STEAM-H: Science, Technology, Engineering, Agriculture,

Mathematics & Health

Norou Diawara Editor

Modern
Statistical
Methods for
Spatial and
Multivariate Data
STEAM-H: Science, Technology, Engineering,
Agriculture, Mathematics & Health
STEAM-H: Science, Technology, Engineering,
Agriculture, Mathematics & Health

Series Editor
Bourama Toni
Department of Mathematics
Howard University
Washington, DC, USA

This interdisciplinary series highlights the wealth of recent advances in the pure
and applied sciences made by researchers collaborating between fields where
mathematics is a core focus. As we continue to make fundamental advances in
various scientific disciplines, the most powerful applications will increasingly be
revealed by an interdisciplinary approach. This series serves as a catalyst for these
researchers to develop novel applications of, and approaches to, the mathematical
sciences. As such, we expect this series to become a national and international
reference in STEAM-H education and research.
Interdisciplinary by design, the series focuses largely on scientists and math-
ematicians developing novel methodologies and research techniques that have
benefits beyond a single community. This approach seeks to connect researchers
from across the globe, united in the common language of the mathematical sciences.
Thus, volumes in this series are suitable for both students and researchers in a variety
of interdisciplinary fields, such as: mathematics as it applies to engineering; physical
chemistry and material sciences; environmental, health, behavioral and life sciences;
nanotechnology and robotics; computational and data sciences; signal/image pro-
cessing and machine learning; finance, economics, operations research, and game
theory.
The series originated from the weekly yearlong STEAM-H Lecture series
at Virginia State University featuring world-class experts in a dynamic forum.
Contributions reflected the most recent advances in scientific knowledge and were
delivered in a standardized, self-contained and pedagogically-oriented manner to a
multidisciplinary audience of faculty and students with the objective of fostering
student interest and participation in the STEAM-H disciplines as well as fostering
interdisciplinary collaborative research. The series strongly advocates multidis-
ciplinary collaboration with the goal to generate new interdisciplinary holistic
approaches, instruments and models, including new knowledge, and to transcend
scientific boundaries.

More information about this series at http://www.springer.com/series/15560

Norou Diawara
Editor

Modern Statistical Methods

for Spatial and Multivariate
Data

123
Editor
Norou Diawara
Department of Mathematics and Statistics
Old Dominion University
Norfolk, VA, USA

ISSN 2520-193X ISSN 2520-1948 (electronic)

STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health
ISBN 978-3-030-11430-5 ISBN 978-3-030-11431-2 (eBook)
https://doi.org/10.1007/978-3-030-11431-2

Mathematics Subject Classification: 62Hxx, 62H11, 30C40

© Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

Statistical ideas and concepts have increasing impacts at many levels. From H.G.
Wells’ 1903 book Mankind in the Making, a quote paraphrased by Sam Wilks in his
1950 American Statistical Association speech states: “Statistical thinking will one
day be as necessary for efficient citizenship as the ability to read and write.” Current
disciplinary boundaries encourage interaction between scientists and the sharing
of information and educational resources. Researchers from these interdisciplinary
fields will find this book an important resource for the latest statistical methods for
spatial and multivariate data.
Given the increasingly complex data world we live in, this volume takes
on unique approach with respect to methodology, simulation, and analysis. The
environment provides the perfect setting for exciting opportunities and interdisci-
plinary research and for practical and robust solutions, contributing to the science
community in large. The National Institutes of Health and the Howard Hughes
Medical Institute have strongly recommended that undergraduate biology education
should incorporate mathematics and statistics, physics, chemistry, computer science,
and engineering until “interdisciplinary thinking and work become second nature.”
In that sense, this volume is playing an ever more important role in the physical and
life sciences, engineering and technology, data sciences, and artificial intelligence,
blurring the boundaries between scientific disciplines.
The shared emphasis of these carefully selected and refereed contributed chapters
is on important methods, research directions, and applications of analysis including
within and beyond mathematics and statistics. As such the volume promotes statis-
tical sciences and their applications to physical, life, and data sciences. Statistical
methods for spatial and multivariate data have gained indeed tremendous interest
over the last decades and are rapidly expanding. This book features recent advances
in statistics to include the spatio-temporal aspects, classification techniques, the
multivariate outcomes with zero and doubly inflated data, the copula distributions,
the wavelet kernels for support matrix machines, and feasible algorithmic solutions.
Events are sometimes affected by a set of covariates accounted in space locations
and times. With the influx of big data, statistical tools are identified, tested, and

v
vi Preface

improved to fill in the gaps sometimes found in the environmental, financial, and
healthcare fields.
This volume stretches our boundaries of knowledge for this fascinating and
ongoing area of research. It features the following chapters:
The chapter “Functional Form of Markovian Attribute-Level Best-Worst Dis-
crete Choice Modelling” by Amanda Working, Mohammed Alqawba, and Norou
Diawara provides modeling discrete choice experiments. The challenging parts can
be linked to the large number of covariates, issues with reliability, and the condition
that consumer behaviors is a forward evolving activity/practice. By extending the
idea of stationary process, the authors present a dynamic model with evaluation
under random utility analysis. The simulated and aggregated data examples show
the flexibility and wide applications of the proposed techniques.
The chapter “Spatial and Spatio-temporal Analysis of Precipitation Data from
South Carolina” from David Hitchcock, Haigang Liu, and S. Zahra Samadi presents
both spatial and spatio-temporal models for rainfall in South Carolina during a
period including one of the most destructive storms in state history. The models
proposed have allowed to determine several covariates that affect the rainfall and to
interpret their effects.
The chapter “A Sparse Areal Mixed Model for Multivariate Outcomes, with an
Application to Zero-Inflated Census Data” from Donald Musgrove, Derek S. Young,
John Hughes, and Lynn E. Eberly describes the multivariate sparse areal mixed
model (MSAMM) as an alternative to the multivariate conditional autoregressive
(MCAR) models. The MSAMM is capable of providing superior fit relative to
models provided under independent or univariate assumptions.
The next chapter “Wavelet Kernels for Support Matrix Machines” by Edgard
M. Maboudou-Tchao provides support vector machine techniques to the matrix-
based method support matrix machines (SMM), accepting matrix as input, and
then proposing new wavelet kernels for SMM. Such techniques are very powerful
approximations for nonstationary signals.
In the chapter “Properties of the Number of Iterations of a Feasible Solutions
Algorithm,” the authors Sarah A. Janse and Katherine L. Thompson provide
statistical guidance for the number of iterations by deriving a lower bound on the
probability of obtaining the statistically optimal model in a number of iterations of
algorithm along with the performances of the bound.
Classification techniques are commonly used by scientists and businesses alike
for decision-making. They involve assignment of objects (or information) to pre-
defined groups (or classes) using certain known characteristics such as classifying
emails as real or spam using information in the subject field. In the chapter “A
Primer of Statistical Methods for Classification,” the authors Rajarshi Dey and
Madhuri S. Mulekar describe two soft and four hard classifiers popularly used by
statisticians in practice. To demonstrate their applications, two simulated and three
real-life datasets are used to develop classification criteria. The results of different
classifiers are compared using misclassification rate and an uncertainty measure.
In the chapter entitled “A Doubly-Inflated Poisson Distribution and Regression
Model” by Manasi Sheth-Chandra, N. Rao Chaganty, and R. T. Sabo, doubly
Preface vii

inflated Poisson distribution is presented to account for count inflation at some value
k in addition to that seen at zero, while it was also incorporated into the generalized
linear models framework to account for associations with covariates.
The chapter “Multivariate Doubly Inflated Negative Binomial Distribution Using
Gaussian Copula” by Joseph Mathews, Sumen Sen, and Ishapathik Das presents a
model for doubly inflated count data using the negative binomial distribution, under
Gaussian copula methods. The authors also provide visuals of the bivariate doubly
inflated negative binomial model.
Moran’s Index is a statistic that measures spatial autocorrelation, quantifying the
degree of dispersion (or spread and properties) of components in some location/area.
Recognizing that a single Moran’s statistic may not give a sufficient summary of the
spatial autocorrelation measure, local spatial statistics have been gaining popularity.
Accordingly, the chapter “Quantifying Spatio-temporal Characteristics via Moran’s
Statistics” by Jennifer Matthews, Norou Diawara, and Lance Waller proposes to
partition the area and compute the Moran’s statistic of each subarea.
The book as a whole certainly enhances the overall objective of the series, that
is, to foster the readership interest and enthusiasm in the STEAM-H disciplines
(Science, Technology, Engineering, Agriculture, Mathematics, and Health), to
include statistical and data sciences, and to stimulate graduate and undergraduate
research through effective interdisciplinary collaboration.
The editor of the current volume is affiliated with the Department of Mathematics
and Statistics at Old Dominion University, Norfolk, Virginia. The department has
the unique distinction of being the only one in the Commonwealth of Virginia
Hampton area to offer B.S., M.S., and Ph.D. degrees in Computational and Applied
Mathematics, with an active research program supported by NASA, NSF, EVMS,
JLab, and the Commonwealth of Virginia.

Washington, DC, USA Bourama Toni

Norfolk, VA, USA Norou Diawara
Acknowledgments

We express special gratitude to the STEAM-H series Editor-in-Chief, Bourama

Toni, Professor and Chair, in the Department of Mathematics at Howard Univer-
sity, for the opportunity to produce such a volume. Among the many we thank
and acknowledge are the editors and associate editor at Springer, in particular
Ms. Dimana Tzvetkova. Your inputs have facilitated and made the presentation very
elegant. We are grateful to the many reviewers, both internal and external, for their
professionalism; they raised many questions and challenged the materials featured
in this book with insightful comments and suggestions from concepts to exposition.
We would like to express our most sincere appreciation to all the contributors
for the excellent research work. They all made this volume a reality for the greater
benefit of the community of Statistical Scientists.

ix
Contents

Functional Form of Markovian Attribute-Level Best-Worst Discrete

Choice Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Amanda Working, Mohammed Alqawba, and Norou Diawara
Spatial and Spatio-Temporal Analysis of Precipitation Data from
South Carolina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Haigang Liu, David B. Hitchcock, and S. Zahra Samadi
A Sparse Areal Mixed Model for Multivariate Outcomes,
with an Application to Zero-Inflated Census Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Donald Musgrove, Derek S. Young, John Hughes, and Lynn E. Eberly
Wavelet Kernels for Support Matrix Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Edgard M. Maboudou-Tchao
Properties of the Number of Iterations of a Feasible Solutions Algorithm 95
Sarah A. Janse and Katherine L. Thompson
A Primer of Statistical Methods for Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Rajarshi Dey and Madhuri S. Mulekar
A Doubly-Inflated Poisson Distribution and Regression Model . . . . . . . . . . . . 131
Manasi Sheth-Chandra, N. Rao Chaganty, and Roy T. Sabo
Multivariate Doubly-Inflated Negative Binomial Distribution
Using Gaussian Copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Joseph Mathews, Sumen Sen, and Ishapathik Das
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics . . . . . . 163
Jennifer L. Matthews, Norou Diawara, and Lance A. Waller

xi
Contributors

Mohammed Alqawba Department of Mathematics and Statistics, Old Dominion

University, Norfolk, VA, USA
N. Rao Chaganty Department of Mathematics and Statistics, Old Dominion
University, Norfolk, VA, USA
Ishapathik Das Department of Mathematics, Indian Institute of Technology
Tirupati, Tirupati, India
Rajarshi Dey Department of Mathematics and Statistics, University of South
Alabama, Mobile, AL, USA
Norou Diawara Department of Mathematics and Statistics, Old Dominion Univer-
sity, Norfolk, VA, USA
Lynn E. Eberly Division of Biostatistics, University of Minnesota, Minneapolis,
MN, USA
David B. Hitchcock Department of Statistics, University of South Carolina,
Columbia, SC, USA
John Hughes Department of Biostatistics and Informatics, University of Colorado,
Denver, CO, USA
Sarah A. Janse Center for Biostatistics, The Ohio State University, Columbus,
OH, USA
Haigang Liu Department of Statistics, University of South Carolina, Columbia,
SC, USA
Jennifer L. Matthews Department of Mathematics and Statistics, Old Dominion
University, Norfolk, VA, USA
Edgard M. Maboudou-Tchao Department of Statistics, University of Central
Florida, Orlando, FL, USA

xiii
xiv Contributors

Joseph Mathews Department of Mathematics and Statistics, Austin Peay State

University, Clarksville, TN, USA
Madhuri S. Mulekar Department of Mathematics and Statistics, University of
South Alabama, Mobile, AL, USA
Donald Musgrove Medtronic, Minneapolis, MN, USA
Roy T. Sabo Department of Biostatistics, Virginia Commonwealth University,
Richmond, VA, USA
S. Zahra Samadi Department of Civil and Environmental Engineering, University
of South Carolina, Columbia, SC, USA
Sumen Sen Department of Mathematics and Statistics, Austin Peay State Univer-
sity, Clarksville, TN, USA
Manasi Sheth-Chandra Center for Global Health, Old Dominion University,
Norfolk, VA, USA
Katherine L. Thompson Department of Statistics, University of Kentucky,
Lexington, KY, USA
Lance A. Waller Department of Biostatistics, Rollins School of Public Health,
Emory University, Atlanta, GA, USA
Amanda Working Department of Mathematics and Statistics, Old Dominion
University, Norfolk, VA, USA
Derek S. Young Department of Statistics, University of Kentucky, Lexington, KY,
USA
Functional Form of Markovian
Attribute-Level Best-Worst Discrete
Choice Modeling

Amanda Working, Mohammed Alqawba, and Norou Diawara

1 Introduction and Motivation

In today’s consumer society, people are provided multiple alternatives in product

from which to choose the one that benefits them the most. Such situations include
purchasing a smartphone, choosing a car insurance company, or deciding upon
vacation destinations. Discrete choice experiments (DCEs) are designed to elicit
information from consumers as to why they choose the products or services that
they do. DCEs are applicable in multiple fields such as health system program,
public policy, transportation research, marketing, and economics.
Discrete choice models (DCMs) are the statistical models describing individuals’
preferences for products or services. In traditional DCEs, individuals are given a
series of hypothetical scenarios described by attributes forming what is called a
profile. The set of profiles is the choice set, from which the respondent chooses
the best alternative that suits their needs the most as proposed in Louviere et al.
(2000). Note that traditional DCEs and their models are built around the work done
by Thurstone (1927) and the theoretical basis for it by McFadden (1974).
Although some information are gained from traditional DCEs, they fail to
provide knowledge about the impact of the attributes when comparing the utilities
between alternatives (Flynn et al. 2007). However, best-worst scaling (BWS)
experiments (Marley and Louviere 2005) addressed such issues by asking the
respondents to choose the best and worst alternatives from a choice set instead of

A. Working · M. Alqawba · N. Diawara ()

Department of Mathematics and Statistics, Old Dominion University, Norfolk, VA, USA
e-mail: ndiawara@odu.edu

© Springer Nature Switzerland AG 2019 1

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_1
2 A. Working et al.

just the best as is done in the traditional DCEs. Besides having the information
of what the best and worst are for the respondents, Marley and Louviere (2005)
stated that BWS experiments provide information about the respondent’s ranking of
products.
BWS experiments are divided into three cases: the object case, the profile case,
and the multi-profile case (Louviere et al. 2015). In an object case, a list of objects,
scenarios, or attributes are given to respondents and the latter choose the best and
worst alternative. Unlike in traditional DCEs, no information about the object is
provided to the respondents. In the profile case which is also known as attribute-
level best-worst case, information or attributes about the alternatives are provided.
In this type of experiments, profiles composed of attribute-levels for each attribute
describing a product are determined. From the profiles, respondents are tasked
with choosing the best and the worst attribute-level pair. These experiments seek
to determine the extent to which attribute and their associated attribute-levels affect
consumer behavior. Furthermore, in attribute-level best-worst DCEs the levels of the
attributes are well defined and vary across profiles or products, providing sufficient
information to measure their impact. For example, Knox et al. (2012) study women’s
contraceptives by using unbalanced design, with seven attributes (product, effect
on acne, effect on weight, frequency of administration, contraceptive effectiveness,
effect on bleeding, and cost) with associated number of attribute-levels 8, 3, 4, 4,
8, 9, and 6, respectively. The attribute effect on acne had three levels (worsens,
improves, or no effect). Finally, the multi-profile case most closely resembles the
traditional DCEs in such that a set of profiles describing products are provided to
the respondents and the respondents choose the best and worst products from the
choice set.
In this chapter, extension to the existing work done on partial profile models for
attribute-level best-worst DCEs, or profile-case BWS, is presented. Attribute-level
best-worst data are presented as indicator functions demonstrating the equivalence
of these models to the traditional attribute-level best-worst models. The indicator
functions are then generalized providing an alternative method for accounting for
the attributes of attribute-level models. The functional form of the data definition
provides an adaptive model able to conform changes in the profile or set of attributes
under discrete choice modeling (DCM). We also allow changes in decisions/utilities
over time under Markov decision process (MDP). The conditional logit model is
used in the DCMs.
The chapter is organized as follows: attribute-level best-worst DCMs are intro-
duced in Sect. 2. In Sect. 3, functional form of attribute-level best-worst DCMs is
presented. Section 4 considers Markov decision process (MDPs) with regard to
time sensitive attribute-level best-worst DCEs. Simulated data example of functional
form of Markovian attribute-level best-worst DCMs over time and results are
described in Sect. 5. We end with a conclusion in Sect. 6.
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 3

2 Attribute-Level Best-Worst Discrete Choice Model

2.1 Literature Review

Discrete choice experiments (DCEs) and their modeling describe consumer behav-
iors. Given a set of descriptors about a product, one can estimate the probability an
alternative is chosen provided a statistical model appropriate to the data. However,
these models are limited in the information they provide. According to Lancsar et al.
(2013), there exist only two ways to gain more information from traditional DCEs:
either increase the sample size or increase the number of choice sets evaluated by
respondents with the risk of adding burden on the respondents in the experiments.
Alternatively, Louviere and Woodworth (1991) and Finn and Louviere (1992)
presented best-worst scaling experiments that are modified DCEs designed to elicit
more information about choice behavior than the pick one approach implemented
in the traditional DCEs without the added burden on the respondents.
Although the experiments were presented in the early 1990s, it was not until
Marley and Louviere (2005) that the mathematical probabilities and properties were
formally presented. Marley and Louviere (2005), Marley et al. (2008), and Marley
and Pihlens (2012) provided the probability and properties to best-worst scaling
experiments for the profile case and in multi-profile version; however, utility was not
introduced. Additionally, Lancsar et al. (2013) provided the probability and utility
definition for the multi-profile experiments that include the sequential best-worst
choice from a set of choices. Louviere and Woodworth (1983) stated that orthogonal
main effects and fractional factorial designs provide better parameter estimates than
other designs. In application to best-worst scaling experiments, balanced incomplete
block designs (BIBD) (Louviere et al. 2013; Parvin et al. 2016) and orthogonal main
effects plans (OMEPs) are popular designs (Flynn et al. 2007; Knox et al. 2012;
Street and Knox 2012). These designs and their properties are examined by Street
and Burgess (2007). Louviere et al. (2013) looked at the design of experiments for
best-worst scaling experiments and stated that it is possible to determine individual
parameter estimates for the respondents.
This chapter focuses on the profile case, also known as attribute-level best-worst
DCEs. These experiments seek to determine the extent to which attributes and their
associated attribute-levels impact consumer behavior. Louviere and Timmermans
(1990) introduced hierarchical information integration (HII) for the examination
of the valuation of attributes in DCEs. Under HII, the impact of an attribute
necessitates discerning the various levels of the attribute. An experiment must
be designed in such a way that can measure the different levels varying across
products and determine such an impact. In attribute-level best-worst DCEs, the
levels of attributes are well defined and vary across profiles, or products, providing
sufficient information to measure their impact. Attribute-level best-worst discrete
choice experiments provide more information into consumer’s choices of products
than the usual discrete choice experiments and add more value to the understanding
4 A. Working et al.

of the data (Marley and Louviere 2005). Those models outperform the standard logit
modeling in terms of goodness of fit as mentioned in Hole (2011) in the context of
attribute attendance.
Understanding the impact attribute and attribute-levels have on utility is desir-
able. The guiding ideology in DCEs is that consumers behave in a way to maximize
utility. Understanding the impacts attributes and attribute-levels have on consumer
behavior provides information with regard to developing and advertising a product,
service, or policy to consumers. A preponderance of the literature on attribute-
level best-worst DCEs are empirical studies often in the area of health systems
research and marketing. Examples include Flynn et al. (2007) on seniors’ quality
of life, Coast et al. (2006) and Flynn et al. (2008) on dermatologist consultations,
Marley and Pihlens (2012) on cell phones, Knox et al. (2012, 2013) on choices in
contraceptives for women.
While there exists literature on attribute-level best-worst DCEs, it is rather scarce
compared to the work done on traditional DCEs. In this section, we provide utility
definition and the resulting choice probabilities and properties. We use the utility
definition and choice probabilities to extend the work done by Grossmann et al.
(2009) to fit models on a function of the attributes and attribute-levels to reflect
fluctuation that are inherent in DCE over time.

2.2 Notations, Theory, and Properties

Attribute-level best-worst scaling are modified DCEs designed to elicit the impact
the attributes and attribute-levels have on the utility of a product. As mentioned
by Louviere and Timmermans (1990), an experiment must be designed in a way
to evaluate combinations of attribute-levels to obtain information about attribute
impacts on utility. Best-worst attribute-level DCEs provide such an experimental
design to attain these impacts.
In the attribute-level best-worst DCEs, each product is represented by a profile
x = (x1 , x2 , . . . , xK ), where xk is the attribute-level for the kth attribute Ak
that makes up the product with k = 1, 2, . . . , K. The attribute-levels take values
from 1 to lk for k = 1, 2, . . . , K. The number of possible profiles is given by
Kk=l li . Full factorial designs are generally not used due to the large number of
profiles. Alternatively, OMEP designs are promoted in the literature as efficient and
optimal provide sufficient information to estimate model parameters (Louviere and
Woodworth 1983; Street and Burgess 2007; Street and Knox 2012).
In these experiments, respondents are tasked with choosing a pair of attribute-
levels that contains the best and the worst attribute-level for a given profile. For
every profile, the choice set is then:

Cx = {(x1 , x2 ), . . . , (x1 , xK ), (x2 , x3 ), . . . , (xK−1 , xK ), (x2 , x1 ), . . . , (xK , xK−1 )},

Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 5

where the first attribute-level is considered to be the best and the second is the worst.
From the profile Cx , the respondent determines from the τ = K(K − 1) choices
given which is the best-worst pair.
In our setup, we extend the state of choices as follows. Let there be G profiles
and the associated profiles given as:

x1 = (x11 , x12 , . . . , x1K )

x2 = (x21 , x22 , . . . , x2K )
..
.
xG = (xG1 , xG2 , . . . , xGK ).

The corresponding choice sets for the G profiles are given in Fig. 1. To simplify the
notation, we may interchange C1 with Cx1 , C2 with Cx2 . . . , and CG with CxG .
 
The total number of attribute-levels is L = ki=1 li , and J = K k=1 lk (L − lk ) is
the total number of unique attribute-level pairs in the experiment (Street and Knox
2012). Within each of the G choice sets, thereare τ = K(K − 1) choice pairs.
In the experiment, there is a total of J = K k=1 lk (L − lk ) alternatives. However,
within a choice set there is a total of τ = K(K − 1) choices in each of the G choice
sets evaluated. Each respondent will have made G choices within the experiment.
The response variable representing the choices within each of the choice sets for the

Fig. 1 The G choice sets in an experiment with corresponding choice pairs

6 A. Working et al.

experiment is binary data and denoted as:


1, if sth respondent chooses j th alternative in the ith choice set,
Yisj = (2.1)
0, otherwise,

for i = 1, 2, . . . , G, s = 1, 2, . . . , n and j = 1, 2, . . . , τ .
For the attribute-level best-worst DCEs, the data, X, is composed of indicators for
the best and worst attributes and attribute-levels. Consider the choice pair (xij , xij  )
from the choice set Ci , for i = 1, 2, . . . , G, j = j  , j, j  = 1, 2,
. . . , K, and
1 ≤ xij ≤ lj . Let X be the J × p design matrix, where p = K + K k=1 lk . The
rows of X correspond to the possible choice pairs. Let XA1 , XA2 , . . . , XAK be the
data corresponding to the attributes Ak , k = 1, 2, . . . , K. Then,
⎧
⎪
⎪ 1, if xij ∈ Ak for k = 1, 2, . . . , K,
⎨
XAk = −1, if xij  ∈ Ak for k = 1, 2, . . . , K,
⎪
⎪
⎩ 0, otherwise.

Let XAk xik be the data for the attribute-level 1 ≤ xik ≤ lk within attribute
Ak , ∀k = 1, 2, . . . , K. Referring to the choice pair (xij , xij  ), the corresponding
data for the attribute-levels are given by,
⎧
⎪
⎪ 1, if xij = xik ∈ Ak is the best attribute-level,
⎨
XAk xik = −1, if xij  = xik ∈ Ak is the worst attribute-level,
⎪
⎪
⎩ 0, otherwise.

Marley and Louviere (2005) developed the probability theory for best-worst
scaling experiments including attribute-level best-worst DCEs. In attribute-level
best-worst DCEs, there are two components being modeled, the best choice and
the worst choice of attribute-levels from a profile xi , where i = 1, 2, . . . , G.
Under random utility theory (Marschak 1960), there are random utilities Uij
corresponding to the τ attribute-levels in the choice set and an individual chooses
an alternative with highest utility, i.e., they are not independent. Consider the
choice pair (xij , xij  ), for i = 1, 2, . . . , G, j, j  = 1, 2, . . . , K, and j = j  .
According to Marley and Louviere (2005), the definition of utility consistent with
random utility theory satisfies, Uij = −Uij  and Uijj  = Uij − Uij  for i =
1, 2, . . . , G, j, j  = 1, 2, . . . , K, and j = j  , and Uij = Vij + ij where Vij is
a systematic component and ij is an error term that distributed as type I extreme
value distribution (McFadden 1978). Hence, the definition of utility associated with
the best-worst choice pair under random utility theory is given by:

Uijj  = Uij − Uij  = Vij − Vij  + ij − ij  (2.2)

for i = 1, 2, . . . , G, j, j  = 1, 2, . . . , K, and j = j  .
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 7

The definition of the utilities under the random utility model is unable to
be modeled under the conditional logit model due to the definition of the error
components (Marley and Louviere 2005).
If we assume that the random error terms are independently and identically
distributed type I extreme value distribution, or the Gumbel distribution, then the
choice probability comes directly from the conditional logit. The choice probability
is then,

BWxi (xij , xij  ) = P (Uijj  > Uiqq  , ∀q, q  ∈ Ci )

= P (Vijj  + ijj  > Viqq  + iqq  , ∀q, q  ∈ Ci )
= P (iqq  − ijj  < Vijj  + Viqq  , ∀q, q  ∈ Ci ), (2.3)

where j = j  , j, j  = 1, 2, . . . , K, and i = 1, 2, . . . , G.
Since the error terms come from the type I extreme value distribution, their
difference is a logistic distribution. It follows from McFadden (1974) that the best-
worst attribute-level choice probability is defined by the conditional logit as:

exp(Vijj  )
BWxi (xij , xij  ) =  , (2.4)
exp(Viqq  )
(xiq ,xiq  )∈Ci

where j = j  , j, j  = 1, 2, . . . , K, and i = 1, 2, . . . , G.
Marley et al. (2008) provide essential properties to the above probabilities. They
define the choice probability as:

b(xij )
b(xij  )
BWxi (xij , xij  ) =  b(xij )
, (2.5)
 b(xij  )
∀ xij ,xij  ∈Cxi ,j =j 

where xij is chosen as the best attribute-level, and xij  is the worst attribute-level,
and b is some positive scale function or impact of attribute for j = j  , j, j  =
1, 2, . . . , K, and i = 1, 2, . . . , G. Under the conditional logit, the scale function is
defined as b(xij ) = exp(Vij ), and the probability is as given in Eq. (2.4).
Essential properties of probability hold for Eq. (2.5), as

BWxi (xij , xij  ) ≥ 0, ∀i, j (2.6)

and

BWxi (xi , xj ) = 1, (2.7)


∀ xij ,xij  ∈Cxi ,j =j 

where j, j  = 1, 2, . . . , K, j = j  , and ∀i = 1, 2, . . . , G.
8 A. Working et al.

With such assumptions, the consumer is expected to select choices with higher
i . Attribute-level best-worst models
BWxi values. We denote BWxi (xij , xij  ) as Pjj 
are called maxdiff models because they maximize the difference in utility.
Associated properties of the maxdiff model mentioned in Marley et al. (2008)
are:
1. Invertibility: For profile i,

 Pj  j = Pqq  Pq  q ,
i i i i
Pjj

where 1 ≤ j, j  , q, q  ≤ k and j = j  and q = q  .

2. Reversibility: For profile i and i
 
 Pqq  = Pq  q Pj  j ,
i i i i
Pjj

where xij  = xi  q , and j = j  and q = q  .

3. Reversibility: For profiles i, i , and i ,
   
 Pqq  Prr  = Pr  r Pq  q Pj  j ,
i i i i i i
Pjj

where xij  = xi  q , xi  q  = xi  r , and xij = xi  r  , and j = j  , q = q  , and r = r  .

4. Reversibility: For profiles i, i , i , and i ,
     
 Pqq  Prr  Pww  = Pw  w Pr  r Pq  q Pj  j ,
i i i i i i i i
Pjj

where xij  = xi  q , xi  q  = xi  r , xi  r = xi  w , and xij = xi  w , and j = j  , q =

q  , r = r  and w = w  .
Now, the systematic component in Eq. (2.2) can be expressed as,

Vijj  = Vij − Vij  = (xij − xij  ) β, (2.8)

where j = j  , j, j  = 1, 2, . . . , K, and i = 1, 2, . . . , G. The data xij are indicators

of the attribute xij ∈ Aj and the attribute-level xij . The systematic component Vij
can be written as,

Vij = xij β = βAj + βxij Aj , (2.9)

where j = j  , j, j  = 1, 2, . . . , K, and i = 1, 2, . . . , G.
We assume the error terms come from a type I extreme value distribution and use
the conditional logit model to estimate the p × 1 parameter vector,

β  = (βA1 , βA2 , . . . , βAk , βA1 1 , βA1 2 , . . . , βA1 l1 , . . . , βAk 1 , . . . , βAk lk ). (2.10)

Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 9

The likelihood for estimating the model parameters based on a random sample n
individuals as in Eq. (2.1) is given as:


n 
G 
Y
L(β, Y) = Pijjisj . (2.11)
s=1 i=1 j =j 

Estimation of the parameters is done in SAS® maximizing the likelihood given in

Eq. (2.11).
Attribute and attribute-level data in the experiments are a series of 1 s and

0 s, indicating the attributes and attribute-levels in the choice pair with positive
and negative signs for best and worst attribute-levels, respectively. When fitting a
conditional logit model to the data, parameter estimates for the last attribute and
last attribute-level for each attribute are not retrievable due to singularity issues.
According to Flynn et al. (2007), these parameter estimates are needed to determine
the impact of attribute, which is the essential purpose for experiments of this design.
To estimate these parameters, the following identifiability condition defined on the
parameters of the attribute-levels must be met,

lk
i=1 βi = 0 (2.12)

or

i −1
l

βlk = − βj (2.13)
j =1

for all k = 1, 2, . . . , K (Street and Burgess 2007; Flynn et al. 2007; Grasshoff et al.
2003).
Next, the goal will be to build a functional form of the attributes and the attribute-
levels and estimate the associated model parameters that reflect their utilities.

3 Functional Form of Attribute-Level Best-Worst Discrete

Choice Model

The attribute-level best-worst DCEs are modified traditional DCEs. Models and
theory done for traditional DCEs have not been completely evaluated in terms of
best-worst scaling experiments. It is of interest to us to extend the model built on a
function of the data as presented in Grasshoff et al. (2003, 2004), and Grossmann
et al. (2009) to the attribute-level best-worst DCEs. On extending this work to these
experiments, we provide an additional way to define the systematic component that
provides flexibility not seen in traditional methods.
10 A. Working et al.

Considering functions of the attributes as they enter into the utility function is
not a new idea. Van Der Pol et al. (2014) present the systematic components of the
utility defined as linear functions, quadratic functions, or as stepwise functions of
the attributes. Grasshoff et al. (2013) define the functions as regression functions of
the attributes and attribute-levels in the model.
In the attribute-level best-worst DCEs, a set of G profiles, or products, are
examined. The profiles are given as xi = (xi1 , xi2 , . . . , xiK ), where xij is the
attribute-level in profile i = 1, 2, . . . , G that corresponds to the j th attribute, where
j = 1, 2, . . . , K. The choice task for respondents is to choose the best-worst pair
of attribute-levels. In the experiment, respondents make paired comparisons within
the profiles instead of between as in traditional DCEs.
In the attribute-level best-worst DCEs, the utility of the pairs is composed of
the utility corresponding to the best attribute-level and the worst attribute-level. The
regression functions presented in Grasshoff et al. (2003) are applied to the attributes
and attribute-levels within the respective systematic components. Let f be the set
of regression functions for the best attribute-levels in the pairs, and g the set of
regression functions for the worst attribute-levels in the pairs. The p × 1 parameter
vector β still must satisfy the identifiability condition given in Eq. (2.13).
As noted in Marley and Louviere (2005), the inverse random utility model must
be used so that the properties are met for the conditional logit model. Taking the
systematic component defined in Eq. (2.2), the functional systematic component for
the pair (xij , xij  ) is defined as:

Vijj  = Vij − Vij  = (f(xij  ) − g(xij  )) β, (3.1)

where j, j  = 1, 2, . . . , K, j = j  , and i = 1, 2, . . . , G.
The probability an alternative is chosen depends on the definition of the utility
and the distribution of the error terms. Referring back to Eq. (2.5) under the
conditional logit, the probability is

exp(Vijj  )
BWxi (xij , xij  ) = 
exp(Viqq  )
(xiq ,xiq  )∈Ci

exp(Vij − Vij  )
= 
exp(Viq − Viq  )
(xiq ,xiq  )∈Ci

exp((f(xij ) − g(xij  )) β)

=  , (3.2)
exp((f(xiq  ) − g(xiq  )) β)
(xiq ,xiq  )∈Ci

where i = 1, 2, . . . , G, j, j  = 1, 2, . . . , K, and j = j  .
The forms of the systematic components of the utilities as well as their associated
probabilities depend on the definition of the regression functions f and g. We define
the regression functions used in the traditional attribute-level best-worst model and
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 11

extend the definition of the regression functions to a more general form that provides
flexibility in the model. We present that feasible version in the next subsection
followed by simulated example.

3.1 Regression Functions Definitions

As presented earlier, best-worst DCEs take into account we provided the design,
probabilities, and properties associated with attribute-level best-worst pairs. The
data in these experiments are defined as a series of 1 s, 0 s, and −1 s corresponding
to the best and worst attributes and attribute-levels in a given choice pair. There exist
a set of functions f and g defined on the attribute-level pair that produces traditional
methods.
In the attribute-level best-worst DCEs, a set of G profiles, or products, are
examined. The profiles are given as xi = (xi1 , xi2 , . . . , xiK ), where xij is the
attribute-level in profile i = 1, 2, . . . , G that corresponds to the j th attribute for
j = 1, 2, . . . , K. Let us consider the choice is given as (xij , xij  ), where j = j  ,
j, j  = 1, 2, . . . , K, and i = 1, 2, . . . , G. Let f be the set of regression functions
defined on the best attribute-level in a pair and g be the set of regression functions
defined on the worst attribute-level in a pair.
In the traditional attribute-level best-worst DCE, the regression functions f and g
are defined as indicator functions. The indicator functions are p × 1 vectors. For the
attributes, they are defined as,

1, if xij ∈ Ak ,
IAk (xij ) = (3.3)
0, otherwise,

and for the attribute-levels,


1, if xij = xk for xk ∈ Ak ,
IAk xk (xij ) = (3.4)
0, otherwise.

where j, k = 1, 2, . . . , K and i = 1, 2, . . . , G.
The best-worst systematic component for the pair (xij , xij  ) is given as,

Vijj  = Vij − Vij 

= (f(xij ) − g(xij  )) β
⎡ ⎤

K

lk
= ⎣IAk (xij )βAk + IAk xkj (xij )βAk xkj ⎦
k=1 j =1
12 A. Working et al.

⎡ ⎤

lk
− ⎣IAk (xij  )βAk + IAk xkj  (xij  )βAk xkj  ⎦
k=1 j  =1

K 

= IAk (xij )βAk − IAk (xij  )βAk

k=1

lk

lk 
+ IAk xkj (xij )βAk xkj − IAk xkj (xij  )βAk xkj 
j =1 j  =1

= IAj βAj − IAj  βAj  + IAj xij βAj xij − IAj  xij  βAj  xij  , (3.5)

where j, j  = 1, 2, . . . , K, j = j  , and i = 1, 2, . . . , G.
Rewriting the indicator functions of the Ak and Ak xk , a more general form of the
regression functions can be defined. Let bAk and bAk xk be constants corresponding
to the best attribute and attribute-levels in a pair, and wAk and wAk xk be constants
corresponding to the worst attribute and attribute-levels in a pair, where xk =
1, 2, . . . , lk and k = 1, 2, . . . , K. The regression functions f and g are given as,
⎡ ⎤

lk
f(xij ) = ⎣bAk IAk (xij ) + bAkxk IAk xk (xij )⎦ (3.6)
k=1 j =1

and
⎡ ⎤

lk
g(xij  ) = − ⎣wAk IAk (xij  ) + wAkxk IAk xk (xij  )⎦ (3.7)
k=1 j =1

where j, j  = 1, 2, . . . , K, j = j  , and i = 1, 2, . . . , G.
The above functions are simple linear process which can be used to model
the attribute-level best-worst DCEs. Furthermore, the dependence, or functional
dependence, can be extended by considering
⎡ ⎤

lk
f(xij ) = ⎣fk (Ak ) + fk,j (Ak xkj )⎦ . (3.8)
k=1 j =1

and
⎡ ⎤

lk
g(xij ) = − ⎣gk (Ak ) + gk,j (Ak xkj )⎦ . (3.9)
k=1 j =1
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 13

where fk , gk , fk,j , and gk,j can be linear, nonlinear, or kernel based functional form
for best and worst attributes and attributes-levels, respectively.
The regression functions defined in this way provide flexibility than the tradi-
tional attribute-level best-worst DCEs. Consumer preference in products is con-
stantly changing, new information about the product comes to light or as trends
come and go. Hence, the data collected on a product may be dynamic. The addition
of these constants to the regression functions provides researchers the ability to
scale the data to reflect current trends or changes in the products. For example,
let us consider the products being modeled are pharmaceuticals such as in the
contraceptives as proposed in Knox et al. (2012, 2013). If new information about
a brand of contraceptives posing a health risk was discovered, then using regression
functions, it is possible to update the model to reflect this change. Assuming the
change is to remove the brand. The attribute-level associated with the brand may
have bkxk = wkxk = 0, where xk = 1, 2, . . . , lk and k = 1, 2, . . . , K to represent its
removal from the market. For all the pairs this attribute-level was in, the information
the choice pair provides in terms of the other attributes and attribute-levels would
remain intact. The model would be estimated again and the parameter vector,
β, would provide the updated impact of the attributes and attribute-levels in the
experiment.

3.2 Data Example

In the simulated example an empirical setup is considered. We assume K = 3

attributes with l1 = 2, l2 = 3, and l3 = 4 attribute-levels in an unbalanced design.
There are 2 × 3 × 4 = 24 possible  profiles, or products, in this experiment. The total
number of attribute-levels is L = ki=1 li = 9, and the total number of choice pairs

is J = K k=1 lk (L − lk ) = 52.
We simulated data for n = 300 respondents for 24 profiles. Each choice set has
τ = K(K − 1) = 6 alternatives to choose from. Using the parameters given in
Table 1, we simulated data in R. The data was then exported from R into the SAS®
environment. Using the SAS® procedure called MDC (multinomial discrete choice),
the conditional logit model was fitted to the data. The parameter estimates for the
generated data are given in Table 1. The parameter estimates are close to the original
parameters for this example.
We consider an example where the model is built on the regression functions f
and g of the data. We define f and g as given in Eqs. (3.6) and (3.7). The weights
used in the regression functions are given as: bA1 = wA1 = −2, bA2 = wA2 = 5,
bA3 = wA3 = 1, bA1 1 = wA1 1 = bA1 2 = wA1 2 = −2, bA2 1 = wA2 1 = bA2 2 =
wA2 2 = bA2 3 = wA2 3 = 5 and bA3 1 = wA3 1 = bA3 2 = wA3 2 = bA3 3 = wA3 3 =
bA3 4 = wA3 4 = 1.
The conditional logit model is fit to the data and the resulting parameter estimates
are given in Table 1. The parameter estimates provide the adjusted attribute and
14 A. Working et al.

Table 1 Parameters and parameter estimates for simulated example

Estimates Functional form
Parameters β β̂ SE β̂ SE
βA1 −2.0000 −2.0711 0.0621 0.9787 0.0289
βA2 1.5000 1.5248 0.0438 0.3042 0.0082
βA3 * * * * *
βA1 1 −2.0000 −2.0308 0.0619 0.9838 0.0288
βA1 2 2.0000 2.0308 * −0.9838 *
βA2 1 1.9900 2.0970 0.0804 0.3864 0.0148
βA2 2 −0.2900 −0.3567 0.0482 −0.0548 0.0092
βA2 3 −1.7000 −1.7403 * −0.3316 *
βA3 1 −0.9200 −0.8914 0.0407 −0.8867 0.0410
βA3 2 −0.1800 −0.1805 0.0368 −0.1806 0.0368
βA3 3 0.5000 0.4911 0.0369 0.4966 0.0366
βA3 4 0.6000 0.5808 * 0.5707 *

attribute-level impacts. In Sect. 5, we will evaluate the changes in expected utility

for this weighted data in comparison to the original data and model.
We can see the impact of weighting as a reciprocal change in the impact of
attribute 1 is noticed as its value goes from −2 to 0.9787. We will use these
functional forms and included time in them in Sect. 5 in the simulated example with
two scenarios.
We utilize the new definition of the systematic components in the modeling
of attribute-level best-worst DCEs across time. In Sect. 4, we extend the work
done here to Markov decision processes. The generalized form of the systematic
components we provided allows for the evaluation of hypothetical future scenarios.

4 Markov Decision Process

Markov decision processes (MDPs) are sequential decision-making processes. A

decision process is said to be Markovian if the future depends on the present and
not the past. In that sense, a Markov process is a memoryless practice. MDPs seek
to determine the policy, or set of decision rules, under which maximum reward over
time is obtained. According to Puterman (2014), decision processes are defined by
the set (S, R, D), where S is the set of states, R is the set of rewards, and D is the
set of possible decisions for each time step. Let st ∈ S be the states occupied at
time t, rt (st ) be the rewards associated with st , and dt (rt , st ) is the decision based
on possible rewards and states at time t, where 0 ≤ t ≤ T . The rewards are defined
as the expected gain, or loss, associated with the state. With regard to DCEs, the
states are the choice pairs and the rewards are the utility associated with the choice
in alternative.
The definition of time is important in the methods for mapping the decision
processes. These processes may be discrete or continuous in time with finite or
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 15

infinite horizon. For the purpose of this chapter, our interest is in discrete time,
finite horizon MDPs, that is t = 1, 2, . . . , T where T < ∞. Numerical methods
such as dynamic programming are used to estimate the expected rewards for this
type of MDPs.
As the decision process is Markovian, the transition probability to the next state
st+1 based solely on the decision made at the current state, st , is p(st+1 |st ), where
t = 1, 2, . . . , T (Puterman 2014). The transition probabilities are the drivers of this
sequential decision-making process. The decision process maps the movement from
one state to another over time, t, based on rewards received and the optimal decision
set. The optimal decision rule is known as the policy, δ = (d1∗ , d2∗ , . . . , dT∗ ), where
dt∗ is the decision at time t = 1, 2, . . . , T that yields the maximum expected reward
(Puterman 2014).
While there exists some literature on the application of MDPs in traditional
DCEs, we have not encountered any work in the literature to extend these methods
to best-worst scaling DCEs. In this chapter, we extend the use of MDPs to Case 2
of best-worst scaling models, the attribute-level best-worst DCEs.
In traditional MDPs, the value functions are computed for each of the J
alternatives, or products. At each time point, t = 1, 2, . . . , T , the decision dt is to
choose the alternative that provides the maximum expected utility given information
about the state st = (xt , t ), where xt is the set of K attributes. The decision made
is between alternatives in the traditional DCEs. In attribute-level best-worst DCEs,
the experiments model choices within products not between products.
In attribute-level best-worst DCEs, there are K attributes describing a product
each with lk levels, where k = 1, 2, . . . , K. The total number of products in these

K
experiments is lk . The products are represented in the experiment by a profile.
k=1
The profile corresponding to the ith product is given as xi = (xi1 , xi2 , . . . , xiK ),
where xik is the attribute-level corresponding to the attribute Ak for k = 1, 2, . . . , K
for i = 1, 2, . . . , G. Within each choice set there are τ = K(K − 1) choices. A
respondent is asked to evaluate G choice sets in the experiment.
MDPs model the decision process for respondents over multiple time points.
For attribute-level best-worst DCEs, the model is built within the choice sets
corresponding to each of the G choice sets. In traditional DCEs, there are J
alternatives evaluated at each time point producing J value functions at each time
point. Attribute-level best-worst DCEs require a respondents to evaluate a series
of G choice sets each with τ choices, thus there are τ value functions for each
choice set in attribute-level best-worst MDPs. Our interest is to further model the
sequence of decisions made by introducing the time element into the experiments.
For attribute-level best-worst DCEs, we consider discrete time finite horizon MDPs
where:
• G choice sets are modeled across time.
• xtijj  = (xij , xij  ) are the attributes and attribute-levels corresponding to the
choices in set Ci , i = 1, 2, . . . , G, j = j  , j, j  = 1, 2, . . . , K, and t =
1, 2, . . . , T .
16 A. Working et al.

• The decision set depends on the choice set, called Di , and we evaluate dit ∈ Di ,
where i = 1, 2, . . . , G and 1 ≤ dti ≤ τ .
• The set of possible states in the experiment depends on the choice set, called Si ,
where sit = (xijt , xijt  ) ∈ Si , where j = j  , j, j  = 1, 2, . . . , τ , i = 1, 2, . . . , G,
and 1 ≤ sti ≤ τ .
• Transition probabilities depend on a set of parameters θ that are assumed known,
or estimated from data (Arcidiacono and Ellickson 2011).
• Transition probability matrices, Pst s  , are dependent on the choice set being
i i
evaluated.
In attribute-level best-worst DCEs, the MDPs model the choices in attribute-level
pairs within choice sets over time. Therefore, the transition probabilities and value
functions must be defined within the choice sets. Bellman (1954) utilized dynamic
programming to evaluate the value function, also known as Bellman’s equation, at
each time step. Rust (1994, 2008) presented the use of dynamic programming for
evaluating DCEs as MDPs.
The value function for DCEs defined by Bellman’s equation is given as:

T

  
Vt (xt , t ) = max Psst  γ t −t U (xt  , dt  ) + (dt  )|xt
dt ∈D
t  =t
 

T
t  −t
= max E γ U (xt  , dt  ) + (dt  )|xt , t , (4.1)
dt ∈D
t  =t

where dt ∈ D is the decision at time t, U (xt , dt ) is the derived iterated/expected

utility, t is the associated error term at time t, where t = 1, 2, . . . , T , and discount
utility rate is given by γ ∈ (0, 1) (Bellman 1954).
The value functions are computed recursively via dynamic programming. To
determine the value function, backwards recursion must be used. At the last time
point, T , the value function is the utilities associated with the different states.
Adopting for the attribute-level best-worst DCEs eventually the value function is
given by,


Vit (sit , dit ) = U (sit , dit ) + i
γ Vt+1 (si t+1 , dit+1 )Pst s  , (4.2)
i i

si t+1 ∈Si

where t = 1, 2, . . . , T , U (sit , dit ) represents the utility associated with the state sit
and decision dit , and discount utility rate is given by γ ∈ (0, 1) and i = 1, 2, . . . , G.
The decision dit = (xij , xij  ) is a choice pair within Ci , where i = 1, 2, . . . , G,
j, j  = 1, 2, . . . , K, and j = j  . In the attribute-level best-worst DCEs, there
will be τ = K(K − 1) value functions per each of the G choice sets. One of the
disadvantages of these experiments is the “curse of dimensionality” (Rust 2008). As
the number of attributes, attribute-levels, and profiles grow in the experiment, the
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 17

estimation process becomes exponentially more difficult as dynamic programming

requires an explicit discretization of the states, decisions, and the value function as
seen in Eq. (4.2) depends on the utility and transition probabilities over time. The
ability to direct the system, via the transition probabilities, when it is of a higher
dimension becomes difficult, if not impossible.
In the following subsections, we provide definitions and insights regarding these
components to the value function. In Sect. 3, we provided a functional form of the
utility that we can apply in these time dependent processes. Furthermore, we define
dynamic transition probabilities that we apply to attribute-level best-worst DCEs. In
Sect. 5, simulations of MDPs for the attribute-level best-worst DCEs are provided.

4.1 Utility

Marschak (1960) presented random utility theory defining utility to include a

systematic component Vij and an unobserved component ij for i = 1, 2, . . . , n
and j = 1, 2, . . . , J . A consumer chooses the alternative that provides them with
the maximum utility. The utility function for traditional DCEs is given as:

Uij = Vij + ij = xij β + ij ,

for i = 1, 2, . . . , n and j = 1, 2, . . . , J . For MDPs in the traditional DCEs, the

utility is then,

Ut (xt , dt ) = xt β + t , (4.3)

where t = 1, 2, . . . , T . Common models to determine the parameter estimates of β

are conditional logit, generalized extreme value distributions, and probit models. We
consider the conditional logit model in this chapter. Stenberg et al. (2007) provided
that the definition of utility/reward in MDPs maybe constant over time or time
dependent/dynamic in nature.
The definition of utility in attribute-level best-worst DCEs that meets the
necessary independence from irrelative alternative (IIA) condition (Luce 1959) for
the conditional logit model is given in Eq. (2.2) under the inverse random utility
theory presented in Marley and Louviere (2005). For the corresponding choice pair
 = (xij , xij  ) ∈ Ci the corresponding utility is given as,
t
xijj t t t

 = Vij − Vij  + ijj  = Vijj  + ijj  ,

t t t t t t
Uijj

where i = 1, 2, . . . , G, j, j  = 1, 2, . . . , K, and j = j  .
Referring back to Sect. 3, the systematic component is defined as a model built
on functions of the best and worst attribute-levels in the pair, using Eq. (3.1),

Vijj  = (ft (xij ) − gt (xij  )) β, (4.4)

18 A. Working et al.

where β are the attribute and attribute-level coefficients, j, j  = 1, 2, . . . , K,

j = j  , i = 1, 2, . . . , G, and ft and gt , t = 1, 2, . . . , T , are regression functions
defined on the best and worst attributes and attribute-levels, respectively. In the
traditional attribute-level best-worst DCEs, the functions ft and gt are given as
indicator functions of the best and worst attributes and attribute-levels as shown
in Eq. (3.5).
However, using this functional form of the systematic component, we may
consider alternative definitions of the systematic component. In Sect. 3, we provide
a weighted function for ft and gt , given in Eqs. (3.6) and (3.7). Let bA t and bt
k Ak xk
be weights corresponding to the best attribute and attribute-levels in a pair, and wA t
k
and wA t be weights corresponding to the worst attribute and attribute-levels in a
k xk
pair, where xk = 1, 2, . . . , lk , k = 1, 2, . . . , K, and t = 1, 2, . . . , T . The regression
functions ft and gt are given as,
⎡ ⎤

lk
ft (xij ) = ⎣bA
t
I (xij ) + t
bA I (xij )⎦ , (4.5)
k Ak k xk Ak xk
k=1 j =1

and
⎡ ⎤

lk
gt (xij  ) = − ⎣wA
t
I (xij  ) + t
wA I (xij  )⎦ , (4.6)
k Ak k xk Ak xk
k=1 j =1

where j, j  = 1, 2, . . . , K, j = j  , and i = 1, 2, . . . , G.
Defining the systematic components according to the weighted function allows
the utility to change over time. We considered in Sect. 3 an example where an
attribute-level no longer exists in the future. The weighted functions of ft and gt
allowed us to update the parameter estimates, thus the utilities, using these weights.
It is conceivable in the future that an attribute-level scale may need to be adjusted for
possible bettering, worsening, or removal type of conditions for that attribute-level.

4.2 Transition Probabilities

MDPs have infinitely many possible futures able to be considered in the simulations.
The definition of transition probabilities is the vehicle that drives the processes to
these different futures. However, determining transition probabilities for MDPs is a
difficult task. One way for estimating the transition probabilities is using maximum
likelihood estimates (MLEs). An empirical solution to the transition probabilities
may be determined by considering the transition probabilities as a multinomial
distribution (Lee et al. 1968).
In the attribute-level best-worst DCEs, there are τ choices within a choice set Ci ,
where i = 1, 2, . . . , G. There are τ states, and/or decisions, possible at each of the
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 19

time points. The transition probabilities are denoted as Pss  = P (st+1 = s  |st = s),
where st , st+1 ∈ S and S = {1, 2, . . . , τ }. Let Ni be the respondents common to
time t and t + 1 in the experiment and niss  be the number of respondents who chose
s at time t and s  at t + 1, where t = 1, 2, . . . , T and i = 1, 2, . . . , G. The transition
choice probability is given by the multinomial distribution as:

Ni ! n n nisτ
f (pis1 , pis2 , . . . , pisτ ) = p is1 p is2 . . . pisτ , (4.7)
nis1 !nis2 ! . . . nisτ ! is1 is2

and log-likelihood is given as

⎛ ⎞
⎜ N! ⎟
τ
⎜ i ⎟
log(L) = log ⎜ τ ⎟ + niss  log(piss  ),
⎝  ⎠
niss  ! 
s =1
s  =1


τ
where s = 1, 2, . . . , τ , i = 1, 2, . . . , G, piss  ≥ 0, and piss  = 1.
s  =1

τ
Due to the constraint piss  = 1, Lagrange multipliers, λ, are used and the
s  =1
Lagrangian function is given as:

τ
G(pss  ) = LL(piss  ) − λ( piss  − 1),
s  =1


τ
where s = 1, 2, . . . , τ , i = 1, 2, . . . , G, piss  ≥ 0, and piss  = 1.
s  =1
Taking the partial derivative of the Lagrangian to determine the MLEs gives us

τ
= piss  where s  = 1, 2, . . . , τ . Under the constraint,
niss 
λ piss  = 1, the value
s  =1

τ
for s, s  = 1, 2, . . . , τ and
niss 
of λ = niss  = Ni . Thus, the MLE for piss  = Ni
s  =1
i = 1, 2, . . . , G.
The MLE of piss  is computationally simple; however, access to the information
needed to compute it may not always be available. To compute the MLE of this
nature, we would need to have respondents evaluate the same choice sets at two time
periods, which is not necessarily an easy task. Furthermore, this is considering the
transition matrix is stationary. It is possible to consider a dynamic transition matrix
 for t = 1, 2, . . . , T . A transition matrix of this
that changes over time, that is pisst

nature would need to have multiple time periods of data for the same respondents
evaluating the same choice sets to compute the empirical probabilities. Instances
where multiple time periods of data for respondents are not possible, one must
consider alternative methods for determining the transition probabilities.
20 A. Working et al.

There are infinitely many possible transition probabilities in MDPs. Common

methods for determining these probabilities is to take a Bayesian approach and the
other is a rational observation according to Rust (2008). Arcidiacono and Ellickson
(2011) indicates that the transition probabilities, Pss  = P (st+1 = s  |st = s, θ )
are a probability function, where the parameters θ are assumed known. Rust (1994,
2008) state that discrete decision processes, as we are considering in the attribute-
level best-worst models, the transition parameters and probabilities are often times
non-parametrically identified. Chadès et al. (2014) applied MDPs to solve problems
in an ecological setting. As they mentioned, to suggest guidance in transition
probabilities would require running several scenarios. To the best of our knowledge,
such technique has not yet been applied to consumer choice experiments with
attribute and attribute-level best-worst experiments.
We provide a definition of the parameters for the transition probabilities under
the rational observation that may be used in stationary or dynamic transition
matrices. This method maintains the researcher’s ability to guide the MDPs in the
direction of their choosing where the transitions occur at a rate determined by the
researcher. In such a way, the researcher is able to consider stationary or dynamic
transition probabilities to model an evolving MDP over time. The researcher may
also determine the amount of time points necessary for the system to converge to
the decision they were working towards.
In attribute-level best-worst choice models, a set of G choice sets are considered
in the experiment. In MDPs, there exists a set of states st ∈ S and possible decisions
in dt ∈ D for t = 1, 2, . . . , T . For attribute-level best-worst MDPs, the possible
states in each choice set are the alternatives, and the decision made at each time
point will also be one of the alternatives. For choice set Ci the state sti and decision
dti are such that 1 ≤ sti , dti ≤ τ where i = 1, 2, . . . , G and t = 1, 2, . . . , T .
Each state s = sti gives rise to a new set of states s  at time t + 1 with transition
t 
 = P (s |s, θ s ), where
t t
probability denoted as Piss

θ ts = (θsA
t
1
t
, θsA2
t
, . . . , θsAK
t
, θsA11
t
, . . . , θsAK lk
)

where sAi indicates involvement of attribute i, and sAi lk indicates involvement of

attribute-level k in attribute i where i = 1, 2, . . . , K and k = 1, 2, . . . , lk .
The parameter estimates determined by fitting the conditional logit model, as

K
described in Sect. 3, produced β̂ a p = K + lk vector. These parameter estimates
k=1
measure the relative impact of each attribute and attribute-level in the decisions
made by respondents. The parameters θ ts are the assumed impacts of the attributes
and attribute-levels in respondents decisions given they currently occupy state s. We
define these parameters as functions of the parameter estimates β̂, where there is a
rate of change in the impacts over time. We define

θ̂ ts = (asA1 (t)β̂A1 , . . . , asAK (t)β̂Ak ,

asA1 1 (t)β̂A1 1 , . . . , asAK lk (t)β̂AK lk ),
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 21

where a  s are the time factor change and β  s are fixed for i = 1, 2, . . . , G, 1 ≤ s ≤
τ , and t = 1, 2, . . . , T . The definition of the as (t) depends on the state s and time
t = 1, 2, . . . , T . We have considered asj (t) = asj t , where if |a | < 1 the impact of
sj
the attribute or attribute-level would be lessening with time, where j = 1, 2, . . . , K.
If asj (t)β̂j = asj t β̂ > 0, then the attribute or attribute-level has a positive impact
j
evolving at the rate asj t over time for j = 1, 2, . . . , K and t = 1, 2, . . . , T . A static,

or non-time dependent, system is considered if asj (t) = 1, where j = 1, 2, . . . , K

and t = 1, 2, . . . , T .
As mentioned, these asj (t) are rates of change that guide the dynamic transition
of the decision process. We can easily consider them to be non-time dependent,
asj (t) = asj , defining the transition probabilities as stationary over time. As was
mentioned earlier, there are infinitely many possibilities in how we define the
transitions. Rust (2008) states that using rational observation to define the transitions
of any possible choice behavior on the respondents is possible. Chadès et al. (2014)
recommend running many scenarios to determine the transition probabilities that
will maximize the expected reward. Our definition also offers infinitely many
possibilities in terms of the definition; however, we defined a rate of change to
consider an evolving system. In this way, the researcher can determine what they
consider feasible rates and see if the system eventually evolves to the decision they
desire and how long it would take to get there.
Now the transition probability is given as,
 
P t (sijj  |si , θ si ) = P (Uijj  > Uikk  , ∀k = k ∈ Ci |si , θ si )
t t t t


= P t (Vijj
t
 + ijj  > Vikk  + ikk  , ∀k = k ∈ Ci |si , θ si )
t t t t


= P t (ikk
t
 < ijj  + Vijj  − Vikk  , ∀k = k ∈ Ci |si , θ si ),
t t t t
(4.8)

where j = j  , j, j  = 1, 2, . . . , τ , i = 1, 2, . . . , G, and t = 1, 2, . . . , T . If we
assume the random error terms are independently and identically distributed as
type I extreme value distribution, the probability would then be found using the
conditional logit, and is given as:
 
P t (sijj  |si , θ si ) = P (Uijj  > Uikk  , ∀k = k ∈ Ci |si , θ si )
t t t t t

t )
exp(Vijj 
=  t )
exp(Vikk 
k,k  ∈Ci

exp((ft (xij ) − gt (xij  )) θ tsi )

=  , (4.9)
exp((ft (xik ) − gt (xik  )) θ tsi )
k,k  ∈C i

where j = j  , k = k  , j, j  = 1, 2, . . . , τ , i = 1, 2, . . . , G, and t = 1, 2, . . . , T .
22 A. Working et al.

The transition matrix is then a τ × τ matrix of the form,

⎛ ⎞
t Pt
Pi11 t
. . . Pi1τ
i12
⎜Pt Pt t ⎟
⎜ i21 i22 . . . Pi2τ ⎟
⎜ ⎟
⎜ . . ... . ⎟  t 
Pt = ⎜ ⎟ = Piss 
⎜ . . ... . ⎟
⎜ ⎟
⎝ . . ... . ⎠
Piτt 1 Piτt 2 . . . Piτt τ


τ
where i = 1, 2, . . . , G, s, s  = 1, 2, . . . , τ , and where t
Piss  = 1. The
s  =1
transition matrix may be either stationary or dynamic in nature. In our definition
of θ tsi , this is determined by the rate asi j (t), where i = 1, 2, . . . , G, 1 ≤ j ≤ p, and
t = 1, 2, . . . , T . In Sect. 5, we provide simulations under stationary and dynamic
transition probabilities and make comparisons.

5 Simulation Example

In the simulated example, an empirical setup is considered. We assume K = 3

attributes with l1 = 2, l2 = 3, and l3 = 4 attribute-levels in an unbalanced design.
There are 2 × 3 × 4 = 24 possible  profiles, or products, in this experiment. The total
number of attribute-levels is L = ki=1 li = 9, and the total number of choice pairs

is J = K k=1 lk (L − lk ) = 52.
Louviere and Woodworth (1983), Street and Knox (2012), and Grasshoff et al.
(2004) discussed the benefits in using orthogonal arrays. Generally, orthogonal
experimental designs are utilized in attribute-level best-worst DCEs due to the
large number of profiles in a full factorial design. There is a package in R called
DoE.design that creates full factorial and orthogonal designs for a given set of
attributes and attribute-levels. To obtain an orthogonal design, the oa.design function
is used. For this experiment, the orthogonal design returned the full factorial design,
so we used the full set of 24 profiles when simulating this data.
We simulated data for n = 300 respondents for 24 profiles. Each choice set
has τ = K(K − 1) = 6 choices to choose from. Using the parameters given in
Table 1, we simulated data in R. The data was then exported from R into the SAS®
environment. Using the SAS® procedure called MDC (multinomial discrete choice),
the conditional logit model was fitted to the data. The parameter estimates for the
generated data are given in Table 1. The parameter estimates are close to the original
parameters for this example. Using the parameter estimates, the choice utilities were
computed and are used to determine the expected utility/value function. The best
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 23

Table 2 Choice pairs with Best attribute Level Worst attribute Level Utility
the highest utility in the
experiment 2 1 1 1 12.3633
2 2 1 1 8.8012
3 4 1 1 7.6931

Table 3 Choice pairs with Best attribute Level Worst attribute Level Utility
the lowest utility in the
experiment 1 1 2 1 −9.2594
1 1 2 2 −6.5358
1 1 3 4 −5.7929

and worst 3 choice pairs along with their utilities are presented in Tables 2 and
3, respectively. The opposite of the pairs with the highest utilities have the lowest
utilities.

5.1 Scenario 1

We ran the simulation under this scenario with an advantageous proposed structure.
The intent is to validate/justify our relative performance over time under stationary
sparsity.
In this example, respondents are assumed to make similar decisions at each
decision epoch that they made at the previous time point. The transition parameters
θ tsi where sit = (xij , xij  ) are defined as for the attributes as,
⎧
⎨ 1.7|βAk |, if xij ∈ Ak ,
θsti Ak = −1.7|βAk |, if xij  ∈ Ak , (5.1)
⎩
βAk , otherwise,

and for the attribute-levels,

⎧
⎨ 1.7|βAk xik |, if xij = xik where xik ∈ Ak ,
θsti Ak xik = −1.7|βAk xik |, if xij  = xik where xik ∈ Ak , (5.2)
⎩
βAk xik , otherwise,

where j = j  , j, j  , k = 1, 2, . . . , K, 1 ≤ xk ≤ lk , and i = 1, 2, . . . , G. The

transition parameters do not change with time, so the transition matrix is stationary.
The goal of this scenario was to design the transition probabilities in a way that the
choice made at t is most likely to be made at t + 1. If we considered asi m (t) = βm
for i = 1, 2, . . . , G, and m = 1, 2, . . . , p, then the system would remain static and
24 A. Working et al.


K
every row of the transition matrix would be the same. Recall that p = K + lk =
k=1
12 is the number of parameters. We consider 1.7|βm | when a state or choice pair
at time t + 1 has the same best attribute and attribute-level as the state occupied
at time t, and −1.7|βm | when a state or choice pair at time t + 1 has the same
worst attribute and attribute-level as the state occupied at time t. We consider |βm | to
control the direction of the impact making sure it is positive for the best attribute and
attribute-level of si and use −|βm | to make sure its negative for the worst attribute
and attribute-level of si . We use 1.7 to increase the impact of the best and worst
attributes and attribute-levels of si . The definition of asi m (t) in this way insures that
states with common best and worst attributes and attribute-levels as the present state
occupied, sit = (xij , xij  ), have a greater probability of being transitioned to, where
i = 1, 2, . . . , G, j = j  , j, j  = 1, 2, . . . , K, and t = 1, 2, . . . , T . The weights
associated with the attributes and attribute-levels are selected as: bA1 = wA1 = −2,
bA2 = wA2 = 5, bA3 = wA3 = 1, bA1 1 = wA1 1 = bA1 2 = wA1 2 = −2, bA2 1 =
wA2 1 = bA2 2 = wA2 2 = bA2 3 = wA2 3 = 5 and bA3 1 = wA3 1 = bA3 2 = wA3 2 =
bA3 3 = wA3 3 = bA3 4 = wA3 4 = 1.
Referring back to Sect. 4, the systematic component as a function of the best and
worst attribute-level in the pair is as in Eq. (4.4),

Vijj  = (ft (xij  ) − gt (xij  )) β,

where ft and gt , as in Eqs. (4.5) and (4.6) with profile choice pairs shown in Fig. 2.
The value function/expected utilities for Profile 1 are displayed in Fig. 3, with
legend displayed in Fig. 2 along with the difference in the value functions over
time. Choice pair (x22 , x12 ), where x22 is the 2nd level of attribute 2 is the best and
x12 is the 2nd level of attribute 1 is the worst, is the choice with the highest expected
utility. The opposite pair (x12 , x22 ) is the worst choice pair. The pair (x34 , x22 ) has
a sharp drop between time t = 3 and t = 4 because of the change in the weights
applied to the attributes and attribute-levels from Eq. (3.1).
The model applied here views the attribute-level best-worst DCEs as sequential
leading to partial separation best-worst choices over time. Validity is guided by the
transition probabilities under Scenario 1, the participants follow the same choice
preferences. In Table 4, the transition probabilities are generally highest on the

Fig. 2 Legend corresponding

to Fig. 3
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 25

Fig. 3 Expected discounted utility and their differences over time for Profile 1
26 A. Working et al.

Table 4 Stationary transition matrix in Scenario 1 for Profile 1

(x12 , x22 ) 0.9837 0.0000 0.0136 0.0000 0.0000 0.0027
(x22 , x12 ) 0.0000 0.8924 0.0000 0.1074 0.0003 0.0000
(x12 , x34 ) 0.0038 0.0000 0.9932 0.0000 0.0030 0.0000
(x34 , x12 ) 0.0000 0.0038 0.0000 0.9613 0.0000 0.0003
(x22 , x34 ) 0.0000 0.4289 0.0004 0.0002 0.5705 0.0000
(x34 , x22 ) 0.0001 0.0004 0.0000 0.7113 0.0000 0.2882

diagonal and the same at each time period as we would expect in this setup. As
expected the trend in the utility is kept.

5.2 Scenario 2

In Scenario 2, respondents are allowed to make similar decisions at each time epoch
with a different rate of change, making the transition probabilities dynamic. The
transition parameters θ tsi where sit = (xij , xij  ) are defined as for the attributes as,
⎧ t
⎨ 1.7 |βAk |, if xij ∈ Ak ,
θsti Ak = −1.7t |βAk |, if xij  ∈ Ak , (5.3)
⎩
βAk , otherwise,

and for the attribute-levels,

⎧ t
⎨ 1.7 |βAk xik |, if xij = xik where xik ∈ Ak ,
θsti Ak xik = −1.7t |βAk xik |, if xij  = xik where xik ∈ Ak , (5.4)
⎩
βAk xik , otherwise,

where j = j  , j, j  , k = 1, 2, . . . , K, 1 ≤ xk ≤ lk , and i = 1, 2, . . . , G.
We ran the simulation under this scenario with advantageous proposed hybrid
structure as shown above using the functional form as described in Scenario 1 with
profile choice pairs shown in Fig. 4. The transition matrix at time t = 1 is kept
the same as it was Scenario 1 in Table 4, and subsequent transition probabilities at
time t = 2, 3, and 4 are given in Tables 5, 6, and 7, respectively. The transition
probabilities are highest on the diagonal verifying the direction we wanted in the
transitions. The value function/expected utilities for Profile 1 are displayed in Fig. 5,
with legend displayed in Fig. 4 along with the difference in value functions. Choice
pair (x22 , x12 ), where x22 is the 2nd level of attribute 2 is the best and x12 is the 2nd
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 27

Fig. 4 Legend corresponding

to Fig. 5

Table 5 Dynamic transition matrix in Scenario 2 at time t = 2 for Profile 1

(x12 , x22 ) 0.9985 0.0000 0.0015 0.0000 0.0000 0.0000
(x22 , x12 ) 0.0000 0.9873 0.0000 0.0127 0.0000 0.0000
(x12 , x34 ) 0.0002 0.0000 0.9998 0.0000 0.0000 0.0000
(x34 , x12 ) 0.0000 0.0019 0.0000 0.9981 0.0000 0.0000
(x22 , x34 ) 0.0000 0.0337 0.0001 0.0000 0.9663 0.0000
(x34 , x22 ) 0.0000 0.0000 0.0000 0.2082 0.0000 0.7918

Table 6 Dynamic transition matrix in Scenario 2 at time t = 3 for Profile 1

(x12 , x22 ) 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
(x22 , x12 ) 0.0000 0.9997 0.0000 0.0003 0.0000 0.0000
(x12 , x34 ) 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
(x34 , x12 ) 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000
(x22 , x34 ) 0.0000 0.0002 0.0000 0.0000 0.9998 0.0000
(x34 , x22 ) 0.0000 0.0000 0.0000 0.0058 0.0000 0.9942

Table 7 Dynamic transition matrix in Scenario 2 at time t = 4 for Profile 1

(x12 , x22 ) 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000
(x22 , x12 ) 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000
(x12 , x34 ) 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
(x34 , x12 ) 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000
(x22 , x34 ) 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000
(x34 , x22 ) 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000

level of attribute 1 is the worst, still remains the choice with the highest expected
utility as in Scenario 1. The opposite pair (x12 , x22 ) is the worst choice pair. We
also notice more shifts in expected utility than in the previous scenarios for Profile
1. Scaling the data makes the utilities shift in much more extreme values.
28 A. Working et al.

Fig. 5 Expected discounted utility and their differences over time for Profile 1

6 Conclusion

The attribute-level best-worst DCMs considered in this chapter provide a gen-

eral insight for modeling complex dependencies of time evolving decisions. The
decisions for the large data set are guided by a functional form of the expected
Functional Form of Markovian Attribute-Level Best-Worst Discrete Choice Modeling 29

utility under identifiability constraints. Profile specific trends are displayed and
pattern behaviors are exhibited. We highlighted compelling situations that allow
shrinkage towards referenced choices and show efficiency data examples to make
inferences on the best-worst decisions of interest. Our simulated and aggregated data
examples show the flexibility and wide applications of our proposed techniques. Our
methodology is easily reproducible. The functional dependency and time evolving
structure may accommodate additional arrangements and setups.
A potential area of concern in the application of MDPs for attribute-level best-
worst DCEs is the curse of dimensionality as mentioned in Rust (2008). Since the
number of attributes, attribute-levels, and profiles grow quickly in the experiment,
the estimation process becomes exponentially more difficult. DCEs with larger
number of attributes and attribute-levels have more choice sets and pairs to model
across time. For discrete processes as is considered in the attribute-level best-worst
DCEs, the amount of information that needs to be stored becomes overwhelming.
The ability to guide the system becomes difficult due to the increased number of
states and choice sets considered. These issues should be considered when using
MDPs.
Extensions of this work may include interactions of choice pairs under different
correlation structures. The first order Markov dependency structure presented here
may be extended to higher order decision processes under stationary and dynamic
transition probabilities. Extensions to the continuous time scale case are being
explored.

References

Arcidiacono, P., Ellickson, P.B.: Practical methods for estimation of dynamic discrete choice
models. Annu. Rev. Econ. 3(1), 363–394 (2011)
Bellman, R.: The theory of dynamic programming. Bull. Am. Math. Soc. 60(6), 503–515 (1954)
Chadès, I., Chapron, G., Cros, M.-J., Garcia, F., Sabbadin, R.: Mdptoolbox: a multi-platform
toolbox to solve stochastic dynamic programming problems. Ecography 37(9), 916–920 (2014)
Coast, J., Salisbury, C., De Berker, D., Noble, A., Horrocks, S., Peters, T., Flynn, T.: Preferences
for aspects of a dermatology consultation. Br. J. Dermatol. 155(2), 387–392 (2006)
Finn, A., Louviere, J.J.: Determining the appropriate response to evidence of public concern: the
case of food safety. J. Public Policy Mark. 11(2), 12–25 (1992)
Flynn, T.N., Louviere, J.J., Peters, T.J., Coast, J.: Best–worst scaling: what it can do for health care
research and how to do it. J. Health Econ. 26(1), 171–189 (2007)
Flynn, T.N., Louviere, J.J., Marley, A.A., Coast, J., Peters, T.J.: Rescaling quality of life values
from discrete choice experiments for use as QALYs: a cautionary tale. Popul. Health Metr.
6(1), 6 (2008)
Grasshoff, U., Grossmann, H., Holling, H., Schwabe, R.: Optimal paired comparison designs for
first-order interactions. Statistics 37(5), 373–386 (2003)
Grasshoff, U., Grossmann, H., Holling, H., Schwabe, R.: Optimal designs for main effects in linear
paired comparison models. J. Stat. Plan. Inference 126(1), 361–376 (2004)
Grasshoff, U., Grossmann, H., Holling, H., Schwabe, R.: Optimal design for discrete choice
experiments. J. Stat. Plan. Inference 143(1), 167–175 (2013)
30 A. Working et al.

Grossmann, H., Grasshoff, U., Schwabe, R.: Approximate and exact optimal designs for paired
comparisons of partial profiles when there are two groups of factors. J. Stat. Plan. Inference
139(3), 1171–1179 (2009)
Hole, A.R.: A discrete choice model with endogenous attribute attendance. Econ. Lett. 110(3),
203–205 (2011)
Knox, S.A., Viney, R.C., Street, D.J., Haas, M.R., Fiebig, D.G., Weisberg, E., Bateson, D.: What’s
good and bad about contraceptive products? Pharmacoeconomics 30(12), 1187–1202 (2012)
Knox, S.A., Viney, R.C., Gu, Y., Hole, A.R., Fiebig, D.G., Street, D.J., Haas, M.R., Weisberg, E.,
Bateson, D.: The effect of adverse information and positive promotion on women’s preferences
for prescribed contraceptive products. Soc. Sci. Med. 83, 70–80 (2013)
Lancsar, E., Louviere, J., Donaldson, C., Currie, G., Burgess, L.: Best worst discrete choice
experiments in health: methods and an application. Soc. Sci. Med. 76, 74–82 (2013)
Lee, T.C., Judge, G., Zellner, A.: Maximum likelihood and Bayesian estimation of transition
probabilities. J. Am. Stat. Assoc. 63(324), 1162–1179 (1968)
Louviere, J., Timmermans, H.: Stated preference and choice models applied to recreation research:
a review. Leis. Sci. 12(1), 9–32 (1990)
Louviere, J.J., Woodworth, G.: Design and analysis of simulated consumer choice or allocation
experiments: an approach based on aggregate data. J. Mark. Res. 20(4), 350–367 (1983)
Louviere, J.J., Woodworth, G.G.: Best-worst scaling: a model for the largest difference judgments.
Working Paper. University of Alberta (1991)
Louviere, J.J., Hensher, D.A., Swait, J.D.: Stated Choice Methods: Analysis and Applications.
Cambridge University Press, Cambridge (2000)
Louviere, J., Lings, I., Islam, T., Gudergan, S., Flynn, T.: An introduction to the application of
(case 1) best–worst scaling in marketing research. Int. J. Res. Mark. 30(3), 292–303 (2013)
Louviere, J.J., Flynn, T.N., Marley, A.A.J.: Best-Worst Scaling: Theory, Methods and Applications.
Cambridge University Press, Cambridge (2015)
Luce, R.D.: On the possible psychophysical laws. Psychol. Rev. 66(2), 81 (1959)
Marley, A.A., Louviere, J.J.: Some probabilistic models of best, worst, and best–worst choices. J.
Math. Psychol. 49(6), 464–480 (2005)
Marley, A., Pihlens, D.: Models of best–worst choice and ranking among multiattribute options
(profiles). J. Math. Psychol. 56(1), 24–34 (2012)
Marley, A., Flynn, T.N., Louviere, J.: Probabilistic models of set-dependent and attribute-level
best–worst choice. J. Math. Psychol. 52(5), 281–296 (2008)
Marschak, J.: Binary choice constraints on random utility indicators. Technical Report 74 (1960)
McFadden, D.: Conditional logit analysis of qualitative choice behavior. In: Zarembka, P. (ed.)
Frontiers in Econometrics, pp. 105–142. Academic Press, New York (1974)
McFadden, D.: Modeling the choice of residential location. Transp. Res. Rec. 673, 72–77 (1978)
Parvin, S., Wang, P., Uddin, J.: Using best-worst scaling method to examine consumers’ value
preferences: a multidimensional perspective. Cogent Bus. Manag. 3(1), 1199110 (2016)
Puterman, M.L.: Markov Decision Processes: Discrete Stochastic Dynamic Programming. Wiley,
New York (2014)
Rust, J.: Structural estimation of Markov decision processes. Handb. Econ. 4, 3081–3143 (1994)
Rust, J.: Dynamic programming. In: Durlauf, S.N., Blume, L.E. (eds.) The New Palgrave
Dictionary of Economics. Palgrave Macmillan, Ltd, London (2008)
Stenberg, F., Manca, R., Silvestrov, D.: An algorithmic approach to discrete time non-
homogeneous backward semi-Markov reward processes with an application to disability
insurance. Methodol. Comput. Appl. Probab. 9(4), 497–519 (2007)
Street, D.J., Burgess, L.: The Construction of Optimal Stated Choice Experiments: Theory and
Methods, vol. 647. Wiley, New York (2007)
Street, D.J., Knox, S.A.: Designing for attribute-level best–worst choice experiments. J. Stat.
Theory Pract. 6(2):363–375 (2012)
Thurstone, L.L.: A law of comparative judgment. Psychol. Rev. 34(4), 273 (1927)
Van Der Pol, M., Currie, G., Kromm, S., Ryan, M.: Specification of the utility function in discrete
choice experiments. Value Health 17(2), 297–301 (2014)
Spatial and Spatio-Temporal Analysis of
Precipitation Data from South Carolina

Haigang Liu, David B. Hitchcock, and S. Zahra Samadi

1 Introduction

Spatial and spatio-temporal data are everywhere: we encounter them on TV, in

newspapers, on computer screens, on tablets, and on plain paper maps. As a
result, researchers in diverse areas are increasingly faced with the task of modeling
geographically referenced and temporally correlated data.
The geostatistical analysis of spatial data involves point-referenced data, where
Y (s) is a random vector at a location s ∈ Rr , where s varies continuously
over D, a fixed subset of Rr that contains an r-dimensional rectangle of positive
volume (Banerjee et al. 2014). The sample points are measurements of some
phenomenon such as precipitation measurements from meteorological stations or
elevation heights. The geostatistical analysis models a surface using the values from
the measured locations to predict values for each location in the landscape.
Spatial statistics methods have been frequently used in applied statistics as
well as water resources engineering. The work of Thiessen (1911) was the first
attempt in using interpolation methods in hydrology. Sharon (1972) used an average
of the observations from a number of rain gages to obtain estimates of the
areal rainfall. Soon after, Benzécri (1973), Delfiner and Delhomme (1975), and
Delhomme (1978) applied the various geostatistical methods such as variograms
and kriging methods in modeling rainfall. The work of Troutman (1983), Tabios

H. Liu · D. B. Hitchcock ()

Department of Statistics, University of South Carolina, Columbia, SC, USA
e-mail: haigang@email.sc.edu; hitchcock@stat.sc.edu
S. Z. Samadi
Department of Civil and Environmental Engineering, University of South Carolina, Columbia,
SC, USA
e-mail: samadi@cec.sc.edu

© Springer Nature Switzerland AG 2019 31

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_2
32 H. Liu et al.

and Salas (1985), Georgakakos and Kavvas (1987), Isaaks and Srivastava (1989),
Kumar and Foufoula-Georgiou (1994), Deidda (2000), Ferraris et al. (2003), Ciach
and Krajewski (2006), Berne et al. (2009), Ly et al. (2011), and Dumitrescu
et al. (2016) further advanced the application of geostatistical methods in rainfall
prediction. The theoretical basis of the geostatistical approach was strengthened
using Bayesian inference via the Markov Chain Monte Carlo (MCMC) algorithm
introduced by Metropolis et al. (1953). MCMC was subsequently adapted by
Hastings (1970) for statistical problems and further applied by Diggle et al. (1998) in
geostatistical studies. Recent developments in MCMC computing now allow fully
Bayesian analyses of sophisticated multilevel models for complex geographically
referenced data. This approach also offers full inference for non-Gaussian spatial
data, multivariate spatial data, spatio-temporal data, and solutions to problems such
as geographic and temporal misalignment of spatial data layers (Banerjee et al.
2014).
The data we are studying are monthly rainfall data measured across the state
of South Carolina from the start of 2011 to the end of 2015. The precipitation
record in 2015 is of particular interest because a storm in October 2015 in North
America triggered a high precipitation event, which caused historic flash flooding
across North and South Carolina. Rainfall across parts of South Carolina reached
500-year-event levels (NBC News, October 4, 2015). Accumulations reached
24.23 in. near Boone Hall (Mount Pleasant, Charleston County) by 11:00 a.m.
Eastern Time on October 4, 2015. Charleston International Airport saw a record
24-h rainfall of 11.5 in. (290 mm) on October 3 (Santorelli, October 4, 2015).
Some areas experienced more than 20 in. of rainfall over the 5-day period. Many
locations recorded rainfall rates of 2 in. per hour (National Oceanic and Atmospheric
Administration (NOAA), U.S. Department of Commerce, 2015).
The extraordinary rainfall event was generated by the movement of very moist
air over a stalled frontal boundary near the coast. The clockwise circulation around
a stalled upper level low over southern Georgia directed a narrow plume of tropical
moisture northward and then westward across the Carolinas over the course of
4 days. A low pressure system off the US southeast coast, as well as tropical
moisture related to Hurricane Joaquin (a category 4 hurricane) was the underlying
meteorological cause of the record rainfall over South Carolina during October 1–5,
2015 (NOAA, U.S. Department of Commerce 2015).
Flooding from this event resulted in 19 fatalities, according to the South Car-
olina Emergency Management Department, and South Carolina state officials said
damage losses were 1.492 billion dollars (NOAA, U.S. Department of Commerce
2015). The heavy rainfall and floods, combined with aging and inadequate drainage
infrastructure, resulted in the failure of many dams and flooding of many roads,
bridges, and conveyance facilities, thereby causing extremely dangerous and life-
threatening situations.
The chapter is arranged as follows: in Sect. 2, we give an overview of our
precipitation data, in conjunction with some other variables, e.g., sea surface
temperature, which might help explain the behavior of the precipitation. In Sect. 3,
we introduce the kriging method to analyze the precipitation using a pure spatial
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 33

analysis. In Sect. 4, some methods in seasonal trend removal are discussed. In

Sect. 5, the Gaussian process is introduced to build a spatio-temporal model.

2 Data Description

2.1 Overview

The original data used in this research are the daily precipitation records in South
Carolina from National Oceanic and Atmosphere Administration (NOAA) between
2011 and 2015. The original data files include daily precipitation, maximum
temperature, and minimum temperature, along with the latitude, longitude, and
elevation of each observation’s location.
In addition, to investigate the effect of El Niño-Southern Oscillation (ENSO)
activity on precipitation, we have calculated an index based on the monthly sea
surface temperature (SST). The derivation of our index is given in Sect. 2.3.

2.2 Data Preprocessing

We collected 281 unique meteorological locations in South Carolina with varying

completeness of data. For instance, if we look at the most recent 5 years (2011–
2015), 31 locations do not have any record of precipitation while 65 locations have
a complete record. The other 185 locations contain missing data ranging from 30%
to less than 5% of the total data set size.
In Fig. 1, we plot all the meteorological locations with an available precipitation
record on October 3, 2015, when the storm struck South Carolina. Note that
smoothing is necessary since most of observations are clustered in several regions.
See Bivand et al. (2008) for more information about the sp package, which provides
a comprehensive solution for spatial data visualization.
We aggregate the daily records into monthly variables. The monthly maximum
of precipitation is calculated since we are interested in capturing the extreme
rainfall behavior which might lead to flooding subsequently. The monthly midrange
temperature, which reflects the general warmth of that month, is computed by
averaging the highest and the lowest daily temperature for that month.
To incorporate more temperature information, we find the range of daily maxima
over a month. We similarly obtain the range of the daily minima. Lastly, for each
location, we also find an overall range, the difference of the maximum and minimum
temperature of that month.
In the data set, several variables, e.g., precipitation, elevation, and temperature
have missing values. We replace each missing observation with the weighted
average of its neighbors. The weights are determined by the distance between
34 H. Liu et al.

Fig. 1 The meteorological

locations with available
record of precipitation on
October 3, 2015

locations.
n In other words, if we denote the missing value at s∗ by Y (s∗ ), then
i=1 w(si )Y (si ) can be used as the imputed value, where

 n

||s∗ − si || ||s∗ − si ||
w(si ) = K / K . (1)
h h
i=1

Note that ||si − s∗ || refers to the haversine distance rather than the Euclidean
distance. We impute missing data based on neighboring observations because doing
so takes the spatial correlation into consideration.

2.3 A Sea Surface Temperature (SST)-Related Variable

El Niño-Southern Oscillation (ENSO) is an irregular variation in winds and sea

surface temperature (SST) over the tropical eastern Pacific Ocean, affecting much
of the tropic and subtropics. Like other climate indices, ENSO occurs irregularly
and is associated with changing in physical pattern of temperature and precipitation.
Figure 2 gives the plot of sea surface temperature for ocean locations off the coast
of South Carolina in June 2015. In this figure, dark colors correspond to cooler sea
temperature values. Scientists believe that the ENSO has a significant influence on
precipitation and hence controls flood magnitude and frequency. We thus include
an SST-based index as a proxy for the ENSO activity. Since our rainfall data are
observed for inland locations, we must define our index related to SST for such
inland locations, rather than for off-shore locations where sea temperature is actually
measured.
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 35

Fig. 2 The sea surface temperature near South Carolina

For any inland location si at a given month, we build an index based on the SST
values of the nearest n adjacent ocean observation points {zj }, where j = 1, . . . , n.
Denote this SST-based index as W (si ) for the ith inland location. It follows that
n 
1
wj
W (si ) = n SST(zj ), (2)
n l=1 wl
j =1

where the weight wj can be determined by the kernel function K(||si − zj ||) for
j = 1, . . . , n, which is symmetric around 0. We use the standard normal density
as the kernel function. The kernel function includes a bandwidth h, thus making
||s −z ||
wj = h1 K( i h j ). The bandwidth parameter h is set to 0.25 times the range of all
of the distances.
Additionally, we simplify the calculation by considering only locations within a
certain threshold. Figure 3 gives a demonstration to calculate the SST-related index
for Columbia, South Carolina. We first determine the sea temperature records to be
included based on a 300-mile threshold. For the included measurements, we find
their weights by calculating their distance to Columbia, and derive the SST-related
index based on (2). Note that the closer a location is to the coast, the more sea surface
temperature records are used to derive an SST-related index for that location.
36 H. Liu et al.

Fig. 3 A demonstration of the calculation of the SST-related variable. The red points are the
observations that are included in the calculation

3 Precipitation Modeling: A Spatial Perspective

In this section, we use a spatial model for the rainfall data without considering
the temporal aspect. Since geostatistical data feature a strong correlation between
adjacent locations, we start by modeling the covariance structure with a variogram,
and then we propose two methods of predicting the rainfall for new location.

3.1 Describing the Spatial Structure: Variogram

We assume that our spatial process has a mean, μ(s) = E(Y (s)), and that the
variance of Y (s) exists for all s ∈ D. The process Y (s) is said to be Gaussian
if, for any n ≥ 1 and any set of sites {s1 . . . , sn }, Y = (Y (s1 ), . . . , Y (sn ))T has
a multivariate normal distribution. Moreover, the process is intrinsic stationary if,
for any given n ≥ 1, any set of n sites {s1 , . . . , sn } and any h ∈ Rr , we have
E[Y (s + h) − Y (s)] = 0, and E[Y (s + h) − Y (s)]2 = Var(Y (s + h) − Y (s)) = 2γ (h)
(Banerjee et al. 2014).
In other words, E[Y (s + h) − Y (s)]2 only depends on h, and not the particular
choice of s. The function 2γ (h) is then called the variogram, and γ (h) is called
the semivariogram. Another important concept is that of an isotropic variogram. If
the semivariogram function γ (h) depends upon the separation vector only through
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 37

its length ||h|| (distance between observations), then the variogram is isotropic.
Otherwise, it is anisotropic. Isotropic variograms are popular because of simplicity,
interpretability, and, in particular, because a number of relatively simple parametric
forms are available as candidates for the semivariogram, e.g., linear, exponential,
Gaussian, or Matérn (or K-Bessel).
A variogram model is chosen by plotting the empirical semivariogram, a simple
nonparametric estimate of the semivariogram, and then comparing it to the various
theoretical parametric forms (Matheron 1963). For demonstration purposes, we
choose the precipitation values of October 13 in 2015, shortly after the flood struck
South Carolina. Assuming intrinsic stationarity and isotropy, the Matérn model is
used due to its better fit to the empirical semivariogram. The correlation function
of this model allows control of spatial association and smoothness. See Fig. 4 for a
plot of this fit.

Fig. 4 The empirical and parametric (Matérn) variogram for the precipitation values in October
13, 2015
38 H. Liu et al.

3.2 Spatial Prediction

Inverse Distance Weighted Interpolation

We use inverse distance weighting (IDW) (Bivand et al. 2008) to compute a spatially
continuous rainfall estimate as a weighted average for a given location s0 ,

w(si )Z(si )
Ẑ(s0 ) =  , where w(si ) = ||si − s0 ||−p .
w(si )

In other words, the weight of a given observed location is based on its Lp -distance
to the interpolation location. If location s0 happens to have an observation, then
the observation itself will be used to avoid the case of infinite weights. The weight
assigned to data points will be more influenced by neighboring points when they are
more clustered. The best p found by cross validation for the analysis of our data set
is approximately 2.5.
Although this method does not incorporate the covariates, it still possesses some
desirable features. For instance, we can make a prediction for the rainfall amount at
every single location with a latitude and longitude.

Linear Gaussian Process Model (Kriging)

Since our precipitation data in the study are geostatistical data, we may employ
a linear Gaussian process model (Cressie 1993). We start by defining the spatial
process at location s ∈ Rd as

Z(s) = X(s)β + w(s), (3)

where X(s) is a set of p covariates associated with each site s, and β is a p-

dimensional vector of coefficients. Spatial dependence is imposed via the residual
terms, i.e., w(s). Specifically, we model {w(s) : s ∈ Rd } as a zero mean
Gaussian process. In other words, the vector w = (w(s1 ), . . . , w(sn ))T follows
w| ∼ Nn (0, ()). We assume  to be a symmetric and positive definite matrix
in order to end up with a sensible distribution. To ensure these conditions, ()
can be treated as a function of  with certain constraints, which are tantamount to
specifying a variogram model.
Among several variogram structures, e.g., spherical, Gaussian, exponential, etc.
we choose the exponential covariance with parameters  = (ψ, κ, φ), where
ψ, κ, φ > 0. The exponential covariance () has the form

() = ψI + κH (φ), whereH (φ) = exp(−||si − sj ||)/φ).

Note that ||si − sj || is the Euclidean distance between location i and j . Another type
of distance, Geodesic, takes the curvature of the earth’s surface into consideration.
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 39

We use Euclidean distance since most of our distances are between South Carolina
counties and the effects of curvature are thus negligible.
The exponential model enjoys a simple interpretation. The “nugget” in a
variogram graph is represented by ψ in this model, and this nugget is also the
variance of the non-spatial error. Moreover, κ and φ dictate the scale and range of
the spatial dependence, respectively. Also note that the exponential model assumes
the covariance and hence dependence between two locations decreases as distance
between locations increases, which is sensible for the study of rainfall behavior.
Letting Z = (Z(s1 ), . . . , Z(sn ))T , we estimate the multivariate normal distribu-
tion for Z after parameter estimation. To find the unknown parameters  and β,
we use Bayesian methods implemented by the spTimer package in R (Bakar and
Sahu 2015), which requires users to provide sensible prior information based on
sample variogram graphs. Note that this model fitting process will collapse if we
start with initial values far from the true value.

Monte Carlo Simulation for Kriging

Predictions of the process, Z∗ = (Z(s∗1 ), . . . , Z(s∗m ))T , where s∗i is the ith new
location, can be obtained via the posterior predictive distribution
!
π(Z∗ |Z) = π(Z∗ |Z, , β)π(, β|Z)ddβ,

by sampling from the posterior predictive distribution in two steps:

• Step 1: Simulate  , β  ∼ π(, β|Z) by the Metropolis–Hastings algorithm.
• Step 2: Simulate Z∗ | , β, Z from a multivariate normal density.
For step 1, it suffices to find the posterior distribution π(, β|Z) based on (1) and
(2). The posterior distribution has low dimension as long as we do not have many
covariates. The major challenge is that since covariance parameters might be highly
correlated, one must expect autocorrelation issues in the sampler, which can be
alleviated by a block updating scheme, a scheme that generates multiple covariance
parameters in a single Metropolis–Hastings step.
For step 2, the joint distribution of Z and Z∗ is given by
     
Z μ1 11 12
|, β ∼ N ,
Z∗ μ2 21 22

based on which one can find the conditional distribution of Z∗ | , β, Z. According

to Anderson (2003), it follows that
−1
E(Z∗ | , β, Z) = μ2 + 21 11 (Z − μ1 ),
−1
Var(Z∗ | , β, Z) = 22 − 21 11 12 .
40 H. Liu et al.

Hence, one can obtain simulated observations that follow a given covariance
structure by iterating between step 1 and step 2. Bivand et al. (2008) suggest the
method of sequential simulation: (1) compute the conditional distribution with our
given data, (2) draw a value from this conditional distribution, (3) add this value into
the data set, and (4) repeat steps (1)–(3).
As Z becomes a larger matrix as more data are generated, the algorithm
becomes more and more expensive. Many strategies are proposed for reducing the
considerable computational burden posed by matrix operations, including the use
of covariance functions (Hughes and Haran 2013) as well as setting a maximum
number of neighbors (Bivand et al. 2008). In our study, we used the maximum
number of neighbors with the nearest 40 observations.
We illustrate prediction by modeling rainfall in South Carolina on October 13,
2015 with a kriging model that assumes an exponential spatial covariance structure.
Using the Monte Carlo approach described above, we predict by simulating from the
posterior predictive distribution. This can be done repeatedly to give a sense of the
variability associated with the spatial predictions. Figure 5 demonstrates ten simu-

Fig. 5 Ten simulated precipitation heat maps based on kriging. The darker color indicates heavier
precipitation and vice versa. A consistent look reveals a robust performance of the kriging model
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 41

lated predictions of the spatial distribution of rainfall amounts in a small rectangular

spatial area in the northwest corner of South Carolina. The darker color indicates
heavier predicted precipitation and the lighter color a small predicted rainfall. The
consistent pattern across all ten simulations reveals a robust performance of the
kriging model. A pointwise prediction at any spatial location could be obtained by
averaging the predicted rainfall values at that location across all ten simulations.

4 Seasonal Trend Removal

We now analyze the geostatistical rainfall data across time. Due to the nature of our
rainfall data, the seasonality is of particular interest when we model the temporal
trend. We propose two methods to remove the seasonal trend in this section.

4.1 Harmonic Regression

To remove the seasonal trend, one approach is to fit a first-order harmonic regression
model with terms sin(x) and cos(x). In addition, we set x = 2π t if the period is
1. In our case, it is justifiable to set the period as 12 since the monthly rainfall is
measured, and thus x = (π/6)t is used. Hence, one can regress the precipitation
y against dependent variables sin((π/6) t) and cos((π/6) t). The omnibus F-test to
test for the usefulness of the trigonometric terms in this multiple regression model
gives a p-value close to 1, which confirms the existence of seasonality.
One can also use a second-order harmonic model to capture more complex
behavior, in which two more terms, sin[(4π/ω)t] and cos[(4π/ω)t] are included,
where ω is the periodic parameter. However, for our rainfall data, it is unnecessary
to include these two other terms since we observe no great improvement in model
fit by introducing the extra terms (see Fig. 6).

Fig. 6 The fitted model based on the first- and second-order harmonic models. The dotted line
corresponds to the second-order model, and the solid red line corresponds to the first-order model
42 H. Liu et al.

4.2 Seasonality Indicator

Another approach to model seasonality in the spatio-temporal model is the seasonal

means model. Specifically, one indicator variable will be 1 if the record is collected
from January to March, and will be 0 otherwise. Similarly, another dummy variable
indicates the month April to June while a third dummy variable indicates July to
September. Lastly, if all three variables are 0, then the observation is from the last
3 months of the year. Note that one could also include dummy variables for months
in a similar way if necessary, but we have found that it is sufficient to model the
means of the four seasons.

5 Precipitation Modeling: A Spatio-Temporal Perspective

In this section, we discuss how to model spatio-temporal data with two different
methods, the Gaussian process (GP) model and autoregressive (AR) model. The
latter model is an extension of the Gaussian process model obtained by introducing
an autoregressive term.

5.1 Gaussian Process (GP) Model

The independent Gaussian process (GP) model (Cressie and Wikle 2015; Gelfand
et al. 2010) is specified hierarchically in two stages,

Zt = μt +  t (4)
μt = Xt β + ηt , (5)

in which Zt = (Z(s1 , t), . . . , Z(sn , t))T , which defines the response variable for
all n locations at time t. It is known that s1 , . . . , sn can be indexed by latitude and
longitude. In the first layer, Zt is defined by a simple mean model plus a pure white
noise term,  t . We therefore assume that

 t ∼ N(0, σ2 In ), (6)

where the σ2 is the pure error variance and In is the identity matrix.
The second level models μt as the sum of fixed covariates and random effects at
time t. The fixed term, Xt β, comes from the covariates, and ηt is the spatio-temporal
random effects, ηt = (η(s1 , t), . . . , η(sn , t))T . Similar to  t , ηt also follows a
multivariate normal distribution whose mean vector is 0. However, ηt has a more
complicated covariance matrix than does  t .
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 43

We use the exponential function to specify the correlation matrix of the random
effects. The correlation strength is solely based on the distance between si and sj ,
which is given by

 η = ση2 H (φ) + τ 2 In ,

where H (φ) = exp(−||si − sj ||)/φ), and ||si − sj || indicates the spatial distance
between location i and j . This function is used to determine each element in the
matrix Sη , where  η = ση2 Sη . This parameterization allows ση2 to capture the
invariant spatial variance, and Sη is used to capture the spatial correlation.
The posterior distribution involves three layers, i.e., the prior distribution for
parameters, the mean model, and the random effects model. We will set aside the
prior for later discussion and use π(θ) = π(β, ν, φ, ση2 , σ2 ) to refer to the prior in
general. Thus the posterior is given by


N
g(θ , μ|Z) = π(θ) × fn (Zt |μt , σ2 )gn (μt |β, υ, φ, ση2 ). (7)
t=1

To be specific, we use fn (·) and gn (·) to indicate an n-dimensional distribution

function. In this case, each of them is a multivariate normal distribution, and n is
the number of locations in the data set and N is the number of time points. μt is the
vector of random effects for time t and we use μ on the left-hand side to refer to the
collection of all random effects.
Since both Zt and μt follow a multivariate normal distribution, their density
functions are given as follows:

1 1
fn (Zt |μt , σ2 ) =" exp − 2 (Zt − μt )T (Zt − μt ) , (8)
(2π ) |σ In |
n 2 2σ
 
1 1 T −1
gn (μt |S η , ση , β) = #
2
exp − 2 (μt − Xt β) S η (μt − Xt β) ,
(2π )n |σ 2 S | 2ση
η η
(9)

Thus the posterior distribution is given by plugging (8) and (9) into (7). The
logarithm of the joint posterior distribution of the parameters for this Gaussian
process model is given by

N
N
logπ(σ2 , ση2 , μ, β, υ, φ|Z) ∝ log σ2 − (Zt − μt )T (Zt − μt )
2 2σ2
t=1
 
N 1

N
1
− log |ση2 S η | − − 2 (μt − Xt β)T S −1
η (μt − Xt β) + log π(θ).
2 2ση2 2ση
i=1
44 H. Liu et al.

We specify the prior π(θ) to reflect the assumption that β, ν, φ, ση2 , and σ2
are mutually independent, so the joint prior is the product of the marginal prior
densities, which are given as follows: All the parameters describing the mean,
e.g., β and ρ (see Sect. 5.2) are given independent normal prior distributions, with
the prior on ρ truncated to have support on (−1, 1). We assume φ and ν both
follow uniform distributions, while the prior for the precision (inverse of variance)
parameter is a gamma distribution. We choose the hyperparameters to make these
prior distributions very diffuse.

5.2 Autoregressive (AR) Model

In this section, we introduce the autoregressive model (Sahu and Bakar 2012). The
hierarchical AR(1) model is given as follows:

Zt = μt +  t
μt = ρμt−1 + Xt β + ηt ,

where ρ denotes the unknown temporal correlation parameter assumed to be in

the interval (−1, 1). Obviously, for ρ = 0, these models reduce to the GP model
described in Sect. 5.1.
The autoregressive model requires specification of the initial term, the first
random effect, which has mean β 0 and covariance matrix σ02 S 0 . The correlation
matrix S 0 is obtained using the exponential correlation function. The derivation of
the posterior distribution is similar to that in GP model with ρ = 0. The logarithm
of the posterior distribution of the parameters is now given by

N
N
logπ(σ2 , ση2 , μ, β, υ, φ|Z) ∝ log σ2 − (Zt − μt )T (Zt − μt )
2 2σ2
t=1
N
− log |ση2 S η |
2
 
1

N
1 T −1
− − 2 (μt − ρμt−1 − Xt β) S η (μt − ρμt−1 − Xt β)
2ση2 2ση
i=1
1 1
− log |σ02 S 0 | − (μ0 − β 0 )T S −1
0 (μ0 − β 0 ) + log π(θ )
2 2σ02

Note that β 0 is only a mean vector for the initial random effect term, which is
different from β, which refers to regression coefficients corresponding to covariates
X. In other words, the terms in the last line (except log π(θ)) derive from the initial
random effect term.
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 45

5.3 Model Fitting

In this section, we fit the AR(1) model with monthly precipitation data from the
beginning of year 2011 to the end of year 2015. A natural log transformation was
initially applied to the precipitation to improve the model fit and ensure positive
predicted rainfall values once we back-transform by exponentiating the predicted
log-rainfall values. We include temperature range, sea surface temperature, and
elevation as monthly covariates.
We initially found that ordinary temperature measurements such as the monthly
average temperature were not apparently related to precipitation after accounting for
the season and thus we did not include these in the model. However, measurements
of variability in temperature over each month, e.g., the range of daily maxima and
the range of daily minima over a month, were believed to have an effect on precip-
itation and thus we include these to determine whether their effects are significant.
We also include a flood-year indicator as a dummy variable, where data
from 2015 is labeled as 1 and otherwise 0, to account for the unusual October
precipitation amounts in this year. Interaction terms involving the dummy variable
were also tested, none of which were statistically significant and were thus removed
from the final model. The acceptance rate from Metropolis step for all parameters is
42.97% and a brief summary of model fitting details is given as follows:

-----------------------------------------------------
Model: AR
Call: LOG ~ RANGE_OVERALL + RANGE_LOW + RANGE_HIGH
+ SST + ELEVATION + SST * RANG E_LOW + Year2015

Iterations: 5000
nBurn: 1000
Acceptance rate: 29.76
-----------------------------------------------------
Parameters
Mean Median SD Low2.5p Up97.5p
(Intercept) 0.3635 0.3689 0.1363 0.0894 0.6265
RANGE_OVERALL -0.0006 -0.0006 0.0017 -0.0039 0.0027
RANGE_LOW 0.0017 0.0017 0.0030 -0.0040 0.0078
RANGE_HIGH 0.0006 0.0007 0.0011 -0.0016 0.0028
SST -0.0057 -0.0058 0.0045 -0.0142 0.0033
ELEVATION 0.0001 0.0001 0.0001 0.0000 0.0002
Year2015 0.0808 0.0810 0.0180 0.0450 0.1154
RANGE_LOW:SST -0.0001 -0.0001 0.0001 -0.0003 0.0001
rho 0.0756 0.0757 0.0151 0.0466 0.1054
sig2eps 0.0054 0.0054 0.0002 0.0051 0.0057
sig2eta 0.0764 0.0739 0.0121 0.0617 0.1073
phi 0.0501 0.0502 0.0090 0.0322 0.0659
-----------------------------------------------------
46 H. Liu et al.

Fig. 7 The residual plot and QQ plot from AR(1) prediction

The dummy variable for year 2015 is significant. After back-transforming, we

can say the predicted monthly rainfall for 2015 is exp(0.0808) = 1.084 times
greater than the predicted monthly rainfall in other years, holding other predictors
fixed. This is consistent with the flooding event in the fall of 2015. Another finding
is that elevation might be an explanatory factor to the rainfall since higher elevation
relates to higher volumes of precipitation. In addition, a statistically significant and
positive ρ indicates that a rainy month might tend to precede another one. On the
other hand, the SST has a marginally negative effect on the rainfall prediction but is
not significant based on the 95% credible interval.
We also obtain the residuals and the QQ plot in Fig. 7. There is no obvious pattern
in the residual plots. However, the residuals show deviations in the tails to some
extent from normality based on the QQ plot on the right panel, which indicates a
heavy-tailed error distribution and lack of symmetrical pattern (e.g., Samadi et al.
2017).

6 Model Comparison: State-Space Model vs. Gaussian

Process

Another framework for spatio-temporal data analysis is the dynamic state-space

model. A formulation of the spatio-temporal framework (Stroud et al. 2001) is
specified as follows:

yt (s) = xt (s)T β t + ut (s) + t (s), t (s) ∼ N (0, τt2 )

β t = β t−1 + ηt , ηt ∼ N(0,  η )
μt (s) = μt−1 (s) + wt (s), wt (s) ∼ GP (0, Ct (·, θ t )).
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 47

Fig. 8 The 95% confidence interval for β1 (the SST-related variable) and β2 (elevation) over
12 months in 2015

Here xt (s) is a p × 1 vector of predictors and β t is a p × 1 vector of coefficients.

The GP (0, Ct (·, θ t )) denotes a spatial Gaussian process with covariance function
Ct (·, θ t ). We further specify Ct (s1 , s2 ; θ t )) = σt2 ρ(s1 , s2 ; φt ), where θ t = {σt2 , φt }
and ρ(·; φ) is a correlation function with φ controlling the correlation decay.
The same response variable and covariates with AR(1) model are used when
fitting the state-space model. The R package spBayes (Finley et al. 2007) provides
a framework to sample from parameters and posterior. The 95% credible interval for
sea surface temperature and elevation are plotted for all 12 months in 2015.
The state-space model allows for a more detailed monthly look of the effect
of covariates. For instance, one can conclude that, based on Fig. 8, the SST-based
variable effects the rainfall amount in a more significant manner during the first
few months of the year. These results strengthen the previous findings of Häkkinen
(2000), Mehta et al. (2000), Wang et al. (2006), and Dima and Lohmann (2010), and
further support the hypothesis that the variability of North Atlantic SST is coherent
with the fluctuations of the rainfall pattern and occurrence. In other words, intense
ocean–atmosphere coupling exists in the North Atlantic, particularly during winter.
In contrast, elevation is more related to the precipitation in June and October, when
heavier rainfall data are observed. This covariate specifies a convective mode that is
widely recognized as an important contributor to the probability and type of severe
convective rainfall during summer and early fall in the southeast region. The residual
plot and the QQ plot for the state-space are shown in Fig. 9. We see the heavy-tailed
error pattern is still apparent in this model, based on the QQ plot.
48 H. Liu et al.

Fig. 9 The residual plot and QQ plot from the state-space model

7 Discussion

We have presented both spatial and spatio-temporal models for rainfall in South Car-
olina during a period including one of the most destructive storms in state history.
Our models have allowed us to determine several covariates that affect the rainfall
and to interpret their effects. In particular, the flood year of 2015 was an important
indicator of rainfall and elevation also had a positive significant effect on precip-
itation. There was a significant positive correlation in rainfall measurements over
time. Finally, our novel SST index provided some evidence that cooler nearby sea
temperatures corresponded to higher rainfall at in land sites although this SST effect
was not significant at the 0.05 level based on a 95% credible interval for its effect.
A spatial prediction at a new location and a temporal prediction at a future time
point can be obtained based on the posterior predictive distribution for Z(s0 , t  ),
where s0 denotes a new location and t  is a future time point. Further details
regarding these predictions are provided in Cressie and Wikle (2015) for the GP
models, and Sahu and Bakar (2012) for the AR models.
A limitation of the study, and a direction for future research, is that the model
does not account for the apparent heavy-tailed nature of the errors. Methods
involving generalized extreme value distribution (Rodríguez et al. 2016) could
possibly be adapted to this model to help handle this heavy-tailed error structure,
but such research is still relatively new in the spatio-temporal modeling literature.

References

Anderson, T.W.: An Introduction to Multivariate Statistical Analysis. Wiley, Hoboken (2003)

Bakar, K.S., Sahu, S.K.: spTimer: spatio-temporal Bayesian modeling using R. J. Stat. Softw.
63(15), 1–32 (2015)
Spatial and Spatio-Temporal Analysis of Precipitation Data from South Carolina 49

Banerjee, S., Carlin, B.P., Gelfand, A.E.: Hierarchical Modeling and Analysis for Spatial Data.
CRC Press, Boca Raton (2014)
Benzécri, J.P.: L’Analyse des Données. Dunod, Paris (1973)
Berne, A., Delrieu, G., Boudevillain, B.: Variability of the spatial structure of intense Mediter-
ranean precipitation. Adv. Water Resour. 32(7), 1031–1042 (2009)
Bivand, R.S., Pebesma, E.J., Gomez-Rubio, V., Pebesma, E.J.: Applied Spatial Data Analysis with
R. Springer, New York (2008)
Ciach, G.J., Krajewski, W.F.: Analysis and modeling of spatial correlation structure in small-scale
rainfall in central Oklahoma. Adv. Water Resour. 29(10), 1450–1463 (2006)
Cressie, N.: Statistics for Spatial Data. Wiley, New York (1993)
Cressie, N., Wikle, C.K.: Statistics for Spatio-Temporal Data. Wiley, New York (2015)
Deidda, R.: Rainfall downscaling in a space-time multifractal framework. Water Resour. Res.
36(7), 1779–1794 (2000)
Delhomme, J.P.: Kriging in the hydrosciences. Adv. Water Resour. 1(5), 251–266 (1978)
Delfiner, P., Delhomme, J.P.: Optimum Interpolation by Kriging. Ecole Nationale Supérieure des
Mines, Paris (1975)
Diggle, P.J., Tawn, J.A., Moyeed, R.A.: Model-based geostatistics. J. R. Stat. Soc.: Ser. C: Appl.
Stat. 47(3), 299–350 (1998)
Dima, M., Lohmann, G.: Evidence for two distinct modes of large-scale ocean circulation changes
over the last century. J. Clim. 23(1), 5–16 (2010)
Dumitrescu, A., Birsan, M.V., Manea, A.: Spatio-temporal interpolation of sub-daily (6 h)
precipitation over Romania for the period 1975–2010. Int. J. Climatol. 36(3), 1331–1343 (2016)
Ferraris, L., Gabellani, S., Rebora, N., Provenzale, A.: A comparison of stochastic models for
spatial rainfall downscaling. Water Resour. Res. 39(12), 1368 (2003). https://doi.org/10.1029/
2003WR002504
Finley, A.O., Banerjee, S., Carlin, B.P.: spBayes: an R package for univariate and multivariate
hierarchical point-referenced spatial models. J. Stat. Softw. 19(4), 1 (2007)
Gelfand, A.E., Diggle, P., Guttorp, P., Fuentes, M.: Handbook of Spatial Statistics. CRC Press,
Boca Raton (2010)
Georgakakos, K.P., Kavvas, M.L.: Precipitation analysis, modeling, and prediction in hydrology.
Rev. Geophys. 25(2), 163–178 (1987)
Häkkinen, S.: Decadal air-sea interaction in the North Atlantic based on observations and modeling
results. J. Clim. 13(6), 1195–1219 (2000)
Hastings, W.K.: Monte Carlo sampling methods using Markov chains and their applications.
Biometrika 57(1), 97–109 (1970)
Hughes, J., Haran, M.: Dimension reduction and alleviation of confounding for spatial generalized
linear mixed models. J. R. Stat. Soc. Ser. B Stat Methodol. 75(1), 139–159 (2013)
Isaaks, H.E., Srivastava, R.M.: Applied Geostatistics. Oxford University Press, New York (1989)
Kumar, P., Foufoula-Georgiou, E.: Characterizing multiscale variability of zero intermittency in
spatial rainfall. J. Appl. Meteorol. 33(12), 1516–1525 (1994)
Ly, S., Charles, C., Degre, A.: Geostatistical interpolation of daily rainfall at catchment scale: the
use of several variogram models in the Ourthe and Ambleve catchments, Belgium. Hydrol.
Earth Syst. Sci. 15(7), 2259–2274 (2011)
Matheron, G.: Principles of geostatistics. Econ. Geol. 58(8), 1246–1266 (1963)
Mehta, V., Suarez, M., Manganello, J.V., Delworth, T.D.: Oceanic influence on the North Atlantic
oscillation and associated northern hemisphere climate variations: 1959–1993. Geophys. Res.
Lett. 27(1), 121–124 (2000)
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E.: Equation of state
calculations by fast computing machines. J. Chem. Phys. 21(6), 1087–1092 (1953)
National Oceanic and Atmosphere Administration, U.S. Department of Commerce: Service
assessment: the historic South Carolina floods of October 1–5, 2015. www.weather.gov/
media/publications/assessments/SCFlooding_072216_Signed_Final.pdf (2015). Accessed 4
Dec 2017
50 H. Liu et al.

Rodríguez, S., Huerta, G., Reyes, H.: A study of trends for Mexico city ozone extremes: 2001–
2014. Atmósfera 29(2), 107–120 (2016)
Sahu, S.K., Bakar, K.S.: Hierarchical Bayesian autoregressive models for large space–time data
with applications to ozone concentration modeling. Appl. Stoch. Model. Bus. Ind. 28(5), 395–
415 (2012)
Samadi, S., Tufford, D., Carbone, G.: Estimating hydrologic model uncertainty in the presence of
complex residual error structures. Stoch. Environ. Res. Risk Assess. 32(5), 1259–1281 (2018)
Sharon, D.: Spatial analysis of rainfall data from dense networks. Hydrol. Sci. J. 17(3), 291–300
(1972)
Stroud, J.R., Müller, P., Sansó, B.: Dynamic models for spatio-temporal data. J. R. Stat. Soc. Ser.
B Stat. Methodol. 63(4), 673–689 (2001)
Tabios III, Q.G., Salas, J.D.: A comparative analysis of techniques for spatial interpolation of
precipitation. Water Resour. Bull. 21(3), 365–380 (1985)
Thiessen, A.H.: Precipitation averages for large areas. Mon. Weather Rev. 39(7), 1082–1084 (1911)
Troutman, B.M.: Runoff prediction errors and bias in parameter estimation induced by spatial
variability of precipitation. Water Resour. Res. 19(3), 791–810 (1983)
Wang, C., Enfield, D.B., Lee, S.K., Landsea, C.W.: Influences of the Atlantic warm pool on western
hemisphere summer rainfall and Atlantic hurricanes. J. Clim. 19(12), 3011–3028 (2006)
A Sparse Areal Mixed Model
for Multivariate Outcomes,
with an Application to Zero-Inflated
Census Data

Donald Musgrove, Derek S. Young, John Hughes, and Lynn E. Eberly

1 Introduction

The Committee on National Statistics assembled the Panel to Review the 2010
Census to suggest general priorities for research in preparation for the 2020
Census. In their first interim report (Cook et al. 2011) the Panel laid out three
recommendations, the first of which highlighted “four priority topic areas, in order
to achieve a lower cost and high-quality 2020 Census.” A theme across these priority
areas was the effective use of Census Bureau databases (e.g., geographic databases
and databases built with administrative records) to achieve operational objectives.
In addition to implementing recommendations from the Panel to Review the 2010
Census, the Census Bureau is placing increasing emphasis on accurate model-based
predictions as a way to more generally conduct efficient and cost-effective surveys
(U.S. Census Bureau 2015).
One of the Census Bureau’s most prominent databases is the Master Address File
(MAF), which is a continually updated inventory of all known living quarters in the
USA and its island territories. The MAF is used as a sampling frame for various
Census Bureau surveys, including the decennial Census. The MAF comprises

D. Musgrove
Medtronic, Minneapolis, MN, USA
D. S. Young ()
Department of Statistics, University of Kentucky, Lexington, KY, USA
e-mail: derek.young@uky.edu
J. Hughes
Department of Biostatistics and Informatics, University of Colorado, Denver, CO, USA
L. E. Eberly
Division of Biostatistics, University of Minnesota, Minneapolis, MN, USA

© Springer Nature Switzerland AG 2019 51

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_3
52 D. Musgrove et al.

approximately 194 million records, and includes geographic information, residential

status codes, and addresses. As discussed in Young et al. (2017), there is interest in
building census-block-level models to (1) help understand coverage errors resulting
from a particular MAF extract, which is what defines a survey sampling frame,
and (2) identify blocks with a large number of expected housing unit “adds” and
“deletes,” which could aid targeted address canvassing operations.
Our analysis uses data from the 2010 Census. We are interested in revealing
causes of address deletes, which are defined by the Census Bureau as addresses that
were deleted from the base count because they did not correspond to valid housing
units (a house, an apartment, a mobile home, a group of rooms, or a single room that
is occupied, or intended for occupancy, as a separate living area). The data are zero-
inflated since the aggregated number of deletes has over 90% zeros across the entire
USA. Some zeros arise from those areas in which changes in mailing addresses are
unlikely. Another source of zeros is areas prone to redevelopment, in which case
there is a high “risk” of future address changes. We include various geographic,
demographic, and operational variables as predictors in our model. This work could
lead to a more efficient and cost-effective decennial Census.
The source of the data we analyze is a publicly available Census Bureau dataset
called the Planning Database (PDB) (http://goo.gl/LlcwY7). These data include
variables and counts from the 2010 Census and the 2009–13 American Community
Survey (ACS). The data are aggregated at the block-group level rather than the block
level. A census block is the smallest geographic unit used by the Census Bureau.
Blocks are typically bounded by streets or creeks/rivers. Within a city, a block
corresponds to a city block. In a rural area, blocks may be large and irregularly
shaped and bounded by features including roads, streams, or transmission lines.
Census blocks are not delineated based on population. A block group, on the other
hand, comprises multiple blocks and contains between 600 and 3000 people. The
PDB comprises approximately 220,000 block groups and 300 variables.
Young et al. (2017) developed zero-inflated models to reveal predictors of
housing unit adds and deletes, but their data were from a particular MAF extract.
Their work highlighted the potential success of using model-based predictions in the
decision-making process of costly Census Bureau operations. However, the models
of Young et al. were developed based on data collected just before and just after the
Census Bureau’s 2010 address canvassing operation. Thus, their results illustrated
what could have been done for 2010 operations using model-based strategies.
Regardless, their work served as a proof-of-concept that model-based strategies
should be investigated using other federal databases. For example, the PDB is
an excellent candidate because it includes more geographic and demographic
variables than are available from a traditional MAF extract. We also note that the
models developed in Young et al. (2017) were non-spatial. Failure to account for
consequential spatial dependence can lead to erroneous inference, e.g., confidence
intervals for regression coefficients may be too narrow. Thus, we develop a sparse
areal mixed model for multivariate outcomes to reflect the type of features in the
data discussed above.
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 53

Multivariate areal (i.e., spatially aggregated) outcomes are common in a number

of fields. Disease mapping provides what is perhaps the canonical example: a
number of disease outcomes (e.g., incidence, prevalence, and rate) that share a
collection of spatially varying risk factors. Such outcomes are likely to exhibit
dependence both within and across areal units, in which case applying a multivariate
model instead of multiple univariate models can yield improved inference. For our
Census data analysis, we assess a collection of spatially varying covariates when
modeling the number of address deletes. We accomplish this goal by developing a
new mixed model for multivariate areal data.
Traditional mixed models for multivariate areal outcomes face two formidable
challenges: (1) spatial confounding (Clayton et al. 1993; Reich et al. 2006), which
can lead to erroneous regression inference, and (2) an immense computational
burden. Our approach addresses both of these challenges by extending the sparse
areal mixed model (SAMM) of Hughes and Haran (2013). The model is further
specialized to handle zero-inflated data.
The remainder of this chapter is organized as follows. In Sect. 2 we develop
our model, the multivariate sparse areal mixed model (MSAMM). In Sect. 3 we
discuss the hyperpriors used and computational details for the MSAMM. In Sect. 4
we specialize the MSAMM to handle zero-inflated data and apply the resulting
model to zero-inflated Census data for the state of Iowa. In Sect. 5 we conclude
with a summary and a sketch of future work. We provide derivations, computational
details, and extended simulation results in an appendix.

2 Our Sparse Mixed Model for Multivariate Areal Data

In this section we develop our multivariate sparse areal mixed model (MSAMM).
Our approach is similar to the approach of Bradley et al. (2015) in that we, too,
employ the orthogonal, multiresolutional spatial basis described by Hughes and
Haran (2013) (see also Griffith (2003) and Tiefelsdorf and Griffith (2007)). This
basis, known as the Moran (1950) basis, is appealing from a modeling point of view
and also permits efficient computing.

2.1 Review of Univariate CAR Models

To motivate our development of the MSAMM, we begin by reviewing conditional

autoregressive (CAR) models for univariate areal data. The transformed conditional
mean vector for these models is given by

g(μ) = g {E (y | β, φ)} = Xβ + φ, (1)

where g is a link function; y = (y1 , . . . , yn ) are the outcomes, the ith of which
is associated with the ith areal unit; X is an n × p design matrix; β is a p-vector
54 D. Musgrove et al.

of regression coefficients; and φ = (φ1 , . . . , φn ) are spatially dependent random

effects. Note that the outcomes are assumed to be independent conditional on the
random effects. Marginally, however, the outcomes are spatially dependent because
the random effects are spatially dependent.
The various CAR models are distinguished by the distribution assigned to φ. The
most popular specification for φ is the intrinsic CAR model (Besag and Kooperberg
1995):
 τ
p(φ | τ ) ∝ exp − φ  Qφ ,
2
where p(·) denotes a prior distribution, τ is a smoothing parameter, and the
precision matrix Q is equal to D−A, where D is the diagonal matrix with the degrees
(i.e., number of neighbors) of the areal units on its diagonal and A is the adjacency
matrix for the underlying graph G, i.e., A is the binary matrix that encodes the
adjacency structure among the areal units. Since Q is singular, the intrinsic CAR
is improper. A proper alternative has precision matrix Q(ρ) = D − ρA, where ρ
is constrained to the interval [0, 1). The parameter ρ can be considered a range
parameter, but its effect on the marginal dependence structure is complex and often
pathological (Wall 2004; Assunção and Krainski 2009). Additional proper CAR
specifications exist. For instance, the prior proposed by Leroux et al. (2000) offers
an alternative specification of the spatial random effects resulting in a proper prior
distribution.
Traditional CAR models present serious challenges. First, there is often multi-
collinearity between the spatial random effects and the fixed-effects predictors. This
characteristic, which is known as spatial confounding (Clayton et al. 1993), often
leads to (1) biased estimation of regression coefficients and (2) substantial variance
inflation that may make important covariates appear insignificant. Additionally,
CAR random effects permit patterns of spatial repulsion, which we do not expect to
observe in the types of data to which these models are usually applied.
Second, computation for CAR models can be extremely burdensome due to (1)
the high dimensionality of φ and (2) the nature of φ’s posterior distribution. It
is well known that a univariate Metropolis–Hastings algorithm for sampling from
the posterior distribution of φ leads to a slow mixing Markov chain because the
components of φ exhibit strong a posteriori dependence. This has led to a number
of approaches for updating φ in a block(s). Constructing proposals for these updates
is challenging, and the faster mixing comes at the cost of increased running time per
iteration, see, for instance, Knorr-Held and Rue (2002), Haran et al. (2003), and
Haran and Tierney (2012).

2.2 Review of Hughes and Haran’s SAMM

To alleviate spatial confounding, eliminate patterns of spatial repulsion, and greatly

reduce computing time and storage requirements, Hughes and Haran (2013) intro-
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 55

duced their sparse areal mixed model (SAMM). In signal processing, statistics, and
related fields, it is not uncommon to use the term “sparse” to refer to representation
of a signal in terms of a small number of generating elements drawn from an
appropriately chosen domain (Donoho and Elad 2003). We use the term “sparse”
in precisely this sense, since our model accomplishes spatial smoothing by using
q  n Moran basis vectors (as opposed to traditional CAR models, which have
approximately n spatial random effects). The SAMM can be developed as follows.
Reich et al. (2006) showed that the traditional CAR models are spatially
confounded in the sense that the random effects can “pollute” the regression
manifold C(X), which can lead to a biased and variance-inflated posterior for β.
To see this, first let P be the orthogonal projection onto C(X), so that In − P is the
orthogonal projection onto C(X)⊥ . Now eigendecompose P and In − P to obtain
orthogonal bases Kn×p and Ln×(n−p) for C(X) and C(X)⊥ , respectively. Then (1)
can be rewritten as

g(μ) = Xβ + Kγ + Lδ,

where γ p×1 and δ (n−p)×1 are random coefficients. This form shows that K is the
source of the confounding, for K and X have the same column space.
Since the columns of K are merely synthetic predictors (i.e., they have no
scientific meaning), Reich et al. (2006) recommend removing them from the model.
The resulting model (henceforth the RHZ model) has

g(μ) = Xβ + Lδ,

so that spatial smoothing is restricted to the orthogonal complement of C(X). In a

subsequent paper, Hodges and Reich (2010) referred to this technique as restricted
spatial regression (RSR).
RSR is not only an effective remedy for confounding but also speeds computing.
Because the columns of L are orthogonal, the RHZ model’s random effects are
approximately a posteriori uncorrelated. This yields a fast-mixing Markov chain,
and the cost per iteration is reduced because a simple spherical normal proposal
is sufficient for updating the random effects. But fitting the RHZ model to large
areal datasets is still quite burdensome computationally because the random effects
remain high dimensional.
By taking full advantage of the underlying graph G, Hughes and Haran (2013)
were able to greatly reduce the number of random effects while also improving
regression inference. Hughes and Haran (2013) begin by defining the so-called
Moran operator for X with respect to G: (In − P)A(In − P). This operator appears in
the numerator of a generalized form of Moran’s I , a popular nonparametric measure
of spatial dependence for areal data (Moran 1950):

n y  (In − P)A(In − P)y

IX (A) = .
1 A1 y  (In − P)y
56 D. Musgrove et al.

Fig. 1 Three Moran basis vectors, exhibiting spatial patterns of increasingly finer scale

Boots and Tiefelsdorf (2000) showed that (1) the (standardized) spectrum of a
Moran operator comprises the possible values for the corresponding IX (A), and (2)
the eigenvectors comprise all possible mutually distinct patterns of clustering resid-
ual to C(X) and accounting for G. The positive (negative) eigenvalues correspond
to varying degrees of positive (negative) spatial dependence, and the eigenvectors
associated with a given eigenvalue (ωi , say) are the patterns of spatial clustering that
data exhibit when the dependence among them is of degree ωi .
In other words, the eigenvectors of the Moran operator form a multiresolutional
spatial basis for C(X)⊥ that exhausts all possible patterns that can arise on G.
Since we do not expect to observe repulsion in the phenomena to which these
models are usually applied, we can use the spectrum of the operator to discard all
repulsive patterns, retaining only attractive patterns for our analysis. By retaining
only eigenvectors that exhibit positive spatial dependence, we can usually reduce
the model dimension by at least half a priori. Hughes and Haran (2013) showed that
a much greater reduction is often possible in practice, with 50–100 eigenvectors
being sufficient in many cases. Three example Moran vectors are shown in Fig. 1.
Let Mn×q contain the first q  n eigenvectors of the Moran operator. Then the
SAMM has first stage

g(μ) = Xβ + Mδ s ,

where δ s (“s” for “sparse”) is a q-vector of random coefficients that are assumed to
be jointly Gaussian:

δ s ∼ N {0, (τ Qs )−1 }, (2)

where Qs = M QM. This implies p + q + 1 unknowns, compared to p + n + 1

for the traditional model and p + (n − p) + 1 = n + 1 for the RHZ model. This
dramatic reduction in dimension speeds computation considerably, allowing even
the largest areal datasets to be analyzed quickly (in minutes or hours rather than
days or weeks).
Since δ s are regression coefficients, one may be tempted to assign δ s a spherical
Gaussian prior instead of the abovementioned prior. This would be a mistake,
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 57

however, for (2) is not arbitrary (see Reich et al. (2006) and/or Hughes and
Haran (2013) for derivations) but is, in fact, very well suited to the task at hand.
Specifically, two characteristics of (2) discourage overfitting even when q is too
large for the dataset being analyzed. First, the prior variances are commensurate with
the spatial scales of the predictors in M. This shrinks toward zero the coefficients
corresponding to predictors that exhibit small-scale spatial variation. Additionally,
the correlation structure of (2) effectively reduces the degrees of freedom in the
smoothing component of the model.

2.3 Review of Multivariate CAR Models

A number of multivariate CAR (MCAR) models have been developed (Carlin and
Banerjee 2003; Gelfand and Vounatsou 2003; Jin et al. 2005; Martinez-Beneito
2013). These models have the same drawbacks as their univariate counterparts, but
of course entail even more burdensome computation. Thus it is desirable to develop
a SAMM for multivariate outcomes. We begin by reviewing the MCAR model that
is the multivariate analog of the traditional univariate CAR model described above.
Suppose we observe multiple outcomes at each areal unit and that each outcome
has its own regression component and collection of spatial effects. Specifically,
for j ∈ {1, . . . , J } we have outcomes y j = (y1j , . . . , ynj ) , design matrix Xj ,
regression coefficients β j , and spatial effects φ j = (φ1j , . . . , φnj ) . Then the
transformed conditional mean vectors are given by

gj (μj ) = Xj β j + φ j .

Now collect the φ j to form

= (φ 1 , . . . , φ J ) , and put
 $
1   −1
p(
| ) ∝ exp −
 ⊗ Q
,
2

where  is a J × J covariance matrix consisting of J (J + 1)/2 unknown variance

and covariance parameters, and ⊗ denotes the Kronecker product. The j th diagonal
entry of  is proportional to the variance of the spatial effects corresponding to the
j th outcome. The jj  off-diagonal entry is proportional to the covariance between
the j th and j  th spatial effects within an areal unit.
Should we require a different precision matrix for each φ j , the prior on
can be
generalized as
 $
1
p(
| ) ∝ exp −
 bdiag(R1 , . . . , RJ ) ( −1 ⊗ In )bdiag(R1 , . . . , RJ )
,
2

where bdiag(·) denotes a block diagonal matrix and Rj (j = 1, . . . , J ) is such that

Rj Rj = Qj .
58 D. Musgrove et al.

2.4 The Multivariate SAMM (MSAMM)

Recently, Bradley et al. (2015) introduced the Moran’s I (MI) prior, which is a
multivariate spatiotemporal model based on the SAMM. We introduce a multivariate
model that uses a similar prior but is strictly for spatial data. We call our model the
multivariate SAMM (MSAMM). The MSAMM serves as the foundation for the
zero-inflated count model that we focus on below in Sect. 4.
Construction of the MSAMM is of course analogous to construction of the
SAMM. For j ∈ {1, . . . , J }, let Pj = Xj (Xj Xj )−1 Xj , and let Mj be a matrix,
the columns of which are the first q eigenvectors of (In − Pj )A(In − Pj ). Denote
the prior precision matrix as Qsj = Mj QMj , and let Rsj be the upper Cholesky
triangle of Qsj so that Rsj Rsj = Qsj . Then the MSAMM can be specified as
%  &
gj (μj ) = gj E y j | β j , δ sj = Xj β j + Mj δ sj
  $
1
p( | ) ∝ exp −  R  −1 ⊗ Iq R ,
2

where  = (δ s1 , . . . , δ sJ ) and R = bdiag (Rs1 , . . . , RsJ ). Once again  is a

J × J covariance matrix consisting of J (J + 1)/2 unknown parameters. Thus the
MSAMM has J (p +q +(J +1)/2) unknowns, while the MCAR has J (p +n+(J +
1)/2). This is a considerable reduction so long as q  n. Moreover, the MSAMM,
like the SAMM, alleviates spatial confounding; permits simple, fast updating of the
spatial random effects; and yields a fast-mixing Markov chain.
In our application and simulation study, we used the same design matrix for all
dimensions, i.e., we used Xj = X for all j , which implies a single prior precision
matrix Qs . In this case, the above specification simplifies to
%  &
gj (μj ) = gj E y j | β j , δ sj = Xβ j + Mδ sj (3)
 $
1 
p( | ) ∝ exp −   −1 ⊗ Qs  .
2

In either case the precision matrix is invertible, and so the prior distribution is proper.

3 Hyperpriors and Computation for the MSAMM

Although using a truncated Moran basis dramatically reduces the time required to
draw samples from the posterior, and the space required to store those samples, this
approach does incur the substantial up-front burden of computing and eigendecom-
posing (In −Pj )A(In −Pj ). The efficiency of the former can be increased by storing
A in a sparse format and parallelizing the matrix multiplications. And we can more
efficiently obtain the desired basis vectors by computing only the first q eigenvectors
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 59

of (In − Pj )A(In − Pj ) instead of doing the full eigendecomposition. This can be

done using the Spectra library (Qiu 2017), for example.
ind
We use independent Gaussian priors for the regression coefficients: β j ∼
N (0, 10, 000 Ip ). There are several appealing choices for the prior distribution
of . We follow Huang and Wand (2013) and use the hierarchical half-t prior
distribution. The hierarchical half-t prior relies on an inverse-Wishart distribution
such that the diagonal elements of the scale matrix are given inverse-gamma prior
distributions. The prior scale of the inverse-Wishart is 2νV, which implies ν + J − 1
degrees of freedom. We let V = diag (1/v1 , . . . , 1/vJ ), where vj has an inverse-
gamma prior with shape 1/2 and scale 1/ζj2 . The hyperparameters are given values
of ν = 2 and ζj = 105 (j = 1, . . . , J ). This approach yields a conjugate prior
distribution. Alternatives include the inverse-Wishart prior, the covariance matrix
separation strategy (Barnard et al. 2000), and the LKJ prior (Lewandowski et al.
2009).
The full conditional distribution of the spatial effects  does not have a closed
form, and so we use a Metropolis update to draw samples from the posterior. When
a common design matrix is used for all dimensions, we can speed computation by
reparameterizing the spatial effects. Suppose we have  ∼ N (0,  ⊗Iq ). Let Rs be
the upper Cholesky triangle of Qs , and let Ws = R−1  −1
s so that Ws Ws = Qs . Then
it is easy to see that (IJ ⊗ Ws ) and  have the same distribution. We replace M
with Ms = MWs and work with  instead of . We provide details in an appendix.
We also show that a similar reparameterization holds when multiple design matrices
are required.
For our simulation study and application to the Census data, we used fixed-
width analysis (Flegal et al. 2008), in which samples are drawn until all Monte
Carlo standard errors are smaller than some pre-selected threshold. We used the
batchmeans package (Haran and Hughes 2016) for R (Ihaka and Gentleman
1996) to compute Monte Carlo standard errors. Two hundred thousand samples
were sufficient to ensure that all Monte Carlo standard errors were less than 0.05.
The MCMC estimates stabilized after approximately 100,000 iterations. The time
required to fit the MSAMM was approximately equal to the time required to fit
independent SAMMs.
Our software was written in R and C++ (Stroustrup 2013). Our use of C++ was
aided greatly through the use of the Rcpp package (Eddelbuettel and Francois
2011). Most of the numerical linear algebra was carried out using the Armadillo
C++ library (Sanderson 2010), which we accessed by way of the RcppArmadillo
package (Eddelbuettel and Sanderson 2014).

4 An MSAMM for Zero-Inflated Data

In this section we specialize the MSAMM to handle zero-inflated count data.

Zero-inflated count models are inherently two-component mixture models, where
one component is a point mass at 0 and the other is a discrete distribution. In
60 D. Musgrove et al.

recent years, many novel finite mixture models have been developed to incorporate
spatial dependencies. Alfó et al. (2009) used finite mixture models to analyze
multiple, spatially correlated, counts, where the dependence among outcomes is
modeled using a set of correlated random effects. Green and Richardson (2002)
developed a class of hidden Markov models in the spatial domain to analyze spatial
heterogeneity of count data on a rare phenomenon. Neelon et al. (2015) developed a
broad class of Bayesian two-part models for the spatial analysis of semicontinuous
data. Torabi (2016) proposed a hierarchical multivariate mixture generalized linear
model to simultaneously analyze spatial normal and non-normal outcomes. Zero-
inflated count models are often applied in non-spatial settings, e.g., in manufacturing
(Lambert 1992), where defective materials are rare and the number of defects is
assumed to follow a Poisson distribution, and in the hunger-for-bonus phenomenon
that occurs in risk assessment for filed insurance claims (Boucher et al. 2009).
Spatial zero-inflated count models have been applied to various types of data,
including animal sightings (Agarwal et al. 2002; Ver Hoef and Jansen 2007; Recta
et al. 2012), plant distribution (Rathbun and Fei 2006), tornado reports (Wikle and
Anderson 2003), and emergency room visits (Neelon et al. 2013).

4.1 Models for Zero-Inflated Counts

Two common approaches to modeling zero-inflated counts are the hurdle model
and the zero-inflated-Poisson (ZIP) model (Lambert 1992). For a hurdle model, the
outcome is 0 with probability 1 − π , and with probability π the outcome arose from
a zero-truncated Poisson (ZTP) distribution (Cohen 1960; Singh 1978). Formally,
the hurdle model is of the form

P(y = 0) = 1 − π
exp(−λ) λk
P(y = k) = π (k ∈ N : k ≥ 1).
1 − exp(−λ) k!

The ZIP model, on the other hand, is given by

P(y = 0) = (1 − π ) + π exp(−λ)
λk
P(y = k) = π exp(−λ) (k ∈ N : k ≥ 1).
k!
Let us compare and contrast the hurdle and ZIP models informally. Each model
can be viewed as comprising a binary process (the incidence process) and a counting
process (the prevalence process). For the hurdle model there is only one source of
zeros, namely the binary process. If the binary outcome is 0, no count is observed.
If the binary outcome is 1, a nonzero count is observed. The ZIP model differs in
that it posits two sources of zeros. If the binary outcome is 0, no count is observed.
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 61

If the binary outcome is 1, a (possibly 0) count is observed. This difference has a

practical effect: the hurdle model can accommodate both zero-inflation and zero-
deflation while the ZIP model can accommodate only zero-inflation (Neelon et al.
2013).
It may also be the case that one or the other of the models makes more sense
given the phenomenon of interest. Consider emergency room visits, for example,
as in Neelon et al. (2013). A hurdle model is appropriate here; the binary outcome
registers whether a given subject visits an emergency room while the count outcome
records the number of visits. As a second example, consider animal sightings (Agar-
wal et al. 2002; Recta et al. 2012). A type of animal may not be sighted in a given
region even though the animal is likely present in the region. A ZIP model is appro-
priate in this scenario. We adopt the hurdle model for the remainder of the paper.

4.2 A Spatial Poisson Hurdle Model Based on the MSAMM

In the context of our MSAMM, an appropriate hurdle model can be specified as

follows. Suppose we have n areal units, in which case our outcomes are y1 , . . . , yn .
Although the outcomes are univariate, we employ a bivariate MSAMM for the mean
structure, i.e., for the pairs of incidence probabilities and prevalence rates (πi , λi )
(i = 1, . . . , n). For the sake of simplicity, suppose that the same design matrix,
X, is appropriate for both the incidence process and the prevalence process. This
implies an MSAMM of the form given in (3) (as we showed in Sect. 2.3, this model
can easily be extended to accommodate different design matrices). Thus the linear
predictors for the ith areal unit can be specified as

ηi1 = x i β 1 + mi δ s1 (incidence process)

ηi2 = x i β 2 + mi δ s2 (prevalence process).

Using the logit and log link functions, respectively, gives

exp(ηi1 )
πi =
1 + exp(ηi1 )
λi = exp(ηi2 ),

where πi is the probability of incidence for the ith areal unit, and λi is the ZTP rate
for the ith areal unit. The within-unit covariance matrix  is of course 2 × 2 for this
model. Clearly, this model accommodates (1) spatial dependence among areal units,
and (2) dependence between the incidence process and the prevalence process. For
the case of a positive off-diagonal value in , the latter source of dependence implies
that a higher probability of incidence is associated with a higher prevalence rate. As
we will see in the next section, our MSAMM hurdle model’s ability to accommodate
consequential dependence between πi and λi permits improved inference and fit.
62 D. Musgrove et al.

4.3 Simulated Application of the MSAMM Hurdle Model

To assess the performance of our areal hurdle model, we carried out a simulation
study. We simulated data for the 2600 census block groups of the US state of Iowa
(Fig. 2). We included an intercept term and, as a covariate, the percentage of housing

1.0
0.9
0.8
0.7
0.6
(a) 0.5
0.4
0.3
0.2
0.1
0.0

38.0

32.6

27.1

21.7
(b)
16.3

10.9

5.4

0.0

Fig. 2 A single simulated zero-inflated dataset for the census block groups of Iowa. The propor-
tion of renters in each block group was used as a covariate. Panel (a) displays the probabilities of
incidence. Panel (b) displays the ZTP rates, where a given rate is nonzero only if the underlying
binary outcome is equal to 1
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 63

units occupied by renters (see Sect. 4.4). We used β 1 = (−1, 1) and β 2 = (2, −1) .
These values for β 1 indicate that block groups with a high percentage of renters
will be more likely to take nonzero values than block groups with a low percentage
of renters. The values of β 2 indicate that as the proportion of renters increases, the
ZTP rate decreases. We used 4 and 8 as the diagonal elements of , and we used
five different values for the within-unit correlation: ρ = 0, 0.2, 0.4, 0.6, 0.8.
We constructed Qs as detailed in Sect. 2.3 and used the same Qs for both
processes. Eigendecomposition of the Moran operator yielded 1000 basis vectors
exhibiting patterns of positive spatial dependence. We used the first q = 250
eigenvectors to construct M. This choice of q allowed the responses to exhibit
both small- and large-scale spatial variation. We then simulated spatial effects
 = (δ s1 , δ s2 ) from a zero-mean Gaussian distribution with covariance  ⊗ Q−1 s .
For the ith block group we simulated yi1 from the Bernoulli distribution with
success probability πi = logit−1 (x i β 1 + mi δ s1 ). Conditional on yi1 = 1, we drew
yi2 from the zero-truncated Poisson distribution with rate λi = exp(x i β 2 + mi δ s2 ).
Finally, we let yi = 0 if yi1 = 0, or yi = yi2 if yi1 = 1.
We analyzed 1000 simulated datasets for each of the five correlations. To assess
the importance of modeling the dependence within areal units, we applied both
the MSAMM and independent SAMMs to each dataset. Key results are shown in
Table 1. Extended results, including credible interval coverage rates, are included in
an appendix. We see that neglecting within-unit dependence leads to larger biases
and, for some parameters, larger mean squared errors, especially for larger values
of ρ.
Note that we did not compare the performance of our model to the performance
of one or more hurdle MCAR models, for three reasons. First, to the best of our
knowledge, no hurdle MCAR model has been implemented in software. Second, any
MCAR model is spatially confounded for the same reason that the univariate CAR
models are spatially confounded. And third, fitting any MCAR model would be
terribly burdensome computationally (with respect to both running time and storage
requirement) for the same reason that fitting any univariate CAR model would be
burdensome. Hence, pitting a hurdle MCAR model against our hurdle MSAMM
would have led to no new knowledge.

4.4 Application of the Hurdle Model to the Iowa Census Data

In this section we apply our areal hurdle model to address deletes from the 2010 US
Census within the state of Iowa. Recall that a delete is defined as an address that was
deleted from the base count because it did not correspond to a valid housing unit.
64 D. Musgrove et al.

Table 1 Results for the simulation study

MSAMM Independent SAMMs
Par. Truth Mean est. Bias MSE Mean est. Bias MSE
β11 −1 −0.998 0.002 0.006 −0.995 0.005 0.006
β12 1 0.994 −0.006 0.058 0.991 −0.009 0.057
β21 2 2.003 0.003 0.001 2.003 0.003 <0.001
β22 −1 −1.001 −0.001 0.009 −1.001 −0.001 0.009
σ12 4 3.970 −0.030 1.357 3.703 −0.297 1.395
σ22 8 8.274 0.274 0.939 8.153 0.153 0.847
ρ 0 <0.001 <0.001 0.014 – – –
β11 −1 −0.998 0.002 0.006 −0.995 0.005 0.006
β12 1 0.994 −0.006 0.058 0.991 −0.009 0.057
β21 2 2.004 0.004 <0.001 2.012 0.012 0.001
β22 −1 −1.000 <0.001 0.008 −1.003 −0.003 0.008
σ12 4 4.009 0.009 1.336 3.703 −0.297 1.395
σ22 8 8.248 0.248 0.933 8.112 0.112 0.838
ρ 0.2 0.201 0.001 0.013 – – –
β11 −1 −0.998 0.002 0.006 −0.995 0.005 0.006
β12 1 0.994 −0.006 0.058 0.991 −0.009 0.057
β21 2 2.004 0.004 <0.001 2.020 0.020 0.001
β22 −1 −1.000 <0.001 0.008 −1.004 −0.004 0.008
σ12 4 4.012 0.012 1.345 3.703 −0.297 1.395
σ22 8 8.229 0.229 0.945 8.071 0.071 0.836
ρ 0.4 0.405 0.005 0.012 – – –
β11 −1 −0.998 0.002 0.006 −0.995 0.005 0.006
β12 1 0.995 −0.005 0.058 0.991 −0.009 0.057
β21 2 2.005 0.005 <0.001 2.028 0.028 0.002
β22 −1 −1.000 <0.001 0.007 −1.007 −0.007 0.007
σ12 4 4.060 0.060 1.309 3.703 −0.297 1.395
σ22 8 8.238 0.238 0.943 8.018 0.018 0.827
ρ 0.6 0.604 0.004 0.010 – – –
β11 −1 −1.000 <0.001 0.006 −0.995 0.005 0.006
β12 1 0.996 −0.004 0.058 0.991 −0.009 0.057
β21 2 2.007 0.007 <0.001 2.036 0.036 0.002
β22 −1 −1.000 <0.001 0.007 −1.008 −0.008 0.007
σ12 4 4.165 0.165 1.263 3.703 −0.297 1.395
σ22 8 8.197 0.197 0.940 7.944 −0.056 0.824
ρ 0.8 0.793 −0.007 0.006 – – –
We analyzed 1000 simulated zero-inflated datasets for each value of ρ. MSE denotes mean squared
error
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 65

0.8
Fig. 3 Histogram of number
of address deletes for the
2600 block groups of Iowa
from the 2010 Census.
Approximately 75% of the
outcomes are zeros and 4% of

0.6
the outcomes are greater
than 10

Observed percent
0.4
0.2
0.0

0 1 2 3 4 5 6 7 8 9 10
Number of deletes

Figure 3 displays a histogram (truncated at 10) showing the zero-inflation present in

the data.
In the count component of the model, along with covariates, we introduced an
offset calculated via external standardization. Specifically, the offset is the natural
log of the total number of addresses in the block group to which a Census form was
delivered. We used the same covariates in the binary and count components of the
model. Each covariate was a proportion. We used the following covariates and an
intercept.
• RURAL_POP: proportion of population living outside of an urban area or urban
cluster
• OCCP_HU: proportion of housing units classified as the usual place of residence
of the individual or group living there
• RENTER_OCCP_HU: proportion of occupied housing units that are not owner
occupied, whether they are rented or occupied without payment of rent
• TEA_MAIL: proportion of addresses that received a Census form in the mail and
occupants were instructed to complete and return the form
• FIRST_FRM: proportion of addresses where the first form mailed was completed
and returned
As in the simulation study, we used the first 250 eigenvectors of the Moran operator,
and we applied both the MSAMM and independent SAMMs.
The results are shown in Table 2. First, we see that the correlation ρ was esti-
mated as 0.73, with a 95% highest posterior density (HPD) interval of (0.25, 0.91).
66 D. Musgrove et al.

Table 2 Iowa address delete results for the MSAMM versus independent SAMMs
MSAMM Independent SAMMs
Predictor/parameter Posterior mean 95% CI Posterior mean 95% CI
Intercept 7.23 (−4.43, −9.83) 7.48 (−4.93,10.53)
RURAL _ POP 2.72 (−2.18, −3.30) 2.70 (−2.13, −3.29)
OCCP _ HU 0.68 (−2.68, −4.60) 0.25 (−3.21, −3.61)
RENTER _ OCCP _ HU −3.51 (−5.73, −1.41) −3.60 (−6.01, −1.16)
TEA _ MAIL −6.95 (−7.54, −6.37) −6.98 (−7.60, −6.40)
FIRST _ FRMS −7.10 (−9.93, −4.28) −6.86 (−9.82, −3.86)
σ12 0.67 (−0.30, −2.04) 0.94 (−0.21, −2.52)
Intercept 3.99 (−3.50, −4.50) 3.92 (−3.07, −4.69)
RURAL _ POP −0.00 (−0.18, −0.15) 0.02 (−0.14, −0.18)
OCCP _ HU 2.52 (−1.07, −3.59) 2.00 (−0.48, −3.17)
RENTER _ OCCP _ HU −1.47 (−2.28, −0.68) −1.16 (−1.99, −0.16)
TEA _ MAIL −2.03 (−2.22, −1.83) −1.98 (−2.24, −1.73)
FIRST _ FRMS −6.11 (−7.39, −4.65) −5.50 (−7.00, −2.93)
σ22 7.74 (−4.99,11.47) 8.24 (−5.37,12.01)
ρ 0.73 (−0.25, −0.91) –
pD 258.85 172.31
DIC −7745 −7735
Results for the binary components of the models are shown in the top portion of the table. Results
for the count components are shown in the bottom portion. CI denotes credible interval

This suggests that the MSAMM, as opposed to independent SAMMs, is appropriate

for these data. This claim is further supported by the fact that the correlation model
yields a lower deviance information criterion (DIC) value (Spiegelhalter et al. 2002).
(A difference in DIC of 10 may not seem substantial but is, in fact, enormous if
relative likelihood (Burnham et al. 2011) is used as one’s basis for comparison:
exp(−10/2) = exp(−5) ≈ 0.007.)
Regarding the regression coefficients for the binary component, we see that all
but OCCP_HU have 95% credible intervals that exclude zero. This suggests that, in
Iowa, the proportion of occupied housing units is not predictive of the occurrence
of deletes within a block group. We see that RENTER_OCCP_HU, TEA_MAIL, and
FIRST _ FRM offer a “protective effect” against the occurrence of deletes while
RURAL _ POP is associated with an increased likelihood of deletes. We also see that
the variation in the spatial effects for the binary component appears to be small:
σ̂12 = 0.67 (0.30, 2.04).
The regression for the count component tells a somewhat different story. All coef-
ficients save RURAL_POP have 95% credible intervals that exclude zero, suggesting
that the proportion of a block group that is rural is not associated with the number
of deletes. Similar to the binary component, we see that RENTER_OCCP_HU,
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 67

TEA _ MAIL , and FIRST _ FRM offer a “protective effect” against the number of
deletes, but now OCCP_HU is associated with a greater number of deletes. Evi-
dently, the spatial process is not as smooth for the count component since σ̂22 =
7.74 (4.99, 11.47).
The results obtained using our areal hurdle model could be valuable to the
Census Bureau. When using the most recent covariate values obtained (such as from
official government surveys or administrative records), predictions using our model
could help to characterize deletes, which is an indication of demographic change
or stability in an area. The spatial component of our model can assist in designing
efficient and cost-effective address updating operations. For example, it can inform
Census Bureau personnel as to clusters of block groups that are candidates for
updating in a non-decennial Census setting. Focusing on adjacent regions within
a cluster will be advantageous over assessing sets of block groups that might have
only “stable” block groups as neighbors. Such clustering will not always be captured
accurately by non-spatial zero-inflated models.

5 Discussion and Future Work

Our proposed methods for handling multivariate and zero-inflated areal data
offer improved regression inference while greatly reducing computing time and
storage requirements. Our simulation study illustrates the benefit of accounting for
dependence within areal units as well as among areal units. This is not surprising:
in general, multivariate data call for multivariate methods.
The count distribution used for our model is the Poisson, which requires the
assumption of equi-dispersion. Of course, the data could be heavily over- or under-
dispersed, in which case other distributions could be developed in our MSAMM
setup, such as the negative binomial or the Conway–Maxwell–Poisson distribution.
Both of these distributions have an additional parameter that characterizes the
dispersion, which could possibly depend on spatially varying covariates. These
different models would be novel, but would require additional numerical work to
demonstrate how well they improve the fits.
Application of our areal Poisson hurdle model to zero-inflated Census data
provided a superior fit relative to that provided by independent univariate models,
at no extra computational cost. Most importantly, our methodology provides a
compelling framework for understanding dynamic features of the USA, which could
aid the planning of various Census Bureau operations. Moreover, our methodology
could be extended to handle additional data challenges faced by the Census Bureau.
For example, a spatiotemporal extension of our MSAMM could be useful for
analyzing data from historical databases being developed by the Census Bureau.
In such a model, time-dependent covariates could be viewed as driving the deletion
of housing units.
68 D. Musgrove et al.

Appendix: Supplementary Materials

Multivariate Spatial Effect Reparameterization

For the multivariate sparse areal mixed model (MSAMM), when the design matrices
are the same across multivariate outcomes, i.e., X1 = X2 = · · · = XJ , the first and
second stages can be written as
%  &
gj E y j | β j , δ sj = Xβ j + Mδ sj (j = 1, . . . , J )

p ( | ) = N 0,  ⊗ Q−1 s ,

 
where  = δ s1 , . . . , δ sJ , each δ sj is q × 1,  is the J × J covariance matrix, and
Qs is the q × q spatial precision matrix.
Computation can be eased considerably as follows. Let Rs be the upper Cholesky
triangle of Qs , and let Ws = R−1 such that Ws Ws = Q−1 s . Then, for  =
  
 
s  
ψ s1 , . . . , ψ sJ , each ψ sj is q × 1, and  |  ∼ N 0,  ⊗ Iq , we have that
(IJ ⊗ Ws )  and  have the same distribution conditional on . This is easy to see
since E {(IJ ⊗ Ws ) } = (IJ ⊗ Ws ) E () = 0 and
 
cov {(IJ ⊗ Ws ) } = (IJ ⊗ Ws )  ⊗ Iq (IJ ⊗ Ws )
=  ⊗ Q−1
s .

Hence, the model’s first and second stages can now be written as
%  &
gj E y j | β j , ψ sj = Xβ j + MWs ψ sj (j = 1, . . . , J )
 
p ( | ) = N 0,  ⊗ Iq .

Now suppose that X1 = X2 = · · · = XJ . Then we have

%  &
gj E y j | β j , δ sj = Xj β j + Mj δ sj
 '  (−1 
 −1
p ( | ) = N 0, R  ⊗ Iq R ,

 
where  = δ s1 , . . . , δ sJ , R = bdiag (Rs1 , . . . , RsJ ), and Rsj Rsj = Qsj , where
Rsj is the upper Cholesky triangle of Qsj . For ease of exposition, let J = 2 (the
following easily extends to the case when J > 2). The prior distribution of the
spatial effects can be written
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 69

⎡ ) *−1 ⎤
   R
δ s1 0 Rs1 0 s1 0
| ∼ N⎣ ,  −1 ⊗ Iq ⎦
δ s2 0 0 Rs2 0 Rs2

+   ,
0 Ws1 0   Ws1 0 
= N ,  ⊗ Iq ,
0 0 Ws2 0 Ws2

   −1
where Wsj = R−1 −1
sj (j = 1, 2), and we have used the fact that Rsj = Rsj .
Now, suppose we have
  $
ψ s1 0
| ∼ N ,  ⊗ Iq .
ψ s2 0

Using basic properties of the multivariate normal distribution, we have that

  )   *
Ws1 0 ψ s1 0 Ws1 0   Ws1 0 
| ∼ N ,  ⊗ Iq .
0 Ws2 ψ s2 0 0 Ws2 0 Ws2

Then, since
  
Ws1 0 ψ s1 Ws1 ψ s1
= ,
0 Ws2 ψ s2 Ws2 ψ s2

we can apply a reparameterization similar to the case where design matrices are
equivalent across the outcomes. Thus we can specify the first and second stages of
the model as
%  &
gj E y j | β j , ψ sj = Xj β j + Mj Wsj ψ sj
 
p ( | ) = N 0,  ⊗ Iq .

Extended Simulation Results

Table 3 provides complete results for our simulation study.

 70

Table 3 Extended results for our simulation study

MSAMM Independent SAMMs
Par. Truth Mean est. Bias (%) MSE Coverage Mean est. Bias (%) MSE Coverage
β11 −1 −0.998 0.002 (0.002) 0.006 0.948 −0.995 0.005 (0.005) 0.006 0.942
β12 1 0.994 −0.006 (0.006) 0.058 0.930 0.991 −0.009 (0.009) 0.057 0.931
β21 2 2.003 0.003 (0.001) 0.001 0.953 2.003 0.003 (0.002) <0.001 0.962
β22 −1 −1.001 −0.001 (0.001) 0.009 0.958 −1.001 −0.001 (0.001) 0.009 0.955
σ12 4 3.970 −0.030 (0.008) 1.357 0.945 3.703 −0.297 (0.074) 1.395 0.906
σ22 8 8.274 0.274 (0.034) 0.939 0.975 8.153 0.153 (0.019) 0.847 0.977
ρ 0 <0.001 <0.001 (–) 0.014 0.952 – – – –
β11 −1 −0.998 0.002 (0.002) 0.006 0.945 −0.995 0.005 (0.005) 0.006 0.942
β12 1 0.994 −0.006 (0.006) 0.058 0.929 0.991 −0.009 (0.009) 0.057 0.931
β21 2 2.004 0.004 (0.002) <0.001 0.959 2.012 0.012 (0.006) 0.001 0.944
β22 −1 −1.000 <0.001 (<0.001) 0.008 0.953 −1.003 −0.003 (0.003) 0.008 0.949
σ12 4 4.009 0.009 (0.002) 1.336 0.950 3.703 −0.297 (0.074) 1.395 0.906
σ22 8 8.248 0.248 (0.031) 0.933 0.970 8.112 0.112 (0.014) 0.838 0.974
ρ 0.2 0.201 0.001 (0.006) 0.013 0.954 – – – –
β11 −1 −0.998 0.002 (0.002) 0.006 0.943 −0.995 0.005 (0.005) 0.006 0.942
β12 1 0.994 −0.006 (0.006) 0.058 0.929 0.991 −0.009 (0.009) 0.057 0.931
β21 2 2.004 0.004 (0.002) <0.001 0.962 2.020 0.020 (0.010) 0.001 0.930
β22 −1 −1.000 <0.001 (<0.001) 0.008 0.954 −1.004 −0.004 (0.004) 0.008 0.951
σ12 4 4.012 0.012 (0.003) 1.345 0.949 3.703 −0.297 (0.074) 1.395 0.906
σ22 8 8.229 0.229 (0.029) 0.945 0.970 8.071 0.071 (0.009) 0.836 0.966
ρ 0.4 0.405 0.005 (0.011) 0.012 0.949 – – – –
β11 −1 −0.998 0.002 (0.002) 0.006 0.945 −0.995 0.005 (0.005) 0.006 0.942
D. Musgrove et al.
β12 1 0.995 −0.005 (0.005) 0.058 0.932 0.991 −0.009 (0.009) 0.057 0.931
β21 2 2.005 0.005 (0.003) <0.001 0.966 2.028 0.028 (0.014) 0.002 0.903
β22 −1 −1.000 <0.001 (<0.001) 0.007 0.957 −1.007 −0.007 (0.007) 0.007 0.950
σ12 4 4.060 0.060 (0.015) 1.309 0.940 3.703 −0.297 (0.074) 1.395 0.906
σ22 8 8.238 0.238 (0.030) 0.943 0.976 8.018 0.018 (0.002) 0.827 0.973
ρ 0.6 0.604 0.004 (0.006) 0.010 0.945 – – – –
β11 −1 −1.000 <0.001 (<0.001) 0.006 0.945 −0.995 0.005 (0.005) 0.006 0.942
β12 1 0.996 −0.004 (0.004) 0.058 0.933 0.991 −0.009 (0.009) 0.057 0.931
β21 2 2.007 0.007 (0.003) <0.001 0.963 2.036 0.036 (0.018) 0.002 0.853
β22 −1 −1.000 <0.001 (<0.001) 0.007 0.959 −1.008 −0.008 (0.008) 0.007 0.946
σ12 4 4.165 0.165 (0.041) 1.263 0.945 3.703 −0.297 (0.074) 1.395 0.906
σ22 8 8.197 0.197 (0.025) 0.940 0.979 7.944 −0.056 (0.007) 0.824 0.965
ρ 0.8 0.793 −0.007 (0.009) 0.006 0.953 – – – –
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . .
71
72 D. Musgrove et al.

References

Agarwal, D.K., Gelfand, A.E., Citron-Pousty, S.: Zero-inflated models with application to spatial
count data. Environ. Ecol. Stat. 9(4), 341–355 (2002)
Alfó, M., Nieddu, L., Vicari, D.: Finite mixture models for mapping spatially dependent disease
counts. Biom. J. 51(1), 84–97 (2009). http://dx.doi.org/10.1002/bimj.200810494
Assunção, R., Krainski, E.: Neighborhood dependence in Bayesian spatial models. Biom. J. 51(5),
851–869 (2009)
Barnard, J., McCulloch, R., Meng, X.L.: Modeling covariance matrices in terms of standard
deviations and correlations, with application to shrinkage. Stat. Sin. 10(4), 1281–1312 (2000)
Besag, J., Kooperberg, C.: On conditional and intrinsic autoregression. Biometrika 82(4), 733–746
(1995)
Boots, B., Tiefelsdorf, M.: Global and local spatial autocorrelation in bounded regular tessellations.
J. Geogr. Syst. 2(4), 319 (2000)
Boucher, J.P., Denuit, M., Guillen, M.: Number of accidents or number of claims? An approach
with zero-inflated Poisson models for panel data. J. Risk Insur. 76(4), 821–846 (2009)
Bradley, J.R., Holan, S.H., Wikle, C.K.: Multivariate spatio-temporal models for high-dimensional
areal data with application to longitudinal employer-household dynamics. Ann. Appl. Stat. 9(4),
1761–1791 (2015)
Burnham, K.P., Anderson, D.R., Huyvaert, K.P.: AIC model selection and multimodel inference in
behavioral ecology: some background, observations, and comparisons. Behav. Ecol. Sociobiol.
65(1), 23–35 (2011)
Carlin, B.P., Banerjee, S.: Hierarchical multivariate CAR models for spatio-temporally correlated
survival data (with discussion). In: Bayarri, M., Berger, J., Bernardo, J., Dawid, A., Heckerman,
D., Smith, A., West, M. (eds.), Bayesian Statistics 7, pp. 45–63. Oxford University Press, New
York (2003)
Clayton, D., Bernardinelli, L., Montomoli, C.: Spatial correlation in ecological analysis. Int. J.
Epidemiol. 22(6), 1193–1202 (1993)
Cohen, A.C.: Estimating the parameter in a conditional Poisson distribution. Biometrics 16(2),
203–211 (1960)
Cook, T., Norwood, J., Cork, D., Panel to Review the 2010 Census, Committee on National
Statistics, Division of Behavioral and Social Sciences and Education, National Research
Council: Change and the 2020 Census: Not Whether But How. National Academies Press,
Washington, D.C. (2011)
Donoho, D.L., Elad, M.: Optimally sparse representation in general (nonorthogonal) dictionaries
via 1 minimization. Proc. Natl. Acad. Sci. 100(5), 2197–2202 (2003)
Eddelbuettel, D., Francois, R.: Rcpp: Seamless R and C++ integration. J. Stat. Softw. 40(8), 1–18
(2011)
Eddelbuettel, D., Sanderson, C.: RcppArmadillo: Accelerating R with high-performance C++
linear algebra. Comput. Stat. Data Anal. 71, 1054–1063 (2014)
Flegal, J.M., Haran, M., Jones, G.L.: Markov chain Monte Carlo: can we trust the third significant
figure? Stat. Sci. 23(2), 250–260 (2008)
Gelfand, A.E., Vounatsou, P.: Proper multivariate conditional autoregressive models for spatial data
analysis. Biostatistics 4(1), 11–15 (2003)
Green, P.J., Richardson, S.: Hidden Markov models and disease mapping. J. Am. Stat. Assoc.
97(460), 1055–1070 (2002). https://doi.org/10.1198/016214502388618870
Griffith, D.A.: Spatial Autocorrelation and Spatial Filtering: Gaining Understanding Through
Theory and Scientific Visualization. Springer, Berlin (2003)
Haran, M., Hughes, J.: batchmeans: consistent batch means estimation of Monte Carlo standard
errors. Denver (2016)
Haran, M., Hodges, J., Carlin, B.: Accelerating computation in Markov random field models for
spatial data via structured MCMC. J. Comput. Graph. Stat. 12(2), 249–264 (2003)
A Sparse Areal Mixed Model for Multivariate Outcomes, with an Application. . . 73

Haran, M., Tierney, L.: On automating Markov chain Monte Carlo for a class of spatial models.
Preprint (2012). arXiv:12050499
Hodges, J., Reich, B.: Adding spatially-correlated errors can mess up the fixed effect you love.
Am. Stat. 64(4), 325–334 (2010)
Huang, A., Wand, M.: Simple marginally noninformative prior distributions for covariance
matrices. Bayesian Anal. 8(2), 439–452 (2013)
Hughes, J., Haran, M.: Dimension reduction and alleviation of confounding for spatial generalized
linear mixed models. J. R. Stat. Soc. Ser. B Stat. Methodol. 75(1), 139–159 (2013)
Ihaka, R., Gentleman, R.: R: a language for data analysis and graphics. J. Comput. Graph. Stat. 5,
299–314 (1996)
Jin, X., Carlin, B.P., Banerjee, S.: Generalized hierarchical multivariate CAR models for areal data.
Biometrics 61(4), 950–961 (2005)
Knorr-Held, L., Rue, H.: On block updating in Markov random field models for disease mapping.
Scand. J. Stat. 29(4), 597–614 (2002)
Lambert, D.: Zero-inflated Poisson regression, with an application to defects in manufacturing.
Technometrics 34(1), 1–14 (1992)
Leroux, B.G., Lei, X., Breslow, N.: Estimation of disease rates in small areas: a new mixed model
for spatial dependence. Inst. Math. Appl. 116, 179–191 (2000)
Lewandowski, D., Kurowicka, D., Joe, H.: Generating random correlation matrices based on vines
and extended onion method. J. Multivar. Anal. 100(9), 1989–2001 (2009)
Martinez-Beneito, M.A.: A general modelling framework for multivariate disease mapping.
Biometrika 100(3), 539–553 (2013)
Moran, P.: Notes on continuous stochastic phenomena. Biometrika 37(1/2), 17–23 (1950)
Neelon, B., Ghosh, P., Loebs, P.F.: A spatial Poisson hurdle model for exploring geographic
variation in emergency department visits. J. R. Stat. Soc. Ser. A Stat. Soc. 176(2), 389–413
(2013)
Neelon, B., Zhu, L., Neelon, S.E.B.: Bayesian two-part spatial models for semicontinuous data
with application to emergency department expenditures. Biostatistics 16(3), 465–479 (2015)
Qiu, Y.: Spectra: sparse eigenvalue computation toolkit as a redesigned ARPACK. http://spectralib.
org (2017)
Rathbun, S.L., Fei, S.: A spatial zero-inflated Poisson regression model for oak regeneration.
Environ. Ecol. Stat. 13(4):409–426 (2006)
Recta, V., Haran, M., Rosenberger, J.L.: A two-stage model for incidence and prevalence in point-
level spatial count data. Environmetrics 23(2), 162–174 (2012)
Reich, B., Hodges, J., Zadnik, V.: Effects of residual smoothing on the posterior of the fixed effects
in disease-mapping models. Biometrics 62(4), 1197–1206 (2006)
Sanderson, C.: Armadillo: an open source C++ linear algebra library for fast prototyping and
computationally intensive experiments. Technical Report; NICTA (2010)
Singh, J.: A characterization of positive Poisson distribution and its statistical application. SIAM
J. Appl. Math. 34(3), 545–548 (1978)
Spiegelhalter, D.J., Best, N.G., Carlin, B.P., Van Der Linde, A.: Bayesian measures of model
complexity and fit. J. R. Stat. Soc. Ser. B Stat. Methodol. 64(4), 583–639 (2002)
Stroustrup, B.: The C++ Programming Language. Pearson Education, New Jersey (2013)
Tiefelsdorf, M., Griffith, D.A.: Semiparametric filtering of spatial autocorrelation: the eigenvector
approach. Environ. Plan. A 39(5), 1193 (2007)
Torabi, M.: Hierarchical multivariate mixture generalized linear models for the analysis of spatial
data: an application to disease mapping. Biom. J. 58(5), 1138–1150 (2016)
U.S. Census Bureau: 2020 Census operational plan: a new design for the 21st century (2015)
Ver Hoef, J.M., Jansen, J.K.: Space-time zero-inflated count models of harbor seals. Environ-
metrics 18(7), 697–712 (2007)
Wall, M.: A close look at the spatial structure implied by the CAR and SAR models. J. Stat. Plan.
Inference 121(2), 311–324 (2004)
74 D. Musgrove et al.

Wikle, C.K., Anderson, C.J.: Climatological analysis of tornado report counts using a hierarchical
Bayesian spatiotemporal model. J. Geophys. Res. Atmos. (1984–2012) 108(D24), 1–15 (2003).
https://doi.org/10.1029/2002JD002806
Young, D.S., Raim, A.M., Johnson, N.R.: Zero-inflated modelling for characterizing coverage
errors of extracts from the US Census Bureau’s Master Address File. J. R. Stat. Soc. Ser. A
Stat. Soc. 180(1), 73–97 (2017)
Wavelet Kernels for Support Matrix
Machines

Edgard M. Maboudou-Tchao

1 Introduction

Binary classification is a supervised machine learning technique that classifies

the elements of a certain set into two groups based on a classification rule.
Classification problems occur in different areas such as text categorization, web
page classification, news classification, spam mail filtering, and discriminating
tumorous genes from non-tumorous genes. Support vector machine (SVM) (Cortes
and Vapnik 1995) is a powerful method based on statistical learning theory and has
been proven to perform better than existing methods in many aspects.
Typical classification methods use vectors as input samples. In the case that input
samples are second-order tensors or matrices, they need to be vectorized first before
applying classical classification techniques. This practice will destroy the structure
information of the data matrix as well as the correlation between the variables. Also,
the dimension of the resulting matrix vectorized will be large and this can cause a
dimensionality issue. Consequently, computation time will increase significantly.
Support matrix machine was introduced for binary classification of matrices (Shi
and Zhang 2009). Luo et al. (2015) proposed a penalized support matrix machine by
using a spectral elastic net regularization, which combines the Frobenius norm and
nuclear norm to constrain the regression matrix. Maboudou-Tchao (2017) proposed
support matrix data description (SMDD) for one-class classification of matrices. Xia
and Fan (2016) suggested a penalized least squares support matrix based on bilevel
programming by using both the Frobenius norm and nuclear norm. Zheng et al.
(2017) proposed sparse support matrix by regularizing a combination of nuclear
norm and 1 norm of the regression matrix.

E. M. Maboudou-Tchao ()
Department of Statistics, University of Central Florida, Orlando, FL, USA
e-mail: edgard.maboudou@ucf.edu

© Springer Nature Switzerland AG 2019 75

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_4
76 E. M. Maboudou-Tchao

In case of nonlinear classification problem, the input space is mapped into a

higher dimensional feature space so that the problem becomes linearly separable.
However, the mapping function does not have to be known. Instead, SVMs use the
kernel trick, which actually inverts the chain of processes by choosing a kernel rather
than choosing a mapping first. A necessary and sufficient condition for the kernel to
be a valid kernel is to satisfy the condition given by Mercer’s theorem (Mercer
1909). Many Mercer kernels are available, including Gaussian and polynomial
kernels. Recent studies proposed Mercer kernels based on wavelet techniques.
Zhang et al. (2004) used Morlet wavelet kernel as kernel for SVM. Wu and Zhao
(2006) also used Morlet wavelet kernel for least squares SVM.
Support matrix machine (SMM) is a matrix version of SVM and is based on the
matrix space. It accepts directly a matrix as inputs without the need of vectorization.
By constructing a classifier in the matrix space, the data structure information
is retained and it helps overcome the overfitting problem encountered mostly in
vector-based learning. This paper will discuss first support matrix machine without
using regularization, support matrix regression, and then extend wavelet kernels
of SVMs to wavelet kernels for support matrix machines. A new support matrix
machine learning algorithm can therefore be built using these new wavelet kernels
for matrices.

2 Overview of Support Matrix Machine (SMM)

The standard support vector machines (SVM) aim at finding the optimal hyper-
plane that maximizes the margin between two classes. SVM are solved using
quadratic programming methods. SVM are easily extended to accept matrix as
input. For the two-class classification problem, let the training set be D =
((X1 , y1 ), . . . , (XN , yN )) ∈ (X × {−1, 1})N , where N is the number of matrices
and X ⊆ Rn ⊗ Rp is an original matrix input space. Rn and Rp are two vector
spaces. SMM consists in solving the following primal problem:

1
N
minimize tr(W W) + C ξj ,
W 2
j =1
  (1)
subjectto yj tr(W ϕ(Xj )) + b ≥ 1 − ξj , j = 1, 2, . . . , N
ξj ≥ 0, j = 1, 2, . . . , N.

where W ∈ R n×p is the matrix of regression coefficients, tr(.) is the trace operator,
ξj are the slack variables, the parameter C > 0 is introduced to control the influence
of the slack variables, and ϕ is a function mapping data to a higher dimensional
Hilbert space.
Wavelet Kernels for Support Matrix Machines 77

This optimization problem is convex with respect to the primal variables W, b,

and ξ since each of the f0 , f1 , . . . , fN are convex where

1
N
f0 (W, b, ξ ) = tr(W W) + C ξj
2
j =1

and
 
fj (W, b, ξ ) = 1 − ξj − yj tr(W ϕ(Xj )) + b ≤ 0, j = 1, . . . , N.

Next, since the inequality constraints

 are affine and there always exist some
ξj ≥ 0 and yj tr(W ϕ(Xj )) + b ≥ 1 − ξj , Slater’s conditions are met and thus
strong duality holds. It follows that the duality gap is zero and the optimal values
of the primal and dual problems are equal. Consequently, it ensures that the original
problem (primal) can be solved through the Lagrange dual problem, which is usually
easier to solve than the primal.
The Lagrangian of the given problem is

N
-   .
N
L(W, b, α, ξ ) = tr(W W) + C ξj − αj yj tr(W ϕ(Xj )) + b − 1 + ξj − γj ξj
2
j =1 j =1 j =1

N
  

N
 
= tr(W W) + αj 1 − yj tr(W ϕ(Xj )) + b + ξj C − αj − γj ,
2
j =1 j =1
(2)
where αj and γj are positive Lagrange multipliers. To construct the dual problem,
we need to determine the optimal W, ξ , and b in terms of the dual variables. We
achieve this by differentiating the Lagrangian with respect to the primal variables.

∂L
N
= 0 =⇒ W − αj yj ϕ(Xj ) = 0,
∂W
j =1

∂L
N
= 0 =⇒ αj yj = 0,
∂b
j =1

∂L
= 0 =⇒ C − αj − βj = 0,
∂ξj

So, the key takeaways are


1. W = N j =1 αj yj ϕ(Xj ),
N
2. j =1 αj yj = 0,
3. βj = C − αj
78 E. M. Maboudou-Tchao

It follows now that the Kuhn–Karush–Tucker (KKT) conditions are

(i) Primal feasibility:
 
1 − ξj − yj tr(W ϕ(Xj )) + b ≥ 0

and

ξj ≥ 0,

(ii) Dual feasibility:

αj ≥ 0

and

γj ≥ 0,

(iii) Complementary slackness:

-  .
αj 1 − ξj − yj tr(W ϕ(Xj )) + b = 0

and

γj ξj = 0.

Now, from the first equation of the complementary slackness condition, the
objects for which αj = 0 are not on the margin and do not impact the value of W.
On the other hand, the objects for which αj > 0 do impact the value of W. These
matrices Xj corresponding to αj > 0 are the support matrices. The support matrices
that correspond to matrices located on the decision boundary, with 0 < αj < C,
are the margin support matrices. The other support matrices, with αj = C, are the
non-margin support matrices.
The next step is to maximize the dual problem. Plugging W into the Lagrangian
L and taking into account that βj = C − αj , the dual problem becomes

N
1

N
Maximize αj − yi yj αi αj tr(ϕ(Xi ) ϕ(Xj )),
α 2
j =1 i,j =1

subject to 0 ≤ αj ≤ C ∀j, (3)

N
αj yj = 0.
j =1

Now, if we let K(X, Y) represent the inner product tr(ϕ(X) ϕ(Y)) in a higher
dimensional space, the dual problem becomes
Wavelet Kernels for Support Matrix Machines 79

N
1

N
Maximize αj − yi yj αi αj K(Xi , Xj ),
α 2
j =1 i,j =1

subject to 0 ≤ αj ≤ C ∀j, (4)

N
αj yj = 0.
j =1

The optimization problem is formulated as a strongly convex quadratic problem

(QP) whose dual is also a QP. This can be solved easily using any quadratic
programming software.
Once the optimal values αj∗ are obtained, the optimal matrix W∗ is found by
plugging αj∗ into the equation of W, that is

N
W∗ = αj∗ yj ϕ(Xj ). (5)
j =1

The next step is to evaluate the offset b. b can be found by using a support matrix
Xj and the complementary slackness condition. Alternatively, this can also be
achieved using the set of all support matrices by finding an average over all support
matrices as
⎛ ⎞
1
⎝

b= yi − αj yj K(Xi , Xj )⎠ .
∗
(6)
Ns
i∈S j ∈S

Each new matrix point X0 is classified by evaluating

  N 

N
 


y0 = sgn αi yi tr ϕ(Xi ) ϕ(X0 ) + b = sgn αi yi K(Xi , X0 ) + b .
i=1 i=1
(7)

3 Support Matrix Regression (SMR)

Support matrix machines can easily be applied to matrix regression problems by

using an alternative loss function (Smola 1996). The key for that loss function is
to have a distance measure. The commonly used loss functions are the quadratic
loss function based on the conventional least squares error criterion, Laplacian loss
function that is less sensitive to outliers than the quadratic loss function, polynomial
loss function, piecewise polynomial loss function, Huber loss function that is a
robust loss function with optimal properties when the underlying distribution of the
80 E. M. Maboudou-Tchao

data is unknown, and the -insensitive loss function. The first three loss functions
do not produce sparseness in the support vectors. To address that issue (Vapnik
1995) suggested the -insensitive loss function as an approximation to Huber’s loss
function that enables a sparse set of support vectors to be obtained. The -insensitive
loss function is defined as:
)
0 if |f (x) − y| < 
L (y) = (8)
|f (x) − y| −  otherwise.

3.1 Basic Idea

Let the training set be D = ((X1 , y1 ), . . . , (XN , yN )) ∈ (X × R), where N is the

number of matrices and X ⊆ Rn ⊗ Rp is an original matrix input space. Rn and
Rp are two vector spaces. SMR consists to find a function f (x) that has at most 
deviation from the observed yi for all the training data, and at the same time it is as
flat as possible. f (x) will take the form:

f (x) = tr(W ϕ(Xj )) + b (9)

where W ∈ X , tr(.) is the trace operator, and b ∈ R.

This consists in solving the following convex optimization primal problem:

1
 N
minimize tr(W W) + C ξj+ + ξj− ,
W 2
j =1
 
subject to yj − tr(W ϕ(Xj )) + b ≤  + ξj+ , (10)
 
tr(W ϕ(Xj )) + b − yj ≤  − ξj− ,

ξj+ , ξj− ≥ 0, j = 1, 2, . . . , N.

where C > 0 is a pre-specified value and determines the trade-off between the
flatness of f and the amount up to which deviations larger than  are tolerated, and
ξj+ , ξj− are slack variables representing upper and lower constraints on the outputs
of the system.

3.2 Lagrange Dual Problems

Using similar arguments to the previous section, Slater’s conditions are met and thus
strong duality holds. It follows that the duality gap is zero and the optimal values
of the primal and dual problems are equal. Consequently, it ensures that the original
Wavelet Kernels for Support Matrix Machines 81

problem (primal) can be solved through the Lagrange dual problem, which is usually
easier to solve than the primal. To solve the primal problem (10), we construct the
Lagrangian by using Lagrange multipliers. The Lagrangian is

1
 N
 N
Lp = tr(W W) + C ξj+ + ξj− − αi+  + ξj+ − yj + tr(W ϕ(Xj )) + b
2
j =1 j =1

N 
N 
− αi−  − ξj− + yj − tr(W ϕ(Xj )) − b − ηj+ ξj+ + ηj− ξj−
j =1 j =1
(11)

It follows from the saddle point condition that the partial derivatives of Lp with
respect to the primal variables (W, b, ξj+ , ξj− ) have to vanish for optimality.

∂Lp
N
= 0 =⇒ W = (αj+ − αj− )ϕ(Xj ), (12)
∂W
j =1

∂Lp
N
= 0 =⇒ (αj+ − αj− ) = 0, (13)
∂b
j =1

∂Lp
= 0 =⇒ C − αj+ − ηj+ = 0. (14)
∂ξj+

∂Lp
= 0 =⇒ C − αj− − ηj− = 0. (15)
∂ξj−

Substituting these in the Lagrangian (11) yields the dual problem:

1
+
N
Maximize − (αi − αi− )(αj+ − αj− )tr(ϕ(Xi ) ϕ(Xj ))
α + ,αj− 2
i,j =1

N
+ αj+ (yj − ) − αj− (yj + ),
j =1 j =1 (16)
subject to 0 ≤ αj+ , αj− ≤ C ∀j,

N
(αj+ − αj− ) = 0.
j =1
82 E. M. Maboudou-Tchao

The KKT conditions are

(i) Primal feasibility:

yj − tr(W ϕ(Xj )) − b −  − ξj+ ≤ 0, (17)

tr(W ϕ(Xj )) + b − yj −  − ξj− ≤ 0, (18)

ξj+ ≥ 0 and ξj− ≥ 0. (19)

(ii) Dual feasibility:

αj+ ≥ 0 and αj− ≥ 0 (20)

ηj+ ≥ 0 and ηj− ≥ 0. (21)

(iii) Complementary slackness:


αj+ yj − tr(W ϕ(Xj )) − b −  − ξj+ = 0, (22)

αj− tr(W ϕ(Xj )) + b − yj −  − ξj− = 0, (23)

(C − αj+ )ξj+ = 0 and (C − αj− )ξj− = 0. (24)

From the KKT conditions, some important conclusions can be made:

• Samples (Xj , yj ) with αj+ = C or αj− = C lie outside the -insensitive tube
around f (x).
• αj+ αj− = 0 meaning that there can never be a set of dual variables αj+ , αj− which
are both nonzero.
• For αj+ , αj− ∈ (0, C), the slack variables ξj+ or ξj− would correspondingly be
zero.
By looking at the complementary slackness conditions (Eqs. (22) and (23)), the
matrices Xj with αj+ = 0 or αj− = 0 are not needed. Next, the matrices Xj , for
which the Lagrange multipliers are nonzero, are needed to determine W and are
called support matrices (SM).
To compute the bias term b, one has to choose a support matrix Xs . Then

b = ys −  − tr(W ϕ(Xs )), αs+ ∈ (0, C), (25)

b = ys +  − tr(W ϕ(Xs )), αs− ∈ (0, C). (26)

So for a new observation Z, its prediction is obtained by using the regression

function given by

 
f (Z) = (αj+ − αj− )tr ϕ(Xi ) ϕ(Xj ) + b. (27)
SMs
Wavelet Kernels for Support Matrix Machines 83

4 Non-spherical Decision Boundaries

It is not very clear in the SMM methods proposed by Luo et al. (2015) and Zheng
et al. (2017), the kernel function used. We propose one alternative of choice for
the kernel function. The optimization problem (16) only involves the patterns ϕ(X)
through the computation of inner products in feature space. There is no need to
compute the features ϕ(X) when one knows how to compute the dot products
directly. Instead of actually mapping each instance to a higher dimensional space
using a mapping function ϕ, Boser et al. (1992) propose to directly choose a
kernel function K(X, Y) that represents an inner product tr(ϕ(X) ϕ(Y)) in some
unspecified high dimensional space.
The key idea of the kernel technique, or the so-called kernel trick, is to invert the
chain of arguments, i.e., choose a kernel K rather than a mapping before applying
a learning algorithm. It is clear that not any symmetric function K can serve as a
kernel. The necessary and sufficient conditions of K : X × X → R to be a kernel
are given by Mercer’s theorem.
Theorem 4.1 (Mercer’s Theorem) Suppose K is a symmetric function such that
the integral operator
!
(TK f )(.) = K(., x)f (x) dx
X

is positive semidefinite, that is,

! !
K(y, x)f (x)f (y) dx dy ≥ 0,
X X

for all integrable functions f . Let ψi ∈ L2 (X ) be an eigenfunction

/ of Tk associated
with the eigenvalue λi ≥ 0 and normalized such that ||ψi ||2 = X ψi2 (x) dx = 1,
i.e.,
!
∀x ∈ X , K(x, y)ψi (y) dy = λi ψi (x).
X

Then
1. λi ∈ 1 , i ∈ N
2. ψi ∈ L∞ (X )
3. K can be expanded in a uniformly convergent series, i.e.,
∞

K(x, y) = λi ψi (x)ψi (y)

i=1

holds for all x, y ∈ (X).

84 E. M. Maboudou-Tchao

Mercer’s theorem not only gives necessary and sufficient conditions for K to be
a kernel, but also suggests a constructive way of obtaining features φi from a given
kernel K.
Theorem 4.2 (Mercer Kernels) The function K : X × X → R is a Mercer kernel
if, and only if, for each l ∈ N and X = (x1 , . . . , xl ) ∈ X l , the l × l matrix K =
 l
K(xi , xj ) i,j =1 is positive semidefinite.
Theorem 4.3 (Mercer Condition, Mercer 1909) The symmetry function K(x, y)
/ function if and only if: for all function f = 0 which satisfies the
is a valid kernel
condition of X f 2 (x)dx < ∞, we need to satisfy the condition:
! !
K(y, x)f (x)f (y) dx dy ≥ 0. (28)
X X

If we can find a Mercer kernel K that computes an inner product in the feature
space F we are interested
 in, we can use the kernel evaluations K(X, Y) to replace
the inner products tr ϕ(X) ϕ(Y) in the LS-SMM algorithm. Note that obtaining a
Mercer kernel in a matrix space is not as easy as in a vector space.

4.1 Wavelet Kernels for Support Matrix Machines

We will construct wavelet kernels that are admissible support matrix kernels, i.e.,
satisfy Mercer conditions. The support matrix kernel
 function can be described as
the inner product of two matrices, K(X, Y) = tr ϕ(X) ϕ(Y) . Instead of working
with matrices, we will use vectors instead. The support vector kernel function can
also be described as translation-invariant kernels such as K(x, y) = K(x − y)
(Burges 1999). A function is an admissible support vector kernel function if it
satisfies the condition of Mercer (Theorem 4.3). However, it is very challenging
to decompose translation-invariant functions as the product of two functions and
then prove that they satisfy Mercer condition. So the next result gives us necessary
and sufficient conditions for translation-invariant kernels to be admissible support
vector kernels.
Theorem 4.4 (Smola et al. 1998; Burges 1999) The translation-invariant kernel
function is an admissible support vector kernel function if and only if the Fourier
transform of k(x) satisfies
!
−m
F [k(ω)] = (2π ) 2 exp (−j (ωx)) k(x)dx ≥ 0. (29)
Xm
Wavelet Kernels for Support Matrix Machines 85

The goal of wavelet analysis is to approximate a function by a family of functions

generated by dilations and translations of a function (x), called the base wavelet
or mother wavelet. The wavelet function group is defined as:

x−c
 a,c (x) = (a)−1/2  (30)
a

where a ≥ 0 and c, x ∈ R. a is the dilation factor and c is a translation factor.

If the wavelet function of one dimension is (x), using tensor theory (Zhang
and Benveniste 1992), the multidimensional wavelet function can be defined as:


n 
p
 m (X) = (xij ) (31)
i=1 j =1

% &
where X = xij i,j ∈ Rn ⊗ Rp is a matrix with entries xi,j .
We can build the admissible kernel function for matrix as
Theorem 4.5 Let  be a base wavelet or mother wavelet, let a ≥ 0 be the dilation,
and c ∈ R be the translation. If X, Y ∈ Rn ⊗ Rp are two matrices, then the wavelet
kernels for matrices are


n 
p  
xij − c yij − c
K(X, Y) =   . (32)
a a
i=1 j =1

Proof We just need to show that the wavelet kernels for matrices satisfy the
condition of Mercer, i.e., are admissible support matrix kernels.
First, let vec(X) = [x11 , . . . , xnp ] and vec(Y) = [y11 , . . . , ynp ] , and let set
x = [x1 , . . . , xN ] = vec(X) and y = [y1 , . . . , yN ] = vec(Y), where N = n × p,
then


n 
p  
xij − c yij − c
K(X, Y) =  
a a
i=1 j =1


N  
xi − c yi − c
=  
a a
i=1

= K(x, y).
86 E. M. Maboudou-Tchao

Now, ∀f ∈ L2 (RN ),
!! ! 
N  ! 
N 
xi − c yi − c
K(x, y)f (x)f (y) dx dy =  dx  dy
RN ⊗RN RN a RN a
i=1 i=1
!  2

N
xi − c
=  dx ≥ 0.
RN a
i=1



Therefore, K(X, Y) satisfies Mercer condition and is admissible support matrix
kernel. Consequently, it follows that we can build translation-invariant kernels as
follows:


n 
p 
xij − yij
K(X, Y) =  . (33)
a
i=1 j =1

A necessary and sufficient condition for translation-invariant kernels to be

admissible is to satisfy the condition of Mercer (Theorem 4.4).

Mexican Hat Wavelet Kernel for Support Matrix

Now, we give an existing wavelet kernel function, the Mexican hat wavelet or
Sombrero wavelet, that can be used to construct translation-invariant wavelet
kernels. Note that the Mexican hat wavelet is sometimes called Marr wavelet and its
mother wavelet is
 2
x
(x) = (1 − x 2 ) exp − . (34)
2

Figure 1 represents a 2-D plot of the Mexican Hat wavelet kernel function.
Theorem 4.6 Let a ≥ 0, if X, Y ∈ Rn ⊗ Rp are two matrices, then the Mexican
Hat wavelet kernel function for matrices is defined as:
   

n 
p
xij − yij 2
(xij − yij )2
K(X, Y) = 1− exp − , (35)
a 2a 2
i=1 j =1

and is an admissible support matrix kernel.

Wavelet Kernels for Support Matrix Machines 87

Mexican Hat wavelet kernel

1.0
0.5
K(X)
0.0
−0.5

−4 −2 0 2 4
X

Fig. 1 Mexican Hat wavelet kernel function

Proof According to Theorem 4.4, one just needs to show that the Fourier transform
of Mexican hat wavelet is nonnegative. Let x = [x1 , . . . , xN ] = vec(X) with N =
n × p, then it follows that


N x
i
K(x) = 
a
i=1
N   x 2 
 i 1  xi 2
= 1− exp − .
a 2 a
i=1
88 E. M. Maboudou-Tchao

The integral term

!
I= exp(−j (ωx))K(x)dx
RN
! N 
  
  x 2 1  xi 2
i
= exp(−j (ωx)) 1− exp − dx
RN a 2 a
i=1
N !
   
 ∞  xi x2 x2
= exp −j aωi 1 − i2 exp − i 2 dxi
a a 2a
i=1 ∞
N !
   
 ∞  xi x2 x2 x
i
=a N
exp −j aωi 1 − i2 exp − i 2 d
−∞ a a 2a a
i=1
N !
   
 ∞ x2 x2 xi x
i
=a N
1 − i2 exp − i 2 − j aωi d
a 2a a a
i=1 −∞
N !
 
 ∞ xi2 xi x
i
=a N
exp − 2 − j aωi d
−∞ 2a a a
i=1
! ∞ 2  
xi xi2 xi x
i
− 2
exp − 2 − j aωi d
−∞ 2a 2a a a
N ! 
 ∞ 1
=a N
exp − xi2 + 2j aωi xi dxi
2
i=1 −∞
! ∞ 
1
− xi2 exp − xi2 + 2j aωi xi dxi
−∞ 2
N  ! ∞ 
1 2 2 2 1
=a N
exp − j a ωi exp − (xi + j aωi )2 dxi
2 −∞ 2
i=1
! ∞ 
1
− xi2 exp − (xi + j aωi )2 dxi
−∞ 2
  N
1 2 2
2 
N
= a exp − j a
N
ωi {G1 (ω) − G2 (ω)} .
2
i=1 i=1

/∞  /∞
where G1 (ω) = −∞ exp − 12 (xi + j aωi )2 dxi and G2 (ω) = −∞ xi2 exp

− 12 (xi + j aωi )2 dxi .
Wavelet Kernels for Support Matrix Machines 89

Using partial integration, we get

G1 (ω) = (2π )1/2 and G2 (ω) = (2π )1/2 (1 − a 2 ω2 ). (36)

Then


N
{G1 (ω) − G2 (ω)} = (2π )N/2 a 2N ωi2N . (37)
i=1

and so
  N
1 2 2
2  2
N
I = (2π ) N/2 3N
a exp − j a ωi ωi . (38)
2
i=1 i=1

It follows therefore that

  N
1 2 2
2  2
N
F (K(ω)) = (2π ) N/2
I = (2π ) a N 3N
exp − j a ωi ωi . (39)
2
i=1 i=1

and F (K(ω)) ≥ 0 since a ≥ 0 and the proof is completed. 



Morlet Wavelet Kernel for Support Matrix

Another wavelet kernel function that can be used is Morlet wavelet kernel function.
Its mother wavelet is defined as:
 2
x
(x) = cos(ω0 x) exp − . (40)
2

Figure 2 represents a 2-D plot of the Morlet wavelet kernel function.

Similarly, we define the Morlet wavelet kernel function as follows.
Theorem 4.7 Let a ≥ 0, if X, Y ∈ Rn ⊗ Rp are two matrices, then the Morlet
wavelet kernel function for matrices is defined as:
    

n 
p
xij − yij 
n 
p
xij − yij (xij − yij )2
K(X, Y) =  = cos ω exp − .
a a 2a 2
i=1 j =1 i=1 j =1
(41)
and is an admissible support matrix kernel function.
90 E. M. Maboudou-Tchao

Morlet wavelet kernel

1.0
0.8
0.6
0.4
K(X)

0.2
0.0
−0.2

−4 −2 0 2 4
X

Fig. 2 Morlet wavelet kernel function

Proof Using Theorem 4.4, we need to show that the Fourier transform of Morlet
wavelet is nonnegative. Let x = [x1 , . . . , xN ] = vec(X) with N = n × p, then it
follows that


N x
i
K(x) = 
a
i=1
N   x 
 i 1  xi 2
= cos ω0 exp − .
a 2 a
i=1
Wavelet Kernels for Support Matrix Machines 91

The integral term

!
J = exp(−j (ωx))K(x)dx
RN
! N 
  
  x 1  xi 2
i
= exp(−j (ωx)) cos ω0 exp − dx
RN a 2 a
i=1
N !
 
 ∞   x xi2 xi x
i i
=a N
cos ω0 exp − 2 − j aωi d
−∞ a 2a a a
i=1
N !
 
 ∞ x2 xi
=a N
exp − i 2 − j aωi
−∞ 2a a
i=1
1  x
i
 xi  xi
× exp ω0 + exp −ω0 d
2 a a a
N ! ∞
 
 a N  2
xi x
i
= exp − 2 + (j ω0 − j aωi ) dxi
2 2a a
i=1 −∞
! ∞  
xi2 x
i
+ exp − 2 − (j ω0 + j aωi ) dxi
−∞ 2a a


N    $
(2π )1/2 (ω0 + aωi )2 (ω0 − aωi )2
= aN exp − + exp − .
2 2 2
i=1

It follows now that

F (K(ω)) = (2π )N/2 J

N 
 a N    $ (42)
(ω0 + aωi )2 (ω0 − aωi )2
= exp − + exp − .
2 2 2
i=1

and F (K(ω)) ≥ 0 since a ≥ 0 and the proof is completed.



The left panel of Fig. 3 shows 3-D plot of the Mexican Hat kernel function while
the right panel displays a 3-D plot of Morlet kernel. The two plots look distinctly
different from each other.
92 E. M. Maboudou-Tchao

Mexican Hat kernel Morlet wavelet kernel

0.8 0.8
1.0 1.0

0.6 0.6

K(X, Y)
0.5
K(X, Y)

0.5
0.4 0.4
0.0

0.0 0.2
0.2
4 −4
2 −2
−4 0.0
−4 0 0.0 0
−2 −2
Y

Y
0 −2 2 0
X 2 2 X −0.2
−0.2
4 −4 4 4

−0.4
−0.4

Fig. 3 3-D plot of wavelet kernels. (left) Mexican Hat wavelet kernel, (right) Morlet wavelet
kernel

5 Applications

We illustrate the use of the two wavelet kernel functions on two datasets. We apply
WSMM to EEG and image classification problems. The EEG alcoholism dataset
is concerned with the relationship between genetic predisposition and tendency
for alcoholism. The study involved two groups of subjects: an alcoholic and a
control group. Each subject was exposed to a stimulus while voltage values were
measured from 64 channels of electrodes placed on the subject’s scalp for 256 time
points. So each subject has measurements of electrical scalp activity, which form a
256 × 64 matrix. There are 77 subjects from the alcoholic group and 45 subjects
from the control group. In our application, we used both 10 subjects from the
alcoholic and control groups. The performance comparison was assessed in terms
of the classification accuracy. Both the Mexican Hat wavelet kernel function for
matrices (Eq. (35)) and Morlet wavelet kernel function for matrices (Eq. (41)) yield
an accuracy of 95%.
The second dataset used is the INRIA person dataset. This dataset was proposed
to detect whether or not people exist in an image. Each color image is converted
into a 160 × 96 gray level image and the pixel values are used as an input matrix
without any advanced feature extraction technique. We use a small subset of the
dataset. The training set has 60 positives and 30 negatives for a total of 90 matrices.
The test set consists of 55 positives and 25 negatives for a total of 80 matrices. The
Mexican Hat wavelet kernel function for matrices (Eq. (35)) gives a classification
accuracy of 87.5% while Morlet wavelet kernel function for matrices, (Eq. (41))
yields a classification accuracy of 88.5%.
Wavelet Kernels for Support Matrix Machines 93

6 Conclusion

This article proposed some new kernel functions of support matrix machines, Mex-
ican Hat, and Morlet wavelet kernel functions. These kernel functions were used
to map matrices from the low dimensional matrix space to some high dimensional
space. We establish by proving that these kernels are valid or admissible kernels.
The method was successfully applied to EEG and INRIA image classification with
good performances.

References

Boser, B.E., Guyon, I., Vapnik, V.N.: A training algorithm for optimal margin classifiers. In:
Proceedings of the Fifth Annual Workshop of Computational Learning Theory, vol. 5, pp. 144–
152. ACM, Pittsburgh (1992)
Burges, C.J.C.: Geometry and invariance in kernel based methods. In: Scholkopf, B., Burges,
C.J.C., Smola, A.J. (eds.) Advances in Kernel Methods–Support Vector Learning, pp. 89–116.
MIT, Cambridge (1999)
Cortes, C., Vapnik, V.: Support vector networks. Mach. Learn. 20, 273–297 (1995)
Luo, L., Xie Y., Zhang, Z., Li, W.-J.: Support matrix machines. In: Proceedings of the 32nd
International Conference on Machine Learning (ICML), pp. 928–947 (2015)
Maboudou-Tchao, E.M.: Kernel methods for changes detection in covariance matrices. Commun.
Stat. Simul. Comput. (2017). http://dx.doi.org/10.1080/03610918.2017.1322701
Mercer, J.: Functions of positive and negative type and their connection with the theory of integral
equations. Philos. Trans. R. Soc. Lond. A 209, 415–446 (1909)
Shi, W., Zhang, D.: Support matrix machine for large-scale data set. In: International Conference
on Information Engineering and Computer Science, 2009. ICIECS 2009. 20, pp. 1191–1199
(2009)
Smola, A.J.: Regression estimation with support vector learning machines. Master’s thesis,
Technische Universitat Munchen (1996)
Smola, A., Scholkopf, B., Muller, K.-R.: The connection between regularization operators and
support vector kernels. Neural Netw. 11, 637–649 (1998)
Vapnik, V.: The Nature of Statistical Learning Theory. Springer, New York (1995). ISBN 0-387-
94559-8
Wu, F., Zhao, Y.: Least squares support vector machine on Morlet wavelet kernel function and its
application to nonlinear system identification. Inf. Technol. J. 5(3), 439–444 (2006)
Xia, W., Fan, L.: Least squares support matrix machines based on bilevel programming. Int. J.
Appl. Math. Mach. Learn. 4(1), 1–18 (2016)
Zhang, Q., Benveniste, A.: Wavelet networks. IEEE Trans. Neural Netw. 3(6), 889–898 (1992)
Zhang, L., Zhou, W., Jiao, L.: Wavelet support vector machine. IEEE Trans. Syst. Man Cybern. B
Cybern. 34, 34–39 (2004)
Zheng, Q., Zhu, F., Qin, J., Chen, B., Heng, P.H.: Sparse support matrix machine. Pattern Recogn.
76, 1–12 (2017)
Properties of the Number of Iterations
of a Feasible Solutions Algorithm

Sarah A. Janse and Katherine L. Thompson

1 Introduction

Many approaches exist to identify interaction effects in data sets with small
to moderate size. Classical statistical methods suggest considering all pairwise
combinations of possible explanatory variables in the proposed logistic regression or
linear regression model, and selecting a set of variables based on either hypothesis
tests or a model selection criterion. Although the theory supporting these techniques
is developed, often the data sets of interest have an inordinate number of possible
explanatory variables to consider in higher-order interactions using the conventional
implementations of classical methods.
For example, genomic data is also unique in its complexity due to intricate
dependencies among genes and traits, often in the form of information from external
influences or genetic makeup that are unaccounted for during analysis. In particular,
interaction effects among genes contribute to epistasis, which is especially difficult
to identify in genomic data (Moore and Williams 2009). These interactions have
been targets of recent analyses of genomic data although development of methods
to identify higher-order interaction effects has been more limited due to compu-
tational concerns (Gemperline 1999). For example, even using the second largest

Electronic Supplementary Material The online version of this chapter (https://doi.org/10.1007/

978-3-030-11431-2_5) contains supplementary material, which is available to authorized users.

S. A. Janse
Center for Biostatistics, The Ohio State University, Columbus, OH, USA
K. L. Thompson ()
Department of Statistics, University of Kentucky, Lexington, KY, USA
e-mail: katherine.thompson@uky.edu

© Springer Nature Switzerland AG 2019 95

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_5
96 S. A. Janse and K. L. Thompson

supercomputer at IBM (a 262,144 core machine), Goudey et al. performed analyses

examining two-way interactions on a large data set of 1.1 million single nucleotide
polymorphisms (SNPs) from 2000 samples in less than 10 min, and project that
analyzing three-way interactions on the same computer would take approximately
5.8 years (Goudey et al. 2015). Thus, exhaustive searches using anything less than
a massive supercomputer are highly impractical for this type of data, especially in
the case of higher-order interactions.
In contrast, stochastic search algorithms address the computational concern
of the exhaustive search methods by employing some aspect of randomness in
order to perform a non-exhaustive search over the possible explanatory variables.
However, these methods may not produce the same result every time, and thus
may fail to identify the truly optimal model according to the criterion used. In
addition, most exhaustive and stochastic searches produce a single “best” model
or set of explanatory variables with respect to some hypothesis testing or model
selection criterion. In any given data set, there may be another, nearly best model
that is more practical than the statistically best model. Only considering the
single statistically optimal solution leaves little room to consider more practically
meaningful combinations of variables without further experimentation.
Thus, room for improvement exists in implementing methods that are fast and
flexible in their ability to detect both main and interaction effects in a model,
and reliable in detecting effects that are not only statistically, but also practically,
significant. FSA may produce several nearly optimal solutions from different
iterations of the algorithm, rather than a single optimal solution. This variability
in FSA produces multiple results for consideration to glean practically reasonable
conclusions from the data, rather than ending with a single (statistically) optimal
solution. In the latter case, one solution may be optimal and practically nonsensical,
while another nearly optimal solution exists that is biologically relevant. FSA will
show the analyst both solutions for consideration during analysis.
Due to the stochastic nature of FSA, one issue that arises when implementing the
algorithm is the choice of number of iterations. Here, each replication is referred
to as a random start and begins with an arbitrary model based on the desired mth-
order interaction. The number of random starts must be chosen by the user. Thus,
we derive a bound on the probability of obtaining the statistically optimal solution
in a set number of random starts of FSA that can be used to select this number.
This allows users to choose a bound such that they obtain the statistically optimal
solution with a desired probability, prior to beginning data analysis.

2 Background

Issues from the complex nature of interaction effects, coupled with the size of
data, cause theoretical and computational problems when classical methods are
applied using standard implementations. To address these limitations, some recent
work has been focused on revisiting versions of the Feasible solutions algorithm
Properties of the Number of Iterations of a Feasible Solutions Algorithm 97

(FSA) first popularized by Doug Hawkins at the University of Minnesota in the

early 1990s (Miller 1984; Hawkins and Olive 1999). Several versions of FSA
exist (Hawkins 1993, 1994a,b, 1993), but we are focused on the current version
of FSA, which is used for subset selection. Common algorithms and methods
exist to find the best subset of predictors that adequately explain the response
variable. Forward selection, backward selection, and stepwise selection are common
automatic variable selection techniques. LASSO and ridge regression are common
penalized regression techniques that are used for subset selection with many
variables. Exhaustive search checks all possible combinations for the possible model
structures. These are currently available in the form of R packages for linear and
generalized linear models in leaps (Lumley and Miller 2004) and glmnet (Friedman
et al. 2009).
Here, we study the properties of FSA, which is designed to find interactions
when the number of predictors is large (Lambert et al. 2018). FSA searches the
set of all possible interaction effects to identify those that improve the predictive
model for a given response. Issues from the complex nature of interaction effects,
coupled with the size of big data, cause theoretical and computational problems
when classical methods are applied using standard implementations. One advantage
of applying FSA in these cases is that it provides more than one feasible solution, or
candidate set of explanatory variables, for a particular data analysis. Providing a set
of solutions increases the likelihood of finding practically significant associations
rather than solely statistically significant associations.
Specifically, FSA is carried out as follows:
1. Randomly choose m variables from the possible p predictors and compute a
specified objective function, i.e., R 2 .
2. Consider exchanging one of the m selected explanatory variables from the current
model with another explanatory variable in the data set.
3. If an improvement exists, make the exchange that improves the objective function
the most.
4. Keep making exchanges until the objective function does not improve. The
explanatory variables included in the resulting model are called a feasible
solution.
5. Repeat steps (1)–(4) for the number of random starts specified to find additional
feasible solutions.
A feasible solution is optimal in that no one exchange of a variable in the
model for another outside of the model can improve the selected criterion function.
Not only does FSA provide a set of feasible solutions, but it is often more
computationally efficient than standard exhaustive approaches due to its stochastic
nature. Other advantages of the algorithm include the ability to analyze both linear
and logistic regression models, as well as being able to implement several different
optimization criteria. In this work, we focus on the required number of iterations, or
replications, required by FSA to produce the statistically optimal model.
98 S. A. Janse and K. L. Thompson

3 Methods

In FSA, each random start begins with an arbitrary model with a fixed number of
predictors and proceeds by taking steps to better models based on some optimization
criteria, e.g., R 2 . The algorithm proceeds until it reaches an optimal model for a
given random start. Thus, each random start, or replication of FSA, will have at least
one step, but often times will have several more. FSA is not guaranteed to identify
the optimal solution, but as the analyst increases the number of random starts, FSA
is more likely to do so. Thus, we need enough random starts to obtain the optimal
solution with some probability. However, the larger the number of random starts,
the longer the time it will take FSA to run. Therefore, it would be highly useful to
have information regarding how many random starts to choose in order to obtain
the optimal solution with some probability while still maintaining computational
efficiency.
As the number of explanatory variables, p, in a data set increases, it is more
difficult to identify the optimal solution and will require more random starts. We
propose choosing the number of random starts as a function of p. As p goes to
infinity, the probability that the optimal solution is identified by FSA is bounded
below. The limit described in Theorem 1 holds for FSA in the case of considering
m-way interactions.
Theorem 1 In the case of using FSA to find a statistically significant m-way
interaction in a predictive model, as the number of potential explanatory variables,
p, goes to infinity, a lower bound on the probability of identifying the statistically
optimal model in cp random starts, where 0 < c < 1, is 1 − e−cm .
2

Lemma
 
k tx
lim 1 + = etk
x→∞ x

Proof of Theorem 1 Let p be the number of possible explanatory variables we are

choosing from, c be a constant  such
 that 0 < c < 1, and cp be the pnumber
 of
random starts. Then there are p−m m pairs of variables out of the total m possible
pairs that do not contain any of the variables in the optimal solution, consisting of
m variables. Note that, if you randomly start with m − 1 out of the m variables in
the statistically optimal solution, you are guaranteed to obtain the optimal solution.
Then, the probability of not identifying the optimal solution in the first step of a
given random start is
p−m
m
p (1)
m
Properties of the Number of Iterations of a Feasible Solutions Algorithm 99

and so the probability of obtaining the optimal solution in the first step of a given
random start is
p−m
1 − m
p . (2)
m

For a given random start, FSA completes at least one step, and often more than
one step, before reaching a feasible solution. Since we are only considering finding
the statistically optimal solution after the first step and not considering the cases
where we could find the optimal solution in later steps, Eq. (2) will be a lower bound
on the probability of identifying the statistically optimal solution in a single random
start. So, the probability of obtaining the statistically
 p−m optimal solution
 p−m in cpat least
cp
( m ) ( m )
one of the cp random starts is greater than 1 − , where is the
(mp ) (mp )
probability that none of the random starts identify the optimal solution in the first
step of FSA. So we consider
  p−m cp
lim 1 − m p
p→∞
m
 p−m cp
= lim 1 − lim m
p
p→∞ p→∞
m
 
(p − m)! m!(p − m)! cp
= 1 − lim
p→∞ m!(p − 2m)! p!
 
(p − m)!(p − m)! cp
= 1 − lim
p→∞ p!(p − 2m)!
 cp
(p − m)!
= 1 − lim
p→∞ (p − 2m)!p(p − 1) · · · (p − m + 1)
 
(p − m)(p − m − 1) · · · (p − 2m + 1) cp
= 1 − lim .
p→∞ p(p − 1) · · · (p − m + 1)

Notice that both the numerator and denominator in the limit statement contain m
quantities. Thus we can write the last line above as
     
p − m cp p − m − 1 cp p − 2m + 1 cp
= 1 − lim ···
p→∞ p p−1 p−m+1
 cp  cp  
p−m p−m−1 p − 2m + 1 cp
= 1 − lim lim · · · lim
p→∞ p p→∞ p−1 p→∞ p − m + 1
 cp  cp  cp
m m m
= 1 − lim 1 − lim 1 − · · · lim 1 − .
p→∞ p p→∞ p−1 p→∞ p−m+1
100 S. A. Janse and K. L. Thompson

Then we have
 
m cp
lim 1 − = e−cm
p→∞ p

by the lemma with t = c and k = −m. Next,

 
m cp
lim 1 −
p→∞ p−1
 c(p−1)  
m m c
= lim 1 − 1− .
p→∞ p−1 p−1
 c(p−1)
Since limp→∞ 1 − p−1
m
= e−cm by the lemma with t = c and k = −m
 c
and limp→∞ 1 − p−1
m
= 1, we have

 cp
m
lim 1− = e−cm .
p→∞ p−1

Next,
 cp
m
lim 1−
p→∞ p−m+1
 c(p−m+1)  c(m−1)
m m
= lim 1 − 1− .
p→∞ p−m+1 p−m+1
 c(p−m+1)
Since limp→∞ 1 − p−m+1
m
= e−cm by the lemma with t = c and
 c(m−1)
k = −m and limp→∞ 1 − p−m+1
m
= 1, we have

 cp
m
lim 1− = e−cm .
p→∞ p−m+1

So,
 p−m cp
1 − lim m
p = 1 − e−cm × e−cm × · · · e−cm (mtimes)
p→∞
m

= 1 − e−c(m ) .
2
Properties of the Number of Iterations of a Feasible Solutions Algorithm 101

Fig. 1 In this plot, the dots

1.2
l

Probabilty of Obtaining Optimal Solution

c = 0.01
show the exact value of the l c = 0.02
lower bound for varied values l c = 0.10

1.0
of c, and the lines show the
ll
l c = 0.20
l
l l c = 0.40
asymptotic lower bound on l
ll
lll
llllll

0.8
llllllllllllllllll
the probability of getting the l
lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll

statistically optimal solution l

l
with m = 2. The lower bound l
ll

0.6
lll
llllll
llllllllllllllll
llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll
is attained very quickly and l

the probability of identifying l

0.4
l
ll
the statistically optimal lll
llllll
lllllllllllllllllllllll
llllllllllllllllllllllllllllllllllllllllllllllllllllllllllll
solution increases as the
0.2
number of random starts
increases, as expected l
ll
llllllllll
llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll
ll
lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll
0.0

1 20 40 60 80 100
p

Thus,

P(Obtaining the statistically optimal model in cp random starts using FSA)

 p−m cp
≥1− m
p ,
m

and note that:

  p−m cp
= 1 − e−cm .
2
lim 1 − m p
p→∞
m

Figure 1 shows how the calculated probability of obtaining the optimal solution
approaches the lower bound derived above for 5 values of c with m = 2. It can
be seen that the lower bound is attained very quickly and thus is appropriate when
considering data sets with a large number of explanatory variables, p. It is also clear
that the probability of obtaining the statistically optimal solution increases as the
number of starts increases, as is expected.

4 Results

Simulation studies were performed for both quantitative and binary response
variables to examine the outcomes of utilizing the lower bound derived above. These
simulations were followed by a real data analysis to demonstrate the use of the lower
bound in practice.
102 S. A. Janse and K. L. Thompson

4.1 Simulations

Quantitative trait data were simulated as the sum of two covariates and their
interaction under the typical regression model for values of p of 50, 100, 1000, and
2500. One hundred data sets were simulated for each value of p. Binary trait data
were simulated in an analogous manner. Simulations parameters are as follows:
• Quantitative response variable (lmFSA)
– Xij ∼ U (0, 1) for i = 1, . . . , n and j = 1, .., p
– Yi = 5 + Xi1 + Xi2 + 2Xi1 Xi2 + i where i ∼ N (0, 1)
• Binary response variable (glmFSA)
– Xij ∼ U (0, 1) for i = 1, . . . , n and j = 1, . . . , p
Xi1 +Xi2 +2Xi1 Xi2
– πi = e Xi1 +Xi2 +2Xi1 Xi2
)
(1+e )
1 with probability πi
– Yi =
0 with probability 1 − πi
FSA was used to provide a set of feasible solutions for every simulated data
set via the implementation in Lambert et al. (2018). Exhaustive search was then
performed to find the statistically optimal solution using R 2 and AI C as the
criteria functions for the quantitative and binary response variables, respectively.
The numbers of random starts chosen for FSA were values of c including 0.01,
0.02, 0.1, 0.2, and 0.4 with each value of p. Then, for each simulation setting, the
percentage of simulated data sets producing the statistically optimal solution using
FSA was calculated. These percentages, along with the lower bound from Sect. 3,
are plotted in Fig. 2.
Figure 2 shows the results from 100 simulated data sets for both methods in FSA
for four values of p and five values of c. Note that the asymptotic lower bound
proposed here depends only on m and c. The red dots represent these lower bounds
for each value of c. The yellow diamonds, blue dots, green squares, and black
triangles represent the percentage of 100 simulations with p = 50, 100, 1000, and
2500, respectively, when m = 2 [where FSA was able to identify the statistically
optimal solution]. It is clear from Fig. 2 that the lower bound is often much lower
than the observed probability and is thus very conservative, but does provide a good
guidance as to the number of random starts needed to produce at least one feasible
solution containing the statistically optimal solution.

4.2 Real Data Example

Data were collected in a genome-wide association study using 288 outbred mice in
a study that aimed to identify, or map, locations along the genome called SNPs that
influence HDL cholesterol, systolic blood pressure, triglyceride levels, glucose, or
Properties of the Number of Iterations of a Feasible Solutions Algorithm 103

1.0 lmFSA Results glmFSA Results

1.0
l
l l
l
l
l
Probabilty of Obtaining Optimal Solution

Probabilty of Obtaining Optimal Solution

0.8

0.8
l l
l
l
0.6

0.6
ll
l l
0.4

0.4
l l

l Lower Bound l Lower Bound

0.2

0.2
p = 50 p = 50
p = 100 p = 100
l l p = 1000 l l p = 1000
l l
0.0

0.0
p = 2500 p = 2500

1 10 20 40 1 10 20 40
% of p (c) % of p (c)

Fig. 2 Simulation results for the probability of getting the optimal solution in 100 simulations
with a quantitative response variable (left) and a binary response variable (right): for each value
of c, the lower bounds are represented by the red dots and the probability of identifying the
statistically optimal solution for the four values of p is represented by yellow diamonds (p = 50),
blue dots (p = 100), green squares (p = 1000), and black triangles (p = 2500). Both the left and
right plots show that the lower bound is valid for all values of p in the simulation study

urinary albumin-to-creatinine ratios (Zhang et al. 2012). Our goal was to determine
if SNPs or interactions of SNPs were associated with HDL levels. Information from
3045 SNPs on chromosome 11 were analyzed for this real data analysis.
Using the lower bound in Theorem 1, if we want the probability of obtaining the
statistically optimal solution including a 2-way interaction to be at least 95%, then
we need to solve the following equation for c:

1 − e−c2 = 0.95
2

⇐⇒ c = 0.7489331

Since we have p = 3045, the number of random starts we need is 0.75(3045) =

2283.75, or 2284 random starts.
The exhaustive search of the 3045 SNPs on chromosome 11 took approximately
11 h in total on a large cluster without parallelization. Using this method, we found
that the statistically optimal solution includes an interaction between mb2863979
and mb87344525 and corresponds to a value of R 2 = 0.1256, which can be found in
Table 1. FSA took approximately half an hour to perform 2284 random starts when
parallelized on a larger cluster, using 16 cores. There were a total of 33 feasible
solutions identified through FSA, including the statistically optimal solution, which
is presented in bold in the subset of FSA results in Table 2. (Full FSA results can
104 S. A. Janse and K. L. Thompson

Table 1 The exhaustive search produced the single statistically optimal solution with R 2 =
0.1256308 (column 3)
Variable 1 Variable 2 R2
mb2863979 mb87344525 0.1256308
Columns 1 and 2 show the SNPs that were identified in this model

Table 2 FSA produced 33 feasible solutions and a subset of those are shown here, including the
statistically optimal solution denoted in bold with R 2 = 0.1256308 (column 3)
Variable 1 Variable 2 Times chosen by FSA R2
mb104327194 mb91638370 42 0.0957401
mb13136127 mb31255782 898 0.1245719
mb28636979 mb87344525 107 0.1256308
mb111935889 mb43233761 25 0.1065257
mb62443411 mb99541026 23 0.1088855
mb112250554 mb96331482 56 0.1123864
Columns 1 and 2 show the SNPs that were identified in each of the models

be found in the supplemental materials.) Out of the 2284 replications of FSA, the
statistically optimal solution was identified in 107 of the replications, showing that
the number of random starts used here was sufficient.

5 Discussion and Conclusion

Although FSA addresses limitations of existing modeling techniques, little is

known about its theoretical properties. To address one aspect of this limitation,
we have provided a way to determine the number of iterations of FSA needed
to obtain the statistically optimal solution of an m-way interaction model with
a certain probability in Theorem 1. For example, when considering a two-way
interaction model, if you would like the probability of obtaining the statistically
optimal solution to be at least 80%, then you would need to choose the number of
random starts of FSA to be 40% of the number of possible explanatory variables
in your data set. This lower bound on the probability of obtaining the statistically
optimal solution can be easily implemented by data analysts running FSA. Further,
simulation study and real data analysis demonstrated the validity and usefulness of
this lower bound.
The work here provides a foundation for further study of theoretical properties of
FSA. For instance, the simulation study results show that the derived lower bound
is conservative. Thus, in future work, we aim to tighten the bound. However, in this
case, the conservative lower bound does provide statistical guidance to FSA users.
In addition, little is known about how conservative this bound is in the presence
of smaller effect sizes, which will increase the impact of this work by providing
more specific guidance to the data analyst. Knowing how to choose the number of
Properties of the Number of Iterations of a Feasible Solutions Algorithm 105

times to run FSA will improve the computational usability of FSA by allowing the
user to choose fewer random starts based on the desired likelihood of obtaining the
statistically optimal solution while still being computationally feasible, and continue
providing a valid alternative to exhaustive search methods.

References

Friedman, J., Hastie, T., Tibshirani, R.: glmnet: Lasso and elastic-net regularized generalized linear
models. R package version 1(4) (2009)
Gemperline, P.J.: Computation of the range of feasible solutions in self-modeling curve resolution
algorithms. Anal. Chem. 71(23), 5398–5404 (1999)
Goudey, B., Abedini, M., Hopper, J.L., Inouye, M., Makalic, E., Schmidt, D.F., Wagner, J., Zhou,
Z., Zobel, J., Reumann, M.: High performance computing enabling exhaustive analysis of
higher order single nucleotide polymorphism interaction in genome wide association studies.
Health Inf. Sci. Syst. 3(1), 1 (2015)
Hawkins, D.M.: The feasible set algorithm for least median of squares regression. Comput. Stat.
Data Anal. 16(1), 81–101 (1993)
Hawkins, D.M.: A feasible solution algorithm for the minimum volume ellipsoid estimator in
multivariate data. Comput. Stat. 8, 95–95 (1993)
Hawkins, D.M.: The feasible solution algorithm for least trimmed squares regression. Comput.
Stat. Data Anal. 17(2), 185–196 (1994)
Hawkins, D.M.: The feasible solution algorithm for the minimum covariance determinant estimator
in multivariate data. Comput. Stat. Data Anal. 17(2), 197–210 (1994)
Hawkins, D.M., Olive, D.J.: Improved feasible solution algorithms for high breakdown estimation.
Comput. Stat. Data Anal. 30(1), 1–11 (1999)
Lambert, J., Gong, L., Elliot, C.F., Thompson, K., Stromberg, A.: rFSA: an R package for finding
best subsets and interactions. R J. 10(2), 295–308 (2018)
Lumley, T., Miller, A.: Leaps: regression subset selection. R package version 2 (2004)
Miller, A.J.: Selection of subsets of regression variables. J. R. Stat. Soc. Ser. A Gen. 147(3), 389–
425 (1984)
Moore, J.H., Williams, S.M.: Epistasis and its implications for personal genetics. Am. J. Hum.
Genet. 85(3), 309–320 (2009)
Zhang, W., Korstanje, R., Thaisz, Staedtler, F., Harttman, N., Xu, L., Feng, M., Yanas, L., Yang,
H., Valdar, W., et al.: Genome-wide association mapping of quantitative traits in outbred mice.
G3: Genes Genomes Genetics 2(2), 167–174 (2012)
A Primer of Statistical Methods
for Classification

Rajarshi Dey and Madhuri S. Mulekar

1 Introduction

Classification is a process of assigning a new subject or item to one of the G

known groups or classes on the basis of how closely specific characteristics of this
subject/item match with those of the groups. For example, on the basis of specific
protein levels measured for a patient, an oncologist can determine with certain
confidence whether or not that patient has a certain type of cancer; using a pixel-
based satellite image, a geographer can classify land cover into different categories
such as water, forested wetland, and upland forest; or using certain admissions
criteria, a university can classify applicants as accepted or non-accepted into their
program.
There are many different methods (or rules) used to achieve this goal of
classification of subjects/items into different classes. Note that classification is
not to be confused with clustering as classification involves assigning items to a
known number of groups with specific characteristics, whereas clustering involves
forming groups of items with similar characteristics when the existing number of
groups is unknown. In a world of machine learning and computation, the process of
classification is referred to as a supervised learning whereas the method of clustering
is an example of unsupervised learning. For example, consider a chicken farm that
packages chicken eggs. Before packaging, each egg has to be classified as medium,
large, extra-large, or jumbo. This is a case of classification as the classes are well
defined based on the egg size and machines can be taught to properly classify eggs
based on their size. Sometimes clustering or other pattern recognition methods are

R. Dey · M. S. Mulekar ()

Department of Mathematics and Statistics, University of South Alabama, Mobile, AL, USA
e-mail: rajarshidey@southalabama.edu; mmulekar@southalabama.edu

© Springer Nature Switzerland AG 2019 107

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_6
108 R. Dey and M. S. Mulekar

used by scientists as a first step towards classification in which existence of number

of groups is determined using multiple characteristics of its constituents. In this
article, we discuss classification only in the sense of supervised learning, i.e., a
procedure in which discriminating variables or functions are used to predict group
membership.
Classification into one of the two groups (i.e., G = 2) is known as a binary clas-
sification and is relatively easier to implement. The earlier classification procedures
developed were mostly binary classification methods. But with the development of
computing and technology, more literature on different aspects of classification has
become available (Hastie et al. 2001).
Broadly speaking there are two types of classifiers: hard and soft. As Wahba
(2002) described, the soft classifiers also known as probabilistic classifiers typically
provide conditional probability of membership for each of G (G ≥ 2) groups for
each new subject to be classified and then put the subject in the group with the
largest probability of membership. In contrast, hard classifiers only provide a hard
classification boundary like a fence around a property for each group based on the
explanatory variables. A new subject with characteristics within the boundary of a
group is assigned to that particular group.
The development of a classification procedure involves two major steps, classi-
fication and validation. In the first step, a classification rule or an algorithm also
known as a classifier is developed and in the second step, performance of this
classifier is evaluated. Dataset used in the first step to develop a classification rule
is known as the training dataset and includes information about group membership
(response or output) of each subject along with other random variables (explanatory
variables or features). A similar dataset is used in the second step to validate the
classifier developed in the first step and is known as the validation dataset.
Let us assume that all outcomes are independently observed for random response
variable Y and a vector of random explanatory variables X = (X1 , X2 , . . . , Xp ).
Let {(xi , yi ) : i = 1, 2, . . . n} indicate a training dataset consisting of n independent

measurements. Each measurement is a (p + 1)-tuple where xi = (x1i , x2i , . . . , xpi )
is a vector of outcomes for p explanatory variables and yi (i = 1, 2, . . . , n) is the
outcome for response variable, which only shows membership of ith unit to a certain
class.
Consider the famous iris dataset by R.A. Fisher (Anderson 1935) containing 150
data points with four explanatory variables and three classes of iris. The explanatory
variables are sepal-length, sepal width, petal-length, and petal-width of three types
of irises, namely setosa, virginica, and versicolor. The goal is to identify type of
iris using sepal and petal measurements. This is a classification problem. Sepal-
length and petal-length distributions in Fig. 1 show that although setosa tend to
have lowest and virginica tend to have highest sepal-lengths the separation among
three classes based on sepal-length is not clear due to considerable overlap and
hence sepal-length by itself is not a good classifier for these three types of irises.
On the other hand, petal-length is clearly able to distinguish setosa from the other
two but not between virginica and versicolor possibly resulting in misclassification.
Hence there is need for more than one predictor to reduce misclassification. Figure 2
A Primer of Statistical Methods for Classification 109

class of Iris
Iris-virginica
Iris-versicolor
Iris-setosa

4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0

sepal-length

class of Iris
Iris-virginica
Iris-versicolor
Iris-setosa

1.6 2.4 3.2 4.0 4.8 5.6 6.4

petal-length

Fig. 1 Distributions of sepal- and petal-lengths of three varieties of iris

Fig. 2 A scatterplot of iris petal-length vs petal-width

110 R. Dey and M. S. Mulekar

shows a scatterplot of iris petal-length versus petal-width. Once again setosa are
clearly separated from the other two varieties, but there is still some overlap between
virginica and versicolor possibly leading to misclassification.
In this article, we aim to provide basic description of the most well-known
and commonly used classification methods that are used to develop classifiers
(or classification rules) based on relation between the response variable Y and
explanatory variables X, which then are used to assign new objects to these known
groups based on observed x’0 . Two soft classifiers (logistic regression and naïve
Bayes estimator) and four hard classifiers (linear discriminant analysis, support
vector machines, K nearest neighbor, and classification trees), respectively, are
described in Sects. 2 and 3 along with their strengths and weaknesses. Some
discussion assessing performance of these classifiers for five different datasets, three
real and two simulated, is provided in Sect. 4. Some concluding remarks about
choice of classifiers in practice are provided in Sect. 5.

2 Soft Classifiers

Intuitively, a soft classifier should appeal to anyone who likes to incorporate the
uncertainty of outcome provided by classifiers because it also shows the likelihood
of a new observation being a member of different classes. Here two most commonly
used soft classifiers, namely logistic regression and naïve Bayes classifiers are
discussed.

2.1 Logistic Regression

As described by Cramer (2003), the first use of logistic function in logistic

regression was traced to modeling population growth rate in the nineteenth century
Africa. Berkson (1944) suggested the use of logistic probability density function
(pdf) instead of normal pdf in certain bioassay procedures. He also coined the term
logit model to describe the resulting model. Later many researchers in statistics and
epidemiology started working on what would eventually become one of the most
widely used methods in classification, namely the logistic regression, particularly
with G = 2 groups. Cox (1969) is considered one of the pioneers in binary logistic
regression. More generalized versions of logistic regression, which can classify new
items into G (G ≥ 2) groups, are credited to Gurland et al. (1960), Mantel (1966),
and Theil (1969).
For the sake of simplicity, let’s start with the case of binary logistic regression
with two classes being coded as 1 and 2 (i.e., yi = 1 or 2). For π i1 = P(yi = 1| xi ),
i = 1, 2, . . . , n and under the assumption that the response variable yi has a
A Primer of Statistical Methods for Classification 111

Bernoulli distribution with parameters π i1 , the logistic model is given by,

 p
exp β0 + j =1 βj xj i
πi1 = E (yi = 1|xi ) =  p (1)
1 + exp β0 + j =1 βj xj i

where β 0 , β 1 , . . . , β p are (p + 1) regression coefficients. Note that in a binary case,

π i2 = 1 − π i1 .
Alternatively this model can be presented as,


p
πi1
logit (πi1 ) = log = β0 + βj xj i (2)
1 − πi1
j =1

The regression parameters β j , j = 0, 1, . . . , p are estimated from the available

training dataset. The maximum likelihood (ML) estimates βˆj , j = 0, 1, . . . , p are
obtained by maximizing the likelihood function

  
n
y
L β0 , β1 , . . . , βp = πi i (1 − πi )n−yi .
i=1

Since there are no closed-form solutions available for maximizing this likelihood
function, iterative algorithms are used to obtain the ML estimates of regression
parameters. According to Agresti (2013), the most popular choices for iterative
algorithms are either the Newton–Raphson algorithm (Tjalling 1995) or itera-
tively reweighted least square (IRWLS) algorithm (Burrus et al. 1994). However,
sometimes due to the use of too many explanatory variables or highly correlated
explanatory variables, these algorithms fail to converge resulting in failure to
estimate parameters. Another counter-intuitive situation sometimes occurs when
there is a complete separation between two classes using some linear combination
of explanatory variables. More information on estimating parameters of logistic
regression is available in Menard (2002).
A simple extension of logistic regression from binary to multiclass classification
is known as multinomial logistic regression. The multinomial logistic model is
given by,
 p
exp β0g + j =1 βjg xj i
πig =   p , g = 1, 2, . . . , G − 1 and i = 1, 2, . . . , n.
1 + G−1
g=1 exp β0g + j =1 βjg xj i
(3)

Extending notation used in the binary case to G ≥ 2 groups, we can write

π ig = P(yi = g| xi ), for i = 1, 2, . . . , n and g = 1, 2, . . . , G. Although it does
not matter which category is chosen as baseline, generally category G is used as a
112 R. Dey and M. S. Mulekar

G−1
baseline and π iG can be obtained using the fact that πiG = 1 − g=1 πig . From
the point of estimation, there are (p + 1)(G − 1) model parameters to be estimated.
For estimating these parameters, the ML estimation or the maximum a posteriori
(MAP) methods are commonly used (Murphy 2012). Estimation method MAP is
similar to ML in the sense that it chooses that value of parameter which maximizes
the value of a mathematical function, in this case the posterior distribution of the
parameter itself. Most of the times, a closed-form solution is not available, hence
different algorithms are used for estimation and IRWLS is a popular choice among
practitioners.
If the G ≥ 2 classes are ordered using an ordinal response variable, an alternative
popular model often used in practice is the proportional-odds cumulative logit
model. For example, consider a typical Likert scale question where the responders
are asked to grade certain experience on a scale of 1 to 5 with 1 being the worst
rating and 5 being the best. It might be of interest to determine if there exist some
explanatory variables that can explain how the responders rate their experience. First
developed by Snell (1964), this model is given by,
g
πic

p
Lig = log Gc=1 = β0g + βj xj i , for g = 1, 2, . . . , G and i = 1, 2, . . . , n.
c=g+1 πic j =1
(4)

Here, Lig represents the log-odds of two cumulative probabilities. A manageable

number of total (G − 1 + p) parameters are to be estimated from this model.
Typically, ML estimates of parameters of this model are obtained using iterative
algorithms such as IRWLS and majorization-minimization (Lange 2016).

2.2 Naïve Bayes Classifier

Naïve Bayes (NB) is a family of soft classifiers that uses the Bayes theorem (Bayes
1763) along with a very strong assumption of independence among explanatory
variables which is often unrealistic. However, this classifier works very well in
the presence of dependencies among many categorical explanatory variables (Rish
2001) and is quite fast to execute even with large datasets.
NB classifier differs from the logistic regression classifier in terms of how
the probability π ig is modeled. When using a logistic regression classifier,
π ig = P(yi = g| xi ) is modeled directly from data. On the other hand, when using
a Naïve Bayes classifier, first the estimates for P(xi | yi = g) are obtained from
data and then assuming independence among explanatory variables, π ig is modeled
using Bayes theorem as,


p
 
πig ∝ P (yi = g) P xj i |yi = g , g = 1, 2, . . . , G and i = 1, 2, . . . , n. (5)
j =1
A Primer of Statistical Methods for Classification 113

The estimate for P(yi = g) can be obtained from the training set as the proportion
of training set observations that belong to class g (g = 1, 2, . . . , G). The estimates
for P(xji | yi = g) are typically obtained via ML estimation technique. A new
observation is assigned to a group for which probability π ig is maximum among
all G groups.
Estimating parameters from the likelihood function depends on how the likeli-
hood, P(xji | yi = g), i = 1, 2, . . . , n, j = 1, 2, . . . , p, and g = 1, 2, . . . , G
is modeled parametrically. If Xj is a continuous random variable, then the popular
choice of distribution is normal (Gaussian) such that (Xj | Y = g)∼N(μg ,  g ). If Xj
is a categorical random variable with m categories, then the most commonly used
distribution is multinomial, i.e., (Xj | Y = g)∼Multinomial(1, φ 1g , . . . , φ mg ) for one
φmlg , l = 1, 2, . . . m is the probability associated with the l category
trial where th

such that l=1 φl = 1.

The NB classifier is remarkably effective considering the assumptions needed to
obtain the probabilities are almost always wrong (Hand and Yu 2001). This method
is a building block to what is commonly known as a Bayesian spam filter (Nigam
et al. 2000) used by the email providers. A semi-parametric version of NB classifier
performs much better when the explanatory variables are obviously non-normal
(Soria et al. 2011).

3 Hard Classifiers

Hard classifiers typically do not provide a probability of group association. In

other words, there is no uncertainty associated with classification because classifier
provides a hard boundary between groups and exactly for this reason some
researchers like to use them. Four commonly used classifiers discussed here are
linear discriminant analysis, K nearest neighbor, support vector machines, and
classification trees.

3.1 Linear Discriminant Analysis

Linear discriminant analysis (LDA) is a hard classification method. Statistical

literature indicates that LDA is one of the first methods developed for classification
and its basic idea originated from none other than Fisher (1936). The basic idea
behind LDA is to determine that linear combination of explanatory variables which
will magnify the difference between two classes making it easier to achieve correct
classification. The generalization of this idea for classification into G(G > 2) classes
is credited to Rao (1948). The NB classifier is similar to LDA in nature (Hand and
Yu 2001), although in LDA the aim is obtain a classifier while in NB there is more
emphasis on identifying a class with the maximum posterior probability.
Fisher (1936) proposed a classification rule for two groups which involved
 
determining a vector r that maximizes function δ(r) = (r r)−1 (r (μ1 − μ2 ))2
114 R. Dey and M. S. Mulekar

under the assumption that (X| Y = g)∼N(μg , ) for g = 1, 2. This is equivalent

to finding a hyperplane that provides a solution to equation
 
P (Y = 1|X)
log = 0. (6)
P (Y = 2|X)

Using Bayes’ rule, we can write,

P (X|Y = g)
P (Y = g|X) = P (Y = g) for g = 1, 2
P (X)

where pg = P(Y = g) for g = 1, 2 is the overall class probability and can be estimated
from the training data. Under the assumption that the explanatory variables are
multivariate normal, the hyperplane can be found by solving the following equation
for r,
 
p1 1  ’ −1
log + r’  −1 (μ1 − μ2 ) − μ1  μ1 − μ2’  −1 μ2 = 0. (7)
p2 2

Solution to (7) leads to a linear classifier (or a linear boundary between two
groups) because Eq. (6) is a linear function of explanatory variables. The first step
in LDA is to estimate the mean vectors (μ1 and μ2 ) and variance–covariance matrix
() from the training dataset. For any new observation, x0 , one can estimate x0
from (8) as,
  
 
ˆ x0 = log

p̂1 ˆ −1 μ̂1 − μ̂2 − 1 μ̂’1 
+ x0’  ˆ −1 μ̂1 − μ̂’2 
ˆ −1 μ̂2 . (8)
p̂2 2

ˆ x0 value a new observation x0 is assigned to one of the two groups

Using this 
as follows:

Group 1 if  ˆ x0 > 0
Assign x0 to ˆ
Group 2 if x0 < 0.

In cases where the assumption of homoscedasticity of variance–covarinace

matrix is not justified and a more general underlying assumption is that
(X| Y = g)∼N(μg ,  g ) for g = 1, 2, a quadratic classifier (i.e., a quadratic function)
is used to describe a boundary between two classes. This procedure is known as
quadratic discriminant analysis (QDA) (Hastie et al. 2001). In QDA, the hyperplane
can be obtained by solving (9) for r.
  
log p1
p2− 12 log | 1|
| 2 | + r’  −1 μ −  −1 μ
1 1 2 2
  (9)
− 12 r’  −1
1 −  −1
2 r − 1
2 μ ’  −1 μ − μ’  −1 μ = 0
1 1 1 2 1 2
A Primer of Statistical Methods for Classification 115

As can be seen from (7) and (9), a QDA requires more parameters to be estimated
from the training dataset, precisely (2 + 2p + 2p2 ) parameters for QDA compared
to (2 + 2p) for LDA. That can lead to a serious issue if the training dataset is small.
To overcome this issue, Srivastava et al. (2007) proposed an effective Bayesian
solution.
A simpler method under the assumption of homoscedasticity of variance–
covariance matrices is to use Mahalanobis distance (Mahalanobis 1936) for clas-
sification. For any new observation, x0 , a linear discriminant function LDFg
is computed for each group (see (10)) under the assumption that μg and  ,
respectively, are the unknown mean vector and variance–covariance matrix of X.

ˆ −1 1 ’ ˆ −1
LDF g (x0 ) = x0’  0 μ̂g − μ̂g  0 μ̂g + p̂ g (10)
2

where  ˆ 0 is the pooled estimate of the common variance–covariance matrix 

and p̂g is the estimate of probability pg = P(Y = g) for g = 1, 2, . . . , G
obtained from the training data. Then the new observation x0 is assigned to the
group with the highest discriminant function value, i.e., the group corresponding to
max{LDFg (x0 ), g = 1, 2, . . . , G}.
Although LDA is quite effective in many situations (Hand 2006), in some
situations the joint pdf of explanatory variables differs considerably from the
multivariate normal distribution. In such cases semi-parametric LDA technique
derived by Mai and Zou (2015) under the assumption of sparse variance–covariance
matrix is more effective.

3.2 K Nearest Neighbor

Assumed to have originated in long past, the history of K(1 < K < n) nearest
neighbor (KNN) classification is not really that well known. In modern times,
Sebestyen (1962) described this method as proximity algorithm and Nilsson (1965)
called it the minimum distance classifier. Cover and Hart (1967) were the first to
name this algorithm as the nearest neighbor and that name became popular.
Although mostly used as a hard classifier, KNN can be used as a soft classifier
too. The idea behind KNN is quite simple and no parametric assumption is required.
Given a training dataset of size n(n > K), this classification algorithm starts when

a new observation, x0 = (x01 , . . . , x0p ) , is recorded with known values for all
explanatory variables but unknown class. The first step is to calculate the K nearest
neighbors in terms of the explanatory variables. Using some well-defined distance
measure, distance di = d(x0 , xi ), i = 1, 2, . . . , n between this new observation
and each observation from the training dataset is calculated and these distances
are ordered as d(1) ≤ d(2) ≤ . . . ≤ d(n) . Considering the lowest K distances,
{d(1) , d(2) , . . . , d(K) }, the class membership of these closest K neighbors in the
training dataset is determined. Then the new observation is placed in the class that
116 R. Dey and M. S. Mulekar

has the largest number of these K neighbors. For example, suppose kg of the nearest

K neighbors belong to group g(g = 1, 2, . . . , G) such that G g=1 kg = K, then
the new observation is placed in the group c if kc = max {kg , g = 1, 2, . . . , G}.
Note that there is a possibility that no such unique maximum exists for a given
new observation and a chosen K, thus resulting in ties. Although not exactly a
group inclusion probability, these nearest neighbors can be used to provide a group
membership indicator of the new observation using relative fractions (kg /K), g = 1,
2, . . . , G.
Now the question is: how to choose value of K, the number of nearest neighbors
to be used? Given a large dataset one can always use cross-validation and choose the
K value corresponding to the lowest misclassification rate in the validation dataset.
Note that choice of a too small value for K indicates that the space generated
by the explanatory variables is divided into many small subspaces and the class
membership of a new observation depends on which subspace the new observation
belongs to. In that case outliers in the original dataset can create problems in
predicting the class membership of a new observation that is close to the outlier
resulting in a higher variance in prediction. However, choice of a large value for K
basically leads to division of the training data space into G smooth subspaces which
in turn creates the problem of misclassification of any outlier of these√ subspaces and
subsequently higher bias in prediction. As a rule of thumb, K = n is considered
to be a sensible choice for number of classes in practice. If the number of groups in
the data is 2 (i.e., G = 2), then K should be an odd number to avoid the possibility
of ties in group membership indicators.
The most popular choice for a distance measure is the Euclidean distance which

for a new observation, x0 = (x01 , . . . , x0p ) , is calculated as,
0
1

1 p  2
di = 2 x0i − xj i , i = 1, 2, . . . , n. (11)
j =1

Some other distance measures used commonly in practice are Hamming distance
(Hamming 1950) and Chebyshev distance (Grabusts 2011). Chomboon et al.
(2015) looked at eleven different distance measures and found that the Euclidean,
Chebyshev, and Mahalanobis distance measures perform well. For a synopsis of
different distance measures, please refer to Mulekar and Brown (2014). Different
explanatory variables tend to have different range of possible values and some
distance measures such as the Euclidean distance tend to be affected by the
range of measurements. Hence in practice, datasets are typically normalized before
classification to reduce the influence of explanatory variables with larger range of
measurements. When using a dataset with a large number of explanatory variables,
to reduce the computation time, a dimension reduction technique such as principal
component analysis (PCA) is used. To overcome the problem of choosing a value
for K, Samworth (2012) suggested the use of weighted nearest neighbor algorithm in
which instead of choosing K nearest neighbors (i.e., essentially assigning a weight
A Primer of Statistical Methods for Classification 117

of 1/K to K nearest neighbors and 0 to the remaining observations in the training

dataset while assigning a class to the new observation), all observations in the
training dataset are assigned a weight using some optimal weighting scheme. When
dealing with a big dataset, an approximation to the method of nearest neighbors
proposed by Har-Peled et al. (2012) is useful.

3.3 Support Vector Machine

Support vector machine (SVM) is a class of hard classifiers. For a binary clas-
sification with p explanatory variables, an SVM classifier constructs a (p − 1)-
dimensional hyperplane in the pth dimension to maximize the margin. Here margin
refers to the distance between the observation closest to the boundary of a group
and the remaining groups. Points on or closest to the boundary of decision surface
are called support vectors and they are used in learning models associated with
the classification algorithm. The idea behind SVM is to find that hyperplane which
provides the maximum margin from support vectors among infinitely many possible
hyperplanes that can separate two groups provided the two groups are completely
separable. For a binary classification, one hyperplane known as the maximum
margin hyperplane is constructed. For G > 2 groups, more than one such maximum
margin hyperplanes need to be created to separate groups and a combination of these
hyperplanes is used for the classification of a new observation.
Consider the case of binary classification, and assume that there actually exists a
linear hyperplane of the form

p
W (X) = w0 + wj Xj (12)
j =1

that can perfectly differentiate between two classes. Then a method described by
Vapnik and Lerner (1963) can be used to find a maximum margin hyperplane.
Maximum margin hyperplane is a hyperplane for which W(X) = 0. In SVM, only
support vectors obtained using the training data are used to estimate the coefficients
of explanatory variables in (12). Since the decision surface differentiates the classes
completely, the linear function in (12) should be positive for one group and negative
for another. Without any loss of generality, assume that for support vector(s) in
group 1, Ŵ (X) = −1 and for those in groupp 2, Ŵ (X) = 1. In order to maximize
the margin, it is sufficient to minimize j =1 wj2 subject to vi Ŵ (xi ) ≥ 1, i =
1, 2, . . . , n where vi = − 1 if yi = 1 and vi = 1 if yi = 2. Thus this hyperplane
can be obtained by minimizing the Lagrangian formulation,

n
L=− ai (vi W (xi ) − 1)
i=1
118 R. Dey and M. S. Mulekar

where ai ( i = 1, 2, . . . , n) are Lagrange multipliers. Once this hyperplane is

estimated, Ŵ (x0 ) is computed for any new observation x0 and the new observation
is assigned to the group 1 if the Ŵ (x0 ) < 0 and to group 2 if Ŵ (x0 ) > 0.
In many practical situations, a perfectly differentiating hyperplane does not
exist. For such situations, Cortes and Vapnik (1995) proposed a modification
to the maximum margin hyperplane to differentiate between two groups. They
proposed estimating
 the hyperplane
with the help of a hinge loss function,
h (x) = max 0, 1 − v Ŵ (x) . Note that unlike a linearly separable case where
νi Ŵ (x0 ) ≥ 1 ∀i; for linearly non-separable cases, the possibility of νi Ŵ (x0 ) < 0
exists for a few support vectors. So, the hinge loss function is 0 for such support
vectors and this is used to penalize such support pvectors while estimating the
decision boundary. Thus, instead of minimizing j =1 wj2 subject to vi Ŵ (xi ) ≥
1, i = 1, 2, . . . , n, function

p
1

n 
θ wj2 + max 0, 1 − vi Ŵ (xi ) (13)
n
j =1 i=1

is minimized where θ is a penalty parameter. This is known as the soft version of

SVM, although this is not a soft classifier.
Since a linear classifier does not always exist, researchers extended the idea
of SVM to find a non-linear classifier. Using ideas first promoted by Aizerman
et al. (1964), Boser et al. (1992) proposed the use of kernelization to obtain a
non-linear classifier by improving a classifier obtained via SVM. The trick is to
use a transformation Z of X such that the new transformed explanatory variables
Z(X) provide a better classifier than the one provided by X. Then, proceed to
obtain an SVM based on the new transformed explanatory variables, Z. A carefully
chosen transformation Z can possibly result in a linear classifier. Note that Z is not
observed or calculated from data but it is replaced by the kernel function, κ , such
that κ (xi , xl ) = z’i zl for i = l = 1, 2, . . . , n. There are many kernel functions in
use, but the most used Gaussian kernel (Schölkopf et al. 1997) is given by,
⎛ ⎞

p
 2
κ (xi , xl ) = exp ⎝−ξ xij − xlj ⎠ .
j =1

To develop a multiclass SVM classifier (i.e., for G > 2), there are few options
available. In a method known as one-against-all, G SVM classifiers are obtained
for each class separately and a new observation is assigned to the class chosen
by maximum number of these classifiers (Bottou et al. 1994). In another method
known as one-against-one (Kressel 1998), G(G − 1)/2 SVM classifiers each
separating a pair of classes are obtained and a new observation is assigned to the
class that is predicted by the most classifiers (Kressel 1998). Hsu and Lin (2002)
who compared their performances concluded that one-against-one performs better
A Primer of Statistical Methods for Classification 119

than one-against-all in most of the situations that they studied. There are many
modifications of SVM proposed by researchers from different fields that work better
in certain specific situations. Typically, SVM works really well if there exists a good
separation between classes or when the number of explanatory variables is large
compared to the sample size of the training dataset. SVM is not computationally
effective when using a very large training dataset.

3.4 Classification Trees

Classification trees (CT) are methods used to partition the space of explanatory
variables into disjoint subsets and assign a class to each subset by minimizing some
measure of misclassification also known as impurity. It is a visually pleasing method
and can be easily as well as effectively described to those from the non-scientific
communities. CT can handle large datasets as well as missing data, and it can easily
ignore bad explanatory variables. However, sometimes depending on the dataset CT
can produce a really bad partition of the space of explanatory variables leading to
high misclassification rates.
CT produces a flowchart or tree structure starting with a root node (one
explanatory variable) and then, proceeds with splits (internal nodes) until no split
is deemed necessary (leaf nodes). Each leaf node is assigned to a class. There are
many algorithms on how to select a root node, how to split a node, how many splits
of each node are needed, and when to stop splitting a node to make it a leaf node.
An example of classification tree is shown in Fig. 3. It shows classification of a
random sample of n = 78 from iris data by R.A. Fisher (Anderson 1935) using JMP
12 into one of the three classes using two explanatory variables petal-width and
petal-length.
The root node is typically chosen with an explanatory variable that provides
the lowest rate of misclassification. This is easily achieved when the number of
explanatory variables is small. For example, let there be two classes (G = 2) and
one explanatory variable (p = 1). Consider the rule for using two complementary
subgroups Ag created by a split with the explanatory variable such that, yi = g if
x1i ∈ Ag , g = 1, 2. For each split a Gini impurity measure is computed as,
⎛ ⎞

1
I (CT ) = ⎝1 − q
g (j )⎠
g=0 j =0

where

n  
I yi = j ; x1i ∈ Ag
q
i=1
g (j ) =  

n
I x1i ∈ Ag
i=1
120 R. Dey and M. S. Mulekar

Fig. 3 Classification tree using a random sample of iris data (n = 78)

is the misclassification rate for group g and the splits are chosen such that the
Gini impurity measure is minimized (Witten et al. 2011). As shown in Fig. 3,
LogWorth for each model defined as −log10 (p-value) is also another measure used
to decide where to split. The p-value can be based on a chi-square test for a split that
maximizes LogWorth value.
The first classification tree algorithm was proposed by Messenger and Mandell
(1972). However, Breiman et al. (1984) have provided what became the most
popular classification tree algorithm, namely classification and regression trees
(CART). Several improved versions of it were proposed later and are still used in
practice. Many modifications of the CART method have been proposed for various
reasons, but mainly because CART produces biased and high-variance trees, i.e.,
changing the training set can drastically change the tree diagram. A few Bayesian
versions of this algorithm are also available in the literature (Chipman et al. 1998;
Denison et al. 1998). Loh (2009) provides a classification algorithm, called GUIDE
which is computationally faster and incorporates nearest neighbor algorithm to
improve the CT.
To reduce the variance in CT, two new methods were proposed, namely the
bagging method (Breiman 1996) and the random forests method (Breiman 2001).
Note that these methods do not produce a tree diagram but they focus on obtaining
many classification trees from the training data so that a new observation that needs
to be classified is put into the class suggested by majority of these trees. Bagging is
simply achieved by obtaining bootstrap samples with replacement from the training
A Primer of Statistical Methods for Classification 121

data. Random forests are similar to bagging but in each tree only a randomly
√
chosen subset of typically p number of explanatory variables is considered when
determining the nodes.
Freund and Schapire (1997) introduced the concept of boosting which aims to
reduce bias in a CT. This is achieved by refitting the data into trees with higher
weights for misclassified data points. In the initial calculation of the first CT, all
data points are given equal weight. However, the weights are updated after each
iteration and the impurity measure is updated by assigning higher weights to the
misclassified observations. Then the final classifier is selected via weighted average
of the trees.

4 Assessment and Comparison of the Performance

of Classifiers

The performance of a classifier is typically judged by cross-validating the classifi-

cation rule with a separate dataset of size s, called the validation data. Sometimes
cross-validation is also used to estimate unknown parameters such as the number of
neighbors to be considered in KNN method. In the absence of a separate validation
data, the idea of Jackknife sampling (Quenouille 1949, 1956; Tukey 1958) is used to
obtain a K-fold cross-validation. In this special case of cross-validation, the training
dataset is divided into K smaller datasets of equal size, and (K − 1) of them are used
as the training data and the remaining Kth one as the validation dataset. This process
is repeated K times until each of them is used once as the validation data.
The simplest performance measure of a classifier is the misclassification rate
R(0 ≤ R ≤ 1), which is the proportion of validation sample that is misclassified.
Hence a small value of R(R → 0) is an indication of more accurate classification.
Although a very simple metric, this is an effective measure of performance. It works
well as long as the cost of misclassification for and sample sizes from all classes are
relatively similar. If sample sizes differ considerably, then the use of an uncertainty
coefficient is recommended (Mills 2011). Uncertainty coefficient U is calculated as
U = (H − Hc )/H (0 ≤ U ≤ 1) where

G
H =− P (Y = g) log (P (Y = g))
g=1

and

G   
Hc = − P Y = g, Ŷ = l log P Y = g|Ŷ = l
g=1 l=1
122 R. Dey and M. S. Mulekar

can be estimated from the training data. Larger values of U indicate a better
classifier. If the accuracy of classification for only one class is very important, then
one can calculate the sensitivity (also known in medicine as the true positive rate
or in machine learning as the recall rate) for that class. Sensitivity also takes values
between 0 and 1 but a good classifier is expected to have a higher sensitivity. The
sensitivity for class g can be calculated as,
n
i=1 I (yi = g, y
3i = g)
seng = n n
i=1 I (yi = g, 3
yi = g) + i=1 I (yi = g, y
3i = g)

where I is the indicator variable taking values 1 or 0 depending on whether the

condition is satisfied or not.
Besides misclassification rate, sensitivity, and uncertainty coefficient, there are
many other performance measures available to judge classification methods, a
detailed discussion of which is provided by Hand (2012).
Some articles dedicated to comparison of different classification methods are
available. The earlier research was mostly focused on comparing logistic regression
against LDA (Hosmer et al. 1983 and McLachlan and Byth 1979). Their outcomes
indicate that LDA is a better performer if the explanatory variables are normally
distributed but the advantage diminishes as sample size becomes larger. Meshbane
and Morris (1996) recommended that QDA should be used instead of LDA if the
distributions of explanatory variables are skewed. After comparing outcomes using
classification tree and KNN, Liu and White (1995) concluded that KNN performs
better than classification tree unless the number of explanatory variables is large.
A study by Bhattacharya et al. (2011) compared SVM to logistic regression for
detecting credit card fraud and found no advantage in using more complicated
method like SVM over simpler logistic regression. Finch and Schneider (2006)
compared performances of logistic regression, discriminant analysis, and classifi-
cation trees based on simulated data while Kiang (2003) compared performances
of logistic regression, LDA, KNN, and classification tree based on a separate set
of simulated data. Asparoukhov and Krzanowski (2001) compared performances
of all but one (namely SVM) classifiers mentioned in this paper using five real-life
datasets for binary classification. They also discussed the effect of choosing different
sized training set along with changing the number of explanatory variables. Steel et
al. (2000) argue that simply comparing the methods is not completely meaningful
unless the model selection process (i.e., the choice of explanatory variables in the
final prediction model) is included. The only conclusion that can be drawn from
all these studies is that there is no winner among all methods that work in every
situation very effectively. The performance of a classifier depends very much on the
dataset for which a classification is needed.
In this section, we compare the performance of six classifiers discussed earlier
using two simulated datasets and three real-life datasets with respect to the
misclassification rates and uncertainty measure U. All computations were done
using available R packages rpart, e1071, class, naivebayes, and MASS. Of the two
simulated datasets, one is visually separable albeit not linearly while the other is
A Primer of Statistical Methods for Classification 123

Fig. 4 Visualization of the simulated datasets

Table 1 Comparison of misclassification rates for different classifiers for five datasets
Iris Skin Glass Bivariate normal Bivariate uniform
Logistic 0.09804 0.0401 NA 0.288 0.495
LDA 0.03922 0.0505 0.2364 0.300 0.495
NB 0.03922 0.0245 0.4545 0.304 0.077
KNN 0.03922 0.0048 0.1818 0.002 0.000
SVM 0.05882 0.0051 0.3273 0.312 0.008
CT 0.07843 0.0232 0.2545 0.320 0.096

Table 2 Comparison of uncertainty measure for different classifiers for five datasets
Iris Skin Glass Bivariate normal Bivariate uniform
Logistic 0.71463 0.72119 NA 0.12654 0.00116
LDA 0.86424 0.71310 0.52570 0.11903 0.00116
NB 0.88589 0.80570 0.43788 0.11408 0.66483
KNN 0.88589 0.94258 0.62021 0.96238 1.00000
SVM 0.79531 0.93916 0.25449 0.06982 0.93602
CT 0.78020 0.80141 0.52453 0.10053 0.54546

not separable using any reasonable curve and provides a challenge in terms of
classification (see Fig. 4). Both datasets have two explanatory variables as presented
in scatterplots in Fig. 4 where each class is represented by separate point type and
color. For NB classifier,
√ Gaussian prior was used. For KNN classifier, the next larger
odd integer to n was used as the value of K, except in one real data example
(skin data), where this value was too large due to large dataset. The observed
misclassification rates for six methods and five examples are listed in Table 1 and
the uncertainty coefficients are listed in Table 2.
124 R. Dey and M. S. Mulekar

Example 1 (Iris) Consider the famous iris dataset by R.A. Fisher (Anderson 1935)
described in the Introduction. Fifty observations are available for each type of
iris. Of the 150 measurements available, a training dataset of 99 observations was
created with 33 observations each from three groups. The misclassification rate
was estimated based on the remaining 51 observations that constituted a validation
sample. Misclassification rates lower than 0.10 (Table 1) and uncertainty measures
over 0.70 (Table 2) show that all the methods did a commendable job of correct
classification for this data. However, NB, KNN, and LDA are slightly better than
other classifiers.
Example 2 (Skin) Refer to the skin segmentation dataset from the UCI machine
learning repository (Bhatt et al. 2009). This dataset contains 245,057 observations
randomly sampled from photos of faces of people of different age group, gender,
and color. Of those, 50,589 observations are for samples of skin while the rest are
for samples of non-skin parts of the face. The three explanatory variables in this
example are RGB triplet, i.e., red, green, and blue colors used in displaying images.
RGB values are typically given as an integer value in the range of 0–255, and
combined together they determine the color of the image which in this case is part
sampled. Distribution of RGB pixels for skin data is presented in a 3-dimensional
plot in Fig. 5. Without rotating the plot around three axes it is difficult to tell if
there is clear distinction or some overlap between two groups, skin and non-skin. A
training sample of size 150,000 was used, out of which 30,000 were skin samples.

Fig. 5 Visualization of RGB pixel distribution for skin data

A Primer of Statistical Methods for Classification 125

Tables 1 and 2 show that KNN and SVM perform best for this data, followed by NB
and CT. To save computation time, K = 19 was used for KNN algorithm.
Example 3 (Glass) Consider the glass dataset from UCI machine learning repos-
itory (Lichman 2013). The original dataset describes six types of glass samples
along with the refractive index and weight percent of oxides formed with sodium,
magnesium, aluminum, silicon, potassium, calcium, barium, and iron in the sample.
Since some of the classes have small sample sizes, only three types of glass
were used for classification purpose in this example. They are float-processed
building window glasses (70 measurements), nonfloat-processed building window
glasses (76 measurements), and non-window headlamp glasses (29 measurements).
Fifty samples each from two classes of building window glasses and 20 samples
from headlamp were used as the training data. Simulations to obtain parameters
for a multinomial logistic regression failed due to non-convergence of iterations.
Outcomes in Tables 1 and 2 indicate that KNN performs the best followed by CT
and LDA.
Example 4 (Bivariate Uniform) Consider two independent univariate uniform
distributions, namely Xi ∼ Uniform(−1.25, 1.25) for i = 1, 2. A sample of 3000
observations was generated with seed 1234. With the unit circle providing the class
boundary, the i-th observation is assigned to group 1 if x1i2 + x 2 ≤ 1 and to group 2
2i
otherwise. The first 2000 observations generated were used as a training data and the
remaining 1000 as a validation sample. In this training dataset, 1012 observations
were from group 1 and the remaining 998 from group 2. In the validation dataset,
520 observations were from group 1 and the remaining 480 from group 2. Note that
in this situation a linear classifier is not supposed to perform well because of non-
normal distributions and that is reflected in the misclassification rates listed in Table
1 and uncertainty measures listed in Table 2. Logistic regression and LDA seems
to be only as good as a coin toss in this situation whereas KNN and SVM perform
admirably well.
Example 5 (Bivariate Normal) Now consider the bivariate normal populations. A
sample of 1500 observations from two homoscedastic bivariate normal distributions
that differ only in mean vector was generated using the mvtnorm package in R
with seed 5678, resulting in a total sample of size 3000. The difference in the
mean vectors and the variance–covariance matrix used in the simulation were,
respectively,

   
1.0 1.0 0.5
μ1 − μ2 = and = .
1.0 0.5 1.0

The first 1000 rows were used as a training sample for both groups (resulting in a
total sample size of 2000) and the remaining 500 as the validation sample (resulting
total sample size 1000). Tables 1 and 2 show KNN as a clear winner while the other
classifiers are almost equally bad. Although we expected LDA to perform better, to
our surprise the results say otherwise.
126 R. Dey and M. S. Mulekar

5 Concluding Remarks

In this paper, the basic ideas that dominate the world of statistical classification were
described. Detailed discussions of them are scattered in different textbooks, but none
discusses them all together. For example, logistic regression is discussed in detail
by Kleinbaum and Klein (2010), LDA by McLachlan (2004), SVM by Steinwart
and Christmann (2008), classification trees by Breiman et al. (1984), and different
classification methods by Izenman (2008) and James et al. (2013).
For data with highly correlated explanatory variables or a large number of
explanatory variables, the use of some dimension reduction technique such as
principal component analysis, low variance filter, and high correlation filter before
classification is recommended (Farcomeni and Greco 2015). In cases where p > n,
dimension reduction becomes necessary. Alternatively, although random forests
method is not a dimension reduction technique for explanatory variables, in cases
√
where p < n this method can be effectively used without reducing dimension of
explanatory variables.
A very basic question on this topic should be about the preference for any
particular classification method. Alternatively, should there be preference for a
certain classification method over the others. It depends on the circumstances.
There is no single method that stands out as the best. Typically for complex
problems in which the misclassification rate is higher among all classifiers, the
use of soft classifiers is recommended. However, hard classifiers remain popular
as their outcomes are easier to interpret in practice. Also hard classifiers like SVM
and KNN generally provide good outcomes as seen from the situations discussed
here. In this age of computation, the most recent research emphasis is on effective
ways of implementing bagging and random forests (James et al. 2013) which can
be computationally more effective than other classifiers like KNN. Liu et al. (2011)
describe a suave large-margin unified machine that combines margin-based hard and
soft classifiers, and that hard classifiers tend to perform better than soft classifiers
when the classes are either easily separable or when the training sample size is
relatively small compared to number of explanatory variables.
Research over the years has led to the development of many classifiers. As a
result, the toolbox from which a classifier can be chosen provides an extensive list of
options which to some extent depends on software used by and the computing power
available for the researcher. Also comparative performance of different classifiers
is changing with changing technology and results of past studies might lead to
different conclusions with the current technology. Thus, one can entertain the idea
of using all possible classifiers and assign a new observation to a class assigned by
most of the classifiers.
A Primer of Statistical Methods for Classification 127

References

Agresti, A.: Categorical Data Analysis. Wiley, Hoboken (2013)

Aizerman, M.A., Braverman, E.M., Rozonoer, L.I.: Theoretical foundations of the potential
function method in pattern recognition learning. Autom. Remote. Control. 25, 821–837 (1964)
Anderson, E.: The irises of the Gaspe Peninsula. Bull. Am. Iris Soc. 59, 2–5 (1935)
Asparoukhov, O.K., Krzanowski, W.J.: A comparison of discriminant procedures for binary
variables. Comput. Stat. Data Anal. 38, 139–160 (2001)
Bayes, T.: An essay towards solving a problem in the doctrine of chances. Philos. Trans. 53, 370–
418 (1763)
Berkson, J.: Applications of the logistic function to bioassay. J. Am. Stat. Assoc. 9, 357–365 (1944)
Bhatt, R.B., Sharma, G., Dhall, A., Chaudhury, S.: Efficient Skin Region Segmentation Using Low
Complexity Fuzzy Decision Tree Model. IEEE-Indicon, Ahmedabad (2009)
Bhattacharya, S., Sanjeev, J., Tharakunnel, K., Westland, J.C.: Data mining for credit card fraud: a
comparative study. Decis. Support. Syst. 50, 602–613 (2011)
Boser, B.E., Guyon, I.M., Vapnik, V.N.: A training algorithm for optimal margin classifiers. In:
Proceedings of the Fifth Annual Workshop on Computational Learning Theory – COLT ’92. p.
144 (1992)
Bottou, L., Cortes, C., Denker, J.S., Drucker, L., Guyon, I., Jackel, L., LeCun, Y., Muller, U.A.,
Sackinger, E., Simard, P., Vapnik, V.N.: Comparison of classifier methods: a case study in
handwriting digit recognition. Int. Conf. Pattern Recognit. 2, 77–87 (1994)
Breiman, L.: Bagging predictors. Mach. Learn. 24(2), 123–140 (1996)
Breiman, L.: Random forests. Mach. Learn. 45, 5–32 (2001)
Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J.: Classification and Regression Trees.
Wadsworth, Belmont (1984)
Burrus, C.S., Barreto, J.A., Selesnick, I.W.: Iterative reweighted least squares design of FIR filters.
IEEE Trans. Signal Process. 42(11), 2922–2936 (1994)
Chipman, H.A., George, E.I., McCulloch, R.E.: Bayesian CART model search. J. Am. Stat. Assoc.
93, 935–948 (1998)
Chomboon, K., Pasapichi, C., Pongsakorn, T., Kerdprasop, K., Kerdprasop, N.: An empirical study
of distance mreics for K-nearest neighbor algorithm. 3rd International Conference on Industrial
Application Engineering, 280–285 (2015)
Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. 20(3), 273–297 (1995)
Cover, T., Hart, P.: Nearest neighbor pattern classification. IEEE Trans. Inf. Theory. 13(1), 21–27
(1967)
Cox, D.R.: Analysis of Binary Data. Chapman and Hall, London (1969)
Cramer, J.S.: The origins and development of the logit model. In: Cramer, J.S. (ed.) Logit Models
from Economics and Other Fields, pp. 149–158. Cambridge University Press, Cambridge
(2003)
Denison, D.G.T., Mallick, B.K., Smith, A.F.M.: A Bayesian CART algorithm. Biometrika. 85,
363–377 (1998)
Farcomeni, A., Greco, L.: Robust Methods for Data Reduction. CRC Press, Boca Raton (2015)
Finch, W.H., Schneider, M.K.: Misclassification rates for four methods of group classification.
Educ. Psychol. Meas. 66(2), 240–257 (2006)
Fisher, R.A.: The use of multiple measurements in taxonomic problems. Ann. Eugenics. 7(2), 179–
188 (1936)
Freund, Y., Schapire, R.E.: A decision-theoretic generalization of on-line learning and an applica-
tion to boosting. J. Comput. Syst. Sci. 5(1), 119–139 (1997)
Grabusts, P.: The choice of metrics for clustering algorithms. Proceedinhs of the 8th International
Scientific and Practical Conference, 11, 70–76 (2011)
Gurland, J., Lee, I., Dahm, P.A.: Polychotomous quantal response in biological assay. Biometrics.
16, 382–398 (1960)
128 R. Dey and M. S. Mulekar

Hamming, R.W.: Error detecting and error correcting codes. Bell Syst. Tech. J. 29(2), 147–160
(1950)
Hand, D.J.: Classifier technology and the illusion of progress. Stat. Sci. 21, 1–14 (2006)
Hand, D.J.: Assessing the performance of classification methods. Int. Stat. Rev. 80, 400–414 (2012)
Hand, D.J., Yu, K.: Idiot’s Bayes - not so stupid after all? Int. Stat. Rev. 69(3), 385–399 (2001)
Har-Peled, S., Indyk, P., Motwani, R.: Approximate nearest neighbor: towards removing the curse
of dimensionality. Theory Comput. 8, 321–350 (2012)
Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data Mining,
Inference and Prediction. Springer, New York (2001)
Hosmer, T., Hosmer, D.W., Fisher, L.L.: A comparison of the maximum likelihood and dis-
criminant function estimators of the coefficients of the logistic regression model for mixed
continuous and discrete variables. Commun. Stat. 12, 577–593 (1983)
Hsu, C.W., Lin, C.J.: A comparison of methods for multi-class support vector machines. IEEE
Trans. Neural Netw. 13(2), 415–425 (2002)
Izenman, A.J.: Modern Multivariate Statistical Techniques: Regression, Classification and Mani-
fold Learning. Springer, New York (2008)
James, G., Witten, D., Hastie, T., Tibshirani, R.: An Introduction to Statistical Learning: With
Applications in R. Springer, New York (2013)
Kiang, M.: A comparative assessment of classification methods. Decis. Support. Syst. 35, 441–454
(2003)
Kleinbaum, D.G., Klein, M.: Logistic Regression: A Self-learning Text, 3rd edn. Springer, New
York (2010)
Kressel, U.H.G.: Pairwise classification and support vector machines. In: Advances in Kernel
Methods: Support Vector Learning, pp. 255–268. MIT Press, Cambridge (1998)
Lange, K.: MM Optimization Algorithms. SIAM, Philadelphia (2016)
Lichman, M.: UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. University of
California, Irvine 2013
Liu, W.Z., White, A.P.: A comparison of nearest neighbor and tree-based methods of non-
parametric discriminant analysis. J. Stat. Comput. Simul. 53, 41–50 (1995)
Liu, Y., Zhang, H.H., Wu, Y.: Hard or soft classification? Large-margin unified machines. J. Am.
Stat. Assoc. 106(493), 166–177 (2011)
Loh, W.Y.: Improving the precision of classification trees. Ann. Appl. Stat. 3, 1710–1737 (2009)
Mahalanobis, P.C.: On the generalized distance in statistics. Proceedings of the National Institute
of Science in India, 2(1), 49–55, (1936)
Mai, Q., Zou, H.: Semiparametric sparse discriminant analysis in ultra-high dimensions. J.
Multivar. Anal. 135, 175–188 (2015)
Mantel, N.: Models for complex contingency tables and polychotomous response curves. Biomet-
rics. 22, 83–110 (1966)
McLachlan, G.J.: Discriminant Analysis and Statistical Pattern Recognition. Wiley-Interscience,
New York (2004)
McLachlan, G.J., Byth, K.: Expected error rates for logistic regression versus normal discriminant
analysis. Biom. J. 21, 47–56 (1979)
Menard, S.: Applied Logistic Regression Analysis, 2nd edn. Sage Publications, Thousand Oaks
(2002)
Meshbane, A., Morris, J.D.: A method for selecting between linear and quadratic classification
models in discriminant analysis. J. Exp. Educ. 63(3), 263–273 (1996)
Messenger, R., Mandell, L.: A modal search technique for predictive nominal scale multivariate
analysis. J. Am. Stat. Assoc. 67, 768–772 (1972)
Mills, P.: Efficient statistical classification of satellite measurements. Int. J. Remote Sens. 32, 6109–
6132 (2011)
Mulekar, M.S., Brown, C.S.: Distance and similarity measures. In: Alhaji, R., Rekne, J. (eds.)
Encyclopedia of Social Network and Mining (ESNAM), pp. 385–400. Springer, New York
(2014)
Murphy, K.P.: Machine Learning: A Probabilistic Perspective. MIT Press, Cambridge (2012)
A Primer of Statistical Methods for Classification 129

Nigam, K., McCallum, A., Thrun, S., Mitchell, T.: Text classification from labeled and unlabeled
documents using EM. Mach. Learn. 39(2-3), 103–134 (2000)
Nilsson, N.: Learning Machines: Foundations of Trainable Pattern-Classifying Systems. McGraw-
Hill, New York (1965)
Quenouille, M.H.: Problems in plane sampling. Ann. Math. Stat. 20(3), 355–375 (1949)
Quenouille, M.H.: Notes on bias in estimation. Biometrika. 43(3-4), 353–360 (1956)
Rao, R.C.: The utilization of multiple measurements in problems of biological classification. J. R.
Stat. Soc. Ser. B. 10(2), 159–203 (1948)
Rish, I.: An empirical study of the naive Bayes classifier. In: IJCAI Workshop on Empirical
Methods in AI, Sicily, Italy (2001)
Samworth, R.J.: Optimal weighted nearest neighbour classifiers. Ann. Stat. 40(5), 2733–2763
(2012)
Schölkopf, B., Sung, K., Burges, C., Girosi, F., Niyogi, P., Poggio, T., Vapnik, V.: Comparing
support vector machines with Gaussian kernels to radial basis function classifiers. IEEE Trans.
Signal Process. 45, 2758–2765 (1997)
Sebestyen, G.S.: Decision-making Process in Pattern Recognition. McMillan, New York (1962)
Snell, E.J.: A scaling procedure for ordered categorical data. Biometrics. 20, 592–607 (1964)
Soria, D., Garibaldi, J.M., Ambrogi, F., Biganzoli, E.M., Ellis, I.O.: A non-parametric version of
the naive Bayes classifier. Knowl.-Based Syst. 24(6), 775–784 (2011)
Srivastava, S., Gupta, M.R., Frigyik, B.A.: Bayesian quadratic discriminant analysis. J. Mach.
Learn. Res. 8, 1287–1314 (2007)
Steel, S.J., Louw, N., Leroux, N.J.: A comparison of the post selection error rate behavior of the
normal and quadratic linear discriminant rules. J. Stat. Comput. Simul. 65, 157–172 (2000)
Steinwart, I., Christmann, A.: Support Vector Machines. Springer, New York (2008)
Theil, H.: A multinomial extension of the linear logit model. Int. Econ. Rev. 10(3), 251–259 (1969)
Tjalling, J.Y.: Historical development of the Newton-Raphson method. SIAM Rev. 37(4), 531–551
(1995)
Tukey, J.W.: Bias and confidence in not quite large samples. Ann. Math. Stat. 29(2), 614–623
(1958)
Vapnik, V., Lerner, A.: Pattern recognition using generalized portrait method. Autom. Remote.
Control. 24, 774–780 (1963)
Wahba, G.: Soft and hard classification by reproducing Kernel Hilbert space methods. Proc. Natl.
Acad. Sci. 99, 16524–16530 (2002)
Witten, I., Frank, E., Hall, M.: Data Mining. Morgan Kaufmann, Burlington (2011)
A Doubly-Inflated Poisson Distribution
and Regression Model

Manasi Sheth-Chandra, N. Rao Chaganty, and Roy T. Sabo

1 Introduction

The Poisson distribution is the standard choice for modeling probabilities of count
data and can be modeled against covariates using the generalized linear models
framework (McCulloch and Searle 2001). A common extension of these models
is to account for over-dispersion, or situations where the count variability exceeds
the count mean, see Cameron and Trivedi (1998) for a summary of approaches to
verify and analyze over-dispersed count data. Another extension of the basic Poisson
model is to account for excess zeros. Both Cohen (1963) and Johnson and Kotz
(1969) describe the zero-inflated Poisson (ZIP) distribution, and Lambert (1992)
extended this distribution to account for covariates that could simultaneously model
the counts as well as the probability of particular zeros being in excess. More details
on Poisson regression models and zero-inflated models are provided by Hall (2000),
Bae et al. (2005), Coxe et al. (2009), and Hall and Shen (2010).
One example of count data is a patient’s length of stay (LOS) before hospital
discharge, which is often used as a simple count measure of the cost of care
(Marazzi et al. 1998). The use of LOS as a measure of cost assumes that a patient
with a larger LOS is of more cost to a hospital than a patient with a shorter LOS.

M. Sheth-Chandra
Center for Global Health, Old Dominion University, Norfolk, VA, USA
e-mail: msheth@odu.edu
N. R. Chaganty ()
Department of Mathematics and Statistics, Old Dominion University, Norfolk, VA, USA
e-mail: rchagant@odu.edu
R. T. Sabo
Department of Biostatistics, Virginia Commonwealth University, Richmond, VA, USA
e-mail: rsabo@vcu.edu

© Springer Nature Switzerland AG 2019 131

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_7
132 M. Sheth-Chandra et al.

Table 1 Observed length of LOS Frequency LOS Frequency LOS Frequency

stay (LOS)
0 55 5 20 10 3
1 35 6 13 11 1
2 35 7 8 12 4
3 75 8 4 13 0
4 40 9 5 14 1

This measure is a natural candidate for a ZIP model, as many patients are served in
outpatient scenarios for which LOS is recorded as zero (LOS = 0). However, other
frequencies can be inflated as well. Consider the data presented in Table 1, which
shows a sample LOS data. While the frequency of zero-valued LOS is certainly
high, we notice that LOS = 3 is also inflated relative to all other values. This may
be due to patients staying three nights at the hospital to receive inpatient treatment
followed by recovery time. As another example, consider the dental epidemiology
data presented in Bohning et al. (1997, 1999). Here, the decayed, missing, filled teeth
(DMFT)-index is used to measure the dental status of 1013 school children of age
7, which is a count of the number of decayed, missing, or filled teeth. This study
focused on the eight deciduous molars, and the DFMT was measured at baseline
and at 1 year. The goal of the study was to compare the absolute change in DMFT
(δ(DMF T ) = |DMF T1 −DMF T2 |) between four types of decay prevention, their
combination, and a control, in the presence of other covariates. In this case zero was
inflated as most of the children exhibited no change in dental status, and as such the
zero count—corresponding to no improvement and/or consistent dental care—was
inflated. However, δ(DF MT ) scores of one were also inflated. The count of one is
for children that showed improvement in only one cavity.
In both the length of stay and dental epidemiology examples, there were inflated
instances of count k > 0 in addition to an inflation of zeros. As ZIP models can
only account for excess zeros, here we introduce a doubly-inflated Poisson (DIP)
distribution. This distribution can probabilistically model excessive zeros as well as
excess counts at some positive integer value k. This model can also be adapted to
the generalized linear models framework to account for covariates in a regression-
like setting in a similar manner as the ZIP model. The added advantage of the DIP
is that it can model covariates against the preponderance of both excess zeros and
k values in addition to the count outcome. Lin and Tsai (2013) studied excessive
zero and k responses in the context of health survey data. However, they did not
consider method of moments estimation and comparisons with maximum likelihood
using asymptotic relative efficiency criteria. They also did not derive analytically the
elements for Fisher information as we did in this paper.
This chapter is organized as follows. The DIP distribution is presented in
Sect. 2. In addition to the distributional form and parameterization of the DIP
model, likelihood and moment-based estimation processes are outlined and their
asymptotic efficiencies are compared. The hospital LOS data are used to exemplify
this distribution. The DIP model is extended to the generalized linear model
A Doubly-Inflated Poisson Distribution and Regression Model 133

framework in Sect. 3, where models are developed for subject-specific (ungrouped)

and frequency (grouped) data. These models are then used to analyze the dental
epidemiology data. A brief summary concludes the manuscript in Sect. 4, including
an outline of additional extensions of the DIP model. Finally, the Appendix contains
some analytical first and second- order partial derivatives of the log-likelihood
functions.

2 Doubly-Inflated Poisson Distribution

2.1 Parameterization

The doubly-inflated Poisson model can be constructed as a mixture of binomial

(2, p) and Poisson (λ) distributions where 0 < p < 1 and λ > 0, which can be
interpreted as a mixture of a Poisson process and two degenerate distributions with
masses concentrated at zero and at a positive integer value k. The probability mass
function of a DI P (p, λ) random variable Y can be written as:
⎧
⎪
⎨p + q f1 (0; λ),
⎪ for y = 0;
2 2

f (y; p, λ) = 2pq + q f1 (k; λ), for y = k;

2 (1)
⎪
⎪
⎩q 2 f (y; λ), for y = 1, 2, . . . = k.
1

where q = 1 − p and f1 (y; λ) = λy exp(−λ)/(y !) is the Poisson mass function

with mean λ. This simple choice of modeling two probabilities with one parameter
and its compliment guarantees identifiability. Properties of the distribution (1)
can be found in Sheth-Chandra (2011). Based on (1), the probabilities of inflated
instances of y = 0 and y = k outside the standard Poisson are p2 and 2pq,
respectively. The probabilities for all counts governed by the standard Poisson
distribution are then scaled downward by q 2 . Note that as p → 0 this model reduces
to the ordinary Poisson distribution. A doubly-inflated model could be created with
separate proportions p1 and p2 to account for inflations at y = 0 and y = k
respectively, but the additional parameter adds considerable complexity to both the
model parameterization and estimation and as such this model will not be considered
here.

2.2 Methods of Estimation

Maximum Likelihood

Assuming our data consists of independent count responses yi , i = 1, . . . , n,

generally governed by a Poisson(λ) distribution but with inflated counts at 0 and k.
134 M. Sheth-Chandra et al.

The log-likelihood function of the DI P (p, λ) model given in (1) can be written as

(p, λ|y) = n0 log[f (0; p, λ)] + nk log[f (k; p, λ)] + log[f (yi ; p, λ)],
i: yi =0,k
(2)

where n0 represents the number of zero counts and nk represents the number of
k counts. Differentiating log-likelihood function (2) with respect to the p and
λ (assuming k is known) and solving the resulting score functions for those
parameters yield the maximum likelihood (ML) estimates (3 p, 3λ). The Newton-
Raphson algorithm can be used to find numerical solutions as the score equations are
p) and σ 2 (3
not in the closed form. The asymptotic variances σ 2 (3 λ) of the maximum
likelihood estimates p3 and 3 λ, respectively, are obtained by taking the diagonal
elements of the inverse of the Fisher information matrix. The analytical formulas for
the elements of the Fisher information matrix are given in Appendix “Information
Matrix”.

Method of Moments

To obtain moment estimates for p and λ, we note that the first two population
 
moments are μ1 = E(Y ) = 2pqk + q 2 λ and μ2 = E(Y 2 ) = 2pqk 2 + q 2 λ + λ2 .

We equate these to their corresponding sample moments y 1 = n1 ni=1 yi and

y 2 = n1 ni=1 yi2 , which yields the following two equations:
)
y1 = 2pqk + q 2 λ
  (3)
y2 = 2pqk 2 + q 2 λ + λ2 .

Solving Eq. (3) numerically for p and λ using Newton-Raphson yields moment
p, 4
estimators (4 λ) for p and λ, respectively. If we let
⎛ ∂μ ∂μ ⎞
1 1
⎜ ∂p ∂λ ⎟
D = ⎝ ∂μ ∂μ ⎠ ,
2 2
∂p ∂λ

a matrix of first-order partial derivatives of the population moments, and letting 

be the covariance matrix of Y and Y 2 given by

V ar(Y ) Cov(Y, Y 2 )
= , (4)
Cov(Y 2 , Y ) V ar(Y 2 )
A Doubly-Inflated Poisson Distribution and Regression Model 135

then the asymptotic covariance matrix for the moment estimators (4p, 4
λ) is given
−1
 T −1
by A = (D)  D (see Theorem A.1 in Chaganty and Shi (2004)). The
diagonal elements of covariance matrix A are the asymptotic variances σ 2 (4
p) and
σ 2 (4 4 and 4
λ) of the moment estimators p λ, respectively.

2.3 Asymptotic Relative Efficiencies

To compare the performance of the ML and moment estimators we calculate

asymptotic relative efficiencies (ARE), taking the form of ratios of the asymptotic
variances e(4 p, p3) = σ 2 (3 p) and e(4
p)/σ 2 (4 λ, 3
λ) = σ 2 (3 λ)/σ 2 (4
λ), respectively.
Ratios less than 1 imply the ML estimators are more efficient than the moment
estimators, ratios greater than 1 imply the moment estimators are more efficient
than the ML estimators, and ratios close to 1 imply the two estimators are equally
efficient.
Table 2 presents efficiencies for p (e(4
p, p3)) and λ (e(4λ, 3λ)) for a DI P (p, λ)
model with inflated 0’s and 3’s. The efficiencies are calculated for various values of
p ∈ (0, 1) and λ ∈ [3, 9]. Inspection of Table 2 shows an inverse relationship
between the efficiencies for p and λ. For small values of p and λ—implying
low rates of inflation and a low mean for the Poisson counts—the ML estimators
are much more efficient than the moment estimators at estimating p than for λ.
Conversely, the ML and moment estimators are nearly as efficient in estimating
p for large values of p and λ, but the ML estimators are much more efficient in
estimating λ than the moment estimators for those same scenarios. In general, since
the efficiencies for estimating λ are nowhere greater than 0.8, the ML estimators can
be said to outperform the moment estimators overall.
Table 3 presents AREs e(4 p, p 3) and e(4
λ, 3λ) for the DIP model with inflated
0’s and 6’s. A similar pattern to that shown in Table 2 is also shown here, with the
efficiencies for p and λ having an inverse relationship for small and large values of

Table 2 Asymptotic relative efficiencies for DIP (p, λ) model for inflated 0’s and 3’s
e(4 3)
p, p e(4
λ, 3
λ)
p λ=3 λ=5 λ=7 λ=9 λ=3 λ=5 λ=7 λ=9
0.1 0.0766 0.0859 0.4343 0.6629 0.7709 0.2219 0.5807 0.6988
0.2 0.2926 0.2592 0.6384 0.8174 0.2930 0.3895 0.6172 0.6973
0.3 0.0819 0.4404 0.7524 0.8854 0.2784 0.4869 0.6215 0.6712
0.4 0.2594 0.5853 0.8205 0.9223 0.4875 0.5419 0.6089 0.6302
0.5 0.4475 0.6953 0.8659 0.9454 0.5946 0.5746 0.5892 0.5829
0.6 0.6106 0.7808 0.9007 0.9614 0.6555 0.5951 0.5678 0.5352
0.7 0.7427 0.8499 0.9297 0.9735 0.6938 0.6086 0.5469 0.4907
0.8 0.8478 0.9075 0.9551 0.9835 0.7197 0.6180 0.5276 0.4495
0.9 0.9319 0.9569 0.9784 0.9921 0.7384 0.6248 0.5101 0.4132
136 M. Sheth-Chandra et al.

Table 3 Asymptotic relative efficiencies for DIP (p, λ) model for inflated 0’s and 6’s
e(4 3)
p, p e(4
λ, 3
λ)
p λ=3 λ=5 λ=7 λ=9 λ=3 λ=5 λ=7 λ=9
0.1 0.1877 0.0152 0.0023 0.1468 0.3659 0.6813 0.0178 0.3197
0.2 0.3803 0.0274 0.1170 0.3544 0.5086 0.6890 0.3013 0.4399
0.3 0.5413 0.1692 0.2978 0.5081 0.4549 0.6916 0.4411 0.4854
0.4 0.6653 0.3464 0.4619 0.6211 0.3350 0.6928 0.5071 0.5055
0.5 0.7617 0.5090 0.5969 0.7105 0.2449 0.6934 0.5442 0.5158
0.6 0.8363 0.6461 0.7074 0.7851 0.1866 0.6939 0.5676 0.5216
0.7 0.8939 0.7595 0.7991 0.8491 0.1486 0.6942 0.5837 0.5251
0.8 0.9382 0.8537 0.8764 0.9053 0.1229 0.6944 0.5953 0.5273
0.9 0.9727 0.9328 0.9426 0.9552 0.1047 0.6945 0.6041 0.5288

Fig. 1 Log-likelihood of LL
LOS data using DIP model 500
with parameters (p, λ)

1000

1500

2000

2500

1.0
0.8
12 0.6
10 8 0.4
6 4 0.2
2 0.0
LAMBDA P

the target parameters. However, in this case we see that the efficiencies for λ here
are never larger than 0.7, implying that the moment estimators are not nearly as
efficient as the ML estimators.

2.4 Hospital Length of Stay Example

Returning to the LOS data from Table 1, recall that there were inflated counts
of patients who are received in outpatient settings (n0 = 55), as well as inflated
counts of patients receiving inpatient care for 3 days (n3 = 75). This phenomenon
of double inflation makes the DI P (p, λ) a natural choice to model the inflation
and count parameters. The log-likelihood for this model is presented in Fig. 1, with
A Doubly-Inflated Poisson Distribution and Regression Model 137

the peak matching the numerical ML estimates of p 3 = 0.161 and 3 λ = 3.604.

The standard errors for the two estimates are 0.023 and 0.126, respectively, and the
maximum log-likelihood is −666.0, at the ML estimates.

3 Doubly-Inflated Poisson Regression Modeling

In this section we allow parameters in the DI P (p, λ) to be modeled by covariates

in the generalized linear model framework McCulloch and Searle (2001). We focus
on two general situations: one where the ith subject has a unique vector of l
covariates xi = (xi1 , . . . , xil )T , and another where the covariates occur in s < n
distinct combinations, such that several subjects are linked to the same covariate
combination. In both cases, the covariates are separately linked to parameters p
and λ via vectors of coefficients γ and β, respectively. Assuming canonical link
functions for a binomial proportion and Poisson count, the structural component
of the covariates and coefficients is modeled using logit and log link functions,
respectively, such that for the ith subject we have
p
logit(pi ) = x i γ = γ0 + γ1 xi1 + . . . + γmp ximp
(5)
log(λi ) = x λi β = β0 + β1 xi1 + . . . + βmλ ximλ
p
where x i is the ith row of the set of covariates used to estimate pi , and x λi is the ith
row of the set of covariates used to estimate λi . Here mp and mλ denote the number
of covariates used for modeling p and λ, respectively, and these sets of covariates
need not be identical for the two models in (5). Note exp(γl ) is the multiplicative
effect on the binomial probability from a 1-unit increase in xil at the fixed levels
of the other covariates. If xil is a categorical indicator variable, then exp(γl ) is the
conditional odds ratio of the inflated counts. Similarly, a one-unit increase in xil has
a multiplicative impact of exp(βl ) on λi , and is equal to the relative risk if xil is a
categorical indicator.

3.1 Subject-Specific Data

When the count data y1 , . . . , yn are accompanied by subject-specific sets of

covariates (as shown in Table 4), the log-likelihood function (2) of the DI P (p, λ)
138 M. Sheth-Chandra et al.

Table 4 General layout of Subject Response Covariates

raw count data
1 y1 x11 . . . x1l
2 y2 x21 . . . x2l
.. .. .. .. ..
. . . . .
n yn xn1 ... xnl

regression model can be rewritten as


 exp(−λi )λki
(pi , λi |y) = log pi2 + qi2 exp(−λi ) + log 2pi qi + qi2
k!
{i:yi =0} {i:yi =k}

 exp(−λi )λyi
i
+ log qi2 , (6)
yi !
{i:yi =0,k}

which can be reexpressed in terms of the regression parameters γ and β as

n 

p p
(γ , β) = −2 log 1 + exp(x i γ ) + log exp(2x i γ ) + exp(− exp(x λi β))
i=1 {i:yi =0}

 exp(kx λi β − exp(x λi β))
p
+ log 2 exp(x i γ ) +
k!
{i:yi =k}


+ yi x λi β − exp(x λi β) − log(yi !) .
{i:yi =0,k}

γ ,3
The maximum likelihood estimate (3 β) is the solution to the equations

∂(γ , β) ∂(γ , β)
= 0, = 0.
∂γ ∂β

These regression parameters can be solved iteratively using Newton-Raphson. In

large samples, the MLEs (3 γ, 3β) are approximately normal with means (γ , β).
The covariance matrix of the estimates can be obtained by taking the inverse of
the negative Hessian matrix calculated using the ML estimates. The standard errors
are simply the square root of the diagonal elements of the covariance matrix.

3.2 Grouped Frequency Data

Rather than covariates consisting of subject-specific measurements, we could

alternatively face situations where there are only l = 1, . . . , s (s < n) sets of
A Doubly-Inflated Poisson Distribution and Regression Model 139

Table 5 General layout of Response Covariates

grouped data
n01 , n11 , . . . , nm1 x11 . . . x1u
n02 , n12 , . . . , nm2 x21 . . . x2u
.. .. .. ..
. . . .
n0s , n1s , . . . , nms xs1 ... xsu

covariate patterns (e.g., factorial experiments). As shown in Table 5, here our data
consists of the frequencies n0l , n1l , . . . , nml , where nj l represents the frequencies
 j(j = 0, . . . , m), where m is the largest observed frequency and
of count
n = sl=1 m j =1 nj l .
The binomial proportion pl and Poisson mean λl can be parameterized by using
the logit and log link functions similar to that provided in (5) as follows:
p
logit(pl ) = x l γ and log(λl ) = x λl β, (7)

p
where x l is the set of covariates in the lth covariate pattern used to model p, x λl is
the set of covariates in the lth covariate pattern used to model λ, and where γ and
β are again the vectors of corresponding regression coefficients. The log-likelihood
function then becomes


 exp(−λl )λkl
(pl , λl |nj l ) = nj l log pl2 + ql2 exp(−λl ) + nj l log 2pl ql + ql2
k!
{l:j =0} {l:j =k}

m  exp(−λ )λj
l l
+ nj l log ql2 ,
j!
j =1
{l: =k }

which can be reexpressed in terms of the regression parameters γ and β as

m 
p
(γ , β) = −2 nj l log 1 + exp(x l γ )
l:j =0


p
+ nj l log exp(2x l γ ) + exp(− exp(x λl β))
{l:j =0}

 exp(kx λl β − exp(x λl β))
p
+ nj l log 2 exp(x l γ ) +
k!
{l:j =k}

m 
+ nj l j x λl β − exp(x λl β) − log(j !) .
j =1
{l: =k }
140 M. Sheth-Chandra et al.

γ ,3
The ML estimates (3 β) are again found as the solution to the first-order derivatives

∂(γ , β) ∂(γ , β)
= 0, = 0.
∂γ ∂β

Estimated standard errors of the regression parameters can be found by evaluating

the inverse information matrix at the ML estimates.

3.3 Dental Epidemiology Example

We now return to the dental epidemiology data (Bohning et al. 1997, 1999), a subset
of which is presented in Table 6. The aim of the study was to compare six methods of
school-based dental care: 1. oral health education, 2. enrichment of the school diet
with rice bran, 3. mouthwash with 0.2% of NaF solution, 4. oral hygiene, 5. all of
the four treatments, and 6. a standard care control. Gender and race/ethnicity groups
(White, Black, Others including predominantly Hispanic) were also considered.
Inspection across all 1013 child measures shows inflated counts at 0 and 1.
Thus, the DI P (p, λ) regression model is used to assess whether the δ(DMF T )
counts were associated with the treatment and covariates, accounting for possible
inflation at 0 or 1. For simplicity, both p and λ were modeled against the same
set of covariates (treatment, gender, and race/ethnicity). The regression parameter
estimates are provided in Table 7. We see that the combination of all four treatments
lowers the likelihood of inflated counts relative to the control, Black race/ethnicity
leads to a lower likelihood of inflated counts relative to children with Other or
Hispanic race/ethnicities, and there was no effect on the probability of inflation
due to gender. For the traditional Poisson distribution, we see that the education
treatment leads to lower expected counts than the control, and neither gender nor
race/ethnicity had a significant relationship with the expected δ(DMF T ) counts.

Table 6 Dental ID Treatment Gender Ethnicity δ(DMFT)

epidemiology data
1 1 Male White 0
2 2 Female White 1
3 4 Female Black 2
4 5 Female Other 0
.. .. .. .. ..
. . . . .
1012 3 Male Black 3
1013 2 Male Other 0
A Doubly-Inflated Poisson Distribution and Regression Model 141

Table 7 Regression parameter estimates for dental epidemiology data

Logit link for p Log link for λ
Parameter Est. (S.E.) p-value Est. (S.E.) p-value
Constant 1.484 (0.677) 0.029 0.878 (0.190) 0.001
Treatment
Educ 1.161 (0.807) 0.151 −0.451 (0.169) 0.008
Enrich 0.847 (0.772) 0.273 −0.254 (0.180) 0.158
Rinse −0.928 (0.544) 0.088 −0.003 (0.159) 0.983
Hygiene −0.760 (0.572) 0.185 −0.041 (0.200) 0.838
All −1.195 (0.543) 0.028 −0.097 (0.167) 0.561
Gender
Male 0.077 (0.236) 0.774 0.113 (0.085) 0.184
Ethnicity
White −0.181 (0.354) 0.609 0.175 (0.109) 0.111
Black −0.766 (0.362) 0.035 0.060 (0.137) 0.659

4 Conclusion

In this chapter we have introduced a doubly-inflated Poisson (DIP) distribution to

model the probabilities of count data that feature inflation at some value k in addition
to inflation at zero. Both maximum likelihood and moment estimators were derived,
though the latter did not estimate the Poisson mean parameter as efficiently as
the likelihood-based estimator. The DIP distribution was also incorporated into the
generalized linear models framework, and likelihood-based estimation procedures
were developed to estimate the regression coefficients.
A natural extension would be to model the inflation at value k with an additional
parameter than that used to model the inflation at zero (as mentioned earlier).
Though this parameterization would allow better modeling of both inflationary
tendencies, the added complexity could complicate the computations needed for
parameter estimation. The trade-off in this instance between more accurate mod-
eling and computational complexity remains to be seen. Other extensions include
allowing inflation at three or more values, and incorporating a process to determine
whether non-zero inflation exists, as opposed to “knowing” that a certain value
exhibits the inflation. Extensions to correlated counts measures are also possible.

Acknowledgements Dr. Manasi Sheth-Chandra’s research was partially supported by a Modeling

and Simulation graduate student award at Old Dominion University.
142 M. Sheth-Chandra et al.

Appendix: Analytical Derivatives for DIP(p, λ) Distribution

Information Matrix

Let Y be distributed as DI P (p, λ) given in (1). The Fisher information matrix for
this distribution is

I (p) I (p, λ)
I= , (8)
I (p, λ) I (λ)

where the elements of I are defined as

 
∂ 2 log(f (y; p, λ))
I (p) = −E
∂p 2
   2
−2 1 + exp(−λ) p 2 + q 2 exp(−λ) − 2p − 2q exp(−λ)
= 
p 2 + q 2 exp(−λ)
⎡ k  ⎤
2 exp(−λ)λ exp(−λ)λk  exp(−λ)λk 2
⎢ 2pq + q −4+2 − 2 − 4p − 2q ⎥
−⎢ ⎣
k!

k!
k
k! ⎥
⎦
exp(−λ)λ
2pq + q 2
k!
 
exp(−λ)λk
+ 2 1 − exp(−λ) − , (9a)
k!
 
∂ 2 log(f (y; p, λ))
I (p, λ) = −E
∂λ∂p

−2pq exp(−λ) 2q 2 exp(−λ)λk k
=  − 1− , (9b)
p 2 + q 2 exp(−λ) 2pq(k!) + q exp(−λ)λ
2 k λ

and

∂ 2 log(f (y; p, λ))
I (λ) = −E
∂λ2

−p2 q 2 exp(−λ) q2 exp(−λ)λk−1
= + 1−
p2 + q 2 exp(−λ) λ (k − 1)!
 2
q 2 exp(−λ)λk k exp(−λ)λk
+ × −1 q2
2pq(k!) + q 2 exp(−λ)λk λ k!
 
k k 2
2 exp(−λ)λ k
−q −1 − 2 . (9c)
k! λ λ
A Doubly-Inflated Poisson Distribution and Regression Model 143

The covariance matrix of Y and Y 2 is


V ar(Y ) Cov(Y, Y 2 )
=
Cov(Y, Y 2 ) V ar(Y 2 )

where
 2
V ar(Y ) = 2pqk 2 + q 2 (λ2 + λ) − 2pqk + q 2 λ ,
   2
V ar(Y 2 ) = 2pqk 4 + q 2 λ4 + 6λ3 + 7λ2 + λ − 2pqk 2 + q 2 λ + λ2 , and

Cov(Y 2 , Y ) = Cov(Y, Y 2 )
     
= 2pqk 3 + q 2 λ3 + 3λ2 + λ − (2pqk + q 2 λ) 2pqk 2 + q 2 λ + λ2 .

Score Equations for DIP(p, λ) Regression

The partial derivatives for the log-likelihood function for DI P (p, λ) regression of
raw counts are

∂(γ , β)
n p p
x i exp(x i γ )
p p
2x i exp(2x i γ )
= −2 p + p
∂γ
i=1
1 + exp(x i γ ) {i:y =0} exp(2x i γ ) + exp(− exp(x λi β))
i

p
2x i
p
exp(x i γ )
+ ,
{i:yi =k} p exp(kx λi β − exp(x λi β))
2 exp(x i γ ) +
k!
and

∂(γ , β)
−x λi exp(x λi β − exp(x λi β))
= p
∂β
{i:yi =0}
exp(2x i γ ) + exp(− exp(x λi β))


kx λi − x λi exp(x λi β) exp(kx λi β − exp(x λi β))
+ 
{i:yi =k} k! 2 exp(x p γ ) +
exp(kx λi β − exp(x λi β))
i
k!


+ yi x λi − x λi exp(x λi β) .
yi =y
{i:y=0,k }
144 M. Sheth-Chandra et al.

The partial derivatives for the log-likelihood function for DI P (p, λ) regression of
grouped frequencies are

∂(γ , β)
m p p
x exp(x l γ )
p p
2x l exp(2x l γ )
= −2 nj l l p + nj l p
∂γ
l:j =0
1 + exp(x l γ ) {l:j =0} exp(2x l γ ) + exp(− exp(x λl β))

p p
2x l exp(x l γ )
+ nj l ,
{l:j =k} p exp(kx λl β − exp(x λl β))
2 exp(x l γ ) +
k!
and

∂(γ , β)
x λl exp(x λl β − exp(x λl β))
=− nj p
∂β
{l:j =0}
exp(2x l γ ) + exp(− exp(x λl β))
 λ 

kx l − x λl exp(x λl β) exp(kx λl β − exp(x λl β))
+ nj l 
{l:j =k} p exp(kx λl β − exp(x λl β))
k! 2 exp(x l γ ) +
k!

m
 
+ nj l j x λl − x λl exp(x λl β) .
j =1
{l: =k }

References

Bae, S., Famoye, F., Wulu, J.T., Bartolucci, A.A., Singh, K.P.: A rich family of generalized Poisson
regression models with applications. Math. Comput. Simul. 69, 4–11 (2005)
Böhning, D., Dietz, E., Schlattmann, P.: Zero-inflated count models and their applications in public
health and social science. In: Applications of Latent Trait and Latent Class Models in the Social
Sciences, pp. 333–344. Wasemann, Münster (1997)
Böhning, D., Dietz, E., Schlattmann, P., Mendonça, L., Kirchner, U.: The zero-inflated Poisson
model and the decayed, missing, and filled teeth index in dental epidemiology. J. R. Stat. Soc.
162, 195–209 (1999)
Cameron, A.C., Trivedi, P.K.: Regression Analysis of Count Data. Cambridge, New York (1998)
Chaganty, N.R., Shi, G.: A note on the estimation of autocorrelation in repeated measurements.
Commun. Stat. Theory Methods 33, 1157–1170 (2004)
Cohen, A.C.: Estimation in mixtures of discrete distributions. In: Proceedings of the International
Symposium on Discrete Distributions, Montreal, pp. 373–378. Statistical Pub. Society, Calcutta
(1963)
Coxe, S., West, S., Aiken, L.S.: The analysis of count data: a gentle introduction to Poisson
regression and its alternative. J. Pers. Assess. 91, 121–136 (2009)
Hall, D.: Zero-inflated Poisson and binomial regression with random effects: a case study.
Biometrics 56, 1030–1039 (2000)
Hall, D., Shen, J.: Robust estimation for zero-inflated Poisson regression. Scand. J. Stat. 37,
237–252 (2010)
A Doubly-Inflated Poisson Distribution and Regression Model 145

Johnson, N.L., Kotz, S.: Distributions in Statistics: Discrete Distributions. Houghton Mifflin,
Boston (1969)
Lambert, D.: Zero-inflated Poisson regression with an application to defects in manufacturing.
Technometrics 34, 1–14 (1992)
Lin, T.H., Tsai, M.-H.: Modeling health survey data with excessive zero and K responses. Stat.
Med. 32, 1572–1583 (2013)
Marazzi, A., Paccaud, F., Ruffieux, C., Beguin, C.: Fitting the distributions of length of stay by
parametric models. Med. Care 36, 915–927 (1998)
McCulloch, C.E., Searle, S.R.: Generalized, Linear, and Mixed Models. Wiley, New York (2001)
Sheth-Chandra, M.: The doubly inflated Poisson and related regression models. PhD Dissertation.
Old Dominion University, Norfolk (2011)
Multivariate Doubly-Inflated Negative
Binomial Distribution Using Gaussian
Copula

Joseph Mathews, Sumen Sen, and Ishapathik Das

1 Introduction

Multivariate count data appear in many fields of study such as econometrics,

ecology, criminology, and epidemiology. One popular alternative to the Poisson
distribution for modeling such multivariate count data is the negative binomial
distribution. Unlike the Poisson distribution, the negative binomial distribution
has a separate mean and variance parameter making it an attractive choice for
overdispersed or underdispersed data. In the case of multivariate, zero-inflated data,
both the zero-inflated Poisson (ZIP) distribution model and zero-inflated negative
binomial (ZINB) distribution model are used. In certain cases, the multivariate
count data contain an additional inflation point. The ZIP and ZINB models are
then extended to a doubly or multiple inflation model. Forms of such doubly or
multiple inflation models using the multivariate Poisson distribution are given by
Karlis and Ismail (2005), Sen et al. (2017), and Sengupta et al. (2015). Bivariate
zero-inflated negative binomial models exist in the literature such as So et al. (2011)
and Ismail and Faroughi (2017). However, little is found on a doubly or multiple
inflation model using the multivariate negative binomial distribution.

J. Mathews · S. Sen ()

Department of Mathematics and Statistics, Austin Peay State University, Clarksville, TN, USA
e-mail: jmathews6@my.apsu.edu; sens@apsu.edu
I. Das
Department of Mathematics, Indian Institute of Technology Tirupati, Tirupati, India
e-mail: ishapathik@iittp.ac.in

© Springer Nature Switzerland AG 2019 147

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_8
148 J. Mathews et al.

Expressions of the multivariate negative binomial distribution exist in the

literature such as Doss (1979) and Antzoulakos and Philippou (1991). However,
such forms are given by the probability generating function and are very com-
plicated in practice. Therefore, copula methods for modeling multivariate count
data are a common alternative (Karlis and Nikoloulopoulos 2009; Brechmann
et al. 2011). A copula is a probability function commonly used to form the
joint probability distribution of X1 , X2 , . . . , Xp random variables with marginal
distributions F1 , F2 , . . . , Fp , respectively. Moreover, a copula preserves the depen-
dence structure between the p random variables in a tractable form. Here, we use
copula methods to obtain a multivariate form of a doubly-inflated negative binomial
distribution.
This paper is organized as follows. Section 2 will cover the definition of a copula
and the fundamental Sklar’s theorem (Sklar 1959). Section 3 will cover multivariate
discrete distributions using copula methods and Sklar’s theorem. In Sect. 4 we
construct and provide properties of the doubly-inflated negative binomial model
and end with a description of a data simulation algorithm. In Sect. 5 we review
maximum likelihood estimation, provide an expression for the likelihood function
of the proposed model, and outline a method for estimating the model’s parameters.
Also, we give parameter and mean squared error (MSE) results for simulated data.
In Sect. 6 we apply the model to the DoctorAUS dataset from the Ecdat package
in R (Bolker 2016) and provide parameter estimates along with their respective
asymptotic standard error. We end with a conclusion in Sect. 7.

2 The Multivariate Copula and Sklar’s Theorem

A copula is a multivariate distribution function with uniform margins on the

interval [0, 1]. Given specified marginal distributions, one can obtain a multivariate
probability distribution through a coupla, see Joe (2014) and Song (2007).
Definition 1 A p-dimensional copula is a function C : [0, 1]p → [0, 1] with the
following properties:
1. C(1, . . . , ai , . . . , 1) = ai , ∀i = 1, 2, . . . , p and ai ∈ [0, 1]
2. C(a1 , a2 , . . . , ap ) = 0 if at least one ai = 0 for i = 1, 2, . . . , p.
3. For any ai,1 , ai,2 ∈ [0, 1] with ai,1 ≤ ai,2 , for i = 1, 2, . . . , p:

2
... (−1)j1 +j2 +...+jp C(a1,j1 , a2,j2 , . . . , an,jp ) ≥ 0.
j1 =1 j2 =1 jp =1

Using Sklar’s theorem, one can form a multivariate probability distribution

function using known univariate marginal distributions and a unique copula.
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 149

Theorem 1 (Sklar’s Theorem) Let X1 , X2 , . . . , Xp be random variables with

marginal distribution functions F1 , F2 , . . . , Fp and joint cumulative distribution
function F , then the following holds:
1. There exists a p-dimensional copula C such that for all x1 , x2 , . . . , xp ∈ R

F (x1 , x2 , . . . , xp ) = C(F1 (x1 ), F2 (x2 ), . . . , Fp (xp ))

2. If X1 , X2 , . . . , Xp are continuous, then the copula C is unique. Otherwise,

C can be uniquely determined on a p-dimensional rectangle with dimension
Range(F1 ) × Range(F2 ) × . . . × Range(Fp ).
See Joe (2014) and Nelsen (2006) for more information of copula’s and
associated properties.
One very common copula is the Gaussian copula due to its ability to preserve the
correlation structure of two variables.
Definition 2 The Gaussian copula is given by the function:

C(u1 , u2 , . . . , up |R(r)) = R (−1 (u1 ), −1 (u2 ), . . . , −1 (up )), (1)

where −1 is the inverse CDF of a standard normal and R is the joint cumulative
distribution function of a standard multivariate normal distribution with covariance
matrix equal to the correlation matrix R.
Definition 3 The Gaussian copula density is defined as:

1 1
c(u1 , u2 , . . . , up |R(r)) = √ exp(− U T × (R(r)−1 − Ip ) × U ), (2)
|R(r)| 2

where U = (−1 (u1 ), −1 (u2 ), . . . , −1 (up ))T

A plot of a bivariate Gaussian copula with r = 0.85 is given   in Fig. 1. If the
dimension of the random variable X is p, then there are p2 elements in the
association matrix R. As p increases, the number of parameters increases. If a
model contains too many parameters to estimate, then computational challenges
may be experienced. Therefore, we assume a structured correlation matrix R of two
kinds.
1. Equi-correlation structure: Under this structure, we assume

R(r) = r11t − (1 − r)Ip

150 J. Mathews et al.

12

10

8
Density

0
1
0.8
0.6
0.4
1
0.2 0.8
0.6
0.4
u2 0 0.2
0
u1

Fig. 1 A bivariate Gaussian copula density with r = 0.85

where Ip is a p-dimensional identity matrix, r ∈ (− p−1 1

, 1), and 1 is a p-
dimensional column vector of ones. It follows from Olkin and Pratt (1958) that:
1 r
R −1 (r) = Ip − 11t (3)
1−r (1 − r){1 + (p − 1)r}

2. AR-1 structure: Under this structure, the (i, j )th element of R(r) is given by
r |i−j | , with r ∈ (−1, 1). The inverse of this matrix is given below (Chaganty
1997)

1
R −1 (r) = (Ip − r 2 M2 − rM1 ), (4)
1 − r2

where M2 = diag(0, 1, . . . , 1, 0) and M1 is a tridiagonal matrix with 0 on the

main diagonal and 1 on the upper and lower diagonals.
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 151

3 Discrete Multivariate Density Function Using the Gaussian

Copula

A copula may also be used to derive a joint distribution for discrete data. Given a
set of discrete, marginal distributions F1 (y1 |θ1 ), F2 (y2 |θ2 ), . . . , Fp (yp |θp ), one can
obtain the following joint probability mass function of Y = (Y1 , . . . , Yp ):

fC (Y|, R(r)) = P (Y1 = y1 , Y2 = y2 , . . . , Yp = yp |, R(r))

2
= ... (−1)j1 +j2 +...+jp C (u1,j1 , . . . , up,jp |R(r))
j1 =1 j2 =1 jp =1

(5)

where  = (θ1 , . . . , θp ), C (·|R(r)) is the Gaussian copula distribution function

with association matrix R(r). Note that uj 1 = Fj (yj ) and uj 2 = Fj (yj −), where
Fj (yj −) is the left-hand limit of Fj at yj equal to Fj (yj − 1).
Let Y1 , . . . , Yn be n random variables where the ith random variable Yi has
a negative binomial distribution with probability and size parameters pi and si ,
respectively. The distribution function be Fsi ,pi (yi ) and probability mass function
given by:

yi + si − 1 y
fN B (yi |si , pi ) = P (Yi = yi ) = (1 − pi )si pi i , (6)
yi

where si > 0, yi ∈ N and pi , ∈ [0, 1] for i = 1, . . . n. Using Sklar’s theorem and

Eq. (5) one can derive the joint probability mass function fC (Y1 = y1 , . . . Yn =
yn ). For bivariate case plots of the joint probability mass function are given in Fig. 2
with different correlation values.

4 Multivariate Doubly-Inflated Negative Binomial Model

Modeling count data are very popular in statistics. Both the Poisson and negative
binomial distributions are very popular to model count data. One simple approach
of introducing correlation among count variables is through common additive error
models. Kocherlakota and Kocherlakota (2001) and Johnson et al. (1997) provided
a detailed discussion for the one-factor multivariate Poisson model. Along this line
of study, Winkelmann (2000) proposed a multivariate negative binomial regression
model. In the case of zero and doubly-inflated count data, zero-inflated and doubly-
inflated multivariate Poisson models are available (Sen et al. 2017; Agarwal et al.
2002; Lee et al. 2009). In this section, we shortly review the model construction
process. For details, refer to Sen et al. (2017).
152 J. Mathews et al.

0.02

0.005 Z 0.01
Z

8
8 6
0 6 4 X
8 4 0
6 X 8 2
4 2 6
y 4
2 2 0
0 0 V
(a) (b)

Fig. 2 Bivariate negative binomial density using a Gaussian copula with s1 = 10, p1 = 0.50 and
s2 = 15, p2 = 0.60. (a) r = 0.10. (b) r = 0.90

4.1 Construction of Doubly-Inflated Model

Consider a latent random variable Z with the following probability distribution:

⎧
⎪
⎪ pr1 if z = 2
⎪
⎪
⎨p if z = 1
r2
P (Z = z) = (7)
⎪
⎪ 1−q if z = 0
⎪
⎪
⎩
0 elsewhere

where pr1 , pr2 ∈ (0, 1) and q = pr1 + pr2 < 1. Also, let Y = (Y1 , . . . , Yp ) be a
multivariate negative binomial random variable with mass function fC constructed
using a Gaussian copula as mentioned in Eq. (5).
The distribution functions of Y|Z and (Y, Z) are given below:
⎧
⎪
⎪ if z = 2, y = (0, . . . , 0)
⎨1
f1 (Y|Z) = 1 if z = 1, y = (k1 , . . . , kp ) (8)
⎪
⎪
⎩f if z = 0, y = (y1 , . . . , yp )
C (y1 , . . . , yp )

⎧
⎪
⎪ if z = 2, y = (0, . . . , 0)
⎨pr1
f2 (Y, Z) = pr2 if z = 1, y = (k1 , . . . , kp ) (9)
⎪
⎪
⎩(1 − q)f if z = 0, y = (y1 , . . . , yp )
C (y1 , . . . , yp )
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 153

Thus, the probability mass function of the multivariate doubly-inflated negative

binomial distribution is given by :
⎧
⎪
⎨pr1 + (1 − q)fC (0, . . . , 0|R(r))
⎪ if y = (0, . . . , 0)
fR(r) (Y|) = pr2 + (1 − q)fC (k1 , . . . , kp |R(r)) if y = (k1 , . . . , kp ) (10)
⎪
⎪
⎩(1 − q)f
C (y1 , . . . , yp |R(r)) if y = (y1 , . . . , yp ),

where  = (pr1 , pr2 , p1 , . . . , pp , s1 , . . . sp ). The parameters pr1 and pr2 corre-

spond to the points of inflation in the cells (0, . . . , 0) and (k1 , . . . , kp ), respectively.

4.2 Bivariate Doubly-Inflated Negative Binomial

In the previous section, a multivariate doubly-inflated model is developed using a

Gaussian copula. In order to simplify things we consider a bivariate
  doubly-inflated
negative binomial model with correlation structure R(r) = 1r 1r . For bivariate case
consider the bivariate random vector Y = (Y1 , Y2 ). Using Eq. (10), the joint mass
function of Y is given below.
First consider the case when Y = (0, 0), then Eq. (5) reduces to:

fC (Y|, R(r)) = P (Y1 = 0, Y2 = 0)

= C (Fs1 ,p1 (0), Fs2 ,p2 (0)) − C (Fs1 ,p1 (−1), Fs2 ,p2 (0))
− C (Fs1 ,p1 (0), Fs2 ,p2 (−1)) + C (Fs1 ,p1 (−1), Fs2 ,p2 (−1))
= C (Fs1 ,p1 (0), Fs2 ,p2 (0))
= C ((1 − p1 )s1 , (1 − p2 )s2 ).
(11)

This follows from the fact that Fs1 ,p1 (−1) = Fs2 ,p2 (−1) = 0 and the definition of
a copula. Using Eq. (10), the mass function of the bivariate doubly-inflated negative
binomial is given by:
⎧
⎪
⎨pr1 + (1 − q)C ((1 − p1 ) , (1 − p2 ) ),
⎪ y = (0, 0)
s1 s2

fR(r) (Y|) = pr2 + (1 − q)fC (k1 , k2 |R(r)), y = (k1 , k2 ) (12)

⎪
⎪
⎩(1 − q)f (y , y |R(r)), otherwise.
C 1 2

A plot of Eq. (12) is given in Fig. 3.

154 J. Mathews et al.

Fig. 3 Bivariate 0.4

doubly-inflated negative
binomial model with
s1 = 10, s2 = 15, p1 =
0.50, p2 = 0.60 and inflation
points at (0, 0) and (7, 2)

Z 0.2

0 0
0 2
2
4
4 X
6
y 6

4.3 Properties of Multivariate Doubly-Inflated Negative

Binomial

In this section we discuss important properties of the proposed distribution in

Eq. (10). Expression for marginal mass functions and nth order moments are
provided.

Marginal Distributions

From the multivariate distribution mentioned in Eq. (10), marginal distribution of Yi

is given by:
⎧
⎪
⎨pr1 + (1 − q)(1 − pi ) , yi = 0
⎪ si

mi (yi ) = pr2 + (1 − q)fN B (ki ), yi = ki (13)

⎪
⎪
⎩(1 − q)f (y ), yi = N − {{0}, {ki }},
NB i

where fN B) (.) is the mass function of a negative binomial distribution, with

parameters p1 and s1 , defined in Eq. (6). Plot of the mass function is provided in
Fig. 4.
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 155

0.4

0.4
0.3

0.3
Density

Density
0.2

0.2
0.1
0.1

0.0
0.0

0 5 10 15 20 0 5 10 15 20
Y1 Y2

Fig. 4 Mass functions of Y1 and Y2 where k1 = 3 and k2 = 5

Using Eqs. (10) and (13), the expected value of the distribution can be computed
as below:
∞

∞

∞

E(Yi ) = ··· ··· yi fR(r) (y1 , . . . , yp |)

y1 =0 yi =0 yp =0
∞

∞

∞

∞

∞

= yi ··· ··· fR(r) (y1 , . . . , yp |)

yi =0 y1 =0 yi−1 =0 yi+1 =0 yp =0
∞

= yi mi (yi )
yi =0
∞

= 0mi (0) + ki mi (ki ) + yi m(yi )

x=1
x=0,ki
∞

= ki {pr2 + (1 − q)fN B (ki )} + yi {(1 − q)fN B (yi )}

x=1
x=0,ki
∞

= ki pr2 + (1 − q) yi fN B (yi )
yi =0
pi s i
= ki pr2 + (1 − q) (14)
1 − pi

In a similar fashion, we can find an expression for E(Yin ):

E(Yin ) = kin pr2 + (1 − q)E(Nin ), (15)

156 J. Mathews et al.

where Ni is the negative binomial random variable with parameters pi and si . Using
Eqs. (14) and (15), we can find an expression for the variance:

pi si (1 + pi si )
V (Yi ) = ki2 pr2 + (1 − q)
(1 − pi )2
(16)
pi s i 2
− [ki pr2 + (1 − q)( )]
1 − pi

4.4 Data Simulation Algorithm

A description of a data simulation algorithm for simulating data from a multivariate

distributions using Gaussian copula can be found in Joe (2014) and Nelsen (2006).
We are interested in simulating data from the multivariate doubly-inflated negative
binomial model given in Eq. (10). The steps for the algorithm are as follows:
1. Simulate Z from the categorical distribution given in Eq. (8)
2. If Z = 2, then y = (0, . . . , 0).
3. If Z = 1, then y = (k, . . . , k).
4. If Z = 0, then use the algorithm described in Joe (2014) to simulate n
observations from the multivariate mass function defined in Eq. (5).
In particular, we used this algorithm to simulate from bivariate doubly-inflated neg-
ative binomial distribution. Below are histograms of the marginals of a simulation
with n = 1000 samples (Fig. 5).

5 Parameter Estimation Using Maximum Likelihood

Estimation

Now we will review and use maximum likelihood techniques to estimate the
parameters of our model.

5.1 The Likelihood Function

The doubly-inflated negative

  binomial probability mass function defined in (14) is
dependent on 2 + 2p + p2 parameters. Notice that the number of parameters to
estimate increases quickly as the number of dimensions increases—posing potential
computational problems for high-dimensional data. Thus, we use a structured
correlation matrix as defined in Eqs. (3) and (4) to avoid these computational
problems. We will use the method of maximum likelihood to estimate the 2 +
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 157

0.10

0.10
0.08

0.08
0.06

0.06
Frequency

Frequency
0.04

0.04
0.02

0.02
0.00

0.00
0 10 20 30 40 50 60 0 10 20 30 40 50 60

y1 y2

Fig. 5 Histogram of marginal distributions for a bivariate doubly-inflated negative binomial model
with s1 = 10, s2 = 15, p1 = 0.30, p2 = 0.40 and inflation point at k = 10

 
2p + p2 parameters. Hence, by (14) the likelihood function for the doubly-inflated
negative binomial using a Gaussian copula is given by:

LR(r) (|Y ) = pr 1 + (1 − q)C ((1 − p1 )s1 , . . . , (1 − pp )sp |R(r))
i:Yi =(0,...,0)

× pr 2 + (1 − q)fC (k, . . . , k|R(r))
i:Yi =(k,...,k)

× (1 − q)fC (y1i , . . . , yni |R(r)) (17)
i:Yi =(y1i ,...,yni )

The maximum likelihood parameter estimates are given by the derivative of the
log of the likelihood equation in (17). Applying the logarithm function to both sides
of (17), we obtain the following:

lR(r) (|Y ) = pr 1 + (1 − q)C ((1 − p1 )s1 , . . . , (1 − pp )sp |R(r))

i:Yi =(0,...,0)

+ pr 2 + (1 − q)fC (k, . . . , k|R(r))

i:Yi =(k,...,k)

+ (1 − q)fC (y1i , . . . , yni |R(r)) (18)

i:Yi =(y1i ,...,yni )
158 J. Mathews et al.

A complicated expression such as (18) will have no closed form expressions

for estimators by solving the score equations. Therefore, we use quasi Newton
Raphson (Davidon 1991) to obtain estimates. Also, as estimating parameters from
a multivariate Gaussian copula can be computationally challenging, convergence
issues arise through direct maximization of the likelihood function. Thus, we will
split the estimation process into several steps. The inference function for margins
(IFM) method proposed by Joe (2005) is a popular method for estimating copula
parameters.
Notice that the algorithm below maximizes the likelihood function in two
steps:
1. Start with initial values pr 1 0 and pr 2 0 . These are found by finding the proportion
of inflation points (0, . . . , 0) and (k1 , . . . , kp ) in the data, respectively.
2. Obtain initial estimates for s i = (s1i , . . . , spi ) and p i = (p1i , . . . , ppi ) by applying
the method of maximum likelihood to the marginal distributions of the data.
3. At the ith step, i = 1, 2, . . . , use the initial values from step 2 and maximize the
log-likelihood function lR(r) (pr 1 0 , pr 2 0 , s i , p i , r) with respect to r and obtain r̂.
4. Using r̂ from step 3, maximize the log-likelihood function lR(r) (pr 1 0 , pr 2 0 , s i ,
pi , r̂) with respect to pr 1 , pr 2 , s, and p to obtain pˆr 1 , pˆr 2 , ŝ, p̂.
5. Index i = i + 1 and use the estimates pˆr 1 , pˆr 2 , ŝ, and p̂ as initial values. Repeat
steps 2 and 3 to get updated values.
6. Repeat steps 2, 3, and 4 until convergence.

5.2 Simulation Results

Now the proposed algorithm is used to estimate the doubly-inflated negative

binomial model parameters. First, the data simulation algorithm from Sect. 4.2 is
used to simulate data from a bivariate, doubly-inflated negative binomial model. The
estimation algorithm from Sect. 5.1 is then used to estimate the seven parameters.
The results for sample sizes 200 and 500 are given in Table 1.

Table 1 Bivariate, Simulation results

doubly-inflated negative
Sample size = 200 Sample size = 500
binomial distribution with
points of inflation (0, 0) and Parameters Estimates MSE Estimates MSE
(3, 2) p1 = 0.4 0.4129 0.2496 0.4167 0.1217
p2 = 0.2 0.2126 0.0658 0.2140 0.3217
s1 = 5 5.6621 4.3250 5.4683 3.2377
s2 = 3 3.2342 1.0639 3.2339 0.7849
pr 1 = 0.30 0.2967 0.0350 0.3022 0.0200
pr 2 = 0.40 0.3972 0.0647 0.4170 0.0040
ρ = 0.6 0.6396 0.2054 0.5591 0.0760
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 159

6 Application

We now fit a bivariate doubly-inflated negative binomial distribution to the Doc-

torAUS dataset. We compare three models: an independent model, a zero-inflated
negative binomial (ZINB) model, and a bivariate doubly-inflated negative binomial
(BDINB) model. For the independent model, we assume that both variables are
independent and a negative binomial distribution is fit to both variables. For the
ZINB model, we assume that both variables are independent and a zero-inflated
negative binomial distribution is fit to both variables using the dzinbinom function
from the emdbook package Bolker (2016) in R. For the BDINB model, we fit a
doubly-inflated negative binomial distribution and impose a dependence structure
on the two variables using Gaussian copula methods. We use Akaike’s information
criterion (AIC) and the Bayesian information criterion (BIC) to compare the models.

6.1 Application to DoctorAUS Dataset

The DoctorAUS dataset can be found in the Ectdat package in R. The data comes
from a study done at an Australian hospital from 1977 to 1978 with n = 5190
observations. We consider the actdays = (0, . . . , 14) and illness = (0, . . . , 5)
variables. Here, actdays is the number of days of reduced activity from illness or
injury in the previous 2 weeks for a given patient and illness is the number of days
a given patient was sick in the previous 2 weeks. A table of counts for the two
variables is given in Table 2. Hence, there appears to be inflation points at (0, 0) and
(1, 0). In order to compute the asymptotic standard errors, we estimate pr 1 and pr 2
by finding the proportion of (0, 0) and (1, 0) counts rather than maximum likelihood
estimation. The results for the parameter estimates are given in Table 3. The results
for the three models are given in Table 4.

Table 2 Counts of actdays × illness variables

Count table
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
0 1543 5 0 0 1 0 0 0 0 0 0 0 0 0 5
1 1375 71 40 26 20 10 6 8 5 2 5 1 2 0 67
2 763 53 28 15 7 18 3 11 5 2 2 1 3 0 35
3 420 25 19 14 8 5 5 10 4 1 1 0 0 2 28
4 200 11 10 8 6 6 1 5 2 1 1 0 0 2 21
5 153 12 11 11 3 1 2 4 1 1 3 0 1 1 32
160 J. Mathews et al.

Table 3 Estimation results Estimation results

for DoctorAUS dataset
Parameters Estimates Standard error
s1 13.37164495 1.442319086
s2 0.11853352 0.005808301
p1 0.85182789 0.013635534
p2 0.06818760 0.004735758
r 0.03349333 0.031864960

Table 4 Model results for Model results

DoctorAUS dataset
Model AIC BIC
Independent model 24,819.30 24,845.52
ZINB model 27,423.81 27,463.13
BDINB model 245,85.85 25,631.74

7 Conclusion

We have proposed a doubly-inflated model using the multivariate negative binomial

distribution. Copula methods were used to preserve the dependence structure of
the model in a tractable form. Maximum likelihood estimation was introduced
and a method for estimating the proposed model’s parameters was given. The
doubly-inflated model obtained the best fit compared to the independent and zero-
inflated model when applied to real data. For high-dimensional data, computational
problems may still occur. Further research includes extending the model to a
regression setting by including covariates.

References

Agarwal, D.K., Gelfand, A.E., Citron-Pousty, S.: Zero-inflated models with application to spatial
count data. Environ. Ecol. Stat. 9(4), 341–355 (2002)
Antzoulakos, D.L., Philippou, A.N.: A note on the multivariate negative binomial distributions of
order k. Commun. Stat. Theory Methods 20, 1389–1399 (1991)
Bolker, B.: emdbook: Ecological Models and Data in R; R package version 1.3.9 (2016)
Brechmann, E.C., Czado, C., Kastenmeier, R., Min, A.: A mixed copula model for insurance claims
and claim sizes. Scand. Actuar. J. 2012(4), 278–305 (2011)
Chaganty, N. Rao.: An alternative approach to the analysis of longitudinal data via generalized
estimating equations. J. Stat. Planning Inference 63, 39–54 (1997)
Davidon, W.C.: Variable metric method for minimization. SIAM J. Optim. 1(1), 1–17 (1991)
Doss, D.C.: Definition and characterization of multivariate negative binomial distribution. J.
Multivar. Anal. 9(3), 460–464 (1979)
Ismail, N., Faroughi, P.: Bivariate zero-inflated negative binomial regression model with applica-
tions. J. Stat. Comput. Simul. 87(3), 457–477 (2017)
Joe, H.: Asymptotic efficiency of the two-stage estimation method for copula-based models. J.
Multivar. Anal. 94, 401–419 (2005)
Multivariate Doubly-Inflated Negative Binomial Distribution Using Gaussian Copula 161

Joe, H.: Dependence Modeling with Copulas. Chapman & Hall/CRC Monographs on Statistics &
Applied Probability. Chapman and Hall/CRC, Boca Raton (2014)
Johnson, N.L., Kotz, S., Balakrishnan, N.: Discrete Multivariate Distributions. Wiley, New York
(1997)
Karlis, D., Ismail, N.: Bivariate Poisson and diagonal inflated bivariate Poisson regression models
in R. J. Stat. Softw. 14(10) (2005)
Karlis, D., Nikoloulopoulos, A.: Modeling multivariate count data using copulas. Commun. Stat.
Simul. Comput. 39(1), 172–187 (2009)
Kocherlakota, S., Kocherlakota, K.: Regression in the bivariate Poisson distribution. Commun.
Stat. 30(5), 815–825 (2001)
Lee, J., Jung, B.C., Jin, S.H.: Tests for zero inflation in a bivariate zero-inflated Poisson model.
Statistica Neerlandica 63, 400–417 (2009)
Nelsen, R.B.: An Introduction to Copulas. Springer Series in Statistics. Springer, Berlin (2006)
Olkin, I., Pratt, J.W.: Unbiased estimation of certain correlation coefficients. Ann. Math. Stat.
29(1), 201–211 (1958)
Sen, S., Sengupta, P., Diawara, N.: Doubly inflated Poisson model using Gaussian copula.
Commun. Stat. Theory Methods 10(0), 1–11 (2017)
Sengupta, Pooja, Chaganty, N.R., Sabo, Roy T.: Bivariate doubly inflated Poisson models with
applications. J. Stat. Theory Pract. 10(1), 202–215 (2015)
Sklar, A.: Fonctions de répartition à n dimensionset leurs marges. Publ. Inst. Statis. Univ. Paris 8,
229–231 (1959)
So, S., Lee, D., Jung, B.C.: An alternative bivariate zero-inflated negative binomial regression
model using a copula. Econ. Lett. 113(2), 183–185 (2011)
Song, P.X.K.: Correlated Data Analysis: Modeling, Analytics, and Applications, 1st edn., vol. 365.
Springer Series in Statistics. Springer, New York (2007)
Winkelmann, R.: Seemingly unrelated negative binomial regression. Oxford Bull. Econ. Stat.
62(4), 553–560 (2000)
Quantifying Spatio-Temporal
Characteristics via Moran’s Statistics

Jennifer L. Matthews, Norou Diawara, and Lance A. Waller

1 Introduction

Moran’s index is a measure of spatial autocorrelation within a domain area; it has

and continues to be applied in many fields. As described by Moran (1950), when
given a set of spatial variates (defined on a two-dimensional discrete area), we
may want to investigate whether there is any evidence that spatial autocorrelation is
present overall or in neighboring clusters based on selected features. For example,
in epidemiology, we may be interested in disease mapping based on the number
of people infected in a given area. Lawson (2009) and Zhou and Lawson (2008)
show evidence of clustered disease maps and develop spatio-temporal disease
surveillance under a Bayesian modeling. Meddens and Hicke (2014) describe
spatial and temporal patterns of tree mortality from beetle ecology and dynamics
in parts of Colorado and Wyoming. The spatial point process experiment in Vaillant
et al. (2011) describes sustainability or occurrence dates of sugarcane viruses. In
environmental setting, growth of species can be modeled individually, but it may be
much more appropriate to include their spatio-temporal dynamics in the ecosystem
(Jones-Todd et al. 2018). In geoscience, patterns have always been of interest. Wang
et al. (2015) proposed geographically weighted statistics as an alternative to the
assumption of homogeneous spatial characteristics, but not adding a time factor.

J. L. Matthews · N. Diawara ()

Department of Mathematics and Statistics, Old Dominion University, Norfolk, VA, USA
e-mail: ndiawara@odu.edu
L. A. Waller
Department of Biostatistics, Rollins School of Public Health, Emory University,
Atlanta, GA, USA

© Springer Nature Switzerland AG 2019 163

N. Diawara (ed.), Modern Statistical Methods for Spatial and Multivariate Data,
STEAM-H: Science, Technology, Engineering, Agriculture, Mathematics & Health,
https://doi.org/10.1007/978-3-030-11431-2_9
164 J. L. Matthews et al.

Indeed, modeling disease spatio-temporal spread using ordinary least squares

or weighted least squares will lead to much bias in the estimates and violation
of model assumptions (Pace and LeSage 2009). A simple global Moran’s value
cannot explain area differences, temporal characteristics, or the spatio-temporal
relationship of some natural ubiquitous phenomena. Data may show local trends
that are not shared globally. This has led many authors such as Anselin (1995) to
partition the global spatial domain and compute Moran’s index for each subdomain,
leading to local indices of spatial association (LISA). However, the impact of time
was not incorporated. Spatio-temporal autocorrelation was first introduced by Cliff
and Ord (1975), and the concept has been explored by many others over the years
(see, for example, Martin and Oeppen 1975; Wang and He 2007; Lee and Li
2017). By adding a time component to the Moran’s statistics, the spatio-temporal
characteristics lead to better understanding of the spread of the phenomena or the
spatial variations.
We propose to evaluate spatio-temporal model association, capturing Moran’s
statistical spatial autocorrelation values over iterative time under a Poisson process.
The model we propose is capable of handling spatial variation and temporal
stochastic/dynamic association. Implementation is provided under a multilevel
Poisson process that is time dependent, and the dependence among the levels is
captured spatially with a rate of increase of the phenomena over time. The rate
of increase is built under incrementing time scales for each subarea. The scale
parameter effectively creates a region of dynamic model phenomena occurrence.
Using Markov chain Monte Carlo techniques, simulated data are generated to
investigate Moran’s statistics. Diagnostics and comparison of the Moran’s statistics
are performed. Clusters are built and local features are presented to enhance the
understanding of Moran’s statistics, trend in time, and functionality of model
analysis.

2 Spatio-Temporal Process

For a measurable space S = Rd , a domain D ⊂ S, and a collection of random

points x1 , x2 , . . . , let N(D) represent the count or number of points xi that capture
location occurrence of some disease/event in the domain D. Then N (D) can be seen
as a measure and is called a point process. For nonoverlapping areas D1 , D2 , . . . ,
N(D1 ), N (D2 ), . . . are mutually independent and for each area D, N (D) follows
a Poisson distribution with mean λD , where λD is the intensity of the process. As a
process, a counting may be associated with it and is defined as:
∞


1 if xi ∈ D,
N (D) = 1D (xi ), where 1D (xi ) = (1)
0 otherwise.
i=1
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 165

To every domain, we consider an associated intensity measure λD as λD =

E(N(D)), the expected number of points that are obtained in D. Such measure
can be assumed to be finite if N(D) < ∞, hence the name finite measure.
We consider the temporal conditional distribution of the occurrence of disease in
the partitioned subareas defined by Voronoi cells induced by a set of points (sites)
xit as described in Kallenberg (2001). These cells are defined as:

D(xit ) = {x ∈ D : μ(xit , x) < μ(xit , z), ∀z ∈ D, μ ∈ N (Rd )},

where N (Rd ) is the class of locally finite measures on R+ and D is the domain
area. Hence D − i t denotes the area around the point xit . Doing so, we have a local
precise measurement of the density and neighborhoods. We then build a partition of
D into subareas called time dependent Voronoi collection of D, Dit where a finite
measure or “locally finite” measure is such that μDit < ∞ for all i ≥ 1.
Later, we consider as measure the nonhomogeneous Poisson point process with
non-constant intensity function λD in time say, λD (t) the expected number of points
in D in interval time of length t, with t ∈ [0, T ] and the main assumption that
points are independent of each other. The nonhomogeneous Poisson process has the
following three important properties as described in Baddeley et al. (2016):
• conditional property: given exactly n points in a region Dkt , these points are
mutually independent and each point has the same probability distribution over
D, with probability density Yk(t) = YD t ∼ fDk (t) = λDk (t)/μD , where μD =
! k

λD (t)dt.
D
• superposition property: If Y1 , Y2 , . . . , Yn are  independent Poisson  random
variables with means E(Yi ) = λi , λi ∈ R+ , then i Yi ∼ P oisson( λi ).
• random thinning property: Suppose that N ∼ P oisson(λ), and that
Y1 , Y2 , . . . , Yn are independent, identically distributed multinomial random
variables with distribution Multinomial (p1 , p2 , . . . , pn ), that is P {Yi = k} =
pk for k = 1, 2, . . . , m. Then the random variables N1 , N2 , . . . , Nm defined by

k
Nk = 1{Yi = k} are independent Poisson random variables with parameters
i=1
E(Nk ) = (λpk ).
Hence λD t = λ(·, Dkt ) is a random variable in [0, +∞] for every Dkt ⊆ D. Such
k
an idea is also described in Thäle and Yukich (2016) with properties of Poisson-
Voronoi generated sets.
The Voronoi cells are defined as a point process Dkt ⊆ D, each Dkt containing a
finite number of simple points and they can be viewed as a countable set of random
points with associated intensity λD = E(N(Di )) the expected number of points
in Di .
166 J. L. Matthews et al.

The cells Dkt define a sequence of original/congruent points and for partition
{Dkt } of D, each Dkt can be associated with the original point xkt . If D(t) denotes the
7
n(t)
collection of Voronoi cells generated at time t, D(t) = t
Dk(t) , the elements of
k(t)=1
D(t) are nested within D(t − 1), i.e., D(t) is a refinement of D(t − 1). They include
convex sets (Kieu et al. 2013), differentiable manifolds with smooth boundary of
finite dimensional Hausdorff-type measure with topology on R.
We formalize the notion of conditional distribution of the count given that a point
has occurred under Palm probability/measure and define it as Px t (Dkt ) and because
k(t)
of time dependence we consider the reduced Palm distribution P! x t (Dk(t) t ) denoted
k(t)
as Qx t ,which is the conditional distribution of the count when the point at x is
k(t)
omitted.
The cells are such that their asymptotic mean and variance are a function of the
weighted surface (Reitzner et al. 2012).
For a random exchangeable sequence of spaces (D1t , . . . , Dnt t ) defined at time t ∈
[ti−1 , ti ), let Qtk(t) denote the empirical distribution of the sets generated from points
t , 1 < k(t) < n(t). Let (Y
xk(t) k(t) ) be the sequence count of outcomes observed in
t
(Dk(t) ), then

 ! 
P Y1 , . . . , Ynt = Qx t , or
i

 !
P Yk(t) , 1 ≤ k(t) ≤ n(t) = Qx(t−1) (Dk(t) )dF (Q),

where Q is the limiting distribution of FN over D.

Observing that the sequence is time dependent, we introduce the model as a
discrete time Markov chain as in Resnick (2002):
 
P Yk(t) |Yk(1) , . . . , Yk(t−1) = P Yk(t) |Yk(t−1)

and define the process as follows:

 !
P Y1t , . . . , Ynt |Y t−1 = Qx1 (D1t ) · · · Qxn (t) (Dnt )dF (Q), or

  ! 
P Yk(t) |Yk(t−1) = Qxk(t) (Dkt (t))dF (Q).
1≤k(t)≤n(t)

For a time period of length T partitioned into m time subintervals t0 = 0 < t1 <
t2 < · · · tr < · · · < tm , consider [tr−1 , tr ), r = 0, 1, . . . , m, 0 ≤ tr ≤ T . The
number of random occurrences or points generated within (tr−1 , tr ) in state space D
are tied to locations which are sequences of nested counts of total number of points
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 167

generated at time tr , denoted n(tr ), and sequence of points k(t1 ), k(t2 ), . . . , k(tm ),
where 1 ≤ k(t1 ) ≤ n(t1 ); 1 ≤ k(t2 ) ≤ n(t2 ); . . . ; 1 ≤ k(tr ) ≤ n(tr ); . . . ; 1 ≤
k(tm ) ≤ n(tm ), and n(tr ) represents the total number of points generated at time
tr within specific locations. The mathematics then requires notation capable of
representing outcomes across multiple time point steps and subareas. The associated
count within Dkt defines a sequence {Y tr } of number of occurrences within [tr−1 , tr )
t
in subarea Dk(t) and follows the Markov property, i.e., given the current state of the
system, we can make predictions about the future state without regard for previous
states. This is a discrete time finite Markov chain.
To minimize complexity, we define the count process {Y t , t ≥ 0} of occurrences
between consecutive times tr−1 and tr to be

PY t (tr − tr−1 ) = P (λ(t)D |D|),

where λD (t) denotes the local intensity for any t within [tr−1 , tr ). Such a process
has stationary transition probabilities and transition matrix of the counts Q = (Qij )
constructed from the transition probabilities, Pij (s), where Qij is the (i, j ) count
from Pij (s) from consecutive time periods t1 , t2 , . . . , tm adjusted with subarea
containing locations i and j , and Qii = 0, ∀ i ∈ S, at each time period.
tr tr−1
We extend Dkt such that they form a nested sequence Dk(t r)
⊂ Dk(t r−1 )
, the
t , 1 ≤ k(t) ≤ n(t). Following Resnick
subarea within time (tr−1 , tr ) and count Yk(t)
(2002), we can define the nested subareas Dk(t) t and each fixed time t, a sequence
{Ek(t) }k≥0 of independent and identically distributed exponential random variables
t

with unit mean such that Ek(t) t t , 1 ≤ k(t) ≤ n(t)

is independent of Y (t) on Dk(t)
and set Yk(t) equal to the count within Dk(t) for interval time [tr−1 , tr ).
t t

By discretizing time, we then define two finite sequences {(Y tr ), (tr )}, where

n(t)
Yt = t
Yk(t) t
with locations Yk(t) for tn < t < tn+1 and {tr }, r = 1, . . . , m
k(t)=1
˜
is the sequence of times when the process is observed. The sequence {Y t } is the
cumulative count of occurrence within interval time [tr−1 , tr ) in a subset of space D
such that tr − tr−1 is conditionally independent and exponential given Y tr−1 . More
precisely, the sequence {Yk(t)
t } is defined as follows:

Y1t1
..
.
t1
Yk(t 1)

..
.
t1
Yn(t 1)
168 J. L. Matthews et al.

n(t 1)
t1
and Y t1 = Yk(t 1)
;
k(t1 )=1

..
.
Y1tr
..
.
tr
Yk(t r)

..
.
tr
Yn(t r)

n(t r)
tr
and Y tr
= Yk(t r)
;
k(tr )=1

..
.
Y1tm
..
.
tm
Yn(t m)

n(t m)
tm
and Y tm
= Yk(t m)
,
k(tm )=1

where n(tr ) is the total number of points generated at time tr .

3 Moran’s Statistics

The global Moran’s index, I , based on a sample of n observations is defined as:

n n
n i=1 j =1 wij Zi Zj
I= n 2
,
i=j wij i=1 Zi

where Zi and Zj are the normalized values of spatial characteristic or feature

indications i and j , and wij is the weight between features i and j (see Moran
1950).
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 169

Using Eq. (1), Vaillant et al. (2011) modeled the propagation of sugarcane yellow
leaf virus with a focus on the spatial spread of disease over six time periods (weeks
6, 10, 14, 19, 23, and 30 in the growing season). For each pair of consecutive
observation dates (ti−1 , ti ), i = 1, 2, . . . , 6, they defined a Moran-type index based
on a nearest neighbor scheme as follows:

Mi = wx,y 1[0,ti−1 ] (Tx )1(ti−1 ,ti ] (Ty ),

(x,y)∈D

where D denotes the discrete set of plant locations, Tx denotes the date (time
variable) of virus detection for plant x, 1[0,ti−1 ] (Tx ) is an indicator whether time
Tx falls in the interval [0, ti−1 ], and wx,y denotes the weights, which are non-zero
only if x and y are neighbors, located within the same subarea. There are several
possible weight functions as we will describe later in this section.
We will utilize a similar space-time definition of Moran’s index. Since the
partitioned areas vary from one time to the next, we define Dkt as subarea k ≥ 1
at time t and then we will define a Moran-type autocorrelation statistic on each
measurable subset Dkt as:

Mkt = wu,u 1[t−1,t) (Tu , Tu ), k ≥ 1, t ≥ 1, (2)

u,u ∈Dkt

where wu,u denotes the spatial weight between points u and u of disease/event
occurrence (wu,u = 0), Tu and Tu denote the time of detection of u and u , and
1[t−1,t) (Tu , Tu ) is an indicator of whether times Tu and Tu both fall in the interval
[t − 1, t).
In the definition of Moran’s autocorrelation index, wu,u represents a spatial
weight between any two distinct event locations generated within disc Dkt , which
could be a function, e.g.,
(i) the inverse distance between two points;
(ii) the inverse distance squared between two points;
(iii) an estimate of the autocorrelation/semivariance statistic;
(iv) the “geographical” weights defined as

e(−dij /r) j = i,
wij =
0 otherwise,

where r is the maximum distance in the minimum tree that spans all points and
does not have any nodes that link back to itself as in Murakami et al. (2017).
As stated in Chen (2012), selection of the weight function objectively is an open
question. Different weight functions have different spatial effects. Later in Sect. 4,
we give reasons why we choose a particular weight function.
170 J. L. Matthews et al.

Also, while values of the Moran-type statistic may not always be between −1
and 1 (the full range depends on the weights), they are essentially in the same spirit
as larger values indicate larger autocorrelation measures.
We will use a modified version of Geary’s C autocorrelation index (Geary 1954;
Sokal et al. 1998) as a comparison to determine and evaluate the trend between the
two measures. We will define it as Ckt :

n−1

Ckt = wu,u 1(0,t−1] (Tu )1(t−1,t] (Tu ),

2W0  t u,u ∈Dk

with wu,u as in Eq. (2) and W0 = wu,u , the sum of weights for all points
u,u ∈D t
generated across all discs at time t.

3.1 Expected Value

Cliff and Ord (1981) derived the moments of Moran’s index given the assumption
that observations are random independent drawings from normal populations. Under
this assumption, the expected value of Moran’s index is given as:

n E i,jwij Zu(t) Zu (t) 1
E(I ) =  =− ,
S0 E Z 2 n − 1
i i


where S0 = wij . To link our modified Moran-type autocorrelation to the global
Moran’s index, it can be viewed as

Mkt = wu,u 1[t−1,t) (Tu , Tu )

u,u ∈Dkt

= wu,u Zu(t) Zu (t) ,

u,u ∈Dkt


1 if u appeared in Dkt in the time period [t − 1, t),
where Zu(t) = 1D t (Tu ) =
k 0 otherwise.
Then Zu(t) is the indication of occurrence of event within the disc centered at an
event from the previous time point.
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 171

Given the assumption that observations are random independent drawings within
the same disc from a given distribution, the expected value of the proposed
spatio-temporal Moran’s autocorrelation statistic can be derived as follows:
⎛ ⎞

E(Mkt ) = E ⎝ wu,u (Zu(t) Zu (t) )⎠

u,u ∈Dkt

 
= wu,u E Zu(t) Zu (t)
u,u ∈Dkt

where u and u are events that occurred under randomization process.

4 Simulation Study

Here we provide an illustration of the theory built in the previous section. We

conduct the simulation using R with the spatstat package and represent the
weights of the Moran’s index as the inverse distance between two points. The weight
function was chosen as described in Chen (2012) since we are modeling spread of
disease which is a large-scale complex system and we intend to capture transitions
globally to locally. We begin with an observed unit area (1 × 1 unit) and generate
a Poisson point processes within that area. The Poisson process is used as general
setup since it is widely applicable and is applicable to many real-world situations
such as disease cases, radioactive decay, and plant propagation. Other processes,
for example, negative binomial and Cox, may be utilized to explain other space-
time point phenomena. Next, we define the Voronoi cells around each point (e.g.,
the locus of locations nearer to that point than any other). Each location site is then
used as the center point to generate a disc with diameter equal to the minimum of the
distance to the Voronoi edge and the distance to the observed area edge (edge of D).
This effectively means that an infected site can only subsequently infect sites that
are closer to it than any other infected site. These Voronoi discs define the subareas
t
Dk(t) for the point process generated at the next time interval (step). In application,
subareas can be tied to locations such as country, state, county, and zip code. The
choice of subareas could be used to determine whether more affluent regions are able
to deal more effectively with disease outbreak, for example. The Moran’s statistic is
calculated at each step as the sum of the inverse distances between an offspring site
and another offspring site. This format will continue through time t = 5 adhering
to the Markov property (t = 5 chosen to follow results from a forestry example
in Meddens and Hicke 2014 or as in Vaillant et al. 2011). Each subarea will be
magnified by a scale parameter, α > 0, such that the new radius is α times the
previous radius with intensity function λ ∗ t, following the spatial resolution of
aggregated grid cells that experienced a 1% tree mortality as in Meddens and Hicke
(2014).
172 J. L. Matthews et al.

Algorithm: Iterative local Moran’s statistics

Procedure:
(i) Define the partition and generate a Poisson process in the area with intensity
λ.
(ii) Calculate the Moran’s statistic.
(iii) Define new subareas with point from previous process as center and radius
equal to the minimum of distance to the edge of the window or distance to
the edge of the Voronoi cell.
(iv) Rescale subarea and iterate local points from a Poisson process with
intensity λ ∗ t.
(v) Compute the local Moran’s statistic for each subarea at that time step.
Repeat steps (iii) through (v).
Stop when all times are reached.
End
Figure 1 shows the original observed area with the seven points generated at
t = 1 alongside the discs generated from these points and the points generated at
t = 2.
The Moran’s value for t = 1 equals 31.65. The number of points generated in
each Voronoi subarea, the rescaled area, and Moran’s values generated at time t = 2
are displayed in Table 1 with λ = 4 and α = 5. For example, at point 1 in Fig. 2a,
we generate a disc of maximum radius within the Voronoi cell centered at the point
generated at the previous time with 10 points as shown in Fig. 2b. The number of
points generated in other Voronoi cells is displayed in Table 1. The Geary’s values
are also computed. The ordering of Moran’s and Geary’s values display resemblance
as they “rank” the strength of correlations in a similar way.

Time t=1 Time t=2

2 2

5 5

3 3

6 6
4 4

1 1

(a) (b)

Fig. 1 Poisson point process plots, λ = 4 (a) t = 1 and (b) t = 2

Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 173

Table 1 Non-zero Moran’s Disc Points Area Moran’s Geary’s C

and Geary’s c values for
λ = 4 and α = 5 at t = 2 1 10 1.064 158.18 2.444
2 7 0.361 129.29 1.998
3 24 3.096 453.09 7.000
4 9 1.266 111.58 1.724
5 10 1.058 116.03 1.793
6 28 3.415 542.49 8.382

Fig. 2 Discs 1–6 (a–f) at t = 3

The process is described at time t = 3 with disc 1 having 10 subareas (see Fig. 3),
disc 2 having 7 subareas, etc. The non-zero Moran’s values for disc 1 at time t = 3
are listed in Table 2.
After five time intervals, the clustering of points generated is displayed in Fig. 3,
with red shades representing the most dense areas and white shades representing
areas with no instances of disease. Such a representation would not have been clearly
observed if the area was not partitioned and if the rate of spread was not tabulated
in space and time. This framework will allow us to detect clusters and estimate the
density.
174 J. L. Matthews et al.

Fig. 3 Apparent clustering

through t = 5

Table 2 Non-zero Moran’s Disc Points Area Moran’s Geary’s C

and Geary’s c values for disc
1 at t = 3 1 28 1.905 779.89 5.101
5 11 1.362 125.23 0.819
8 2 0.065 20.11 0.132
9 15 1.183 337.63 2.208
10 7 0.508 80.63 0.527

5 Model Analysis

Here we investigate any trend in our Moran’s values based on the number of points
generated and area of the disc. Figure 4 shows area and number of points plotted
versus Moran’s value. The plots of Moran’s values for time t = 2 and for disc 1
at t = 3 and t = 4 show an approximate linear trend between Moran’s values and
number of points generated, even though there is no apparent relationship between
area and Moran’s values.
The 3-D plot in Fig. 5 shows that at each successive time point the number of
points generated versus the Moran’s gets progressively larger. This is due to the
number of points generated increasing as time increases, but the area increase is
moderated as time increases. This results in points that are closer together, thus
increasing the sum of inverse distances and the Moran’s value.
Focusing on the six points generated at t = 1 and the associated discs at t = 2,
we investigate the idea of correlation between two consecutive Poisson distributions.
Table 3 shows the number of points generated within each of the original six discs
at times t = 2, 3, 4, 5 and also includes the area of each disc. The total number of
points generated is also shown and each of these points has an associated Moran’s
statistic at the next time point.
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 175

Fig. 4 Trend analysis (a) t = 2, (b) t = 3 for disc 1, and (c) t = 4 for disc 1
2000
1500
Moran
1000

# of points
500

40
30
20
10
0

0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
area
t=1 t=2 t=3 t=4 t=5

Fig. 5 Plot of Moran’s values

Table 3 Counts at Time

t = 2, 3, 4, 5
Disc Area 2 3 4 5
1 1.064 10 66 437 2781
2 0.361 7 11 37 174
3 3.096 24 235 1303 7710
4 1.266 9 24 90 547
5 1.058 10 89 738 4399
6 3.415 28 336 2223 13305
Total points 88 761 4828 28,916
176 J. L. Matthews et al.

6 Conclusion

Our designed model has shown that global properties of spatio-temporal disease
spread can be tuned in by modifying the space and adding time efficiently. The
results show that long range, spread feature, and variability yielded spatial clustering
of occurrence events. To obtain a quantitative analysis of the performance of
the Moran’s values, we compared them with adjusted Geary’s C statistics. We
controlled for targeted subareas and such an approach provides a framework for
understanding “disorganized” or “disordered” evolution of some natural organisms
in the form of cluster analysis. The distribution of subareas can be found based on
our density estimation or concentration under the proposed spatio-temporal Moran’s
index.
Extension to different dynamics other than the Poisson will be considered in
future studies and widen the scope of the spatio-temporal model incorporating
covariates.

Acknowledgements The research for this paper was made possible by financial support from the
U.S. Naval Academy. Material contained herein is solely the responsibility of the authors and is
made available for the purpose of peer review and discussion. Its contents do not necessarily reflect
the views of the Department of the Navy or the Department of Defense.

References

Anselin, L.: Local indicators of spatial association - LISA. Geogr. Anal. 27(2), 93–115 (1995).
https://doi.org/10.1111/j.1538-4632.1995.tb00338.x
Baddeley, A., Rubak, E., Turner, R.: Spatial Point Patterns: Methodology and Applications with R.
CRC Press, Boca Raton (2016)
Chen, Y.: On the four types of weight functions for spatial contiguity matrix. Lett. Spat. Resour.
Sci. 5, 65–72 (2012). https://doi.org/10.1007/s12076-011-0076-6
Cliff, A., Ord, J.: Space-time modelling with an application to regional forecasting. Trans. Inst. Br.
Geogr. 64, 119–128 (1975). https://doi.org/10.2307/621469
Cliff, A., Ord, J.: Spatial Processes - Models and Applications. Pion Limited, London (1981)
Geary, R.C.: The contiguity ratio & statistical mapping. Inc. Stat. 5(3), 115–145 (1954). https://
doi.org/10.2307/2986645
Jones-Todd, C.M., Swallow, B., Illian, J.B., Toms, M.: A spatio-temporal multispecies model of
a semicontinuous response. J. R. Stat. Soc. Ser. C 67(3), 705–722 (2018). https://doi.org/10.
1111/rssc.12250
Kallenberg, O.: Foundations of Modern Probability, 2nd ed. Springer, New York (2001)
Kieu, K., Adamczyk-Chauvat, K., Monod, H., Stoica, R.S.: A completely random T-tessellation
model and Gibbsian extensions. Spat. Stat. 6, 118–138 (2013). http://dx.doi.org/10.1016/j.
spasta.2013.09.003
Lawson, A.: Bayesian Disease Mapping: Hierarchical Modeling in Spatial Epidemiology. Chap-
man & Hall/Crc Interdisciplinary Statistics Series. CRC Press, Boca Raton (2009)
Lee, J., Li, S.: Extending Moran’s index for measuring spatiotemporal clustering of geographic
events. Geogr. Anal. 49, 36–57 (2017). https://doi.org/10.1111/gean.12106
Martin, R.L., Oeppen, J.E.: The identification of regional forecasting models using space:time
correlation functions. Trans. Inst. Br. Geogr. 66, 95–118 (1975). https://doi.org/10.2307/621623
Quantifying Spatio-Temporal Characteristics via Moran’s Statistics 177

Meddens, A.J.H., Hicke, J.A.: Spatial and temporal patterns of Landsat-based detection of tree
mortality caused by a mountain pine beetle outbreak in Colorado, USA. For. Ecol. Manag. 322,
78–88 (2014). https://doi.org/10.1016/j.foreco.2014.02.037
Moran, P.A.P.: Notes on continuous stochastic phenomena. Biometrika 37(1–2), 17–23 (1950).
https://doi.org/10.1093/biomet/37.1-2.17
Murakami, D., Yoshida, T., Seay, H., Griffith, D.A., Yamagata, Y.: A Moran coefficient-based
mixed effects approach to investigate spatially varying relationships. Spat. Stat. 19, 68–69
(2017). https://doi.org/10.1016/j.spasta.2016.12.001
Pace, R.K., LeSage, J.P.: Omitted variable and spatially dependent variables. In: Páez, A., et al.,
(eds.) Progress in Spatial Analysis: Advances in Spatial Science, pp. 17–28. Springer, Berlin
(2009)
Reitzner, M., Spodarev, E., Zaporozhets, D.: Set reconstruction by Voronoi cells. Adv. Appl.
Probab. 44(4), 938–953 (2012). https://doi.org/10.1239/aap/1354716584
Resnick, S.I.: Adventures in Stochastic Processes. Birkhäuser, Boston (2002). https://doi.org/10.
1007/978-1-4612-0387-2
Sokal, R.R., Oden, N.L., Thomson, B.A.: Local spatial autocorrelation in a biological model.
Geogr. Anal. 30(4), 331–354 (1998). https://doi.org/10.1111/j.1538-4632.1998.tb00406.x
Thäle, C., Yukich, J.E.: Asymptotic theory for statistics of Poisson-Voronoi approximation.
Bernoulli 22(4), 2372–2400 (2016). https://doi.org/10.3150/15-BEJ732
Vaillant, J., Puggioni, G., Waller, L.A., Daugrois, J.: A spatio-temporal analysis of the spread of
sugarcane yellow leaf virus. J. Time Ser. Anal. 32, 392–406 (2011). https://doi.org/10.1111/j.
1467-9892.2011.00730.x
Wang, Y.F., He, H.L.: Spatial Data Analysis Method. Science Press, Beijing (2007)
Wang, H., Cheng, Q., Zuo, R.: Quantifying the spatial characteristics of geochemical pattern via
GIS-based geographically weighted statistics. J. Geochem. Explor. 157, 110–119 (2015)
Zhou, H., Lawson, A.B.: EWMA smoothing and Bayesian spatial modeling for health surveillance.
Stat. Med. 27, 5907–5928 (2008). https://doi.org/10.1002/sim.3409