.g., slivers) will result in inaccuracies for the neighbor relations
included in the spatial weights. In practice, it is essential to check the characteristics of the
weights for any evidence of problems. When problems are detected, the solution is to go
188 Contiguity-Based Spatial Weights

Figure 10.9: Weights Manager interface

back to the GIS and fix or clean the topology of the data set. Editing of spatial layers is not
implemented in GeoDa, but this is a routine operation in most GIS software.

10.3 Creating Contiguity Weigths

In GeoDa, the creation and manipulation of spatial weights is carried out by means of a
Weights Manager. This dialog is invoked from the menu as Tools > Weights Manager.
Alternatively, the large W icon on the toolbar can be used, highlighted in Figure 10.1.
The main interface is shown in Figure 10.9. It contains buttons to Create, Load and
Remove weights that are listed under Weights Name. The highlighted weight is the
currently active one (for example, see Figure 10.11). Its summary characteristics are listed
in the table with columns Property and Value.
A second row of buttons allows various operations on existing weights, including Intersec-
tion, Union and Make Symmetric. Finally, at the bottom of the interface are options to
visualize the properties of the weights, in the form of a Histogram, Connectivity Map
and Connectivity Graph.

10.3.1 Weights manager

The actual construction of the weights is implemented through the Weights File Creation
interface, invoked by the Create button. It is shown in Figure 10.10 for the Ceará Zika
sample data set. The interface provides the entry point to all the available options. Two
buttons correspond to the main types, Contiguity Weight or Distance Weight. In this
example, the option for Rook contiguity is checked, with the Contiguity Weight button
active.
Creating Contiguity Weigths 189

Figure 10.10: Rook contiguity in the Weights File Creation interface

10.3.1.1 ID variable
The first item to specify in the dialog is the ID Variable. This variable has to be unique
for each observation. It is a critical element to make sure that the weights are connected to
the correct observations in the data table. In other words, the ID variable is a so-called key
that links the data to the weights.
In GeoDa, it is best to have the ID Variable be integer. In practice, this is often not the case,
and even though the identifier may look like an integer value, it may be stored as a string.
For example, the standard identifiers that come with many open data sets (such as the US
census data) are typically character variables. One way to deal with this problem is to use
the Edit Variable Properties functionality in the table to turn a string into an integer,
as shown in Section 2.4.1.1.
In some instances, there is no easy way to identify an ID variable. In that case, the Add ID
Variable button provides a solution: the added ID variable is simply an integer sequence
number that is inserted into the data table (as always, there must be a Save operation to
make the addition permanent).
For the Ceará Zika data set, code7 is used as the ID variable, as indicated in Figure 10.10.

10.3.2 Rook contiguity

With the Rook contiguity radio button checked, as in Figure 10.10, a click on Create
will start the weights construction process.
First, a file dialog appears in which a file name for the weights must be specified (the file
extension GAL is added automatically). In the example, this file is ceara_r.gal. Since
there are no real metadata in a spatial weights file, it is a good practice to make the file
name something meaningful, so that one can remember what type of weight was created.
190 Contiguity-Based Spatial Weights

Figure 10.11: Rook weights summary interface

Figure 10.12: GAL file contents

For example, an _r added to the name of the data would suggest rook weights. However, as
outlined in Section 10.3.5, if a Project File is saved, several of the characteristics of the
weights (i.e., its metadata) are stored in that file for later reuse.
The weights are immediately computed and written to the GAL file. At the end of this
operation, a success message appears (or an Error message if something went wrong).
A useful option in the weights file creation dialog is the specification of a Precision
threshold (situated right below the Rook contiguity radio button). In most cases, this is
not needed, but in some instances the precision of the underlying shape file is insufficient to
allow for an exact match of coordinates (to determine which polygons are neighbors). When
this happens, GeoDa suggests a default error band to allow for a fuzzy comparison. In most
cases, this will be sufficient to fix the problem.
After the weights are created, the weights manager becomes populated, as shown in Figure
10.11. The name for the file that was just created is now included under the Weights Name.
In addition, under the item Property, several summary characteristics are listed. This is
further examined in Section 10.4.1.
Creating Contiguity Weigths 191

Figure 10.13: Summary properties of first-order queen contiguity

10.3.2.1 GAL file

The GAL weights file is a simple text file that contains, for each observation, the number of
neighbors and their identifiers. The format was suggested in the 1980s by the Geometric
Algorithms Lab at Nottingham University. It achieved widespread use after its inclusion in
SpaceStat (Anselin, 1992), and subsequent adoption by the R spdep package and others.
The one innovation SpaceStat added to the format was the inclusion of a header line, with
some metadata for the weights, such as the number of observations (184), the name of
the shape file from which the weights were derived (zika_ceara), and the name of the
ID variable (code7).6 As illustrated in Figure 10.12, for each observation, the number of
neighbors is listed after its ID. For example, for the municipality with code7 2300101, there
are 3 neighbors, with respective IDs 2302503, 2308302 and 2308401.
The GAL file is a simple text file. It can thus be edited (e.g., change neighbors), although
this is not recommended: it is easy to break the symmetry of the weights.

10.3.3 Queen contiguity

Spatial weights based on the queen contiguity criterion are created in the same way, but
with the Queen contiguity button checked in the dialog of Figure 10.10. If needed, the
Precision threshold option is available as well.
After the weights file is created, e.g., ceara_q.gal, its summary properties are listed in the
weights manager, with the associated file name highlighted, as in Figure 10.13.
6 The 0 at the start of the header line is currently not used.
192 Contiguity-Based Spatial Weights

10.3.4 Higher-order contiguity

Higher order contiguity weights are constructed in the same general manner as the first-order
weights just considered. To accomplish this, a value larger than the default of 1 must be
specified in the Order of contiguity box in Figure 10.10.
One important aspect of higher order contiguity weights is whether or not the lower order
neighbors should be included in the weights structure. This is determined by a check box.
Importantly, as outlined in Section 10.2.3, there is a difference between the two concepts.
The pure higher order contiguity does not include any lower order neighbors, whereas they
are included when the corresponding check box is checked.
Using the second-order for the queen contiguity criterion in the Ceará data, e.g., saved into
ceara_q2.gal, the summary properties are shown in Figure 10.14.
In contrast, an encompassing notion of second-order neighbors would include the first-order
neighbors as well. The properties of such a weights file, e.g., ceara_q2i.gal, are listed in
Figure 10.15. Note how the include lower orders entry is set to true, in contrast with
the entry in Figure 10.14, where it is set to false.
The difference between the two concepts can be easily gathered from the descriptive statistics.
As expected, the inclusive second-order weights are denser. For example, this is illustrated
by the percent non-zero: 6.56% for the pure second-order and 9.51% for the inclusive
second-order. Weights characteristics are considered in more detail in Section 10.4.

10.3.5 Project file

Similar to how custom classifications and grouped variable specifications were earlier saved
to a Project File, the same can be achieved for spatial weights. This is very important,
since otherwise any weights metadata are lost. They are not included in the saved weights
file (see Section 10.3.6).
As before, this is accomplished by invoking File > Save Project and specifying a file
name. A gda file extension will be added.
In Figure 10.16, the respective summary weights properties are listed under the
weights_entries tag in the file.
As in the previous cases, the gda file must be specified as the input file when starting a new
analysis, to make sure the weights are loaded properly. Their properties will immediately
appear in the weights manager dialog.

10.3.6 Using existing weights

When a Project File is not used as the input, several of the summary properties of the
weights are lost. For example, when using the Load option in the weights manager (the
middle button in Figure 10.9) for the second-order queen contiguity file ceara_q2.gal, the
resulting summary is as in Figure 10.17.
In contrast to the entries in Figure 10.14, there is no awareness of the type (set to custom
compared to queen), the symmetry, or order of contiguity. Since the numerical properties
are computed on the fly, they are still included.
In general, it is highly recommended to use project files to keep track of the spatial weights
metadata.
Creating Contiguity Weigths 193

Figure 10.14: Summary properties of second-order queen contiguity

Figure 10.15: Summary properties of inclusive second-order queen contiguity

10.3.7 Space-time weights

The Time Editor considered in Chapter 9, as depicted in Figure 9.2, contains a button in
the lower right corner to Save Space-Time Table/Weights.
194 Contiguity-Based Spatial Weights

Figure 10.16: Weights description in Project File

This is useful in order to trick the cross-sectional data structure of GeoDa to allow for simple
pooled time series/cross section analyses.7 The pooled aspect means that the same coefficients
or model parameters hold for all time periods, which is not necessarily a realistic assumption.
However, it is often the point of departure in an analysis of space-time dynamics.
7 The most common application is to carry out a regression analysis, but this approach can also be used

to compute spatial autocorrelation statistics for the pooled data.

Creating Contiguity Weigths 195

Figure 10.17: Second-order contiguity weights loaded from file

Figure 10.18: Saved space-time table as csv file

The critical element in this endeavor is to create a unique space-time ID variable (STID) so
that the stacked cross-sections (one for each time period) can be handled as if they were one
single cross-sectional data set, with a unique ID for each space-time observation. Analogously,
the spatial weights are block-diagonal, with the same weights structure repeated for each
stacked time period.
This is illustrated for the Oaxaca Development data set from Chapter 9. Note that it is
necessary to have a spatial weights file active in the weights manager for this to work
properly. In the example, this is a first-order queen contiguity weights file using mun as the
ID. A snapshot of the corresponding space-time data table, saved as a csv file, is shown in
Figure 10.18. In addition to the cross-sectional identifier mun, there is a unique space-time
identifier STID, as well as an indicator variable for each time period: T_2000, T_2010
and T_2020. The remaining entries consist of the time-enabled (grouped) variables.
The contents of the corresponding gal file are illustrated in Figure 10.19. The header line
lists 1710 as the number of space-time observations (570 × 3), stoaxaca as the name of the
source file with the data (i.e., the file shown in Figure 10.18), and STID as the key. The
remaining entries follow the same format as before.
196 Contiguity-Based Spatial Weights

Figure 10.19: Space-time GAL contiguity file

10.4 Weights Characteristics

In addition to creating and loading spatial weights, the Weights Manager also contains
functionality to compute and visualize some important properties of the weights. This
includes summary characteristics, as well as the visualization of the connectivity structure
by means of a histogram, map, or graph.

10.4.1 Summary characteristics

Figures 10.11 and 10.13–10.15 show the summary properties for rook, queen, second-order
queen and inclusive second-order queen weights computed for the Ceará example. These
include both metadata as well as descriptive statistics computed from the weights.
The metadata include the type of weights, symmetry, the original spatial layer (file), the
id variable, the order, and, for higher order contiguity, whether include lower orders
is true or false.
The descriptive statistics consist of the number of observations, the min, max, mean
and median number of neighbors, as well as an overall measure of sparsity, in the form of
the % non-zero.
For the four weights considered so far, there is an increasing degree of density, reflected in all
the descriptive statistics. For example, the mean number of neighbors increases from 5.09
for rook, 5.43 for queen, 12.07 for second-order queen, to 17.50 for inclusive second-order
queen.

10.4.2 Connectivity histogram

The Histogram button at the left-hand side of the bottom of the weights manager (Figure
10.9) produces a connectivity histogram. This shows the number of observations for each
value of the cardinality of neighbors, i.e., how many observations have the given number of
neighbors. The graph is illustrated in Figure 10.20 for the rook contiguity associated with
the municipalities of Ceará. It is a standard GeoDa histogram, with a number of options
available, most of which are generic (see Section 7.2.1), and some specific to the connectivity
histogram.
Weights Characteristics 197

Figure 10.20: Rook contiguity histogram

Figure 10.21: Rook and queen contiguity histograms

The overall pattern is quite symmetric, with a mode of 5 (i.e., most municipalities have 5
rook neighbors), but quite a long tail to the right (i.e., a few municipalities have a much
larger number of neighbors). In addition to the visual inspection, the usual statistics of the
distribution can be added to the bottom of the table by means of the View > Display
Statistics option. The descriptive statistics are the same as listed in the weights manager
summary.
Figure 10.21 illustrates the minor differences between rook and queen contiguity. In the
histogram on the right, for queen contiguity, the municipalities with six neighbors are selected.
In the linked contiguity histogram for rook weights on the left, the matching distribution is
shown. Several of the observations with 6 neighbors using the queen criterion, have only
198 Contiguity-Based Spatial Weights

Figure 10.22: Linked contiguity histogram and map

5 or even 4 neighbors for the rook criterion. This highlights the more inclusive nature of
queen contiguity.
In addition, the contiguity histogram can be linked to any other window, such as the
themeless map in the left-hand panel of Figure 10.22. In the contiguity histogram on the
right (queen contiguity), the two observations with 13 neighbors are selected. A closer
examination of the map identifies two very large areal units. Consequently, they will have
more neighbors than the average sized unit.
It is good practice to check the connectivity histogram for any strange patterns, such as
observations with only one neighbor, or neighborless observations (isolates, see Section
11.4.2).
Ideally, the distribution of the cardinalities should be smooth (i.e., no major gaps) and
symmetric, with a limited range. Deviations from symmetry, such as bi-modal distributions
(some observations have few neighbors and some have many) require further investigation.
Any such characteristics will have an immediate effect on the corresponding weights, with
implications for all the computations that use these weights.

10.4.2.1 Neighbor cardinality

A useful option of the connectivity histogram is to save the neighbor cardinality to the
data table as an additional column/variable. This can then be used as an input in further
analysis. For example, the number of neighbors is an important input into the calculation of
the significance of the local join count statistic, covered in Chapter 19.
The Save Connectivity to Table option is invoked in the usual way by right clicking on
the connectivity histogram window. This brings up a small dialog to specify the variable
name (the default is NUM_NBRS).

10.4.3 Connectivity map

The middle button at the bottom of the Weights Manager interface (Figure 10.9) brings
up a Connectivity Map. This is a standard GeoDa themeless map view, with all the usual
features invoked through toolbar icons, including zooming and panning. In addition, the
connectivity map includes a special functionality that highlights the neighbors of any selected
Weights Characteristics 199

Figure 10.23: Connectivity map for queen contiguity

observation. The window header lists the weights matrix to which the connectivity structure
pertains.
In Figure 10.23, this is illustrated using queen contiguity for the Ceará municipalities. As
soon as the pointer is moved over one of the observations, it is selected and its neighbors
are highlighted. Note that this behavior is slightly different from the standard selection in
a map, since it does not require a click to select. Instead, the pointer is in so-called hover
mode. As soon as the pointer moves outside the main map, the complete layout is shown.
In the example, the selection is over the municipality of Santa Quitéria (code7 2312205),
one of the two observations identified as having 13 neighbors. This outcome is listed in the
status bar below the map, as well as the identifying codes for the neighbors. Moving the
pointer to another observation immediately updates the selection and its neighbors.
The Connectivity Map in the weights manager is actually a special case of the Connec-
tivity option in any map. This is considered in more detail in Section 10.4.5.1.

10.4.4 Connectivity graph

The right-most button at the bottom of the Weights Manager interface (Figure 10.9)
brings up a Connectivity Graph. Like the connectivity map, this is a standard GeoDa
map view, with all the usual options invoked through toolbar icons, including zooming and
panning.
On additional feature is that the graph structure that corresponds to the spatial weights
connectivity is superimposed on the map. For the queen contiguity in Ceará, this yields Figure
10.24. Again, the window header lists the corresponding spatial weights file (ceara_q).
As in the connectivity map, there are a few options that allow for some customization of the
graph structure representation. These include the color and thickness of lines, and whether
the background map is shown or not.
The default edge thickness is Normal, but two other options are available, to make the
lines thicker (Strong) or thinner (Light). It is also often useful to change the color of the
graph to something different from the default black. In Figure 10.24, these options have
been used to change the thickness to Strong, the color to blue, the Opacity of the Fill
Color to zero and the Outline Color for Category to black.
200 Contiguity-Based Spatial Weights

Figure 10.24: Connectivity graph for queen contiguity

Figure 10.25: Connectivity graph for queen contiguity without map

The most dramatic of the options is the Hide Map feature. By checking this, the background
map is removed and only the pure graph structure remains, as in Figure 10.25. This clearly
brings out some interesting features of the graph, such as the observation with a single
neighbor (a single edge in the graph) in the extreme north-east (highlighted in red).

10.4.5 Connectivity option

Whenever there is an active weights matrix specified in the weights manager, the Connec-
tivity Option becomes available in any map window as well as in the Selection Tool for
the table.

10.4.5.1 Connectivity option in the map

In the discussion of the various map options in Section 4.5 of Chapter 4, the Connectivity
option was deferred to the current chapter.
Weights Characteristics 201

This option supports operations similar to those in a Connectivity Map, but applied
to any current map. The functionality is invoked by selecting Connectivity > Show
Selection and Neighbors from the options menu, in the usual fashion (right click).
With the option checked, the selection feature works the same as in the Connectivity
Map. The main difference between the functionality in the Connectivity Map and the
Connectivity option in any thematic map is that the latter is updated to the current active
spatial weights file. In other words, as the active weights change, the connectivity structure
is adjusted. In contrast, when selecting the Connectivity Map from the weights manager,
there is only one such map, tied directly to the weights file from which it was invoked.
With the Show Selection and Neighbors option checked, it becomes possible to Change
Outline Color of Neighbors, to Change Fill Color of Neighbors, and implement
other cosmetic adjustments to the look of the selected polygons.

10.4.5.2 Connectivity option in the table

Neighbors of selected observations can also be displayed by means of a a feature of the table
Selection Tool, as long as a weights matrix is active. The Add Neighbors to Selection
button is central in the middle panel of of the Selection Tool in Figure 2.24 of Chapter 2.
This functionality requires that there is a current selection. Also, the Weights box must
be specified. By default, the drop-down list will show the currently active weights, e.g.,
ceara_q.
Clicking on this button will add the neighbors of a current selection to the selection, similar to
the operation of the Append to Current Selection option. Note that once the neighbors
have been added, the current selection has grown to include those observations, so another
Add Neighbors to Selection operation increases the selection set even more. This is not
always what one has in mind.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
11
Distance-Based Spatial Weights

In this second chapter devoted to spatial weights, the focus is on weights that are derived
from a notion of distance. Intrinsically, this is most appropriate for point layers, but it can
easily be generalized to polygons as well, through the use of the polygon centroids.
The chapter starts with a brief overview of distance metrics and the fundamental difference
between points expressed in projected coordinates and points in latitude-longitude decimal
degrees. Whereas for the former the familiar concept of Euclidean distance can be applied,
the latter requires the computation of great circle distance or arc distance. In addition, a
concept of distance in multivariate attribute space is introduced.
This is followed by a review of spatial weights that are based on distance-bands, k-nearest
neighbor weights and the use of generalized concepts of distance.
Next is a discussion of the implementation of the weights functionality in GeoDa, including a
broadening of the strict concept of contiguity to apply to points.
The chapter closes with a discussion of set operations on weights, such as intersection, union
and making symmetric.
The methods are illustrated with a point data set that contains performance measures for
261 Italian community banks for 2011–2017. This data set was used in an analysis of spatial
spillovers in technical efficiency by Algeri et al. (2022). The data are contained in the Italy
Community Banks sample data set. The space-time weights are illustrated with the Oaxaca
Development data.
As mentioned in the previous chapter, to work most effectively with the spatial weights files,
the data sets should first be copied (Save As) to a working directory.

11.1 Topics Covered

• The concept of a distance metric
• Construct distance band spatial weights
• Great circle distance
• The contents of a gwt weights file
• Assess the characteristics of distance-based weights
• Assess the effect of the max-min distance cut-off
• Identify isolates
• Construct k-nearest neighbor spatial weights
• Making a k-nearest neighbor weights matrix symmetric
• Construct contiguity weights for points and distance weights for polygons
• Block weights

DOI: 10.1201/9781003274919-11 203

204 Distance-Based Spatial Weights

• Space-time weights
• Create a general dissimilarity matrix based on multi-attribute distance
• Investigate commonalities among spatial weights (intersection and union)
GeoDa Functions
• Weight File Creation dialog
– distance-band weights
– k-nearest neighbor weights
– great circle distance option
– block weights option
• Weights Manager
– Connectivity Histogram, Map and Graph
– Intersection, Union
– Make symmetric
• Time Editor
– Save Space-Time Weights

11.2 Distance Metrics

Pairs of observations are considered to be neighbors based on how close they are to each
other. To make this precise, there is a need for a formal definition of distance. Two important
concepts are considered: distance between points in a Cartesian coordinate system, and
distance between points located on a sphere. This can be further generalized to distance
between points in multivariate attribute space, i.e., to a non-geographical notion of distance.

11.2.1 Distance in a Cartesian coordinate system

A general concept for the distance between two points i and j, with respective coordinates
(xi , yi ) and (xj , yj ) in a Cartesian coordinate system, is the so-called Minkowski metric:
(1/p)
dpij = (|xi − xj |p + |yi − yj |p ) ,

with p as a general exponent. The Minkowski metric itself is not often used in spatial analysis,
but there are two special cases of great interest.
Arguably the most familiar case is for p = 2, i.e., Euclidean or straight line distance, dij , as
the crow flies:1
 (1/2)
dij = |xi − xj |2 + |yi − yj |2 ,
or, in its more familiar form:

dij = (xi − xj )2 + (yi − yj )2 .

An alternative to Euclidean distance that is sometimes preferred because it lessens the

influence of outliers is the so-called Manhattan block distance. This metric is obtained by
setting p = 1 in the Minkowski expression. This notion only considers movement along the
1 Since |xi − xj |2 is equivalent to (xi − xj )2 , the latter expression is typically used for Euclidean distance.
Distance Metrics 205

east-west and north-south directions, as in the city blocks of Manhattan. Formally, it is

expressed as:
ij = |xi − xj | + |yi − yj |.
dm

11.2.2 Great circle distance

Euclidean inter-point distances are only meaningful when the coordinates are recorded on
a plane with a Cartesian coordinate system. This implies that any point layer must be
projected for Euclidean distances to be appropriate.
In practice, one often works with unprojected points, expressed as decimal degrees of latitude
and longitude. In this case, using a straight line distance measure is inappropriate, since it
ignores the curvature of the earth. This is especially the case for longer distances, such as
when observations span a continent. Also, decimal degrees are not Cartesian coordinates,
and consequently a notion of distance between degrees is meaningless.
The proper distance measure in this case is the so-called arc distance or great circle distance.
This approach takes the latitude and longitude in decimal degrees as input into a conversion
formula.2 Decimal degrees are obtained from the degree-minute-second value as degrees +
minutes/60 + seconds/3600.
The latitude and longitude in decimal degrees are converted into radians as:

Latr = (Latd − 90) ∗ π/180

Lonr = Lond ∗ π/180,

where the subscripts d and r refer respectively to decimal degrees and radians, and π =
3.14159 . . .. With ΔLon = Lonr(j) − Lonr(i) , the expression for the arc distance is:

dij = R ∗ arccos[cos(ΔLon) ∗ sin Latr(i) ∗ sin Latr(j) )

+ cos Latr(i) ∗ cos Latr(j) ],

or, equivalently:

dij = R ∗ arccos[cos(ΔLon) ∗ cos Latr(i) ∗ cos Latr(j) )

+ sin Latr(i) ∗ sin Latr(j) ],

where R is the radius of the earth. In GeoDa, the arc distance is obtained in miles with R =
3959, and in kilometers with R = 6371.
However, it should be noted that these calculated distance values are only approximate,
since the radius of the earth is taken at the equator. A more precise measure would take
into account the actual latitude at which the distance is measured. In addition, the earth’s
shape is much more complex than a sphere. In most instances in practice, the approximation
works fine.

11.2.3 General distance

The distance metric used to construct weights as discussed in Section 11.3 can be readily
extended to a notion of general distance, familiar in regional science theory (Isard, 1969). In
addition, multidimensional scaling can be applied to create coordinates in a multivariate
2 The latitude is the y dimension, and the longitude the x dimension, so that the traditional reference to

the pair (lat, lon) actually pertains to the coordinates as (y,x) and not (x,y).
206 Distance-Based Spatial Weights

variable space (considered in Part II of Volume 2). These artificial locations can then be
used to derive spatial weights based on the distance between them.
In the empirical literature, such weights are often referred to as economic weights, based
on on a notion of economic distance. Simply put, the distance between observation i and j
expressed in terms of an economic (or social) variable z is:

dij = |zi − zj |α ,

where α is a scaling factor, often simply set equal to one.

A general distance measure readily serves as a proxy for the extent and strength of interactions
between locations (e.g., a social network among neighborhoods) by extending the one-
dimensional approach to a multivariate setting. Such a distance measure includes the
difference between two observations according to several dimensions, summarized as a single
value. Formally, with h = 1, . . . , H as the number of variables zh considered to characterize
the socio-economic makeup of each observation, the Euclidean distance between i and j is:

H

dij =  (zih − zjh )2 .
h=1

This approach is often used in economic applications.

An alternative way to derive distance-based weights in a multivariate setting is to employ
multidimensional scaling, or MDS (see Volume 2). This yields a point map of locations in
two-dimensional space, where the location and configuration of the points approximates their
distance relationship in multivariate dimensions. Loosely put, the two dimensions in the
MDS map correspond to the largest eigenvalues and matching eigenvectors of the correlation
matrix for the variables considered. This yields a best fit solution in the sense that the
relative distances between the points on the MDS map are the closest to the actual relative
distances between observations. The distances in MDS space can then be used to define
neighbors in multivariate attribute space.

11.3 Distance-Based Weights

With a definition of distance in hand, neighbors can be defined as those pairs of points that
are within a critical cut-off distance from each other. Alternatively, the k closest points,
or k-nearest neighbors can be selected. These approaches can be readily extended to more
general concepts of distance.

11.3.1 Distance-band weights

The most straightforward spatial weights matrix constructed from a distance measure is
obtained when i and j are considered neighbors whenever j falls within a critical distance
band from i. More precisely, wij = 1 when dij ≤ δ, and wij = 0 otherwise, where δ is a
preset critical distance cutoff.
In order to avoid isolates (islands) that would result from too stringent a critical distance,
the distance must be chosen such that each location has at least one neighbor. Such a
Distance-Based Weights 207

Figure 11.1: Nearest neighbor properties

distance conforms to a max-min criterion, i.e., it is the largest of the nearest neighbor
distances.3
In practice, the max-min criterion often leads to too many neighbors for locations that are
somewhat clustered, since the critical distance is determined by the points that are farthest
apart. This problem frequently occurs when the density of the points is uneven across the
data set, such as when some of the points are clustered and others are more spread out.
This is examined more closely in the illustrations.

11.3.2 K-nearest neighbor weights

With a full matrix of inter-point distances D, each i-th row contains the distances between
a point i and all other points j. With these distances sorted from smallest to largest, as d∗ij ,
a concept of k-nearest neighbors can be defined.
For each i, the k-nearest neighbors are those j with d∗ij for j = 1, k, i.e., the k closest
points (this assumes there are no ties, which is considered below). This is not a symmetric
relationship. Even though dij = dji the sorted order of distances for i is often not the same
as the sorted order for j.
Imagine the situation with three points on a line as A, B and C, with C closer to B than A is
to B. Then B will be the nearest neighbor for A, but A will not be the nearest neighbor for
B. This asymmetry can cause problems in analyses that depend on the intrinsic symmetry
of the weights (e.g., some algorithms to estimate spatial regression models).
Another potential issue with k-nearest neighbor weights is the occurrence of ties, i.e., when
more than one location j has the same distance from i. A number of solutions exist to break
the tie, from randomly selecting one of the k-th order neighbors, to including all of them. In
GeoDa, random selection is implemented.
To illustrate these properties, consider Figure 11.1, portraying six points. The nearest
neighbor relation is shown by an arrow from i to j. For example, E is clearly the nearest
neighbor of C, but C is not the nearest neighbor for E (that is F). The triangle D-A-B
3 The nearest neighbor distance is the smallest distance from a given point to all the other points, or, the

distance from a point to its nearest neighbor.

208 Distance-Based Spatial Weights

illustrates the problem of ties. D is equidistant from A and B, so it is impossible to pick

one or the other as the nearest neighbor. On the other hand, the nearest neighbor relation
between A and B is symmetric (arrow in both directions).
In practice, k-nearest neighbor weights often provide a solution to the problem of isolates,
since they ensure that each observation has exactly k neighbors (see Section 11.4.2). However,
this approach should not be applied uncritically. One drawback of the concept is that it
warps the effect of distance.
For sparsely distributed points, the nearest neighbor (and, by extension, the k nearest
neighbors) can be quite far removed, whereas for densely distributed points, many neighbors
can be found within a small distance band. In terms of interaction, this implies that the
actual distance is of less relevance than the fact that a neighbor is encountered, which is not
always an appropriate assumption. For example, when considering commuting distances, a
destination 40 miles away may be a reasonable neighbor, whereas there may be very little
interaction with a nearest neighbor that is 150 miles away.
Since the k-nearest neighbor relationship is typically asymmetric, it is sometimes desired to
convert this to symmetric spatial weights. One approach to accomplish this is to replace
the original weights matrix W by (W + W )/2, which is symmetric by construction.4 Each
new weight is then (wij + wji )/2. The resulting matrix will have more neighbors than the
original knn weights.
An alternative approach is to only use the instances where wij = wji = 1. This is referred to
as mutual symmetry. The result will be a much sparser weights matrix, with the possibility
of isolates.

11.3.3 General distance weights

With a general concept of distance in attribute space, the implementation of distance-band
and k-nearest neighbor weights is straightforward. However, when using a distance band, the
meaning of the implied max-min distance is not always obvious, since it is a combination of
distances among multiple variable dimensions.
An alternative weights construction is sometimes considered as the inverse of the generalized
distance:
wij = 1/|zi − zj |α = |zi − zj |−α ,

or, in standardized form:

|zi − zj |−α
wij =  −α
.
j |zi − zj |

For example, in the much cited study by Case et al. (1993), two separate weights matrices
were constructed using economic weights. One was based on the difference in per capita
income between states and the other on the difference in the proportion of their Black
population. In another nice example, Greenbaum (2002) uses the concept of a similarity
matrix to characterize school districts that are similar in per-student income. Instead of
using those weights as such, they are combined with a nearest neighbor distance cut-off,
so that the similarity weight is only computed when the school districts are also within a
given order of geographic nearest neighbors (see also the discussion of weights intersection
in Section 11.6.1).
Weights created as functions of distance are further considered in Chapter 12.
4 W is the transpose of the weights matrix W, such that rows of the original matrix become columns in

the transpose.
Implementation 209

Figure 11.2: Italian community bank point layer

11.4 Implementation
As mentioned in the chapter introduction, the distance-based weights functionality is
illustrated with a point layer of 261 Italian community banks, shown in Figure 11.2.5
Distance weights are invoked in the same way as contiguity weights, through the Weights
Manager, as outlined in Section 10.3.1. The Weights File Creation dialog with the
Distance Weight button selected is shown in Figure 11.3. The ID Variable is set to idd.
The dialog controls the various options available to create distance-based weights. These
include the choice between geographic distance – the Geometric centroids button under
Variables – or generalized distance – the Variables button (see Section 11.4.4).
The default for geographic distances is to have the X-coordinate variable as the internally
obtained X-Centroids, and the Y-coordinate variable as the internally obtained Y-
Centroids. However, any other pair of variables can be selected, including non-geographic
ones. The resulting concept of general distance is limited to two dimensions, whereas the
Variables tab allows for higher dimensions as well.
5 The map is customized by selecting Fill Color for Category and Outline Color for Category as

red (instead of the default green), changing the Point Radius to 1, and using Stamen > Toner Lite as
the Basemap (with Transparency set to 30).
210 Distance-Based Spatial Weights

Figure 11.3: Distance-based weights in the Weights File Creation interface

Figure 11.4: Default distance-based weights summary properties

The default Distance metric is Euclidean Distance. Other options, appropriate for
unprojected coordinates, are Arc Distance (mi) and Arc Distance (km), for great circle
distance expressed as either miles or kilometers.
Implementation 211

Figure 11.5: GWT weights file format

The type of weight is set by the Method tab. The Distance band criterion is the default.
Other options are K-Nearest neighbors and Kernel weights.6 The default bandwidth
for distance band weights is the max-min distance. In the current example, this is listed as
about 125 km (the units are meters). The options for inverse distance and Power are
considered in Chapter 12.

11.4.1 Distance-band weights

With the default settings as in Figure 11.3, invoking Create will bring up the familiar
file save dialog. As was suggested for contiguity weights, it is useful to make the file name
reflect the nature of the weights since there are no metadata, unless a project file is created.
In this example, the file name is italy_banks_te_d with the GWT file extension added
automatically.
Once the file is created, it appears in the Weights Manager, as shown in Figure 11.4.
The summary properties indicate the type as threshold, with threshold value as 124660,
distance metric as Euclidean, distance vars as coordinates and distance unit as
meter. The other properties are the same as for contiguity weights (see Section 10.4.1).
Note how the range of the neighbor cardinality is much larger than for contiguity, going
from 1 to 95. This reflects the impact of the max-min default cut-off distances on densely
distributed points. The distance of about 125 km is much larger than the average distance
that separates the community banks.
6 Kernel weights are considered in Chapter 12.
212 Distance-Based Spatial Weights

Figure 11.6: Connectivity histogram for default distance weights

11.4.1.1 GWT file

Distance-based weights are saved in files with a GWT file extension. This format, illustrated
in Figure 11.5, is slightly different from the GAL format used for contiguity weights. It was
first introduced in SpaceStat in 1995, and later adopted by R spdep and other software
packages. The header line is the same as for GAL files, but each pair of neighbors is listed,
with the ID of the observation, the ID of its neighbor and the distance that separates them.
The pairwise distance is currently only included for informational purposes, since GeoDa
does not use the actual distance value in any statistical operations. In other words,
the weights are turned into binary form and row-standardized by default. The only place
where the distances are taken into account is in the computation of distance functions,
covered in Chapter 12.
Note how observation with idd=1 has no less than 28 neighbors, with inter-point distances
ranging from slightly over 20 km to almost 125 km.

11.4.1.2 Weights characteristics

In the same way as for contiguity weights, the characteristics of distance-based weights
can be assessed by means of the Connectivity Histogram, the Connectivity Map and
the Connectivity Graph. These are available through the buttons at the bottom of the
weights manager.
The shape of the connectivity histogram for distance-band weights is typically very different
from that of contiguity-based weights. As illustrated in Figure 11.6, the range in the neighbor
cardinality is much larger than for contiguity weights. The distribution also includes extremes,
with one observation having only one neighbor, and one having 95.
As mentioned, the range in the number of neighbors is directly related to the density of
the spatial distribution of the points. Locations that are somewhat isolated will drive the
determination of the largest nearest neighbor cut-off point (their nearest neighbor distance
Implementation 213

Figure 11.7: Connectivity graph for default distance weights

will be large), whereas dense clusters of locations will encompass many neighbors using this
large cut-off distance.
The properties of the distance band weights can be further investigated by means of the
Connectivity Graph. As before, this is invoked through the right-most button at the
bottom of the weights manager.
The pattern shown in Figure 11.7 highlights how the connectivity is highly uneven, with
very dense areas alternating with much sparser distributions. The connection highlighted in
blue corresponds to the max-min distance. Clearly, this distance is much larger than the
average distance among locations, leading to a highly unbalanced weights matrix.

11.4.2 Isolates
The default distance band ensures that each observation has at least one neighbor, but this
has several undesirable side effects. To create a more balanced weights matrix, one can lower
the distance threshold. However, this will result in observations that do not have neighbors,
isolates or islands.
The Weights File Creation dialog is flexible enough that a specific distance cut-off can
be entered in the box, or the movable button can be dragged to any value larger than the
minimum distance, but smaller than the max-min distance. Sometimes, theoretical or policy
considerations suggest a specific value for the cut-off that may be smaller than the max-min
distance.
When such a threshold is chosen, a warning will appear, pointing out that the specified
cut-off value is smaller than the distance needed to ensure that each observation has at least
one neighbor.
214 Distance-Based Spatial Weights

Figure 11.8: Weights summary properties for distance 73 km

For example, using a cut-off distance of 73 km instead of the default 125 km yields a much
sparser weights matrix (italy_banks_te_d73.gwt), as evidenced by the summary in Figure
11.8. The maximum number of neighbors has been reduced from 95 to 59 (which is still
quite large), but, more importantly, the min neighbors is listed as 0. Compared to the
default, the density has decreased substantially, going from 15.16% non-zero elements to
7.87% non-zero.
The connectivity graph shown in Figure 11.9 illustrates the greater sparsity. As it turns out,
the distance band of 73 km resulted in only two isolates, highlighted in the figure, one near
Aosta, in the north-west, and a second on the island of Sicily. All other observations are
connected. However, there begin to form disconnected entities, i.e., separate components in
the graph that are connected internally, but not between them.
Further lowering the cut-off distance may result in more meaningful groupings, but at the
expense of a growing number of isolates.

11.4.2.1 How to deal with isolates

Since the isolated observations are not included in the spatial weights (in effect, the corre-
sponding row in the spatial weights matrix consists of zeros), they are not accounted for in
any spatial analysis, such as tests for spatial autocorrelation, or spatial regression. For all
practical purposes, they should be removed from such analysis. However, they are fine to be
included in a traditional non-spatial data analysis.
Ignoring isolates may cause problems in the calculation of spatially lagged variables, or
measures of local spatial autocorrelation. By construction, the spatially lagged variable will
be zero, which may suggest spurious correlations.
Implementation 215

Figure 11.9: Connectivity graph for distance band 73 km

11.4.3 K-nearest neighbor weights

K-nearest neighbor (KNN) weights are computed by selecting the corresponding button
in Distance Weight panel of the Weights File Creation interface. The value for the
Number of neighbors (k) is the only option. The default is 4, but 6 is used in the example
(as in Algeri et al., 2022).
The weights (saved as the file italy_banks_te_k6.gwt) are added to the collection contained
in the weights manager. In addition, all its properties are listed, as illustrated in Figure
11.10. Note that the properties now include the number of neighbors instead of the
distance threshold value, as was the case of distance-band weights. Also, symmetry is set
to asymmetric, which is a fundamental difference with distance-band weights.
The properties listed in the weights manager also include the mean and median number of
neighbors, which of course equal k (in the example, they equal 6). The resulting weights
matrix is dramatically less dense relative to the distance-band weights (2.30% compared to
15.16%).
K-nearest neighbors weights are often an alternative when a realistic distance band results in
too many isolates, since each observation has k neighbors by design. However, this approach
is not without some drawbacks and may not be appropriate in all situations.

11.4.3.1 Weights characteristics

Again, the connectivity histogram, map and graph can be used to inspect the neighbor
characteristics of the observations. However, in this case, the histogram doesn’t make much
sense, since all observations have the same number of neighbors (by construction).
216 Distance-Based Spatial Weights

Figure 11.10: Weights summary properties for 6 k-nearest neighbors

In contrast, the connectivity graph, shown in Figure 11.11, clearly demonstrates how each
point is connected to six other points. In the example, this yields a fully connected graph.

11.4.3.2 Caveat
One drawback of the k-nearest neighbor approach is that it ignores the distances involved. The
first k neighbors are selected, irrespective of how near or how far they may be. This suggests
a notion of distance decay that is not absolute, but relative, in the sense of intervening
opportunities (e.g., one considers the two closest grocery stores, irrespective of how far they
may be).
In addition, the requirement to have exactly k neighbors may create artificial connections,
spanning large distances or bridging across natural and other barriers. For example, in the
graph in Figure 11.11, the blue lines highlight some instances where this is the case. In the
case of the northern isolate, the Aosta location becomes connected to observations located
near Milan. For an observation on the north coast of Sicily, two long connections are included
across the straight of Messina to the boot of Italy. A similar distant neighbor is identified
for a location in Calabria to a far away one in the region of Basilicata. It is unlikely that
these connections correspond with meaningful interactions among the community banks.
The relative distance effect should be kept in mind before mechanically applying a k-nearest
neighbor criterion.

11.4.4 General distance matrix

The Variables tab in the distance weights dialog (Figure 11.3) also allows for the compu-
tation of a general weights matrix. This is based on the distance between observations in
Broadening the Concept of Contiguity 217

Figure 11.11: Connectivity graph for 6 k-nearest neighbors

multi-attribute space. Instead of selections for the X and Y coordinates, a drop-down list
with all the variables is available. From this, any number of variables can be selected, not
just two, as in the standard distance case.
In addition, there is a Transformation option that allows the usual standardizations (see
Section 2.4.2.3). The distance metric can be either Euclidean Distance or Manhattan
Distance (great circle distance is not meaningful in this context).
Typically, the variables are first standardized, such that their mean is zero and variance one,
which is the default setting.
All the standard options are available, such as specifying a distance band (in multi-attribute
distance units) or k-nearest neighbors. Such a general distance matrix or dissimilarity matrix
is a required input in the multivariate clustering methods considered in Volume 2.

11.5 Broadening the Concept of Contiguity

The concept of contiguity can be generalized to point layers by converting the latter to a
tessellation, such as Thiessen polygons (see Section 3.3). Queen or rook contiguity weights
can then be created for these polygons, in the usual way.
Similarly, the concepts of distance-band weights and k-nearest neighbor weights can be
generalized to polygon layers by considering their centroids (see Section 3.3.1). The polygons
are represented by their central points and the standard distance computations are applied.
218 Distance-Based Spatial Weights

Figure 11.12: Contiguity for points from Thiessen polygons

Figure 11.13: Weights summary properties for point contiguity (Thiessen polygons)

These operations can be carried out explicitly, by actually creating a separate Thiessen
polygon layer or centroid point layer, and subsequently applying the weights operations.
However, in GeoDa, this is not necessary, since the computations happen under the hood. In
this way, it is possible to create contiguity weights for points or distance weights for polygons
directly in the Weights File Creation dialog.

11.5.1 Contiguity-based weights for points

To illustrate the concept of contiguity weights for a point layer, consider the layout of six
points depicted in Figure 11.12. The Thiessen polygons are drawn as solid lines. This clearly
Broadening the Concept of Contiguity 219

Figure 11.14: Connectivity histogram for point contiguity (Thiessen polygons)

demonstrates how they are perpendicular to the dashed lines that connect the points. The
latter correspond to a queen contiguity relation.
Contiguity weights for a point layer are created in the same way as such weights for polygon
layers. In the Weights File Creation dialog, one selects the Contiguity Weight option
and specifies the type of contiguity. For the Italian community bank example, the summary
properties of a queen contiguity (italy_banks_te_q.gal) are shown in Figure 11.13.7
Compared to the other weights for this point layer, the mean neighbors is 5.78 and the
median neighbors is 6, similar to the k-nearest neighbor results in Figure 11.10. In addition,
the resulting weights are still a bit less dense (or, sparser), with a % non-zero of 2.21%
(compared to 2.30% for k-nearest neighbors).

11.5.1.1 Weights characteristics

These attractive characteristics are also reflected in the connectivity histogram, shown in
Figure 11.14. The graph is nicely balanced with a mode of 5-6, near equal to mean and
median. There is only a slight skewness to the right.
While, at first sight, the queen contiguity weights derived from the Thiessen polygons seem
to have attractive properties, there is also a serious drawback. As depicted in Figure 11.15,
the Thiessen polygons create several artificial contiguities, making connections between point
locations that are totally unrealistic. For example, the connectivity graph shows several
links between locations on the island of Sicily and the mainland, crossing the Tyrrhenian
Sea. Similarly, a location in Trieste (bordering Slovenia) is connected to several points on
the other side of the Adriatic Sea. It is unlikely that these links represent actual interaction
between the community banks.
A potential solution to this problem is to combine the contiguity weights with other criteria,
such as a meaningful distance band, by means of a set operation like intersection (see Section
11.6.1).
7 Note that the file extension is GAL, as for other contiguity weights.
220 Distance-Based Spatial Weights

Figure 11.15: Connectivity graph for point contiguity (Thiessen polygons)

11.5.2 Distance-based weights for polygons

As discussed in Section 3.3.1, a polygon layer has a series of Shape Center options to
add the centroid or mean center information to the data table, display those points on the
map, or save them as a separate point layer. Such points can then be used to construct
distance-based weights in the usual way.
However, as is the case for contiguity weights for points, in GeoDa this operation is carried
out invisibly, so that it is not necessary to create a separate layer. Distance-band weights
and k-nearest neighbor weights are obtained in the same way as for any other point layer
(see Section 11.4).

11.5.3 Block weights

A slightly different concept of neighbor follows when a block structure is imposed, in which all
observations in the same block are considered to be neighbors. This is a form of hierarchical
spatial model, in which all units that share a common higher order level are contiguous.
Familiar examples are counties as part of a state or province, census tracts belonging to a
larger neighborhood, etc. In some contexts, the block structure corresponds to the notion of
local interaction, in the sense of reflecting the average behavior of a reference group.8
The adoption of a block structure became more common in spatial analysis after its applica-
tion in a study of innovation diffusion by Case (1991).9 In her study, individual observations
8 See, for example, Brock and Durlauf (2001), for a formal discussion.
9 In some literature, block weights are referred to as Case weights.
Broadening the Concept of Contiguity 221

Figure 11.16: Block weights in the Weights File Creation interface

Figure 11.17: Connectivity graph for block weights by region

on farmers were grouped into districts, and all the farms within the same district were
considered to be neighbors. However, the neighbor relation did not extend across districts.
In other words, farms that may be physically adjoining but in different districts would not
be considered neighbors, in effect turning each district into an island.
The result of this approach is a block-diagonal spatial weights structure. Within each region
or district, all observations are neighbors (row elements of 1, except for the diagonal), but
there is no neighbor relation between the regions, hence the block-diagonal structure.
222 Distance-Based Spatial Weights

Figure 11.18: Space-time weights GAL file

The block weights structure has important consequences for the properties of estimators
in spatial regression models, but is less useful in an exploratory context, unless used in
combination with other criteria.10

11.5.3.1 Weights characteristics

Block weights are invoked as an option of the Contiguity Weight tab in the Weights
File Creation dialog. A drop-down list shows potential variables that could be used to
define the blocks, as illustrated in Figure 11.16 for the Italian community bank example. In
this instance, the variable REGCODE is selected, which provides an integer value for each
of the 20 regions in Italy.
The resulting spatial weights have quite different characteristics. The matrix is not sparse
(10.28% non-zero), but not as dense as some distance-band weights. The number of neighbors
ranges from 0 (a single observation in a region) to 62. The corresponding graph structure,
shown together with the regional boundaries in Figure 11.17, reveals several separate
components, with a strong internal connection, but no links between them.11 The sole
community bank in the Aosta region is an isolate.

11.5.4 Space-time weights

A special type of block weights is created as an option of the Time Editor. As mentioned in
Section 9.2.1.2, the Save Space-Time Table/Weights button in the interface of Figure
9.2 can generate both a pooled data set as well as a weights file. The latter requires that
there is an active weights file in the Weights Manager.
The resulting file follows a GAL format and repeats the cross-sectional contiguity structure
for each time period, by using a new space-time index, STID. This index matches the values
contained in the corresponding space-time data table. It is a sequential value going from 1
to n × T , where n is the cross-sectional dimension and T is the time dimension.
To illustrate this feature, consider the Oaxaca time-enabled data set from Section 9.2.1 with
a first-order queen contiguity weights active (oaxaca_q.gal). The contents of the space-time
GAL file are illustrated in Figure 11.18.
On the left side, the header line shows how the space-time ID variable, STID takes on values
from 1 to 1710 (570 cross-sectional units × 3 time periods). The same contiguity structure
repeats itself every 570 observations. In the figure, observation 1 has 6 neighbors, with IDs
10 Fora technical discussion of the spatial econometric aspects, see, e.g., Anselin and Arribas-Bel (2013).
11 The figure was created by combining the point layer with a polygon layer for the 20 regions, regions.shp,
using the multilayer functionality (see Section 3.6).
Manipulating Weights 223

362, 298, 264, 218, 178 and 177. On the right hand side, observation 571, which corresponds
to the first cross-sectional unit in the second time period, has the same six neighbors, but
all ID values are incremented by 570. For example, the first neighbor, with STID 932, is
obtained as 362 + 570.
The resulting space-time weights take on a block-diagonal structure in the sense that the
same cross-sectional block repeats itself for each time period. Formally, this corresponds to
the matrix expression I ⊗ W, where I is a T -dimensional identity matrix and W are the
cross-sectional spatial weights.12
The specialized weights structure can be used in a pooled analysis, where the space-time
aspect of the data is hidden. The observations are represented as one large cross-section,
with variables stacked by time period, and spatial weights block-diagonal by time period.
Consequently, all time periods are assumed to have the same parameters, such as the same
spatial autocorrelation coefficient.

11.6 Manipulating Weights

The information in spatial weights can be manipulated in a limited number of ways. Set
operations allow the creation of a new set of weights by union or intersection operations
on two existing weights files. In addition, asymmetric weights, typically k-nearest neighbor
weights, can be made symmetric.

11.6.1 Set operations

Once multiple weights are included in the Weights Manager, it becomes possible to carry
out set operations on any pair. With two weights selected, say italy_banks_te_d73 (for a
distance band of 73 km) and italy_banks_te_k6 (for 6 k-nearest neighbors), the buttons
Intersection and Union become active (placed directly above Property in the dialog).
Figure 11.19 shows the graph structure that results when these two weights are combined
through an Intersection operation.13 The only links that remain are those where some of
the six nearest neighbors fall within 73 km. The new structure illustrates how some of the
drawbacks of the individual weights were remedied.
The very dense connections in the north, shown in Figure 11.9, become much sparser, since
only six of the distance-band neighbors are retained. On the other hand, the artificial
connections in the k-nearest neighbor graph in Figure 11.11 are removed by imposing a 73
km distance constraint. This yields the same two isolates as for the original distance-band,
highlighted in the figure.
Nevertheless, the resulting compromise provides a better reflection of potential interaction
pairs for the community banks.
The Union function operates in a similar way (again, two weights need to be selected to
activate this functionality). An obvious application would be to combine pure first- and
second-order contiguity weights, although this functionality is built-in through the Include
lower orders option for contiguity weights.
12 The symbol ⊗ stands for a Kronecker product. This means that each element of the first matrix is

multiplied by the second matrix. In other words, since the first matrix is a T × T identity matrix, the result
is a block diagonal matrix with a block of W for each time period.
13 This is the operational implementation used by Algeri et al. (2022), with the exception that the two

isolates were removed in the analysis.

224 Distance-Based Spatial Weights

Figure 11.19: Connectivity graph for intersection distance 73 km and knn 6

11.6.2 Make symmetric

As discussed in Section 11.3.2, there are two straightforward ways to turn an asymmetric
k-nearest neighbor matrix into a symmetric one. One approach is inclusive and identifies
a link as long as either wij or wji is non-zero. This is the default solution activated when
the Make Symmetric button is selected in the Weights Manager. As a result, the new
symmetric weights will be denser than the original k-nearest neighbor weights.
The second approach is exclusive in that a link is only counted when it exists in both
directions: wij = wji = 1. This is activated when the mutual box is checked, next to
the Make Symmetric button. The resulting symmetric weights will be sparser than the
original k-nearest neighbor weights.
For example, when applied to the 6 k-nearest neighbor weights for the Italian community
banks, the inclusive symmetric weights have neighbors ranging from 6 to 13 (clearly, the
minimum will be 6, as in the original weights, but the maximum can be more than 6). The
median number of neighbors has increased to 7, with an overall density of 2.8 /%.
In contrast, the exclusive criterion yields weights with 0 to 6 neighbors (in this case, 6 is the
maximum), and thus creates some isolates. The median number of neighbors has decreased
to 5, with an overall density of 1.73 /%.
Which of these transformation is appropriate, will depend on each specific empirical cir-
cumstance. As a rule, a careful consideration of the weights characteristics is strongly
recommended, especially by means of the connectivity graph.
Finally, it is important to save the created weights in a Project File, so that metadata can
be preserved in later analyses.
12
Special Weights Operations

In this third chapter devoted to spatial weights, I consider some more specialized topics,
in the sense that they are less frequently encountered in empirical practice. There are two
main subjects.
First, I discuss two situations where the actual values for the spatial weights take on a
special meaning. So far, only the presence or absence of a neighbor relation has been taken
into account. In this chapter, this is generalized to weights that are transformations of the
pairwise distances between neighbors, i.e., inverse distance functions and kernel weights.
The resulting weights files primarily provide the basis for creating new spatially explicit
variables for use in further analyses, such as in spatial regression specifications.1 The actual
value of the weights themselves is not used in measures of spatial autocorrelation or other
exploratory analyses in GeoDa. As mentioned before, only the existence of a neighbor relation
is taken into account.
There are two important applications for these spatial transformations. One pertains to the
construction of so-called spatially lagged variables. Such variables are used in the exploration
of spatial autocorrelation and in spatial regression analysis. In a second set of applications,
the spatial weights are included as part of rate smoothing operations, as an extension of the
methods covered in Section 6.4.
To illustrate these techniques, I continue to use the point layer from the Italy Community
Banks sample data set and the Oaxaca Development polygon layer.

12.1 Topics Covered

• Compute inverse distance functions
• Compute kernel weights functions
• Assess the characteristics of weights based on distance functions
• Understand the contents of KWT format weights files
• Create a spatially lagged variable as an average or sum of the neighbors
• Create a spatially lagged variable as a window sum or average
• Create a spatially lagged variable based on inverse distance weights
• Create a spatially lagged variable based on kernel weights
• Compute and map spatially smoothed rates
• Compute and map spatial Empirical Bayes smoothed rates

1 The distance functions in GeoDa provide an alternative and more user-friendly way to calculate the

weights included in PySAL and GeoDaSpace (see Anselin and Rey, 2014, for details).

DOI: 10.1201/9781003274919-12 225

226 Special Weights Operations

GeoDa Functions
• Weight File Creation dialog
– inverse distance weights
– kernel weights
– bandwidth options
– diagonal element options
• Table > Calculator > Spatial Lag
– select spatial weights
– row-standardized weights or not
– include diagonal or not
• Map > Rates-Calculated Map
– Spatial Rate
– Spatial Empirical Bayes
• Table > Calculator > Rates
– Spatial Rate
– Spatial Empirical Bayes

12.2 Spatial Weights as Distance Functions

In the creation of distance band or k-nearest neighbor weights, the pairwise distance between
neighbors is included in the GWT file (see Figure 11.5). However, these values are not used
in any actual operations.
The distances themselves are not that useful, but transformations of distance lead to weights
that enable spatial transformations. Two types of such functions are considered here: inverse
distance weights and kernel weights.

12.2.1 Inverse distance weights

Spatial weights based on a distance cut-off can be viewed as representing a step function,
with a value of 1 for neighbors with dij < δ, and a value of 0 for others. As before, dij stands
for the distance between observations i and j, and δ is the bandwidth.
A straightforward extension of this principle is to specify the spatial weights as a continuous
parameterized function of distance itself:

wij = f (dij , θ), (12.1)

with f as a functional form and θ a vector of parameters.

In order to conform to Tobler’s first law of geography, a distance decay effect must be
respected. In other words, the value of the function of distance needs to decrease with
a growing distance. More formally, this is expressed by the requirement that the partial
derivative of the distance function with respect to distance should be negative:

∂wij /∂dij < 0.

Commonly used distance functions are the inverse, with wij = 1/dα
ij (and α as a parameter),
−βdij
and the negative exponential, with wij = e (and β as a parameter). The functions are
Spatial Weights as Distance Functions 227

typically combined with a distance cut-off criterion, such that for dij > δ, it follows that
wij = 0.
In practice, the parameters are often not estimated, but instead set to a fixed value, such
as α = 1 for inverse distance weights (1/dij ), and α = 2 for gravity weights (1/d2ij ). By
convention, the diagonal elements of the spatial weights are set to zero and not computed.
In fact, if the actual value of dii = 0 were used in the inverse distance function, this would
result in a division by zero.
An often overlooked feature in the computation of distance functions is that the resulting
values depend not only on the parameter value and functional form but also on the metric
and scale used for the distance measure. Since the weights are inversely related to distance,
large values for the latter will yield small values for the former, and vice versa. This may be
a problem in practice when the distances are so large (i.e., measured in small units) that the
corresponding inverse distance weights become close to zero, possibly resulting in an all zero
spatial weights matrix. This issue is directly related to the scale of the coordinates. It is a
common concern in practice when a projection yields distances expressed in feet or meters.
In addition, a potential problem may occur when the distance metric is such that distances
take on values less than one. As a consequence, some inverse distance values may be larger
than one, which is typically not a desired result.
A simple rescaling of the coordinates on which the distance computation is based will fix
both problems.

12.2.1.1 Implementation
Inverse distance weights are specified through the Weights File Creation interface, with
the Distance Weight tab selected, as in Figure 11.3. The check box next to Use inverse
distance? enables this functionality. This also activates the Power drop-down list, where a
coefficient can be specified, the default being 1.
The remainder of the interface works in the same fashion as before, with a query for a file
name for the new weights.
Using the Italy Community Banks sample data with the default settings yields a spatial
weights file like italy_banks_te_id1, the summary properties of which are shown in Figure
12.1. The descriptive statistics are identical to those for the distance band weights in Figure
11.4. The only differences are that inverse distance takes on the value of true, and the
power is listed as 1.
By design, GeoDa only takes into account the existence of a neighbor relation. The actual
values of the weights are ignored. Consequently, the properties of any distance function
(irrespective of the power) will be identical to those of the corresponding simple weights for
the same distance band. Also, the connectivity histogram, connectivity map and connectivity
graph will be the same as before. They do not provide any new information.
Inverse distance weights can be calculated for any bandwidth specified. While the default is
the max-min bandwidth, in some applications this is not useful. For example, the bandwidth
can be set to the maximum inter-point distance. As a result, the calculations will be for a
full n × n distance matrix. This is not recommended for larger data sets, but it can provide
a useful point of departure to compute accessibility indices formulated as spatially lagged
variables (see Section 12.3.1).
In addition, the coordinate option is perfectly general, and any two variables contained
in the data set can be specified as X and Y coordinates in the distance band setup. For
228 Special Weights Operations

Figure 12.1: Summary properties of inverse distance weights

example, this allows for the computation of so-called socio-economic weights, where the
inverse operation is applied to the Euclidean distance computed in the attribute domain
between two locations, using any two (but only two) variables.
In contrast, with the Variables tab checked in the Weights File Creation interface, fully
multivariate inverse distance weights can be computed as well, for as many variables as
required. However, for such distances to be meaningful, one has to be mindful of the scale in
which the variables are expressed. Standardization is highly recommended.

12.2.1.2 K-nearest neighbor inverse distance weights

The computation of inverse distance weights extends to the k-nearest neighbor design in a
straightforward manner. The only difference is that the K-Nearest neighbors tab must
be checked. All the operations are the same as for distance band weights.
For example, using the original coordinates with k=6, this yields a GWT file,
e.g., italy_banks_te_idk6.gwt, and with the coordinates in kilometers, a file
italy_banks_te_idk6km.gwt.

12.2.1.3 GWT file

The results of an inverse distance calculation are written to a GWT file, which has the
same structure as before. However, in contrast to the standard case, the third column now
contains the value for the inverse distance rather than the distance itself.
Spatial Weights as Distance Functions 229

Figure 12.2: GWT files with distance for original and rescaled coordinates

To illustrate this, consider the two GWT files just created, using 6 nearest neighbors as the
bandwidth. The left-hand panel of Figure 12.2 shows the outcome when using the default
coordinates (italy_banks_te_idk6.gwt). Note how the value of 2.01549423e-05 for the pair
1-95 is exactly the inverse of the distance of 49,615.6221 (meters), listed on the matching
line of Figure 11.5.
This illustrates the problem just mentioned. Since the projected coordinates for the Italian
bank point file are expressed in meters, the distances are fairly large (e.g., almost 50 km
in the example). As a consequence, the values obtained for inverse distance are very small.
This has undesirable effects for the computation of spatial transformations like a spatially
lagged variable.
The right-hand panel in Figure 12.2 shows the result after the coordinates were rescaled to
kilometers.2 The outcome is simply multiplied by 1,000.

12.2.2 Kernel weights

Kernel smoothing is a core element in nonparametric statistics (Silverman, 1986; Racine,
2019). It is essentially a weighted average of observations, centered on a focal point, with
decreasing weights as the distance to the focal point increases. The magnitude of the weights
is determined by a kernel function, i.e., a decreasing function of the distance to the focal
point (in attribute space), and a bandwidth, i.e., the range within which observations are
weighted.
In spatial analysis, kernel weights are probably best known for their inclusion in geographically
weighted regression, or GWR (Fotheringham et al., 2002). However, they are also an essential
part of non-parametric approaches to model spatial covariance (Hall and Patil, 1994),
considered more closely in Chapter 13.3
More precisely, kernel weights are defined as a function K(z) of the ratio between the
distance dij from i to j, and the bandwidth hi , with z = dij /hi . This ensures that z is
always less than 1. For distances greater than the bandwidth, K(z) = 0.
Two important considerations in the use of kernel weights are the selection of the kernel
function and the determination of the bandwidth.

12.2.2.1 Specification of the kernel weights

Kernel functions determine how quickly the weights get smaller as the distance to the focal
point increases. In GeoDa, there currently is support for five commonly used kernel weights
2 In GeoDa, this is implemented in the Calculator. Assuming that COORD_X and COORD_Y are

the coordinates in meters, a new set of variables XKM and YKM are computed by dividing these values
by 1,000. The new XKM and YKM then need to be entered as respectively X_coordinate variable and
Y_coordinate variable in the Weights File Creation interface.
3 An application in spatial econometrics is the heteroskedasticity and autocorrelation consistent (HAC)

estimation of the regression error variance, due to Kelejian and Prucha (2007). See Anselin and Rey (2014),
for implementation details in GeoDaSpace and PySal.
230 Special Weights Operations

functions:
• Uniform, K(z) = 1/2 for |z| < 1,
• Triangular, K(z) = (1 − |z|) for |z| < 1,
• Quadratic or Epanechnikov, K(z) = (3/4)(1 − z 2 ) for |z| < 1,4
• Quartic, K(z) = (15/16)(1 − z 2 )2 for |z| < 1 and
• Gaussian. K(z) = (2π)(−1/2) exp(−z 2 /2).5
Typically, the value for the diagonal elements of the weights is set to 1, although sometimes
the actual kernel value is used. For example, for a quadratic of Epanechnikov kernel function,
the weight for dij = 0 would in principle be 3/4, although in practice a value of 1 is often
used as well.
Many careful decisions must be made in selecting a kernel function. Apart from the choice
of a functional form for K( ), a crucial aspect is the selection of the bandwidth. In the
literature, the latter is found to be more important than the functional form. Several rules
have been developed to determine an optimal bandwidth (e.g., see Silverman, 1986), although
in practice, trial and error may be more appropriate.
An important choice is whether the bandwidth should be the same for all observations, a
so-called fixed bandwidth, or instead allowed by vary, as a variable bandwidth.
A drawback of fixed bandwidth kernel weights is that the number of non-zero weights can
differ considerably from one observation to the next, especially when the density of the point
locations is not uniform throughout space. This is the same problem encountered for the
distance band spatial weights (Section 11.3.1).
GeoDa includes two different implementations of the concept of fixed bandwidths. One is the
max-min distance used earlier (the largest of the nearest-neighbor distances). The other is
the maximum distance for a given specification of k-nearest neighbors. For example, for a
specified number of nearest neighbors, this is the distance between the k-nearest neighbors
pairs that are the farthest apart. In other words, this is an extension of the max-min concept
to the specific context of k neighbors, rather than nearest neighbor pairs.
To correct for the issues associated with a fixed bandwidth, a variable bandwidth approach can
be taken. It adjusts the bandwidth for each location to ensure equal or near-equal coverage.
One common approach is to take the k-nearest neighbors, and to adjust the bandwidth
for each location such that exactly k neighbors are included in the kernel function. The
bandwidth specific to each location is then any distance larger than its k nearest neighbor
distance, but less than the k+1 nearest neighbor distance.
When the kernel weights are implemented for the nearest neighbor concept, a final decision
pertains to the value of k. In GeoDa, the default value for k equals the cube root of the
number of observations.6 In general, a wider bandwidth gives smoother and more robust
results, so the bandwidth should always be set at least as large as the recommended default.
Some experimentation may be needed to find a suitable combination of bandwidth and/or
number of nearest neighbors in each particular application.
4 Note that the Epanechnikov kernel is sometimes referred to without the (3/4) scaling factor. GeoDa

implements the scaling factor.

5 While the Gaussian kernel is in principle without a bandwidth constraint, in GeoDa it is implemented

with the same bandwidth option as the other kernel functions.

6 This is based on the recommendation by Kelejian and Prucha (2007) in the context of HAC estimation

and typically forms a good starting point

Spatial Weights as Distance Functions 231

Figure 12.3: Kernel weights in Weights File Creation interface

12.2.2.2 Creating kernel weights

Kernel weights are created with the Distance Weight tab in the Weights File Creation
interface. The Method must be specified as Kernel, as shown in Figure 12.3. In this
illustration for the Italian community bank example, the coordinates have been converted to
kilometers, specified as XKM and YKM.
The Kernel functionality brings up several options, as illustrated in the figure. The Ker-
nel function drop-down list contains the five supported functions. In the example, the
Epanechnikov kernel is selected. Furthermore, the actual weights are used for the diagonal
elements, with the radio button checked next to Apply kernel to diagonal weights. The
default is actually to have Diagonal weights = 1.
The next set of options pertains to the bandwidth. The default max-min distance is listed in
the Specify bandwidth box as 124.659981 km. In the example, the Adaptive bandwidth
is checked, with 7 neighbors. A final option is to instead specify a fixed bandwidth with
max knn distance as bandwidth.
The suggested value of k in the interface is based on the cube root criterion. In the example,
this is the cube root of 261, or 6.39, rounded up to the next integer (7). Note that the
convention used in GeoDa is to count the observation itself as one of k, so that in fact only
k-1 true nearest neighbors are taken into account.
The result is saved in a file with a KWT extension, e.g., italy_banks_te_epa.kwt, for
consistency with the approach taken in the PySAL library. Except for the inclusion of the
diagonal element, its structure is the same as a GWT format file.

12.2.2.3 Characteristics of kernel weights

As soon as the weights are created, their properties appear in the weights manager. As
illustrated in Figure 12.4, the descriptive statistics are again the same as for standard knn
232 Special Weights Operations

Figure 12.4: Properties of kernel weights

weights. The differences are in the first six items. The type of weights is given as kernel,
the kernel method is identified as Epanechnikov, with the bandwidth definition (knn
7) and adaptive kernel set to true. It is also indicated that the kernel is applied to the
diagonal elements, since kernel to diagonal is true. Also, as for the k-nearest neighbor
weights, the resulting weights are asymmetric.
Since the connectivity histogram, map and graph ignore the actual weights values and are
solely based on the implied connectivity structure, they are identical to those obtained for
the corresponding k-nearest neighbor weights. Consequently, they are not informative with
respect to the properties of the weights themselves.

12.2.2.4 KWT file

The contents of the KWT file in our example are shown in the right-hand panel of Figure
12.5, compared to the knn distances in the corresponding GWT file on the left (based on
coordinates expressed as kilometers).
A few characteristics of the results should be noted. First, the bandwidth is determined by
the largest distance among the seven neighbors. In the current example, for observation
1, the panel on the left shows the distance to the seventh neighbor (249) as 54.251 km.
This is the bandwidth. As a result, the kernel weight value for the link 1 to 249 will be 0
(right hand panel). The value for z, to be used in the kernel function, is therefor obtained
by dividing the pairwise distances by 54.251. For example, for the link between 1 and
95, who are 49.615 km apart, this yields 0.915. The value for the Epanechnikov kernel is
0.75(1 − z 2 ) = 0.75(1 − 0.836) = 075 × 0.164 = 0.123, the result listed in the third column of
the matching row on the right-hand side.
There are eight entries for observation 1: the seven neighbors as well as the diagonal. For
the latter, listed as the pair 1 to 1, the value is 0.75.
Spatial Transformations 233

Figure 12.5: KWT file

12.3 Spatial Transformations

The main form of spatial transformation considered here is the so-called spatial lag operation.
This summarizes the observations at neighboring locations into a single new variable. The
weights matrix is central in this operation, since it identifies which observations need to be
selected and if any weights need to be applied.

12.3.1 Spatially lagged variables

With a neighbor structure defined by the non-zero elements of the spatial weights matrix
W, a spatially lagged variable is a weighted sum or a weighted average of the neighboring
values for that variable (Anselin, 1988). In the typical notation, the spatial lag of y is then
expressed as W y.
Formally, for observation i, the spatial lag of yi , referred to as [W y]i (the variable W y
observed for location i) is:

[W y]i = wi1 y1 + wi2 y2 + · · · + win yn ,

or,

n
[W y]i = wij yj ,
j=1

where the weights wij consist of the elements of the i-th row of the matrix W, matched up
with the corresponding elements of the vector y.
In other words, the spatial lag is a weighted sum of the values observed at neighboring
locations, since the non-neighbors are not included (those j for which wij = 0). Typically,
the weights matrix is very sparse, so that only a small
 number of neighbors contribute to the
weighted sum. For row-standardized weights, with j wij = 1, the spatially lagged variable
becomes a weighted average of the values at neighboring observations.
In matrix notation, the spatial lag expression corresponds to the matrix product of the n × n
spatial weights matrix W with the n × 1 vector of observations y, or W × y. The matrix
W can therefore be considered to be the spatial lag operator on the vector y.
In a number of applied situations, it may be useful to include the observation at location i
itself in the weights computation. This implies that the diagonal elements of the weights
matrix must be non-zero, i.e., wii = 0. Depending on the context, the diagonal elements
may take on the value of one or equal a specific value, such as for kernel weights where the
kernel function is applied to the diagonal.
234 Special Weights Operations

Figure 12.6: Spatial lag in calculator

An alternative concept of spatial lag is the so-called median spatial lag. Instead of computing
a (weighted) average of the neighboring observations, the median is selected.

12.3.1.1 Creating a spatially lagged variable

The spatial lag computation is part of the Calculator functionality associated with a data
table (Section 2.4.2). It is started from the menu as Table > Calculator, or by right
clicking in a table to bring up the list of options.
The lag operation requires the Spatial Lag tab to be active. The Weight drop-down list
contains all the spatial weights available to the project, with the currently active weights
listed. In the example illustrated in Figure 12.6, a time-enabled version of the Italian bank
data is used, with the 6 k-nearest neighbor weights after intersection with the 73 km distance
band. The spatial weights are contained in the file italy_banks_te_d73_k6.gal (see Section
11.6.1). The spatial lag operator will be applied to the variable SERV (2016), i.e., the time
enabled ratio of net interest income over total operating revenues in 2016.
To compute the spatially lagged variable, first Add Variable is selected to create a name
for the new variable, W1_SERV16 in the example. At this point, the expression for the
calculation of the lag is filled out in the box, listing both the weights and the original
variable.
The default option is to Use row-standardized weights without including the diagonal
elements. The result is added to the data table, as shown in Figure 12.7, in the column
W1_SERV16.
The file italy_banks_te_d73_k6.gal contains the neighbor list for the first observation, with
idd = 1. These are the observations with idd equal to, respectively, 64, 95, 96, 267, 272
and 309 (corresponding to table rows 15, 41, 42, 167, 172 and 209). The value of SERV
(2016) for the first observation is listed as 0.640208. A look up in the data table reveals
the values for the neighboring observations as 0.560816, 0.748250, 0.659608, 0.615197,
0.349620 and 0.701551. The average of these six values is 0.605840, which is exactly the
entry on the first row of the W1_SERV16 column.
As illustrated in Section 11.6.1, the spatial weights intersection yields two isolates, observa-
tions with idd 269 (Valley of Aosta, row 169) and 287 (Sicily, row 187). Since a spatially
lagged variable is undefined for an isolate (an actual computation would yield a value of
zero), the corresponding entries in the table are empty. This avoids complications from
mistakenly associating a value of zero in further calculations, such as in the computation
Spatial Transformations 235

Figure 12.7: Spatially lagged variables added to table

of a spatial autocorrelation coefficient. The outcome of the empty entry is that the isolate
observations are excluded from such computations.
The effect of the spatial lag operation is a form of smoothing of the original data. The mean
is roughly the same (the theoretic expected values are the same), e.g., 0.6363 for the original
variable vs. 0.6378 for the spatial lag, but the standard deviation is much smaller, i.e., 0.0943
for the original vs. 0.0562 for the spatial lag. Similarly, a form of shrinking has occurred:
the original variable has a minimum of 0.3496 and a maximum of 0.9143, whereas for the
spatial lag this has narrowed to 0.5124 to 0.7478.7
A measure of spatial correlation can be computed as the correlation between the original
and the spatially lagged variable. Note that this is NOT the accepted notion of spatial
autocorrelation considered in the literature (see Chapter 13). Nevertheless, as long as it is
not interpreted in the context of a spatial autoregressive model, it remains a useful measure
of the linear association between a variable and the average of its neighbors. In the example,
this value is 0.350.8

12.3.1.2 Spatial lag as a sum of neighboring values

Combining different settings for the row-standardized and Include diagonal options
allows for other notions of spatially lagged variables to be implemented. For example, a
spatial lag as a sum of the neighboring values is obtained by unsetting all options.
The result is listed in column W2_SERV16 of the table (Figure 12.7). For the first
observation, the corresponding value is 3.635042, which is the sum of the values for the six
neighbors.

12.3.1.3 Median spatial lag

The Median spatial lag is obtained by checking the corresponding box (the two other
options become unavailable). In the example, there are six neighbors, so the median is
computed as the midpoint between the third (0.615197) and fourth (0.659608) values in
the ranked set, yielding 0.637403. This is the entry in the first row of the W3_SERV16
column in Figure 12.7.
7 The
statistics can be readily obtained from a box plot of each variable.
8 The
correlation can be calculated wit h the Data > View Standardized Data in a scatter plot of
SERV (2016) on W1_SERV16.
236 Special Weights Operations

12.3.1.4 Spatial window transformations

So far, the spatial lag operations did not include the observations itself. However, by checking
the Include diagonal of weights matrix option, an average or sum over all observations
in the window can be computed. The window consists of the observation and its neighbors,
a notion commonly used in spatial smoothing operations (see also Section 12.4).
The spatial window average is obtained by checking the Use row-standardized weights
option. The result is an average of the seven values (observation and six neighbors), or
0.610750, the entry in the first row of the W4_SERV16 column in Figure 12.7.
Unchecking the row-standardized box yields the spatial window sum. In the example, this
is 4.275251, listed under column W5_SERV16. Clearly, this is equivalent to the addition
of SERV (2016) and W2_SERV16.

12.3.2 Using inverse distance weights – potential measure

The spatial lag operation can also be applied using spatial weights calculated from the
inverse distance between observations. This and kernel weights (Section 12.3.3) are the only
two instances in GeoDa where the actual values of the weights are used in an operation.
As mentioned in Section 12.2.1, the magnitude of the inverse distance weights depends on
the units in which the coordinates are expressed. Therefore, care is needed before using
inverse distance weights in spatial transformations. When the coordinates are expressed in
feet or meters, as is often the case, the corresponding distances will tend to be very large.
Consequently, the inverse distances may take on very small values, and any spatial lag could
end up being essentially zero.
Formally, the spatial lag operation uses the same principle as before. It amounts to a weighted
average of the neighboring values, with the inverse distance function as the weights:
 yj
[W y]i = ,
j
dα
ij

In practice, this calculation is typically combined with the application of a bandwidth. Such
a bandwidth can be expressed in the form of a distance-band weight, with wij = 0 for
dij > δ, with δ as the bandwidth. Formally:
 wij yj
[W y]i = ,
j
dα
ij

As mentioned, α is either 1 or 2. This expression for the spatial lag corresponds to the
concept of potential, familiar in geo-marketing analyses.9 The potential is a measure of how
accessible location i is to opportunities located in the neighboring locations (as defined by
the weights).

12.3.2.1 Default inverse distance weights setting

Spatial lags constructed from inverse distance weights operate in the same way as outlined
in Section 12.3.1.1. The default setting is that both Use row-standardized weights and
Include diagonal of weights matrix are unchecked. This yields the potential measure
as the sum of the neighboring values weighted by the inverse distance (or its square).
9 The concept of population potential goes back to the early literature in social physics, e.g., Stewart

(1947). See also Harris (1954) and Isard (1960) for extensive early discussions.
Spatial Transformations 237

Figure 12.8: Variables computed with inverse distance weights added to table

Using the Italian bank example, with the inverse distance weights computed from the
original coordinates (in meters) with a 6 nearest neighbor bandwidth requires the file
italy_banks_te_idk6.gwt. The result is shown in the column ID_SERV16 of Figure 12.8.
The small values illustrate the problem associated with the original weights.
Instead, with the inverse distances computed from coordinates in kilometers, the result
in column IDK_SERV16 follow (using the weights file italy_banks_te_idk6km.gwt),
essentially a rescaled version of ID_SERV16.
For example, for observation with idd=1, the inverse distance weights are given in the
right-hand panel of Figure 12.2. Combining these weights with the observation values for
the neighbors listed in Section 12.3.1.1 yields the value 0.093700 given in the first row of
the IDK_SERV16 column.

12.3.2.2 Row-standardized inverse distance weights

The original inverse distance weights are highly scale dependent. This can be remedied by
expressing them in row-standardized form. The spatial lag then takes on the standard meaning
of a weighted average of the values at neighboring observations. The main difference with
lags computed for connectivity weights is that now the neighbors are weighted differently. In
contrast, for the straightforward spatial lag calculation of Section 12.3.1.1, all the neighboring
values get the same weight.
The result of applying the row-standardized inverse distance weights from the right-hand
panel in Figure 12.2 yields 0.564334, listed on the first row of the IDR_SERV16 column.
In contrast to the calculation for the unstandardized inverse weights, the result is now
similar in scale to the original variable (and similar to the spatial lags based on connectivity
weights).

12.3.2.3 Spatial window inverse distance sum

Similar to the operations in Section 12.3.1.4, the inverse distance spatial lag can be computed
for all observations in a spatial window. This is accomplished by checking the Include
diagonal of weights matrix option.
However, unlike what is the case for the simple spatial lag, there is a complication on how to
deal with the diagonal value. In the simple spatial lag calculation this value is either added
238 Special Weights Operations

to the neighboring observations, or averaged with them. In the inverse distance case, the
neighbors are weighted, but it is not obvious what the weight should be for the diagonal,
since a distance of zero does not result in a valid weight.
The convention taken in GeoDa is to keep the diagonal value unweighted. For example,
without row-standardization, this amounts to:
 yj
[W y]i = yi + .
j
dα
ij

Again, this can be implemented in combination with a specific bandwidth for the distance.
In some contexts, this may be the desired result, but it is by no means the most intuitive
concept. It should therefore be used sparingly and only when there is a strong substantive
motivation. The corresponding result is listed in the first row of the IDS_SERV16 column.
It is simply the sum of the value for SERV (2016) and IDK_SERV16.
The same problem occurs when using row-standardized weights. Again, in GeoDa the diagonal
is unweighted, which yields: 
[W y]i = yi + wij yj .
j

where wij are the row-standardized inverse distance weights.

12.3.3 Using kernel weights

Spatially lagged variables can also be computed from kernel weights. However, in this
instance, only one of the options with respect to row-standardization and diagonal weights
makes sense. Since the kernel weights are the result of a specific kernel function, they should
not be altered. Also, each kernel function results in a specific value for the diagonal element,
which should not be changed. As a result, the only viable option to create spatially lagged
variables based on kernel weights is to have no row-standardization and to have the diagonal
elements included. Other options are not available in the interface.
Kernel-based spatially lagged variables correspond to a form of local smoothing. They can
be used in specialized regression specifications, such as geographically weighted regression.
The resulting spatially lagged variable is

[W y]i = Kij yj ,
j

When kernel weights are used, the concept of a bandwidth is already implemented in the
calculation of the kernel.
For the example using Epanechnikov weights with 6 nearest neighbors
(italy_banks_te_epa.kwt), the result is listed on the first row of the column EPA_SERV16.
The value of 1.508157 can be obtained by combining the weights in the right-hand panel
of Figure 12.5 with the neighboring observations from Section 12.3.1.1).
Creating spatially lagged variables and their use in spatial rate smoothing are the only
operations in GeoDa where kernel weights can be applied. They are not available for the
analysis of spatial autocorrelation.
Spatial Weights in Rate Smoothing 239

Figure 12.9: Disability rate: raw rate and EB smoothed rate

12.4 Spatial Weights in Rate Smoothing

In the discussion of rate smoothing in Chapter 6, two approaches were left until after
spatial weights were introduced. Spatial rate smoothing and spatial Empirical Bayes
smoothing are extensions of, respectively, the raw rate map and Empirical Bayes smoothing.
The spatial versions are distinct in that they incorporate values for neighboring observations
into the numerator and denominator of the rate calculation. This is made operational through
the use of spatially lagged variables.
These methods will be illustrated with the municipal disability rate in 2020 from the Oaxaca
Development sample data set. In addition to variables, five different spatial weights matrices
will be employed: first-order queen contiguity (oaxaca_q.gal); inclusive second-order queen
contiguity (oaxaca_q2.gal); nearest neighbor weights, with k=6 (oaxaca_k6.gwt); inverse
distance weights with a 6 nearest neighbor bandwidth (oaxaca_idk6.gwt); and block weights
based on the region classification (oaxaca_reg.gal).

12.4.0.1 Oaxaca disability rate

The disability rate is obtained as the ratio of the number of disabled persons (DIS20) over
the population (PTOT20) in 2020.
This rate provides a prime example of variance instability (see Section 6.4.1), due to the
wide range of population totals for municipalities, including several with very small values.
In 2020, the municipal population for the 570 municipalities ranged from 81 to 270,955,
and 47 locations had fewer than 500 inhabitants. The average population was 7,249, but
the median was much smaller, at 2,881, reflecting the effect of upper outliers (there are 62
such outliers, with populations ranging from 14,198 to 270,955). This will yield substantial
variation in the precision of the computed raw rates.
To provide some reference for the later illustrations, Figure 12.9 includes the raw rate map
(Section 6.2.2) in the left-hand panel and the Empirical Bayes (EB) smoothed rate map
(Section 6.4.2.2) in the right-hand panel.
The effect of the smoothing is to eliminate a few upper outliers in the map. Otherwise, the
relative ranking of observations in quartiles is not much affected. However, there is a real
240 Special Weights Operations

impact on the actual values. The range is reduced from 0.0149 to 0.3161 for the raw rate, to
a much narrower 0.0174 to 0.2164 for the smoothed rate. The correlation between the two
rates is 0.984.

12.4.1 Spatial rate smoothing

A spatial rate smoother is a special case of a nonparametric rate estimator, based on the
principle of locally weighted estimation (see, e.g., Waller and Gotway, 2004, pp. 89–90).
Rather than applying a local average to the rate itself, as would be the case in an application
of a spatial window average, the weighted average is applied separately to the numerator
and denominator.
The spatially smoothed rate for a given location i is then given as:
n
j=1 wij Oj
πi = n ,
j=1 wij Pj

where Oj is the event count in location j, Pj is the population at risk and wij are the spatial
weights. In contrast to most procedures covered so far, here the weights typically include
the diagonal element as well, i.e., wii = 0.
Different smoothing functions are obtained for different spatial definitions of neighbors
and/or different weights applied to those neighbors, such as contiguity weights, inverse
distance weights, or kernel weights.
An early application of this principle was the spatial rate smoother outlined in Kafadar
(1996), based on the notion of a spatial moving average or window average (see also Kafadar,
1997). The window average is not applied to the rate itself, but it is computed separately for
the numerator and denominator.
The simplest case boils down to applying the idea of a spatial window sum (Section 12.3.1.4)
to the numerator and denominator (i.e., with binary spatial weights in both, and including
the diagonal term):
Ji
Oi + j=1 wij Oj
πi = Ji ,
Pi + j=1 wij Pj
A map of spatially smoothed rates tends to emphasize broad spatial trends and is useful for
identifying general features of the data. However, it is not suitable for the analysis of spatial
autocorrelation, since the smoothed rates are autocorrelated by construction. It is also not
very insightful for identifying outlying observations, since the values portrayed are regional
averages and not specific to an individual location. By construction, the values shown for
individual locations are determined by both the events and the population sizes of adjoining
spatial units, which can lead to misleading impressions. More specifically, the counts and
populations of small areas are dwarfed by those or larger surrounding areas. Therefore,
inverse distance weights are sometimes applied to both the numerator and denominator, e.g.,
as in the early discussion by Kafadar (1996). In GeoDa this is not implemented in the rates
calculation, but can be carried out manually in the data table (see Section 12.4.1.2).

12.4.1.1 Spatially smoothed rate map

A spatially smoothed rate map can be created from the list of map options by selecting
Rates > Spatial Rate. Alternatively, from the menu, it is obtained as Map > Rates-
Calculated Map > Spatial Rate.
Spatial Weights in Rate Smoothing 241

Figure 12.10: Spatial rate map: first- and second-order queen contiguity

The map construction follows the same approach as for other rate maps (see Section 6.2.2).
The numerator in the ratio is set as the Event Variable, here DIS20, and the denominator
as the Base Variable, here PTOT20. At the bottom of the variable selection dialog are two
drop-down lists. One contains different Map Themes and the other the spatial Weights.
In the example, two Box map are created, using first- and second-order queen contiguity,
oaxaca_q and oaxaca_q2. The left panel of Figure 12.10 illustrates the former, the right
panel the latter.
Compared to the maps in Figure 12.9, a lot of detail is lost, and the emphasis is on broader
regional patterns, which tend to be dominated by the data for the larger municipalities.
Increasing the range of the contiguity to second-order yields somewhat larger regions of
similar values (in the sense of falling in the same quartile).
The spatial rate smoothing yields an even greater reduction in the variable range than
the non-spatial EB. For first-order contiguity, the range becomes 0.0390 to 0.1370, and
for second-order contiguity 0.0468 to 0.1164. In contrast to EB rates, which remain highly
correlated with the original raw rate, this is no longer the case for spatially smoothed rates.
The averaging over spatial units breaks the connection with the original values somewhat.
In the current example, the correlation between raw rate and the two spatial rates is,
respectively, 0.504 for first-order and 0.468 for second-order contiguity. This should be kept
in mind when interpreting the results.
The spatially smoothed rate map retains all the standard options of a map in GeoDa, as
well as the special options of other rate maps (see Section 6.2.2). The default variable name
for the Save Rates option is R_SPAT_RT. Currently, there are no metadata that keep
track of the spatial weights used in the smoothing operation.

12.4.1.2 Spatially smoothed rate in table

As is the case for other rate calculations (see, e.g., Section 6.2.2.1), the rates can also be
computed by means of Table > Calculator. The only difference is that now there is also
a drop-down list to specify the spatial weights. All other operations are identical to the
non-spatial case.
As mentioned, sometimes the use of inverse distance or kernel weights is advocated to lessen
the effect of neighboring values on the computed spatial smoothed rate. While it may be
tempting to carry out the rate calculation using such spatial weights, this does not work
correctly in GeoDa. Figure 12.11 illustrates this issue. It portrays spatially smoothed rates
using two different spatial weights. The map on the left is based on k-nearest neighbor
242 Special Weights Operations

Figure 12.11: Spatial rate map: k-nearest neighbor and inverse distance weights

Figure 12.12: Spatial rate map: inverse distance weights from Calculator

weights, with k = 6, oaxaca_k6,10 the map on the right on inverse distance weights using
the 6 nearest neighbors as a variable bandwidth, oaxaca_idk6. The maps are identical
because only the contiguity information is used in the weights operations.
The proper way to compute spatially smoothed rates for inverse distance or kernel weights
is through the Calculator. A spatially lagged variables must be computed separately for
the numerator and denominator, using either inverse distance or kernel weights. These new
variables are then used to compute a ratio, as a division operation in the calculator. Figure
12.12 shows a box map with the result (the computed variable is ID_RATE). The pattern
clearly differs from that in the maps in Figure 12.11. The effect of large neighbors is lessened
because of the distance decay effect embedded in the inverse distance weights.

12.4.2 Spatial Empirical Bayes smoothing

The spatial Empirical Bayes rate smoothing method is based on the same principle as the
standard Empirical Bayes smoother, covered in Chapter 6. The main difference is that the
10 Since the coordinates for the Oaxaca data set are in decimal degrees latitude-longitude, the weights

were constructed using Arc Distance (km) option in the Weights File Creation dialog.
Spatial Weights in Rate Smoothing 243

Figure 12.13: Spatial Empirical Bayes smoothed rate map using region block weights

reference rate is computed for a spatial window, specific to each individual observation,
rather than taking the same overall reference rate for all observations. This approach only
works when the number of observations in the reference window, as defined by the spatial
weights, is large enough to allow for effective smoothing. This is a context where block
weights become very useful (Section 11.5.3), since they avoid the irregularity in the number
of neighbors on which the reference rate is based. All the observations in the same block
have an identical number of reference points from which the reference rate is computed.11
Similar to the standard EB principle, a reference rate (or prior) is computed from the data.
It is estimated from the spatial window surrounding a given observation.
Formally, the reference mean for location i is then:

j wij Oj
μi =  ,
j wij Pj

with wij as binary spatial weights, and wii = 1.

The local estimate of the prior variance follows the same logic as for EB, but replacing the
population and rates by their local counterparts:
 2
2 j wij [Pj (ri − μi ) ] μi
σi =  − .
w
j ij jP j ij i /(ki + 1)
w P

Note that the average population in the second term pertains to all locations within the
window, therefore, this is divided by ki + 1 (with ki as the number of neighbors of i). As
in the case of the standard EB rate, it is quite possible (and quite common) to obtain a
negative estimate for the local variance, in which case it is set to zero.
The spatial EB smoothed rate is computed as a weighted average of the crude rate and the
prior, in the same manner as for the standard EB rate (Section 6.4.2.2).

12.4.2.1 Spatial Empirical Bayes smoothed rate map

A spatial Empirical Bayes smoothed rate map is selected from the list of map options as
Rates > Spatial Empirical Bayes. Alternatively, from the menu, it is obtained from
11 Since the observation itself is included in the reference window, all observations in the same block will

have the same mean and variance for the prior.

244 Special Weights Operations

Map > Rates-Calculated Map > Spatial Empirical Bayes. The map is created in
exactly the same way as for the spatial rate map.
In the example, shown in Figure 12.13, block weights based on the region to which a
municipality belongs (oaxaca_reg) are used to define the reference window for each
observation. These weights have on average 95 neighbors, ranging from a minimum of 19 to
154. The left panel of the figure illustrates the regional division, whereas the smoothed rates
are mapped in the right panel.
While there are many similarities with the pattern for the standard EB rate, shown in Figure
12.9, there are also a few differences, such as an additional upper outlier and a number of
observations that shift from the first quartile to the second quartile, as well as from the
fourth quartile to the third. The correlation with the original EB rate is very high, at 0.996,
suggesting overall a great similarity between the two results.
All the usual options are in effect. The rate can be saved with a default variable name of
R_SPAT_EBS.

12.4.2.2 Spatial Empirical Bayes smoothed rate in table

Identical to the treatment of the other rates, a spatial Empirical Bayes smoothed rate can
be computed through the table Calculator option, as an item under the Rates tab.
 Part IV

Global Spatial Autocorrelation

Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
13
Spatial Autocorrelation

In this chapter, I begin the discussion of spatial autocorrelation, both in general as well
as focused on the special case of the Moran’s I statistic. I start off with some definitions,
more specifically the notion of spatial randomness as the null hypothesis, and positive and
negative spatial autocorrelation as the alternative hypotheses.
This is followed by an overview of the general concept of a spatial autocorrelation statistic,
with a brief discussion of several specific implementations. Then attention shifts to arguably
the best known such statistic, Moran’s I (Moran, 1948). The formal structure, inference and
interpretation of the statistic are covered, followed by its visualization through the so-called
Moran scatter plot (Anselin, 1996). The chapter closes with an outline of the implementation
of the Moran scatter plot in GeoDa.
To illustrate these concepts, I use the Chicago Community Areas sample data set with socio-
economic variables for the 77 community areas in the city of Chicago (from the American
Community Survey).

13.1 Topics Covered

• Understand spatial randomness
• Appreciate the difference between positive and negative spatial autocorrelation
• Understand the structure of a spatial autocorrelation statistic
• Moran’s I
• Visualize Moran’s I by means of the Moran scatter plot
• Carry out inference using the permutation approach
• Make analyses reproducible using the random seed setting
• Nonlinear LOWESS smooth of the Moran scatter plot
• Brush Moran scatter plot to assess regional Moran’s I
• Appreciate the difference between dynamic weights and static weights in Moran scatter
plot regime regression
GeoDa Functions
• Table > Calculator
– Univariate > Shuffle
• Space > Univariate Moran’s I
– permutation inference
– setting the random seed
– LOWESS smoother of the Moran scatter plot
– brushing the Moran scatter plot
– save results (standardized value and spatial lag)

DOI: 10.1201/9781003274919-13 247

248 Spatial Autocorrelation

Toolbar Icons

Figure 13.1: Moran Scatter Plot | Spatial Correlogram | Cluster Maps

13.2 Spatial Randomness

Spatial randomness is the point of departure in any statistical analysis of spatial pattern. It
is the null hypothesis against which the data are compared.
Spatial randomness is the absence of any systematic spatial structure in the data, i.e., there
are no distinct patterns. The location of a given observation is of no value as a piece of
information, such that the where does not matter. In other words, the notion of spatial
randomness is not very interesting other than as a statistical concept. The main goal of an
analysis is to reject the null hypothesis of spatial randomness in favor of the presence of
some type of pattern. This is accomplished by means of a test statistic.
Spatial randomness can be interpreted in two main ways, which correspond to different
approaches toward spatial analysis. In one, representing a simultaneous view, the observed
spatial pattern is viewed as equally likely as any other spatial pattern. More specifically, it is
the pattern as a whole that is considered. Another view represents a conditional perspective.
It implies that the value observed at a given location does not depend on the values observed
at neighboring locations. Again, this requires that location does not matter. Fundamentally,
the two notions are equivalent, but they result in different statistical approaches.
Before considering specific statistics (see Section 13.4), two perspectives toward making
spatial randomness operational are briefly outlined. One is based on a parametric assump-
tion, specifying observations as independent and identically distributed (i.i.d.). The other
perspective does not assume a particular distribution, although it still requires identical
distributions, especially a constant variance. In such an approach, the location of the ob-
servations may be altered without affecting the information content of the data. This is
referred to as randomization, in the sense that each observed value has the same probability
of being associated with any of the locations, i.e., the map is irrelevant.

13.2.1 Parametric approach

The parametric approach requires the assumption of a specific distribution. Typically, this
is the normal (or Gaussian) distribution, since it has the important property that lack of
correlation also implies independence. For other distributions, this is not the case.
The simplest implementation of spatial randomness is to take a collection of uncorrelated
(independent) standard normal variates. Since nothing spatial is assumed for this distribution,
the result is spatially random.
For example, Figure 13.2 illustrates the layout of 100 standard normal variates on a 10 × 10
square grid, using a quantile map with six categories. Even though one may be tempted
to see structure in the data, this would be purely coincidental, since the map is devoid of
pattern by construction.
Spatial Randomness 249

Figure 13.2: Spatially random observations – i.i.d

Figure 13.3: Spatially random observations – true and randomized per capita income

13.2.2 Randomization
The randomization approach does not require the selection of a specific parametric distribu-
tion for the data. Instead, it takes advantage of the property that under the null hypothesis,
each observation is equally likely to be at any location. By randomly permuting or reshuffling
the data across locations, the concept of spatial randomness can be simulated.
For example, the left hand panel of Figure 13.3 shows the spatial distribution of the income
per capita for the 77 Chicago community areas (from the 2015–2019 ACS) as a quantile map
with eight categories. On the right, the same 77 values are randomly allocated to community
areas, resulting in a totally different map. The true map illustrates the well-known pattern
of higher incomes in the northern part of the city, contrasting with low income in the south
and west. By contrast, the randomized map shows no such pattern. High income areas (dark
brown) are found all over the city, without any apparent structure to their locations.
Parenthetically, the contrast between the two maps illustrates the value of a spatial per-
spective. Taken as such, the statistical distribution of the original and randomized variables
are identical. In Figure 13.4, an eight category histogram highlights this property. In other
words, the a-spatial perspective on the data distribution offered by the histogram for the
two variables cannot account for the important distinction between their spatial patterns.
250 Spatial Autocorrelation

Figure 13.4: True and randomized per capita income – histograms

13.2.3 Simulating spatial randomness

The spatially randomized map in Figure 13.3 is obtained by creating a new variable, taking
advantage of the Calculator functionality in the table. Specifically, as listed in Section 2.4.2.2,
the Univariate operations include a Shuffle function, which implements randomization.
Operationally, the list of index numbers of the observations is randomly reshuffled and the
original observations are assigned to the new locations. For example, if the original list was
1, 2, 3, 4, 5, and the randomized list yields 3, 2, 4, 5, 1, then the value at location 1 will be
assigned to location 3, 2 will remain in place, 3 will move to spot 4, 4 to spot 5 and 5 to
location 1.
The randomization (or permutation) approach is an essential tool to create many versions of
the data under the null hypothesis. This is particularly useful, since, except in highly stylized
situations (e.g., standard normal distribution), it quickly becomes very difficult to obtain
the properties of any spatial autocorrelation statistic by means of an analytical derivation.

13.3 Positive and negative spatial autocorrelation

Rejecting the null hypothesis of spatial randomness suggests that there is some spatial
structure in the data. When the focus in on dependence (as opposed to heterogeneity), the
two alternative hypotheses are positive and negative spatial autocorrelation. Unlike what
holds for the familiar notion of correlation (without space), negative spatial autocorrelation
is not simply the negative of positive spatial autocorrelation, i.e., a negatively sloped linear
relation contrasted with a positively sloped on. Negative spatial autocorrelation yields a
totally different type of spatial pattern.

13.3.1 Positive spatial autocorrelation

Positive spatial autocorrelation implies that like values in neighboring locations occur
more frequently than they would under spatial randomness. This results in an impression
of clustering, the clumping in space of like values. These like values can be either high
(hot spots) or low (cold spots). The key aspect is attribute similarity among neighboring
locations.
Positive and negative spatial autocorrelation 251

Figure 13.5: Positive spatial autocorrelation

Figure 13.6: Negative spatial autocorrelation

Figure 13.5 illustrates strong positive spatial autocorrelation on a 10 × 10 regular grid.1

While the pattern may suggest some clustering, this is very difficult to assess by eye, especially
relative to the true spatially random pattern of Figure 13.2.

13.3.2 Negative Spatial Autocorrelation

Negative spatial autocorrelation is suggested when neighboring values are more dissimilar
than they would be under spatial randomness. This typically yields a checkerboard-like
pattern. The dissimilarity is often associated with spatial heterogeneity, although the latter
is not limited to neighboring units. Negative spatial autocorrelation implies alternating high
and low values, whereas spatial heterogeneity can pertain to larger regions that include
many neighbors.
Figure 13.6 illustrates how difficult it is to distinguish negative spatial autocorrelation
from spatial randomness (Figure 13.2). It portrays a strong degree of negative spatial
autocorrelation on a 10 × 10 regular grid.2 Whereas there are quite a few dark values
surrounded by lighter ones, this is by no means a perfect checkerboard.
1 The pattern is obtained by imposing a spatial autoregressive structure with a coefficient of 0.9 on 100

standard normal variates.

2 The pattern is obtained by imposing a negative spatial autoregressive structure with a coefficient of

−0.9 on 100 standard normal variates.

252 Spatial Autocorrelation

13.4 Spatial Autocorrelation Statistic

Relying on identifying spatial structure in the data by eye is bound to lead to spurious results,
especially since the human brain is wired to find patterns. Instead, a spatial autocorrelation
statistic is used.
A statistic is a summary measure computed from the data that allows one to assess the
extent to which a given null hypothesis holds. Typically, this is accomplished by deriving
the distribution of the statistic if the null hypothesis were true. If the observed value of the
statistic is extreme relative to this so-called null distribution (i.e., in the extreme tails of the
distribution), the null is rejected.
In classical statistics (as opposed to a Bayesian perspective), a fundamental notion in this
regard is the Type I error, or the probability that the null is rejected, when in fact it is
true. In other words, this is the chance one takes to make the wrong decision. A critical
value of the Type I error is the so-called significance or p-value, a target cut-off point. If
the probability that the value of the statistic (as computed from the data) follows the null
distribution is less than the p-value, one typically rejects the null hypothesis. In this context,
the choice of the p-value becomes crucial. Customary, this has often been set to 0.05, but
recently there has been a lot of discussion around this choice in the context of replicability
of scientific experiments. The current consensus is that a p-value should be much smaller
than 0.05, or even be replaced by alternative concepts (Efron and Hastie, 2016; Benjamin
and 72 others, 2018). The choice of a suitable p-value will be revisited in the discussion of
inference for local indicators of spatial association (LISA, Chapter 16).
Spatial autocorrelation is about the coincidence of attribute similarity (how attribute values
are alike) with locational similarity (how close observations are to each other). One aspect of
the spatial case that is distinct from a standard correlation coefficient is that the attribute
similarity pertains to the same variable at different locations, hence auto correlation. More
precisely, it is a univariate measure, in contrast to the bivariate correlation coefficient.
A spatial autocorrelation statistic is then a summary computed from the data that combines
these two notions. The objective is to assess the observed value of the statistic relative to
the distribution of values it would take on under the null hypothesis of spatial randomness.
Formally, such a statistic can be expressed in cross-product form as:

SA = f (xi , xj ).l(i, j),
i j

where the sum is over all pairs of observations, xi and xj are observations on the variable x
at locations i and j, f (xi , xj ) is a measure of attribute similarity and l(i, j) is a measure of
locational similarity between i and j.
Different choices for the functions f and l lead to specific spatial autocorrelation statistics.
A few common options are briefly reviewed next.

13.4.1 Attribute similarity

The measure of attribute similarity f (xi , xj ) is a summary of the similarity or dissimilarity
over all pairs of observations. Typical choices are the cross product, xi .xj , the squared
difference (xi − xj )2 , and the absolute difference, |xi − xj |. The cross product focuses on
Spatial Autocorrelation Statistic 253

similarity, with extreme large or extreme small values indicating greater similarity (i.e., the
product of adjoining large values, or the product of adjoining small values). In contrast, the
squared and absolute differences are measures of dissimilarity, with smaller values indicating
more alike neighbors.

13.4.2 Locational similarity

There are two main ways to incorporate locational similarity in spatial autocorrelation
statistics. One approach uses the spatial weights matrix, discussed at length in Chapters 10
through 12. The corresponding spatial autocorrelation statistics take the generic form:

SA = f (xi , xj ).wij ,
i j

i.e., the sum over all neighbors (i.e., pairs where wij = 0) of a measure of attribute similarity
between them. This approach is most appropriate when the observations pertain to a set of
discrete locations, a so-called lattice data structure.
A different perspective is taken when the observations are viewed as a sample from a
continuous surface, although this approach can also be applied to discrete locations. Instead
of subsuming the neighbor relations in the spatial autocorrelation statistic, a measure
of attribute similarity is expressed as a function of the distance that separates pairs of
observations:
f (xi , xj ) = g(dij ),
where g is a function of distance. The function can take on a specific form, as in a variogram
or semi-variogram, or be left unspecified, in a non-parametric approach. This form of
modeling is representative of geostatistical analysis, which concerns itself with modeling
and interpolation on spatial surfaces (Isaaks and Srivastava, 1989; Cressie, 1993; Chilès and
Delfiner, 1999; Stein, 1999). The discussion of this approach is deferred to Chapter 15.

13.4.3 Examples of spatial autocorrelation statistics

Each combination of a measure of attribute similarity with a spatial weights matrix yields a
different spatial autocorrelation statistic.
An early example of a very generic spatial autocorrelation statistic is the Mantel test (Mantel,
1967), popularized in spatial analysis through the work of Hubert and Golledge (Hubert
et al., 1981, 1985). It combines the elements of a pairwise (dis)similarity matrix A with a
spatial weights matrix W, yielding the Gamma statistic:

Γ= Aij .Wij .
i j

Some commonly used statistics can be viewed as special cases of this statistic, using a
particular expression for Aij .

13.4.3.1 Cross-product
Arguably the most commonly used spatial autocorrelation statistic is Moran’s I (Moran,
1948), based on a pairwise cross-product as the measure of similarity. Moran’s I is typically
expressed in deviations from the mean, using zi = xi − x̄, with x̄ as the mean. The full
expression is:  
j zi .zj .wij /S0
I=
i
 2 , (13.1)
i zi /n
254 Spatial Autocorrelation
 
with S0 = i j wij as the sum of the weights, and the other notation as before.
This statistic is examined more closely in Section 13.5.1.

13.4.3.2 Squared difference

Whereas squared difference is most commonly used as an attribute dissimilarity measure in
geostatistics, it is combined with spatial weights in Geary’s c statistic (Geary, 1954):
  2
j (xi − xj ) wij /2S0
c=
i
 .
zi2 /(n − 1)
Since the numerator consists of a squared difference, it is not necessary to take the deviations
from the mean (the two means would cancel each other out in the difference operation).
In contrast to Moran’s I, which takes values between roughly (not exactly) −1 and +1,
Geary’s c is always positive, with values between 0 and 2. The theoretical mean for Geary’s
c is 1. Because Geary’s c is based on a dissimilarity measure, positive spatial autocorre-
lation corresponds with small values for the statistic (less than 1), and negative spatial
autocorrelation with large values (larger than 1).

13.4.3.3 Absolute difference

Although used less frequently in practice, attribute dissimilarity can also be based on absolute
difference. This yields a slightly more robust measure in the sense that the influence of
outliers is lessened. An early discussion of this criterion was contained in Sokal (1979). It
will not be considered further.

13.5 Moran’s I
Moran’s I is arguably the most used spatial autocorrelation statistic in practice, especially
after its discussion in the classic works by Cliff and Ord (Cliff and Ord, 1973, 1981). In
this section, I briefly review its statistical properties and introduce an approach toward its
visualization by means of the Moran scatter plot (Anselin, 1996). The implementation of
the Moran scatter plot in GeoDa is further detailed in Section 13.6.

13.5.1 Statistic
Moran’s I statistic, given in Equation (13.1), is a cross-product statistic similar to a Pearson
correlation coefficient. However, the latter is a bivariate statistic, whereas Moran’s I is a
univariate statistic. Also, the magnitude for Moran’s I critically depends on the characteristics
of the spatial weights matrix W. This has as the unfamiliar consequence that, as such,
without standardization, the values of Moran’s I statistics are not directly comparable, unless
they are constructed with the same spatial weights. This contrasts with the interpretation
of the Pearson correlation coefficient, where a larger value implies stronger correlation. For
Moran’s I, this is not necessarily the case, depending on the structure of the spatial weights
involved.
The scaling factor for the denominator of the statistic is the familiar n, i.e., the total number
of observations (sample size), but the scaling factor for the numerator is unusual. The latter
is S0 , the sum of all the (non-zero) spatial weights. However, since the cross-product in the
Moran’s I 255

numerator is only counted when wij = 0, this turns out to be the proper adjustment. For
example, dividing by n2 would also account for all the zero cross-products, which would not
be appropriate.
For row-standardized weights, in the absence of unconnected observations (isolates), the
sum of weights equals the number of observations, S0 = n. As a result, the scaling factors in
numerator and denominator cancel out. Therefore, the statistic can be written concisely in
matrix form as:
z Wz
I=  ,
zz
with z as a n × 1 vector of observations as deviations from the mean.3

13.5.2 Inference
Inference is carried out by assessing whether the observed value for Moran’s I is compatible
with its distribution under the null hypothesis of spatial randomness. This can be approached
from either an analytical or a computational perspective.

13.5.2.1 Analytical inference

Analytical inference starts with assuming an uncorrelated distribution, which then allows
for asymptotic properties of the statistic to be derived. An exception is the result for an
exact (finite sample) distribution obtained by Tiefelsdorf and Boots (1995), under fairly
strict assumptions of normality. The statistic is treated as a special case of the ratio of
two quadratic forms of random variables. The resulting exact distribution depends on the
eigenvalues of the weights matrix.
In Cliff and Ord (1973) and Cliff and Ord (1981), the point of departure is either a Gaussian
distribution or an equal probability (randomization) assumption. In each framework, the
mean and variance of the statistic under the null distribution can be derived, which can
then be used to construct an approximation by a normal distribution. The derivation of
moments leads to some interesting and rather surprising results.
In both approaches, the theoretical mean of the statistic is:
1
E[I] = − .
n−1
In other words, under the null hypothesis of spatial randomness, the mean of Moran’s I is
not zero, but slightly negative relative to zero. Clearly, as n → ∞, the mean will approach
zero.
In the case of a Gaussian assumption, the second-order moment is obtained as:
n2 S1 − nS2 + 3S02
E[I 2 ] = ,
S02 (n2 − 1)
    
with S1 = (1/2) i j (wij + wji )2 , S2 = i ( j wij + j wji )2 and S0 as before.
Interestingly, the variance depends only on the elements of the spatial weights, irrespective
of the characteristics of the distribution of the variable under consideration, except that the
latter needs to be Gaussian. As a consequence, as long as the same spatial weights are used,
the variance computed for different variables will always be identical.
3 Expressed in this manner, the statistic takes on the same form as the familiar Durbin-Watson statistic

for serial correlation in regression residuals.

256 Spatial Autocorrelation

This property no longer holds under the randomization assumption. Now, the second-order
moment is obtained as:
n((n2 − 3n + 3)S1 − nS2 + 3S02 ) − b2 ((n2 − n)S1 − 2nS2 + 6S02 )
E[I 2 ] = ,
(n − 1)(n − 2)(n − 3)S02

with b2 = m4 /m22 , the ratio of the fourth moment of the variable over the second moment
squared.
In other words, under randomization, the variance of Moran’s I does depend on the moments
of the variable under consideration, which is a more standard result.
Actual inference is then based on a normal approximation of the distribution under the null.
Operationally, this is accomplished by transforming Moran’s I to a standardized z-variable,
by subtracting the theoretical mean and dividing by the theoretical standard deviation:

I − E[I]
z(I) = ,
SD[I]
with SD[I] as the theoretical standard deviation of Moran’s I. The probability of the result
is then obtained from a standard normal distribution.
Note that the approximation is asymptotic, i.e., in the limit, for an imaginary infinitely large
number of observations. It may therefore not work well in small samples.4

13.5.2.2 Permutation inference

An alternative to the analytical approximation of the distribution is to simulate spatial
randomness by permuting the observed values over the locations. The statistic is calculated
for each of such artificial and spatially random data sets. The total collection of such statistics
then forms a reference distribution, mimicking the null distribution.
This procedure maintains all the properties of the attribute distribution, but alters the
spatial layout of the observations.
The comparison of the Moran’s I statistic computed for the actual observations to the
reference distribution allows for the calculation of a pseudo p-value. This provides a summary
of the position of the statistic relative to the reference distribution. With m as the number
of times the computed Moran’s I from the simulated data equals or exceeds the actual value
out of R permutations, the pseudo p-value is obtained as:
m+1
p= ,
R+1
with the 1 accounting for the observed Moran’s I. In order to obtain nicely rounded figures,
R is typically taken to be a value such as 99, 999, etc.
The pseudo p-value is only a summary of the results from the reference distribution and
should not be interpreted as an analytical p-value. Most importantly, it should be kept
in mind that the extent of significance is determined in part by the number of random
permutations. More precisely, a result that has a p-value of 0.01 with 99 permutations is not
necessarily less significant than a result with a p-value of 0.001 with 999 permutations. In
each instance, the observed Moran’s I is so extreme that none of the results computed for
simulated data sets exceeds its value, hence m = 0.
4 An alternative alternative approximation based on the saddlepoint principle is given in Tiefelsdorf (2002),

but it is seldom used in practice.

Moran’s I 257

Figure 13.7: Reference distribution, Moran’s I for per capita income

Figure 13.7 illustrates a reference distribution based on 999 random permutations for the
Chicago community area income per capita, using queen contiguity for the spatial weights.
The green bar corresponds with the Moran’s I statistic of 0.6011. The dark bars form a
histogram for the statistics computed from the simulated data sets. The theoretical expected
value for Moran’s I depends only on the number of observations. In this example, n = 77,
such that E[I] = −1/76 = −0.0132. This compares to the mean of the reference distribution
of -0.0125. The standard deviation from the reference distribution is 0.0696, which yields
the z-value as:
0.6011 + 0.0125
z(I) = = 8.82,
0.0696
as listed at the bottom of the graph. Since none of the simulated values exceed 0.6011, the
pseudo p-value is 0.001.
In contrast, Figure 13.8 shows the reference distribution for Moran’s I computed from the
randomized per capita incomes from the right-hand panel in Figure 13.3. Moran’s I is -0.0470,
with a corresponding z-value of −0.4361, and a pseudo p-value of 0.357. In the figure, the
green bar for the observed Moran’s I is slightly to the left of the center of the graph. In other
words, it is indistinguishable from a statistic that was computed for a simulated spatially
random data set with the same attribute values.

13.5.3 Interpretation
The interpretation of Moran’s I is based on a combination of its sign and its significance.
More precisely, the statistic must be compared to its theoretical mean of −1/(n − 1). Values
larger than E[I] indicate positive spatial autocorrelation, values smaller suggest negative
spatial autocorrelation.
However, this interpretation is only meaningful when the statistic is also significant. Otherwise,
one cannot really distinguish the result from that of a spatially random pattern, and thus
the sign is meaningless.
258 Spatial Autocorrelation

Figure 13.8: Reference distribution, Moran’s I for randomized per capita income

It should be noted that Moran’s I is a statistic for global spatial autocorrelation. More
precisely, it is a single value that characterizes the complete spatial pattern. Therefore,
Moran’s I may indicate clustering, but it does not provide the location of the clusters. This
requires a local statistic.
Also, the indication of clustering is the property of the pattern, but it does not suggest
what process may have generated the pattern. In fact, different processes may result in the
same pattern. Without further information, it is impossible to infer what process may have
generated the pattern.
This is an example of the inverse problem (also called the identification problem in econo-
metrics). In the spatial context, it boils down to the failure to distinguish between true
contagion and apparent contagion. True contagion yields a clustered pattern as a result
of spatial interaction, such as peer effects and diffusion. In contrast, apparent contagion
provides clustering as the result of spatial heterogeneity, in the sense that different spatial
structures create local similarity.
This is an important limitation of cross-sectional spatial analysis. It can only be remedied by
exploiting external information, or sometimes by using both temporal and spatial dimensions.

13.5.3.1 Comparison of Moran’s I results

In contrast to a Pearson correlation coefficient, the value of Moran’s I is not meaningful
in and of itself. Its interpretation and the comparison of results among variables depend
critically on the spatial weights used. Only when the same weights are used are results
among variables directly comparable. To avoid this problem, a reliable comparison should
be based on the z-values.
For example, using queen contiguity, Moran’s I is 0.601 for the per capita income, but
0.472 using k-nearest neighbor weights, with k=8. This would suggest that the spatial
correlation is stronger for queen weights than for knn weights. However, when considering
the corresponding z-values, this turns out not to be the case. For queen weights, the z-value
Moran’s I 259

Figure 13.9: Moran scatter plot, community area per capita income

is 8.82, whereas for k-nearest neighbor weights, it is 9.46, instead suggesting the latter shows
stronger spatial correlation.
Comparison of Moran’s I among different variables is only valid when the same weights are
used. For example, percent population change between 2020 and 2010 yields a Moran’s I of
0.410 using queen contiguity, with a corresponding z-value of 6.30. In this instance, there is
consistent evidence that income per capita shows a stronger pattern of spatial clustering.

13.5.4 Moran scatter plot

The Moran scatter plot, first outlined in Anselin (1996), consists of a plot with the spatially
lagged variable on the y-axis and the original variable on the x-axis. The slope of the linear
fit to the scatter plot equals Moran’s I.
As pointed out, with row-standardized weights, the sum of all the weights (S0 ) equals the
number of observations (n). As a result, the expression for Moran’s I simplifies to:
   
j wij zi .zj i (zi × wij zj )
I=
i
 2 =  j2 ,
i zi i zi

with z in deviations from the mean.


Upon closer examination, this expression turns out to be the slope of a regression of j wij zj
on zi .5 This becomes the principle underlying the Moran scatter plot. While the plot provides
5 In a bivariate linear regression y = α + βx, the least squares estimate for β is
 
(x × yi )/ i x2i . In
 i i
the Moran scatter plot, the role of y is taken by the spatial lag j
wij zj .
260 Spatial Autocorrelation

Figure 13.10: Positive spatial autocorrelation, high-high

a way to visualize the magnitude of Moran’s I, inference should still be based on a permutation
approach, and not on a traditional t-test to assess the significance of the regression line.
Figure 13.9 illustrates the Moran scatter plot for the income per capita data for Chicago
community areas (based on queen contiguity). The value for income per capita is shown
in standard deviational units on the horizontal axis, with mean at zero. The vertical axis
corresponds to the spatial lag of these standard deviational units. While its mean is close to
zero, it is not actually zero. Hence the horizontal dashed line is typically not exactly at zero.
The slope of the linear fit, in this case 0.611, is listed at the top of the graph.

13.5.4.1 Categories of spatial autocorrelation

An important aspect of the visualization in the Moran scatter plot is the classification of
the nature of spatial autocorrelation into four categories. Since the plot is centered on the
mean (of zero), all points to the right of the mean have zi > 0 and all points to the left have
zi < 0. These values can be referred to respectively as high and low, in the limited sense of
higher or lower than average. Similarly, the values for the spatial lag can be classified above
and below the mean as high and low.
The scatter plot is then easily decomposed into four quadrants. The upper-right quadrant
and the lower-left quadrant correspond to positive spatial autocorrelation (similar values at
neighboring locations). They are referred to as respectively high-high and low-low spatial
autocorrelation. In contrast, the lower-right and upper-left quadrant correspond to negative
spatial autocorrelation (dissimilar values at neighboring locations). Those are referred to as
respectively high-low and low-high spatial autocorrelation.
The classification of the spatial autocorrelation into four types begins to make the connection
between global and local spatial autocorrelation. However, it is important to keep in mind
that the classification as such does not imply significance. This is further explored in the
discussion of local indicators of spatial association in Chapter 16.
Figure 13.10 illustrates the high-high quadrant for the income per capita Moran scatter plot.
The points selected in the left-hand graph are linked to their locations in the map on the
right. Clearly, these tend to be locations in the upper quartile (as well as upper outliers)
surrounded by locations with a similar classification.
Negative spatial autocorrelation of the low-high type is illustrated in Figure 13.11. The
community area selected in the map on the right is in the second quartile, surrounded by
Visualizing spatial autocorrelation with the Moran scatter plot 261

Figure 13.11: Negative spatial autocorrelation, low-high

neighbors that are mostly of a higher quartile. One neighbor in particular belongs to the
upper outliers, which will pull up the value for the spatial lag. The corresponding selection
in the scatter plot on the left is indeed in the low-high quadrant.
One issue to keep in mind when assessing the relative position of the value at a location
relative to its neighbors is that a box map such as Figure 13.3 reduces the distribution to
six discrete categories. These are based on quartiles, but the range between the lowest and
highest observation in a given category can be quite large. In other words, it is not always
intuitive to assess the high-low or low-high relationship from the categories in the map.

13.6 Visualizing spatial autocorrelation with the Moran scatter

plot
In GeoDa, a Moran scatter plot is created from the menu as Space > Univariate Moran’s
I. Alternatively, it can be invoked through the left most icon in the spatial subset of the
toolbar, highlighted in Figure 13.1, followed by selecting Univariate Moran’s I from the
drop-down list.

13.6.1 Creating a Moran scatter plot

The first requirement for the Moran scatter plot is to select a variable, through the fa-
miliar Variable Settings interface. The community area income per capita variable is
INCPERCAP. The Weights need to be chosen from the matching drop-down list.
If the Weights Manager does not contain at least one spatial weight, then one must be cre-
ated or loaded. The example uses first-order queen contiguity, as the file Chicago_2020_q.gal.
Selecting OK brings up the Moran scatter plot shown in Figure 13.9. This graph has the
same features as the standard scatter plot (Section 7.3), except that no statistics are provided
at the bottom. Also, in contrast to what holds for the standard scatter plot, the Regimes
Regression option is not active.
262 Spatial Autocorrelation

13.6.2 Moran scatter plot options

In the usual way, a right click on the graph brings up a list of available options. These are
largely identical to those for a standard scatter plot (Section 7.3.1), with a few exceptions.
Since the variables in the Moran scatter plot are standardized by design, there is no Data
option.
One additional feature is to Save Results. This saves two new variables to the data table,
one with the standardized variable and one with its spatial lag. The default variable names
for these are, respectively, MORAN_STD and MORAN_LAG.

13.6.2.1 Randomization
As such, the slope of the linear fit only provides an estimate of Moran’s I, but it does not
reveal any information about the significance of the test statistic. The latter is obtained
by means of the Randomization option, which implements a permutation approach. Four
hard-coded choices are provided for the number of permutations: 99, 199, 499 and 999.
In addition, there is an Other option, which allows for any value up to 99,999. As a
consequence, the smallest result that can be obtained for the pseudo p-value is 0.00001.
This implies that out of 99,999 randomly generated data sets, none yielded a computed
Moran’s I that equals or exceeds the one obtained from the actual data.
The result for a choice of 999 is shown in Figure 13.7, with an associated pseudo p-value of
0.001 for the spatial distribution of the Chicago community area per capita income data.
The graph provides strong evidence of spatial clustering of similar values.
As detailed in Section 13.5.2.2, the reference distribution is accompanied by several descriptive
statistics, such as the theoretical (E[I]) and empirical means, the standard deviation and the
associated z-value.
The Run button allows for repeated simulations in order to assess the sensitivity of the result
to the particular run of random numbers that is used. For a large number of permutations
such as 999, one run is typically sufficient.

13.6.2.2 Reproducing results - the random seed

In order to facilitate replication, the default setting in GeoDa is to use a specific seed for
the random number generator. This is evidenced by the check mark next to Use Specified
Seed in the Randomization options. The default seed value used is shown when selecting
the Specify Seed . . . box, where it can also be changed.
The random seed can also be set globally in the Preferences, under the System tab,
selected under the GeoDa item on the menu. At the bottom of the dialog, under the
item Method:, the box for using the specified seed is checked by default, and the value
123456789 is used as the default seed. This can be changed by typing in a different value.
The same random seed is used in all operations in GeoDa that rely on some type of random
permutations, such as all flavors of the Moran scatter plot, as well as all the local spatial
autocorrelation statistics. This ensures that the results will be identical for each analysis
that uses the same sequence of random numbers.
However, due to the way the random number generator is implemented in each operating
system, there may be slight differences in the results depending on the hardware, especially
Visualizing spatial autocorrelation with the Moran scatter plot 263

Figure 13.12: Moran scatter plot LOWESS smoother

when a different number of CPU cores is used.6 In order to control for any such discrepancies,
it is possible to Set the number of CPU cores manually in the Preferences.

13.6.2.3 LOWESS smoother

For the Moran scatter plot, the default is a Linear Smoother, but just as for the standard
scatter plot, it is also possible to use a nonlinear smoother. The View > LOWESS
Smoother works in exactly the same way as for the standard scatter plot.
With the View > Linear Smoother option turned off, only the nonlinear fit from the
local regression is shown on the Moran scatter plot, as in Figure 13.12. The LOWESS
smoother can be explored to identify potential structural breaks in the pattern of spatial
autocorrelation. For example, in some parts of the data set, the curve may be very steep
and positive, suggesting strong positive spatial autocorrelation, whereas in other parts, it
could be flat, indicating no autocorrelation.
The local smoother is fit through the points in the Moran scatter plot by utilizing only those
observations within a given bandwidth. As for the standard scatter plot, the default is to use
0.20 of the range of values. This can be specified by selecting the View > Edit LOWESS
Parameters item in the options menu, in the usual way.
Using the default bandwidth, the nonlinear smoother in Figure 13.12 suggests some instability
in the global spatial autocorrelation for values of z between 0 and 0.80, where the generally
positive slope turns into a negative one for a brief interval. Through linking (and brushing),
the corresponding observations can be identified on the associated map.
6 The implementation of the random numbers in GeoDa takes advantage of multi-threading, a built-in

form of parallel computing.

264 Spatial Autocorrelation

Figure 13.13: Brushing the Moran scatter plot

13.6.3 Brushing the Moran scatter plot

An important option to assess the stability of the spatial autocorrelation throughout the
data set, or, alternatively, to search for potential spatial heterogeneity, is the exploration of
the scatter plot through dynamic brushing.
This is invoked by selecting Regimes Regression in the View option. As in the standard
scatter plot, with this option selected, statistics for the slope and intercept are recomputed as
observations are selected and unselected. However, this is implemented somewhat differently
for the Moran scatter plot.
As soon as this option is activated, three Moran scatter plots are shown, as in Figure 13.13.
The one to the left (in red) is for the selected observations. The one to the right is for the
complement, referred to as unselected (in black).
In the figure, 18 observations from the high-high and high-low quadrants have been selected,
as indicated in the status bar. These yield a Moran’s I of 0.035, illustrated by the plot on
the left. The Moran’s I for the complement is 0.750. While there is no significance associated
with these measures, they suggest the presence of spatial heterogeneity.
The distinguishing characteristics of the selection in the Moran scatter plot (and any
associated brushing) is that the spatial weights are dynamically updated for both the selection
and the unselected observations. More precisely, for each of the subset of observations, selected
and unselected, the spatial weights are adjusted to remove links to observations belonging
to the other subset. In other words, there are no edge effects and each subset is treated as a
self-contained entity.
Visualizing spatial autocorrelation with the Moran scatter plot 265

Figure 13.14: Brushing and linking the Moran scatter plot

Figure 13.15: Static spatial weights in Moran scatter plot

As a result, the slope of the Moran scatter plots on the left and the right are as if the
data set consisted only of the selected/unselected points. This is a visual implementation
of a regionalized Moran’s I, where the indicator of spatial autocorrelation is calculated for
a subset of the observations (Munasinghe and Morris, 1996). As the brush is moved over
the graph, the selected and unselected scatter plots are updated in real time, showing the
corresponding regional Moran’s I statistics in the panels.
An alternative perspective is to approach the brushing starting with the map, as in Figure
13.14. In the left panel, 22 community areas are selected, with the associated observations
highlighted in the Moran scatter plot to the right. The Moran’s I corresponding to the
selection is 0.006, with 0.583 for the complement, again suggesting somewhat of a structural
break.
266 Spatial Autocorrelation

13.6.3.1 Static and dynamic spatial weights

The behavior of the Moran scatter plot with Regimes Regression turned on is to have
the spatial weights updated dynamically, in order to remove any edge effects. This contrasts
with the treatment in a conventional scatter plot constructed with the standardized variable
and its spatial lag.
In Figure 13.15, the two variables are used that were formed by the Save Results option,
i.e., MORAN_STD on the horizontal axis and MORAN_LAG on the vertical axis.
Since this is a standard scatter plot, the Regimes Regression option is on by default,
resulting in the statistics for selected and unselected at the bottom.
The graph is for the same 22 observations as selected in Figure 13.14. However, in contrast
to that case, there is no dynamic adjustment of the spatial weights.
More precisely, this means that the spatial lag for a selected/unselected observation is
computed from all its original neighbors, including the ones that belong to the other subset.
In the example, the corresponding slopes are, respectively, 0.218 and 0.590, compared to
0.006 and 0.583 for the dynamic weights. The associated Chow test shows strong evidence
of structural instability.
However, this result needs to be interpreted with caution, since the weights structure has
not been adjusted to reflect the actual breaks within the data set. Specifically, the areas
at the edge of the selection are still connected to their neighbors, whereas this is not the
case in the dynamic weights approach. The standard scatter plot ignores boundary effects,
in the sense that it includes neighbors outside the regimes. In some contexts, this may
be the interpretation sought, whereas in others the dynamic weights adjustment is more
appropriate.
14
Advanced Global Spatial Autocorrelation

In this chapter, I continue with the treatment of global spatial autocorrelation. First, I
consider some special applications of the Moran scatter plot principle to time differenced
variables and to proportions. These cases provide some short cuts to the computations
needed to construct the graph, but otherwise are identical to the standard plot covered in
Chapter 13.
Next, I cover the topic of bivariate spatial correlation, a common source of confusion in
practice. After presenting the principle behind a bivariate Moran scatter plot, attention
focuses on the special bivariate situation where the observation on a variable at one point
in time is related to its neighbors in a different time period, i.e., space-time correlation. A
critical review is provided of different perspectives and their relative merits and drawbacks.
To illustrate these methods, I go back to the Oaxaca Development sample data set.

14.1 Topics Covered

• Visualize spatial correlation among time-differenced values
• Correct Moran’s I for the variance instability in rates
• Visualize bivariate spatial correlation
• Assess different aspects of bivariate spatial correlation
• Alternative computations of space-time autocorrelation
GeoDa Functions
• Space > Differential Moran’s I
– saving the first differences and spatial lag
• Space > Moran’s I with EB Rate
– saving the rate and spatial lag
• Space > Bivariate Moran’s I
– variable selection
– permutation inference
– saving the spatial lag

14.2 Specialized Moran Scatter Plots

In this section, two special applications of the Moran scatter plot principle are considered.
Each pertains to a different type of situation, both commonly encountered in practice.

DOI: 10.1201/9781003274919-14 267

268 Advanced Global Spatial Autocorrelation

In the first example, the data for the same variable are available at different points in time.
Rather than having to explicitly compute change values between two points in time, the
differential Moran scatter plot provides a short cut, in the sense that the differences are
computed under the hood, as part of the scatter plot construction.
The second situation is the familiar case where the variable of interest is a rate or proportion.
As discussed at length in Chapter 6, the inherent variance instability of such data is
something that needs to be accounted for. This is especially important in the context of
spatial autocorrelation, since an important requirement for such statistics to be valid, like
Moran’s I, is to have a constant variance. The Morans’ I with EB Rate implements the
necessary standardization.

14.2.1 Differential Moran scatter plot

The differential Moran scatter plot is a short cut to compute Moran’s I for the first difference
of a variable measured at two points in time. Specifically, for a variable y observed for each
location i at times t and t − 1, the spatial autocorrelation coefficient is computed for the
difference yi,t − yi,t−1 .
The difference operation takes care of any temporal correlation that may be the result of
a spatial fixed effect, i.e., one or more variables determining the value of yi that remain
unchanged over time. Examples of such fixed effects are locational advantages (e.g., a port
city), or regulatory differences (state income tax vs. no income tax).
More formally, with μi as the spatial fixed effect associated with location i, the value at
each location for times t and t − 1 can be expressed as the sum of some intrinsic value u
and the fixed effect:
yi,t = ui,t + μi ,
and
yi,t−1 = ui,t−1 + μi ,
where μi is unchanged over time (hence, fixed).
The presence of μi in both time periods induces a temporal correlation between yi,t and
yi,t−1 , above and beyond the intrinsic correlation between ui,t and ui,t−1 . Even when ui,t
and ui,t−1 are uncorrelated, the correlation between yi,t and yi,t−1 would be E[μ2i ], or the
variance of the fixed effect.
Taking the first difference eliminates the fixed effect and ensures that any remaining correla-
tion is solely due to u:
yi,t − yi,t−1 = ui,t + μi − ui,t−1 − μi = ui,t − ui,t−1 .

A differential
 Moran’s I is then the slope in a regression of the spatial lag of the difference,
i.e., j wij (yj,t − yj,t−1 ) on the difference (yi,t − yi,t−1 ).
In GeoDa, the slope computation is applied to the standardized value of the difference (i.e.,
the standardization of yi,t − yi,t−1 ), and not to the difference between the standardized
values.

14.2.1.1 Implementation
The differential Moran scatter plot functionality is the third item in the drop-down list
activated by the Moran scatter plot toolbar icon (Figure 13.1). Alternatively, it can be
started from the main menu as Space > Differential Moran’s I.
Specialized Moran Scatter Plots 269

Figure 14.1: Access to health care in 2010 and 2020

Figure 14.2: Change in access to health care between 2010 and 2020

Note that this option is only available for data sets with grouped variables (see Section 9.2.1).
Without grouped variables, a warning message is generated.
The functionality is illustrated with a time enabled version of the Oaxaca Development
sample data set (see Chapter 9). The relevant variables include p_P614NS, p_JOB and
p_PHA, among others.
Before proceeding with the actual differential Moran graph, Figure 14.1 illustrates the
spatial pattern of access to health care in 2010 and 2020, p_PHA, as box maps.1 The
classification is relative and therefore does not reflect that the values for access to health
care have improved rather dramatically between these two years: for example, the mean
changed from 53.4% to 75.8%. The spatial distribution shows some differences, especially in
the center of the map, where several municipalities moved from a higher quartile (brown) to
a lower quarter (blue), including 14 lower outliers in 2020.
1 Notethat the item in the variable selection list is initially given as p_PHA (all times). After selection,
this becomes, respectively, p_PHA (2010) and p_PHA (2020).
270 Advanced Global Spatial Autocorrelation

Figure 14.3: Moran scatter plot for access to health care in 2010 and 2020

A box map for the first difference between 2020 and 2010 (computed in the table) is shown
as Figure 14.2. In contrast to the map for the individual years, the first differences include
both lower (2) and upper outliers (8), i.e., areas that improved much less or much more than
others, relatively speaking. It is for this spatial pattern that the differential Moran’s I is
computed.
To provide some context, the individual Moran scatter plots for p_PHA in 2010 and 2020
are shown in Figure 14.3, computed using queen contiguity weights (Oaxaca_q.gal). Both
Moran’s I statistics are positive and highly significant, suggesting strong clustering. For
2010, the statistic is 0.221, with an associated z-value of 8.371, for 2020, it is 0.286, with an
associated z-value of 11.058. Using 999 permutations, both cases yield a pseudo p-value of
0.001.
After starting the differential Moran’s I functionality, a Differential Moran Variable
Settings interface provides the means to select the variable and time interval. The drop-
down list by Select variable contains only the grouped variables in the data set, such as
p_PHA for the example. Next, two time periods need to be selected. The values for the
time periods are determined by their definition in the Time Editor (see Section 9.2.1). For
the illustration, the years 2020 and 2010 are selected (in this order). Finally, the Weights
drop-down list contains the available spatial weights, here using oaxaca_q.
The resulting differential Moran scatter plot is shown in Figure 14.4. Moran’s I is 0.279, with
an associated z-value of 10.611 (from a randomization), again suggesting a strong pattern of
clustering. In other words, not only is there clustering in the values of the percent health
access, there is also clustering in their change over the ten year period. However, the global
Moran’s I provides no information as to where the clustering may happen, nor whether it is
driven by high or by low values. This is a common misinterpretation of the global Moran’s I
statistic.
The differential Moran scatter plot has the same options as the regular Moran scatter
plot, with one exception. The Save Results option provides a way to not only save the
Standardized Data (as MORAN_STD) and the Spatial Lag (as MORAN_LAG),
as in the regular case, but also Diff Values (as DIFF_VAL), the value of the difference
before standardization. Clearly, applying a regular Moran scatter plot to this new variable
provides the same result as the differential Moran scatter plot. The latter simply saves the
effort of first having to construct the differences.
Specialized Moran Scatter Plots 271

Figure 14.4: Differential Moran scatter plot for access to health care 2020–2010

14.2.2 Moran scatter plot for EB rates

In Chapter 6, the issue of the intrinsic variance instability associated with rates or proportions
was introduced. The proposed solution was to utilize the Bayesian idea of borrowing strength
through various smoothing operations, including an Empirical Bayes (EB) procedure (Section
6.4.2.2).
One may be tempted to carry out spatial autocorrelation tests on EB smoothed rates,
but this is not appropriate, since their construction already induces a degree of correlation.
Instead, Assuncao and Reis (1999) suggested a slightly different procedure to correct Moran’s
I spatial autocorrelation test statistic for varying population densities across observational
units, when the variable of interest is a rate or proportion. In this approach, the spatial
autocorrelation is not computed for a smoothed version of the original rate, but for a
transformed standardized random variable. In other words, the crude rate is turned into a
new variable that has a constant variance. The mean and variance used in the transformation
are computed for each individual observation.
The principle is similar to the shrinkage estimator presented in Section 6.4.2.2, in that it
adjusts the crude rate with a mean and variance obtained from a prior distribution. As in the
case of EB smoothing, these moments are computed from the data, hence the designation as
empirical Bayes.
As before, the point of departure is the crude rate, ri = Oi /Pi , where Oi is the count of
events at location i and Pi is the corresponding population at risk.
The rationale behind the Assunção-Reis approach is to standardize each ri as
ri − β
zi =  ,
α + (β/Pi )
272 Advanced Global Spatial Autocorrelation

using an estimated mean β and standard deviation α + β/Pi . The parameters α and β
are related to the prior distribution for the risk, similar to the treatment in Section 6.4.2.2.
In practice, the prior parameters are estimated by means of the so-called method of moments
(e.g., Marshall, 1991), yielding the following expressions:
β = O/P,
 
with O as the total event count ( i Oi ), and P as the total population count ( i Pi ), and

α=[ Pi (ri − β)2 ]/P − β/(P/n),
i

with n as the total number of observations (in other words, P/n is the average population).
One problem with the method of moments estimator is that the expression for α could yield
a negative result. In that case, its value is typically set to zero, i.e., α = 0. However, in
Assuncao and Reis (1999), the value for α is only set to zero when the resulting estimate for
the variance is negative, that is, when α + β/Pi < 0. Slight differences in the standardized
variates may result depending on the convention used. In GeoDa, when the variance estimate
is negative, the original crude rate is used.
Also, after the EB adjustment, the rates are further standardized to have a mean of zero
and a variance of one, in the usual fashion for a Moran scatter plot.

14.2.2.1 Implementation
The Moran scatter plot for standardized rates is invoked from the main menu as Space >
Moran’s I with EB Rate, or as the fourth item in the drop-down list from the Moran
scatter plot toolbar icon, Figure 13.1.
To illustrate this procedure, the same disability rate example for Oaxaca municipalities is
used as in Section 12.4.0.1, i.e., the ratio of the number of disabled persons (DIS20) over
the municipal population (PTOT20) in 2020. The crude rate and EB smoothed rate are
mapped in Figure 12.9.
The left panel of Figure 14.5 shows the Moran scatter plot obtained for the crude rate,
again using queen contiguity for the spatial weights. The Moran statistic is 0.2649, with an
associated z-value (from 999 permutations) of 10.354.
The right hand panel is for the EB Moran scatter plot. This is obtained through the same
variable selection interface as for spatially smoothed rates in Section 12.4.1.1, with a variable
for the numerator (DIS20) and for the denominator (PTOT20), as well as the spatial
weights (oaxaca_q).
The result for the EB Moran’s I is only slightly different from the crude rate, yielding a
statistic of 0.2680, with associated z-value of 10.452. However, the values along the horizontal
axis show the effect of the standardization, with a much smaller range than for the crude
rate.
In practice, the differences between Moran’s I for the crude rate and EB standardized rate
tend to be minor.
The EB standardized Moran scatter plot has all the same options as the conventional Moran
scatter plot, except again for the Save Results feature. In addition to the Standardized
Data (the values used in the calculation of the Moran scatter plot), and their Spatial Lag,
the EB Rates (i.e., before the standardization used in the Moran computation) can be
saved, with default variable name MORAN_EB.
Bivariate Spatial Correlation 273

Figure 14.5: Moran scatter plot for raw rate and EB Moran scatter plot

14.3 Bivariate Spatial Correlation

The concept of bivariate spatial correlation is complex and often misinterpreted. It is typically
considered to be the correlation between one variable and the spatial lag of another variable,
as originally implemented in the precursor of GeoDa (described in Anselin et al., 2002).
However, this formulation does not necessarily take into account the inherent correlation
between the two variables. More precisely,
 in a bivariate Moran’s I, the correlation is between
xi at location i and its neighbors j wij yj , where x and y are different variables. As
formulated, this concept does not take into account the correlation between xi and yi , i.e.,
between the two variables at the same location.
As a result, a bivariate Moran’s I statistic is often interpreted incorrectly, as it may overesti-
mate the spatial aspect of the correlation that instead may be due mostly to the in-place
correlation.2

14.3.1 Bivariate Moran scatter plot

In its initial conceptualization in Anselin et al. (2002), a bivariate Moran scatter plot extends
the idea of a Moran scatter plot with a variable on the horizontal axis and its spatial lag on
the vertical axis to a bivariate context. The fundamental difference is that in the bivariate
case the spatial lag pertains to a different variable. In essence, this notion of bivariate spatial
correlation measures the degree to which the value for a given variable at a location is
correlated with its neighbors for a different variable.
As in the univariate Moran scatter plot, the interest is in the slope of the linear fit to the
scatter plot. This yields a Moran’s I-like statistic as:
 
i( j wij yj × xi )
IB =  2 ,
i xi

or, the slope of a regression of W y on x.

2 Based on some strong assumptions, Lee (2001) provides an alternative that considers a separation

between the correlative aspect and the spatial aspect. This is not pursued further.
274 Advanced Global Spatial Autocorrelation

As before, all variables are expressed in standardized form, such that their means are zero
and their variance one. In addition, the spatial weights are row-standardized.
Note that, unlike in the univariate autocorrelation case, the regression of x on W y also yields
an unbiased estimate of the slope, providing an alternative perspective on bivariate spatial
correlation (see Section 14.4). In the case of the regression of x on W x, the explanatory
variable W x is endogenous, so that the ordinary least squares estimation of the linear fit
is biased. However, with W y referring to a different variable, the so-called simultaneous
equation bias becomes a non-issue, and OLS in a regression of x on W y has all the standard
properties (such as unbiasedness).
A special case of bivariate spatial autocorrelation is when the variable is measured at two
points in time, say zi,t and zi,t−1 , as in Section 14.2.1. The statistic then pertains to the
extent to which the value observed at a location at a given time is correlated with its value
at neighboring locations at a different point in time.

The natural interpretation of this concept is to relate zi,t to j wij zj,t−1 , i.e., the correlation
between a value at t and its neighbors in a previous time period:
 
i( j wij zj,t−1 × zi,t )
IT =  2 ,
i zi,t

which expresses the effect of neighbors in t − 1 on the present value.

Alternatively, and maybe less intuitively,
 one can relate the value at a previous time period
zt−1 to its neighbors in the future, j wij zt , as:
 
i( j wij zj,t × zi,t−1 )
IT =  2 ,
i zi,t−1

or the effect of a location at t − 1 on its neighbors in the future.

While formally correct, this may not be a proper interpretation of the dynamics involved.
In fact, the notion of spatial correlation pertains to the effect of neighbors on a central
location, not the other way around. While the Moran scatter plot seems to reverse this logic,
this is purely a formalism, without any consequences in the univariate case. However, when
relating the slope in the scatter plot to the dynamics of a process, this representation should
be interpreted with caution.
A possible source of confusion is that the proper regression specification for a dynamic
process would be as: 
zi,t = β1 wij zj,t−1 + ui ,
j

with ui as the usual error term, and not as:


zi,t−1 = β2 wij zj,t + ui ,
j

which would have the future predicting the past. This contrasts with the linear regression
specification used (purely formally) to estimate the bivariate Moran’s I, for example:

wij zj,t−1 = β3 zi,t + ui .
j
Bivariate Spatial Correlation 275

Figure 14.6: Bivariate Moran scatter plot for access to health care 2020 and its spatial lag in
2010

In terms of the interpretation of a dynamic process, only β1 has intuitive appeal. However,
in terms of measuring the degree of spatial correlation between past neighbors and a current
value, as measured by the linear fit in a Moran scatter plot, β3 is the correct interpretation.
In the univariate case, only the specification with the spatially lagged variable on the left
hand side yields a valid estimate. As a result, for the univariate Moran’s I, there is no
ambiguity about which variables should be on the x-axis and y-axis. In contrast, in the
bivariate case, both options are valid, although with a different interpretation.
Inference is again based on a permutation approach, but with an important difference. Since
the interest focuses on the bivariate spatial association, the values for x and y are fixed at
their locations, and only the remaining values for y are randomly permuted. In the usual
manner, this yields a reference distribution for the statistic under the null hypothesis that
the spatial arrangement of the remaining y values is random. It is important to keep in mind
that since the focus is on the correlation between the x value at i and the y values at the
neighboring locations, the correlation between x and y at location i is ignored.

14.3.1.1 Creating a bivariate Moran scatter plot

A bivariate Moran scatter plot is created as the second item in the drop-down list activated
by the Moran scatter plot toolbar icon (Figure 13.1). Alternatively, it can be started from
the main menu as Space > Bivariate Moran’s I.
The variables and spatial weights are selected through the Bivariate Moran Variable
Settings dialog. This allows for the selection of the First Variable (X), the Second
Variable (Y) and the Weights from corresponding drop-down lists. Since the version of
the Oaxaca Development data set used here is time enabled, there are also drop-down lists
for the Time period.
276 Advanced Global Spatial Autocorrelation

To illustrate this functionality, the same variables are used as in Section 14.2.1, again with
queen contiguity spatial weights (oaxaca_q). For the bivariate Moran’s I, the x-variable is
then p_PHA (2020) and the y-variable its spatial lag in 2010, W_pPHA (2010).
The resulting graph is shown in Figure 14.6. The linear fit is almost horizontal, reflected
by a Moran’s I of −0.008. Randomization, using 999 permutations, yields a pseudo p-value
of 0.363, clearly not significant. This result may seem surprising, given the strong spatial
autocorrelation found for p_PHA in each of the individual years (Figure 14.3). One possible
explanation for this finding is that whereas there is strong evidence of clustering in each
year, the location of the clusters may be different. Therefore, the relationship between the
value at a given location and its neighbors in a different year may be weak, as is the case in
this example.
All the same options as for the univariate Moran scatter plot apply here as well. Also, the
standardized value of the x variable and the spatial lag of the y-variable (applied to its
standardized value) can be saved to the table, in the same way as for the univariate Moran
scatter plot.

14.4 The Many Faces of Space-Time Correlation

In contrast to the univariate Moran scatter plot, where the interpretation of the linear fit is
unequivocally Moran’s I, there is no such clarity in the bivariate case. In addition to the
interpretation offered above, which is a traditional Moran’s I-like coefficient, there are at
least four additional perspectives that are relevant. Each is briefly considered in turn.

14.4.1 Serial (temporal) correlation

A first interpretation is the regression/correlation between the variable under consideration
at two points in time, i.e., the linear relationship between zi,t and its time lag zi,t−1 . Since all
variables are expressed as standardized entities, regression and correlation are equivalent.3
The result for the relationship between p_PHA (2020) and p_PHA (2010) is shown
in Figure 14.7. The slope of the regression line is slightly positive and highly significant
(p-value of 0.000) at 0.157. This suggests an overall positive correlation between the values
at the two points in time. However, the overall fit of the regression is very poor, with an R2
of only 0.025.

14.4.2 Serial (temporal) correlation between spatial lags

A second perspective is the linear relation between the spatial lag
of the variable of interest
between the two time periods, i.e., a regression of the variable j wij yj,t on j wij yj,t−1 ,
or W_pPHA (2020) on W_pPHA (2010). In some sense, if there is a strong temporal
correlation between the variable in-situ, as well as spatial correlation in each of the time
periods, one may expect a temporal correlation between their neighbors in the form of the
associated spatial lags. However, this assumes that the location of the clusters in unchanging
over time, a property that the global Moran’s I is unable to assess.
3 The regression can be carried out by selecting Data > View Standardized Data in the scatter

plot, or by using the saved standardized values and their spatial lags from the Moran scatter plot for each
individual year (by means of the Save Results option).
The Many Faces of Space-Time Correlation 277

Figure 14.7: Correlation between access to health care in 2020 and 2010

The result is shown in Figure 14.8. The regression slope of 0.073 is not significant, with a
p-value of 0.132. Also, the overall fit is very poor, given the R2 of 0.004. This is no surprise,
given the ball-like shape of the scatter plot. In other words, while there is a relationship
between the health access for two different time periods, this does not hold for the neighbors
(spatial lags). This finding provides additional support for the notion that the clusters in the
two years may be in different locations. If they were in the same locations, then the spatial
lags for the two time periods would tend to be correlated.

14.4.3 Space-time regression

A third perspective is offered by a regression of the value of a variable at time t on the value
of its neighbors at time t − 1, formally, a regression of yi,t on j wij yj,t−1 . This is the most
natural formulation of a space-time regression, expressing how the values at neighboring
locations at the previous time period diffuse to the location at the next time period.
In Figure 14.9, this is illustrated for a regression of p_PHA (2020) on W_pPHA (2010).
Again, the scatter plot is highly circular, yielding an almost horizontal slope for the linear
fit. The regression coefficient of −0.028 is not significant, with a p-value of 0.718. The R2 is
actually zero!
In other words, here we find no direct relationship between the value at a location and its
neighbors in a previous location, consistent with the lack of significance of the bivariate
Moran’s I in Figure 14.6.
278 Advanced Global Spatial Autocorrelation

Figure 14.8: Correlation between the spatial lag of access to health care in 2020 and the
spatial lag in 2010

Figure 14.9: Correlation between access to health care in 2020 and its spatial lag in 2010
The Many Faces of Space-Time Correlation 279

Figure 14.10: OLS regression of pPHA 2020 on pPHA 2010 and W PHA 2010

14.4.4 Serial and space-time regression

A final perspective is offered by including both in-situ correlation and space-time correlation
 specification. Specifically, this is expressed as a regression of yi,t on both
in a regression
xi,t−1 and j wij xj,t−1 . In the absence of temporal error autocorrelation, standard ordinary
least squares regression will yield unbiased estimates of the coefficients.
The results for a regression of p_PHA (2020) on p_PHA (2010) and W_pPHA
(2010) are shown in Figure 14.10. All variables are in standardized form. Purely as an
illustration, this uses the regression functionality of GeoDa, a discussion of which is beyond
the scope of this book.4
The specification offers yet a different perspective on the space-time relationships. Both
regression coefficients are significant, but the coefficient of the space-time lag is negative and
only weakly significant, with a p-value of 0.04. The overall adjusted R2 is only 0.03. This
would suggest that after correcting for in-situ correlation, the effect of neighbors in the past
was negative, rather than positive.
Needless to say, focusing the analysis solely on the bivariate Moran scatter plot provides
only a limited perspective on the complexity of the dynamics of space-time patterns. In
practice, the full space-time regression may provide the most reliable insights.

4 See Anselin and Rey (2014) for specifics on both the methods and their implementation in GeoDa.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
15
Nonparametric Spatial Autocorrelation

In this chapter, the spatial weights matrix is abandoned in the formulation of a spatial
autocorrelation statistic. Instead, a measure of attribute (dis)similarity is related directly
to the distance that separates pairs of observations, in a non-parametric fashion. Two
approaches covered are the spatial correlogram and the smoothed distance scatter plot.
To illustrate these methods, I will use the Italy Community Banks point layer sample data
set.

15.1 Topics Covered

• Analyze the range of spatial autocorrelation by means of a spatial correlogram
• Assess the sensitivity of the spatial correlogram to its settings (maximum distance,
number of bins, smoother)
• Analyze the range of spatial autocorrelation by means of the smoothed distance scatter
plot
• Assess the sensitivity of the smoothed distance scatter plot to its settings (maximum
distance, number of bins, smoother)
• Address the computation of the correlogram and distance plots for large(r) data sets by
relying on random sampling
GeoDa Functions
• Space > Spatial Correlogram
– variable selection
– selecting the number of bins
– selecting the maximum distance
– using a random sample of locations
– changing the smoothing parameters
• Space > Distance Scatter Plot
– variable selection
– selecting the number of bins
– selecting the maximum distance
– using a random sample of locations
– changing the smoothing parameters

DOI: 10.1201/9781003274919-15 281

282 Nonparametric Spatial Autocorrelation

Toolbar Icons

Figure 15.1: Moran Scatter Plot | Spatial Correlogram | Cluster Maps

15.2 Non-Parametric Approaches

An alternative to the spatial weights and cross-product formulation of a spatial autocorrela-
tion statistic is to establish the connection between attribute (dis)similarity and locational
similarity in a different way (Section 13.4.2). This is implemented by expressing a measure
of attribute (dis)similarity directly as a function of the distance that separates pairs of
observations:
f (xi , xj ) = g(dij ),
where g is a function that expresses the distance decay in the strength of association between
the values observed at a pair of locations i, j.
In this chapter, two approaches are outlined that let the function g be specified in a
non-parametric way, directly from the data. This contrasts with the standard practice in
geo-statistics, where a specific function (semi-variogram) is fit. The spatial correlogram
approach uses a cross-product measure of attribute similarity, whereas the smoothed distance
scatter plot is based on a squared difference dissimilarity.

15.3 Spatial Correlogram

A spatial correlogram expresses the covariances or correlations between all pairs of observa-
tions as a function of the distance that separates them. The estimate is obtained from a
non-parametric local regression fit, such as LOWESS (Section 7.3.2) or kernel regression. Its
properties are based on the theory of nonparametric estimators of the autocovariance for a
stationary random field, first outlined in Hall and Patil (1994; see also Bjornstad and Falck,
2001).
With the variables expressed in standardized form as z, this approach boils down to a local
regression:
zi .zj = g(dij ) + ,
where dij is the distance between a pair of locations i, j,  is an error term and g is the
non-parametric function to be determined from the data. Since the variables are standardized,
the result is a spatial correlogram which shows how the strength of correlation varies with
the distance separating the observations.
In GeoDa, this is implemented by first arranging pairs of observations by distance bin, similar
to the approach taken in the estimation of a semi-variogram in geostatistics. The correlogram
is then the nonlinear curve fit to the average correlation for all pairs of observations in a
distance bin by means of a LOWESS regression .
Spatial Correlogram 283

Figure 15.2: Spatial distribution of loan loss provisions – Italian banks 2016

The main interest in the spatial correlogram is to determine the range of interaction, i.e.,
the distance within which the spatial correlation is positive. In addition, following Tobler’s
law (Tobler, 1970), the strength of the correlation should decrease with distance.
Beyond the range of interaction, the correlation should become negligible, although in
practice it can still show fluctuations.1

15.3.1 Creating a spatial correlogram

To illustrate the spatial correlogram, I use the Italy Community Banks sample data set
from Chapter 11 in time-enabled form. Specifically, the spatial pattern of the loan loss
provisions over customer loans is investigated for 2016 (LLP (2016)). The presumption
is that this variable addresses similar local risk environments, which would suggest spatial
autocorrelation.
A quantile map with eight categories is shown in Figure 15.2. The suggestion of pat-
terning is confirmed by a Moran’s I of 0.179, with associated z-value of 5.41 (based on
999 permutations) for the six neighbor k-nearest neighbor weights truncated at 73 km
(italy_banks_te_d73_k6.gal).2
The spatial correlogram functionality is invoked by selecting the middle icon in the spatial
analysis group, as in Figure 15.1, and choosing Spatial Correlogram. Alternatively, from
the menu, it can be started by selecting Space > Spatial Correlogram (the item near
the very bottom of the list of options).
1 Some autoregressive spatial processes induce a pattern of alternating positive and negative autocorrela-

tions, but a further consideration of this is beyond the current scope.

2 The Moran scatter plot calculation in GeoDa removes the two isolates from the computations.
284 Nonparametric Spatial Autocorrelation

Figure 15.3: Correlogram parameter selection

This brings up a Correlogram Parameters dialog with the default parameter settings
and a graph in the background. However, this initial graph is almost never informative at
this point, since it shows a correlogram for the first variable, which is usually an ID value.
The initial layout of the interface is shown in the left-hand panel of Figure 15.3. The items
at the top of the dialog are the Variable (available from a drop-down list), the Distance
metric (default is Euclidean Distance, but Arc Distance is also available, expressed
either as miles or as kilometers). In the example, since the data set is time enabled, the
Variable also has an associated Time drop-down list. For the initial default setting, LLP
(2016) and 2016 are used with Euclidean Distance (given the projection used, this
distance is in meters).
Next follows the Number of Bins. The non-parametric correlogram is computed by means
of a local regression on the spatial correlation estimated from the pairs that fall within each
distance bin. The number of bins determines the distance range of each bin. This range is
the maximum distance divided by the number of bins. The more bins are chosen, the more
fine-grained the correlogram will be. However, this also potentially can lead to too few pairs
in some bins (the rule of thumb is to have at least 30).
The number of pairs contained in each bin is a result of the interplay between the number of
bins and the maximum distance. As the default, All Pairs are used, with Max Distance
unchecked. In the example, with 261 observations, this yields: [2612 − 261]/2 = 33930 pairs,
as listed in the Estimated Pairs box.
In many instances in practice, using all pairs may not be a good choice. For example, when
there are many observations, the number of pairs quickly becomes unwieldy and the program
will run out of memory. Also, the correlations computed for pairs that are far apart are not
that meaningful, since they should be zero due to Tobler’s law. Therefore it is often practical
to truncate the observations to only those pairs that are within a reasonable distance range.
The bottom half of the dialog provides options for fine-tuning these choices. These are
considered in Sections 15.3.2.1 and 15.3.2.2.
With the default set as in the left-hand panel of Figure 15.3, the spatial correlogram shown
in Figure 15.4 is created.
Spatial Correlogram 285

Figure 15.4: Default spatial correlogram – LLP (2016)

15.3.1.1 Interpretation
The top half of the graph in Figure 15.4 is the actual correlogram that depicts how the
spatial autocorrelation changes with distance. Hovering the pointer over each blue dot gives
the spatial autocorrelation associated with that distance band in the status bar. For example,
the second dot corresponds with an autocorrelation of −0.024 for observations within the
distance band from 119.4 km to 238.9 km, as shown in the status bar (the distances in the
graph are expressed in meters). The first dot corresponds with an autocorrelation of 0.081.
The intersection between the correlogram and the dashed zero axis, which determines the
range of spatial autocorrelation, happens in the midpoint of the second range, or roughly
around 179 km. Beyond that range, the autocorrelation fluctuates around the zero line.
The bottom half of the graph consists of a histogram that shows the number of pairs of
observations in each bin. Hovering the pointer over a given bin shows in the status bar
how many pairs are contained in the bin. In the example, each bin has more than sufficient
observation pairs. Even the last bin, which seems small (a function of the vertical scale),
contains 102 pairs to compute the average autocorrelation.

15.3.2 Spatial correlogram options

The spatial correlogram has four main options, invoked in the usual way by right clicking on
the graph.
The Change Parameters item brings back the Correlogram Parameters dialog and
allows for fine tuning of the number of bins, maximum distance and randomization options
(Sections 15.3.2.1 and 15.3.2.2).
The Edit LOWESS Parameters option manipulates the bandwidth and other technical
parameters that can be specified for any LOWESS smoother. The default is to use a
bandwidth of 0.20, which works well in most situations. As usual, a larger bandwidth will
yield a (slightly) smoother curve, and a smaller bandwidth will work in the opposite way.
This option works in exactly the same manner as for the standard LOWESS nonlinear
smoother (Section 7.3.2.1).
286 Nonparametric Spatial Autocorrelation

Figure 15.5: Spatial correlogram – LLP, 150 km cut off, 10 bins

The View > Display Statistics option generates a list of descriptive statistics for each
bin. The computed autocorrelation is provided, as well as the distance range for the bin
(lower and upper bound), and the number of pairs used to compute the statistic. In addition,
there is a summary with the minimum and maximum distance, the total number of pairs
and an estimate for the range, i.e., the distance at which the estimated autocorrelation first
becomes zero (see the discussion of Figures 15.5 and 15.6 for an illustration). The other
View options are the familiar Set Display Precision, Set Display Precision on Axes
and Show Status Bar. The latter is active by default.
Finally, Save Results provides a record of the descriptive statistics of the correlogram in a
text file (with file extension csv). The file contains the information listed at the bottom of
the graph when View > Display Statistics has been selected, with a column matching
each bin. The file includes the estimates of the spatial autocorrelation, bin ranges, number
of observations in each bin, as well as the summary.

15.3.2.1 Maximum distance and number of bins

In most situations in practice, the default use of all the distance pairs is not informative. In
addition, there is a good reason to limit the maximum distance considered in the selection
of pairs, since correlations at large(r) distances are both sparser (fewer pairs in a bin, which
leads to less precise estimates) and supposed to be near zero (Tobler’s law).
The Correlogram Parameters dialog contains options to set the Max Distance. When
this option is checked, half of the largest inter-observation distance is used.3 In the example,
that would be about 597 km, resulting in 17095 Estimated Pairs, about half of the
number when all pairs are used (not shown).4
Instead of using a rule of thumb such as half the total distance, the maximum distance can
be specified explicitly, as in the right-hand panel of Figure 15.3. Here, a value of 150000
3 This follows one of the most common rules of thumb from geostatistics, e.g., as in Journel and Huijbregts

(1978) and Deutsch and Journel (1998).

4 Since the estimated number of pairs is computed on the fly, as the maximum distance is adjusted by

means of the slider, it is only an approximation. The actual number of pairs used in the calculation of the
pairwise correlations is given in the status bar of the correlogram.
Spatial Correlogram 287

Figure 15.6: Spatial correlogram – LLP, 150 km cut off, 15 bins

(150 km) is entered. Combined with the default Number Bins of 10, and with View >
Display Statistics activated, this yields the correlogram in Figure 15.5.
The graph header lists the cutoff but also gives that all pairs have been used. This should
be interpreted in the sense that all pairs that meet the distance cutoff are used in the
computations, as opposed to a random sample (Section 15.3.2.2).
The distance range for each bin is now 15 km (15,000 in the graph) and, as a result, the
range is estimated to be between 45 km and 60 km, contrasted with more than double the
figure obtained in Figure 15.4. Also, the initial spatial autocorrelation, for pairs within 15
km of each other is 0.506, compared to 0.081 for the first range in the default setting. The
descriptive statistics depict the change of the estimated autocorrelation coefficient with
distance at a much more fine grained level than before. The correlogram shows the desired
shape, decreasing with distance until the range is reached and more or less flat afterward.
The computations are based on 6,579 pairs, less than a fifth of the original total.
A further refinement can be obtained by increasing the number of bins for the same maximum
distance. In Figure 15.6, the correlogram is depicted for 15 bins with the 150 km cut-off (the
settings shown in the right-hand panel of Figure 15.3), resulting in a range of 10 km for each
bin. The zero autocorrelation is estimated to occur between 50 and 60 km, an even narrower
band than for Figure 15.5. The autocorrelation in the first 10 km range is 0.503, basically
the same as in the first 15 km range shown before. The overall shape of the correlogram also
remains the same.
Overall, this suggests quite a strong pattern of positive spatial autocorrelation at short
distances, but negligible association beyond distances of 60 km. This highlights a different
aspect of spatial autocorrelation than can be provided by a statistic like Moran’s I, which
considers all pairs that meet the spatial weights criterion.

15.3.2.2 Random sampling

A final feature of the correlogram is the ability to compute the spatial correlations for a
sample of locations, which reduces the number of pairs used in the calculations. This is
especially useful when the data set is large, in which case the number of pairs could quickly
become prohibitive.
288 Nonparametric Spatial Autocorrelation

In the example used here, this is not really necessary, since the default sample size of
1,000,000 that is used to generate the random sample exceeds the current total number of
pairs in the data set. To implement this option, the Random Sample radio button needs
to be checked and a sample size specified if the default of 1 million is not desired. Also, to
allow exact replication, the Use Specified Seed option should be checked (the seed can be
adjusted by means of the Change button).
In practice, the sampling approximation is quite good, as long as the selected sample size is
not too small relative to the original data size.

15.4 Smoothed Distance Scatter Plot

The smoothed distance scatter plot (Anselin and Li, 2020) offers an alternative approach to
relating a measure of attribute dissimilarity to distance. It is similar in spirit to the familiar
semi-variogram from geostatistics (Isaaks and Srivastava, 1989; Cressie, 1993; Chilès and
Delfiner, 1999; Stein, 1999) but also different in some important respects.
In geostatistics, the point of departure for the semi-variogram is to express the magnitude
of the variance of the difference between observations at two points in space as a function of
the distance that separates them:
1
γ(h) = Var[zs − zs+h ],
2
where γ is the semi-variogram function and h is the separation between two observations.
Exploiting assumptions of spatial stationarity (such as constant mean and constant variance),
this results in a simpler expression that relates the expected value of the squared difference
to distance:
1
γ(h) = E[zs − zsh ]2 .
2
An empirical variogram is estimated from the data by computing the average squared
difference for several distance bins, similar to the computation of the spatial correlogram
(Section 15.3.2.1). A number theoretical variograms have been proposed, such as the spherical,
exponential and Matérn (see, e.g., Banerjee et al., 2015, for an overview). These differ in
terms of characteristics, such as the range, at which point the semi-variogram becomes flat,
and the shape of the function between the origin and the range.
The smoothed distance scatter plot is similar in spirit in that it relates a measure of
dissimilarity to the distance separating the observations, but it is not based on the typical
assumptions from geostatistics. It is a simple scatter plot of the distance in attribute space
(on the vertical axis) against the distance in geographic space (on the horizontal axis). The
distance in attribute space for a single variable z can be expressed as:

vij = (zi − zj )2 ,

with i and j as the observation locations. Note that this is not the squared difference, as in
the semi-variogram, but its square root, as a direct measure of distance.
The geographic Euclidean distance between the observation point is then the familiar:

dij = (xi − xj )2 + (yi − yj )2 ,

with (xi,j , yi,j ) as the observation coordinates.

Smoothed Distance Scatter Plot 289

Figure 15.7: Smoothed distance scatter plot

Whereas the geographic distance has an immediate interpretation in terms of distance units
(e.g., kilometers), this is not the case for the attribute distance matrix. In practice, it is
therefore advisable to first standardize the variable, so that attribute distance is expressed
in standard deviational units.
An example of the resulting scatter plot is shown in Figure 15.7 for the variable LLP (2016)
from the Italy Community Banks data set. The graph shows a slow increase up to about
120 km, after which it is mostly flat (but see Section 15.4.3 for more detailed options and
interpretation).
The idea of combining a measure of distance in attribute space with geographic distance
goes back at least to an example in Oden and Sokal (1986). They implemented a Mantel
test (Section 13.4.3) to assess the similarity between the elements of a geographic distance
matrix and a variable dissimilarity matrix. However, their approach is less flexible than the
smoothed distance scatter plot and does not lend itself readily to an extension to multiple
variables (Section 15.4.1).
One may be tempted to apply a linear fit to the scatter plot as an intuitive measure of
the association between the two variables. However, this would imply a linear relationship
between the two, whereas Tobler’s law suggests a non-linear distance decay, with a range
beyond which there is no association, in the same fashion as shown for the spatial correlogram.
However, in contrast to the correlogram, the smoothed distance scatter plot increases with
distance until the range is reached, since it involves a measure of dissimilarity.
A non-parametric local regression fit is applied to the smoothed distance scatter plot to
extract the overall pattern. Specifically, the implementation in GeoDa is based on the loess
approach (Cleveland et al., 1992).
290 Nonparametric Spatial Autocorrelation

As for the spatial correlogram, the main interest in this graph is to obtain an estimate of
the range of interaction, i.e., the point where the curve ceases to increase with distance and
begins to flatten out.

15.4.1 Multivariate extension

As explored in more depth in Chapter 18, a spatial autocorrelation measure based on a
cross-product (like Moran’s I) is difficult to extend to multiple variables. In contrast, as
demonstrated in Anselin (2019a), this is not the case for statistics that use a squared
difference (or its square root) as an indicator of attribute dissimilarity.
More specifically, the univariate attribute distance is readily extended to k variables as:

vij = (zki − zkj )2 ,
k

for k variables zk observed at locations i and j. This Euclidean distance in multi-attribute

space will play an important role in the treatment of clustering considered in Volume 2.
Instead of a Euclidean distance, a Manhattan metric could be used. Neither of these takes
into account the inherent correlation among the variables, which could be addressed by
means of a Mahalanobis distance, although this requires a reliable estimate of the variance
matrix (more precisely, the inverse of the variance matrix).5 This approach is not considered
further.
The multivariate smoothed distance plot replaces the univariate attribute distance by its
multivariate counterpart on the vertical axis. In all other respects, the method is the same
as the univariate case.

15.4.2 Creating a smoothed distance scatter plot

The smoothed distance plot is invoked as the second option from the spatial correlogram
icon shown in Figure 15.1, or as Space > Distance Scatter Plot from the menu. The
dialog that follows allows for the selection of the variable(s) and the settings to be used in
the analysis. In contrast to the spatial correlogram, multiple variables can be selected. To
illustrate the univariate case, the same variable as before is employed, LLP (2016). The
extension to multiple variables is straightforward, the only difference being that more than
over variable must be selected in the dialog. All other options and interpretations are the
same as for the univariate case.
Other important initial settings in the dialog are the Variable Distance, either Euclidean
Distance (the default) or Manhattan Distance, and the Geographic Distance, as
Euclidean Distance (the default, for projected coordinates), or two versions of Arc
Distance (in miles or kilometers, for coordinates as latitude-longitude).
In addition, a few options need to be set to make the plot operational. As argued for the
spatial correlogram, it is usually not advised to compute the attribute distance for all the
observation pairs in the data set. By default, the Max Distance box is checked, which
results in 1/2 maximum pairwise distance to be used as the largest pairwise distance,
following standard practice in empirical variography (Journel and Huijbregts, 1978; Deutsch
and Journel, 1998). In the example, this is about 597 km (compared to a maximum distance

5 The Mahalanobis distance between vectors xi and xj is dij = (xi − xj ) Σ−1 (xi − xj ), with Σ as the
covariance matrix.
Smoothed Distance Scatter Plot 291

Figure 15.8: Smoothed distance scatter plot with 150 km distance cut-off

of some 1,200 km). As in the spatial correlogram, this yields 17095 pairs to be considered,
compared to the 33,930 pairs in the default case. A different option is to use 1/2 of the
diagonal of the bounding box, another rule of thumb often used for empirical variograms.
The All Pairs button is checked by default. The alternative is to use Random Sample,
which works in the same way as for the spatial correlogram.
These default settings yield the plot in Figure 15.7, but without the actual points, which are
not shown by default. Using a more reasonable cut-off distance of 150 km yields the graph
in Figure 15.8, based on the computation for 6579 observation pairs. In this plot, the range
seems to be between 70 and 80 km, slightly higher than for the spatial correlogram in Figure
15.6.
Further options to refine the visualization are described in Sections 15.4.3 and 15.4.3.1.

15.4.3 Smoothed distance scatter plot options

The plot has several options that are invoked in the usual fashion, by right clicking on the
graph. There are five items:
• LOESS Setup
• View
• Color
• Axis Option
• Save Results.
The LOESS Setup determines the settings for the nonparametric local regression and is
considered in Section 15.4.3.1.
292 Nonparametric Spatial Autocorrelation

Figure 15.9: Smoothed distance scatter plot with 150 km distance cut-off and customized
Y-Axis

The default View setting is to show the status bar. View > Show Data Points adds
all the points to the scatter plot. As mentioned, this is turned off by default. The Color
options allow the color to be specified for the regression line, the scatter plot points and the
background.
An important option for effective visualization is to adjust the range on the Y-axis. This
is implemented through Axis Option > Adjust Value Range of Y-Axis. It allows the
minimum and maximum values for the Y-Axis to be customized, resulting in a clearer view
of the shape of the graph. For example, setting the maximum at 1.5 in Figure 15.8, yields
the plot shown in Figure 15.9.
In addition, the Axis Option can be used to specify the display precision, in the usual way.
Finally, Save Results will create a comma-separated value (csv) file that contains all the
scatter plot points with their X and Y coordinates. This file can then be used as input to
more advanced statistical analyses.

15.4.3.1 Loess settings

The loess approach implements a local polynomial regression method as outlined in Cleveland
et al. (1992). The implementation in GeoDa is based on the code contained in the netlib
library. The local regression uses a subset of the observations within a close attribute
distance from each observation point to compute the slopes in a polynomial regression, which
are then combined to produce a smooth function of predicted values. Since the smoothed
distance scatter plot only has one explanatory variable (geographic distance), the loess
implementation is fairly straightforward.
The LOESS Setup option allows for three important parameters to be set. The first is
the Degree of smoothing (span), i.e., the share of total observations that is used. The
Smoothed Distance Scatter Plot 293

Figure 15.10: Smoothed distance scatter plot with 150 km distance cut-off and span 0.5

Figure 15.11: Smoothed distance scatter plot with 150 km distance cut-off, span 0.5, linear
fit

netlib default is 0.75, which is also the default value in GeoDa. This corresponds to the
formal smoothing parameter α which is used in some other software implementations. The
294 Nonparametric Spatial Autocorrelation

Figure 15.12: Bivariate smoothed distance scatter plot

observations within the maximum range from the fit point are weighted as in a kernel
regression.6
The Degree of the polynomials is 2 by default, for a quadratic model, but 1 (linear)
is another (less often used) option. Finally, the Family setting selects the type of fitting
procedure. The default is gaussian, which boils down to a least squares fit. The alternative
is a sometimes more robust symmetric option.
To illustrate the sensitivity of the graph to some of these settings, Figure 15.10 shows the
effect of setting the span to 0.50, for a more local fit.
Figure 15.11 illustrates the effect of a linear fit vs a quadratic fit (with the same span of 0.5).
In practice, some experimentation may be necessary to yield the most effective graph.
Finally, Figure 15.12 illustrates a smoothed distance scatter plot for two variables, technical
input efficiency in 2016, TE_IN (2016) and loan loss provision LLP (2016). The span is
set to 0.5, with a quadratic function and the vertical axis truncated at 2. The graph shows a
gradually increasing function up to about 70 km, after which it is more or less horizontal.
This suggests the presence of a spatial relationship of the tuples of the two variables observed
within this range.

6 The netlib implementation uses a tricubic weighting scheme, i.e., (1 − (distance/maximum distance)3 )3 .
 Part V

Local Spatial Autocorrelation

Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
16
LISA and Local Moran

This chapter is the first in a series where attention shifts to the detection of the location
of clusters and spatial outliers. These are collections of observations that are surrounded
by neighbors that are more similar to them or more different from them than would be
expected under spatial randomness. The general principle by which this is approached is
based on the concept of Local Indicators of Spatial Association, or LISA (Anselin, 1995).
I begin by outlining the basic idea underlying the concept of LISA. This is followed by an
in-depth coverage of its most common application, in the form of the Local Moran statistic.
This statistic becomes a powerful tool to detect hot spots, cold spots, as well as spatial
outliers when combined with the classification of spatial autocorrelation in the Moran scatter
plot. The ultimate result is a local cluster map. An extensive discussion of its properties and
interpretation is provided, with special attention to the notion of significance.
To illustrate these methods, I will employ the Oaxaca Development sample data set.

16.1 Topics Covered

• Understand the concept of a Local Indicator of Spatial Autocorrelation
• Identify clusters with the Local Moran cluster map and significance map
• Interpret the spatial footprint of spatial clusters
• Assess the significance by means of a randomization approach
• Assess the sensitivity of different significance cut-off values
• Interpret significance by means of Bonferroni bounds and the False Discovery Rate
(FDR)
• Assess potential interaction effects by means of conditional cluster maps
GeoDa Functions
• Space > Univariate Local Moran’s I
– significance map and cluster map
– permutation inference
– setting the random seed
– selecting the significance filter
– saving LISA statistics
– select all cores and neighbors
– local conditional map

DOI: 10.1201/9781003274919-16 297

298 LISA and Local Moran

Toolbar Icons

Figure 16.1: Moran Scatter Plot | Spatial Correlogram | Cluster Maps

16.2 LISA Principle

As pointed out in the discussion of global spatial autocorrelation, statistics designed to
detect the presence of such correlation, such as Moran’s I and Geary’s c, are formulated
to reject the null hypothesis of spatial randomness, typically in favor of an alternative of
clustering (occasionally in favor of negative spatial autocorrelation). Such clustering is a
characteristic of the complete spatial pattern, but it does not provide an indication of the
location of the clusters.
An early suggestion of a test for local spatial autocorrelation was formulated by Getis and
Ord (1992) (see Chapter 17). This was generalized in Anselin (1995), where the concept
of a local indicator of spatial association, or LISA was introduced. A LISA is viewed as
having two important characteristics. First, it provides a statistic for each location with
an assessment of significance. Second, and less importantly, it establishes a proportional
relationship between the sum of the local statistics and a corresponding global statistic.
As shown in section 13.4, most global spatial autocorrelation statistics can  beexpressed
as a double sum over the i and j indices, for all pairs of observations, as i j gij . The
 be, for each observation (location) i, the sum of the
local form of such a statistic would then
relevant expression over the j index, j gij .
More precisely, and as covered previously, a spatial autocorrelation statistic consists of a
combination of a measure of attribute similarity between a pair of observations, f (xi , xj ),
with an indicator for geographical or locational similarity, in the form of spatial weights,
wij . For a global statistic, this takes on the form:

wij f (xi , xj )
i j

. A generic expression for a local indicator of spatial association is then:


wij f (xi , xj ).
j

For each different measure of attribute similarity f , a different statistic for global spatial
autocorrelation results. Consequently, there will be a corresponding LISA for each such
global statistic. First, the local counterpart of Moran’s I is considered.

16.3 Local Moran

The Local Moran (Anselin, 1995) is by far the most commonly used LISA statistic. It
follows from a straightforward application to the global Moran statistic of the principles just
outlined.
Local Moran 299

16.3.1 Formulation
As discussed in Chapter 13, the Moran’s I statistic is obtained as:
 
j zi .zj .wij /S0
I=
i
 2 , (16.1)
i zi /n

with the variable of interest (z) expressed as deviations from the mean, and S0 as the sum of
all the weights. In the row-standardized case, the latter equals the number of observations,
n. As a result, as shown in the discussion of the Moran scatter plot, the Moran’s I statistic
simplifies to:  
wij zi zj
I=
i
j 2 .
i zi

Using the logic just outlined, a corresponding Local Moran statistic would consist of the
component in the double sum that corresponds to each observation i, or:

j wij zi zj
Ii =  2 .
i zi

In this expression, the denominator is fixed and can thus further be ignored. To keep the
notation
 simple, it can be replaced by a, so that the Local Moran expression becomes
a. j wij zi zj . After some re-arranging, a simple expression is:

Ii = a × zi wij zj ,
j

or, the (scaled) product of the value at location i with its spatial lag, the weighted sum or
the average of the values at neighboring locations.

A specialcase occurs when the variable z is fully standardized. Then its variance i zi2 /n = 1,
so that i zi2 = n. Consequently, the global Moran’s I can be written as:
   
I= (a.zi wij zj ) = (1/n) (zi wij zj ),
i j i j

or the average of the local Moran’s I. This case illustrates a direct connection between the
local and the global Moran’s I.
Significance can be based on an analytical approximation, but, as argued in Anselin (1995),
this is not very reliable in practice.1 A preferred approach consists of a conditional per-
mutation method. This is similar to the permutation approach considered in the Moran
scatter plot, except that the value of each zi is held fixed at its location i. The remaining n-1
z-values are then randomly permuted to yield a reference distribution for the local statistic,
one for each location.
The randomization operates in the same fashion as for the global Moran’s I (see Section
13.5.2.2), except that the permutation is carried out for each observation in turn. The result
is a pseudo p-value for each location, which can then be used to assess significance. Note
that this notion of significance is not the standard one, and should not be interpreted that
way (see the discussion in section 16.5).
1 Further discussion of analytical results can be found in Sokal et al. (1998a), for an asymptotic approach,

and in Tiefelsdorf (2002), where a saddlepoint approximation is suggested.

300 LISA and Local Moran

Figure 16.2: Significance map and cluster map

Assessing significance in and of itself is not that useful for the Local Moran. However, when
an indication of significance is combined with the quadrant location of each observation
in the Moran Scatter plot, a very powerful interpretation becomes possible. The combined
information allows for a classification of the significant locations as High-High and Low-Low
spatial clusters, and High-Low and Low-High spatial outliers. It is important to keep in mind
that the reference to high and low is relative to the mean of the variable, and should not be
interpreted in an absolute sense. The notions of clusters and outliers are considered more
in-depth in section 16.4.

16.3.2 Implementation
The Univariate Local Moran’s I is started from the Cluster Maps toolbar icon, the
right-most icon in the spatial correlation group, highlighted in Figure 16.1. This brings up a
drop-down list, with the univariate Local Moran as the top-most item. Alternatively, this
option can be selected from the main menu, as Space > Univariate Local Moran’s I.
Either approach brings up the familiar Variable Settings dialog which lists the available
variables as well as the default weights file, at the bottom (e.g., oaxaca_q in the example).
As in Chapter 14, the variable is p_PHA(2010), the 2010 percentage of the municipal
population with access to health care. In the example, this variable is time enabled, which
results in the selected year (2010) being listed in the dialog as well.
As a reference, a box map reflecting the spatial distribution of this variable is shown in the
left-hand panel of Figure 14.1 (see also the left-hand panels in Figures 16.6 and 16.8 in this
Chapter).

16.3.2.1 Significance map and cluster map

Clicking OK at this point brings up a dialog to select the number and types of graphs and
maps to be created. The default is to provide just the Cluster Map, which is typically the
most informative graphic. In addition, a Significance Map and a Moran Scatter Plot
option is included as well.
To illustrate the Local Moran, selecting both cluster map and significance map brings
up Figure 16.2. This visualizes the result of the default randomization test, using 999
permutations and a p-value cut-off of 0.05.
Local Moran 301

Figure 16.3: Significance map and cluster map – 99999 permutations

The significance map on the left shows the locations with a significant local statistic, with
the degree of significance reflected in increasingly darker shades of green. The map starts
with p < 0.05 and shows all the categories of significance that are meaningful for the given
number of permutations. In the example, since there were 999 permutations, the smallest
pseudo p-value is 0.001, with nine such locations (the darkest shade of green).
The cluster map augments the significant locations with an indication of the type of spatial
association, based on the location of the value and its spatial lag in the Moran scatter plot
(see section 16.4). In this example, all four categories are represented, with dark red for the
High-High clusters (40 in the example), dark blue for the Low-Low clusters (48 locations),
light blue for the Low-High spatial outliers (12 locations) and light red for the High-Low
spatial outliers (16 locations).

16.3.2.2 Randomization options

The Randomization option is the first item in the options menu for both the significance
map and the cluster map. It operates in the same fashion as in the Moran scatter plot. As
before, up to 99,999 permutations are possible (for each observation in turn), preferably
using a specified random seed to allow replication.
The effect of the number of permutations is typically marginal relative to the default of 999.
In the example, selecting 99,999 results in one more significant location (117 vs 116). As
shown in the left-hand panel of Figure 16.3, there are now 70 locations significant at p <
0.05 (compared to 72), 34 locations at p < 0.01 (compared to 35), 11 locations at p < 0.001
(compared to 9) and one each at p < 0.0001 and p < 0.00001 (the smallest possible p-value).
They are illustrated by different shades of green in the significance map.
The cluster map, in the right-hand panel, is similarly only marginally affected. There are two
new observations in the High-High category, and one of the High-Low outliers disappears.
Otherwise, the map is the same as for 999 permutations.

16.3.2.3 Saving the Local Moran statistic

The third item in the options menu for the Local Moran is to Save Results.
This brings up a dialog with three potential variables to save to the table: Lisa Indices,
Clusters and Significance. The default variable names are LISA_I, LISA_CL and
LISA_P. The first are the actual values for the local statistics, which are typically not
302 LISA and Local Moran

that useful. The clusters are identified by an integer that designates the type of spatial
association: 0 for non-significant (for the current selection of the cut-off p-value, e.g., 0.05 in
the example), 1 for High-High, 2 for Low-Low, 3 for Low-High and 4 for High-Low.
The default variable names would typically be changed, especially when more than one
variable is considered in an analysis (or different spatial weights for the same variable).

16.3.2.4 Other options

In addition to the three options considered so far, there are several specialized operations for
the cluster and significance map. These are covered in section 16.4 on clusters and spatial
outliers (Select All) and 16.6 on conditional cluster maps (Show As Conditional Map).
The other options are customary, and include Selection Shape, Color and Show Status
Bar, Save Selection, Copy Image to Clipboard and Save Image As. In addition,
when the data set is time-enabled, the Time Variable Options are available. Finally, the
various options associated with the Connectivity item were covered in Section 10.4.5.1.
These other options operate in the same way as discussed before.

16.4 Clusters and Spatial Outliers

A critical link in the interpretation of significant Local Moran statistics is the connection
between the Moran scatter plot and the cluster map. As shown in Chapter 14, the Moran
scatter plot provides a classification of spatial association into four categories, corresponding
to the location of the observations in the four quadrants of the plot. These categories are
referred to as High-High, Low-Low, Low-High and High-Low, relative to the mean, which is
the center of the graph. It is important to keep in mind that there is a difference between a
location being in a given quadrant of the plot, and that location being a significant local
cluster or spatial outlier.
The cluster map provides a way to interpret the results and classify observations. However,
this is only valid for those locations where the local statistic is significant.

16.4.1 Clusters
Clusters are centered around locations with a significant positive Local Moran statistic, i.e.,
locations with similar neighbors. In the terminology of the Moran scatter plot, these are
either High-High or Low-Low locations.
Figure 16.4 illustrates the connection between the Moran scatter plot and the cluster map.
On the left, the 230 observations in the High-High quadrant of the Moran scatter plot for
p_PHA (2010) (using queen contiguity) are selected. Through the process of linking, the
corresponding observations are highlighted in the cluster map on the right. It is clear that
only a fraction of the locations in the High-High quadrant are actually significant (42),
shown as the bright red municipalities in the map. The non-significant ones are identified in
a grey color, as illustrated by the locations within the red rectangle on the map.
The reverse logic works as well, as shown in Figure 16.5. The High-High locations are selected
in the cluster map on the right (by clicking on the red rectangle in the legend) and the
corresponding 42 observations are highlighted in the Moran scatter plot on the left. Again,
Clusters and Spatial Outliers 303

Figure 16.4: High-High observations in Moran scatter plot

Figure 16.5: Significant High-High observations in Moran scatter plot

this confirms that just belonging to the High-High quadrant in the Moran scatter plot does
not imply significance.
A similar exercise can be carried out for observations in the Low-Low quadrant, which also
correspond with positive local spatial autocorrelation.
The interpretation of clusters is not always that straightforward (see also Section 16.5.3),
especially after inspecting a choropleth map of the variable of interest. In some instances, it
is very clear that an observations is selected as High-High when it is in the top category and
surrounded by neighbors that are also in the top category. For example, in Figure 16.6, the
location selected within the black rectangle on the cluster map (with the arrow pointed at
it) belongs to the top quartile in the box map on the left. In addition, its neighbors all also
belong to the top quartile, hence a natural interpretation as a cluster.
However, in other instances, the identification as a High-High (or similarly, Low-Low) cluster
may seem counter-intuitive. Consider the High-High location identified with the blue arrow
on the cluster map. In the box map to the left, that location is shown as belonging to the
second quadrant, i.e., below the median, which would seem to contradict the characterization
304 LISA and Local Moran

Figure 16.6: High-High cluster location

as high. However, as it turns out, for 2010 the median value of access to health care (56.9) is
above the mean (53.4), so that it is possible for an observation to be below the median, but
still above the mean. Moreover, the simplification of the continuous distribution into a small
number of discrete map categories may be confusing. In this particular example, the spatial
lag is well above the mean, even though the neighbors of the location in question consist of
a mix of municipalities from all four categories on the map. It is important to keep in mind
that the spatial lag is the average of the neighbors, and its value can be easily influenced by
a few extreme values. In addition, the categories on the map may contain observations with
very disparate values (depending on the range in the given category), which may yield less
than intuitive insights into the similarity of neighbors.

16.4.2 Spatial outliers

Spatial outliers are locations with a significant negative Local Moran statistic, i.e., locations
with dissimilar neighbors. In the terminology of the Moran scatter plot, these are observations
in the off-diagonal quadrants, i.e., High-Low or Low-High locations.
Again, the significant spatial outliers are only a subset of the observations in the respective
off-diagonal quadrant of the Moran scatter plot. For example, in Figure 16.7, the selected
High-Low spatial outliers in the cluster map on the right correspond to only 15 of the 106
observations in the lower left quadrant, as shown in the left-hand panel.
Here also, there may be some counter intuitive results. On the one hand, in the case of the
selected High-Low outlier in the purple rectangle in the right-hand panel of Figure 16.8, the
interpretation is straightforward. As seen in the left-hand panel, it consists of an observation
in the top quartile (indicated by the dark brown color), surrounded by neighbors in the
first or second quartile (blue colors), corresponding with much lower values. However, the
interpretation of the spatial outlier in the blue rectangle is less straightforward. As shown in
the left-hand panel, this corresponds with an observation in the second quadrant, i.e., below
the median, but, as before, above the mean, hence classified as high. This observation is
surrounded by neighbors in the first quadrant (i.e., with lower values), but one neighbor
belongs to the top quadrant. Again, what matters here is the average of neighboring values,
not the classifications within which they fall.
Significance and Interpretation 305

Figure 16.7: Significant High-Low observations in Moran scatter plot

Figure 16.8: High-Low spatial outlier location

16.5 Significance and Interpretation

The interpretation of the results in a cluster map needs to proceed with caution. Because
the permutation test is carried out for each location in turn, the resulting p-values no
longer satisfy the classical assumptions. More precisely, because the same data are re-used
in several of the tests, the problem is one of multiple comparisons. Moreover, the p-values of
the local statistic may be affected by the presence of global spatial autocorrelation, e.g., as
investigated by Sokal et al. (1998b), Ord and Getis (2001), among others. Consequently, the
interpretation of the significance of the clusters and outliers is fraught with problems.
In addition, the cluster map only provides an indication of significant centers of a cluster,
but the notion of a cluster involves both the core as well as the neighbors.
These issues are considered more closely in the remainder of the current section.
306 LISA and Local Moran

Figure 16.9: Significance map – p < 0.01

Figure 16.10: Cluster map – p < 0.01

16.5.1 Multiple comparisons

An important methodological issue associated with the local spatial autocorrelation statistics
is the selection of the p-value cut-off to properly reflect the desired Type I error. Not
only are the pseudo p-values not analytical, since they are the result of a computational
permutation process, but they also suffer from the problem of multiple comparisons (for a
detailed discussion, see de Castro and Singer, 2006; Anselin, 2019a). The bottom line is that
a traditional choice of 0.05 is likely to lead to many false positives, i.e., rejections of the null
when in fact it holds.
Significance and Interpretation 307

There is no completely satisfactory solution to this problem, and no strategy yields an

unequivocal correct p-value. A number of approximations have been suggested in the
literature, of which the best known are the Bonferroni bounds and False Discovery Rate
approaches. However, in general, the preferred strategy is to carry out an extensive sensitivity
analysis, to yield insight into what may be termed interesting locations, rather than significant
locations, following the suggestion in Efron and Hastie (2016).
In GeoDa, these strategies are implemented through the Significance Filter option (the
second item in the options menu).

16.5.1.1 Adjusting p-values

Adjusting the cut-off p-value as an option of the Significance Filter is straightforward.
One can select one of the pre-selected p-values from the list, i.e., 0.05, 0.01, 0.001, or 0.0001.
For example, choosing 0.01 immediately changes the locations that are displayed in the
significance and cluster maps, as shown in Figures 16.9 and 16.10. The number of significant
locations drops from 117 to 47 (for 99,999 permutations), with the 70 observations initially
labeled as significant at p < 0.05 removed from the significance map.
Whereas for p < 0.05 the High-High and Low-Low clusters were of comparable size, for p
< 0.01 there are almost twice as many Low-Low clusters as High-High. Similarly, for the
spatial outliers, which were roughly equally represented for p < 0.05, for p < 0.01 there are
more than twice as many High-Low outliers than Low-High outliers.
A sensitivity analysis by adjusting the cut-off p-value in this manner provides some insight
into which selected clusters and spatial outliers stand up to high scrutiny vs. which are more
likely to be spurious.
A more refined approach to select a p-value is available through the Custom Inference
item of the significance filter. The associated interface provides a number of options to deal
with the multiple comparison problem.
The point of departure is to set a target α value for an overall Type I error rate. In a multiple
comparison context, this is sometimes referred to as the Family Wide Error Rate (FWER).
The target rate is selected in the input box next to Input significance. Without any other
options, this is the cut-off p-value used to select the observations. For example, this could
be set to 0.1, a value suggested by Efron and Hastie (2016) for big data analysis. In the
example, this value is kept to the conventional 0.05.

16.5.1.2 Bonferroni bound

The first custom option in the inference settings dialog is the Bonferroni bound procedure.
This constructs a bound on the overall p-value by taking α and dividing it by the number of
multiple comparisons. In the LISA context, the latter is typically taken as the number of
observations, n, although alternatives have been proposed as well. These include the average
or maximum number of neighbors, or some other measure of the range of interaction.
With the number of observations at 570, the Bonferroni bound in the example would be
α/n = 0.00008772, the cut-off p-value to be used to determine significance.
Checking the Bonferroni bound radio button in the dialog immediately updates the signifi-
cance and cluster maps. This typically is a very strict criterion. For example, in the cluster
map in Figure 16.11, only two observations meet this criterion, in the sense that their pseudo
p-value is less than the cut off.
308 LISA and Local Moran

Figure 16.11: Cluster map – Bonferroni bound

16.5.1.3 False Discovery Rate

In practice, the application of a Bonferroni bound is often too strict, in that very few (if
any) observations turn out to be significant. In part, this is due to the use of the total
number of observations as the adjustment factor, which is almost certainly too large, hence
the resulting critical p-value will be too small.
A slightly less conservative option is to use the False Discovery Rate (FDR), first proposed
by Benjamini and Hochberg (1995). It is based on a ranking of the pseudo p-values from the
smallest to the largest. The FDR approach then consists of selecting a cut-off point as the
first observation for which the pseudo p-value is not smaller than the product of the rank
with α/n. With the sorted order as i, the critical p-value is the value in the sorted list that
corresponds with the sequence number imax , the largest value for which pimax ≤ i × α/n.
To illustrate this method, a few additional columns are added to the data table in the
example. First, as shown in Figure 16.12, the column LISA_p containing the pseudo
p-values is sorted in increasing order (signified by the > symbol next to the variable name).
Next, a column with the sorted order is added, as pOrder.2 A new column is added with
the FDR reference value. This is the product of the sequence number in pOrder with α/n,
which for the first observations is the same as the Bonferroni bound, shown as 0.000087 in
the table. The value for the second observation is then 2 × 0.000087 or 0.000174.3
The full result is shown in Figure 16.12, with the observations that meet the FDR criterion
highlighted in yellow. For the 17th ranked observation (location 55), the pseudo p-value of
0.00139 still meets the criterion of 0.001479, but for the 18th ranked observation, this is no
longer the case (0.001730 > 0.001556).
2 This is accomplished by means of the Table Calculator, using Special > Enumerate.
3 Again, this is accomplished in the Table Calculator. First, the new column/variable FDR is set to
the constant value α/n, using Univariate > Assign. Next, the FDR column is multiplied with the value
for pOrder, using Bivariate > Multiply.
Significance and Interpretation 309

Figure 16.12: False Discovery Rate calculation

Figure 16.13: Cluster map – FDR

Checking the False Discovery Rate radio button will update the significance and cluster
maps accordingly, displaying only 17 significant locations, as shown in Figure 16.13. At this
point, the High-High and Low-Low clusters are again fairly balanced (respectively, 6 and
7), but there are no longer any Low-High spatial outliers (there are 4 remaining High-Low
outliers).

16.5.2 Interpretation of significance

As mentioned, there is no fully satisfactory solution to deal with the multiple comparison
problem. Therefore, it is recommended to carry out a sensitivity analysis and to identify the
stage where the results become interesting. A mechanical use of 0.05 as a cut off value is
definitely not the proper way to proceed.
310 LISA and Local Moran

Figure 16.14: Cores and neighbors

Also, for the Bonferroni and FDR procedures to work properly, it is necessary to have a
large number of permutations, to ensure that the minimum p-value can be less than α/n.
Currently, the largest number of permutations that GeoDa supports is 99,999, and thus the
smallest possible p-value is 0.00001. For the Bonferroni bound to be effective, α/n must be
greater than 0.00001, or n should be less than α/0.00001. This is not due to a characteristic
of the data, but to the lack of sufficient permutations to yield a pseudo p-value that is small
enough.
In practice, this means that with α = 0.01, data sets with n > 1000 cannot have a single
significant location using the Bonferroni criterion. With α = 0.05, this value increases to
5000, and with α = 0.1 to 10,000.
The same limitation applies to the FRD criterion, since the cut-off for the first sorted
observation corresponds with the Bonferroni bound (1 × α/n).
Clearly, this drives home the message that a mechanical application of p-values is to be
avoided. Instead, a careful sensitivity analysis should be carried out, comparing the locations
of clusters and spatial outliers identified for different p-values, including the Bonferroni and
FDR criteria, as well as the associated neighbors to suggest interesting locations that may
lead to new hypotheses and help to discover the unexpected.

16.5.3 Interpretation of clusters

Strictly speaking, the locations shown as significant on the significance and cluster maps are
not the actual clusters, but the cores of a cluster. In contrast, in the case of spatial outliers,
the observations identified are the actual locations of interest.
In order to get a better sense of the spatial extent of the cluster, the Select All. . . option
offers a number of ways to highlight cores, their neighbors, or both.
The first option selects the cores, i.e., all the locations shown as non-white in the map. This
is not so much relevant for the cluster or significance map, but rather for any maps that are
linked. The selection of the Cores will select the corresponding observations in any other
map or graph window.
The next option does not select the cores themselves, but their Neighbors. Again, this is
most relevant when used in combination with linked maps or graphs.
Conditional Local Cluster Maps 311

Figure 16.15: Conditional cluster map

The third option selects both the cores and their neighbors (as defined by the spatial weights).
This is most useful to assess the spatial range of the areas identified as clusters. For example,
in Figure 16.14, the Cores and Neighbors option is applied to the significant locations
obtained using the FDR criterion. The non-significant neighbors are shown in grey on the
cluster map on the right, with all the matching cores and neighbors highlighted in the box
map on the left.
A careful application of this approach provides insight into the spatial range of interaction
that corresponds with the respective cluster cores.

16.6 Conditional Local Cluster Maps

A final option for the Local Moran statistic is that the cluster maps can be incorporated in
a conditional map view, similar to the conditional maps considered in Section 8.4.3. This is
accomplished by selecting the Show As Conditional Map option (the fourth entry in the
options menu). The resulting dialog is the same as for the standard conditional map.
In Figure 16.15, the cluster map for access to health care is turned into two micromaps,
conditioned by the percentage of the population that speaks an indigenous language on the
horizontal axis (p_IND10). An additional variable could be taken into account to condition
on the vertical axis, but that is not considered here. The number of intervals (Horizontal
Bin Breaks) was changed to two (from the default three), using Natural Breaks as the
criterion to determine the cut-off value.
The point of departure is that, in the absence of any interaction effect, the two micromaps
should have the same mix of High-High and Low-Low clusters and spatial outliers. In the
example, that does not seem to be the case, with a larger share of the Low-Low clusters on
312 LISA and Local Moran

the right (i.e., for more indigenous communities), and the bulk of the High-High clusters on
the left.
It should be noted that this example is purely illustrative of the functionality available
through the conditional cluster map feature, rather than as a substantive interpretation. As
always, the main focus is on whether the micromaps suggest different patterns, which would
imply an interaction effect with the conditioning variable(s).
17
Other Local Spatial Autocorrelation
Statistics

This chapter continues the exploration of local spatial autocorrelation statistics. First,
several extensions of the Local Moran are introduced, such as the Median Local Moran, the
Differential Local Moran and a specialized version that deals with the variance instability
in rates or proportions, the EB Local Moran. In addition, a local version of Geary’s c
statistic is discussed. The chapter also takes an in-depth look at an important class of local
statistics introduced by Getis and Ord. These statistics are not a LISA in a strict sense, but
nevertheless are effective tools to discover local hot spots and cold spots.
The chapter closes with a brief discussion of the comparative merits of the local statistics
covered so far.
These methods are again illustrated by means of the Oaxaca Development sample data set.

17.1 Topics Covered

• Assess the sensitivity of Local Moran to the use of a median spatial lag instead of an
average spatial lag
• Identify clusters and outliers in the change of a variable over time
• Correct the local Moran statistic for variance instability in rates
• Identify clusters and outliers by means of the Local Geary statistic
• Identify hot spots and cold spots by means to the Gi and Gi* statistics
• Carry out sensitivity analysis
• Compare the results yielded by different LISA
GeoDa Functions
• Space > Median Local Moran’s I
• Space > Differential Local Moran’s I
• Space > Local Moran’s I with EB rate
• Space > Univariate Local Geary
• Space > Local G
• Space > Local G*

DOI: 10.1201/9781003274919-17 313

314 Other Local Spatial Autocorrelation Statistics

17.2 Extensions of the Local Moran

Three extensions of the Local Moran are considered. Two of these are local extensions of the
specialized Moran scatter plots considered in Section 14.2.1 for the differential Moran scatter
plot, and in Section 14.2.2 for the Moran scatter plot with EB rates. The third extension
uses a different concept of spatial lag. Instead of the average of the values in neighboring
locations, the Median Local Moran uses the median. This is considered first.
As was the case for their global counterparts, the other two extensions are primarily shortcuts
to facilitate computations. They are straightforward applications of the Local Moran statistic
to pre-computed variables. In one case, it is the difference between observations at two periods
in time, yielding the Differential Local Moran, in the other case, rates are standardized to
control for variance instability, resulting in the EB Local Moran.

17.2.1 Median Local Moran

In the discussion
 of spatial autocorrelation statistics so far, a spatially lagged variable was
defined as j wij zj , or the average of the values observed at the neighboring locations.
However, as is well known, the average may be sensitive to the presence of outliers. This
may pull the average up (or down), even when many of the neighbors do not have high (low)
values, creating a potentially misleading impression of a cluster or spatial outlier.
An alternative can be based loosely on the idea of a median smoother (e.g., Wall and Devine,
2000). In the latter, the value at a location (typically a rate) is replaced by the median of
the neighboring locations (see also Section 12.3.1.3). In the context of the Local Moran, the
median of the neighbors is used in the place of the average as a median spatial lag.
Consequently, the Median Local Moran becomes:

IiM = zi × med(zj , j ∈ Ni ),

where Ni is the neighbor set of location i (i.e., those locations for which wij = 0).
Inference and interpretation are identical to that for the original Local Moran, based on a
conditional permutation approach (see Section 16.5).

17.2.1.1 Implementation
The Median Local Moran is invoked as the second item in first group of the Cluster Maps
drop-down list from the toolbar, or as Space > Univariate Median Local Moran’s I
from the menu.
This brings up the usual variable selection dialog. All the options are the same as for the
conventional Local Moran, i.e., the randomization, significance filter and saving of the results
(see Chapter 16).
These features are illustrated using the variable p_PHA(2010) (or, p_PHA10) from the
Oaxaca data set, with queen contiguity. The significance map for 99,999 permutations is
shown in Figure 17.1, with the result for the Median Local Moran in the right-hand panel,
compared to the conventional Local Moran on the left.
Overall, with a p < 0.05 cut-off, there are 18 fewer significant locations for the Median
Local Moran compared to the conventional version. These are distributed over the different
Extensions of the Local Moran 315

Figure 17.1: Significance maps conventional and Median Local Moran

categories as 10 fewer for p < 0.05, 5 fewer for p < 0.01, 2 fewer for p < 0.001 and one
at p < 0.00001, but none at p < 0.0001. However, the changes work in both directions,
with some locations becoming significant for the Median Local Moran, that were not for
the conventional Local Moran, and the other way around. In addition, there are changes in
the category of significance in both directions (both less and more significant). For example,
the blue highlight in Figure 17.1 points to a municipality that was not significant for the
conventional Local Moran, but becomes so at p < 0.01 for the Median Local Moran. In
contrast, the red highlighted location was significant for the conventional Local Moran and
becomes insignificant for the median version.

17.2.1.2 Clusters and spatial outliers

The classification of significant locations for the Median Local Moran uses the same principle
as for the conventional Local Moran, but now the reference points are the median of the
(standardized) variable and the median of its median spatial lag. In a scatter plot of the
median lag on the original variable, the center point is drawn at the location of the means,
not the medians. Consequently, the resulting quadrants are not appropriate to derive the
categories of spatial association. For example, for the standardized variable computed from
p_PHA10, the mean is obviously 0 by construction, but the median is 0.1613. Similarly,
the mean and median for the median spatial lag (on the vertical axis) are not the same,
with the latter equal to 0.1786.
As a result, the classification into High-High or Low-Low clusters or spatial outliers must be
done relative to the median values and not relative to the means shown in a conventional
scatter plot.
This is further illustrated in Figures 17.2 and 17.3 for spatial clusters, and Figures 17.4 and
17.5 for spatial outliers.1
In the left-hand panel of Figure 17.2, the observations are selected that are above the median
for the (standardized) variable and above the median for the median spatial lag. Clearly, this
selection does not match the conventional High-High criterion based on the means. In the
cluster map on the right, the subset of the selected observations that is also significant (see
Figure 17.1) is classified as High-High in red, the non-significant locations are colored grey.
A similar operation is carried out (under the hood) to find the Low-Low clusters, shown in
1 The selection is obtained using two steps in the Selection Tool: first selecting the observations that

meet the criterion for the variable, followed by a Select From Current Selection for the median spatial
lag.
316 Other Local Spatial Autocorrelation Statistics

Figure 17.2: High-High clusters in Median Local Moran

Figure 17.3: Low-Low clusters in Median Local Moran

Figure 17.3. In this case, since the median is above the mean for both axes, the selection
includes observations that fall in different quadrants for the conventional Local Moran.
The same principle is applied to the spatial outliers, illustrated in Figures 17.4 and 17.5.
Again, the classification into the respective categories differs from what the conventional
scatter plot would yield. The significant observations correspond with locations classified as
spatial outliers in the cluster map on the right.

17.2.1.3 Median Moran cluster map

A comparison of the cluster map obtained by the conventional Local Moran and the Median
Local Moran (for 99,999 permutations, with p < 0.05) is illustrated in Figure 17.6. In the
current example, the difference in significance results in clusters that are much more limited
in range compared to the conventional Local Moran. For example, both the High-High
cluster highlighted in the blue rectangle and the Low-Low cluster highlighted in the red
rectangle are much smaller for the Median Local Moran. In addition, the classification of
the significant locations changes as well. For example, in the green rectangle, two High-Low
outliers for the conventional Local Moran become part of a Low-Low cluster.
Overall, a comparison of the results for the Median Local Moran to the conventional Local
Moran provides insight into the sensitivity of the results to potential outliers. It should
Extensions of the Local Moran 317

Figure 17.4: Low-High spatial outliers in Median Local Moran

Figure 17.5: High-Low spatial outliers in Median Local Moran

Figure 17.6: Cluster maps conventional and Median Local Moran

be part of a standard sensitivity analysis, together with an assessment of different p-value

cut-offs.
318 Other Local Spatial Autocorrelation Statistics

17.2.2 Differential Local Moran

The Differential Local Moran statistic is the local counterpart to the Differential Moran
scatter plot, discussed in Section 14.2.1. Instead of using the observations on a variable at
two different time periods separately, this statistic is based on the change over time, i.e., the
difference between yt and yt−1 . Note that this is the actual difference and not the absolute
difference, so that a positive change will be viewed as high, and a negative change as low.
The differences are used in standardized form, i.e., they are not the differences between the
standardized variable at two points in time, but the standardized differences between the
original values for the variable.
The formal expression for this statistic follows the same logic as before, and consists of the
cross product of the difference between yt and yt−1 at i with the associated spatial lag:


IiD = a(yi,t − yi,t−1 ) wij (yj,t − yj,t−1 ).
j

The scaling constant a can be ignored. In essence, this is the same as the conventional
Local Moran applied to the difference, but the implementation is based on selecting the two
variables and computing the difference under the hood, rather than computing the difference
separately.
As before, inference is based on conditional permutation. All the usual caveats hold about
multiple comparisons and the choice of a p-value. In all respects, the interpretation is the
same as for the conventional Local Moran statistic.

17.2.2.1 Implementation
The Differential Local Moran is invoked as the fourth item in first group of the Cluster
Maps drop-down list from the toolbar, or as Space > Differential Local Moran’s I
from the menu.
As in the global case, the variables under consideration must be time-enabled (grouped) in
the data table. The variable selection dialog is slightly different from the standard interface,
but the same as for the differential Moran scatter plot.
Continuing to use the (time-enabled) Oaxaca data set, first, the variable of interest is selected
(here, p_PHA), and then the two time periods are chosen (here, 2020 and 2010). Note
that the system is agnostic about the actual time periods, so that any combination can be
selected. The statistic is computed for the difference between the time period specified as
the first item and that given as the second item. In the example, the spatial weights are
oaxaca_q.
To provide some context, Figure 17.7 shows the cluster maps for the conventional Local
Moran for p_PHA in 2010 and 2020 (using 99,999 permutations and p < 0.05). The local
patterns in the two years are very different, with many High-High clusters from 2010 labeled
as Low-Low in 2020, and vice versa. Note that the classification is relative to the mean,
which has improved considerably between the two years (from 53.4% to 75.8%).
The significance map and cluster map for the Differential Local Moran (99,999 permutations
with p < 0.05) are shown in Figure 17.8. Note how the High-High clusters correspond to
locations that were Low-Low in 2010, but High-High in 2020, suggesting a grouping of large
increases. Reversely, the Low-Low clusters for the difference correspond to locations that
were High-High in 2010, but Low-Low in 2020, suggesting a grouping of small increases, or
even decreases.
Extensions of the Local Moran 319

Figure 17.7: Cluster maps for Local Moran in 2010 and 2020

Figure 17.8: Differential Local Moran significance and cluster maps

All the options, such as randomization and significance filter are the same as for the
conventional local Moran and will not be further discussed here. The only slight difference is
in how the results are saved. Similar to the functionality for the differential Moran scatter
plot, the Save Results option includes an item to save the actual change variable (in raw
form, not in standardized form). In the dialog, this corresponds to the Diff Values item
with default variable name DIFF_VAL2. The other options are the same as for all local
spatial autocorrelation statistics, i.e., the value of the statistic (LISA_I), cluster type
(LISA_CL) and p-value (LISA_P).
Once the difference is saved as a separate variable, it can be used in a conventional univariate
Local Moran operation.

17.2.2.2 Interpretation
The significance and cluster maps for a Differential Local Moran identify the locations
where the change in the variable over time is matched by similar/dissimilar changes in the
surrounding locations. It is important to keep in mind that the focus is on change, and there
is no direct connection to whether this changes is from high or from low values.
Two situations can be distinguished, depending on whether the change variable takes on
both positive and negative values, or when all the changes are of the same sign (i.e., all
observations either increase or decrease over time).
320 Other Local Spatial Autocorrelation Statistics

When both positive and negative change values are present, the High-High locations will
tend to be locations with a large increase (positive change), surrounded by locations with
similar large increases. The Low-Low locations will be observations with a large decrease
(negative change), surrounded by locations with similar large decreases. Spatial outliers will
be locations where an increase is surrounded by a decrease and vice versa.
When all changes are of the same sign, the interpretation of High-High and Low-Low is
relative to the mean. When all the changes are positive, large increases surrounded by other
large increases will be labeled High-High, whereas small(er) increases surrounding by other
small(er) increases will be labeled Low-Low. When all the signs are negative, small(er)
decreases surrounded by other small(er) decreases will be labeled as High-High, whereas
large decreases surrounded by other large decreases will be labeled Low-Low.

17.2.3 Local Moran with EB Rate

The third extension of the Local Moran pertains to the special case where the variable
of interest is a rate or proportion. As discussed for the Moran scatter plot in Section
14.2.2, the resulting variance instability can cause problems for the Moran statistic. The EB
standardization suggested by Assuncao and Reis (1999) for the global case can be extended
to the local statistic in a straightforward manner. The statistic has the usual form, but is
computed for the standardized rates, z:


IiEB = azi wij zj ,
j

The standardization of the raw rate ri is the same as before and is repeated here for
completeness (for a more detailed discussion, see Section 14.2.2):

ri − β
zi = 
α + (β/Pi )

with β as an estimate of the mean and the denominator as an estimate of the standard
error.2
All inference and interpretation is the same as for the conventional case and is not further
pursued here.

17.2.3.1 Implementation
The local Moran functionality for standardized rates is invoked as the last item in the Moran
group on the Cluster Maps toolbar icon. Alternatively, it can be selected from the menu
as Space > Local Moran’s I with EB Rate.
Since the rate standardization is computed as part of the operation, the variable selection
interface is similar to that used for rate maps. To illustrate this feature, the rate is computed
with DIS20 (number of people with disabilities in 2020) as the Event Variable, and
PTOT20 (total population in 2020) as the Base Variable, as done earlier in Section
14.2.2.1. A box map of the raw rate is shown in the left hand panel of Figure 12.9. The
spatial weights are queen contiguity, oaxaca_q.
2 To recap, β =
 
O / i Pi , where Oi is the number of events at i and Pi is the population at risk.
i i
2
The estimate of α = [ i Pi (ri − β) ]/P − β/(P/n),, with n as the total number of observations, such that
P/n is the average population. Note that the estimate of α can be negative, in which case it is set to zero.
Extensions of the Local Moran 321

Figure 17.9: Significance maps – raw rate and EB rate

Figure 17.10: Cluster maps – raw rate and EB rate

The resulting significance map is shown in the right-hand panel of Figure 17.9, next to the
corresponding map for the crude rate (for 99,999 permutations with p < 0.05). Compared
to the map for the crude rate, there is one more significant location. Interestingly, the main
difference is at the high end of the significance, where the EB Local Moran has slightly more
significant locations. For example, the location highlighted in the blue rectangle is significant
at p < 0.00001 for EB Local Moran, but at p < 0.001 for the crude rate. Similarly, the
location highlighted in the red rectangle is significant at p < 0.001 for the EB Local Moran,
but at p < 0.01 for the crude rate. Since the same number of permutations is used in both
cases, the p-values are comparable. Overall, however, the differences are minimal and rather
subtle, confirming the findings for the Moran scatter plot.
The differences between the cluster map for the crude rate and for the EB Local Moran,
shown in Figure 17.10, are equally subtle. The EB Local Moran cluster map has three more
locations in the High-High group, and one less in each of the Low-High and High-Low spatial
outliers. Otherwise, the identified locations and groups match exactly.

17.2.3.2 Saving the results

As for the EB Moran scatter plot, the Save Results option includes an item to save
the actual EB rate. This operation is similar to the EB Rate standardization that can
be computed with the Table Calculator option (see Section 6.4.3.2). In the dialog, it
corresponds to the EB Rates item, with default variable name LISA_EB. The other
322 Other Local Spatial Autocorrelation Statistics

options are the same as for all local spatial autocorrelation statistics, i.e., the value of the
statistic, cluster type and p-value, with the same default variable names.

17.3 Local Geary

The Local Geary statistic, first outlined in Anselin (1995), and further elaborated upon in
Anselin (2019a), is a Local Indicator of Spatial Association (LISA) based on a measure of
attribute similarity that differs from the cross-product. As in its global counterpart, the
focus is on squared differences, or, rather, dissimilarity. In other words, small values of
the statistics suggest positive spatial autocorrelation, whereas large values suggest negative
spatial autocorrelation.
As introduced in Section 13.4.3.2, the Geary c statistic of global spatial autocorrelation
(Geary, 1954) takes on the following form:
  2
j wij (xi − xj ) /2S0
c= 
i
2
,
i (xi − x̄) /(n − 1)
 
with S0 = i j wij , and where the x in the numerator do not need to be in standardized
form, due to the squared difference. The statistic has a mean value of 1 under the null
hypothesis of spatial randomness. Significant values less than 1 indicate positive spatial
autocorrelation and values larger than 1 negative spatial autocorrelation.
After controlling for the parts in the expression that do not change with i, a local version of
the statistic can be found as (for technical details, see Anselin, 1995):

ci = wij (xi − xj )2 ,
j

or as: 
ci = (1/m2 ) wij (xi − xj )2 ,
j

with m2 = zi2 /n. Again, because of the squared difference, there is no need to standardize
x.
The sum over all the local statistics is:
  
ci = n[ wij (xi − xj )2 / zi2 ].
i i j i

In comparison, Geary’s global c statistic can be reformulated as:

 
c = [(n − 1)/2S0 ][ wij (xi − xj )2 / zi2 ].
i j i

This establishes the connection between the local and the global as:

ci = 2nS0 /(n − 1)c.
i

Hence, the Local Geary is a LISA statistic in the sense established in Section 16.2.
Local Geary 323

Figure 17.11: Local Geary significance and cluster maps

Closer examination reveals that the Local Geary statistic consists of a weighted sum of the
squared distance in attribute space for the geographical neighbors of observation i. Since
there is no cross-product involved, there is no direct relation to linear similarity. In other
words, since the Local Geary uses a different criterion of attribute similarity, it may detect
patterns that escape the Local Moran, and vice versa.
As for the Local Moran, analytical inference is based on an approximation and generally not
very reliable. Instead, the same conditional permutation procedure as for the Local Moran
is implemented. The results are interpreted in the same way, with the caveat regarding the
p-values and the notion of significance.

17.3.1 Implementation
the Local Geary can be invoked from the Cluster Maps toolbar icon, as the first item in
the fourth block in the drop-down list. Alternatively, it can be started from the main menu,
as Space > Univariate Local Geary.
The subsequent step is the same as before, bringing up the Variable Settings dialog that
contains the names of the available variables as well as the spatial weights. Everything
operates in the same way for all local statistics.
The final dialog offers window options. In the case of the local Geary, there is no Moran
scatter plot option, but only the Significance Map and the Cluster Map. The default is
that only the latter is checked, as before.
With the variable p_PHA(2010) (or p_PHA10), queen contiguity (oaxaca_q) and
99,999 permutations, the significance map and cluster map for the Local Geary statistic,
using p < 0.05 are shown in Figure 17.11.
The significance map uses the same conventions as before, but the cluster map is different.
There are four different types of associations, three for clusters and one for spatial outliers.
The classification is further elaborated upon in Section 17.3.2. Comparison with the results
for a Local Moran statistic is considered in Section 17.3.3.
Overall, there are 108 locations indicating positive spatial autocorrelation, contrasted with
only 16 locations for negative spatial autocorrelation.
All the options operate the same for all local statistics, including the randomization setting,
the selection of significance levels, the selection of cores and neighbors, the conditional map
324 Other Local Spatial Autocorrelation Statistics

Figure 17.12: Local Geary – High-High clusters

Figure 17.13: Local Geary – Low-Low clusters

Figure 17.14: Local Geary – Other clusters

option, as well as the standard operations of setting the selection shape and saving the
image.
Local Geary 325

Figure 17.15: Local Geary – Spatial outliers

17.3.1.1 Saving the results

The Save Results option again includes the possibility to save the statistic itself (Geary_I),
the cluster indication (Geary_CL) and the significance (Geary_P). The code for the
cluster classification for the Local Geary is 0 for not significant, 1 for a High-High cluster
core, 2 for a Low-Low cluster core, 3 for other positive spatial autocorrelation and 4 for
negative spatial autocorrelation.

17.3.2 Clusters and spatial outliers

The interpretation of significant locations in terms of the type of association is not as
straightforward for the Local Geary as it was for the Local Moran. In essence, this is because
the attribute similarity is a squared difference and not a cross-product, and thus has no
direct correspondence with the slope in a scatter plot. Nevertheless, the linking capability
within GeoDa can be exploited to accomplish classification, albeit an incomplete one.
The locations identified as significant and with the Local Geary statistic smaller than its
mean, suggest positive spatial autocorrelation (small differences imply similarity). Those
observations that can be classified in the upper-right or lower-left quadrants of a matching
Moran scatter plot, are identified as High-High or Low-Low. However, given that the squared
difference can cross the mean, there may be observations for which such a classification is
not possible. Those are referred to as other positive spatial autocorrelation.
For negative spatial autocorrelation (large values imply dissimilarity), it is not possible to
assess whether the association is between High-Low or Low-High outliers, since the squaring
of the differences removes the sign.
This is illustrated in Figures 17.12 through 17.15.
Figure 17.12 shows the case where significant local positive spatial autocorrelation (small
Local Geary) corresponds with points in the upper right quadrant of the Moran scatter
plot. This matches the concept of High-High clusters introduced before. The corresponding
locations are labeled as such in the Local Geary cluster map.
The case where significant local positive spatial autocorrelation corresponds with points
in the lower right quadrant of the Moran scatter plot is illustrated in Figure 17.13. This
matches the concept of Low-Low clusters, with the matching observations classified as such
in the cluster map.
326 Other Local Spatial Autocorrelation Statistics

Figure 17.16: Local Geary and Local Moran cluster maps

The third case of significant positive spatial autocorrelation is illustrated in Figure 17.14.
There are six observations for which the Moran scatter plot points appear in the lower right
and upper left quadrants, suggesting a mismatch between the cross-product and squared
difference attribute measures. Those locations are classified as Other Positive in the cluster
map.
Finally, observations with significant negative spatial autocorrelation (large Local Geary)
cannot be classified by type of spatial outlier. They are labeled as Negative in the cluster
map, as shown in Figure 17.15. All but one of these can also be found in the negative spatial
autocorrelation quadrants of the Moran scatter plot, but one is in the lower left quadrant,
again suggesting a mismatch between the cross-product and squared difference.

17.3.3 Comparing Local Geary and Local Moran

To gain further insight into the (subtle) difference between the measures of attribute
(dis)similarity for the Local Moran and Local Geary statistics, the indications of clusters
and spatial outliers obtained by each are considered more closely in what follows. Of course,
the findings are not general, and pertain only to the empirical illustration considered here.
However, it is useful to carry out such a comparison, especially when interest lies in potential
nonlinear associations, which the Local Moran is not able to capture.
Overall, the Local Geary cluster map identifies somewhat more significant locations, yielding
124 such observations relative to 117 for the Local Moran (at p < 0.05), as shown in Figure
17.16.

17.3.3.1 Clusters
Figure 17.17 identifies the matches for the 51 High-High cluster locations in the Local
Geary cluster map (on the right) in the Local Moran cluster map (on the left). Only 16 of
the 42 Local Moran High-High cluster locations belong to the set identified for the Local
Geary, but there are no mismatches in the sense of Low-Low clusters identified. Instead, the
non-matches are all not significant for the Local Moran, indicated by their grey color in the
map. Similarly, the significant High-High locations in the Local Moran cluster map that are
not identified with Local Geary are not significant in the latter (not shown).
As shown in Figure 17.18, only a subset of the 48 Low-Low clusters for the Local Moran
correspond to the 51 such clusters identified by the Local Geary. The other locations are not
Local Geary 327

Figure 17.17: Local Geary and Local Moran – High-High clusters

Figure 17.18: Local Geary and Local Moran – Low-Low clusters

Figure 17.19: Local Geary and Local Moran – Other positive clusters

significant. However, in the reverse order, there is one exception. One of the locations given
as a Low-Low cluster by the Local Moran is identified as a spatial outlier (negative) by the
Local Geary. The location in question is highlighted in a red rectangle in Figure 17.16. The
others are again not significant for Local Geary.
328 Other Local Spatial Autocorrelation Statistics

Figure 17.20: Local Geary and Local Moran – Spatial outliers

Figure 17.21: Local Moran and Local Geary – Spatial outliers

Finally, the six other spatial clusters identified in the Local Geary cluster map correspond
to non-significant locations for the Local Moran, as shown in Figure 17.19.
A careful consideration of the differences and similarities between the identified cluster
locations for the two local statistics sheds light on the extent to which the association may
be nonlinear. This should be followed by an in-depth inspection of the actual statistics and
associated attribute values.

17.3.3.2 Spatial outliers

The Local Geary cluster map indicates 16 locations with significant negative spatial auto-
correlation, compared to 27 such outliers for Local Moran. As shown in Figure 17.20, all
but one of the spatial outliers for Local Geary match outliers for Local Moran. The same
location mentioned earlier is part of a Low-Low cluster for the Local Moran, highlighted by
the red rectangle in the maps.
As Figure 17.21 illustrates, the spatial outliers in the Local Moran cluster map all match
such outliers for Local Geary or are not significant in the Local Geary cluster map, indicated
by the 11 grey polygons.
Getis-Ord Statistics 329

17.4 Getis-Ord Statistics

An early class of statistics for local spatial autocorrelation was proposed by Getis and Ord
(1992), and further elaborated upon in Ord and Getis (1995). It is derived from a point
pattern analysis logic. In its earliest formulation, the statistic consisted of a ratio of the
number of observations within a given range of a point to the total count of points. In a
more general form, the statistic is applied to the values at neighboring locations (as defined
by the spatial weights). There are two versions of the statistic. They differ in that one takes
the value at the given location into account, and the other does not.
The Gi statistic consist of a ratio of the weighted average of the values in the neighboring
locations, to the sum of all values, not including the value at the location (xi ).

j=i wij xj
Gi = 
j=i xj

In contrast, the G∗i statistic includes the value xi in both numerator and denominator:

∗ j wij xj
Gi =  .
j xj

Note that in this case, the denominator is constant across all observations and simply consists
of the total sum of all values in the data set. The statistic is the ratio of the average values
in a window centered on an observation to the total sum of observations.
The interpretation of the Getis-Ord statistics is very straightforward: a value larger than the
mean (or, a positive value for a standardized z-value) suggests a High-High cluster or hot
spot, a value smaller than the mean (or, negative for a z-value) indicates a Low-Low cluster
or cold spot. In contrast to the Local Moran and Local Geary statistics, the Getis-Ord
approach does not consider negative spatial autocorrelation (spatial outliers).3
Inference can be derived from an analytical approximation, as given in Getis and Ord (1992)
and Ord and Getis (1995). However, similar to what holds for the Local Moran and Local
Geary, such approximation may not be reliable in practice. Instead, conditional random
permutation can be employed, using an identical procedure as for the other statistics.

17.4.1 Implementation
The implementation of the Getis-Ord statistics is largely identical to that of the other local
statistics. The Local G and Local G* options can be selected from the second group in
the drop-down menu generated by the Cluster Maps toolbar icon. Alternatively, they can
be invoked from the menu as Space > Local G or Space > Local G*.
The next step brings up the Variable Settings dialog, followed by a choice of windows to
be opened. The latter is again slightly different from the previous cases. The default is to use
row-standardized weights and to generate only the Cluster Map. The Significance
Map option needs to be invoked explicitly by checking the corresponding box. In contrast
to previous cases, it is also possible to compute the Getis-Ord statistics using binary (not
row-standardized) spatial weights, yielding a simple count of the neighboring values.
3 When all observations for a variable are positive, as is the case in our examples, the G statistics

are positive ratios less than one. Large ratios (more precisely, less small values since all ratios are small)
correspond with High-High hot spots, small ratios with Low-Low cold spots.
330 Other Local Spatial Autocorrelation Statistics

Figure 17.22: Gi* signficance map and cluster map

Continuing with the same example, using p_PHA10 (or p_PHA(2010)) from the Oaxaca
data set, with queen contiguity weights, 99,999 permutations and p < 0.05 yields the
significance and cluster maps for the G∗i statistic shown in Figure 17.22. In this example,
the results are identical to those for the Gi statistic, which is not shown separately.
Overall, 117 locations are identified as significant, the same total as for the Local Moran. In
contrast to the latter, the cluster map for the Getis-Ord statistics only takes on two colors,
red for hot spots (High-High), and blue for cold spots (Low-Low).
The options menu contains the same items as for the other local statistics, such as the
randomization setting, the selection of significance levels, the selection of cores and neighbors,
the conditional map option, as well as the standard operations of setting the selection shape
and saving the image.

17.4.1.1 Saving the results

The Save Results option includes the possibility to save the statistic itself (G or G_STR),
the cluster indication (C_ID) and the significance (PP_VAL). For the Getis-Ord statistics,
there are only three cluster categories, with observations taking the value of 0 for not
significant, 1 for a High-High cluster and 2 for a Low-Low cluster.

17.4.2 Clusters and outliers

The interpretation of High-High and Low-Low clusters for the Getis-Ord statistics is somewhat
different from that of the Local Moran or Local Geary. As illustrated in Figures 17.23 and
17.24, the emphasis in the Getis-Ord statistics is on the magnitude of the neighbors.
The cluster maps on the right have, respectively, the observations in the High-High and
Low-Low clusters selected. The corresponding observations are highlighted in the Moran
scatter plot on the left.
In Figure 17.23, the High-High cluster observations all have spatial lags (average of neighbors)
above the mean, whereas the variable itself takes values both above and below its mean. For
the Low-Low cluster observations in Figure 17.24, all the spatial lags are below the mean,
with again the variable itself taking values both above and below the mean.
Consequently, the Getis-Ord statistics include significant locations as High-High clusters
that would be classified as Low-High spatial outliers according to the Local Moran. Similarly,
the Getis-Ord statistics include significant locations as Low-Low clusters that would be
Getis-Ord Statistics 331

Figure 17.23: Gi* High-High clusters

Figure 17.24: Gi* Low-Low clusters

classified as High-Low spatial outliers according to the Local Moran. The extent to which
this affects the interpretation of the spatial extent of hot spots or cold spots depends on
the relative importance of spatial outliers in the sample, but it can lead to quite different
conclusions between the two types of statistics.

17.4.3 Comparing G statistics and Local Moran

The Getis-Ord statistics typically identify the same locations as significant as the Local
Moran, but their interpretation is different. This is primarily due to the fact that the
Getis-Ord statistics do not account for spatial outliers. How much this affects results in
practice depends on each particular instance, but it is something one should be aware of.

17.4.3.1 Clusters
Figures 17.25 and 17.26 show the hot spots and cold spots selected in the Getis-Ord cluster
map on the right and identify the corresponding locations in the Local Moran cluster map
on the left. As mentioned, all the significant locations match exactly, but their classification
does not.
In Figure 17.25, the 54 hot spots in the Getis-Ord cluster map match the 42 High-High
clusters in the Local Moran cluster map as well as the 12 Low-High spatial outliers, similar
to what the Moran scatter plot in Figure 17.23 indicated.
332 Other Local Spatial Autocorrelation Statistics

Figure 17.25: Gi* and Local Moran High-High clusters

Figure 17.26: Gi* and Local Moran Low-Low clusters

On the other hand, in Figure 17.26, the 63 cold spots in the Getis-Ord cluster map correspond
with the 48 Low-Low clusters in the Local Moran cluster map, in addition to the 15 High-Low
spatial outliers. Again, this confirms the indication from the Moran scatter plot in Figure
17.24.

17.4.3.2 Spatial outliers

Figures 17.27 and 17.28 illustrate the same properties, but approached from the significant
spatial outliers in the Local Moran cluster map on the left.
In Figure 17.27, the 12 locations of significant Low-High spatial outliers are selected in the
map on the left, and their corresponding observations identified in the Getis-Ord cluster
map on the right. All locations are classified as High-High hot spots.
In Figure 17.28, the reverse is illustrated. The 15 significant High-Low spatial outliers
are selected in the Local Moran cluster map on the left. Their matching locations in the
Getis-Ord cluster map on the right are all classified as Low-Low cold spots.
Which Local Statistic to Use? 333

Figure 17.27: Gi* and Local Moran Low-High spatial outliers

Figure 17.28: Gi* and Local Moran High-Low spatial outliers

17.5 Which Local Statistic to Use?

In practice, the choice between the three main classes of local spatial autocorrelation statistics
covered so far may seem a bit bewildering. They each provide a slightly different perspective
on the notion of clusters and spatial outliers. The main difference is between the Local
Moran and the Local Geary, with the latter potentially picking up nonlinear relationships
which the Local Moran is not able to. However, the extent to which this matters in practice
depends on the particular application.
When the results between the two approaches differ much, a closer examination is in order.
For example, this could consist of inspecting the values of identified clusters and their
neighbors to look for the potential impact of outliers. Such an investigation may also be in
order when the results for the conventional and Median Local Moran differ much.
The Getis-Ord statistics are different in that they do not account for spatial outliers. Whether
this matters depends on the particular context. In general, they tend to identify the same
locations as significant as the Local Moran.
For all the local statistics, a careful consideration of significance is in order. Locations that
remain identified as clusters or outliers under different criteria (as well as using different
334 Other Local Spatial Autocorrelation Statistics

statistics) are likely interesting locations. On the other hand, observations that move in and
out of significance as the criteria change are likely spurious. The only way to address this
problem confidently is though a careful sensitivity analysis.
18
Multivariate Local Spatial
Autocorrelation

In this chapter, the concept of local spatial autocorrelation is extended to the multivariate
domain. This turns out to be particularly challenging, due to the difficulty in separating
spatial effects from the pure attribute correlation among multiple variables.
Three methods are considered. First, a bivariate version of the Local Moran is introduced,
which, similar to what is the case for its global counterpart, needs to be interpreted with
great caution. Next, an extension of the Local Geary statistic to the multivariate Local
Geary is considered, proposed in Anselin (2019a). The final approach is not based on an
extension of univariate statistics, but uses the concept of distance in attribute space, in the
form of a Local Neighbor Match Test (Anselin and Li, 2020).
To illustrate these methods, the Chicago SDOH sample data set is employed. It contains
observations on socio-economic determinants of health in 2014 for 791 census tracts in
Chicago (for a detailed discussion, see Kolak et al., 2020).

18.1 Topics Covered

• Understand the issues related to extending a spatial autocorrelation measure to a
multivariate setting
• Identify clusters and outliers with the bivariate extensions of the Local Moran statistic
• Identify clusters and outliers with the Multivariate Local Geary
• Understand the issues pertaining to interpreting the results of the multivariate Local
Geary
• Identify clusters by means of the local neighbor match test
GeoDa Functions
• Space > Bivariate Local Moran’s I
• Space > Multivariate Local Geary
• Space > Local Neighbor Match Test

18.2 The Multivariate Spatial Autocorrelation Problem

Designing a spatial autocorrelation statistic in a multivariate setting is fraught with difficulty.
The most common statistic, Moran’s I, is based on a cross-product association, which is in

DOI: 10.1201/9781003274919-18 335

336 Multivariate Local Spatial Autocorrelation

the same spirit as a bivariate correlation statistic. As a result, it is difficult to disentangle

whether the correlation between multiple variables at adjoining locations is due to the
correlation among the variables in-situ, or a similarity due to being neighbors in space.
Early attempts at extending Moran’s I to multiple variables focused on principal components,
as in the suggestion by Wartenberg (1985), and later work by Dray et al. (2008). However,
these proposals only dealt with a global statistic. A more local perspective along the same
lines is presented in Lin (2020), although it is primarily a special case of a geographically
weighted regression, or GWR (Fotheringham et al., 2002).
Lee (2001) outlined a way to separate a bivariate Moran-like spatial correlation coefficient
into a spatial part and a Pearson correlation coefficient. However, this approach relies on
some fairly strong simplifying assumptions that may not be realistic in practice.
An alternative perspective is offered in Anselin (2019a). The central idea underlying this
approach is to focus on the distance between observations in both attribute and geographical
space. A multivariate local spatial autocorrelation statistic then assesses the match between
those two concepts of distance.
More formally, the squared multi-attribute Euclidean distance between a pair of observations
i, j on k variables is given as:


k
d2ij = ||xi − xj || = (xih − xjh )2 ,
h=1

with xi and xj as vectors of observations. In some expressions, the squared distance will be
preferred, in others, the actual distance (dij , its square root) will be used.
In this approach, the overarching objective is to identify observations that are close in both
multi-attribute space and in geographical space, i.e., those pairs of observations where the
two types of distances match.

18.3 Bivariate Local Moran

The treatment of the Bivariate Local Moran’s I closely follows that of its global counterpart
(see Section 14.3.1). In essence, as outlined in more detail in Anselin et al. (2002), it is
intended to capture the relationship between the value for one variable at location  i, xi ,
and the average of the neighboring values for another variable, i.e., its spatial lag j wij yj .
Apart from a constant scaling factor (that can be ignored), the statistic is the product of
xi with the spatial lag of yi (i.e., j wij yj ). In order to make this operational and easier
to interpret, both variables should be standardized, such that their means are zero and
variances equal one. The Bivariate Local Moran is then:


IiB = xi wij yj ,
j

in the usual notation.

As is the case for its global counterpart, this statistic needs to be interpreted with caution,
since it ignores the correlation between xi and yi at location i (see also Section 14.3.1).
Bivariate Local Moran 337

A special case of the Bivariate Local Moran statistic is when the same variable is compared
in neighboring locations at different points in time. The most meaningful application is
 time period t, zt , and the other variable is for the neighbors in the
where one variable is for
previous time period, j wij zt−1 . This formulation measures the extent to which the value
at a location in a given time period is correlated with the values at neighboring locations in
a previous time period, or an inward influence. An alternative view is to consider zt−1 and
the neighbors at the current time, j wij zt . This would measure the correlation between a
location and its neighbors at a later time, or an outward influence. The first specification is
more accepted, as it fits within a standard space-time regression framework.
Inference proceeds similar to the global case, but is now conditional upon the tuple xi , yi
observed at location i. This somewhat controls for the correlation between x and y at i. The
remaining values for y are randomized and the statistic is recomputed to create a reference
distribution. The usual caveats regarding the interpretation of significance apply here as
well.

18.3.1 Implementation
The Bivariate Local Moran I is invoked as the third item in the drop-down list associated with
the Cluster Maps toolbar icon, or, from the menu, as Space > Bivariate Local Moran’s
I. The next dialog is the customary Variable Settings, which now has two columns, one for
First Variable (X), and one for Second Variable (Y). Since the Bivariate Local Moran
is not symmetric, the order in which the variables are specified matters. At the bottom of
the dialog, the Weights need to be selected.
To illustrate this feature, two variables are used from the Chicago SDOH sample data set:
the percentage children in poverty in 2014 (ChldPvt14), and a crowded housing index
(EP_CROWD). The spatial weights are nearest neighbor, with k = 6 (Chi-SDOH_k6).
Before proceeding with the actual bivariate analysis, the univariate characteristics of the
spatial distribution of each variable are considered more closely.

18.3.1.1 Univariate description of two variables

Figure 18.1 shows a box map for each of the variables. On the left, for ChldPvt14, there
are no outliers, so the box map corresponds with a simple quartile map. In contrast, the
map for EP_CROWD has 27 upper outliers, somewhat scattered around the city.
The correspondence between the two spatial distributions is highlighted in the co-location
map in Figure 18.2.1 This map shows the census tracts where the observations for both
variables fall in the same quartile category. In this example, this is appropriate, since both
variables work in the same direction (higher values are worse conditions). The classifications
match for 284 out of 791 observations, or 36%. The bulk of the matches is for the lowest
quartile (i.e., better conditions), where 109 out of 198 observations match. About half that
many locations also match for the three other categories (clearly, there are no matches for
the outliers, since those are only present in one of the maps). The co-location map provides
an initial sense of the degree of spatial correspondence between the two variables. More
formally, the (non-spatial) correlation coefficient between the two is 0.392.
1 The map is constructed after saving the map categories for both box maps, followed by Map >

Co-location Map, using the classification codes and a Box Map classification for the co-location map.
338 Multivariate Local Spatial Autocorrelation

Figure 18.1: Box maps – Child Poverty and Crowded Housing

Figure 18.2: Co-location of Child Poverty and Crowded Housing

18.3.1.2 Univariate Local Moran for each variable

To provide further context, the cluster maps for the univariate Local Moran for each variable
(using 99,999 permutations with a cut-off p-value of 0.01) are shown in Figure 18.3. In both
maps, the largest share consists of Low-Low cluster cores (118 for child poverty and 102
for crowded housing), but for child poverty this is almost equally balanced by High-High
cores (101). For crowded housing, there are clearly much fewer High-High cores (67). In
both instances, the cluster cores constitute a small number of larger regions, with very few
spatial outliers.
The correspondence between the two cluster maps is further highlighted in Figure 18.4, which
shows a co-location map for the cluster classifications.2 Almost 3/4 of the Low-Low cluster
cores match between the two variables, but this is only the case for 15 High-High cores.
Since a cluster consists of the core and its neighbors, the actual match has a somewhat wider
reach than just the cluster cores. None of the spatial outliers occur in the same location.
2 This is accomplished after saving the cluster classifications for both cluster maps, followed by Map >

Co-location Map, using the cluster codes and a LISA Map classification for the co-location map.
Bivariate Local Moran 339

Figure 18.3: Local Moran cluster map – Child Poverty and Crowded Housing

Figure 18.4: Co-location of Local Moran for Child Poverty and Crowded Housing

18.3.1.3 Bivariate analysis

The cluster maps for the Bivariate Local Moran are shown in Figure 18.5, using 99,999
permutations and a p-value cut-off of 0.01. The map on the left is for child poverty surrounded
by crowded housing, the map on the right for crowded housing surrounded by child poverty.
Clearly, the cluster maps are different, in accordance with the lack of symmetry for this
statistic. Both maps are characterized by many more spatial outliers than in the univariate
cluster maps, highlighting locations where a high/low value for one variable is surrounded
by neighbors with the opposite, i.e., low/high values.
The common cluster cores from the univariate maps (Figure 18.4) are present in both
bivariate maps. In fact, a co-location map based on the two bivariate maps yields the exact
same result as Figure 18.4. This makes intuitive sense, since a bivariate High-High cluster
requires a High value for one variable (cluster core) to be surrounded by High values for the
other variable (neighbors part of the cluster surrounding the matching cluster core). The
same holds for the Low-Low clusters.
340 Multivariate Local Spatial Autocorrelation

Figure 18.5: Bivariate Local Moran cluster map – Child Poverty and Crowded Housing

Arguably, the more interesting information follows from the location of the spatial outliers,
especially the Low-High outliers, of which there are 15 for child poverty-crowded housing
and 38 for crowded housing-child poverty. These are census tracts where a Low (good) value
for one indicator is surrounded by a High (bad) value for the other, or vice versa, more so
than expected under spatial randomness. This is information the univariate cluster maps
cannot provide.

18.3.1.4 Options
The Bivariate Local Moran has all the same options as the conventional Local Moran, as
detailed in Chapter 16. The cluster codes associated with saved results are the same as well.

18.3.2 Interpretation
The interpretation of the Bivariate Local Moran needs to be carried out very carefully. Aside
from the usual caveats about multiple comparisons and p-values, the association between
one variable and a different variable at neighboring locations needs to consider the in-situ
correlation between the two variables as well. As the discussion in the previous sections
illustrates, it is best to combine the bivariate analysis with a univariate analysis for each
variable. In addition, it is important to consider both directions of association. This will
tend to reveal strong local clustering among the two variables as well as instances where
their spatial patterns do not coincide in the form of bivariate spatial outliers.

18.4 Multivariate Local Geary

In Anselin (2019a), a multivariate extension of the Local Geary statistic is proposed. This
statistic measures the extent to which neighbors in multi-attribute space (i.e., data points that
are close together in the multidimensional variable space) are also neighbors in geographical
space. While the mathematical formalism is easily extended to many variables, in practice
one quickly runs into the curse of dimensionality.
In essence, the Multivariate Local Geary statistic measures the extent to which the average
distance in attribute space between the values at a location and the values at its geographic
Multivariate Local Geary 341

neighboring locations are smaller or larger than what they would be under spatial randomness.
The former case corresponds to positive spatial autocorrelation, the latter to negative spatial
autocorrelation.
An important aspect of the multivariate statistic is that it is not simply the superposition of
univariate statistics. In other words, even though a location may be identified as a cluster
using the univariate Local Geary for each of the variables separately, this does not mean that
it is also a multivariate cluster, and vice versa. The univariate statistics deal with distances
in attribute space projected onto a single dimension, whereas the multivariate statistics are
based on distances in a higher dimensional space. The multivariate statistic thus provides an
additional perspective to measuring the tension between attribute similarity and locational
similarity.
The Multivariate Local Geary statistic is formally the sum of individual Local Geary statistics
for each of the variables under consideration. For example, with k variables, indexed by h,
the corresponding expression is:


k 
cM
i = wij (xhi − xhj )2 .
h=1 j

This measure corresponds to a weighted average of the squared distances in multidimensional

attribute space between the values observed at a given geographic location i and those at its
geographic neighbors j ∈ Ni (with Ni as the neighborhood set of i).
To achieve comparable values for the statistics, one can correct for the number of variables
involved, as cM
i /k. This is the approach taken in GeoDa.

Inference can again be based on a conditional permutation approach. This consists of

holding the tuple of values observed at i fixed, and computing the statistic for a number of
permutations of the remaining tuples over the other locations. Note that because the full
tuple is being used in the permutation, the internal correlation of the variables is controlled
for, and only the spatial component gets altered.
The permutation results in an empirical reference distribution that represents a computational
approach at obtaining the distribution of the statistic under the null. The associated pseudo
p-value corresponds to the fraction of statistics in the empirical reference distribution that
are equal to or more extreme than the observed statistic.
Such an approach suffers from the same problem of multiple comparisons mentioned for
all the other local statistics. In addition, there is a further complication. When comparing
the results for k univariate Local Geary statistics, these multiple comparisons need to be
accounted for as well.
For example, for each univariate test, the target p-value of α would typically be adjusted
to α/k (with k variables, each with a univariate test), as a Bonferroni bound. Since the
multivariate statistic is in essence a sum of the statistics for the univariate cases, this would
suggest a similar approach by dividing the target p-value by the number of variables (k).
Alternatively, and preferable, a FDR strategy can be pursued. The extent to which this
actually compensates for the two dimensions of multiple comparison (multiple variables and
multiple observations) remains to be further investigated.3
Consequently, the interpretation of significance is difficult. In practice, extensive sensitivity
analysis is recommended, using small p-values or an FDR strategy.
3 See Anselin (2019a) for further technical discussion.
342 Multivariate Local Spatial Autocorrelation

Figure 18.6: Box map and Local Geary cluster map – Uninsured

18.4.1 Implementation
The Multivariate Local Geary is invoked as the second item in third block in the drop-
down list associated with the Cluster Maps toolbar icon, or, from the menu, as Space
> Multivariate Local Geary. The next dialog is a slightly different Multi-Variable
Settings, which now allows for the selection of several variables (not limited to two). As
usual, the spatial Weights need to be specified at the bottom of the dialog.
To illustrate this functionality, three variables are used from the Chicago SDOH sample
data set. In addition to the same two as for the Bivariate Local Moran, the percentage
children in poverty in 2014 (ChldPvt14), and a crowded housing index (EP_CROWD),
the percentage without health insurance (EP_UNINSUR) is included as well. The latter
has a correlation of 0.486 with child poverty and 0.632 with crowded housing. The spatial
weights are again nearest neighbor, with k = 6 (Chi-SDOH_k6).
As before, the univariate properties of these variables are considered first, but now from the
perspective of the Local Geary.

18.4.1.1 Univariate analysis of multiple variables

Figure 18.6 includes a box map for the percent uninsured on the left and a univariate Local
Geary cluster map on the right (for 99,999 permutations with a 0.01 p-value cut-off). The
box map contains three upper outliers, but otherwise boils down to the classification of a
quartile map. The Local Geary cluster map reveals larger regions of High-High and Low-Low
clusters, with very little evidence of negative spatial autocorrelation (only three spatial
outliers).
Box maps for the other two variables were included in Figure 18.1. The corresponding cluster
maps for the Local Geary (with 99,999 permutations and a p-value cut-off of 0.01) are shown
in Figure 18.7. The results are similar to the broad clusters revealed for the Local Moran
(Figure 18.3), but with several local differences.
The commonality in local clusters between the three variables is highlighted in the co-location
map in Figure 18.8. Overall, the three variables have 39 Low-Low and 10 High-High cluster
cores in common, reflecting the overlap in regions of low and high values found in the
univariate cluster maps.
Multivariate Local Geary 343

Figure 18.7: Local Geary cluster maps – Child Poverty, Crowded Housing

Figure 18.8: Co-location of Local Geary

18.4.1.2 Multivariate Local Geary analysis

The significance map and cluster map for the Multivariate Local Geary, with 99,999 permu-
tations and a p-value cut-off of 0.01 are shown in Figure 18.9. The main item that stands
out is the very large number of significant observations, something surely due to the use of
an inappropriate p-value cut-off. In all, 342 positive cluster cores are identified, but only two
locations that suggest negative spatial autocorrelation. Even for a p-value cut-off of 0.0001,
there are still 84 positive cluster cores. A careful sensitivity analysis is in order.

18.4.1.3 Options
All the options of the significance and cluster maps considered before remain the same.
Specifically, the Save Results option offers the same choices as for the univariate Local
Geary (Section 17.3.1.1).
344 Multivariate Local Spatial Autocorrelation

Figure 18.9: Multivariate Local Geary cluster map – Child Poverty, Crowded Housing,
Uninsured

Figure 18.10: Multivariate Local Geary cluster map and Univariate Local Geary co-location

18.4.2 Interpretation
To shed further light on where interesting locations can be found, the results for a range
of different p-value cut-offs can be investigated. In Figure 18.10, the 23 observations that
remain significant under the Bonferroni bound (p = 0.0000126) are linked to the univariate
co-location map from Figure 18.8. Only about half (12) of these locations match between
the two, pointing to other dimensions of association beyond the univariate overlap.
Further insight can be gained by focusing closer on those observations that are identified by
the Multivariate Local Geary, but not by the univariate overlap. This then becomes much
more of an exploratory exercise than a clean p-value interpretation. It should therefore be
carried out with caution.
Local Neighbor Match Test 345

Figure 18.11: Multivariate Local Geary cluster in PCP

18.4.2.1 Multivariate clusters and PCP

Further insight into what the Multivariate Local Geary identifies as interesting locations
can be gained from the parallel coordinate plot in the left-hand panel of Figure 18.11. The
three observations in the multivariate local cluster identified using the Bonferroni bound in
the right hand map are shown as three very close lines in the parallel coordinate plot. This
suggests that neighbors in geographic space are also close neighbors in multi-attribute space.
The connection between neighbors in geographic space and in multi-attribute space is further
leveraged by the concept of a Local Neighbor Match Test, considered next.

18.5 Local Neighbor Match Test

An alternative approach to visualize and quantify the trade off between geographical and
attribute similarity was suggested by Anselin and Li (2020) in the form of what is called
a Local Neighbor Match Test. The basic idea is to assess the extent of overlap between
k-nearest neighbors in geographical space and k-nearest neighbors in multi-attribute space.
In GeoDa, this becomes a simple intersection operation between two k-nearest neighbor
weights matrices (see Section 11.6.1). One matrix is derived from the distances in multi-
attribute space, the other using geographical distance. With the intersection in hand, the
probability that an overlap occurs between the two neighbor sets can be quantified. This
corresponds to the probability of drawing v common neighbors from the k out of n − 1 − k
possible choices as neighbors, a straightforward combinatorial calculation.
More formally, the probability of v shared neighbors out of k is:

p = C(k, v).C(N − k, k − v)/C(N, k),

where N = n − 1 (one less than the number of observations), k is the number of nearest
neighbors considered in the connectivity graphs, v is the number of neighbors in common
and C is the combinatorial operator. Alternatively, a pseudo p-value can be computed based
on a permutation approach, although that is not pursued in GeoDa.
346 Multivariate Local Spatial Autocorrelation

Figure 18.12: K nearest neighbor connectivity graphs

The degree of overlap can be visualized by means of a cardinality map, a special case of a
unique values map. In this map, each location indicates how many neighbors the two weights
matrices have in common. In addition, different p-value cut-offs can be employed to select
the significant locations, i.e., where the probability of a given number of common neighbors
falls below the chosen p.
In practice, the value of k may need to be adjusted (increased) in order to find meaningful
results. In addition, the k-nearest neighbor calculation becomes increasingly difficult to
implement in very high attribute dimensions, due to the empty space problem (Section
8.2.1).
The idea of matching neighbors can be extended to distances among variables obtained from
dimension reduction techniques, such as multidimensional scaling, covered in Volume 2 (see
also the more detailed discussion in Anselin and Li, 2020).
In contrast to the approach taken for the Multivariate Local Geary, the local neighbor match
test focuses on the pairwise distances directly, instead of converting these into a weighted
average. Both measures have in common that they focus on squared distances in attribute
space, rather than a cross-product as in the Moran statistics.

18.5.1 Implementation
The Nearest Neighbor Match test is invoked as the last item in the drop-down list associated
with the Cluster Maps toolbar icon, or, from the menu, as Space > Local Neighbor
Match Test. This is followed by a dialog to select the variables.
In addition to the variable names, the dialog includes options for four important parameters.
The Number of Neighbors specifies the range for which the match is explored, i.e.,
the value of k in the k-nearest neighbor weights that are computed under the hood. In
the example, 6 is used, to allow comparison with the other approaches considered in this
chapter. Next is the Variable Distance Function, with Euclidean as the default, but
Manhattan distance as the other option. A third option pertains to the Geographic
Distance Metric. Here again, Euclidean Distance is the default, but Arc Distance is
available as well, for the case where the map layer is unprojected. Finally, Transformation
Local Neighbor Match Test 347

Figure 18.13: Intersection of K nearest neighbor connectivity graphs

offers six options to adjust the variables: Raw, Demean, Standardize (Z), Standardize
(MAD), Range Adjust and Range Standardize. The default of Standardize (Z) is
the recommended approach. With this information, the two k-nearest neighbor weights are
constructed and the cardinality of the intersection computed.
In the illustration, the same three variables are employed as for the Multivariate Local Geary.
Before moving to the actual test, the logic of the approach is detailed.

18.5.1.1 Intersection between connectivity graphs

Figure 18.12 shows the connectivity graphs corresponding to geographic 6 nearest neighbors
and the multi-attribute 6 nearest neighbors (see Section 11.4.4). The graph on the left takes
on a familiar form, but the graph on the right suggests that a lot of multivariate connections
span large geographic distances.
The intersection between the two weights yields the connectivity graph illustrated in Figure
18.13.4 The result is very sparse, with most intersections consisting of a single link. This
graph structure forms the point of departure for the Local Neighbor Match Test.

18.5.1.2 Testing the local neighbor match

The degree of local neighbor match is visualized by means of a cardinality map, as shown in
Figure 18.14. This is a special case of a unique values map, where the categories correspond
to the number of matching neighbors. In this example, there are 101 observations with a
single match, and 10 observations with two matches.
The map is accompanied with a prompt to save the cardinality (default variable card) and
the associated probability (default variable cpval). Note that no probabilities are computed
for the cases where no matches occur.
4 This is accomplished by means of the Intersection command in the Weights Manager.
348 Multivariate Local Spatial Autocorrelation

Figure 18.14: Local neighbor match cardinality map

Figure 18.15: Local neighbor match significance map

Even though the largest number of matches is only two, such an occurrence is very rare
under spatial randomness, corresponding with a p-value of 0.0007. The 10 such identified
observations are highlighted in Figure 18.15, with the matching links shown in red.
Figure 18.16 provides a better sense of the rarity of the coincidence between geographic and
multi-attribute nearest neighbors. The two sets of neighbors are shown for the 10 highlighted
Local Neighbor Match Test 349

Figure 18.16: Local neighbor significant matches

Figure 18.17: Local neighbor match and Multivariate Local Geary

locations. It is clear that several nearest neighbors in attribute space are not so close in
geographic space, confirming the tension between the two notions of similarity.

18.5.2 Local neighbor match and Multivariate Local Geary

A final perspective on what insights the local neighbor match provides follows from a
comparison with the interesting locations identified by means of the Multivariate Local
Geary. In Figure 18.17, the 10 locations in the Local Neighbor Match Test map on the
left are linked to their corresponding observations in the Multivariate Local Geary cluster
map on the right. Of the 10, six are also identified by the Multivariate Local Geary, but
350 Multivariate Local Spatial Autocorrelation

four are not. Those four are instances where the average distances in attribute space to the
geographic neighbors are much larger than the individual distances for the nearest neighbors
in attribute space (as opposed to all the neighbors).
This can be assessed by investigating the linkage structure for the affected observations in
the right hand panel of Figure 18.16. For example, for the census tract identified by the
green rectangle in Figure 18.17, several of the nearest attribute neighbors are far away in
geographic space, as illustrated by the link highlighted by the blue arrow in Figure 18.16, as
well as its other links beyond the first two neighbors.
By focusing on a different aspect of the locational-attribute similarity trade-off, the Local
Neighbor Match Test provides yet another avenue to explore multivariate local clusters.
However, as pointed out, this approach does not scale well to a large number of variables,
since the empty space problem creates a large computational burden for the calculation of
k-nearest neighbors in high-dimensional attribute space.
19
LISA for Discrete Variables

So far, the application of local spatial autocorrelation statistics has been to continuous
variables. In this chapter, discrete variables are considered, and, more specifically, binary
variables. To address this context, a univariate Local Join Count statistic (Anselin and Li,
2019) and its extension to a multivariate setting are introduced. The latter allows for a
distinction between situations where the two discrete variables can co-occur (i.e., take the
value of 1 for the same location), and where they cannot (no co-location).
The principle behind the Local Join Count statistic is broadened by applying it to a subset
of observations on a continuous variable that satisfy a given constraint. The most common
application of this idea is to a specific quantile of the observations, leading to the concept of
a Quantile LISA (Anselin, 2019b).
The Chicago SDOH sample data set is again used to illustrate these methods.

19.1 Topics Covered

• Identify clusters in binary variables by means of the local join count statistic
• Distinguish between co-location and no co-location in a bivariate binary variable case
• Identify multivariate co-location clusters in binary variables with the local join count
statistic
• Apply the Quantile LISA concept to simplify the assessment of univariate and multivariate
clusters
GeoDa Functions
• Space > Univariate Local Join Count
• Space > Bivariate Local Join Count
• Space > Co-location Join Count
• Space > Univariate Quantile LISA
• Space > Multivariate Quantile LISA

19.2 Univariate Local Join Count Statistic

For binary variables, coded as 0 and 1, the global spatial autocorrelation statistic of choice
is the join-count statistic (see Cliff and Ord, 1973). This statistic consists of counting the
joins that correspond to occurrences of value pairs at neighboring locations. The three
cases are joins of 1 − 1 (so-called BB joins, for black-black), 0 − 0 (so-called WW joins, for

DOI: 10.1201/9781003274919-19 351

352 LISA for Discrete Variables

white-white) and 0 − 1 (so-called BW joins). The former two are indicators of positive spatial
autocorrelation, the latter of negative spatial autocorrelation.
The primary interest lies in identifying co-occurrences of uncommon events, i.e., situations
where observations that take on the value of 1 constitute much less than half of the sample.
The definition of what is 1 or 0 can easily be reversed to make sure this condition is met.
Therefore, the focus is on the BB join counts. While this is not an absolute requirement, the
way the inference is obtained requires that the probability of obtaining a large number of
like neighbors is small, and thus can form the basis for rejecting the null hypothesis. When
the proportion of observations with 1 is larger than half, then the probability of a small
number of neighbors with a value of 1 will be small, which is counter the overall logic.1
With the variable xi at location i taking either the value of 1 or 0, a global BB join count
statistic can be written as: 
BB = wij xi xj ,
i j

where wij are the elements of a binary spatial weights matrix. In other words, a join is
counted when wij = xi = xj = 1. In all other instances, either when there is a mismatch in
x between i and j, or a lack of a neighbor relation (wij = 0), the term on the right-hand
side does not contribute to the double sum.
Following the logic in Anselin (1995), Anselin and Li (2019) recently introduced a local
version of the BB join count statistic as:

BBi = xi wij xj ,
j

where xi,j can only take on the values of 1 and 0, and, again, wij are the elements of a
binary spatial weights matrix (i.e., not row-standardized).
The statistic is only meaningful for those observations where xi = 1, since for xi = 0 the
result will always equal zero. When xi = 1, it corresponds to the sum of neighbors for which
the value xj = 1. In this sense, it is similar in spirit to to the local second-order analysis for
point patterns outlined in Getis (1984) and Getis and Franklin (1987), where the number of
points are counted within a given distance d of an observed point. The distance cut-off d
could readily form the basis for the construction of the spatial weights wij , which yields the
join count statistic as a count of events (points) within the critical distance from a given
point (xi = 1).
The main difference between the two concepts is the underlying data structure: in the point
pattern perspective, the locations themselves are considered to be random, whereas the
local join count statistic is based on a lattice perspective. The latter considers a finite set of
known locations, for which both events (xi = 1) and non-events (xi = 0) are observed. In
point patterns analysis, one does not know the locations where events might have happened,
but did not.
In addition, the local join count statistic also has the same structure as the numerator in the
local Gi statistic of Getis and Ord (1992), when applied to binary observations (and with
a binary weights matrix). The numerator in this statistic is j wij xj , which is identical
to the multiplier in the local join count statistic. However, the difference between the two
statistics is that the local Gi includes the neighbors with xj = 1 for all locations, including
the ones where xi = 0. Such observations are ignored in the computation of the local join
1 In practice, this can easily be detected when the results show locations with 1 like neighbor to be

significant, and locations with more like neighbors not to be significant.

Univariate Local Join Count Statistic 353

Figure 19.1: Unique values Black-Hispanic tracts

count statistic as outlined above. In a sense, the local join count statistic could thus be
considered a constrained form of the local Gi statistic, limited to observations where xi = 1.
Inference can be based on a hypergeometric distribution, or, as before, on a permutation
approach. Given a total number of events in the sample of n observations as P , the magnitude
of interest is the number of neighbors of location i for which xi = 1, i.e., conditional upon
observing 1 at this location. The number of neighbors with xj = 1 is represented by pi . The
probability of observing exactly pi = p, conditional upon xi = 1 follows the hypergeometric
distribution for n − 1 data points and P − 1 events:
P −1N −P −2
p k −p
Prob[pi = p|xi = 1] = N −1i ,
ki

where ki is the number of neighbors for observation i.

Instead, a conditional permutation procedure can be followed to compute a pseudo p-value, in
the usual fashion. This is the preferred approach since the probability given by the previous
expression underestimates uncertainty, because it ignores the uncertainty associated with
observing xi = 1.
In practice, the permutation approach does not require any parametric assumptions. It is
formulated as a classical one-sided hypothesis test against the null hypothesis of spatial
randomness. In what follows, only the conditional permutation approach is considered.
Further technical details are provided in Anselin and Li (2019).

19.2.1 Implementation
To illustrate the univariate Local Join Count statistic, two variables are considered that
correspond to census tracts in Chicago with a predominant ethnic make-up. More precisely,
these are the tracts where the Hispanic population makes up more than 50% (Hisp), and
the tracts where the majority population is Black (Blk).2 A unique values map for these two
2 Note that in the original data set, the binary variables HISP50PCT and BLCK50PCT are not

computed correctly. In fact, it turns out these variables are computed in the data set using a 49% cut-off
percentage, which does not preclude co-location. The variables Blk and Hisp use the correct cut-off, so that
co-location (majority population is more than one ethnic group) is precluded by construction.
354 LISA for Discrete Variables

Figure 19.2: Local join count significance map

variables is given in Figure 19.1. The pattern reveals the strong degree of racial segregation,
a well-known characteristics of the population distribution in Chicago.
The univariate local join count statistic is invoked in the third group in the Cluster Maps
drop-down list from toolbar icon, or from the menu, as Space > Univariate Local Join
Count. This is followed by a Binary Variable Settings dialog, where the variable is
selected and the spatial weights matrix is specified. In this illustration, the spatial weights
are queen contiguity (Chi_SDOH_q). This is one of the few instances in GeoDa where
the spatial weights are not row-standardized and instead are used as binary weights.

19.2.1.1 Local significance map

The local significance maps for the Univariate Local Join count statistic for respectively
Hisp and Blk are shown in the panels of Figure 19.2. The result is for 99,999 permutations
and a 0.01 p-value cut-off. The strong impression of local clustering given by the unique
values map is confirmed by these significance maps. Unlike what is the case for other local
statistics, there is no separate cluster map, since only positive spatial autocorrelation is
considered, and there is no notion of High-High or Low-Low. The map shows those cores
of clusters with a value of 1 for the variable of interest, where the probability of finding
the observed number of neighbors with a value of 1 is highly unlikely. As before, the actual
notion of a cluster would include those neighbors.
All the usual options for a significance map are available (see Chapter 16).

19.2.1.2 Saving the results

The Save Results option saves the statistic, i.e., the number of BB joins (default variable
name JC), the total number of neighbors (NN) and the pseudo p-value (PP_VAL). These
values are only included for those observations where xi = 1. They can be used to more
closely assess the structure of specific clusters.
Bivariate Local Join Count Statistic 355

19.3 Bivariate Local Join Count Statistic

In a bivariate binary context, the two variables under consideration, say x and z, again only
take on a value of 0 and 1.
Two different situations are considered, referred to as co-location and no co-location. In
the first, it is possible for both variables to take the same value at a location. In other
words, at location i, it is possible to have xi = zi = 1. For example, this would be the case
when the variables pertain to two different types of events that can occur in all locations,
such as the presence of several non-exclusive characteristics (e.g., a city block that is both
low-density and commercial, or a housing unit that is classified as single family/multifamily
and owned/rented).
In the no co-location case, whenever xi = 1, then zi = 0 and vice versa. In other words, the
two events cannot happen in the same location. This would be case where the classification
of a location is exclusive, i.e., if a location is of one type, it cannot be of another type (e.g.,
majority ethnicity, a zoning classification, or a case-control design). As before, it is important
to make sure that the second variable, z, is a rare occurrence, appearing in less than half of
the sample.
In GeoDa, the no-colocation case is referred to as Bivariate Local Join Count, since this
is the only setting in which it is operational. As conceived, the no co-location setup does not
work for more than two variables.
In its general form, the expression for a Bivariate Local Join Count statistic given in Anselin
and Li (2019) is: 
BJCi = xi (1 − zi ) wij zj (1 − xj ),
j

with wij as unstandardized (binary) spatial weights.

The roles of x and z may be reversed, but the statistic is not symmetric, so that the results
will tend to differ whether x or z is the focus. In addition, it is important to keep in mind
that the statistic will only be meaningful when the proportion of the second variable (for
the neighbors) in the population is small.
Since the condition of no co-location ensures that 1 − zi = 0 whenever xi = 1, and vice versa,
the statistic simplifies to: 
BJCi = xi wij zj ,
j

whenever xi = 1, which are the only locations of interest.

Inference can be based on a hypergeometric distribution, considering P observations with
xi = 1 and Q observations with zj = 1. Again, a more robust alternative is based on a
permutation approach.
A pseudo p-value can be obtained from a one-sided conditional permutation test. This is
implemented by carrying out a series of ki draws (with ki as the number of neighbors for
i) for each location i where xi = 1 (and thus zi = 0). The draws are without replacement
from n − 1 data tuples (xj , zj ) of which Q observations have z = 1 (since zi = 0) and P − 1
observations have x = 1. In practice, we only need to draw the zj , since the matching xj
in the tuple are zero by construction. The number of times the resulting local join count
statistic equals or exceeds the observed value yields a pseudo p-value, in the usual way.
356 LISA for Discrete Variables

Figure 19.3: Bivariate local join count significance map

Figure 19.4: Local join count spatial outlier

19.3.1 Implementation
The Bivariate Local Join Count statistic is invoked from the third group in the drop-down
list associated with the Cluster Maps toolbar icon, or, from the menu, as Space >
Bivariate Local Join Count. Next is a Variable Settings dialog from which the two
binary variables are selected. The First Variable (X) (x) is selected from the left-hand
column in the dialog, the Second Variable (Y) (z) is taken from the right-hand column.
Two cases are illustrated in Figure 19.3. In the left-hand panel, the first variable is Hisp
with the second variable as Blk. In the right-hand panel, it is the other way around.
Both significance maps are based on queen contiguity (Chi_SDOH_q) and use 99,999
permutations with a 0.05 significance cut-off. Clearly, the statistic is not symmetric.
Co-Location Local Join Count Statistic 357

The statistic picks up the rare occasions where a tract with an ethnic majority of one type
is surrounded by tracts with a majority of the other type. In the example, this is only the
case in very few instances. Hisp is surrounded by Blk in 12 locations, but 11 of those are
only a p = 0.05. Blk is surrounded by Hisp in only 7 instances, but only one of these is
highly significant at p < 0.00001. As shown in Figure 19.4, where the neighbors are selected
in the unique values map on the left, this is a clear example of a spatial outlier, where a
Black majority tract is surrounded by all Hispanic majority tracts.
All the same options as before are available, including saving the results.

19.4 Co-Location Local Join Count Statistic

The second extension of the Local Join Count statistic to multiple variables considers
co-location. This allows two or more variables to take a value of 1 at the same location, i.e.,
xi = zi = 1.
For two variables, a co-location cluster requires that an observation for which xi = zi = 1
coincides with neighbors for which xj = zj = 1 as well. For two variables, the corresponding
Local Join Count statistic takes the form (Anselin and Li, 2019):

CLCi = xi zi wij xj zj ,
j

with wij as unstandardized (binary) spatial weights. As before, there are P observations
with xi = 1 and Q observations with zi = 1 out of a total of n.
A conditional permutation approach can be constructed for those locations with xi = zi = 1.
The permutation consists of draws of ki pairs of observations (xj , zj ) from the remaining set
of n − 1 tuples, which contain P − 1 observations with xj = 1 and Q − 1 observations with
zj = 1. In a one-sided test, the number of times are counted where the statistic equals or
exceeds the observed join count value at i.
The extension to more than two variables is mathematically straightforward. At each location
i, k variables are considered, i.e., xhi , for h = 1, . . . , k, with Πkh=1 xhi = 1, which enforces
the co-location requirement.
The corresponding statistic is then:

CLCi = Πkh=1 xhi wij Πkh=1 xhj .
j

The implementation of a conditional permutation strategy follows as a direct generalization of

the two-variable co-location case. However, for a large number of variables, such co-locations
become less and less likely, and a different conceptual framework may be more appropriate.

19.4.1 Implementation
The Co-Location Local Join Count statistic is invoked from the third group in the drop-
down list associated with the Cluster Maps toolbar icon, or, from the menu, as Space >
Co-location Local Join Count. This is followed by a Multi-Variable Settings dialog,
where the variables to be considered can be specified. At the bottom of the dialog is the
customary drop-down list with the spatial weights.
358 LISA for Discrete Variables

Figure 19.5: Co-location local join count

To illustrate this statistic, the two variables are Blk, tracts with majority Black population,
and CAR, tracts where more than 50% of the commutes happen by car. The spatial weights
are again queen contiguity, Chi_SDOH_q.
A co-location map of the two binary variables is shown in the left-hand panel of Figure
19.5. Of the 791 tracts, 167 are both majority Black and majority commute by car. Of the
remainder, 254 are neither and 370 show a mismatch of the majorities. The corresponding
Co-Location Local Join Count significance map is shown in the right-hand panel, using
99,999 permutations and a cut-off of 0.05. At 0.05, there are 90 cores of clusters that overlap,
out of the 167, but at 0.01, only 57 of those remain. They confirm the impression of high
dependence on commuting by car in majority Black neighborhoods.
All the usual options are available.

19.5 Quantile LISA

The Local Join Count statistics are designed to deal with binary variables. The principles
behind these statistics can be readily extended to the situation where a subset of observations
on a continuous variable is selected based on a specific value interval. The most straightforward
example is when all the observations in a given quantile are assigned a value of 1, and the
remaining observations become 0. As a result, assessing local spatial autocorrelation in this
case can be approached as an extension of the Local Join Count statistics.
In Anselin (2019b), this case is referred to as Quantile LISA or quantile local spatial
autocorrelation. In a bivariate or multivariate setting, the quantile LISA often serves as a
viable alternative to a Bivariate Local Moran or a Multivariate Local Geary, especially when
the focus is on extremes in the distribution.
As discussed in Section 18.3, the continuous (linear) association between two variables that is
measured by the Bivariate Local Moran suffers from the problem of in-situ correlation. The
quantile local spatial autocorrelation sidesteps this problem by converting the continuous
variable to a binary variable that takes the value of 1 for a specific quantile. By considering
Quantile LISA 359

either co-location (positive in-situ correlation) or no co-location (negative in-situ correlation),

the correlation between the two variables can be controlled for.
For example, the spatial association between two continuous variables can be simplified
by focusing on the local autocorrelation between binary variables for the upper quintile.
Since the bivariate (or multivariate) Local Join Count statistic enforces co-location, the
problem of in-situ correlation is controlled for. This provides an alternative to other bivariate
and multivariate local spatial autocorrelation statistics, albeit with a loss of information,
associated with going from the continuous variable to a binary category.
In addition, the quantile LISA approach can also be employed to assess negative spatial
autocorrelation, such as between the upper and lower quantile of the same variable. This
becomes a special case of a No-Colocation Bivariate Local Join Count statistic.
The quantile approach focuses on a subset of a distribution and thus constitutes a loss
of information. However, in practice, this loss of information is often compensated for by
superior insight and a focus on the most important interesting locations.
Formally, a continuous variable y with a cumulative density function F (y) yields a sequence
of ranked observations as y1 , y2 , . . . , yn . A new binary variable x is created that takes on
the value xi = 1 for all yi for which

yql ≤ yi < yqu ,

with yql and yqu as lower and upper bounds for a given quantile.3 For all other observations,
xi = 0.
The new x variable (or set of such variables in a multivariate case) then forms the basis for
analysis by means of a Local Join Count statistic, or one of its multivariate extensions.

19.5.1 Implementation
The Univariate and Multivariate Quantile LISA are invoked from the next to last block in
the drop-down list associated with Cluster Maps icon on the toolbar, or, from the menu,
as Space > Univariate Quantile LISA and Space > Multivariate Quantile LISA
respectively.
The Quantile LISA Dialog contains several options. In the univariate case, the continuous
variable needs to be specified, together with the spatial weights. Next follow the criteria to
create the binary form of the variable: the Number of Quantiles, Select a Quantile for
LISA and Save Quantile Selection in Field. The default number of quantiles is 5, for
quintiles. The quantile selection consists of a drop-down list with appropriate values, i.e.,
in this example 1 to 5, with 1 for the lowest quantile (in this case, quintile), and 5 for the
highest. The resulting binary variables is added to the data table with a default variable
name of QT.
In the multivariate case, a similar interface allows for the creation of one binary variable
at a time, with for each the number of categories and the order of the selected category,
saved under default variables QTk, with k as the sequence number. Each newly created
variable must be moved to the right-hand panel of the dialog by means of the > button
(and, conversely, can be removed by means of the < button).
3 In general, this may be applied to any interval, not just a given quantile.
360 LISA for Discrete Variables

Figure 19.6: Univariate Quantile LISA

A check box determines whether No co-location must be enforced. The default is to allow
co-location. A warning message results in case of no overlap when there should be overlap,
and vice versa.
With the spatial weights selected, the analysis is run and a significance map is created.
This is illustrated with three specific cases: a univariate Quantile LISA, and a bivariate and
multivariate example. All the standard options are available.

19.5.2 Univariate Quantile LISA

To illustrate the univariate Quantile LISA case, consider the upper quintile of the variable
ChldPvt14, the children poverty rate. A quartile map (but not a quintile map) was shown
earlier in Figure 18.1. The upper quintile is converted into a binary variable ChildPvt5,
shown in the unique values map in the left-hand panel of Figure 19.6.
The right-hand panel shows the significance map using 99,999 permutations with a p-value
cut-off of 0.05, for queen contiguity (Chi_SDOH_q). The highlighted locations form a
subset of the cores identified by means of the Local Moran map in the left-hand panel of
Figure 18.3. Of the 101 High-High cluster cores at p=0.01 in the Local Moran cluster map,
61 are retained in the Quantile LISA significance map. Three of these are highly significant
at p=0.00001.
Obviously, Low-Low clusters and spatial outliers are not considered. Nevertheless, the
Quantile LISA approach identifies essentially the same interesting High-High cluster locations
as the analysis that uses the full continuous distribution.

19.5.3 Bivariate and Multivariate Quantile LISA

The bivariate case allows for both no co-location (box checked) and co-location, as specified
through the Multivariate Quantile LISA Dialog. The No co-location option only
works in the bivariate case and provides a way to address negative spatial autocorrelation,
i.e., spatial outliers. For example, this can be used to assess whether observations in a top
quantile are surrounded by locations in a bottom quantile, or vice versa (as pointed out, the
statistic is not symmetric).
Quantile LISA 361

Figure 19.7: Spatial outliers with Quantile LISA

The co-location option is the default and it allows to identify clusters of observations that
belong to specified quantiles for different variables. Typically, this will be the top or bottom
quantiles, but the tool is sufficiently flexible to allow any combination (provided it makes
sense). A critical constraint is that there need to be co-location of the quantiles for all
the variables considered. For example, if three variables are taken into account, then there
must be locations (at least one), where the respective quantiles coincide. Just as for the
multivariate local join count, this becomes harder to satisfy as more variables are being
considered.

19.5.3.1 Spatial outliers

The No co-location option is illustrated in Figure 19.7 for the lower and upper quintiles of
the variable ChldPvt14, using 99,999 permutations with p = 0.05 and queen contiguity
spatial weights. Only one census tract can be identified that belongs to a bottom quintile and
is surrounded by more neighbors from the top quintile than likely under spatial randomness.
Moreover, the evidence is very weak and only holds for p = 0.05. Note that the same location
was also identified as a Low-High spatial outlier in the Local Moran cluster map in Figure
18.3.
The reverse assessment, for top quintiles surrounded by bottom quintiles, does not yield a
single significant location.

19.5.3.2 Multivariate co-location

The multivariate co-location case is illustrated with the same three variables as used for the
Multivariate Local Geary in Section 18.4.1, but now for queen contiguity spatial weights. The
three variables, child poverty (ChldPvt14), crowded housing (EP_CROWD) and lack
of health insurance (EP_UNINSUR) are converted into indicator variables with a value
of 1 for those observations belonging to the top quintile. The corresponding Multivariate
Quantile LISA significance map is shown in the left-hand panel of Figure 19.8, for 99,999
permutations and p = 0.05.
362 LISA for Discrete Variables

Figure 19.8: Multivariate Quantile LISA and Multivariate Local Geary

Eight cluster cores are identified, of which three at p = 0.001. The cluster map on the right
shows the significant locations for the corresponding Multivariate Local Geary (using the
Bonferroni bound for 0.01 and 99,999 permutations). The Multivariate Local Geary cluster
map includes 35 cluster cores. Of those, only two are also identified by the Multivariate
Quantile LISA. The six others, are not found to be significant in the Multivariate Local
Geary cluster map.
In contrast to the large number of significant locations obtained with the Multivariate Local
Geary, the Quantile version focuses on a much smaller (sub)set of significant observations
and may thus provide clearer insight into the interesting locations.
20
Density-Based Clustering Methods

In this last chapter dealing with local patterns, density-based clustering methods are
considered. These approaches search the data for high density subregions of arbitrary shape,
separated by low-density regions. Alternatively, the elevated regions can be interpreted as
modes in the spatial distribution over the support of the observations.
The methods covered form a transition between the local spatial autocorrelation statistics
and the regionalization methods considered in Volume 2. They pertain primarily to point
patterns, but can also be extended to a full multivariate setting. At first sight, density-based
clustering methods may seem similar to spatially constrained clustering, but they are not
quite the same. In contrast to the regionalization methods considered in Volume 2, the result
of density based clustering does not necessarily yield a complete partitioning of the data. In
a sense, the density-based cluster methods are thus similar in spirit to the identification of
clusters by means of local spatial autocorrelation statistics, although they are not formulated
as hypothesis tests. Therefore, these methods are included in the discussion of local spatial
autocorrelation, rather than with the regionalization methods considered in Volume 2.
Attempts to discover high density regions in the data distribution go back to the classic
paper on mode analysis by Wishart (1969), and its refinement in Hartigan (1975). In the
literature, these methods are also referred to as bump hunting, i.e., looking for bumps (high
regions) in the data distribution.
In this chapter, the focus is on the application of density-based clustering methods to
the geographic location of points, but the methods can be generalized to locations in
high-dimensional attribute space as well.
Four approaches are considered. First is a simple heat map as a uniform density kernel
centered on each location. The logic behind this graph is similar to that of Openshaw’s
Geographical Analysis Machine (Openshaw et al., 1987) and the approach taken in spatial
scan statistics (Kulldorff, 1997), i.e., a simple count of the points within the given radius.
This is also the main idea behind the Getis-Ord local statistics considered in Chapter 17.
The remaining methods are all related to DBSCAN (Ester et al., 1996), i.e., Density-Based
Spatial Clustering of Applications with Noise. Both the original DBSCAN is outlined, as
well as its improved version, referred to as DBSCAN*, and its Hierarchical version, referred
to as HDBSCAN, or, sometimes, HDBSCAN* (Campello et al., 2013, 2015).
The methods are illustrated with the Italy Community Banks sample data set that contains
the locations of 261 community banks.

DOI: 10.1201/9781003274919-20 363

364 Density-Based Clustering Methods

20.1 Topics Covered

• Interpret the results of a uniform kernel heat map
• Understand the principles behind DBSCAN
• Set the parameters for a DBSCAN approach
• Interpret the clustering results yielded by DBSCAN
• Understand the difference between DBSCAN and DBSCAN*
• Analyze the dendrogram generated by DBSCAN*
• Understand the logic underlying HDBSCAN
• Interpret the condensed tree and clustering results yielded by HDBSCAN
• Understand soft clustering and outlier identification in HDBSCAN
GeoDa Functions
• Map option > Heat Map
• Clusters > DBscan
• Clusters > HDBscan
Toolbar Icons

Figure 20.1: Clusters > DBSCAN | HDBSCAN

20.2 Heat Map

The Heat Map as implemented in GeoDa is a simple uniform kernel for a given bandwidth. A
circle with radius equal to the bandwidth is centered on each point in turn, and the number
of points within the radius is counted. There is no correction for distance decay. The density
of the points is reflected in the color shading.
The Heat Map follows the same principle as the idea underlying the Geographical Analysis
Machine or GAM of Openshaw et al. (1987), where the uniform kernel is computed for a
range of bandwidths to identify high intensity clusters. However, in the implementation
considered here, no significance is assigned, and the heat map is used primarily as a visual
device to highlight differences in the density of the points.

20.2.1 Implementation
Unlike the other methods considered in this Chapter, the Heat Map is not invoked from the
cluster menu, but as the Heat Map option on any point map. This brings up a dialog with
various options. The top item is to Display Heat Map, which, when checked, brings up
the default, with the Heat Map on Top item active.
The default Heat Map is constructed for the max-min bandwidth that ensures that each
location has at least one other location within its circle. This is the same default as used to
Heat Map 365

Figure 20.2: Default heat map

specify the distance-band spatial weights (see Section 11.3.1). In most applications, this is
not very informative, as shown in Figure 20.2, which depicts the locations contained in the
Italy Community Banks sample data set.1 The Heat Map Bandwidth Setup Dialog in
the example indicates a value of 124,660 (meters), the same as for the distance-band spatial
weights used as an illustration in Chapter 11. The map can be customized using the various
available options, of which the most important is to Specify Bandwidth.

20.2.1.1 Options
In Figure 20.3, a customized version is shown, where advantage is taken of several of the
options. First, the bandwidth is set to 73,000 (meters), the same as in Section 11.4.2 (Figure
11.8). In addition, the Change Fill Color is used to set the color to blue, with the Change
Transparency to 0.95.
These settings provide a much clearer indication of the varying density over the map, with
different gradations of the color corresponding to high and low density regions. In particular,
the area in the North of the country is highlighted as a high density location.
A final option, Specify Core Distance, only applies to the output of an HDBSCAN
clustering routine (see Section 20.5.5.3).
1 The map shown in the figure has the default point colors changed to black, with the point size as 3. In

addition a second layer consisting of the outlines of the Italian regions is added, of which the Only Map
Boundary option is selected.
366 Density-Based Clustering Methods

Figure 20.3: Heat map with bandwidth 73 km

20.3 DBSCAN
The DBSCAN algorithm was originally outlined in Ester et al. (1996) and Sander et al.
(1998) and was more recently elaborated upon in Gan and Tao (2017) and Schubert et al.
(2017). Its logic is similar to that just outlined for the uniform kernel. In essence, the method
again consists of placing circles of a given radius on each point in turn, and identifying those
groupings of points where a lot of locations are within each others range.

20.3.1 Important concepts

The DBSCAN algorithm introduces some specialized terminology to characterize the con-
nections between points as well as the ensuing network structure.

20.3.1.1 Core, Border and Noise points

Points are classified as Core, Border or Noise depending on how many other points are
within a critical distance band, the so-called Eps neighborhood. To visualize this, Figure
20.4 contains nine points, each with a circle centered on it with radius equal to Eps.2 Note
that in DBSCAN, any distance metric can be used, not just Euclidean distance as in this
illustration.
2 The figure is loosely based on Figure 1 in Schubert et al. (2017).
DBSCAN 367

Figure 20.4: DBSCAN Core, Border and Noise points

A second critical concept is the number of points that need to be included in the distance
band in order for the spatial distribution of points to be considered as dense. This is the
so-called minimum number of points, or MinPts criterion. In the example, this is set to four.
However, in contrast to the convention used before in defining k-nearest neighbors (Section
11.3.2), the MinPts criterion includes the point itself. So a MinPts = 4 corresponds to a
point having 3 neighbors within the Eps radius .3
In Figure 20.4, all red points with associated red circles have at least four points within the
critical range (including the central point). They are labeled as Core points (the letter C in
the graph) and are included in the cluster (connected by the dashed blue lines).
The points in magenta (labeled B), with associated magenta circles have some points within
the distance range (one, to be precise), but not sufficient to meet the MinPts criterion. They
are potential Border points and may or may not be included in the cluster.
Finally, the blue point (labeled N) with associated blue circle does not have any points
within the critical range and is labeled Noise. Such a point cannot become part of any
cluster.

20.3.1.2 Reachability
A point is directly density reachable from another point if it belongs to the Eps neighborhood
of that point and is one of MinPts neighbors of that point. This is not a symmetric
relationship.
In the example, any red point C is directly density reachable from at least two other red
points. Any such point pairs are within each others critical range. However, for border points
B, the relationship only holds in one direction, as shown by the arrow. They are directly
density reachable from a neighbor C, but since the B points only have one neighbor, their
range does not meet the minimum criterion. Therefore, the neighbor C is not directly density
reachable from B.
3 This definition of MinPts is from the original paper. In some software implementations, the minimum

points pertain to the number of nearest neighbors, i.e., MinPts − 1. GeoDa follows the definition from the
original papers.
368 Density-Based Clustering Methods

Figure 20.5: Connectivity graph for Eps = 20

Figure 20.6: DBSCAN clusters for Eps = 20 and MinPts = 4

A chain of points in which each point is directly density reachable from the previous one is
called density reachable. In order to be included, each point in the chain has to have at least
MinPts neighbors and could serve as the core of a cluster. All the points labeled C in the
figure are density reachable, highlighted by the dashed connections between them.
In order to decide whether a border point should be included in a cluster, the concept of
density connected is introduced. A point becomes density connected if it is connected to a
density reachable point. For example, the points B are each within range of a point C that
is itself density reachable. As a result, the Border points B become included in the cluster.

20.3.2 DBSCAN algorithm

In DBSCAN, a cluster is defined as a collection of points that are density connected and
maximize the density reachability. The algorithm starts by randomly selecting a point
and determining whether it can be classified as Core – with at least MinPts in its Eps
neighborhood – or instead as Border or Noise. A Border point can later be included in a
cluster if it is density connected to another Core point. It is then assigned the cluster label
of the core point and no longer further considered. One implication of this aspect of the
algorithm is that once a Border point is assigned a cluster label, it cannot be assigned to a
DBSCAN 369

different cluster, even though it might actually be closer to the corresponding core point. In
a later version, labeled DBSCAN* (considered in Section 20.4), the notion of border points
is dropped, and only Core points are considered to form clusters.
The algorithm systematically moves through all the points in the data set and assesses their
range. The search for neighbors is facilitated by using an efficient spatial data structure,
such as an R* tree. When two clusters are density connected, they are merged. The process
continues until all points have been evaluated.
Two critical parameters in the DBSCAN algorithm are the distance range and the minimum
number of neighbors. In addition, sometimes a tolerance for noise points is specified as well.
The latter constitute zones of low density that are not deemed to be interesting.
In order to avoid any noise points, the critical distance must be large enough so that every
point is at least density connected. An analogy is the specification of a max-min distance band
in the creation of spatial weights, which ensures that each point has at least one neighbor. In
practice, this is typically not desirable, but in some implementations a maximum percentage
of noise points can be set to avoid too many low density areas.4
Ester et al. (1996) recommend that MinPts be set at 4 and the critical distance adjusted
accordingly to make sure that sufficient observations can be classified as Core. As in other
cluster methods, some trial and error is typically necessary. Having to find a proper value
for Eps is often considered a major drawback of the DBSCAN algorithm.

20.3.2.1 Illustration
To illustrate how the DBSCAN proceeds, a toy example of nine points is depicted in Figure
20.5. The inter-point distances are included on the graph. The typical point of departure is to
consider a connectivity structure that ensures that each point has at least one neighbor, i.e.,
the max-min critical distance. In the example, this distance is 29.1548, with the corresponding
number of neighbors in the connectivity graph ranging from 1 to 5.
The concept of Noise points (i.e., unconnected points for a given critical distance) becomes
clear when the Eps distance is set to 20. In Figure 20.5, this results in one unconnected
point, labeled 3, shown by the dashed blue line between 3 and 5. The green lines correspond
with inter-point distances that do not meet the Eps criterion either, but they do not result
in isolates. The effective connectivity graph for Eps = 20 is shown in solid blue.
With MinPts as 4 (i.e., 3 neighbors for a core point), the algorithm proceeds through each
point, one at a time. For example, it could start with point 1, which only has one neighbor
(point 2) and thus does not meet the MinPts criterion. Therefore, it is initially labeled
as Noise. Next, point 2 is considered. It has 3 neighbors, meaning that it meets the Core
criterion, and therefore it is labeled cluster 1. All its neighbors, i.e., 1, 4 and 5 are also
labeled cluster 1 and are no longer considered. Note that this changes the status of point 1
from Noise to Border. Point 3 has no neighbors and is therefore labeled Noise.
Next is point 6, which is a neighbor of point 4 that belongs to cluster 1. Since 6 is therefore
density connected, it is added to cluster 1 as a Border point. The same holds for point 7,
which is similarly added to cluster 1.
Point 8 has two neighbors, which is insufficient to reach the MinPts criterion. Even though
it is connected to point 7, which belongs to the cluster, 7 is not a Core point, so therefore 8
4 GeoDa currently does not support this option.
370 Density-Based Clustering Methods

Figure 20.7: DBSCAN clustering settings

is not density connected and is labeled Noise. Finally, point 9 has only point 8 as neighbor
and is therefore labeled Noise as well.
This results in one cluster consisting of 6 points, shown as the dark blue points in Figure
20.6. The solid blue lines show the connections between the Core points, the dashed lines
show the Border points. The light blue points for 3, 8 and 9 indicate Noise points.

20.3.3 Implementation
DBSCAN is invoked from the cluster toolbar icon (Figure 20.1) as the first item in the density
cluster group, or from the menu as Clusters > DBScan . This brings up the DBScan
Clustering Settings dialog, which combines two functions. One is the specification of
variables and various cluster options, shown in the left-hand panel of Figure 20.7. The other
is the presentation of Summary characteristics of the resulting clusters, and, for DBSCAN*,
the Dendrogram (see Section 20.4.2.1).
The Select Variables panel includes the variables that must be chosen as X and Y
coordinates. For the Italy Community Banks sample data set, the variables XKM and
YKM are specified, which correspond with the projected coordinates expressed in kilometers
(the original is in meters). The selected Method is DBScan.
The selection of the coordinates immediately populates some of the fields in the Parameters
panel. The default Transformation is Standardize (Z), with the Distance Threshold
(epsilon) as 0.443177. However, in this instance, the original coordinates should be used,
hence the Raw option is specified in Figure 20.7. This yields a critical distance of 124.659981.
DBSCAN 371

Figure 20.8: Default DBSCAN cluster map

As it turns out, a slight correction is needed, and the value is rounded to 124.660 in the
dialog.
The other options consist of the Min Points, which is initially left to the default of 4,
and the Distance Function, set to Euclidean (the other option is Manhattan distance).
The Min Cluster Size option is not available for DBSCAN.
A final item to be specified is the variable that will contain the cluster identifier (Save
Cluster in Field), with CL as the default. In GeoDa the identifiers are assigned in decreasing
order of the size of the cluster, i.e., with 1 for the largest cluster. Points that remain as
Noise points are assigned a value of 0.

20.3.3.1 Cluster map and summary

The resulting cluster map for the default settings is shown in Figure 20.8. Only two clusters
are identified, one very large one, consisting of 242 bank locations on the mainland, the
other, a very small one of 19 banks on the island of Sicily. Given the choice of the max-min
bandwidth, there are no Noise points.
The Summary in the right-hand panel of Figure 20.7 lists the parameters used, the cluster
centers and within-cluster sum of squares, as well as an overall summary of the fit, the ratio
of between to total sum of squares. The value of 0.288 is very low, which is not surprising,
given the large number of points contained in the main cluster, which results in a high
within-cluster sum of squares. The specifics of these summary measures are discussed in
more detail in the treatment of spatial clustering in Volume 2.

20.3.3.2 Sensitivity analysis

In most applications, the default parameters are just a point of departure and considerable
sensitivity analysis is needed to assess the effect of changes in the settings. Specifically, the
372 Density-Based Clustering Methods

Figure 20.9: DBSCAN cluster map, Eps=50, Min Pts=10

two main parameters of Distance Threshold and Min Points play a crucial role. As an
illustration, the results of a cluster analysis with a bandwidth of 50 km and the minimum
points set to 10 are given in Figure 20.9. In addition to the cluster points, identified by
different colors, the connectivity graph for a 50 km distance band is shown as well. Seven
clusters are identified, ranging in size from 110 to 10 observations. A total of 70 locations
are labeled Noise, either because they become isolates (not connected to the graph), or they
have insufficient neighbors. The overall fit of the clusters is much better, with a between to
total ratio of 0.957 (not shown), due to a grouping of locations that are relatively close by.

20.3.3.3 Border points and minimum points

In some instances, the outcome of a DBSCAN algorithm can lead to seemingly counter
intuitive results. For example, using the Italian bank data with a 50 km bandwidth and 8
minimum points yields 10 clusters and a good fit of 0.960. However, cluster 10 consists of
only 6 observations, less than the minimum points requirement (results not shown). This is
a result of the peculiar way in which Border points are treated. Specifically, as mentioned
above, they are assigned to the first set of density reachable points to which it is density
connected, even though it may be closer to a later cluster. The specific implementation of
the algorithm does not change the allocation, but does recognize the later clusters as such,
even with less than minimum points.
DBSCAN* 373

20.4 DBSCAN*
One of the potentially confusing aspects of DBSCAN is the inclusion of so-called Border
points in a cluster. The DBSCAN* algorithm, outlined in Campello et al. (2013) and further
elaborated upon in Campello et al. (2015), does away with the notion of border points, and
only considers Core and Noise points.
Similar to the approach in DBSCAN, a Core object is defined with respect to a distance
threshold (Eps) as the center of a neighborhood that contains at least Min Points other
observations. All non-core objects are classified as Noise (i.e., they do not have any other
point within the distance range Eps). Two observations are Epsilon reachable if each is part
of the Epsilon-neighborhood of the other. Core points are density-connected when they are
part of a chain of Epsilon-reachable points. A cluster is then any largest (i.e., all eligible
points are included) subset of density connected pairs. In other words, a cluster consists of a
chain of pairs of points that belong to each others Epsilon-neighborhoods.

20.4.1 Mutual reachability distance

An alternative perspective on the concept of Epsilon-neighborhoods is to consider the
duality between density of points within a given radius and the distance needed to reach a
given number of neighbor points. More specifically, the requirement of having k neighbors
(Minimum Points − 1) within a given Epsilon distance range can be replaced by considering
the k-nearest neighbor distance for a point. The smaller the distance required to reach the
k-nearest neighbor, the denser the point distribution will be, and vice versa. In DBSCAN*
and HDBSCAN (see Section 20.5), this is called the core distance (dcore ) for a given k. The
larger the core distance, the less dense is the point distribution around the reference location.
This consideration leads to a new concept of distance between two points A and B, defined
as the mutual reachability distance. This is a compromise between the actual inter-point
distance, dAB and the density of the points, reflected in the respective core distances, dcore (A)
and dcore (B). Specifically, even though A and B may be close, if they are in low density
regions, their inter-point distance will be replaced by the larger of dcore (A) and dcore (B),
effectively pushing the points further away from each other. Formally, the mutual reachability
distance between A and B is defined as:

dmr (A, B) = max[dcore (A), dcore (B), dAB ].

This mutual reachability distance replaces the original distance dAB in the connectivity
graph that forms the basis for the determination of clusters (e.g., a graph like Figure 20.5).
Unless A and B are mutual k-nearest neighbors, this effectively replaces the inter-point
distances by a core distance.5
In DBSCAN*, the adjusted connectivity graph is reduced to a minimum spanning tree or
MST (see Section 3.4) to facilitate the derivation of clusters. Rather than finding a single
solution for a given distance range as in DBSCAN, the longest edges in the MST are cut
sequentially to provide a solution for each value of inter-point mutual reachability distance.
This boils down to finding a cut for smaller and smaller values of the core distance. It yields
a hierarchy of clusters for decreasing values of the core distance.
5 The largest k-nearest neighbor distance will always be larger than the inter-point distance, unless A and

B are mutual k-nearest neighbors, in which case dmr (A, B) = dcore (A) = dcore (B) = dAB .
374 Density-Based Clustering Methods

Figure 20.10: DBSCAN-Star dendrogram, Eps=50, Min Pts=10

In practice, one starts by applying a cut between two edges that are the furthest apart. This
either results in a single point splitting off (a Noise point), or in the single cluster to split
into two (two subtrees of the MST). Subsequent cuts are applied for smaller and smaller
distances. Decreasing the critical distance is the same as increasing the density of points,
referred to as λ in this context (see Section 20.5).
As Campello et al. (2015) show, applying cuts to the MST with decreasing distance (or
increasing density) is equivalent to applying cuts to a dendrogram obtained from single
linkage hierarchical clustering using the mutual reachability distance as the dissimilarity
matrix (hierarchical clustering methods are considered in Volume 2).
DBSCAN* derives a set of flat clusters by applying a cut to the dendrogram associated with
a given distance threshold. The main difference with the original DBSCAN is that border
points are no longer considered. In addition, the clustering process is visualized by means of
a dendrogram. However, in all other respects it is similar in spirit, and it still suffers from
the need to select a distance threshold.
DBSCAN* 375

Figure 20.11: DBSCAN-Star cluster map, Eps=50, Min Pts=10

20.4.2 Implementation
DBSCAN* is invoked in the same way as DBSCAN, but with DBScan* checked as the
Method in the DBScan Clustering Settings dialog (Figure 20.7). All other settings are
the same, except that now the Min. Cluster Size option becomes available. However, in
practice, this is typically set equal to the Min Points parameter and it is seldom used.
The functionality operates largely the same as for DBSCAN, except that the first result
for a given Min Points is a Dendrogram, shown in the right panel of the dialog. Figure
20.10 depicts the dendrogram for the same settings as in Figure 20.9, i.e., with a threshold
distance of 50 km and 10 minimum points. However, the results are very different. Only two
clusters are identified, compared to 7 clusters before. They are visualized by selecting the
Save/Show Map button at the bottom of the dialog.
The resulting cluster map is shown in Figure 20.11. One cluster is very large, containing 100
observations, the other is small, with only 12 observations. A total of 149 observations are
classified as Noise. Relative to the DBSCAN solution in Figure 20.9, the fit declined to a
ratio of 0.894.

20.4.2.1 Exploring the dendrogram

The main advantage of DBSCAN* over DBSCAN is a much more straightforward treatment
of Border points (they are all classified as Noise), and an easy visualization of the sensitivity
to the threshold distance by means of a dendrogram. Clusters can be explored for different
core distances. As long as Min Points (and the minimum cluster size) remains the same,
the dendrogram can be used to assess the cluster structure for different threshold distance
cut values. This is implemented by moving the cut-off line in the dendrogram. The distance
on the horizontal axis is the mutual reachability distance or core distance, and it replaces
the original inter-point distances.
376 Density-Based Clustering Methods

Figure 20.12: Exploring the dendrogram

For example, in the right-hand panel of Figure 20.12, the dashed red line corresponding to
the cut-off distance is moved to the left, to a distance of 73 km (also used in Section 11.4.2).
The corresponding cluster map is shown in the left-hand panel. The number of clusters has
doubled to four, ranging in size from 151 to 11 observations, with 53 unclustered Noise
points. The fit ratio decreases to 0.855.
The map may give the impression that seemingly disconnected points are contained in the
same cluster. In the example, the points within the red rectangle contain both observations
that belong to cluster 1 (dark blue) as well as unclustered points (light blue). This seeming
contradiction is due to the fact that the core distance is used in the dendrogram, and not
the actual inter-point distance.
The dendrogram needs to be reconstructed for each different value of Min Points, since
this critically affects the mutual reachability and core distances.
The main disadvantage of DBSCAN*, just as for DBSCAN, remains that the same fixed
threshold distance must be applied to all clusters, yielding a so-called flat cluster. This is
remedied by HDBSCAN.

20.5 HDBSCAN
HDBSCAN was originally proposed by Campello et al. (2013), and more recently elaborated
upon by Campello et al. (2015) and McInnes and Healy (2017).6 As in DBSCAN*, the
6 This method is variously referred to as HDBSCAN or HDBSCAN*. Here, for simplicity, the former will

be used.
HDBSCAN 377

Figure 20.13: Level set

algorithm implements the notion of mutual reachability distance. However, rather than
applying a fixed value of the cut-off distance to produce a cluster solution, a hierarchical
process is implemented that finds an optimal cluster combination, using a different critical
cut-off distance for each cluster. In order to accomplish this, the concept of cluster stability
or persistence is introduced. Optimal clusters are selected based on their relative excess of
mass value, which is a measure of their persistence.
HDBSCAN keeps the notion of Min Points from DBSCAN and also uses the concept of core
distance of an object (dcore ) from DBSCAN*. The core distance is the distance between
an object and its k-nearest neighbor, where k = Min Points − 1 (in other words, as for
DBSCAN, the object itself is included in Min Points).
Intuitively, for the same k, the core distance will be much smaller for densely distributed
points than for sparse distributions. Associated with the core distance for each object x
is the concept of density or λp , as the inverse of the core distance. As the core distance
becomes smaller, λ increases, indicating a higher density of points.
As in DBSCAN*, the mutual reachability distance is employed to construct a minimum
spanning tree, or, rather, the associated dendrogram from single linkage clustering. From
this, HDBSCAN derives an optimal solution, based on the notion of persistence or stability
of the cluster.

20.5.1 Important concepts

Several new concepts are introduced in the treatment of HDBSCAN.

20.5.1.1 Level set

A first general concept is that of a level set associated with a given density level λ. This
is introduced in the context of bump hunting, or the search for high density regions in a
general distribution f . Formally, the level set associated with density λ in the associated
distribution is:
L(λ; f ) = {x|P rob(x) > λ},
i.e., the collection of data points x for which the probability density exceeds the given
threshold. A subset of such points that contains the maximum number of points that are
connected (so-called connected components) is called a density-contour cluster of the density
function. A density-contour tree is then a tree formed of nested clusters that are obtained by
varying the level λ (Hartigan, 1975; Müller and Sawitzki, 1991; Stuetzle and Nugent, 2010).
378 Density-Based Clustering Methods

To visualize this concept, the data distribution can be represented as a two-dimensional

profile, as in Figure 20.13 (loosely based on the example in Stuetzle and Nugent, 2010). A
level set consists of those data points (i.e., areas under the curve) that are above a horizontal
line at p. This can be thought of as islands sticking out above the ocean, where the ocean
level corresponds with p.
For the lowest p-values in the figure, up to around 0.03, all the land mass is connected and it
appears as a single entity. As the p-value increases, part of the area gets submerged and two
separate islands appear, with a combined land mass (level set) that is smaller than before.
Each of these land masses constitutes a density-contour cluster. As the p-value becomes
larger than around 0.08, the island to the right disappears, and only a single island remains,
corresponding to a yet smaller level set. Finally, for p > 0.10, this splits into two separate
islands.
The level set establishes the connection between the density and the cluster size.

20.5.1.2 Persistence and excess of mass

As the value of the density λ increases, or, equivalently, as the core distance becomes smaller,
larger clusters break into sub-clusters (the analogy with rising ocean levels), or shed Noise
points. In the current context and in contrast to the previous discussion, a Noise point
dropping out is not considered a split of the cluster. The cluster remains, but only becomes
smaller.
The objective is to identify the most prominent clusters as those that continue to exist
longest. In the Figure, it corresponds with the area under the curve while a particular cluster
configuration exists. This is formalized through the concept of excess of mass (Müller and
Sawitzki, 1991), or the total density contained in a cluster after it is formed. The point of
formation is associated with a minimum density, or λmin .
In the island analogy, λmin would correspond to the ocean level where the land bridge is
first covered by water, such that the islands appear for the first time as separate entities.
The excess of mass would be the volume of the island above that ocean level, as illustrated
in Figure 20.13.
The stability or persistence of a cluster is formalized through a notion of relative excess
of mass of the cluster. For each cluster Ci , there is a minimum density, λmin (Ci ), where
it comes into being. In addition, for each point xj that belongs to the cluster, there is a
maximum density λmax (xj , Ci ) after which it no longer belongs to the cluster. For example,
in Figure 20.13, the cluster on the left side comes into being at p = 0.03. At p = 0.10, it
splits into two new clusters. So, for the large cluster, λmin , or the point where the cluster
comes into being is 0.03. The cluster stops existing (it splits) for λmax = 0.10. In turn, p =
0.10 represents λmin for the two smaller clusters. In this example, all the members of the
cluster exit at the same time, but in general, this will not be the case. Individual points
that fail to meet the critical distance as it decreases (λ increases) will fall out. Such points
will therefore have a different λmax (xj , Ci ) from those that remain in the cluster. They are
treated as Noise points, while the cluster itself continues to exist.
The persistence of a given cluster Ci is then the sum of the difference between the λmax and
λmin for each point xj that belongs to the cluster at its point of creation:

S(Ci ) = [λmax (xj , Ci ) − λmin (Ci )],
xj ∈Ci
HDBSCAN 379

Figure 20.14: Core distance and lambda for each point

where λmin (Ci ) is the minimum density level at which Ci exists, and λmax (xj , Ci ) is the
density level after which point xj no longer belongs to the cluster.
The persistence of a cluster becomes the basis to decide whether to accept a cluster split or
to keep the larger entity as the cluster.

20.5.1.3 Condensed dendrogram

A condensed dendrogram is obtained by evaluating the stability of each cluster relative to the
clusters into which it splits. If the sum of the stability of the two descendants (or children)
is larger than the stability of the parent node, then the two children are kept as clusters
and the parent is discarded. Alternatively, if the stability of the parent is larger than that
of the children, the children and all their descendants are discarded. In other words, the
large cluster is kept and all the smaller clusters that descend from it are eliminated. Here
again, a split that sheds a single point is not considered a cluster split, but only a shrinking
of the original cluster (which retains its label). Only true splits that result in sub-clusters
that contain at least two observations are considered in the simplification of the tree.
Sometimes a further constraint is imposed on the minimum size of a cluster, which would
pertain to the splits as well. As suggested in Campello et al. (2015), a simple solution is to
set the minimum cluster size equal to Min Points.
For example, in the illustration in Figure 20.13, the area in the left cluster between 0.03 and
0.10 (in blue) is larger than the sum of the areas in the smaller clusters obtained at 0.10 (in
purple). As a result, the latter would not be considered as separate clusters. The condensed
dendrogram yields a much simpler tree representation and an identification of clusters that
maximize the sum of the individual cluster stabilities.

20.5.2 HDBSCAN algorithm

The HDBSCAN algorithm has the same point of departure as DBSCAN*, i.e., a minimum
spanning tree or dendrogram derived from the mutual reachability distance. To illustrate the
logic, the same toy example is employed as in Figure 20.5. To keep matters simple, the Min
Cluster Size and Min Points are set to 2. As a result, the mutual reachability distance
is the same as the nearest neighbor distance. Since the core distance for each point is its
smallest distance to have one neighbor, the maximum of the two core distances and the
actual distance will always be the actual distance. While this avoids one of the innovations
of the algorithm, it still allows to highlight its distinctive features.
The nearest neighbor distances can be taken from the connectivity graph in Figure 20.5,
which yields the core distance for each point. The λp value associated with this core distance
is its inverse. The results for each point in the example are summarized in Figure 20.14.
380 Density-Based Clustering Methods

Figure 20.15: Minimum spanning tree connectivity graph

The inter-point distances form the basis for a minimum spanning tree (MST), which is
essential for the hierarchical nature of the algorithm. The corresponding connectivity graph
is shown in Figure 20.15, with both the node identifiers and the edge weights listed (the
distance between the two nodes the edge connects). The HDBSCAN algorithm essentially
consists of a succession of cuts in the connectivity graph. These cuts start with the highest
edge weight and move through all the edges in decreasing order of the weight. The process
of selecting cuts continues until all the points are singletons and constitute their own cluster.
Formally, this is equivalent to carrying out single linkage hierarchical clustering based on
the mutual reachability distance matrix.

20.5.2.1 Minimum spanning tree

An alternative visualization of the pruning process is in the form of a tree, shown in Figure
20.16. In the root node, all the points form a single cluster. The corresponding λ level is set
to zero. As λ increases, or, equivalently, the core distance decreases, points fall out of the
cluster, or the cluster splits into two sub-clusters. This process continues until each point is
a leaf in the tree (i.e., it is by itself).
In Figure 20.15, the longest edge is between 3 and 5, with a core distance of 29.15 for point
3. The corresponding λ is 0.034, for which point 3 is removed from the tree and becomes a
leaf. In the GeoDa implementation of HDBSCAN, singletons are not considered to be valid
clusters, so node C1 becomes the new root, with λmin reset to zero.
The next meaningful cut is for λ = 0.055, or a distance of 18.03. This creates two clusters,
one consisting of the six points 1–7, but without 3, labeled C2, and the other consisting of 8
and 9, labeled C3. The process continues for increasing values of λ, although in each case
the split involves a singleton and no further valid clusters are obtained. This leads to the
consideration of the stability of the clusters and the condensed tree.

20.5.2.2 Cluster stability and condensed tree

A closer investigation of the cuts in the tree reveals that the λp values associated with each
point (Figure 20.14) correspond to the λ value where the corresponding point falls out of a
clusters in the tree. For example, for point 7, this happens for λ = 0.055, or a core distance
of 18.03.
HDBSCAN 381

Figure 20.16: Pruning the MST

Figure 20.17: Cluster stability

The stability or persistence of each cluster is defined as the sum over all the points of the
difference between λp and λmin , where λmin is the λ level where the cluster forms.
For example, cluster C3 is formed (i.e., splits off from C1) for λ = 0.055463 and consists of
points 8 and 9. Therefore, the value of λmin for C3 is 0.055463. The value of λp for points 8
and 9 is 0.066667, corresponding with a distance of 15.0 at which they become singletons
(or leaves in the tree). The contribution of each point to the persistence of the cluster C3
382 Density-Based Clustering Methods

is thus 0.066667 – 0.055463 = 0.011204. The total value of the persistence of cluster C3 is
therefore the sum of this contribution for both points, or 0.022408. Similar calculations yield
the persistence of cluster C2 as 0.041400. This is summarized in Figure 20.17. For the sake
of completeness, the results for cluster C1 are included as well, although those are ignored
by the algorithm, since C1 is reset as root of the tree.
The condensed tree only includes C1 (as root), C2 and C3. For example, the split of C2 into
C4 and 7 is based on the following rationale. Since the persistence of a singleton is zero,
and C4 includes one less point than C2, but with the same value for λmin , the sum of the
persistence of C4 and 7 is less than the value for C2, hence the condensed tree stops there.
The condensed tree allows for different values of λ to play a role in the clustering mechanism,
potentially corresponding with a separate critical cut-off distance for each cluster.

20.5.3 Cluster membership

As outlined in the previous section, a point becomes part of a cluster for λmin , which
corresponds of a cut-off distance dmax . Once a point is part of a cluster, it remains there
until the threshold distance is small enough such that it becomes Noise, which happens for
dp,core . This may be the same or be different from the smallest distance where the cluster
still exists, or dmin . The strength of the cluster membership of point p can now be expressed
as the ratio of the final cut-off distance to the core distance for point p, dmax /dp,core , or,
equivalently, as the ratio of the density that corresponds to the point’s core distance to the
maximum density, or, λp /λmax .
The ratio gives a measure of the degree to which a point belongs to a cluster. For example,
for points that meet the dmin thresholds, i.e., observations in the highest density region, the
ratio will be 1. Noise points are assigned 0. More interestingly, for points that are part of
the cluster when it forms, but drop out during the pruning process, the ratio will have a
value smaller than 1. Therefore, the higher the ratio, the stronger the membership of the
point in the corresponding cluster.
This ratio can be exploited as the basis of a soft or fuzzy clustering approach, in which
the degree of membership in a cluster is considered. This is beyond the current scope.
Nevertheless, the ratio, or probability of cluster membership, remains a useful indicator of
the extent to which the cluster is concentrated in high density points.

20.5.4 Outliers
Related to the identification of density clusters is a concern to find points that do not belong
to a cluster and are therefore classified as outliers. Parametric approaches are typically
based on some function of the spread of the underlying distribution, such as the well-known
3-sigma rule for a normal density.7
An alternative approach is based on non-parametric principles. Specifically, the distance of
an observation to the nearest cluster can be considered as a criterion to classify outliers.
Campello et al. (2015) proposed a post-processing of the HDBSCAN results to characterize
the degree to which an observation can be considered an outlier. The method is referred to
as GLOSH, which stands for Global-Local Outlier Scores from Hierarchies.
7 These approaches are sensitive to the influence of the outliers on the estimates of central tendency and

spread. This works in two ways. On the one hand, outliers may influence the parameter estimates such that
their presence could be masked, e.g., when the estimated variance is larger than the true variance. The
reverse effect is called swamping, where observations that are legitimately part of the distribution are made
to look like outliers.
HDBSCAN 383

The logic underlying the outlier detection is closely related to the notion of cluster membership
just discussed. In fact, the probability of being an outlier is the complement of the cluster
membership.
The rationale behind the index is to compare the density threshold for which the point is
still attached to a cluster (λp in the previous discussion) and the highest density threshold
λmax for that cluster. The GLOSH index for a point p is then:
λmax − λp
GLOSHp = ,
λmax
or, equivalently, 1 − λp /λmax , the complement of the cluster membership for all but Noise
points. For the latter, since the cluster membership was set to zero, the actual λmin needs
to be used.
Given the inverse relationship between density and distance threshold, an alternative
formulation of the outlier index is as:
dmin
GLOSHp = 1 − .
dp,core

20.5.5 Implementation
HDBSCAN is invoked from the Menu or the cluster toolbar icon (Figure 20.7) as the second
item in the density clusters subgroup, Clusters > HDBSCAN. The overall setup is
very similar to that for DBSCAN, except that there is no threshold distance. Again, the
coordinates are taken as XKM and YKM.
The two main parameters are Min Cluster Size and Min Points. These are typically
set to the same value, but there may be instances where a larger value for Min Cluster
Size is desired. In the illustration, these values are set to 10. The Min Points parameter
drives the computation of the core distance for each point, which forms the basis for the
construction of the minimum spanning tree. As before, Transformation is set to Raw and
Distance Function to Euclidean.
The Run button brings up a cluster map and the associated dendrogram. The cluster map,
shown in Figure 20.18, reveals 7 clusters, ranging in size from 67 to 10, with 69 Noise points.
The Summary button reveals an overall fit of 0.970, which is the best value achieved so far.
In and of itself, the raw dendrogram is not of that much interest. One interesting feature of
the dendrogram in GeoDa is that it supports linking and brushing, which makes it possible
to explore what branches in the tree subsets of points or clusters belong to (selected in one
of the maps). The reverse is supported as well, so that branches in the tree can be selected
and their counterparts identified in other maps and graphs.
In Figure 20.19, the 19 observations are selected that form cluster 5, consisting of bank
locations on the island of Sicily. The corresponding branches in the dendrogram are high-
lighted. This includes all the splits that are possible after the formation of the final cluster,
in contrast to the condensed tree, considered next.

20.5.5.1 Exploring the condensed tree

An important and distinct feature of the HDBSCAN method is the visualization of cluster
persistence (or stability) in the condensed tree or condensed dendrogram. This is invoked by
means of the middle button in the results panel. The tree visualizes the relative excess of
mass that results in the identification of clusters for different values of λ. Starting at the
384 Density-Based Clustering Methods

Figure 20.18: HDBSCAN cluster map

Figure 20.19: HDBSCAN dendrogram

root (top), with λ = 0, the left hand axis shows how increasing values of λ (going down)
results in branching of the tree. The shaded areas are not rectangles, but they gradually
taper off, as single points are shredded from the main cluster.
The optimal clusters are identified by an oval with the same color as the cluster label in the
cluster map. The shading of the tree branches corresponds to the number of points contained
HDBSCAN 385

Figure 20.20: HDBSCAN condensed tree

Figure 20.21: HDBSCAN cluster membership

in the branch, with lighter suggesting more points. In Figure 20.20, the same 19 members of
cluster 5 are selected as before. In contrast to the dendrogram, no splits occur below the
formation of the cluster, highlighted in red.
In GeoDa, the condensed tree supports linking and brushing as well as a limited form of zoom
operations. The latter may be necessary for larger data sets, where the detail is difficult
386 Density-Based Clustering Methods

Figure 20.22: HDBSCAN core distance heat map

to distinguish in the overall tree. Zooming is implemented as a specialized mouse scrolling

operation, slightly different in each system. For example, on some track pads, this works
by moving two fingers up or down. There is no panning, but the focus of the zoom can be
moved to specific subsets of the graph.

20.5.5.2 Cluster membership

The Save button provides the option to add the core distance, cluster membership probability
and outlier index (GLOSH) to the data table. The default variable names are HDB_CORE
for the core distance associated with each point, HDB_PVAL for the degree of cluster
membership and HDB_OUT for the outliers. These variables can be included in any type
of map. For example, Figure 20.21 shows the degree of cluster membership as a natural
breaks map with 6 intervals. Clearly, the observations near the center of the cluster (smaller
core distances) carry a darker color and locations further away from the core, are lightly
colored.
A similar map can be created for the outlier probability. This will show the reverse shading,
with more remote points colored darker.

20.5.5.3 Core distance map

With the the core distance for each point saved to the data table (HDB_CORE), in GeoDa
it is possible to create a Heat Map of the core distances on top of the Cluster Map.8 This
8 Note that for this to work properly, the coordinates used for the point map and the coordinates employed

in the HDBSCAN algorithm must be the same. In the current example, this means that COORD_X and
COORD_Y should be used instead of their kilometer equivalents. The cluster solutions are the same, since
it is a simple rescaling of the coordinates.
HDBSCAN 387

is one of the Heat Map options, invoked by right clicking on the cluster point map. The
core distance variable must be specified in the Specify Core Distance option. This then
yields a heat map as in Figure 20.22, where the core distances associated with each cluster
are shown in the corresponding cluster color.
The overall pattern in the figure corresponds to gradations in the cluster membership map.
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
 Part VI

Epilogue
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
21
Postscript – The Limits of Exploration

In this first volume, devoted to exploring spatial data, I have outlined a progression of
techniques to aid in finding interesting patterns of dependence and heterogeneity in geospatial
data, supported by the GeoDa software. The highly interactive process of discovery moves
from simple maps and graphs to a more structured visualization of clusters, outliers and
indications of potential structural breaks.
The main topic in this volume is the investigation of spatial correlation, i.e., the match or
tension between attribute similarity and locational similarity. In that quest, the primary
focus is on the identification of univariate locations of clusters and spatial outliers by means
of a range of local indicators of spatial association (LISA).
For a single variable, Tobler’s first law of geography suggests the omnipresence of positive
spatial autocorrelation, with a distance decay effect for the dependence. This is indeed
observed for many variables in empirical applications. However, Tobler’s law does not
necessarily generalize to a bivariate or multivariate setting, where the situation is considerably
more complex (Anselin and Li, 2020). As discussed in more detail in Chapter 18, in a
multivariate setting, not only is there the tension between attribute and locational similarity,
but there is an additional dimension of inter-attribute similarity, which is typically not
uniform across space.
The GeoDa software is designed to make this process of discovery easy and intuitive. However,
this may also unintentionally facilitate potentially pointless clicking through the range of
interfaces until one finds an outcome one likes, without fully understanding the limitations
of the techniques involved. The extensive discussion of the methods in the book is aimed at
remedying the second aspect. Nevertheless, some caution is needed.
First and foremost, exploration is not the same as explanation. Interesting patterns may be
discovered, but that does not necessarily identify the process(es) that yielded the patterns.
As stated strongly by Heckman and Singer (2017): “data never speak for themselves.” As
argued in the discussion of global spatial autocorrelation in Section 13.5.3, spatial data are
characterized by the inverse problem, in the sense that the same pattern may be generated by
very different spatial processes. More specifically, there is an important distinction between
true contagion and apparent contagion. In cross-sectional data, both types of processes yield
a clustered pattern. However, the pattern that follows from apparent contagion is generated
by a spatially heterogeneous process and not a dependent process, as is typically assumed.
Without further information, such as a time dimension, it is impossible to distinguish between
the two generating processes.
The interpretation and validation of outcomes of an exploratory (spatial) data analysis has
been the subject of growing discussion in the literature, as reviewed in Section 4.2.1. The
exploratory process is neither inductive nor deductive, but rather abductive, involving a
move back and forth between data analysis, hypothesis generation and reformulation, as

DOI: 10.1201/9781003274919-21 391

392 Postscript – The Limits of Exploration

well as the addition of new information, an approach sometimes referred to as the “Sherlock
Holmes method” (Gahegan, 2009; Heckman and Singer, 2017).
How such a discovery process is carried out has become increasingly important with the
advent of big data, and the associated argument for data-driven discovery as the fourth
paradigm in scientific reasoning (see, e.g., Hey et al., 2009; Gahegan, 2020).
In this final chapter, I want to briefly consider three broader issues that run the risk of
getting lost in the excitement of the discovery process driven by interactive software such as
GeoDa: (1) the potential pitfalls of data science and their implications for scientific reasoning;
(2) the limitations intrinsic to spatial analysis; and (3) reproducible research.
A critical aspect of the abductive approach is how to deal with surprising results, or with
results that run counter to pre-conceived ideas. As outlined in detail in Nuzo (2015), among
others, it is easy to fool oneself by finding patterns where there are none, by focusing on
explanations that fit one’s prior convictions, or by failing to consider potential alternative
hypotheses. This can result in confirmation bias (i.e., finding what one sets out to find), as
well as disconfirmation bias (tendency to reject results that are counter to one’s priors). In
GeoDa, the extensive use of the permutation approach to represent spatial randomness in
the data is one way to partially address this concern, but by itself, it is insufficient.
Even without nefarious motivations (e.g., driven by the pressure to publish), problems
such as p-hacking (searching until significant results are found), HARKing (hypotheses
after the results are known), JARKing (justifying after the results are known), and the like
can unwittingly infiltrate an exploratory spatial data analysis. This has led to extensive
discussions in the literature as to how these issues affect the process of scientific discovery
(among others, Kerr, 1998; Simmons et al., 2011; Gelman and Loken, 2014; Gelman and
Hennig, 2017; Rubin, 2017). Sound scientific reasoning is a way of reasoning that looks for
being wrong, a systematic and enduring search for what might be wrong and an iterative
process that goes back and forth between potential explanations and evidence.1 Ideally, the
process of data exploration should be guided by these principles.
A critical notion in this regard is the researcher degrees of freedom, a reference to the many
decisions one makes with respect to the data that are included, the hypotheses considered
(and not considered), and methods selected, in the so-called garden of forking paths (Gelman
and Loken, 2014). In the context of the methods considered here and implemented in GeoDa,
aspects such as the selection of spatial scale, how to deal with outliers, whether to apply
imputation for missing values, the choice of spatial weights, the treatment of unconnected
observations (islands or isolates), and various tuning parameters are prime examples of
decisions one has to make that may affect the outcome of the data exploration. Careful
attention to these decisions, ideally accompanied by a sensitivity analysis can partially
remedy the problem.
Finally, as in any data science, spatial data science similarly may suffer from well-known
generic pitfalls, for example as outlined in the book by Smith and Cordes (2019). Their book
specifically lists “using bad data, putting data before theory, worshipping math, worshipping
computers, torturing data, fooling yourself, confusing correlation with causation, regression
toward the mean and doing harm” as examples of such pitfalls. Any serious student of
exploratory spatial data analysis should become familiar with these potential traps and learn
to recognize and avoid them.
1 For examples of how this applies to data science and spatial data science, see, e.g., https://puttingscien

ceintodatascience.org.
 393

In addition to these aspects associated with data science in general, spatial data science
also faces its own special challenges. These include the ecological fallacy (Robinson, 1950;
Goodman, 1953; 1959), the modifiable areal unit problem, or MAUP (Openshaw and Taylor,
1979), the change of support problem (Gotway and Young, 2002), as well as the more general
issue of the importance of spatial scale (Goodchild, 2011; Oshan et al., 2022). Many of
these are just special cases of well-known challenges associated with any type of statistical
analysis. For example, the ecological fallacy was first raised in sociology and it cautions
against interpreting the results of aggregate (spatial) analysis to infer individual behavior.
The change of support problem concerns the combination of data at various spatial scales,
such as point data (e.g., measurements by environmental sensors) and areal data (e.g., health
outcomes at the census tract level), and associated issues of data aggregation and imputation,
shared with the attention to scale in geographical analysis.
The MAUP stands out as particular to spatial analysis. It includes both aspects of data
aggregation as well as spatial arrangement, or, zonation. In essence, MAUP suggests that
different spatial scales and different areal boundaries will yield different and sometimes
conflicting statistical results. In many instances in the social sciences, the boundaries are
pre-set (e.g., administrative districts) and there is little one can do about it, other than being
careful in phrasing the findings of an analysis. This is particularly important when areal
boundaries do not align with the spatial scale of the processes investigated. For example,
in the U.S., census tracts are typically assumed to correspond with neighborhoods, even
though this is seldom the case. Similarly, counties are identified with labor markets, but this
is clearly invalid in metro areas (consisting of multiple counties) or in the sparsely populated
large counties in the west.
In terms of the spatial exploration, this means that the discovery of clusters needs to be
interpreted with caution. A cluster may be nothing more than an indication of poor spatial
scale (e.g., spatial units of observation much smaller than the process of interest), especially
when the size of the areal units is heterogeneous across the data set. On the other hand,
when individual spatial observations are available, the regionalization methods from Volume
2 may be applied to group them into meaningful spatial aggregates.
MAUP and its associated challenges does not invalidate spatial analysis, but it may limit the
relevance of some of the findings. Where possible, sensitivity analysis should be implemented.
A second type of limit to spatial analysis follows from the enormous advances made in
machine learning and artificial intelligence. For example, deep learning techniques (e.g.,
Goodfellow et al., 2016) are able to identify patterns without much prior (spatial) structure,
as long as they are trained on very large data sets. GeoAI pertains to the application of
these new methods to the analysis of geospatial data (Janowicz et al., 2020). Most current
applications have been in the physical domain, such as land cover classification by means of
remote sensing, landscape feature extraction and way finding. The focus in GeoAI has been
on so-called feature engineering, i.e., identifying those spatial characteristics that should
be included in the training data sets to yield good predictive performance. To date, much
of GeoAI can be viewed as applying data science to spatial data, rather than spatial data
science as conceived of in this book. Nevertheless, this raises an important question about
the relevance of spatial constructs and spatial thinking in AI. Does this imply that spatial
analysis as considered here will become irrelevant?
While it may be tempting to reach this conclusion, the powerful new deep learning methods
are not a panacea for all empirical situations. Specifically, in order to perform well, these
techniques require very large training data sets, consisting of millions (and even billions)
of data points. Such data sets are still rather rare in the contexts in empirical practice,
especially in the social sciences. In addition, the objective in exploratory spatial data analysis
394 Postscript – The Limits of Exploration

is to discover patterns and generate hypotheses, not prediction, the typical focus in AI. So,
the insights gained from the methods covered in this book will remain relevant for some time.
Nevertheless, the extent to which spatial structure and explicit spatial features contribute to
the performance of deep learning methods, or, instead becomes discovered by these methods
(and thus irrelevant) still remains to be answered.
A final concern pertaining to the exploration of spatial data is the extent to which it can
be reproducible and replicable. Reproducibility refers to obtaining the same results in a
repetition of the analysis on the same data (by the researchers themselves or by others),
replicability to obtaining the same type of results using different data.
The growing demands for transparency and openness in the larger scientific community
(so-called TOP guidelines, for transparency and openness promotion) have resulted in
requirements of open data, open science and open software code. These concerns have
also been echoed in the context of spatial data science (among others, by Rey, 2009; 2023;
Singleton et al., 2016; Brunsdon and Comber, 2021; Kedron et al., 2021). Common approaches
to accomplish this in spatial data analysis is the codification of workflows by means of visual
modeling languages, e.g., as implemented in various GIS software (see also Kruiger et al.,
2021). More formally, data, text and code are combined in notebooks, such as the well-known
Jupyter (originally for Python) and RStudio (originally for R) implementations (see Rowe
et al., 2020).
To what extent does the highly dynamic and interactive visualization in GeoDa meet these
requirements? At one level, it does not and cannot. The thrill of discovery can easily result in
a rapid succession of linked graphs and maps, in combination with various cluster analyses, in
a fairly unstructured manner that does not lend itself to codification in workflows. However,
this intrinsic lack of reproducibility can be remedied to some extent.
For example, it may be possible to embed elementary analyses (e.g., a Local Moran) into
a specific workflow (e.g., select variable, specify weights, identify clusters), although this
will necessarily result in a loss of spontaneity. Alternatively, as outlined in Appendix C,
the geodalib architecture allows specific analyses to be embedded in notebook workflows
through the new RGeoDa and PyGeoda wrappers.
In addition, there are some simple ways to obtain a degree of reproducibility for major
aspects of the analysis, such as the specification of spatial weights, custom classifications and
various variable transformations through the project file (see, e.g., Sections 4.6.3, 9.2.1.1 and
10.3.5). Nevertheless, obtaining full reproducibility for dynamic and interactive visualization
remains a challenge.
In spite of these challenges, I believe that the exploratory perspective developed in the book
and implemented in the GeoDa software can be very efficient at generating useful insights. As
long as the approach is applied with caution, such as by including ample sensitivity analyses,
pitfalls can be avoided. Time will tell.
The second volume of this introduction to spatial data science is devoted to what is referred
to in the machine learning literature as unsupervised learning. The objective is to reduce
the complexity in multivariate data by means of dimension reduction, both in the attribute
dimension (e.g., principal components and multidimensional scaling), as well as in the
observational dimension (clustering and regionalization). The distinctive perspective offered
is to emphasize the relevance of spatial aspects, by spatializing the results of classic methods,
either through mapping in combination with linking and brushing, or by imposing spatial
constraints in the clustering algorithms themselves. This builds upon the foundations
developed in the current volume.
A
Appendix A – GeoDa Preference Settings

The Preferences item on the main GeoDa menu includes several options that can be set to
fine tune the appearance and performance of the software. They appear under two different
tabs, one for System and one for Data. The contents are shown in Figures A.1 and A.2. In
a MAC OSX operating system, the preferences are accessed under the GeoDa item on the
menu bar. In Windows and Linux, the item is under File.
The System preferences pertain to detailed settings for the map and plot properties, support
for multiple languages, general appearance and computational customization.
The map and plot appearance settings control the transparency of selected items, font size,
and an option to revert to the legacy method of highlighting selected observations using
cross-hatching (this was the method used in legacy GeoDa version 0.9.5-i).
In addition to the default English, GeoDa currently has language customization in all menus
and dialogs for Simplified Chinese, Russian, Spanish, Portuguese and French, with more
languages slated to be added in the future.
Several minor options determine the appearance of a few dialogs dealing with data entry,
and other options such as automatic crash detection and software updates, as well as Retina
display support (on Mac only). Finally, some features are the default settings for the random
number seed, and the number of cores and/or GPU used in computation. Most of these are
best left to their default specification.
Under the Data tab, the most useful setting is the supported formats for dates and time (see
also Section 2.4.1.2 in Chapter 2). The other items pertain to the interaction with remote
data base servers, as well as the numerical stopping criterion used in the auto-weighting
method for spatially constrained clustering (covered in Volume 2). Here as well, the default
should be fine for most situations in practice.

DOI: 10.1201/9781003274919-A 395

396 Appendix A – GeoDa Preference Settings

Figure A.1: Preferences settings – System

Appendix A – GeoDa Preference Settings 397

Figure A.2: Preferences settings – Data

Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
B
Appendix B – Menu Structure

This Appendix lists the complete menu structure of GeoDa, with the exception of items
pertaining to spatial regression (covered in Anselin and Rey, 2014), but including the
clustering methods covered in Volume 2. The listing below shows the organization for a
MacOSX operating system. In Windows and Linux, Preferences appear under the File
menu. Otherwise, the structure is identical across operating systems.
• GeoDa
– About GeoDa
– Preferences
– Quit GeoDa
• File
– New from Recent
– New
– Save
– Save As
– Save Selected As
– Open Project
– Save Project
– Project Information
– Close
• Edit
– Time
– Category
• Tools
– Weights Manager
– Shape
∗ Points from Table
∗ Create Grid
– Spatial Join
– Dissolve
• Table
– Aggregate
– Merge
– Selection Tool
– Invert Selection
– Clear Selection
– Save Selection
– Move Selected to Top
– Calculator
– Add Variable
– Delete Variable(s)

DOI: 10.1201/9781003274919-B 399

400 Appendix B – Menu Structure

– Edit Variable Properties

– Encode
∗ (encoding options)
– Setup Number Formatting
• Map
– Themeless Map
– Quantile Map
∗ (number of quantiles)
– Percentile Map
– Box Map (Hinge=1.5)
– Box Map (Hinge=3.0)
– Standard Deviation Map
– Unique Values Map
– Co-location Map
– Natural Breaks Map
∗ (number of categories)
– Equal Intervals Map
∗ (number of categories)
– Custom Breaks
∗ Create New Custom
– Rates-Calculated Map
∗ Raw Rate
∗ Excess Risk
∗ Empirical Bayes
∗ Spatial Rate
∗ Spatial Empirical Bayes
– Conditional Map
– Cartogram
– Map Movie
• Explore
– Histogram
– Box Plot
– Scatter Plot
– Scatter Plot Matrix
– Bubble Chart
– 3D Scatter Plot
– Parallel Coordinate Plot
– Averages Chart
– Conditional Plot
∗ Map
∗ Histogram
∗ Scatter Plot
∗ Box Plot
• Clusters
– PCA
– MDS
– t-SNE
– K Means
– K Medians
– K Medoids
– Spectral
Appendix B – Menu Structure 401

– Hierarchical
– DBScan
– HDBScan
– SC K Means
– SCHC
– skater
– redcap
– AZP
– max-p
– Cluster Match Map
– Make Spatial
– Validation
• Space
– Univariate Moran’s I
– Bivariate Moran’s I
– Differential Moran’s I
– Moran’s I with EB Rate
– Univariate Local Moran’s I
– Bivariate Local Moran’s I
– Differential Local Moran’s I
– Local Moran’s I with EB Rate
– Local G
– Local G*
– Univariate Local Join Count
– Bivariate Local Join Count
– Co-location Join Count
– Univariate Local Geary
– Multivariate Local Geary
– Univariate Quantile LISA
– Multivariate Quantile LISA
– Local Neighbor Match Test
– Spatial Correlogram
– Distance Scatter Plot
• Time
– Time Player
– Time Editor
Taylor & Francis
Taylor & Francis Group
http://taylorandfrancis.com
C
Appendix C – Scripting with GeoDa via
the geodalib Library

The architecture of the GeoDa software is based on a very tight integration of the graphical user
interface and the actual computations. While effective in a traditional desktop environment,
this approach becomes less efficient when moving the functionality to a different computing
platform, such as a browser in a web-GIS, or a cyberGIS environment (Anselin et al., 2004;
Wang et al., 2013).
An alternative to the GUI-driven approach toward spatial data science taken in desktop
GeoDa is to leverage open-source software development environments, such as R (Pebesma
and Bivand, 2023) or Python (Rey et al., 2023). Such environments facilitate scripting and
stress reproducibility, which has become of increased relevance and importance in spatial
data science (see the discussion in Chapter 21).
The GeoDa desktop environment does not lend itself to scripting. It is also ill-suited for
repeated execution of the same application, such as in a simulation experiment. Finally, apart
from the limited record in the project file, there is no explicit way to ensure reproducibility.
In light of these limitations of the original design, a major refactoring effort was embarked
upon to separate the user interaction in the software from the core computational functionality,
and to collect the latter in a library, named libgeoda. The library contains the same C++
code as in the computations underlying desktop GeoDa, but has a more limited range of
functionality. The focus has been on methods that are (still) unique to GeoDa, such as some
of the recent LISA statistics. In addition, applications are included where the reliance on
C++ yields large performance improvements in terms of speed and scalability. Examples
include weights creation, and permutation tests from this Volume, as well as regionalization
methods covered in Volume 2.
The libgeoda library has a clearly defined Application Programming Interface (API),
which allows other C++ code to access its functionality directly. In fact, this is what
currently happens under the hood for part of desktop GeoDa, and in the experimental
web-GeoDa (jsgeoda, implemented through javascript). In addition to achieving a more
flexible interaction with different graphical user interface implementations, the API also
allows other software, such as R or Python programs to access the functionality through
well-defined wrapper code. The overall architecture is illustrated in Figure C.1.
The primary focus in this effort so far has been to create an R package, rgeoda, and a Python
module, pygeoda. These provide easy access to the functionality in libgeoda through a
native interface and designated middleware. The interaction between R and Python and the
C++ library is implemented under the hood, so that from a user’s perspective, everything
works natively as in any other R package or Python module.
As shown in Figure C.1, the core of the libgeoda library consists of three broad categories
of functionality: spatial weights, LISA, and spatial clustering (regionalization). In addition,

DOI: 10.1201/9781003274919-C 403

404 Appendix C – Scripting with GeoDa via the geodalib Library

Figure C.1: libgeoda architecture

there are a number of helper functions, such as support for different map classifications (to
facilitate visualization) and variable standardization (for use in the cluster routines). The
functionality in pygeoda and rgeoda is the same, with only minor differences to reflect the
particular characteristics of each software environment. For example, in Python, methods
are attributes of a class (e.g., a spatial weights class) and invoked as such. In contrast, in R,
the typical approach is to apply a function to an object to extract the relevant information
(e.g., spatial weights characteristics).
Extensive details and specific examples can be found in Anselin et al. (2022) and in the
documentation on the GitHub site. Since the functionality of libgeoda mimics what has
been covered in the book, it is not further considered here. The software development effort
is ongoing.
Bibliography

Akbari, K., Winter, S., and Tomko, M. (2023). Spatial causality: A systematic review on
spatial causal inference. Geographical Analysis, 55: 56–89.
Algeri, C., Anselin, L., Forgione, A. F., and Migliardo, C. (2022). Spatial dependence in the
technical efficiency of local banks. Papers in Regional Science, 101:385–416.
Amaral, P., de Carvalho, L. R., Rocha, T. A. H., da Silva, N. C., and Vissoci, J. R. N. (2019).
Geospatial modeling of microcephaly and zika virus spread patterns in Brazil. PLoS ONE,
14.
Andrienko, G., Andrienko, N., Keim, D., MacEachren, A., and Wrobel, S. (2011). Challenging
problems of geospatial visual analytics. Journal of Visual Languages and Computing,
22:251–256.
Andrienko, N., Lammarsch, T., Andrienko, G., Fuchs, G., Keim, D., Miksch, S., and Rind, A.
(2018). Viewing visual analytics as model building. Computer Graphics Forum, 37:275–299.
Angrist, J. and Pischke, J.-S. (2015). Mastering ’Metrics, The Path from Cause to Effect.
Princeton University Press, Princeton, New Jersey.
Anselin, L. (1988). Spatial Econometrics: Methods and Models. Kluwer Academic Publishers,
Dordrecht, The Netherlands.
Anselin, L. (1990). What is special about spatial data? Alternative perspectives on spatial
data analysis. In Griffith, D. A., editor, Spatial Statistics, Past, Present and Future, pages
66–77. Institute of Mathematical Geography, (IMAGE), Ann Arbor, MI.
Anselin, L. (1992). SpaceStat, a Software Program for Analysis of Spatial Data. National
Center for Geographic Information and Analysis (NCGIA), University of California, Santa
Barbara, CA.
Anselin, L. (1994). Exploratory spatial data analysis and geographic information systems. In
Painho, M., editor, New Tools for Spatial Analysis, pages 45–54. Eurostat, Luxembourg.
Anselin, L. (1995). Local indicators of spatial association — LISA. Geographical Analysis,
27:93–115.
Anselin, L. (1996). The Moran scatterplot as an ESDA tool to assess local instability in
spatial association. In Fischer, M., Scholten, H., and Unwin, D., editors, Spatial Analytical
Perspectives on GIS in Environmental and Socio-Economic Sciences, pages 111–125. Taylor
and Francis, London.
Anselin, L. (1998). Exploratory spatial data analysis in a geocomputational environment. In
Longley, P. A., Brooks, S., Macmillan, B., and McDonnell, R., editors, Geocomputation: A
Primer, pages 77–94. John Wiley, New York, NY.
Anselin, L. (1999). Interactive techniques and exploratory spatial data analysis. In Lon-
gley, P. A., Goodchild, M. F., Maguire, D. J., and Rhind, D. W., editors, Geographical

405
406 Bibliography

Information Systems: Principles, Techniques, Management and Applications, pages 251–

264. John Wiley, New York, NY.
Anselin, L. (2005a). Exploring Spatial Data with GeoDa: A Workbook. Spatial Analysis
Laboratory (SAL). Department of Geography, University of Illinois, Urbana-Champaign,
IL.
Anselin, L. (2005b). Spatial statistical modeling in a GIS environment. In Maguire, D. J.,
Batty, M., and Goodchild, M. F., editors, GIS, Spatial Analysis and Modeling, pages
93–111. ESRI Press, Redlands, CA.
Anselin, L. (2012). From spacestat to cybergis, twenty years of spatial data analysis software.
International Regional Science Review, 35:131–157.
Anselin, L. (2019a). A local indicator of multivariate spatial association, extending Geary’s
c. Geographical Analysis, 51:133–150. doi:10.1111/gean.12164.
Anselin, L. (2019b). Quantile local spatial autocorrelation. Letters in Spatial and Resource
Sciences, 12:155–166.
Anselin, L. (2020). Spatial data science. In Richardson, D., editor, The International
Encyclopedia of Geography: People, the Earth, Environment, and Technology. Wiley, New
York, NY.
Anselin, L. (2021). Spatial models in econometric research. In Hamilton, J., editor, Oxford
Research Encyclopedia of Economics and Finance. Oxford University Press.
Anselin, L. and Arribas-Bel, D. (2013). Spatial fixed effects and spatial dependence in a
single cross-section. Papers in Regional Science, 92:3–17.
Anselin, L., Kim, Y.-W., and Syabri, I. (2004). Web-based analytical tools for the exploration
of spatial data. Journal of Geographical Systems, 6:197–218.
Anselin, L. and Li, X. (2019). Operational local join count statistics for cluster detection.
Journal of Geographical Systems, 21:189–210.
Anselin, L. and Li, X. (2020). Tobler’s law in a multivariate world. Geographical Analysis,
52:494–510.
Anselin, L., Li, X., and Koschinsky, J. (2022). GeoDa, from the desktop to an ecosystem for
exploring spatial data. Geographical Analysis, 54:439–466.
Anselin, L., Lozano-Gracia, N., and Koschinky, J. (2006a). Rate transformations and
smoothing. Technical report, Spatial Analysis Laboratory, Department of Geography,
University of Illinois, Urbana, IL.
Anselin, L. and Rey, S. J. (2014). Modern Spatial Econometrics in Practice, A Guide to
GeoDa, GeoDaSpace and PySAL. GeoDa Press, Chicago, IL.
Anselin, L. and Smirnov, O. (1996). Efficient algorithms for constructing proper higher
order spatial lag operators. Journal of Regional Science, 36:67–89.
Anselin, L., Syabri, I., and Kho, Y. (2006b). GeoDa, an introduction to spatial data analysis.
Geographical Analysis, 38:5–22.
Anselin, L., Syabri, I., and Smirnov, O. (2002). Visualizing multivariate spatial correlation
with dynamically linked windows. In Anselin, L. and Rey, S., editors, New Tools for Spatial
Data Analysis: Proceedings of the Specialist Meeting. Center for Spatially Integrated Social
Science (CSISS), University of California, Santa Barbara. CD-ROM.
Bibliography 407

Assuncao, R. M. and Reis, E. A. (1999). A new proposal to adjust Moran’s I for population
density. Statistics in Medicine, 18:2147–2161.
Banerjee, S., Carlin, B. P., and Gelfand, A. E. (2015). Hierarchical Modeling and Analysis
for Spatial Data, 2nd Edition. Chapman & Hall/CRC, Boca Raton.
Bavaud, F. (1998). Models for spatial weights: A systematic look. Geographical Analysis,
30:153–171.
Becker, R. A. and Cleveland, W. (1987). Brushing scatterplots. Technometrics, 29:127–142.
Becker, R. A., Cleveland, W., and Shyu, M.-J. (1996). The visual design and control of
Trellis displays. Journal of Computational and Graphical Statistics, 5:123–155.
Becker, R. A., Cleveland, W., and Wilks, A. (1987). Dynamic graphics for data analysis.
Statistical Science, 2:355–395.
Bellman, R. E. (1961). Adaptive Control Processes. Princeton University Press, Princeton,
N.J.
Benjamin, D. and 72 others (2018). Redefine statistical significance. Nature Human Behavior,
2:6–10.
Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: A practical
and powerful approach to multiple testing. Journal of the Royal Statistical Society B,
57:289–300.
Bjornstad, O. N. and Falck, W. (2001). Nonparametric spatial covariance functions: Estima-
tion and testing. Environmental and Ecological Statistics, 8:53–70.
Blommestein, H. J. (1985). Elimination of circular routes in spatial dynamic regression
equations. Regional Science and Urban Economics, 15:121–130.
Brewer, C. A. (1997). Spectral schemes: controversial color use on maps. Cartography and
Geographic Information Systems, 49:280–294.
Brewer, C. A. (2016). Designing Better Maps. A Guide for GIS Users, 2nd Edition. ESRI
Press, Redlands, CA.
Brewer, C. A., Hatchard, G., and Harrower, M. A. (2003). ColorBrewer in print: a catalog
of color schemes for maps. Cartography and Geographic Information Science, 30:5–32.
Brock, W. A. and Durlauf, S. N. (2001). Discrete choice with social interactions. Review of
Economic Studies, 59:235–260.
Brunsdon, C. and Comber, L. (2015). Geocomputation, a Practical Primer. Sage, Thousand
Oaks, CA.
Brunsdon, C. and Comber, L. (2019). An Introduction to R for Spatial Analysis and Mapping,
2nd Edition. Sage, Thousand Oaks, CA.
Brunsdon, C. and Comber, L. (2021). Opening practice: supporting reproducibility and
critical spatial data science. Journal of Geographical Systems, 23:477–496.
Buja, A., Cook, D., Hofmann, H., Lawrence, M., Lee, E.-K., Swayne, D. F., and Wickham,
H. (2009). Statistical inference for exploratory data analysis and model diagnostics.
Philosophical Transactions of the Royal Society A, 367:4361–4383.
Buja, A., Cook, D., and Swayne, D. (1996). Interactive high dimensional data visualization.
Journal of Computational and Graphical Statistics, 5:78–99.
408 Bibliography

Campello, R. J., Moulavi, D., and Sander, J. (2013). Density-based clustering based on
hierarchical density estimates. In Pei, J., Tseng, V. S., Cao, L., Motoda, H., and Xu, G.,
editors, Advances in Knowledge Discovery and Data Mining. PAKDD 2013. Lecture Notes
in Computer Science, Vol. 7819, pages 160–172.
Campello, R. J., Moulavi, D., Zimek, A., and Sandler, J. (2015). Hierarchical density
estimates for data clustering, visualization, and outlier detection. ACM Transactions on
Knowledge Discovery from Data, 10,1.
Carr, D. B. and Pickle, L. W. (2010). Visualizing Data Patterns with Micromaps. Chapman
& Hall/CRC, Boca Raton, FL.
Case, A. (1991). Spatial patterns in household demand. Econometrica, 59:953–965.
Case, A., Rosen, H. S., and Hines, J. R. (1993). Budget spillovers and fiscal policy interde-
pendence: Evidence from the states. Journal of Public Economics, 52:285–307.
Chen, C., Härdle, W., and Unwin, A. (2008). Handbook of Data Visualization. Springer,
Berlin, Germany.
Chilès, J.-P. and Delfiner, P. (1999). Geostatistics, Modeling Spatial Uncertainty. John Wiley
& Sons, New York, NY.
Chow, G. (1960). Tests of equality between sets of coefficients in two linear regressions.
Econometrica, 28:591–605.
Clayton, D. and Kaldor, J. (1987). Empirical Bayes estimates of age-standardized relative
risks for use in disease mapping. Biometrics, 43:671–681.
Cleveland, W. S. (1979). Robust locally weighted regression and smoothing scatterplots.
Journal of the American Statistical Association, 74:829–836.
Cleveland, W. S. (1993). Visualizing Data. Hobart Press, Summit, NJ.
Cleveland, W. S., Grosse, E., and Shyu, W. M. (1992). Local regression models. In Chambers,
J. M. and Hastie, T. J., editors, Statistical Models in S, pages 309–376. Wadsworth and
Brooks/Cole, Pacific Grove, CA.
Cleveland, W. S. and McGill, M. (1988). Dynamic Graphics for Statistics. Wadsworth,
Pacific Grove, CA.
Cliff, A. and Ord, J. K. (1973). Spatial Autocorrelation. Pion, London.
Cliff, A. and Ord, J. K. (1981). Spatial Processes: Models and Applications. Pion, London.
Comber, L. and Brunsdon, C. (2021). Geographical Data Science and Spatial Analysis. An
Introduction in R. Sage, Thousand Oaks, CA.
Cressie, N. (1993). Statistics for Spatial Data. Wiley, New York.
Dasu, T. and Johnson, T. (2003). Exploratory Data Mining and Data Cleaning. John Wiley,
Hoboken, NJ.
de Berg, M., Cheong, O., van Kreveld, M., and Overmars, M. (2008). Computational
Geometry, Algorithms and Applications, 3rd Edition. Springer Verlag, Berlin.
de Castro, M. C. and Singer, B. H. (2006). Controlling the false discovery rate: An
application to account for multiple and dependent tests in local statistics of spatial
association. Geographical Analysis, 38:180–208.
Bibliography 409

Deutsch, C. and Journel, A. (1998). GSLIB: Geostatistical Software Library and User’s
Guide. Oxford University Press, New York, NY.
Dorling, D. (1996). Area Cartograms: Their Use and Creation. CATMOG 59, Institute of
British Geographers.
Dray, S., Saïd, S., and Débias, F. (2008). Spatial ordination of vegetation data using a
generalization of Wartenberg’s multivariate spatial correlation. Journal of Vegetation
Science, 19:45–56.
Dykes, J. (1997). Exploring spatial data representation with dynamic graphics. Computers
and Geosciences, 23:345–370.
Dykes, J., MacEachren, A. M., and Kraak, M.-J. (2005). Exploring Geovisualization. Elsevier,
Oxford, United Kingdom.
Efron, B. and Hastie, T. (2016). Computer Age Statistical Inference. Algorithms, Evidence,
and Data Science. Cambridge University Press, Cambridge, UK.
Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. (1996). A density-based algorithm for
discovering clusters in large spatial databases with noise. In KDD-96 Proceedings, pages
226–231.
Farah Rivadeneyra, I. (2017). A geospatial analysis of Mexico’s distal effects on food
insecurity. Master’s thesis, University of Chicago, Chicago, IL.
Fisher, W. D. (1958). On grouping for maximum homogeneity. Journal of the American
Statistical Association, 53:789–798.
Fotheringham, A. S., Brunsdon, C., and Charlton, M. (2002). Geographically Weighted
Regression. John Wiley, Chichester.
Friendly, M. (2008). A brief history of data visualization. In Chun-houh Chen, C., Härdle,
W., and Unwin, A., editors, Handbook of Data Visualization, pages 15–56. Springer, Berlin,
Germany.
Gahegan, M. (2009). Visual exploration and explanation in geography: analysis with lights.
In Miller, H. and Han, J., editors, Geographic Data Mining and Knowledge Discovery,
pages 291–315. Taylor and Francis, Boca Raton, FL.
Gahegan, M. (2020). Fourth paradigm GIScience? prospects for automated discovery and
explanation from data. International Journal of Geographical Information Science, 34:1–21.
Gan, J. and Tao, Y. (2017). On the hardness and approximation of Euclidean DBSCAN.
ACM Transactions on Database Systems (TODS), 42:14.
Gao, S. (2021). Geospatial artificial intelligence (GeoAI). Oxford Bibliographies, Oxford
University Press.
Geary, R. (1954). The contiguity ratio and statistical mapping. The Incorporated Statistician,
5:115–145.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. (2014).
Bayesian Data Analysis, 3rd Edition. Chapman & Hall, Boca Raton, FL.
Gelman, A. and Hennig, C. (2017). Beyond subjective and objective in statistics. Journal of
the Royal Statistical Society A, 180:967–1033.
Gelman, A. and Loken, E. (2014). The statistical crisis in science. American Scientist,
102:460–465.
410 Bibliography

Getis, A. (1984). Interaction modeling using second-order analysis. Environment and

Planning A, 16:173–183.
Getis, A. (2009). Spatial weights matrices. Geographical Analysis, 41:404–410.
Getis, A. and Franklin, J. (1987). Second-order neighborhood analysis of mapped point
patterns. Ecology, 68:473–477.
Getis, A. and Ord, J. K. (1992). The analysis of spatial association by use of distance
statistics. Geographical Analysis, 24:189–206.
Good, I. (1983). The philosophy of exploratory data analysis. Philosophy of Science,
50:283–295.
Goodchild, M. F. (2011). Scale in GIS: an overview. Geomorphology, 130:5–9.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press, Cambridge,
MA.
Goodman, L. A. (1953). Ecological regression and behavior of individuals. American
Sociological Review, 18:663–664.
Goodman, L. A. (1959). Some alternatives to ecological correlation. American Journal of
Sociology, 64:610–625.
Gotway, C. A. and Young, L. J. (2002). Combining incompatible spatial data. Journal of
the American Statistical Association, 97:632–648.
Greenbaum, R. T. (2002). A spatial study of teacher’s salaries in Pennsylvania school
districts. Journal of Labor Research, 23:69–86.
Hall, P. and Patil, P. (1994). Properties of nonparametric estimators of autocovariance for
stationary random fields. Probability Theory and Related Fields, 99:399–424.
Harris, C. (1954). The market as a factor in the localization of industry in the United States.
Annals of the Association of American Geographers, 64:315–348.
Harris, R., Moffat, J., and Kravtsova, V. (2011). In search of ‘W’. Spatial Economic Analysis,
6:249–270.
Harrower, M. A. and Brewer, C. A. (2003). ColorBrewer.org: an online tool for selecting
color schemes for maps. The Cartographic Journal, 40:27–37.
Hartigan, J. A. (1975). Clustering Algorithms. John Wiley, New York, NY.
Haslett, J., Bradley, R., Craig, P., Unwin, A., and Wills, G. (1991). Dynamic graphics
for exploring spatial data with applications to locating global and local anomalies. The
American Statistician, 45:234–242.
Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical Learning,
2nd Edition. Springer, New York, NY.
Heckman, J. and Singer, B. (2017). Abducting economics. American Economic Review,
Papers and Proceedings, 107:298–302.
Hey, T., Tansley, S., and Tolle, K. (2009). The Fourth Paradigm: Data-Intensive Scientific
Discovery. Microsoft Research, Redmond, WA.
Hubert, L. J., Golledge, R., and Costanzo, C. M. (1981). Generalized procedures for
evaluating spatial autocorrelation. Geographical Analysis, 13:224–233.
Bibliography 411

Hubert, L. J., Golledge, R., Costanzo, C. M., and Gale, N. (1985). Measuring association
between spatially defined variables: An alternative procedure. Geographical Analysis,
17:36–46.
Hullman, J. and Gelman, A. (2021). Designing for interactive exploratory data analysis
requires theories of graphical inference. Harvard Data Science Review, 3.
Inselberg, A. (1985). The plane with parallel coordinates. Visual Computer, 1:69–91.
Inselberg, A. and Dimsdale, B. (1990). Parallel coordinates: A tool for visualizing multi-
dimensional geometry. Proceedings of the IEEE Visualization 90, pages 361–378.
Isaaks, E. H. and Srivastava, R. M. (1989). An Introduction to Applied Geostatistics. Oxford
University Press, New York, NY.
Isard, W. (1960). Methods of Regional Analysis. MIT Press, Cambridge, MA.
Isard, W. (1969). General Theory. MIT Press, Cambridge, MA.
James, W. and Stein, C. (1961). Estimation with quadratic loss. Proceedings of the Fourth
Berkeley Symposium on Mathematical Statistics and Probability, 1:361–379.
Janowicz, K., Gao, S., McKenzie, G., Hu, Y., and Bhaduri, B. (2020). GeoAI: spatially
explicit artificial intelligence techniques for geographic knowledge discovery and beyond.
International Journal of Geographical Information Science, 34:625–636.
Jenks, G. (1977). Optimal data classification for choropleth maps. Occasional. paper no. 2,
Department of Geography, University of Kansas, Lawrence, KS.
Journel, A. and Huijbregts, C. (1978). Mining Geostatistics. Academic Press, London.
Kafadar, K. (1996). Smoothing geographical data, particularly rates of disease. Statistics in
Medicine, 15:2539–2560.
Kafadar, K. (1997). Geographic trends in prostate cancer mortality: An application of
spatial smoothers and the need for adjustment. Annals of Epidemiology, 7:35–45.
Kaufman, L. and Rousseeuw, P. (2005). Finding Groups in Data: An Introduction to Cluster
Analysis. John Wiley, New York, NY.
Kedron, P., Frazier, A. A., Trgovac, A. B., Nelson, T., and Fotheringham, A. S. (2021).
Reproducibility and replicability in geographical analysis. Geographical Analysis, 53:135–
147.
Kelejian, H. H. and Prucha, I. R. (2007). HAC estimation in a spatial framework. Journal
of Econometrics, 140:131–154.
Kerr, N. (1998). HARKing: hypothesizing after the results are known. Personality and
Social Psychology Review, 2:196–217.
Kessler, F. C. and Battersby, S. E. (2019). Working with Map Projections: A Guide to Their
Selection. CRC Press, Boca Raton, FL.
Kielman, J., Thomas, J., and May, R. (2009). Foundations and frontiers in visual analytics.
Information Visualization, 8:239–246.
Kolak, M. and Anselin, L. (2020). A spatial perspective on the econometrics of program
evaluation. International Regional Science Review, 43:128–153.
412 Bibliography

Kolak, M., Bhatt, J., Park, Y. H., Padrón, N. A., and Molefe, A. (2020). Quantification of
neighborhood-level social determinants of health in the continental united states. JAMA
Network Open, 3(1):e1919928–e1919928.
Kraak, M. and MacEachren, A. (2005). Geovisualization and GIScience. Cartography and
Geographic Information Science, 32:67–68.
Kraak, M. and Ormeling, F. (2020). Cartography. Visualization of GeoSpatial Data, 4th
Edition. CRC Press, Boca Raton, FL.
Kruiger, J., Kasalica, V., Meerlo, R., Lamprecht, A.-L., Nyamsuren, E., and Schneider, S.
(2021). Loose programming of GIS workflows with geo-analytical concepts. Transactions
in GIS, 25:424–449.
Kulldorff, M. (1997). A spatial scan statistic. Communications in Statistics – Theory and
Methods, 26:1481–1496.
Lancaster, T. (2000). The incidental parameter problem since 1948. Journal of Econometrics,
95:391–413.
Law, M. and Collins, A. (2018). Getting to Know ArcGIS Desktop, 5th Edition. Environmental
Systems Research Institute, Redlands, CA.
Lawson, A. B., Browne, W. J., and Rodeiro, C. L. V. (2003). Disease Mapping with
WinBUGS and MLwiN. John Wiley, Chichester.
Lee, J. A. and Verleysen, M. (2007). Nonlinear Dimensionality Reduction. Springer-Verlag,
New York, NY.
Lee, S.-I. (2001). Developing a bivariate spatial association measure: An integration of
Pearson’s r and Moran’s I. Journal of Geographical Systems, 3:369–385.
Lin, J. (2020). A local model for multivariate analysis: extending Wartenberg’s multivariate
spatial correlation. Geographical Analysis, 52:190–210.
Loader, C. (1999). Local Regression and Likelihood. Springer-Verlag, Heidelberg.
Loader, C. (2004). Smoothing: Local regression techniques. In Gentle, J. E., Härdle, W., and
Mori, Y., editors, Handbook of Computational Statistics: Concepts and Methods, pages
539–563. Springer-Verlag, Berlin.
Lovelace, R., Nowosad, J., and Muenchow, J. (2019). Geocomputation with R. CRC Press,
Boca Raton, FL.
MacEachren, A. M. and Kraak, M.-J. (1997). Exploratory cartographic visualization:
Advancing the agenda. Computers and Geosciences, 23:335–343.
MacEachren, A. M., Wachowicz, M., Edsall, R., Haug, D., and Masters, R. (1999). Con-
structing knowledge from multivariate spatiotemporal data: Integrating geographical
visualization with knowledge discovery in database methods. International Journal of
Geographical Information Science, 13(4):311–334.
Madry, S. (2021). Introduction to QGIS: Open Source Geographic Information System.
Locate Press, Chugiak, AK.
Mantel, N. (1967). The detection of disease clustering and a generalized regression approach.
Cancer Research, 27:209–220.
Marshall, R. J. (1991). Mapping disease and mortality rates using Empirical Bayes estimators.
Applied Statistics, 40:283–294.
Bibliography 413

McCann, P. (2001). Urban and Regional Economics. Oxford University Press, New York,
NY.
McInnes, L. and Healy, J. (2017). Accelerated hierarchical density clustering. In 2017 IEEE
International Conference on Data Mining Workshops (ICDMW), New Orleans, LA.
Monmonier, M. (1989). Geographic brushing: Enhancing exploratory analysis of the scatter-
plot matrix. Geographical Analysis, 21:81–4.
Monmonier, M. (1993). Mapping it Out: Expository Cartography for the Humanities and
Social Sciences. University of Chicago Press, Chicago, IL.
Monmonier, M. (2018). How to Lie with Maps, 3rd Edition. University of Chicago Press,
Chicago, IL.
Moran, P. A. (1948). The interpretation of statistical maps. Journal of the Royal Statistical
Society, B, 10:243–251.
Müller, D. and Sawitzki, G. (1991). Excess mass estimates and tests for multimodality.
Journal of the American Statistical Association, 86:738–746.
Munasinghe, R. L. and Morris, R. D. (1996). Localization of disease clusters using regional
measures of spatial autocorrelation. Statistics in Medicine, 15:893–905.
Newman, M. (2018). Networks. Oxford University Press, Oxford, United Kingdom.
Nuzo, R. (2015). Fooling ourselves. Nature, 526:182–185.
Oden, N. L. and Sokal, R. R. (1986). Directional autocorrelation: an extension of spatial
correlograms to two dimensions. Systematic Zoology, 35:608–617.
Okabe, A., Boots, B., Sugihara, K., and Chiu, S. N. (2000). Spatial Tessellations: Concepts
and Applications of Voronoi Diagrams, 2nd Edition. Wiley, New York, NY.
Openshaw, S., Charlton, M. E., Wymer, C., and Craft, A. (1987). A mark I geographical
analysis machine for the automated analysis of point data sets. International Journal of
Geographical Information Systems, 1:359–377.
Openshaw, S. and Taylor, P. (1979). A million or so correlation coefficients. In Wrigley, N.,
editor, Statistical Methods in the Spatial Sciences, pages 127–144. Pion, London.
Ord, J. K. and Getis, A. (1995). Local spatial autocorrelation statistics: Distributional issues
and an application. Geographical Analysis, 27:286–306.
Ord, J. K. and Getis, A. (2001). Testing for local spatial autocorrelation in the presence of
global autocorrelation. Journal of Regional Science, 41:411–432.
Oshan, T. M., Wolf, L., Sachdeva, M., Bardin, S., and Fotheringham, A. S. (2022). A scoping
review of the multiplicity of scale in spatial analysis. Journal of Geographical Systems,
24:293–324.
Pebesma, E. and Bivand, R. (2023). Spatial Data Science: With Applications in R. Chapman
& Hall/CRC, Boca Raton, FL.
Preston, S. H., Heuveline, P., and Guillot, M. (2001). Demography, Measuring and Modeling
Population Processes. Blackwell Publishers Ltd., Oxford, UK.
Prim, R. (1957). Shortest connection networks and some generalizations. Bell System
Technical Journal, 36:1389–1401.
414 Bibliography

Racine, J. (2019). An Introduction to the Advanced Theory of Nonparametric Econometrics.

A Replicable Approach Using R. Cambridge University Press, Cambridge, United Kingdom.
Rattenbury, T., Hellerstein, J. M., Heer, J., Kandel, S., and Carreras, C. (2017). Principles
of Data Wrangling. Practical Techniques for Data Preparation. O’Reilly, Sebastopol, CA.
Reich, B. J., Yang, S., Guan, Y., Giffin, A. B., Miller, M. J., and Rappold, A. (2021). A review
of spatial causal inference methods for environmental and epidemiological applications.
International Statistical Review, 89:605–634.
Rey, S. J. (2009). Show me the code: Spatial analysis and open source. Journal of Geographical
Systems, 11:191–207.
Rey, S. J. (2023). Big code. Geographical Analysis, 55:211–224.
Rey, S. J., Anselin, L., Pahle, R., Kang, X., and Stephens, P. (2013). Parallel optimal
choropleth map classification in PySAL. International Journal of Geographical Information
Science, 27:1023–1039.
Rey, S. J., Arribas-Bel, D., and Wolf, L. J. (2023). Geographic Data Science with Python.
CRC Press, Boca Raton, FL.
Rey, S. J., Stephens, P., and Laura, J. (2017). An evaluation of sampling and full enumeration
strategies for Fischer Jenks classification in big data settings. Transactions in GIS, 21:796–
810.
Rhyne, T., MacEachren, A., and Dykes, J. (2006). Exploring geovisualization. IEEE
Computer Graphics and Applications, 26:20–21.
Robinson, W. (1950). Ecological correlations and the behavior of individuals. American
Sociological Review, 15:351–357.
Rowe, F., Maier, G., Arribas-Bel, D., and Rey, S. (2020). The potential of notebooks for
scientific publication, reproducibility and dissemination. REGION, 7:E1–E5.
Rubin, M. (2017). When does HARKing hurt? identifying when different types of undisclosed
post hoc hypothesizing harm scientific progress. Review of General Psychology, 21:308–320.
Sander, J., Ester, M., Kriegel, H.-P., and Xu, X. (1998). Density-based clustering in spatial
databases: The algorithm GDBSCAN and its applications. Data Mining and Knowledge
Discovery, 2:169–194.
Schubert, E., Sander, J., Ester, M., Kriegel, P., and Xu, X. (2017). DBSCAN revisited,
revisited: why and how you should (still) use DBSCAN. ACM Transactions on Database
Systems (TODS), 42:19.
Shellito, B. A. (2020). Introduction to Geospatial Technologies, 5th Edition. Macmillan
Learning, New York, NY.
Silverman, B. (1986). Density Estimation for Statistics and Data Analysis. Chapman &
Hall, New York, NY.
Simmons, J., Nelson, L., and Simonsohn, U. (2011). False-positive psychology: undis-
closed flexibility in data collection and analysis allows presenting anything as significant.
Psychological Science, 22:1359–1366.
Singleton, A. and Arribas-Bel, D. (2021). Geographic data science. Geographical Analysis,
53:61–75.
Bibliography 415

Singleton, A., Spielman, S., and Brunsdon, C. (2016). Establishing a framework for open
geographic information science. International Journal of Geographical Information Science,
30:1507–1521.
Slocum, T. A., McMaster, R. B., Kessler, F. C., and Howard, H. H. (2023). Thematic
Cartography and Geovisualization, 4th Edition. CRC Press, Boca Raton, FL.
Smith, G. and Cordes, J. (2019). The 9 Pitfalls of Data Science. Oxford University Press,
Oxford, UK.
Snyder, J. P. (1993). Flattening the Earth: Two Thousand Years of Map Projections.
University of Chicago Press, Chicago, IL.
Sokal, R. R. (1979). Testing statistical significance of geographic variation patterns. System-
atic Zoology, 28:227–232.
Sokal, R. R., Oden, N. L., and Thompson, B. A. (1998a). Local spatial autocorrelation in a
biological model. Geographical Analysis, 30:331–354.
Sokal, R. R., Oden, N. L., and Thompson, B. A. (1998b). Local spatial autocorrelation in
biological variables. Biological Journal of the Linnean Society, 65:41–62.
Stein, M. L. (1999). Interpolation of Spatal Data, Some Theory for Kriging. Springer-Verlag,
New York.
Stewart, J. Q. (1947). Empirical mathematical rules concerning the distribution and
equilibrium of population. Geographical Review, 37:461–485.
Stuetzle, W. (1987). Plot windows. Journal of the American Statistical Association, 82:466–
475.
Stuetzle, W. and Nugent, R. (2010). A generalized single linkage method for estimating the
cluster tree of a density. Journal of Computational and Graphical Statistics, 19:397–418.
Thomas, J. and Cook, K. (2005). Illuminating the Path: The Research and Development
Agenda for Visual Analytics. IEEE Computer Society Press, Los Alamitos, CA.
Tiefelsdorf, M. (2002). The saddlepoint approximation of Moran’s I and local Moran’s Ii ’s
reference distribution and their numerical evaluation. Geographical Analysis, 34:187–206.
Tiefelsdorf, M. and Boots, B. (1995). The exact distribution of Moran’s I. Environment and
Planning A, 27:985–999.
Tobler, W. (1970). A computer movie simulating urban growth in the Detroit region.
Economic Geography, 46:234–240.
Tobler, W. (1973). Choropleth maps without class intervals? Geographical Analysis, 5:262–
265.
Tobler, W. (2004). Thirty five years of computer cartograms. Annals of the Association of
American Geographers, 94:58–73.
Tomlin, C. D. (1990). Geographic Information Systems and Cartographic Modeling. Prentice-
Hall, Englewood Cliffs, NJ.
Tufte, E. R. (1983). The Visual Display of Quantitative Information. Graphics Press,
Cheshire, CT.
Tufte, E. R. (1997). Visual Explanations: Images and Quantities, Evidence and Narrative.
Graphics Press, Cheshire, CT.
416 Bibliography

Tukey, J. (1962). The future of data analysis. Annals of Mathematical Statistics, 33:1–67.
Tukey, J. (1977). Exploratory Data Analysis. Addison Wesley, Reading, MA.
Tukey, J. and Wilk, M. B. (1966). Data analysis and statistics: an expository overview. In
Proceedings of the November 7-10, 1966, Fall Joint Computer Conference, pages 695–709.
ACM.
Wall, P. and Devine, O. (2000). Interactive analysis of the spatial distribution of disease
using a geographic information system. Journal of Geographical Systems, 2(3):243–256.
Waller, L. A. and Gotway, C. A. (2004). Applied Spatial Statistics for Public Health Data.
John Wiley, Hoboken, NJ.
Wang, S. (2010). A cyberGIS framework for the synthesis of cyberinfrastructure, GIS, and
spatial analysis. Annals of the Association of American Geographers, 100:535–557.
Wang, S., Anselin, L., Badhuri, B., Crosby, C., Goodchild, M., Liu, Y., and Nyerges, T.
(2013). CyberGIS software: a synthetic review and integration roadmap. International
Journal of Geographical Information Science, 27:2122–2145.
Wartenberg, D. (1985). Multivariate spatial correlation: A method for exploratory geograph-
ical analysis. Geographical Analysis, 17:263–283.
Wegman, E. J. (1990). Hyperdimensional data analysis using parallel coordinates. Journal
of the American Statistical Association, 85:664–675.
Wegman, E. J. and Dorfman, A. (2003). Visualizing cereal world. Computational Statistics
and Data Analysis, 43(4):633–649.
Wickham, H. (2016). ggplot2 Elegant Graphics for Data Analysis. Springer Nature, London,
UK.
Wickham, H., Cook, D., Hofmann, H., and Buja, A. (2010). Graphical inference for Infovis.
IEEE Transactions on Visualization and Computer Graphics, 16:973–979.
Wishart, D. (1969). Mode analysis: a generalization of nearest neighbor which reduces
chaining effects. In Cole, A., editor, Numerical Taxonomy, pages 282–311, New York, NY.
Academic Press.
Worboys, M. and Duckham, M. (2004). GIS, A Computing Perspective. CRC Press, Boca
Raton, FL.
Index

Page numbers in italic refer to figures.

A Bivariate and multivariate quantile LISA,

Absolute difference, 254 360–361
Aggregation Bivariate local join count statistics, 355–356,
defined, 49 356
dissolve functionality, 50–51, 50–51 Bivariate local Moran
in table, 51 defined, 336
Analytical inference, 255–256 inference, 337
Animation control panel, 98 univariate description
Apparent contagion, 258, 391 for each variable, 338–339, 339
Application Programming Interface (API), variables, 337, 338
403 Bivariate spatial correlation
Arc/great circle distance, 205 Moran scatter plot, 273–275
Asymptotic analysis, 180 variable and spatial lag, 273
Attribute similarity, 179, 252–253 Block weights, 220–222, 221
Autocorrelation statistic, spatial Bonferroni bounds approach, 307–308, 308
absolute difference, 254 Borrowing strength, 108, 109–112, 112
attribute similarity, 252–253 Box map, 87, 89–90, 90, 121, 122, 338
cross-product, 253 Box plot
defined, 252 and box map, 121, 122
locational similarity, 253 options, 122–123
null distribution, 252 for population change 2020–2010, 121,
squared difference, 254 122
type I error, 252 quartiles, 121
Automatic reprojection, 54, 54 Brazilian Index of Urban Structure (IBEU),
Averages chart 83
defined, 134, 171 Bubble chart, 142
Difference-in-Means Test, 172–175 categorical variables, 147–148, 148
map brushing, 136, 136–137, 137 defined, 144
selected and unselected observations, education, basic services, extreme
135–136 poverty, 145
Gapminder web site, 145
B implementation, 145, 145–146, 146
Background Color, 81 options, 146, 147
Bar charts, 118, 118 Bump hunting, 363
Base variable, 104, 105 BW joins, 351–352
Bayes Law, 110–111
BB joins, 351–352 C
Binary categories, 94–95, 94–95 Calculator
Bishop contiguity, 185 bivariate operations, 28
Bivariate analysis, see Scatter plot date and time functions, 28

417
418 Index

special functions, 27 micromaps, 151

interface, 26–27, 27 options, 153, 153
univariate functions, 27 statistical graphs, 153, 153–155
variable standardization, 27–28 Conditional statistical graphs, 153, 153–155
Cartesian coordinate system, 204–205 Connectivity graph, 46, 199–200, 200
Cartogram, 87 Connectivity histogram, 196–198
GeoDa implemention, 96 Connectivity map, 198–199, 199
map type, 96 Connectivity option
variables, 96, 97 in the map, 200–201
Category editor in the table, 201
custom category labels, 83 Contiguity-based weights, 179
description, 81 Contiguity weigths
interface, 81–83, 82 defined, 188
Child poverty and crowded housing higher-order contiguity, 192
bivariate local Moran cluster map, 340 ID variable, 189
box maps, 338 queen contiguity, 191
co-location, 338, 339 rook contiguity, 189–191, 190, 191
local Moran cluster map, 339 weights manager interface, 188, 188
multivariate local Geary cluster map, Coordinate reference system (CRS), 20–21,
344 21
Choropleth map, 67 defined, 39
Circular and redundant paths, 187 for non-geographers, 39–40, 40
Cluster maps, 94, 282 reprojection of current layer, 41
conventional and median local Moran, tessellations, 44–45
317 Correlation matrix, 132, 133
local Moran in 2010 and 2020, 319 Cross-product, 253
median Moran, 316, 317 Current layer, 52
raw and EB rate, 321 Curse of dimensionality
Clusters, 7, 302–304, 303–304, 310–311, 326, defined, 142
327, 331, 332 discrete evaluation points, variable
and spatial outliers, 315–316, 316, dimensions, 143
325–326, 235–326 empty space phenomenon, 143–144, 144
stability and condensed tree, 380–382 Custom categories, 83, 84
Cold spots, 250 cyberGIScience, 1, 403
Co-location local join count, 357–358,
356–358 D
Co-location map, 87, 94–96, 94 Data wrangling, 11
ColorBrewer categorical map palette, 93 Date and time functions, 28
Color scheme, 82 Date/time format, 25
Concept of contiguity Date/time variable, 25
contiguity-based weights, 218 DBSCAN algorithm
Thiessen polygons, 217, 218 border points and minimum points, 372,
weights file creation dialog, 219 372
Condensed dendrogram, 379 cluster map and summary, 371, 371
Condensed tree, 383–386, 384 core, border and noise points, 366–367,
Conditional local cluster maps, 311–312 367
Conditional maps, 63 defined, 368–371, 372
Conditional plots, 142 reachability, 367–368
defined, 151 sensitivity analysis, 371–372
implementation, 152 DBSCAN* algorithm
maps, 155–156, 155 core and noise points, 373
Index 419

dendrogram, 375–376, 375, 376 Exploratory data analysis (EDA), 3, 6, 17,

mutual reachability distance, 373–374, 63, 65, 66
374, 375 Exploratory spatial data analysis (ESDA), 3,
Difference-in-Means Test, 172–175 63, 66
Differential local Moran, 314, 318–320, 319 Extreme value maps
Differential Moran scatter plot, 268–271 box map, 89–90, 90
Dissolved districts map, 51 defined, 88
Dissolved districts table, 51 percentile map, 88–89
Distance-band weights, 206–207, 211–213 standard deviation map, 91–92, 91
Distance-based weights types, 88
band weights, 206–207
general distance weights, 208–209 F
K-nearest neighbor weights, 207–208, False discovery rate approach, 307, 308–309,
207 309
for polygons, 220 Family Wide Error Rate (FWER), 307
Distance metrics File format conversion, 16, 16
Cartesian coordinate system, 204–205 Fixed bandwidth, 230
defined, 204
general distance, 205–206 G
great circle distance, 205 GAL weights file, 191
3-D scatter plot, 142 Garden of forking paths, 392
data cube, 148 General distance matrix, 205–206, 216–217
implementation, 148–149 General distance weights, 208–209
interaction with, 149 Geocomputation, 1
options, 150–151 GeoDa software, 2, 3
selection in, 149–150, 150, 151 cluster analysis, 7
Dynamic box plot options, 167, 167 data entry, 4
Dynamic graphics, 65 data manipulation/table, 5
exploratory data analysis, 6
E functions, 12
EB local Moran, 314 mapping and geovisualization, 5–6
Economic weights, 206 menu structure, 399–401
Effect of smoothing operations/tools, 5
large populations, 113 preference settings 395, 396–397
small populations, 113 sample data, 7–8
Empirical Bayes approach, 111–112, 112 space-time analysis, 6
Empirical Bayes smoothed rate map, spatial autocorrelation analysis, 6–7
112–114 toolbar icons, 4
Empty space phenomenon, 143–144, 144 weights manager, 5
Endogenous classification, 81 Geodetic datum, 38
Equal intervals map, 71–72, 72 Geographic data science, 1
ESRI WorldStreetMap base map, 77–78, 78 Geographic information system (GIS)
Euclidean distance, 203, 210 CRS, see Coordinate reference system
European Petroleum Survey Group (EPSG), (CRS)
39 geodesy, 38
Event variable, 104, 105 geodetic datum, 38
Excess risk, see Relative risk georeferencing process, 38–39
Excess risk maps, 101, 106–108, 107, 108 latitude and longitude, 38
Existing weights, 192, 195 spatial observations, 38
Exogenous classification, 81 GeoJSON format, 13
Geometric centroids, 209
420 Index

Geospatial data science, 1 Higher-order contiguity, 185–187, 186, 192

Geospatial visual analytics, 66 High-high and low-low spatial
Geovisual analytics, 66 autocorrelation, 260
Geovisualization, 66 High-low and low-high spatial
Getis-Ord statistics autocorrelation, 260
and cluster map, 329, 330 Histogram
clusters and outliers, 330, 331 bar charts, 118, 118
inference, 329 basics, 116, 116–118, 117
ratio of weighted average, 329 defined, 116
Getis-Ord statistics vs. local Moran descriptive statistics, 119, 119
high-high clusters, 331, 332 linking map and histogram, 119–120,
low-low clusters, 331, 332 120
spatial outliers, 332, 333 Hot spots, 250
Global Navigation Satellite Systems (GNSS),
38 I
Google Groups Openspace list, 4 ID variable, 189
Great circle distance, 205 Inference
Grid cells, 22, 22 analytical, 255–256
Grouped variables, 269 defined, 255
in data table, 163 hypergeometric distribution353
individual variables, 163 permutation, 256–257
in Project File, 164, 165 Inter-layer linkage, 59, 59
space-time data table, 164–165 Inter-point distance matrix, 48
time editor interface, 162–163 Inter-quartile range (IQR), 89
GWT weights file, 212 Inverse distance weights
default setting, 236–237, 237
H distance-band weight, 236
HARKing, 392 GWT weights file, 228–229, 229
HDBSCAN algorithm K-nearest neighbor, 228
cluster map, 384 maximum inter-point distance, 227
cluster membership, 382, 386 properties of, 228
cluster stability and condensed tree, row-standardized, 237
380–382 spatial weights, 226
condensed dendrogram, 379 spatial window inverse distance sum,
condensed tree, 383–386, 384 237–238
core distance map, 386–387 weights file creation interface, 227
defined, 376–377
dendrogram, 384 J
level set, 377–378, 377 James-Stein paradox, 109
minimum spanning tree, 380, 381 JARKing, 392
minimum spanning tree connectivity
graph, 380 K
mutual reachability distance, 379 Kernel weights
outliers, 382–383 characteristics of, 231–232
persistence and excess of mass, 378–379, creation of, 231
379 Epanechnikov weights, 238
Health care access kernel function, 229
between 2010 and 2020, 269 KWT weights file, 232, 233
in 2010 and 2020, 269 nonparametric statistics, 229
to health care 2020–2010, 271 spatially lagged variables, 238
Heat map, 364–366, 365, 366 specification of, 229–231, 231
Index 421

K-nearest neighbor (KNN) weights, 207–208, M

207, 215, 216, 228, 346 Manhattan block distance, 204
Mantel test, 289
L Map animation, 98–99, 98, 99
Lattice data structure, 253 Map base layer (base map)
Layer position, 53–54 categorical variable, 78–79
Legend detailed base layer options, 77–78, 78
color, 68 options, 75, 76
gradation of values, 68 saving map as an image, 79, 79–80
Libgeoda architecture, 404 Stamen TonerLight base map, 75, 76, 77
Linked animation sanitation-infrastructure, tile concept, 75
99 Map brushing
Linked cartogram and map, 97 averages chart, 136, 136–137, 137
Linked contiguity histogram and map, 199 and scatter plot, 138, 138–139, 139
Linked selections, 59, 59 Map classification, 67–68
Linking and brushing, 66–67 Map layer setting options, 53
Linking map and histogram, 119–120, 120 Mapping and geovisualization, 5–6
Local Geary statistics Mapping options, 68
and cluster maps, 323 Maps
clusters and spatial outliers, 325–326, base layer (base map), see Map base
325–326 layer (base map)
global spatial autocorrelation, 322 legend color, 74, 74
implementation, 323–325 options, 73, 74
Local Geary vs. local Moran toolbar, 75, 75
high-high clusters, 327 Mean absolute deviation (MAD), 28
low-low clusters, 327 Mean center/centroids, 41, 42–44, 43, 44
other positive clusters, 327 Mean squared error (MSE), 110
spatial outliers, 328 Median local Moran, 314–317
Local Indicator of Spatial Association Median spatial lag, 235
(LISA), 252, 298, 391; see also Merging tables, 29–30, 30
Local Moran, Local Geary statistics Micromaps, 151
Local Moran Minimum Spanning Tree (MST), 380, 381
clusters, 302–304, 303–304 connectivity graph of mean centers, 47
with EB rate, 314, 320–322, 321 defined, 46
formulation, 299–300 options, 49
LISA statistic, 298 point coordinates, 47
randomization options, 301 Prim’s algorithm, 46–49, 48
significance and cluster map, 300–301, toy example point coordinates, 47
301 traveling salesperson problem, 46
spatial outliers, 304, 305 Minkowski metric, 204
statistic, 301–302 Modifiable areal unit problem (MAUP), 393
Local Neighbor Match Test, 345–350 Moran scatter plot, 254, 259, 259–261, 282
cardinality map, 348 bivariate spatial correlation, 273–276
multivariate local Geary, 349–350 brushing, 264–266, 265
significance map, 348 community area income per capita
Local neighbor significant matches, 349 variable, 261
Local significance map, 354, 354 for EB rates, 271–273, 273
Locational similarity, 253 high-high cluster location, 304
Location quotient (LQ), 101, 106 high-high observations, 303
LOWESS smoother, 263, 263 high-low observations, 305
Low-low clusters, 316 high-low spatial outlier location, 305
422 Index

LOWESS smoother, 263, 263 Percentile map, 87

randomization, 262 Permutation inference, 256–257
random seed, 262 p-hacking, 392
significant, 303 Point layers, 16–17, 17
static and dynamic spatial weights, 265, Poisson-Gamma model, 111
266 Polygon layers, 14–16, 15
Moran’s I statistics Population and population density, 103
inference, 255–257 Positive spatial autocorrelation, 250–251,
interpretation of, 258–259, 259 251, 260
Pearson correlation coefficient, 254 Posterior distribution, 110
spatial autocorrelation statistic, 254 Prim’s algorithm, 46–49, 48
Multidimensional scaling (MDS), 206 Prior distribution, 110
Multiple categories, 95–96 Project file, 83–85, 85, 192, 194
Multiple comparisons Pseudo p-value calculation, 256
approximations, 307 Python PySAL library, 87
Bonferroni bound, 307–308, 308
cut-off p-value, 307 Q
false discovery rate, 308–309, 309 Quadrat counts, 22
Multiple layers selection, 54–55, 55 Quantile LISA
Multivariate local Geary bivariate and multivariate, 360–361
Bonferroni bound, 341 in-situ correlation, 358–359
box map and local Geary cluster map, multivariate co-location, 361–362
342 spatial outliers, 361
curse of dimensionality, 340 univariate, 360
univariate analysis Quartile map, 89
multiple variables, 342–343 Queen contiguity histograms, 184, 191, 191,
Multivariate spatial autocorrelation problem, 198
335–336 Queries
indicator variable, 33
N save selected observations, 33–33, 33
Natural breaks map, 68, 72–73, 73 save selection indicator variable, 35
Negative spatial autocorrelation, 251, 251, selection on map classification, 35
261 selection tool, 31, 31–33, 32
Neighbor cardinality, 198 spatial selection, 34, 35
Non-parametric approaches, 282 Quintile maps
NSF-funded Center for Spatially Integrated defined, 69
Science, 2 housing index, Ceará, 69
O problem with, 70, 71
Oaxaca Development data set, 141 tied values, 70, 70
Oaxaca disability rate, 239–240
Open Geospatial Consortium (OGC), 13 R
Optimal bandwidth, 230 Randomization approach, 248, 249, 249, 262
Random sampling, 287–288
P Random seed, 262
Parallel coordinate plot (PCP), 141, 142 Rate map interface, 104
brushing, 158, 158–159 Rate smoothing, 101
clusters, 156 borrowing strength, 109–112, 112
clusters and outliers, 159, 159–160 defined, 239
options, 157–158 Oaxaca disability rate, 239–240
standardized variables, 157 spatial empirical Bayes smoothing,
visual data mining frameworks, 156 242–244
Index 423

spatial rate, 240–241, 240–242 Space-time regression, 6, 277, 278

variance instability, 109 Space-time weights, 193–196, 196, 222–223
Raw/crude rate map, 101, 104–105, 105 Spatial analysis, 1
Reference distribution, 256, 257, 258 Spatial autocorrelation analysis, 3, 6–7
Relative risk Spatial correlogram, 282
defined, 106 change parameters, 285
observed rate, 106 dashed zero axis, 285
SMR and LQ, 106 default, 285
Reproducibility, 394 display statistics, 286
Risk surface, 108 edit LOWESS parameters, 285
Rook contiguity histogram, 184, 189–191, functionality, 283
190, 191, 197 loan loss provisions, Italian banks 2016,
Rorschach test, 65 283
Row-standardized inverse distance weights, LOWESS and kernel regression, 282
237 maximum distance and number of bins,
Row-standardizing spatial weights, 183 286–287, 287
number of bins, 284
S parameter selection, 284
Scatter plot random sampling, 287–288
brushing, 138, 138 Spatial data
data options, 127, 126 attribute information, 12, 13
defined, 123 discrete segments, 13
food insecurity on poverty, 124, 124 input, 18, 19
LOWESS parameters, 128, 129, 129 locational information, 12, 13
regimes regression, 124–127, 125 Spatial data science, 1-2
smoother option, local regression, Spatial dependence, 1
128–129 Spatial empirical Bayes smoothing, 242–244
Scatter plot matrix, 130–134, 131–132 Spatial file formats, 13–14, 14
Serial and space-time regression, 279, 279 Spatial fixed effect, 268
Serial (temporal) correlation Spatial heterogeneity
health care access, 2020 and 2010, 277, averages chart, 135–137
278 defined, 1, 134
between spatial lags, 276–277, 278 Spatial join
Sherlock Holmes method, 392 areal unit identifier for point, 56
Show map boundary, 80 assign, 55–57
Shrinkage estimator, 271 community area identifier, 56
Significance and cluster map, 300–301, 301 count, 55, 57–58, 58
Simultaneous equation bias, 274 Spatial lag operation, 233
Single variables, see Box plot; Histogram Spatial layers
Smoothed distance scatter plot, 282 grid cells, 22, 22
attribute dissimilarity to distance, 288 point layers from coordinates, 19–21
bivariate, 294 Spatially extensive variables, 102–104,
empirical variogram, 288 103–104
Italy Community Banks data set, 289 Spatially intensive variables, 102–104,
with 150 km distance cut-off, 291 103–104
LOESS settings, 291–294 Spatially lagged variables, 233–235, 234, 235
LOESS setup, 291, 292 Spatially smoothed rate
Mantel test, 289 map, 240–241
multivariate extension, 290 in table, 241–242
observation locations, 288 Spatial outliers, 89, 304, 305, 328, 328, 332,
Sound scientific reasoning, 292 333, 361
424 Index

Spatial randomness balanced panel, 162

conditional perspective, 248 grouped variables, see Grouped
defined, 248 variables
parametric approach, 248–249, 249 Time player
randomization, 248, 249, 249 box plot over time, 166, 167
simulation of, 250, 250 dialog, 166, 166
simultaneous view, 248 scatter plot with time-lagged variable,
Spatial rate smoothing, 240–241, 240–242 167–169, 169
Spatial similarity, 179 space-time dynamics, 166
Spatial weights; see also Rate smoothing; thematic map over time, 169–171, 170,
Spatial weights matrix 171
asymptotic analysis, 180 Tobler’s first law of geography,
example spatial layout, 181 defined, 1
incidental parameter problem, 180 inverse distance weights, 266
row-standardizing, 183 TOP guidelines, 394
spatial dependence, 180 Treatment effect analysis, 171
types of contiguity, 183–184, 184 True contagion, 258, 391
Spatial weights matrix, 45, 179, 181, 181–182
Spatial window inverse distance sum, U
237–238 Unique values map, 87, 92, 93–94
Spatial window transformations, 236 Univariate functions, 27
Squared difference, 254 Univariate quantile LISA, 360
Stamen TonerLight base map Universal transverse Mercator (UTM) zones,
default map transparency, 76, 76 North America, 40
0.40 map transparency, 75, 76, 77
Standard deviation map, 87, 91–92, 91 V
Standardized mortality rate (SMR), 101, 106 Variable bandwidth, 230
Variable standardization, 27–28
T Variance instability, 101, 109
Table manipulations Voronoi diagrams, 44
calculator
bivariate operations, 28 W
date and time functions, 28 Weights characteristics, 212–213, 213,
description, 26 215–216, 218, 219–220
interface, 26, 27 connectivity histogram, 196–198
special functions, 27 make symmetric, 224
univariate functions, 27 set operations, 223, 224
variable standardization, 27–28 summary, 196
grid layer over chicago community areas, Weights file creation, 209, 213–214
22, 22 Weights manager interface, 5, 188
merging tables, 29–30, 30 World Geodetic System of 1984 (WGS 84),
options, 22, 23 38
variable properties WW joins, 351–352
edit, 23–24, 24
formatting date/time, 24–25, 25 X
operations on table, 25–26, 26 X-centroids, 209
other, 25 X-coordinate variables, 209
Tabular files, 17–18, 19
Tessellations, 44–45 Y
Thiessen polygons, 42, 44, 45 Y-centroids, 209
Time editor Y-coordinate variables, 209