Computational Statistics and Data Analysis 167 (2022) 107349

Contents lists available at ScienceDirect

Computational Statistics and Data Analysis

www.elsevier.com/locate/csda

Hierarchical Bayesian modeling of spatio-temporal

area-interaction processes ✩
Jiaxun Chen a,∗ , Athanasios C. Micheas b , Scott H. Holan b
a
Eli Lilly and Company, Lilly Corporate Center, Indianapolis, IN 46285, United States of America
b
Department of Statistics, University of Missouri, 146 Middlebush Hall, Columbia, MO 65211-6100, United States of America

a r t i c l e i n f o a b s t r a c t

Article history: To model spatial point patterns with discrete time stamps a flexible spatio-temporal area-
Received 6 October 2020 interaction point process is proposed. In particular, this model is suitable for describing
Received in revised form 19 May 2021 the dependency between point patterns over time, when the new point pattern arises
Accepted 13 September 2021
from the previous point pattern. A hierarchical model is also implemented in order to
Available online 20 September 2021
incorporate the underlying evolution process of the model parameters. For parameter
Keywords: estimation, a double Metropolis-Hastings within Gibbs sampler is used. The performance of
Autoregressive prior the estimation algorithm is evaluated through a simulation study. Finally, the point pattern
Bayesian analysis forecasting procedure is demonstrated through a simulation study and an application to
Double Metropolis-Hastings within Gibbs United States natural caused wildfire data from 2002 to 2019.
sampler © 2021 Elsevier B.V. All rights reserved.
Hierarchical model
Spatio-temporal area-interaction process

1. Introduction

Studying point processes over time has become increasingly important in many disciplines. For example, in criminology
or in seismology we wish to predict future occurrences of criminal activity or earthquakes, respectively. The time an event
occurs can be crucial to the understanding of several characteristics of the phenomenon, including the evolution of the point
process, connections with other events in the same domain, or the time of possible observations of future events. As such,
to fully understand the stochastic mechanism of the process of locations as it evolves over time, it is natural to include the
time of occurrence as another coordinate in the event.
Spatio-temporal point process models can either be viewed in continuous or discrete time. The continuous time point
process model assumes that events can be observed within a continuous spatial and temporal domain, such as the self-
exciting point process proposed by Hawkes (1971), space-time Epidemic-Type Aftershock Sequence (ETAS) model introduced
by Zhuang and Ogata (2006), and the multi-scale spatio-temporal area-interaction model developed by Iftimi et al. (2018).
For discrete time models, points are aggregated over time intervals and formed as point patterns at each time stamp. Recent
applications on discrete time models include the log-Gaussian Cox process with dynamical intensity function proposed by
Brix and Diggle (2001) and the spatio-temporal Poisson point process model with Gaussian mixture components developed
by Zhou et al. (2015).

✩
R and C++ source code of the simulation study is included in the Supplementary Material.
* Corresponding author.
E-mail address: [email protected] (J. Chen).

https://doi.org/10.1016/j.csda.2021.107349
0167-9473/© 2021 Elsevier B.V. All rights reserved.
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

One of the advantages of the continuous time model is that the model preserves the accuracy of the time stamp for
each event and it can predict the precise time of future events. However, the predictive temporal domain needs to include
the observed time interval, since these models consider the time stamp of occurrence as a random variable with support
consisting of the whole temporal domain. In other words, the continuous time model generates events in the past and future
simultaneously. On the other hand, discrete time models can directly predict future events without generating past events,
which is computationally more efficient and suitable to the directional feature of time. Based on past events, events in the
next time interval can be easily predicted. However, the exact time of the future event within the time interval cannot be
provided by discrete time models.
Apart from the temporal component, the spatial structure is also crucial for the analysis. A flexible model should be
able to capture the spatial inhomogeneity and interactions between points. An important class of point processes are Gibbs
point processes, such as the Strauss process (Strauss, 1975), the hardcore process (Kelly and Ripley, 1976), the Geyer’s
saturation process (Geyer, 1998) and the area-interaction process (Baddeley and Van Lieshout, 1995). The Strauss process
and its special case, the hardcore process, can be used to model inhibition between points. The saturation process and area-
interaction process can be applied to both repulsive and attractive point patterns. Properties of the Poisson and Markov
point process models have been extensively studied over the past fifty years. Theoretical and practical treatments can
be found in the texts by Ripley (1987), Karr (1991), Cressie (1993), Barndorff-Nielsen et al. (1999), Van Lieshout (2000),
Lantuejoul (2001), Lawson and Denison (2002), Møller (2013), Møller and Waagepetersen (2003), Daley and Vere-Jones
(2003, 2007), Illian et al. (2008), Gelfand et al. (2010), Chiu et al. (2013), Spodarev (2013), Diggle (2013) and Baddeley et al.
(2015).
In particular, Gibbs processes offer a large class of models which allow any type of interaction (attraction or repulsion)
between the point process events, across space and time and, as such, have received more attention over the past few years
(see Dereudre (2019), for a recent review). Moreover, space-time Gibbs processes can describe phenomena with interactions
at different spatial or spatio-temporal scales. For example, in the case of seismic data, different sources of earthquakes
(faults, active tectonic plates and volcanoes) produce events with different displacements (e.g., Siino et al., 2018) and can be
seen as the superposition of background earthquakes and clustered earthquakes (e.g., Pei et al. 2012). Such multi-structure
phenomena have motivated researchers to construct new spatial point process models, e.g., Matern-type point processes
with applications to tree spatial locations, nerve fiber cells, Greyhound bus stations (Rao et al., 2017), in ecology (Wie-
gand et al., 2007; Picard et al., 2009), in epidemiology (Iftimi et al., 2017), and in seismology (Siino et al., 2017; Micheas,
2019; Sørbye et al., 2019), mainly based on Gibbs and Cox point processes, but not exclusively (e.g., Lavancier and Møller,
2016).
In contrast, there are very few spatio-temporal models in the literature; Gabriel et al. (2017) and Raeisi et al. (2019)
modeled the multi-scale spatio-temporal structure of forest fires occurrences using the log-Gaussian Cox processes (LGCP)
and the multi-scale Geyer saturation process, respectively. Additionally, Iftimi et al. (2018) developed a multi-scale area-
interaction model for varicella cases and Illian et al. (2012) modeled the locations of muskoxen herds using the LGCP with
a constructed covariate measuring local interactions.
In this paper, we propose a discrete-time spatio-temporal area-interaction process, where we implement a Markovian
structure for the spatial interaction between the current and the previous patterns. Specifically, the main effect of the Gibbs
point process at the current time depends on the point pattern of the previous time stamp. This dependence structure can
also incorporate the spatial inhomogeneity of the process. For the interaction between points within a specific time stamp,
an area-interaction function is used.
We propose a hierarchical model structure to describe the underlying evolution process. A time indexed parameter
structure and autoregressive structure are implemented in the next level of the hierarchy. The time indexed structure allows
us to evaluate the point process at each time period, while the autoregressive structure focuses more on the evolution of
the underlying process.
For parameter estimation, we implement the double Metropolis-Hastings (DMH) (Liang, 2010) within Gibbs sampler for
the hierarchical model. Thus, we are able to sample posterior realizations of the parameters of the area-interaction process
at each time using the DMH algorithm and sample the realizations of other parameters by a standard Gibbs step. The
DHM replaces the perfect sampler (Møller et al., 2006; Murray et al., 2012) with a standard MCMC sampler. According
to (Park and Haran, 2018), the DMH is the most practical approach for high-dimensional data. Based on the estimates of
the parameters at each time period and the underlying evolution model, we can effectively predict the parameter values.
Consequently, future point patterns can be predicted based on the evolution parameter estimates and the point pattern in
the previous time period.
This paper proceeds as follows. Section 2 begins by defining the proposed spatio-temporal area-interaction point process.
Section 3 provides details about the hierarchical Bayesian formulation including the choices of prior distribution, joint
posterior distribution, and the resulting full conditional distributions. Section 4 provides a sensitivity analysis of the model
parameters. Section 5 presents the results of a simulation study that illustrates the effectiveness of the estimation method
and forecast procedure. Section 6 illustrates an application to the United States natural caused wildfire data from 2002 to
2019. Finally, in Section 7, we provide concluding remarks.

2
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

2. Spatio-temporal area-interaction point process

We introduce a finite spatio-temporal point pattern X on Ws,t ⊂ R2 × Z+ , where Ws is a finite subset of R2 and Z+
denotes the collection of positive integers. Thus, this point process is defined on a finite continuous spatial and discrete
temporal domain.

Definition 1. The density function of the spatio-temporal area-interaction (STAI) process X on R2 × Z+ is


T
1 
nt
nt −m{∪i =
tn
B (xi ,t ,rh )}/(π rh2 )
f (X|β, γ ) = f (X0 ) βtnt g (xi ,t |Xt −1 )γt 1
, (1)
C (βt , γt )
t =1 i =1

where
 
1 1
g (xi ,t |Xt −1 ) = exp − (xi ,t − μi ,t −1 ) T − 1
g (xi ,t − μi ,t −1 ) ,
2 π | g | 2
xi ,t represents the ith point in point pattern Xt , μi ,t −1 = wi (x1,t −1 , . . . , xnt −1 ,t −1 ), t = 1, 2, . . . , T , wi is the vector
of indicators that identify the nearest previous point to xi ,t , i.e., wi = [ I {||xi ,t − x1,t −1 || = min(||xi ,t − x j ,t −1 ||, j =
1, . . . , nt −1 )}, . . . , I {||xi ,t − xnt −1 ,t −1 || = min(||xi ,t − x j ,t −1 ||, j = 1, . . . , nt −1 )}]. nt is the number of points in point pattern
at time t. m represents Lebesgue measure and B (xi ,t , rh ) is the disc centered at xi ,t with radius rh .  g is a positive definite
matrix. The function C (βt , γt ) is the normalizing constant for the model at time t. Notice that X0 is the initial point pattern
(i.e., the observed point pattern at time 0) and any point process model can be used to model f (X0 ).

Lemma 1. The STAI process defined in Definition 1 is a valid process, i.e., f (X|β, γ ) is a density.

Proof. First we note that f (X0 ) is assumed to be the density of any point process model and, therefore, we only need
consider propriety of the evolutions f (Xt |Xt −1 , βt , γt ), so that propriety of the STAI model easily follows using the evolutions
likelihood of (1), i.e.,
   
T 

f (X|β, γ )dX = f (X0 )dX0 f (Xt |Xt −1 , βt , γt )dXt = 1.
t =1

Now in order for the evolution f (Xt |Xt −1 , βt , γt ) to be a proper point process model, we require that
+∞ 
n

t
n −m{∪i =
t n
B (xi ,t ,rh )}/(π rh2 )
C (βt , γt ) = βtnt g (xi ,t |Xt −1 ) γt t 1
< +∞.
nt =0 i =1

Owing to the form of g, we have


nt
g (xi ,t |Xt −1 ) ≤ 1,
i =1

and, therefore, we can write

+∞ 

n
nt −m{∪i =
t
B (xi ,t ,rh )}/(π rh2 )
C (βt , γt ) ≤ βtnt γt 1
.
nt =0

Since the latter is a standard area-interaction model, following Baddeley and Van Lieshout (1995), we have that the right
hand side is finite and, therefore, C (βt , γt ) < +∞, as required. 

The function g (xi ,t |Xt −1 ) in the main effect indicates that the points of the current point pattern will arise from their
neighboring points at the previous time. Specifically, the density of a given location xi ,t follows a normal kernel which is
centered at the nearest point in the previous point pattern. The parameter β affects the number of points at time t, and
parameter γ is the interaction parameter at time t. For γ ∈ (0, 1), the point process shows inhibition between points and
smaller γ values lead to stronger interactions. The STAI process generates clustered point patterns if γ > 1. For γ = 1,
this process is equivalent to the Poisson point process, where the intensity function is determined by the parameter β ,
function g (xi ,t |Xt −1 ), and the point pattern, Xt −1 , at the previous time point. Thus, this model implies that the points of the
current pattern can interact with each other and also interact with the previous point pattern. However, these two types of
interaction functions can be different. The initial point pattern X0 can be generated from any point process.
Algorithm 1 describes the birth-and-death process that can be used to generate point patterns from this process.

3
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Algorithm 1 Birth-and-death process.

for t = 1 to T do
for i = 1 to L do
With probability 0.5 to propose a birth and generate a point x∗i ,t uniformly within W .
Accept the generated point with probability

p (Xt ∪ x∗i ,t )
b(Xt , x∗i ,t ) =
p (Xt )
Or propose a death and randomly remove a point xi ,t from the current point pattern Xt .
Accept the removal of the point with probability

p (Xt \ xi ,t )
d(Xt \ {xi ,t }, xi ,t ) =
p (Xt )

end for
end for

3. Hierarchical Bayesian modeling

We define a Bayesian hierarchical model structure which is suitable for describing the underlying process evolution
and parameter estimation. Additionally, we implement the double Metropolis-Hastings (DMH) algorithm for sampling the
parameters of the area-interaction process and consider two types of structures for the prior distribution of βt and γt , the
time indexed structure and the autoregressive structure. Based on the different structures, we obtain the corresponding
hierarchical models as follows.

3.1. Prior specification

We first consider a time indexed parameter structure, which assumes that the parameters at each time period are
independent of the ones at the other time periods. In other words, the parameters of the area-interaction process will
not evolve over time. Thus, we assume independent time indexed priors for βt and γt at each time t and implement the
standard DMH for parameter estimation. The vague prior distributions of logβt and logγt are both set to N (0, 1000), for
t = 1, . . . , T .
We also propose an autoregressive model for the temporal structure of the parameters of the area-interaction process.
Specifically, we use autoregressive model of order 1 (AR(1)) with nonzero mean as a prior distribution for the logarithms of
βt and γt . That is,
 
logβ1 = b1 |μb1 , σb21 ∼ N μb1 , σb21 ,
 
logβt = bt |bt −1 , φ1 , μ1 , σb2 ∼ N φ1 (bt −1 − μ1 ) + μ1 , σb2 ,
 
logγ1 = ξ1 |μξ1 , σξ21 ∼ N μξ1 , σξ21 ,
and
 
logγt = ξt |ξt −1 , φ2 , μ2 , σ y2 ∼ N φ2 (ξt −1 − μ2 ) + μ2 , σ y2 ,

where φ1 and φ2 are autoregressive coefficients of the underlying AR(1) processes, μ1 and μ2 are the means of the AR(1)
processes, and σb2 and σ y2 are the variances of random errors. A positive autoregressive coefficient means that the main
effect or the interaction parameters at two consecutive time points are positively correlated, indicating that the parameter
values will indeed change gradually over time. This means that the point patterns over time will be similar in terms of
their interactive relationships between points. In contrast, a negative autoregressive coefficient indicates negative correlation
between parameters at two consecutive time points. That is, the parameters will change more dramatically over time and
the generated point patterns will be significantly different from each other at consecutive time points. If the autoregressive
coefficient equals zero, the parameters are independent over time and, therefore, we do not expect any trend in terms of
the change of interaction structures. Parameters μb1 and σb2 are the mean and variance for the distribution of b1 , and μξ1
1
and σξ21 are the mean and variance for the distribution of ξ1 . The prior distributions for these parameters are defined as
follows,

φ1 ∼ N (−1,1) (φ̂1 , σφ21 ), φ2 ∼ U (−1, 1),

2 2
σ ∼ I G (ub , v b ), σ ∼ I G (u y , v y ),
b y

σ ∼ I G (ub1 , v b1 ), σξ21 ∼ I G (u ξ1 , v ξ1 ),
2
b1

4
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

μ1 ∼ N (0, τ1 ), μ2 ∼ N (0, τ2 ),
μb1 ∼ N (0, τb1 ), μξ1 ∼ N (0, τξ1 ),
where u b , v b , u y , v y , u b1 , v b1 , u ξ1 , v ξ1 , τ1 , τ2 τb1 and τξ1 are fixed hyperparameters chosen to form vague hyperprior
distributions.
On average, we note that the estimates of bt have larger biases than the biases of ξt . In order to improve the estimate
of φ1 for small T , the mean of the truncated normal prior is specified as the method of moment estimator φ̂1 , which
is calculated based on the estimated value of bt , t = 1, . . . , T , obtained by stochastic approximation (Geyer, 1998). The
variance of this prior is then provided by the asymptotic distribution of φ̂1 , which is σφ21 = (1 − φ̂12 )/ T . An informative
prior for φ2 does not significantly improve the estimate. Thus, we assume a uniform distribution with support (-1,1) as the
prior distribution of φ2 . Simulation analyses show that the informative prior can have a significant effect on the posterior
distribution when T is small. Thus, we recommend selection of an informative prior be done carefully. For large T , we use
the same uniform prior for both φ1 and φ2 . Moreover, we assume that the parameters r g and  g are fixed and known since
their effects on the configuration of the point process can be confounded. Ideally, we can keep one of these parameters
fixed and estimate the other. This issue is the subject of future research.

3.2. Posterior distribution

In order to implement the posterior sampler, we derive the full posterior distribution of the STAI process associated with
the full conditional distributions as follows:


T
1 
nt
f (b, y , φ1 , φ2 , σb2 , σ y2 , σb21 , σξ21 , μ1 , μ2 , μb1 , μξ1 |X) ∝ exp(bt nt ) g (xti |Xt −1 )
C (bt , ξt )
t =1 i =1
⎛ ⎞ T −1
⎜ 1 ⎟
× exp(−ξt [nt − m{∪ni =t 1 B (xi ,t , rh )}/(π rh2 )]) ⎝  ⎠
2
2πσ b
⎧ ⎫
⎪ T
2⎪
⎛ ⎞ T −1
⎪
⎪ ((bt − μ1 ) − φ1 (bt −1 − μ1 )) ⎪
⎪  
⎨ ⎬ (b1 − μb1 ) ⎜
2
t =2 1 1 ⎟
× exp −  exp − ⎝ ⎠
⎪
⎪ 2 σ 2 ⎪
⎪ 2 2 σ 2 2
⎪
⎩ b ⎪
⎭ 2πσb1 b1 2πσξ

⎧ ⎫
⎪
⎪ T
2⎪
⎪
⎪ ((ξ − μ ) − φ (ξ − μ )) ⎪  
⎨ t 2 2 t −1 2 ⎬ 1 (ξ1 − μξ1 )2
t =2
× exp −  exp −
⎪
⎪ 2σξ2 ⎪
⎪ 2 2σξ2
⎪
⎩ ⎪
⎭ 2πσξ1

× p ( v b ) p ( v ξ ) p ( v b1 ) p ( v ξ1 ) p (φ1 ) p (φ2 ) p (μ1 ) p (μ2 ) p (μb1 ) p (μξ1 ),

where p (. . . ) represents the prior distribution for a given parameter. The full conditional distributions are provided in the
Appendix.
Note that, for bt and ξt , t = 1, . . . , T , the full conditional distributions contain an intractable normalizing constant. Thus,
we use the DMH to obtain posterior samples from the target distribution. The rest of the parameters can be sampled by a
standard Gibbs step. Overall, we implement a DMH within Gibbs sampling algorithm for the STAI process with autoregressive
structure and the algorithm is given by

Algorithm 2 Double Metropolis-Hastings within Gibbs sampling algorithm.

(k) (k)
Given b(k) = bt , ξ (k) = ξt , t = 1, . . . , T at the k-th iteration.
prop (k) 2(k) prop (k) 2(k)
Step 1. Propose φ1 ∼ π (·|b(k) , μ1 , σb) and φ2 ∼ π (·|ξ (k) , μ2 , σξ ).
Step 2. Propose μ1prop ∼ π (·|φ1prop , b(k) , σb2(k) ) and μ2prop ∼ π (·|φ2prop , ξ (k) , σξ2(k) ).
2( prop ) prop prop 2( prop ) prop prop
Step 3. Propose σb ∼ π (·|φ1 , μ1 ) and σξ ∼ π (·|φ2 , μ2 ).
prop (k)
Step 4. Propose b prop
∼ π (·|b ) and ξ
(k)
∼ π (·|ξ ).
Step 5. Generate auxiliary patterns X
using the point pattern generating algorithm.
Step 6. Accept the proposed parameters at t with probability r D M H .
(k+1) (k) (k+1) (k) (k+1) (k) (k+1) (k) (k+1)
else reject and set bt = bt , ξt = ξt . If all the proposed bt and ξt are rejected for t = 1, . . . , T , set φ1 = φ1 , φ2 = φ2 , μ1 =
μ(1k) , μ(2k+1) = μ(2k) , σb2(k+1) = σb2(k) and σξ2(k+1) = σξ2(k) .

5
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

where the form of the acceptance ratio is

π (btprop |φ1(k) , σb2(k) )π (ξtprop |φ2(k) , σξ2(k) )q(bt(k) |btprop )q(ξt(k) |ξtprop )
rDM H =
π (bt(k) |φ1(k) , σb2(k) )π (ξt(k) |φ2(k) , σξ2(k) )q(btprop |bt(k) )q(ξtprop |ξt(k) )
f (Xt
|bt , ξt ) f (Xt |bt
(k) (k) prop prop
, ξt )
× .
f (Xt
|bt
prop prop (k) (k)
, ξt ) f (Xt |bt , ξt )

4. Sensitivity analysis

We provide a sensitivity analysis for several parameters of the STAI to show the effects of different parameter values on
the configuration of the point patterns. This study can suggest reasonable parameter spaces for each parameter that can
then be used when fitting the models to real data where we do not know the truth.

4.1. Sensitivity analysis of  g

In the main effect, the function g (x|Xt −1 ) is a normal kernel centered at the point in the previous point pattern that is
closest to location x. The covariance matrix  g controls the smoothness of the kernel. By assuming a new event is equally
likely to be observed from all the directions of a previous event, we use a diagonal matrix for  g . Figure 1 (Supplementary
Material) shows the initial point pattern from a non-homogeneous Poisson point process, the realizations from the STAI at
t = 1 with  g = 0.01I 2 , and with  g = 0.05I 2 (I 2 is the identity matrix), respectively, given all the other parameters are
fixed. From this figure, we see that the pattern is more spread out as the diagonal elements for  g increase, which implies
that the new event can be generated further away from a previous event. Consequently, a small value should be used for
the diagonal elements in order to preserve the non-homogeneous structure over time.

4.2. Sensitivity analysis of ξ

Next, we present realizations of the STAI process with different interaction levels. To do this, we generate realizations
from the processes with ξ = log(0.01), log(1), log(10), and log(50) for strong repulsion, independence, moderate clustering,
and strong clustering, respectively. The interaction distance is fixed at 0.05 for all of these processes. The matrix  g for the
main effect function is 0.01I 2 . The realizations from the corresponding STAI process at t = 1 are shown in Figure 2 (Supple-
mentary Material). For these processes, smaller values of b are used for larger y to ensure that the generated realizations
have approximately 200 points. In this simulation, the values of b are 7, 3, 1, and -0.35, respectively. As we can see, the
realizations from the processes with larger ξ are more clustered than the Poisson process with mixture intensity and the
realization from the process with ξ = log(0.01) shows significant regularity. The pattern in Figure 2b (Supplementary Mate-
rial) can be considered as a realization from a non-homogeneous Poisson process with high intensity in the neighborhood
area of the previous pattern. Here, we see that the point patterns at the current time period show completely different
structure, even though they are generated based on the same previous pattern.

4.3. Sensitivity analysis of b

To check how the parameter b affects the number of points while the other parameters are fixed, we also conducted
a sensitivity analysis. We chose b = log(0.6), log(0.7), log(0.8), and log(0.9). The other parameters are held fixed at  g =
0.01I 2 and ξ = log(50). Figure 3 (Supplementary Material) shows realizations from these processes at t = 1, which indicates
that the number of points is very sensitive to the value of b. In particular, given the current values of  g and y, the number
of points will drop drastically when b is smaller than log(0.6). Overall, Figure 4 (Supplementary Material) shows that the
number of points increases as the value of b increases from log(0.1) to log(5).

4.4. Analysis of interactions between b, ξ , and rh

In this section, we conduct a robustness study in order to illustrate the effect that different values of the model parame-
ters b, ξ , and rh , have on the generated point patterns. In particular, we consider the following parameter values to simulate
point patterns from the process: b = log(1), log(2), log(4), log(8), ξ = log(0.5), log(1), log(10), log(20), and rh = 0.01, 0.05,
0.1, 0.5. The plots of the realizations are included in the Supplementary Material (Figures 5–8), with each figure corre-
sponding to a different value of the interaction parameter. For each row of a figure, the value of parameter b increases and
we notice that the number of points for each point pattern increases accordingly. For each column, the point patterns are
more clustered as the value of ξ increases. The different combinations of values of b and ξ show that both parameters can
affect the number of points and can generate similar point patterns from different parameter values in a finite window.
This is an indication of the well documented identifiability issues that can occur with such point process (see Chen et al.,
2020).

6
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Table 1
Parameter estimation results for 50 realizations from the STAI process with 1 time
period based on DMH.

Parameter True value Average mean Average median Average SD

b -0.2231 -0.2357 -0.2247 0.3818
ξ 3.9120 3.9316 3.9212 0.4278

Moreover, by comparing the four figures across different interaction scales, we can see that larger rh values exemplify
the effect that the parameter ξ has on the point pattern. That is, for large rh and ξ = log(0.5), we can see that there are
less points and the points are more spread out in the point pattern. In addition, the different combinations of rh and ξ
can generate similar point patterns which can cause identifiability issues. For example, small rh or ξ close to 0 can both
generate point patterns with weak interactions between points. A common approach to avoid this issue is to estimate the
interaction radius rh separately and include it as a fixed parameter in the model, where the other parameters are assumed
unknown and require estimation.

5. Simulation study

This simulation study demonstrates the estimation procedure and results for different model structures. The STAI process
with time indexed parameters for one time period is shown in Section 5.1. The STAI process with time indexed parameters
for five time periods is shown in Section 5.2, and the STAI process with autoregressive parameter structures is included in
Section 5.3.

5.1. Time indexed STAI process with T = 1

We first simulate 50 realizations from a STAI process with only one time period and evaluate the average point estimate
and average standard deviation. The true parameter values are b = log(0.8) = −0.2231, ξ = log(50) = 3.9120, rh = 0.05,
and  g = 0.01I 2 . All 50 realizations are generated based on the same initial point pattern X0 , which is generated from a
non-homogeneous Poisson point process and is shown in Figure 1a (Supplementary Material). For Gibbs point processes,
letting the interaction range be dynamic will cause identifiability issues. For instance, different interaction parameters ξ
and interaction ranges rh can generate similar point patterns. In order to simplify the process, we assume that rh and  g
are fixed and known. For each realization, 10,000 posterior realizations are generated using the DMH sampler with 2000
iterations discarded for burn-in. The average posterior means, medians, and standard deviations are provided in Table 1. The
results show that the average mean and median are good estimators of the parameters for the STAI process, although the
estimate of the parameters based on a single realization may deviate from the true parameter value. Notice that the average
of the standard deviations of b is large, which indicates that the estimate of b might be affected by the identifiability issue
inherited in Gibbs point process, i.e., different parameter values can generate similar point patterns. In this case, a given
point pattern can be generated from processes with several different values of b. The similarity between point patterns can
be evaluated based on several discrepancy measures, such as the integrated difference of K -functions or J -functions (see
Chen et al., 2020). Also, we show that our approach can recover the parameter values of the current point pattern even if
the previous pattern was generated from a different process. Next, we present parameter estimation for the time indexed
STAI with more than one time period.

5.2. Time indexed STAI process with T = 5

In this section, we show the parameter estimation of 50 realizations from a STAI process of 5 time periods under the
time indexed parameter structure. Similar to Section 5.1, we use the same initial point pattern for all 50 realizations and
use the DMH algorithm to generate 10,000 posterior realizations, discarding the first 2000 realizations for burn-in. The
interaction distance rh = 0.05 and  g = 0.01I 2 are fixed for each time period. Table 2 includes the average and standard
deviation for the posterior mean, median, and standard deviation of b and ξ at each time period. The results indicate that
the DMH can provide good estimates of the parameters for all time periods. However, the average standard deviations of bt ,
t = 1, . . . , 5, still indicate possible identifiability issues.

5.3. Evolution STAI process

For the evolution STAI process, we implement an AR(1) prior as described in Section 3.1. We first use 20 time periods as
our observed temporal window. The parameters bt and ξt , t = 1, . . . , 20, are considered latent variables and are generated
by the stationary AR(1) model, where b1 = 0.7 and ξ1 = 2.5 are fixed at t = 1 for this simulation study. The point pattern
in Figure 1a (Supplementary Material) is used as the initial point pattern for all the realizations of this process. In this
simulation, we generate 10 realizations from the AR(1) model with parameters φ1 = 0.25, φ2 = 0.5, μ1 = 0.5, μ2 = 3,
σb2 = 0.05, and σξ2 = 0.1. The hyperparameters are ub = 0.05, v b = 0.05, u ξ = 0.05, v ξ = 0.05, ub1 = 0.05, v b1 = 0.05,

7
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Table 2
The parameter estimation results for 50 realizations from the STAI process with 5 time periods.

Parameter True value Average (SD) mean Average (SD) median Average (SD) SD
b1 log(0.8) = −0.2231 -0.2234(0.38) -0.2142(0.37) 0.3917(0.03)
ξ1 log(50) = 3.912 3.9174(0.42) 3.9083(0.42) 0.4401(0.03)
b2 log(1.2) = 0.1823 0.2141(0.38) 0.2144(0.37) 0.3705(0.04)
ξ2 log(30) = 3.4012 3.3663(0.41) 3.3638(0.41) 0.4147(0.04)
b3 log(2) = 0.6931 0.7293(0.51) 0.7242(0.51) 0.4884(0.06)
ξ3 log(20) = 2.9957 2.9655(0.54) 2.9709(0.54) 0.5140(0.06)
b4 log(0.9) = −0.1054 -0.1255(0.39) -0.1196(0.39) 0.4101(0.06)
ξ4 log(40) = 3.6889 3.7093(0.44) 3.7036(0.44) 0.4485(0.06)
b5 log(3) = 1.0986 1.2675(0.39) 1.2574(0.38) 0.4202(0.06)
ξ5 log(10) = 2.3026 2.1147(0.43) 2.1237(0.42) 0.4608(0.06)

Table 3
The average estimates of the model parameters for 10 realizations from the STAI process with 20 time period
based on DMH.

Parameter True value Average (SD) mean Average (SD) median Average (SD) SD
φ1 0.25 0.2214(0.10) 0.2221(0.10) 0.1122(0.02)
φ2 0.5 0.2389(0.15) 0.2470(0.15) 0.2568(0.03)
μ1 0.5 0.5393(0.14) 0.5317(0.13) 0.1570(0.04)
μ2 3 2.9161(0.18) 2.9434(0.16) 0.4026(0.20)
σb2 0.05 0.0950(0.03) 0.0808(0.03) 0.0593(0.02)
σξ2 0.1 0.1329(0.05) 0.1137(0.05) 0.0850(0.04)

Table 4
The parameter estimation results for patterns from the STAI process with 100 time period
based on DMH.

Parameter True value Mean Median 95% CI

φ1 0.25 0.1465 0.1497 (-0.0998, 0.3702)
φ2 0.5 0.3558 0.3596 (0.1824, 0.5049)
μ1 0.5 0.4590 0.4578 (0.3591, 0.5557)
μ2 3 3.0820 3.0806 (2.9472, 3.2267)
σb2 0.05 0.0465 0.0444 (0.0230, 0.0802)
σξ2 0.1 0.1095 0.1078 (0.0665, 0.1612)

u ξ1 = 0.05, v ξ1 = 0.05, τ1 = 1, τ2 = 5 τb1 = 1, and τξ1 = 5. In Table 3, we show the average and standard deviation of the
posterior mean, median, and standard deviation of the parameters in the AR(1) model mentioned above. As we can see
the DMH within Gibbs sampler can recover the true values of the parameters except for φ2 and σb2 , due to the lack of
information for the temporal process. In order to improve the estimate of the autoregressive coefficient, we extend the time
period, which increases the sample size for the time series model.
We increase the number of time periods T to 100 and repeat the process introduced above. Since calculating the infor-
mative prior for the autoregressive coefficient is computationally inefficient for large T , we use a uniform prior for both φ1
and φ2 . We use the same parameter values for the AR(1) process and same initial point pattern as previously mentioned in
this section. The range of the number of points over 100 time periods is between 24 to 1340. The estimation results for one
realization of the STAI process are shown in Table 4 and the generated latent variable along with the posterior mean and
95% point-wise credible interval is shown in Fig. 1. The point estimates indicate that the autoregressive coefficients are still
underestimated, although the estimates of σb2 and σξ2 are significantly improved. Since the observed point patterns are used
for each time point to generate the auxiliary point pattern in the DHM, we do not expect that the estimation accuracy of
bt and ξt will decrease when T increases. However, estimating the STAI process with large T may require significant com-
putational resources, such as computation time. From Fig. 1, we see that the estimates of bt and ξt over 100 time periods
are performing well. The estimates also indicate that the proposed method can provide reasonable results for point patterns
with different number of points.
In order to assess the goodness-of-fit of the proposed model, we calculate the Bayesian p-value for the point pat-
tern at each time period and the average p-value over 100 time periods is 0.55. The average p-value indicates that the
proposed model is appropriate to describe the observed point pattern. To calculate the Bayesian p-value, we use the inte-
grated difference between estimates of Ripley’s K -function for the auxiliary point pattern and the observed point pattern
as the discrepancy measure and 8000 posterior realizations are used to generate the auxiliary point patterns. The current
goodness-of-fit assessment focuses on the comparison at each time period and provides limited information. A more ef-

8
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 1. True value, posterior mean and 95% CI of the latent variable bt and ξt , t = 1, . . . , 100. True value: black solid line; Posterior mean: blue solid line;
95% CI: red dashed lines. (For interpretation of the colors in the figure(s), the reader is referred to the web version of this article.)

fective goodness-of-fit test based on evaluating similarities between spatio-temporal point processes is subject to future
study.

5.4. Convergence diagnostics

To perform convergence diagnostics, we randomly selected several posterior sample chains from the previous sections
and calculated the Monte Carlo standard error (MCSE) proposed by Flegal et al. (2008). Compared with the corresponding
posterior means, the MCSE for the selected chains are sufficiently small, i.e., all MCSE are less than 0.1. Thus, the comparison
results do not suggest any lack of convergence in the sample chains.

5.5. STAI process forecasting

We also evaluate the forecasting efficiency of the proposed STAI process by comparing the true point pattern and sim-
ulated point patterns of future times. Based on the true evolution process over 100 time periods that was considered in
Section 5.3, we generate the true parameters b101 = 0.421 and ξ101 = 2.842 and the corresponding point pattern at t = 101.
Then, we calculate the predicted parameters b̃101 and ξ̃101 using the posterior mean estimates of the parameters in the
AR(1) processes (see Table 4). Finally, 100 corresponding forecast realizations are generated from the STAI process using
the predicted parameter values. In order to compare the observed (true) point pattern with the forecasts, we calculate the
approximated Papangelou conditional intensity surface for the observed point pattern in Fig. 2a and the average Papangelou
surface for the forecast point patterns in Fig. 2b. The surface of the Papangelou conditional intensity can uniquely determine
a point process and it can be interpreted as the intensity surface for the next event given the current events. However, the
forecast point patterns are usually generated within a certain area that are affected by the previous point pattern and are
highly different than each other. The Papangelou surfaces, calculated based on forecast realizations, will also show signifi-
cant variability. Thus, the averaged Papangelou surface will be smoother and have a significantly larger high-intensity area
than individual Papangelou surfaces.

9
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 2. a: The approximated Papangelou conditional intensity surface for the observed point pattern. b: The average Papangelou surface based on 100
forecast point patterns. The white dots represent the observed point pattern at t = 101.

Fig. 3. a: The nonparametric estimation of the intensity surface for the observed pattern based on the Epanechnikov kernel with bandwidth 0.1. b: The
average nonparametric estimate of the intensity surfaces based on the Epanechnikov kernel with bandwidth 0.1 for 100 forecast point patterns. The white
dots represent the observed point pattern at t = 101.

We also compare the nonparametric estimate of the intensity surface for the observed pattern based on the Epanechnikov
kernel with bandwidth 0.1, shown in Fig. 3a, and the average of the same nonparametric estimate surface for the forecast
patterns, shown in Fig. 3b. Notice that this nonparametric estimate of the surface does not contain information about the
specified model. It is only used to evaluate the similarity between point patterns. The average Papangelou surface of the
predicted point patterns exhibits some deviation from the surface of the observed point pattern. However, most of the
observed points are covered by the high intensity areas in the average Papangelou surface. This result indicates that, on
average, the forecast point patterns can provide a reasonable prediction of the true point pattern.
The nonparametric estimates of the intensity surfaces show similar features as the Papangelou surfaces. The average
surface of the predicted point patterns shows the same hotspots on the bottom left and top right corners, along with some
extra high intensity spots at the other corners of the surface. In addition, the average surface is smoother, since it is the
average of 100 predicted surfaces. Overall, the comparisons based on the Papangelou surfaces and intensity surfaces show
that the proposed model can provide a reasonable forecast for future events from this the process.
For the uncertainty measure of the predictive surfaces, we present the 2.5 and 97.5 pixel-wise percentile surfaces of the
Papangelou intensity and the nonparametric estimate intensity for the forecast realizations in Figs. 4 and 5. The 2.5 and
97.5 pixel-wise percentile surfaces of the Papangelou show that the points of the observed pattern are located in the high
intensity area, which is similar to the average Papangelou surface shown in Fig. 2b. This result confirms that the posterior
predicted realizations can provide reasonable forecasts of the future events. However, due to the dynamic mechanism of
this process, the variation of the predictive surfaces is large.
Based on our simulations, we observed a high variability of this point process in terms of the number and locations of
points and strong correlation between the previous point pattern and the current point pattern. Owing to these features,

10
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 4. For the simulated example with 100 time periods: a: 2.5 pixel-wise percentile of the Papangelou surface based on 100 forecast point patterns; b: 97.5
pixel-wise percentile of the Papangelou surface based on 100 forecast point patterns. The white dots represent the observed point pattern at t = 101.

Fig. 5. For the simulated example with 100 time periods: a: 2.5 pixel-wise percentile of the nonparametric estimate surface based on 100 forecast point
patterns; b: 97.5 pixel-wise percentile of the nonparametric estimate surface based on 100 forecast point patterns. The white dots represent the observed
point pattern at t = 101.

the predicted point patterns rapidly deviate from the true point patterns as the predicted time period increases. Thus, it is
very difficult to make multi-step ahead predictions and, consequently, we recommend making predictions for only one time
point ahead.

6. United States natural caused wildfire data

We apply the proposed model to natural caused wildfire data in the western United States from 2002 to 2019 available
the GEOMAC wildland fire support (https://www.geomac.gov/viewer/viewer.shtml). The data contain the reported locations
and time stamps of the wildfires. We aggregate the events by years and rescale the data to the unit window. The observed
point patterns from 2002 to 2019 are illustrated in the supplementary material (Figures 9-12). Additionally, we treat the
wildfire locations in 2002 as the initial point pattern X0 . Next, we fit the wildfire data from 2003 to 2018 to a STAI process
over 16 time periods. The data in 2019, shown in Fig. 6, is used as the “truth” to compare with the forecasts from the fitted
model. In Fig. 6, we show the polygon that is used as a plausible window for this point process, which approximates the
shape of west coast.
We first implement the Albers equal area projection with standard parallels 30◦ N and 50◦ N to the wildfire data.
Based on a visual inspection, the point patterns over 17 year show similar shapes. Thus, the parameter  g in the main
effect function is set to be 0.005I 2 in order to preserve the similar features over time. The interaction distance of the
area-interaction process is set to rh = 0.025 based on sensitivity analysis. Similar to the simulation study, we also assume
that the parameters of the area-interaction process are generated by some underlying AR(1) processes. Since the data are
aggregated by year, we do not expect significant direction effects. However, an anisotropic kernel can be used for data ag-
gregated based on shorter time intervals in order to incorporate the dominant directions effect. We implement the DMH
within Gibbs sampler algorithm to obtain the 10,000 posterior samples of the model parameters with 2000 iterations dis-
carded for burn-in. The estimation results of the AR(1) processes are shown in Table 5. The posterior means and 95%

11
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 6. The observed natural caused wildfire locations in 2019. The red lines are the boundaries of the polygon that is used as the observation window,
which covers the west United States.

Table 5
The parameter estimation results, using the Albers equal area projection, for
the United States natural caused wildfire data from 2003 to 2018 based on
DMH with  = 0.005I 2 and rh = 0.025.

Parameter Mean Median 95% CI

φ1 0.2284 0.2356 (-0.4061, 0.8924)
φ2 0.3358 0.3561 (-0.3574, 0.9995)
μ1 0.3438 0.3586 (0.0287, 0.5794)
μ2 3.6755 3.9098 (-0.3503, 4.2629)
σb2 0.0338 0.0281 (0.0101, 0.0932)
σ y2 0.0721 0.0569 (0.0153, 0.2257)

point-wise credible intervals of the parameters for the area-interaction process are shown in Fig. 7. Notice that the pos-
terior means of ξt are all larger than 0, which indicates that the locations of wildfire are significantly clustered for every
year.
Using the estimated parameters, we generate 100 predicted parameters b̃2019 and ỹ 2019 for the point process in 2019.
Then, we generate 100 forecast point patterns based on the predicted parameters and calculate the average of the Pa-
pangelou surfaces and nonparametric estimate intensity surfaces for these realizations. These surfaces were introduced in
Section 5.5. The average surface of the Papangelou conditional intensity along with the observed point pattern in 2019 is
shown in Fig. 8a and the average nonparametric estimates of the intensity surfaces with the true pattern are shown in
Fig. 9a. The Papangelou and intensity surfaces show that the true point pattern can be covered by high intensity area based
on the fitted model, which indicates the proposed model can provide reasonable forecasts for possible wildfire locations.
More importantly, this information can help wildland fire management officers to efficiently allocate resources to monitor
the natural environment in order to prevent wildfires.
Moreover, the 2.5 and 97.5 pixel-wise percentile surfaces of the Papangelou and nonparametric estimate intensity for
the forecast point patterns are shown in the Figs. 8b, 8c, 9b, and 9c. The 2.5 and 97.5 pixel-wise percentile surfaces of the
Papangelou show that most of the wildfires in 2019 are located in the high intensity area, which confirm that the posterior
predicted realizations can provide reasonable forecasts of the future wildfire locations. Due to the dynamic mechanism of
this model, the variation of the forecasts is large. More importantly, this information can help wildland fire management
officers to efficiently allocate resources to monitor the natural environment in order to prevent wildfires.

7. Concluding remarks

In this paper, we propose a flexible spatio-temporal point process model that can be used to describe spatial and tem-
poral interactions between points. Specifically, we use a Markovian structure to model the dependence of point patterns

12
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 7. Posterior mean and 95% CI of the latent variables bt and ξt , t = 1, . . . , 16 for the United States natural caused wildfire data from 2003 to 2018 using
the Albers equal area projection. Posterior mean: blue solid line; 95% CI: red dashed lines.

over consecutive time periods, which assumes that the current main effect function is only affected by the point pat-
tern in the previous time period. This structure provides the opportunity to predict the point pattern in the following
time period without the use of covariate information. Unlike continuous time processes, the prediction from our model
does not require generating point patterns across the whole temporal window, and is, thus, more computationally effi-
cient.
For the parameter structure, a time indexed parameter model and a hierarchical model are evaluated in our study. Based
on simulations, we show that the DMH sampler can provide good estimates for the time indexed parameters and the latent
parameters in the hierarchical model. More importantly, we also demonstrate that the parameters of the autoregressive
process model can be recovered by the Bayesian estimation method. These results indicate that it is possible to discover the
underlying structure of a spatio-temporal point process and predict future point patterns by evolving the latent parameters.
However, our results show that it is difficult to recover the autoregressive structures due to the deviations between the true
latent parameter values and their estimates, which is introduced by the sampling algorithm for the parameters of the data
model. Based on the evolution process, we also provide an example of predicting a future point pattern (via a simulation
study). By comparing the surfaces of Papangelou conditional intensity and nonparametric estimate of the intensity for the
true point pattern and predicted patterns, we evaluate the predictions made by the proposed model. The results indicate
that our model can provide reasonable predictions for future events.
For the interaction structure, we use the area-interaction function to describe the interaction level of the points at the
current time. Since the area-interaction process is suitable for repulsive, independent, and clustered events, our model can
be used as a general approach for modeling any point pattern under consideration. Additionally, the interaction function can
be easily replaced by other functions; e.g., if there is strong information to indicate a different interaction function, such as
the Strauss process, say, for forestry data.
The United States natural caused wildfire data show strong dependence between the point patterns at two consecutive
years, which is suitable for the proposed model. Based on the estimated parameters, we provide forecasts for future events
and compared these forecasts with the observed data. The results indicate that the forecasts provide reasonable estimates

13
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 8. The STAI model forecast for wildfire locations in 2019 using the Albers equal area projection with  g = 0.005I 2 and rh = 0.025. a: The average
Papangelou surface for 100 forecast point patterns. b: The 2.5 pixel-wise percentile of the Papangelou surface with edge correction for 100 forecast point
patterns; c: The 97.5 pixel-wise percentile of the Papangelou surface with edge correction for 100 forecast point patterns. The white dots represent the
observed natural caused wildfire locations in 2019.

for future events. Thus, we believe the proposed model can be very useful in terms of describing the process and making
forecasts. Moreover, we believe that the forecast intensity and Papangelou conditional intensity surfaces can provide infor-
mation about high risk areas for wildfires. More importantly, the strategies of managing wildfires can be improved based
on the obtained information.
In principle, a higher order autoregressive model could be implemented; e.g., an AR(2) (for the United States natural
caused wildfire data). However, in practice, significant estimation challenges arise. The main issue manifests in trying to
ensure that the estimated parameters result in a stationary process. For the AR(1) model this is straightforward, as there is
only one constraint on the model coefficients that needs to be satisfied. In contrast, with higher-order autoregressive models
the number of constraints increases. Consequently, this either leads to a less efficient sampling algorithm or explosive
behavior (the case where stationarity is not met). This is a subject of future research.
Overall, the proposed model provides a general framework for evaluating spatio-temporal point processes. In future work,
a more sophisticated parameter evolution structure will be investigated in order to capture more complex features of the
point processes. Moreover, extra covariate information, such as temperature and precipitation level can be included in the
evolution model to improve the prediction. Additionally, it is possible to organize the Bayesian estimation algorithm for
each time period in a parallel structure to improve the computational efficiency.

Acknowledgement

This work was done by Jiaxun Chen as part of his Ph.D dissertation prior to joining Eli Lilly and Company. This research
was partially supported by the U.S. National Science Foundation (NSF) under NSF SES-1853096 and through the Air Force
Research Laboratory (AFRL) Contract No. 19C0067.

14
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Fig. 9. The STAI model forecast for wildfire locations in 2019 using the Albers equal area projection with  g = 0.005I 2 and rh = 0.025. a: The average
nonparametric estimate of the intensity surfaces based on the Epanechnikov kernel with bandwidth 0.075 for 100 forecast point patterns; b: The 2.5 pixel-
wise percentile of the nonparametric estimate of the intensity surface for 100 forecast point patterns; c: The 97.5 pixel-wise percentile of the nonparametric
estimate of the intensity surface for 100 forecast point patterns. The white dots represent the observed natural caused wildfire locations in 2019.

Appendix

The full conditional distributions for the evolution STAI process are shown as follows
  
1 σb + φ1 σb1 n1 σb2 σb2 + σb2 μb1 + φ12 μ1 σb2 + σb2 φ1 (b2 − μ1 )
2 2 2
1
π (b1 |·) ∝ exp − b1 − 1 1 1
.
C (b1 , ξ1 ) 2 σb21 σb2 σb2 + φ12 σb21
  
1 1 1 + φ12 φ1 (bt +1 + bt −1 − 2μ1 ) + σb2nt
π (bt |·) ∝ exp − bt − μ1 − , t = 2, . . . , T .
C (bt , ξt ) 2 σb2 φ12 + 1

1 1 σ y + φ 2 σξ 1
2 2 2
σb2 μb1 + φ12 μ1 σb21 + σb21 φ1 (b2 − μ1 )
π (ξ1 |·) ∝ exp − ξ1 −
C (b1 , ξ1 ) 2 σξ21 σ y2 σb2 + φ12 σb21
!
{n1 − m{∪ni =1 1 B (xi ,1 , rh )}/(π rh2 )}σb21 σb2 +
− .
σb2 + φ12 σb21

1 1 1 + φ22
π (ξt |·) ∝ exp −
C (bt , ξt ) 2 σ y2
!
φ2 (ξt +1 + ξt −1 − 2μ2 ) + σ y2 {nt − m{∪ni =t 1 B (xi ,t , rh )}/(π rh2 )}
× ξt − μ2 − ,
φ22 + 1
t = 2, . . . , T .

15
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

⎛ ⎞

T
2
⎜ σφ1 (bt −1 − μ1 )(bt − μ1 ) + φ̂1 σb2 2 2 ⎟
⎜ t =2
σ σ
b φ1 ⎟
φ1 |· ∼ N (−1,1) ⎜ ⎟.
⎝ 
T 
T ⎠
σφ21 (bt −1 − μ1 )2 + σb2 σφ21 (bt −1 − μ1 )2 + σb 2
t =2 t =2
⎛ ⎞

T
⎜ (ξt −1 − μ2 )(ξt − μ2 ) ⎟
⎜ t =2 σ y2 ⎟
φ2 |· ∼ N (−1,1) ⎜ , ⎟.
⎝ 
T 
T ⎠
(ξt −1 − μ2 )2 (ξt −1 − μ2 )2
t =2 t =2
 
τb1 b1 σb2 τb1
μb1 |· ∼ N , .
σb21 + τb1 σb21 + τb1
   
τ1 (1 − φ1 )
T
σb2 τ1
μ1 |· ∼ N (b t − φ 1 b t − 1 ) , .
( T − 1)(1 − φ1 )2 τ1 + σb2 t =2 ( T − 1)(1 − φ1 )2 τ1 + σb2
 
τξ1 ξ1 σ y2 τξ1
μξ1 |· ∼ N , .
σξ21 + τξ1 σξ21 + τξ1
   
τ2 (1 − φ2 )
T
σ y2 τ2
μ2 |· ∼ N (ξt − φ 2 ξt − 1 ) , .
( T − 1)(1 − φ2 )2 τ2 + σ y2 t =2 ( T − 1)(1 − φ2 )2 τ2 + σ y2
!
1 1
σb21 |· ∼ I G + ub1 , (b1 − μb1 )2 + v b1 .
2 2
 
1

T
2 T −1 2
σ b |· ∼I G + ub . ((bt − μ1 ) − φ1 (bt −1 − μ1 )) + v b .
2 2
t =2
!
1 1
σξ21 |· ∼ I G + u ξ1 , (ξ1 − μξ1 )2 + v ξ1 .
2 2
 
1

T
2 T −1 2
σ y |· ∼I G + u y, ((ξt − μ2 ) − φ2 (ξt −1 − μ2 )) + v y .
2 2
t =2

Appendix A. Supplementary material

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.csda.2021.107349.

References

Baddeley, A., Rubak, E., Turner, R., 2015. Spatial Point Patterns: Methodology and Applications with R. CRC Press.
Baddeley, A.J., Van Lieshout, M., 1995. Area-interaction point processes. Ann. Inst. Stat. Math. 47, 601–619.
Barndorff-Nielsen, O.E., Kendall, W.S., Van Lieshout, M., 1999. Stochastic Geometry, Likelihood and Computation. Chapman and Hall.
Brix, A., Diggle, P.J., 2001. Spatiotemporal prediction for log-Gaussian Cox processes. J. R. Stat. Soc., Ser. B, Stat. Methodol. 63, 823–841.
Chen, J., Micheas, A.C., Holan, S.H., 2020. A comparative study of approximate Bayesian computation methods for Gibbs point processes. J. Stat. Appl. 18,
223–248.
Chiu, S.N., Stoyan, D., Kendall, W.S., Mecke, J., 2013. Stochastic Geometry and Its Applications. John Wiley & Sons.
Cressie, N., 1993. Statistics for Spatial Data, 2nd edition. John Wiley & Sons.
Daley, D.J., Vere-Jones, D., 2003. An Introduction to the Theory of Point Processes, Volume 1: Elementary Theory and Methods. Springer Verlag, New York
Berlin Heidelberg.
Daley, D.J., Vere-Jones, D., 2007. An Introduction to the Theory of Point Processes: Volume II: General Theory and Structure. Springer Science & Business
Media.
Dereudre, D., 2019. Introduction to the theory of Gibbs point processes. In: Stochastic Geometry. Springer, pp. 181–229.
Diggle, P.J., 2013. Statistical Analysis of Spatial and Spatio-Temporal Point Patterns. CRC Press.
Flegal, J.M., Haran, M., Jones, G.L., 2008. Markov chain Monte Carlo: can we trust the third significant figure? Stat. Sci., 250–260.
Gabriel, E., Opitz, T., Bonneu, F., 2017. Detecting and modeling multi-scale space-time structures: the case of wildfire occurrences. J. Soc. Fr. Stat. 158,
86–105.
Gelfand, A.E., Diggle, P., Guttorp, P., Fuentes, M., 2010. Handbook of Spatial Statistics. CRC Press.
Geyer, C., 1998. Likelihood inference for spatial point. In: Stochastic Geometry: Likelihood and Computation, vol. 80, p. 79.
Hawkes, A.G., 1971. Spectra of some self-exciting and mutually exciting point processes. Biometrika 58, 83–90.
Iftimi, A., van Lieshout, M.C., Montes, F., 2018. A multi-scale area-interaction model for spatio-temporal point patterns. Spat. Stat. 26, 38–55.
Iftimi, A., Montes, F., Mateu, J., Ayyad, C., 2017. Measuring spatial inhomogeneity at different spatial scales using hybrids of Gibbs point process models.
Stoch. Environ. Res. Risk Assess. 31, 1455–1469.
Illian, J., Penttinen, A., Stoyan, H., Stoyan, D., 2008. Statistical Analysis and Modelling of Spatial Point Patterns, vol. 70. John Wiley & Sons.

16
J. Chen, A.C. Micheas and S.H. Holan Computational Statistics and Data Analysis 167 (2022) 107349

Illian, J.B., Sørbye, S.H., Rue, H., Hendrichsen, D.K., 2012. Using INLA to fit a complex point process model with temporally varying effects-a case study. J.
Environ. Stat. 3.
Karr, A., 1991. Point Processes and Their Statistical Inference, vol. 7. CRC Press.
Kelly, F.P., Ripley, B.D., 1976. A note on Strauss’s model for clustering. Biometrika, 357–360.
Lantuejoul, C., 2001. Geostatistical Simulation: Models and Algorithms, vol. 1139. Springer Science & Business Media.
Lavancier, F., Møller, J., 2016. Modelling aggregation on the large scale and regularity on the small scale in spatial point pattern datasets. Scand. J. Stat. 43,
587–609.
Lawson, A.B., Denison, D.G., 2002. Spatial Cluster Modelling. CRC Press.
Liang, F., 2010. A double Metropolis–Hastings sampler for spatial models with intractable normalizing constants. J. Stat. Comput. Simul. 80, 1007–1022.
Micheas, A.C., 2019. Cox point processes: why one realisation is not enough. Int. Stat. Rev. 87, 306–325.
Møller, J., 2013. Spatial Statistics and Computational Methods, vol. 173. Springer Science & Business Media.
Møller, J., Pettitt, A.N., Reeves, R., Berthelsen, K.K., 2006. An efficient Markov chain Monte Carlo method for distributions with intractable normalising
constants. Biometrika 93, 451–458.
Møller, J., Waagepetersen, R.P., 2003. Statistical Inference and Simulation for Spatial Point Processes. CRC Press.
Murray, I., Ghahramani, Z., MacKay, D., 2012. MCMC for doubly-intractable distributions. arXiv preprint. arXiv:1206.6848.
Park, J., Haran, M., 2018. Bayesian inference in the presence of intractable normalizing functions. J. Am. Stat. Assoc. 113, 1372–1390.
Pei, T., Gao, J., Ma, T., Zhou, C., 2012. Multi-scale decomposition of point process data. GeoInformatica 16, 625–652.
Picard, N., Bar-Hen, A., Mortier, F., Chadœuf, J., 2009. The multi-scale marked area-interaction point process: a model for the spatial pattern of trees. Scand.
J. Stat. 36, 23–41.
Raeisi, M., Bonneu, F., Gabriel, E., 2019. A spatio-temporal multi-scale model for Geyer saturation point process: application to forest fire occurrences. arXiv
preprint. arXiv:1911.06999.
Rao, V., Adams, R.P., Dunson, D.D., 2017. Bayesian inference for Matérn repulsive processes. J. R. Stat. Soc., Ser. B, Stat. Methodol. 79, 877–897.
Ripley, B.D., 1987. Stochastic Simulation.
Siino, M., Adelfio, G., Mateu, J., Chiodi, M., D’Alessandro, A., 2017. Spatial pattern analysis using hybrid models: an application to the Hellenic seismicity.
Stoch. Environ. Res. Risk Assess. 31, 1633–1648.
Siino, M., D’Alessandro, A., Adelfio, G., Scudero, S., Chiodi, M., 2018. Multiscale processes to describe the eastern Sicily seismic sequences. Ann. Geophys.
Sørbye, S.H., Illian, J.B., Simpson, D.P., Burslem, D., Rue, H., 2019. Careful prior specification avoids incautious inference for log-Gaussian Cox point processes.
J. R. Stat. Soc., Ser. C, Appl. Stat. 68, 543–564.
Spodarev, E., 2013. Stochastic Geometry, Spatial Statistics and Random Fields: Asymptotic Methods, vol. 2068. Springer.
Strauss, D.J., 1975. A model for clustering. Biometrika 62, 467–475.
Van Lieshout, M., 2000. Markov Point Processes and Their Applications. World Scientific.
Wiegand, T., Gunatilleke, S., Gunatilleke, N., Okuda, T., 2007. Analyzing the spatial structure of a Sri Lankan tree species with multiple scales of clustering.
Ecology 88, 3088–3102.
Zhou, Z., Matteson, D.S., Woodard, D.B., Henderson, S.G., Micheas, A.C., 2015. A spatio-temporal point process model for ambulance demand. J. Am. Stat.
Assoc. 110, 6–15.
Zhuang, J., Ogata, Y., 2006. Properties of the probability distribution associated with the largest event in an earthquake cluster and their implications to
foreshocks. Phys. Rev. E 73, 046134.

17