University of Connecticut OpenCommons@UConn
Doctoral Dissertations University of Connecticut Graduate School
5-22-2020
On Bayesian Methods for Spatial Point Processes
Jieying Jiao University of Connecticut - Storrs, [email protected]
Follow this and additional works at: https://opencommons.uconn.edu/dissertations
Recommended Citation Jiao, Jieying, "On Bayesian Methods for Spatial Point Processes" (2020). Doctoral Dissertations. 2443. https://opencommons.uconn.edu/dissertations/2443 On Bayesian Methods for Spatial Point Processes
Jieying Jiao, Ph.D. University of Connecticut, 2020
ABSTRACT
Spatial point pattern data are routinely encountered. A flexible regression model for the underlying intensity is essential to characterizing and understanding the pattern. Spatial point processes are a widely used to model for such data. Additional measurements are often available along with spatial points, which are called marks. Such data can be modeled using marked spatial point processes.
The first part of this dissertation focuses on the heterogeneity of point processes.
We propose a Bayesian semiparametric model where the observed points follow a spa- tial Poisson process with an intensity function which adjusts a nonparametric baseline intensity with multiplicative covariate effects. The baseline intensity is approached with a powered Chinese restaurant process (PCRP) prior. The parametric regression part al- lows for variable selection through the spike-slab prior on the regression coefficients. An efficient Markov chain Monte Carlo (MCMC) algorithm is developed. The performance of the methods is validated in an extensive simulation study and the Beilschmiedia
pendula trees data.
Spatial smoothness is often observed in some environmental spatial point pattern Jieying Jiao, University of Connecticut, 2020 data, and the PCRP may have lower efficiency for such data since it allows more flex- ibility without any spatial constraint. Distance dependent Chinese restaurant process
(ddCRP) can be easily realized by introducing a decay function to Chinese restaurant process. The second part of this dissertation introduces the ddCRP model with Bayesian inference methods, whose performance is illustrated using simulation study.
In the third part, we investigate the marked spatial point process, which is moti- vated by the basketball shot data. We develop a Bayesian joint model of the mark and the intensity, where the intensity is incorporated in the mark’s model as a covariate.
An MCMC algorithm is developed to draw posterior samples from this model. Two
Bayesian model comparison criteria, the modified Deviance Information Criterion and the modified Logarithm of the Pseudo-Marginal Likelihood, are developed to assess the
fitness of different models focusing on the mark. Simulation study and application to
NBA basketball shot data are conducted to show the performance of proposed methods. On Bayesian Methods for Spatial Point Processes
Jieying Jiao
B.S., University of Science and Technology of China, Hefei, China, 2016
A Dissertation Submitted in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy at the University of Connecticut
2020 ii
Copyright by
Jieying Jiao
2020 iii
APPROVAL PAGE
Doctor of Philosophy Dissertation
On Bayesian Methods for Spatial Point Processes
Presented by Jieying Jiao, B.S.
Major Advisor Dr. Jun Yan
Associate Advisor Dr. Haim Bar
Associate Advisor Dr. Dipak Dey
Associate Advisor Dr. Xiaojing Wang
University of Connecticut 2020 iv
To My Parents v Acknowledgments
First and most important, I want to express my sincere gratitude to my advisor, Professor
Jun Yan. He has always been supportive and has given me the freedom to pursue various projects. It would never have been possible for me to take this work to completion without his patient guidance and encouragement.
I would like to thank Professor Haim Bar, Professor Dipak Dey, and Professor Xiao- jing Wang for serving as my general exam and dissertation committee. The discussions with them and questions they proposed helped broaden my research. I also want to thank all the professors that have taught me during my time at UConn. The knowledge
I have learned from them is invaluable.
My sincere thanks also goes to Dr. Guanyu Hu for his generous help and support to my research work. His passion to research motivated me to dig deeper on any problems.
It has been a pleasure to work with him, and I have learned a lot from this process.
My special thanks goes to Professor Peng Zhang for involving me to the cybersecurity project and providing funding to my Ph.D. research. I also want to thank Dr. Yishu
Xue for her help to my research work.
For all my close friends I met at UConn and old friends back in China, I want to thank them for all the help and support they have provided to me. My life as a Ph.D. student would not be this meaningful without them. vi
I am truly grateful to my parents for their immeasurable love and care. I feel so lucky to have parents that trust me unconditionally. The warmth I feel from my parents, my brother and my grandparents is the treasure in my life. vii Contents
Acknowledgmentsv
1 Introduction1
1.1 Spatial Point Process...... 2
1.2 Nonparametric Methods for Intensity Function...... 5
1.3 Dirichlet Process...... 6
1.4 Mixture of Finite Mixtures...... 8
1.5 Prior on Membership Parameter z ...... 10
1.5.1 Stick-Breaking Process...... 11
1.5.2 Chinese Restaurant Process...... 12
1.5.3 Overclustering Problem of DP...... 14
1.6 Spatial Smoothness of the Intensity of Point Process...... 19
1.7 Marked Spatial Point Process...... 23
2 Hierarchical Semiparametric Poisson Point Process Model 26
2.1 Introduction...... 26
2.2 Model Setup...... 28
2.2.1 Spatial Poisson Point Process...... 28
2.2.2 Nonparametric Baseline Intensity...... 29 viii
2.2.3 Hierarchical Semiparametric Regression Model...... 31
2.3 Bayesian Inference...... 32
2.3.1 The MCMC Sampling Scheme...... 32
2.3.2 Inference for the Nonparametric Baseline Intensity...... 37
2.3.3 Selection of r ...... 40
2.4 Simulation Study...... 41
2.5 Point Pattern of Beilschmiedia Pendula ...... 49
2.6 Discussion...... 55
3 Spatial Poisson Point Process with Spatial Smoothness Constraint 57
3.1 Distance Dependent CRP...... 57
3.2 Model Setup...... 58
3.3 Bayesian Inference...... 60
3.4 Simulation Study...... 62
3.5 Discussion...... 71
4 Marked Spatial Point Process 73
4.1 Introduction...... 73
4.2 Shot Chart Data...... 74
4.3 Model Setup...... 77
4.3.1 Marked Spatial Point Process...... 77
4.3.2 Prior Specification...... 79 ix
4.3.3 Bayesian Variable Selection...... 80
4.4 Bayesian Inference...... 81
4.4.1 The MCMC Sampling Schemes...... 81
4.4.2 Bayesian Model Selection for the Mark Model...... 82
4.5 Simulation...... 84
4.5.1 Estimation...... 84
4.5.2 Variable Selection...... 86
4.6 Real Data Analysis...... 92
5 Future Work 100
5.1 Spatially Dynamic Variable Selection...... 100
5.2 Clustered Regression Coefficients Model for Spatial Point Process.... 101
5.3 Multivariate Point Process Model...... 102
A Appendix 105
A.1 BITC Validation Under Gaussian Mixture Model...... 105
Bibliography 110 x List of Tables
1 Simulation estimation results. “SD” is the empirical standard deviation
over 100 replicates, “SD”c is the average of 100 standard deviations cal-
culated using posterior sample. “AR” is the variable selection accuracy
rate over 100 replicates...... 47
2 Posterior means, standard errors (SE), and the 95% HPD credible in-
tervals for the regression coefficients in the analysis of the spatial point
pattern of Beilschmiedia pendula in the BCI data. Covariates with a “*”
are significant...... 53
3 Shot data summary for four players in the 2017–2018 regular NBA season.
The period includes 4 quarters and over time...... 75
4 Summaries of the bias, standard deviation (SD), average of the Bayesian
SD estimate (SD),c and coverage rate (CR) of 95% credible intervals when
Z2 is continuous: ξ = α0 = 0.5, α2 = 1, (β1, β2) = (2, 1) and Z2 ∼ N(0, 1). 87
5 Summaries of the bias, standard deviation (SD), average of the Bayesian
SD estimate (SD),c and coverage rate (CR) of 95% credible intervals
when Z2 is binary: ξ = α0 = 0.5, α2 = 1, (β1, β2) = (2, 1) and Z2 ∼
Bernoulli(0.5)...... 88 xi
6 Percentages of correct decision of the variables in 200 replicates with
continuous Z2. The significant parameters except α1 all equal to 1. All
covariates were generated with standard Normal distribution...... 90
7 Percentages of correct decision of the variables in 200 replicates with
binary Z2. The significant parameters except α1 all equal to 1. Z2 was
generated using Bernoulli(0.5), while other covariates were from standard
normal...... 91
8 Covariates used in the intensity model and the mark model...... 93
9 Summaries of mDIC and mLPML for the mark model and DIC for the
full joint model...... 94
10 Data analysis result using intensity dependent model for Curry and Durant. 97
11 Real data analysis results using intensity dependent model for Harden
and James...... 98
12 Simulation results of parameter estimation for Gaussian Mixture model
with symmetric Dirichlet process prior...... 109 xii List of Figures
1 Chinese restaurant process illustration...... 13
2 Illustration of modified Chinese restaurant process based on MFM.... 16
3 Histograms of Kb and overlaid trace plots of RI from the 100 replicates.
Setting 1 and 2 are on left and right panel, respectively. The optimal
r was selected by the BITC. The RI for each replicate was under the
optimal r for that replicate. The thick lines are the average of the trace
plots over the 100 replicates...... 43
4 Histogram of Kb chosen by LPML and DIC over 100 replicates. Results
of setting 1 and 2 are on left and right panel, respectively...... 44
5 Simulation configurations for baseline intensity, with fitted baseline in-
tensity surfaces. Median and quantiles are calculated out of 100 replicates. 46
6 Boxplots of MSE over the 100 replicates for four models...... 48
7 The locations of Beilschmiedia Pendula and heat maps of the standardized
covariates of the BCI data...... 51
8 BCI data analysis: trace plots of β and K after burnin and thinning... 52
9 Heat map of the fitted baseline intensity surface of the BCI data using
r = 1.4. From top to bottom: 2.5%, 50%, and 97.5% percentiles of
posterior distribution of λ(s)...... 54 xiii
10 Histograms of Kb and overlaid trace plots of RI from the 100 replicates.
Setting 1 and 2 are on left and right panel, respectively. The optimal h
was selected by the LPML or DIC. The thick lines are the average of the
trace plots over the 100 replicates...... 65
11 Simulation configurations for intensity surface, with fitted intensity sur-
faces. Median and quantiles are calculated out of 100 replicates...... 67
12 Histograms of Kb and overlaid trace plots of RI from the 100 replicates.
Setting 1 and 2 are on left and right panel, respectively. The optimal r
was selected by BITC. The thick lines are the average of the trace plots
over the 100 replicates...... 69
13 Simulation configurations for intensity surface, with fitted intensity sur-
faces. Median and quantiles are calculated out of 100 replicates...... 70
14 Shot charts of Curry, Durant, Harden and James in the 2017–2018 regular
NBA season...... 76
15 Intensity fit results of Curry, Durant, Harden and James on the same
scale. Red means higher intensity...... 96
16 Histogram of Kb selection by BITC, LPML and DIC under GM model and
symmetric Dirichlet prior...... 108 1 Chapter 1
Introduction
Spatial point pattern data, which are random locations of certain events of interest in
space (e.g., Diggle, 2013), arise routinely in many field. Such pattern can be, for example,
locations of basketball shooting attempts in sports analytic (Miller et al., 2014), earth-
quake centers in seismology (Schoenberg, 2003), or tree species in forestry (Leininger
et al., 2017; Thurman and Zhu, 2014). We first introduce the spatial point process,
which is often used to model such data. It uses an intensity surface over the inter-
ested region to characterize the probability that certain event happens on any location
of the region. The intensity often varies at different locations, and the main task is
to capture such heterogeneity. Parametric methods are used when there are covariates
available, and nonparametric methods are also actively discussed due to its flexibility.
Most popular nonparametric methods includes Dirichlet process (DP) and Mixture of
Finite Mixtures (MFM) model, which will configure the nonhomogeneous intensity value
as from different groups. One convenient way to set up DP and MFM is to introduce
an index parameter z. The hierarchical model structure makes Bayesian inference nat- ural for such model. The prior distribution on z can then be set up as stick-breaing 2
process or Chinese restaurant Process (CRP). DP or MFM both assume independent
prior distribution for the index parameter z. We then consider the spatial homogeneity
property of the intensity surface displayed by some dataset. Different ways to put spatial
smoothness constraints are introduced. At last, the marked spatial point pattern data
are discussed, and the marked spatial point process is introduced to such data.
1.1 Spatial Point Process
Spatial point pattern data are modeled by spatial point processes (Diggle et al., 1976)
characterized by a quantity called intensity. Within an area B, the intensity on any
location s ∈ B can be represented as λ(s), which is defined as:
E[N(ds)] λ(s) = lim , |ds→0| |ds| where ds is an infinitesimal region around s, |ds| represents its area, and N(ds) shows the number of events happened over ds. Poisson distribution is often used to model count data, so a spatial Poisson point process is widely used for point pattern data. The spatial Poisson point process is defined by the following two conditions (Thurman and
Zhu, 2014): 3
1. the count of data points over any region A follows a Poisson distribution:
N(A) ∼ Poisson(λ(A)), (1.1) Z λ(A) = λ(s)ds. A
2. the joint density f of the event location S = {s1, s2,..., sN } is proportional to the
product of intensity functions λ(si), condition on the total number of events N:
N Y f(S | N) ∝ λ(si). i=1
The second condition says that, condition on the number of events, the locations of events are independent. The likelihood function can then be expressed as:
QN i=1 λ(si) L = R . (1.2) B λ(s)ds
There are some other spatial point processes. Cox processes (Cressie, 1992) is also called doubly stochastic Poisson process. It is an extension of Poisson process, which assumes the intensity of Poisson process is a stochastic process itself. This includes the classic Log Gaussian Cox process (LGCP) (Thurman et al., 2015; Miller et al., 2014), which uses a Gaussian process for the log of intensity in the spatial Poisson point process.
Using a Gaussian Markov random filed (GMRF) in LGCP allows spatial smoothness and accounts for spatial correlation among locations that can not be explained by the 4
covariates (Yue and Loh, 2011). However, such double-stochastic model is often hard
to evaluate and computational challenge to fit (Murray et al., 2012). Gibbs process
(Illian et al., 2008; Chiu et al., 2013) focuses on the dependent point pattern, where the
dependence could be attractive, repulsive, depending on geometric features. Neyman-
Scott process (Cressie, 1992) is a hierarchical model with three levels, and the upper
level is a nonhomogeneous Poisson process. The observed data points are only generated
from the lowest layer of the hierarchy. Due to the simplicity of Poisson process, and
also its advantage that it is theoretically tractable and straightforward to implement
computationally, we will focus on spatial Poisson point process in this dissertation.
When λ(s) varies with different locations, the spatial point process is called non- homogeneous, which is often the case in real life. The primary interest about spatial point process is to characterize the heterogeneous intensity. When spatial covariates are available, most works focus on parametric methods to incorporate these covariates into intensity function. Let X(s) being p dimension vector showing the covariates value on
location s, then a common way to model intensity is to use exponential link function: