Multi-dimensional Point Process Models in R

Roger D. Peng
Department of Statistics, University of California, Los Angeles
Los Angeles CA 90095-1554

Abstract
A software package for fitting and assessing multi-dimensional point process models using the R sta-
tistical computing environment is described. Methods of residual analysis based on random thinning are
discussed and implemented. Features of the software are demonstrated using data on wildfire occurrences
in Los Angeles County, California and earthquake occurrences in Northern California.

1 Introduction
This paper introduces an R package for fitting and assessing multi-dimensional point process models. In par-
ticular, the software is designed for conducting likelihood analysis of conditional intensity models. While the
methodology for applying maximum likelihood to point process models is already well-developed, techniques
for assessing the absolute goodness-of-fit of models are still being actively researched. We describe several
methods of point process residual analysis such as random rescaling, thinning, and approximate thinning, and
discuss their advantages and disadvantages in practice. The package implements two random thinning-based
methods.
Section 2 begins with a brief discussion of the specification and fitting of point process conditional inten-
sity models. Section 3 describes several residual analysis techniques and discusses their various properties.
Section 4 describes some existing R packages for analyzing point data and Section 5 describes the ptproc
package. In particular, we show how features of the R language such as lexical scope and the ability to
manipulate language objects are used to build a compact flexible framework for analyzing point process
models. Finally, Section 6 shows two extended examples of how one might use the ptproc package.
The current version (1.4) of ptproc was written and has been tested with R version 1.7.1. The package
is written in pure R code so it should be usable by anyone with access to an R interpreter.

2 Point Process Models

A point process can be thought of as a random measure N specifying the number of points, N (A), in any
compact set A ⊂ S, where S is simply the domain in which the point process resides. The measure is non-
negative integer-valued and is finite on any finite subset of S. Usually, one assumes that the point process is
simple, so that each point is distinct with probability one. For the purposes of this paper, we will only look
at simple point processes in the space-time domain (i.e. S ⊂ R3 ), with one time dimension and two spatial
dimensions.

1
 A simple point process is conveniently specified by its conditional intensity λ(t, x) ((t, x) ∈ S), defined
as the limiting expectation,
1
lim E [N ((t, t + ∆t) × (x, x + ∆x)) | Ht ]
∆t↓0 ∆t∆x
∆x↓0

where Ht represents the internal history of the process up to time t and N ((t, t + ∆t) × (x, x + ∆x)) is
the number of points in a small neighborhood of (t, x). When it exists, the conditional intensity can be
interpreted as the instantaneous rate of occurrence of events at time t and location x.
Since the conditional intensity completely specifies the finite dimensional distributions of the point pro-
cess N (Daley and Vere-Jones, 1988), modelling N typically involves directly modelling λ. Many conditional
intensity models have been developed for specific applications, most significantly from the field of seismol-
ogy (e.g. Vere-Jones, 1970; Ogata, 1999) and also ecology (Rathbun and Cressie, 1994b).

2.1 Model Fitting

Given a parametrized model for the conditional intensity (with parameter vector θ), the unknown parameters
can be estimated by maximizing the log-likelihood function
Z Z
`(θ) = log λ(t, x; θ) N (dt, dx) − λ(t, x; θ) dt dx (1)
S S

It has been shown that under general conditions, the maximum likelihood estimates are consistent and
asymptotically normal (Ogata, 1978; Rathbun and Cressie, 1994a; Rathbun, 1996). When maximizing the
log-likelihood, care must be taken to ensure that λ is positive at all points in S. One solution to this
potential difficulty is to model log λ rather than model λ directly (see e.g. Berman and Turner, 1992).
Another method is to include a penalty in the specification of the log-likelihood which penalizes against
parameter values which produce negative values of the conditional intensity. For example, one could use the
following modified log-likelihood function

`∗ (θ) = `(θ) − P (θ)

where P (θ) is a suitable penalty function. For example, one could use a smooth penalty function as in Ogata
(1983). An alternative would be simply to add a penalty any time the conditional intensity takes a negative
value. Given an appropriate α > 0, let

P (θ) = α 1{λ(t, x; θ) < 0, (t, x) ∈ S} (2)

where 1{A} is the indicator of the event A.

3 Residual Analysis Methods

A common method of evaluating a point process model is to examine likelihood criteria such as the Akaike
Information Criterion or the Bayesian Information Criterion (e.g. Ogata, 1988; Ogata and Tanemura, 1984;
Vere-Jones and Ozaki, 1982). These criteria provide useful numerical comparisons of the global fit of com-
peting models. For example, a common model for comparison is the homogeneous Poisson model. However,

2
these criteria cannot shed light on the absolute goodness-of-fit of a particular model. In particular, they
cannot identify where a model fits poorly and where it fits well.
Residual analysis in other statistical contexts (such as regression analysis) is a powerful tool for locating
defects in the fitted model and for suggesting how the model should be improved. While the same is true
in point process analysis, one must be careful in how one defines the residuals in this context. For point
processes the “residuals” consist of yet another point process, called the residual process. There exist various
ways of constructing a residual process and we discuss some of those methods in this Section.
The common element of residual analysis techniques is the construction of an approximate homogeneous
Poisson process from the data points and an estimated conditional intensity function λ̂. Suppose we observe
a one-dimensional point process t1 , t2 , . . . , tn with conditional intensity λ on an interval [0, T ]. It is well
known that the points
Zti
τi = λ(s) ds (3)
0

for i = 1, . . . , n form a homogeneous Poisson process of rate 1 on the interval [0, n]. This new point process
is called the residual process. If the estimated model λ̂ is close to the true conditional intensity, then
the residual process resulting from replacing λ with λ̂ in (3) should resemble closely a homogeneous Poisson
process of rate 1. Ogata (1988) used this random rescaling method of residual analysis to assess the fit of one-
dimensional point process models for earthquake occurrences. For the multi-dimensional case, Schoenberg
(1999) demonstrated that for a large class of point processes, the domain can be rescaled in such a way that
the resulting process is again homogeneous with rate 1.
When rescaling a multi-dimensional point process one may encounter two practical difficulties:
1. The boundary of the rescaled domain may be uninterpretable (or unreadable). That is, the residual
process may be homogeneous Poisson but the irregularity of the rescaled domain can make the points
difficult to examine. In particular, an irregular boundary can bias various tests for uniformity which
are sensitive to edge effects.
2. Integration of the conditional intensity function is required. In practice, accurate integration of the
conditional intensity in certain dimensions can be computationally intensive.
Both of these problems can be ameliorated by instead constructing a residual process via random thinning.
Suppose that for all (t, x) ∈ S there exists a value m such that

0<m≤ inf λ(t, x). (4)

(t,x)∈S

Then for each i = 1, . . . , n, we delete the data point (ti , xi ) with probability 1 − m/λ(ti , xi ). The undeleted
points form a homogeneous Poisson process of rate m over the original domain S. The residual process
obtained through this method of thinning will be referred to as the ordinary thinned residual process or
ordinary thinned residuals. For details on random thinning, see Lewis and Shedler (1979) and Ogata (1981).
To construct the ordinary thinned residual process, we only need to evaluate the conditional intensity.
Typically, this is a much simpler task than integrating the conditional intensity. Unfortunately, in some
cases where m is very close to zero, the thinning process can result in very few points. However, because of
the randomness involved in the thinning procedure, one can repeat the thinning many times and examine
the various realizations for homogeneity. Another drawback of this method is that there may not exist such
an m.

3
 A third method for constructing a residual process addresses the problem of having too few points in the
thinned process. Each data point (ti , xi ) is assigned a probability pi such that
1
pi ∝ .
λ(ti , xi )

Then, a weighted subsample of size K (< n) is drawn from the vector {(t1 , x1 ), . . . , (tn , xn )} using the weights
p1 , . . . , pn . As long as K is sufficiently small relative to n, the resulting set of points {(t∗1 , x∗1 ), . . . , (t∗K , x∗K )}
should resemble (approximately) a homogeneous Poisson process of rate K/kSk over the original domain.
The process constructed using this method will be referred to as the approximate thinned residual process, or
simply approximate thinned residuals. This procedure can also be used to generate many realizations of the
residual process and each realization will have exactly K points. Approximate thinned residuals were used
in Schoenberg (2003) to assess the space-time Epidemic-Type Aftershock Sequence model of Ogata (1998).
The residual process is useful for model evaluation purposes because the process is homogeneous Poisson
if and only if the model is equal to the true conditional intensity function. Once the residual process has
been constructed, it can be inspected (graphically) for uniformity and homogeneity. In addition, numerous
statistical tests can be applied to test the process for uniformity (see e.g. Diggle, 1983; Ripley, 1979, and
many others). For example, Ripley’s K function can be used to test for spatial clustering and inhibition (Rip-
ley, 1976). In general, any deviation of the residual process from a homogeneous Poisson process can be
interpreted as a deviation of the model from the true conditional intensity.

4 Point Process Packages for R

There are already a few R packages available on CRAN for analyzing point data. Some notable ones include
the spatstat package of Baddeley and Turner and the splancs package of Rowlingson, Diggle, Bivand,
and Petris. These packages are largely concerned with analyzing two-dimensional point patterns. The
spatstat package of Baddeley and Turner provides data structures and functions for fitting a variety of
spatial statistical models via maximum pseudolikelihood. In particular they use a novel approximation to
construct a simple method for maximizing the pseudolikelihood (see Baddeley and Turner, 2000). The
authors’ concentration on purely spatial patterns (and the associated models) allows for the inclusion of
many useful functions and utilities, all related to spatial analysis. In general, the user does not have to write
much code in order to take advantage of the model fitting routines.
The splancs library is another package for two-dimensional point patterns (see Rowlingson and Diggle,
1993, for details). This package contains many useful functions for computing spatial statistics over irregular
boundaries, simulating point patterns, and doing kernel smoothing. The package has less of an emphasis
on model fitting than spatstat. Some code for analyzing space-time data is available in the package, e.g.
functions for computing space-time K-functions (Diggle et al., 1995).
The choice of package to use will obviously depend on the needs of the user. The package discussed here is
intended to be somewhat more general than those concerned with purely spatial point patterns. However, to
achieve this greater generality, much of the burden of writing code has been shifted to the user. In particular,
we discuss in Section 5 how the user must write code for evaluating and integrating the conditional intensity.
The primary goal for this package is to provide a general framework for organizing multi-dimensional point
process data and fitting arbitrary conditional intensity models. In addition, we provide tools for model doing
model assessment and point process residual analysis.

4
5 The ptproc Package
The ptproc source package and updates can be downloaded from the UCLA Department of Statistics Soft-
ware Repository1 . This package is based in part on the point process portion2 of the Statistical Seismology
Library (SSLib) of Harte (1998). In particular our template for the conditional intensity function is sim-
ilar to that of SSLib. SSLib is focused on time-dependent point process modelling (particularly for jump
type processes) and has many pre-written conditional intensity functions and utilities related to earthquake
analysis. Here we try to generalize some of the concepts in SSLib and provide a framework for analyzing a
variety of multi-dimensional point process data.
After installing the ptproc package it can loaded into R in the usual way:
> library(ptproc)
Multi-dimensional Point Process Models in R (version 1.4)
The ptproc package makes use of S3 style classes and methods. There are three basic classes of point
processes. They are “ptprocInit”, “ptprocFit”, and “ptprocSim”. The “ptprocInit” class is used to
setup a point process model before maximum likelihood fitting. It is used to organize the data, conditional
intensity function, and parameter values. The parameters specified in a “ptprocInit” object are used as
initial values for the optimizer when minimizing the negative log-likelihood. Objects of class “ptprocFit”
are returned by the ptproc.fit function and represent the fitted point process model. Currently, the most
interesting method available for this class is the residuals method, which can be used to compute either
ordinary or approximate thinned point process residuals. Finally, the “ptprocSim” class is used to represent
point process models which will be used solely for simulation.

5.1 The “ptprocInit” Object

The package introduces the class “ptprocInit” to represent an initial point process model and data object.
A “ptprocInit” object contains the data points, the conditional intensity function, parameter values, and
other information required for fitting and evaluating the conditional intensity. The ptproc package contains
some built-in conditional intensity functions but in general, the user will want to specify code for evaluating
and integrating the conditional intensity over the study area. Currently, the “ptprocInit” class has print
and logLik methods associated with it as well as various utility functions for accessing elements of the class.
The constructor function ptproc constructs the point process object. Elements which must be passed as
arguments are:
• pts: A matrix of data points. The matrix should be n × p where n is the number of data points and
p is the number of dimensions in the data.
• cond.int: The conditional intensity function. This should be a valid R function.

• params: A numeric vector of parameters for the model. The values specified here will be used as initial
values for optim when maximizing the log-likelihood function.
Other elements of the “ptprocInit” object include:
1 http://ftp.stat.ucla.edu/Software/7
2 http://www.statsresearch.co.nz/software.html

5
 • fixed: A numeric vector equal in length to params. A NA in the ith position of the fixed vector
indicates that the ith parameter is a free parameter. If an element in the fixed vector is non-NA, then
the corresponding parameter in params is fixed. By default, all parameters are set to be free (i.e. every
element of fixed is NA).
• condition: An R expression. See Section 5.2 for details.
• ranges: A matrix containing upper and lower boundaries for each dimension of the data. The ranges
element specifies the domain of the point process. The matrix will be a 2 × p matrix, where p is the
number of dimensions in the data. The (1, j) element gives the lower bound in the jth dimension and
(2, j) element gives the upper bound in the jth dimension. If ranges is not specified, the minima and
maxima of the data points in each dimension are used.
• data: Other information (usually in the form of a list) that may be needed to evaluate or integrate the
conditional intensity. This may include covariate data, point process marks, or preprocessed values.
The default value for data is NULL.
The most important part of the “ptprocInit” object is the conditional intensity function. This is where
the user has to write the most code. The form of the conditional intensity function should adhere to the
following template.
1. The arguments to the conditional intensity function should be

function (params, eval.pts, pts = NULL, data = NULL, TT = NULL)

where params is a vector of parameters for the model, eval.pts is a matrix of points at which we
want to evaluate the conditional intensity, pts is the matrix containing the original data, data can be
any other information that the conditional intensity function may need, and TT is a matrix denoting
the ranges of integration in each dimension. The object passed through the TT argument should be of
the same form as the ranges element of the “ptprocInit” object.
2. It is the responsibility of the user to make sure that the conditional intensity function can be evaluated
at all of the data points (given a valid set of parameters) and that the integral over the entire domain
can be evaluated. For fitting a model, it is not required that the conditional intensity be evaluated at
all points in the domain; just the data points. However, for plotting and general visualization purposes,
evaluation at all points will likely be useful.
3. The body of the conditional intensity function will generally appear as follows:

my.cond.int <- function (params, eval.pts, pts = NULL, data = NULL, TT = NULL) {
a <- params[1]
b <- params[2]
## Assign other parameters

if(is.null(TT)) {
##
## Evaluate the conditional intensity at eval.pts
##

6
 }
else {
##
## Integrate the conditional intensity over the entire domain
##
}
## return a value: either a vector of values of length ‘nrow(eval.pts)’
## when evaluating the conditional intensity, or a single value when
## integrating.
}

See Appendix A for examples.

5.2 Maximizing the Log-Likelihood

The log-likelihood of a given conditional intensity model can be computed using the logLik method. This
method simply computes the quantity in (1) by summing the conditional intensity values at each data point
and evaluating the integral over the entire domain.
The package function ptproc.fit is used for fitting the conditional intensity model via maximum likeli-
hood. It first calls make.optim.logLik to construct the negative log-likelihood function which will be passed
to the optimizer. In general, there is no need for the user to call make.optim.logLik directly. However,
for models with a small number of parameters it may be useful for exploring the likelihood surface. The
entire point process object is included in the environment of the objective function so that all of the data
and parameters are accessible inside the optimizer. The scoping rules of R (Gentleman and Ihaka, 2000)
make the implementation of this mechanism clean and fairly straightforward. The R function optim is then
called to minimize the negative of the log-likelihood function. optim provides four optimization procedures:
a quasi-Newton method of Broyden, Fletcher, Shanno, and Goldfarb, a conjugate gradient method, the
simplex algorithm of Nelder and Mead (1965), and a simulated annealing procedure based on that of Belisle
(1992). See also Nocedal and Wright (1999) for details on the first two optimization procedures. The user
must choose from these procedures based on the form of the conditional intensity model. For relatively
smooth models with a few parameters, the quasi-Newton and conjugate gradient methods tend to produce
good results quickly. For models with many parameters, the simulated annealing method may be useful for
obtaining a good initial solution. The default Nelder-Mead method tends to produce reasonable results for
a wide class of models.
The function ptproc.fit returns an object of class “ptprocFit”. A “ptprocFit” object is a list contain-
ing the original “ptprocInit” object which was passed to ptproc.fit, a vector of parameters representing
the maximum likelihood estimates and the convergence code returned by optim. If the argument hessian
= TRUE was passed to ptproc.fit, then there is an element hessian which stores the estimated Hessian
matrix around the MLE (otherwise it is NULL).
For the purposes of demonstration, we will fit a homogeneous Poisson model to some simulated data.
This model prescribes a constant conditional intensity over the entire domain (which we will take to be
[0, 1]3 ). That is, λ(t, x) = µ for all (t, x) ∈ [0, 1]3 . We generate the data with the following commands:
> set.seed(1000)
> x <- cbind(runif(100), runif(100), runif(100))
The code for the conditional intensity function is as follows,

7
hPois.cond.int <- function(params, eval.pts, pts = NULL, data = NULL, TT = NULL) {
mu <- params[1]

if(is.null(TT))
rep(mu, nrow(eval.pts))
else {
vol <- prod(apply(TT, 2, diff))
mu * vol
}
}
Finally, we construct the initial point process object with
> ppm <- ptproc(pts = x, cond.int = hPois.cond.int, params = 50,
+ ranges = cbind(c(0,1), c(0,1), c(0,1)))
For this example, 50 was chosen as the initial parameter value for the optimization procedure.
After constructing a “ptprocInit” object with the ptproc function, one can attempt to fit the model
using ptproc.fit. The command
> fit <- ptproc.fit(ppm, method = "BFGS")
minimizes the negative log-likelihood using the BFGS method. Tuning parameters can be passed as a
(named) list to optim via the argument optim.control. For example, it may be desirable to see some
tracing information while fitting the model. One can do this by running
> fit <- ptproc.fit(ppm, optim.control = list(trace = 2), method = "BFGS")
instead. The values in the params element of the “ptprocInit” object are used as the initial parameters
for the optimizer. Certain parameters can be held fixed by setting appropriate values in the fixed vector.
As it was discussed in Section 2.1, it may be necessary to modify the log-likelihood function to include a
penalty term. Normally, including a penalty into the evaluation of the log-likelihood would involve directly
modifying the code for the log-likelihood function. Then for each model which required a different type of
penalty, the log-likelihood function would have to re-written. The ptproc package takes a different approach,
which is to store the penalty term with the particular model object. These user-defined penalties are then
evaluated when the log-likelihood is evaluated. This way, the penalty is identified with the model rather
than the log-likelihood function.
The condition element is included in the “ptprocInit” object for the purpose of including penalties.
By default, it is set to NULL. However, one can include a penalty by using the penalty function. For example,
suppose we wanted to penalize the log-likelihood for negative values of any of the parameters. We could set
> condition(ppm) <- penalty(code = NULL, condition = quote(any(params < 0)))
The penalty function returns an unevaluated R expression and the function condition modifies the “ptprocInit”
object so that the R expression is included. When ppm is passed to ptproc.fit, the expression
if( any(params < 0) )

8
will be evaluated as part of the log-likelihood. If the conditional statement evaluates to TRUE then a penalty of
alpha will be returned before the evaluation of the negative log-likelihood. The value of alpha is an argument
to ptproc.fit and has a default value of zero. If restricting the parameters to be positive guarantees the
positivity of the conditional intensity, then the above code would be an example of implementing the penalty
function in (2).
The entire “ptprocInit” object can be accessed when inserting penalties into the log-likelihood. It can
be accessed using the name ppobj. The vector of model parameters can be accessed separately via the name
params. The situation may arise when several statements must be inserted before a conditional statement
can be evaluated. The code argument to penalty can be used for inserting several statements. There is an
example of the usage of code in Section 6.1.
Finally, we can fit the model with
> fit <- ptproc.fit(ppm, optim.control = list(trace = 2), method = "BFGS", alpha = 1e+9)
initial value -341.202301
final value -360.516878
converged
In this case the computed MLE is 99.83, which is close to the true value of 100.

5.3 Constructing a Residual Process

The ptproc package provides a residuals method for objects of class “ptprocFit” which offers two possible
procedures for generating the residual process. The user is given the option of generating ordinary thinned
residuals or approximate thinned residuals. For ordinary thinned residuals, the user must provide a value
m representing the minimum of the conditional intensity. For approximate thinned residuals, the user must
specify a subsample size K to draw from the original points.
Continuing the example from Section 5.2, we can generate both kinds of residuals with
> r1 <- residuals(fit, type = "ordinary", m = params(fit))
> r2 <- residuals(fit, type = "approx", K = 20)
In this example, since the conditional intensity is constant the minimum is equal to the value of the estimated
parameter. Therefore, we could use m = params(fit) when generating ordinary thinned residuals in the
call to residuals.
The residuals method returns a matrix of points representing the thinned residual process. The number
of rows in the matrix equals the number of points in the thinned process and there is one column for each
dimension in the original dataset. If approximate residuals are used, the number of rows is always equal to
K. These points can then be passed to other functions which test for uniformity. One can generate many
random realizations of residuals by setting the argument R to be greater than 1. If more than one realization
is request, then a list of matrices is returned, with each matrix representing a separate residual process.

5.4 Simulating a Point Process

A point process can be simulated using the random thinning methods of Lewis and Shedler (1979) and Ogata
(1981), given a form for the conditional intensity. These procedures are implemented in the ptproc.sim
function. This method of simulating a point process is very similar to generating a thinned residual process.

9
Here, the user must specify a value M such that

sup λ(t, x) ≤ M < ∞.

(t,x)∈S

Then a homogeneous point process of rate M is generated in the domain S. Suppose there are L points
in this realization. Then for i = 1, . . . , L, each point (ti , xi ) is deleted with probability 1 − λ(ti , xi )/M .
The undeleted points form a realization of a point process with conditional intensity λ. Note that for fitted
models ptproc.sim takes the domain for simulation to be the same domain in which the observed point
process lies.
For simulation, one can either use the ptproc function to construct an object of class “ptprocSim”,
which contains a conditional intensity function and parameter values, or one can coerce an object of class
“ptprocFit” to be of class “ptprocSim” using as.ptprocSim. This latter method may be useful when a
model has been fit and one wishes to evaluate the fit by examining simulations from the model.

10
6 Examples
The various usages of the package are perhaps best demonstrated through examples. In this section we give
an example of fitting a space-time linear model and a one-dimensional Hawkes-type cluster model.

6.1 Fitting a Simple Linear Model to Data

The dataset we use here consists of the times and locations of wildfires in the northern region of Los Angeles
County, California. The spatial locations of the 313 wildfires occurring between 1976 and 2000 are shown in
Figure 1.

● ●
● ● ●
● ● ●
21.0

● ●● ● ●
●●
● ●
● ●●
● ●
● ● ● ● ● ●
● ● ● ● ●
● ●
● ● ● ●
● ●
●
● ●● ● ● ● ●
● ● ●
● ● ● ●●● ● ●
●
● ●
● ●
● ●
20.5

●
● ● ●
● ●● ●
● ● ● ● ●● ●
● ● ● ●● ●● ●
● ● ● ●
● ● ● ● ● ●
● ● ● ●
●● ● ●● ●
Y

● ●● ● ●
●
● ● ●
●● ●
● ●● ● ●● ● ●
● ●● ● ● ● ● ●
● ● ●
20.0

● ●●●● ● ●
● ●
●● ● ●●
● ● ●
● ● ● ● ● ● ●●
●
● ● ●●●● ● ● ● ● ●
● ●● ●
● ●●● ● ●
●● ● ● ● ● ●
●
● ● ●
●● ●
●●● ● ● ● ● ● ●
●
● ● ●● ●●● ●
●● ● ●
● ● ● ●
● ●
● ● ●
● ●● ●
● ● ●●● ●● ● ●
● ● ●
●● ● ● ● ●
●
● ●● ● ●●
● ● ● ●
19.5

● ● ● ●
● ●
●● ● ●● ● ●
●● ● ● ●
● ● ● ●● ●
● ●●
●●●● ● ●

64.0 64.5 65.0 65.5 66.0 66.5

Figure 1: Northern Los Angeles County Wildfires (1976–2000). The direction north is to the towards the
top of the figure.

The dataset is included in the package and can be accessed with the R function data. After loading the
data one should find a 313 × 3 matrix called fires in the global workspace where the first column contains
the times of the wildfires and the second and third columns contain the x and y coordinates, respectively.
The spatial locations are represented with a (scaled) state-plane coordinate system using the NAD 83 datum.
For our dataset one spatial unit corresponds to approximately 18.9 miles. The units of time in this dataset
were produced by the date package and represent the number of days since January 1, 1960.

11
 The model used here is a simple linear model with one parameter for each dimension and a background
parameter. Clearly, such a model is not an ideal way to analyze this dataset. In particular, wildfire occur-
rences can depend on many complex meterological and fuel-related phenomena. However, the model may
serve as a way to describe some broad features of the data (such as spatial and temporal trends) in addition
to demonstrating certain features of the ptproc package. For a more thorough point process analysis of
wildfire data, we refer the reader to Preisler and Weise (2002), Brillinger et al. (2003), or Peng et al. (2003)
The conditional intensity is of the form

λ(t, x, y) = µ + β1 t + β2 x + β3 y. (5)

We first load the data and construct a “ptprocInit” object.

> data(fires)
> ppm <- ptproc(fires, cond.int = linear.cond.int,
+ params = c(mu = .004, beta1 = 0, beta2 = 0, beta3 = 0))
The code for the function linear.cond.int is shown in Appendix A.
Since it is possible to for this conditional intensity function to take negative values we will have to
restrict the parameters somehow. One way would be to restrict all the parameters to be positive. However,
that would likely restrict the model too much by not allowing any negative trend. In this case, since the
conditional intensity is a plane, we can guarantee positivity by testing conditional intensity values at the 8
“corners” of the domain. If the conditional intensity is positive at those corners, then it must be positive
in the middle points. The following statement constructs code to do this evaluation at the corners of the
domain:
> extra.code <- expression(ranges <- as.list(as.data.frame(ranges(ppobj))),
+ corners <- expand.grid(ranges),
+ ci <- evalCIF(ppobj, xpts = corners))
After the conditional intensity has been evaluated at the 8 corners, we must test to see if any of the values
are negative. We will use the penalty and condition functions to modify the condition element of the
“ptprocInit” object.
> condition(ppm) <- penalty(code = extra.code, condition = quote(any(ci < 0)))
The model can be fit without having to worry about negative values of the conditional intensity. We use the
Nelder-Mead optimization method with the default 500 interations and a tracing level of 2. Furthermore,
we set the penalty parameter alpha equal to 105 .
> fit <- ptproc.fit(ppm, optim.control = list(trace=2), alpha = 1e+5)
After the parameters have been estimated, we can print the fitted model
> fit
Model type: LINEAR

Parameter Values:
mu beta1 beta2 beta3
1.610e-01 -8.986e-07 -2.180e-03 -7.746e-05

12
Initial Values:
mu beta1 beta2 beta3
0.004 0.000 0.000 0.000

Fixed Parameters:
mu beta1 beta2 beta3
NA NA NA NA

Condition: expression(ranges <- as.list(as.data.frame(ranges(ppobj))), ...

and check the AIC
> AIC(fit)
[1] 3694.754
Often, the homogeneous Poisson model is a useful null model against which to compare more complex models.
If the more complex model is truly capturing a feature of the data, its AIC value should be much lower than
that of the homogeneous Poisson model. A fitted model’s AIC can be compared against the homogeneous
Poisson model by using summary:
> summary(fit)
Model type: LINEAR

Parameter Values:
mu beta1 beta2 beta3
1.610e-01 -8.986e-07 -2.180e-03 -7.746e-05

Model AIC: 3694.754

H. Pois. AIC: 3759.205
Here, we see that the simple linear model is in fact doing somewhat better than the homogeneous Poisson
model. However, the decrease in AIC is not dramatic.
We can examine the fit of the model further by doing some residual analysis. We first try generating the
ordinary thinned residuals. Since we need to know the minimum of the conditional intensity, the first two
lines of code below compute the conditional intensity at the corners of the domain to find the minimum:
> corners <- expand.grid(as.list(as.data.frame(ranges(fit))))
> ci.corners <- evalCIF(fit, xpts = corners)
> set.seed(100)
> r1 <- residuals(fit, "ordinary", m = min(ci.corners))
> pairs(r1)
The pairs plot of the residual process is shown in Figure 2. Alternatively, we could generate approximate
thinned residuals. Here we will set the subsample size K equal to the number of points we obtained in the
ordinary thinned residual process.
> set.seed(500)
> r2 <- residuals(fit, "approx", K = nrow(r1))
> pairs(r2)

13
The approximate thinned residuals are shown in Figure 3. One can see from both figures that neither residual
process appears to be homogeneous Poisson. Both the ordinary and approximate residuals have a clear trend
from the southwest corner to the northeast corner (see the bottom middle panel) and exhibit some clustering.

14
 64.0 64.5 65.0 65.5

14000
● ● ● ●
● ● ●●

● ● ● ●
● ●

12000
● ●

Time

10000
● ●
● ●
● ●

● ● ● ●
● ● ● ●

8000
● ●
● ● ●● ●● ● ● ● ●● ● ● ●
● ●● ● ● ● ●●
● ● ●●

● ●
● ●
65.5

● ●
● ●
● ●
● ● ● ●
● ●
65.0

● ●
● ●
● ●
X ● ●
64.5

● ● ● ● ●● ● ●
●● ● ●
● ● ● ● ● ●
● ●
● ● ●●
● ●
64.0

● ● ●●

● ● ● ●

● ●

21.0
● ●
● ●
● ● ● ●
● ●

● ●

20.5
● ●
● ● ● ●
● ●

●
● ●
●
Y
20.0

● ●● ● ● ● ●
●● ● ●
● ● ● ●
● ●
● ●
● ●●
●● ●
19.5

● ● ● ●
● ●

8000 10000 12000 14000 19.5 20.0 20.5 21.0

Figure 2: Ordinary thinned residuals from the fitted model in (5)

15
 64.0 64.5 65.0 65.5

● ●

14000
● ● ● ● ● ●
● ●

● ●● ● ● ●
● ● ● ● ● ●

● ●

Time

10000
● ●

● ●

● ● ● ● ● ● ● ●
● ●

8000
● ●
● ●
● ● ● ●
● ● ● ●
● ●
● ●

6000
● ●
● ●
65.5

● ●

● ●
● ● ●
●
● ●
● ●
65.0

● ●
● ● ● ●
● ●
X ● ●
64.5

● ● ● ●
● ● ● ●
● ● ● ●
● ●
● ● ● ●
● ● ● ● ● ●
● ●
64.0

● ● ● ●
● ●

● ●
● ●

● ●
● ●

21.0
● ●
● ●
● ●
● ●
● ●

● ●

20.5
● ●
● ● ●●
● ●
●
●
●
●
Y
● ●
20.0

● ●
● ●
● ●
● ● ●● ● ● ● ●

● ●
19.5

● ●
● ●
● ●
● ●
● ● ● ●

6000 8000 10000 14000 19.5 20.0 20.5 21.0

Figure 3: Approximate thinned residuals (K = 38).

16
 While visual inspection of the residuals can be a useful tool, we may wish to conduct a more systematic
test on either the ordinary or approximate thinned residuals. We will use Ripley’s K-function to test for
spatial clustering and inhibition. The K-function measures, for a given h > 0, the expected proportion of
points within distance h of a point in the process. The splancs package of Rowlingson and Diggle (1993)
contains an implementation of the K-function and has many other useful tools for analyzing spatial point
patterns. We use the khat function in splancs to compute the K-function and the Kenv.csr function to
simulate confidence envelopes. For this example, we use the ordinary thinned residual process for testing
purpose since it appears to have a reasonable number of points in it.
> library(splancs)

Spatial Point Pattern Analysis Code in S-Plus

Version 2 - Spatial and Space-Time analysis

> b <- make.box(fit, 2:3)
> h <- seq(.1, 2, .2)
> K <- khat(r1[,2:3], b, h)
> env <- Kenv.csr(nrow(r1), b, 2000, h)
Instead of plotting the raw K function we plot a standardized version
s
K̂(h)
L̂(h) = −h
π

where K̂(h) is the estimated K function for distance h.

Figure 4(a-b) shows the standardized K function for the original data and the ordinary thinned residuals.
From Figure 4(a) it is clear that the original data are clustered. The dotted red lines are confidence envelopes
for the K function produced by 2000 simulations of a homogeneous Poisson process. The standardized K
function in Figure 4(b) appears quite close to that of a homogeneous Poisson process, but the residual
process still appears to be clustered. This would suggest that the model is not adequately accounting for
some features in the data.
Because of the randomness involved in producing both sets of residuals, a second realization of the residual
process would likely produce a different estimate of the K function. However, one strategy to deal with this
could be to produce many realizations of the residual process, in turn producing many estimates of the K
function. The the range of the K function for the various residual processes can be compared to the range
of the K function for many simulations of a homogeneous Poisson process. If the ranges overlapped closely,
then that would indicate the residual process may be a close to a homogeneous Poisson process.

17
Standardized K function

0.15
0.05
−0.05

0.5 1.0 1.5

Distance
(a)
Standardized K function

0.6
0.2
−0.2

0.5 1.0 1.5

Distance
(b)

Figure 4: Standardized K function for (a) original data and (b) residual process.

18
6.2 Fitting a One-Dimensional Cluster Model
In this section we fit a Hawkes-type cluster model to the times of the earthquake occurrences in Northern
California between 1970 and 2002. The data were obtained from the Northern California Earthquake Data
Center3 . The dataset contains the event times of 849 earthquakes of magnitude greater that 3.0 and occurring
between −121.5 and −121.0 longitude and 36.5 and 37.0 latitude. The units are in days since January 1,
1970. In this example we show two useful diagnostic plots which can be produced from the one-dimensional
residual process. We also show how one might fit a sequence of models when the model can take a variable
number of parameters.
The conditional intensity model we use is

Zt
λ(t) = µ+ g(t − s) N (ds)
−∞
X
= µ+ g(t − ti )
ti <t

where g is the trigger function

K
X
g(z) = ak z k−1 e−cz .
k=1

Here the parameter µ represents the background rate of occurrence while the parameters a1 , . . . , aK and c
control the level of clustering. K represents the order of the trigger function and selecting its value is an
issue we discuss below. More details about this model (and its application to earthquake data) can be found
in Ogata (1988). The R code for this model can be found in Appendix A.
We first construct the “ptprocInit” object (setting K = 1) and then construct a penalty to ensure that
the conditional intensity is positive. Here we simply restrict all of the parameters to be positive. We then
fit the model with ptproc.fit.
> data(cal.quakes)
> rate <- length(cal.quakes) / (365 * 33)

> pstart <- c(mu = rate, C = 1, a = 0.1)

> ppm <- ptproc(pts = cal.quakes, cond.int = hawkes.cond.int, params = pstart)
> condition(ppm) <- penalty(code = NULL, condition = quote(any(params < 0)))
> fit <- ptproc.fit(ppm, optim.control = list(trace = 2), alpha = 1e+5, hessian = TRUE)
> summary(fit)
Model type: HAWKES

Parameter Values:
mu C a
0.05406 11.65518 2.72504

Model AIC: 5191.539

H. Pois. AIC: 6202.561
3 http://quake.geo.berkeley.edu

19
 Conditional intensity (events / day)

8
6
4
2
0

0 2000 4000 6000 8000 10000 12000

Times

Figure 5: Estimated conditional intensity function for the Hawkes model.

The parameter estimate for a indicates that immediately after an earthquake occurs, the conditional intensity
is increased by about 2.7 events per day. The AIC values from the summary output show that the model
does fit better than a homogeneous Poisson model. The estimated conditional intensity is shown in Figure 5.
The plot was constructed by using the package function evalCIF on a grid:
> x <- seq(0, 33*365, len = 12000)
> e <- evalCIF(fit, xpts = x)
> plot(x, e, type = "l", xlab = "Times", ylab = "Conditional intensity (events / day)")
In the call to ptproc.fit we set the argument hessian = TRUE, which directs the optim function to estimate
the Hessian matrix around the MLE. The estimated Hessian matrix is stored in the object returned from
ptproc.fit under the name hessian. We can compute approximate standard errors from the diagonal of
the inverse Hessian matrix. The standard errors for the parameters in this model are
mu C a
0.0022 1.8933 0.4159
A general issue with this kind of cluster model is the choice of K, the order of the polynomial in the
trigger function. In the above example, we arbitrarily chose K = 1. However, we can use the AIC to compare

20
a number of models with different values of K. In the following example, we show how this can be done,
using values of K from 1 to 4. Note that for values of K greater than 2 we provide some scaling information
via the parscale argument to help the optimizer. The use of parscale will typically be necessary to guide
the optimizer towards a good solution (Venables and Ripley, 2002).
> models <- vector("list", length = 4)
> for(k in 1:4) {
+ pstart <- c(mu = rate, C = rate, a = rep(.001, k))
+ ppm <- ptproc(pts = cal.quakes, cond.int = hawkes.cond.int, params = pstart)
+ condition(ppm) <- penalty(code = NULL, condition = quote(any(params < 0)))
+ parscale <- if(k < 3)
+ rep(1, k + 2)
+ else
+ pstart
+ fit <- ptproc.fit(ppm, optim.control=list(trace=3, parscale=parscale, maxit=1000),
+ alpha = 1e+5)
+ models[[k]] <- fit
+ }
> aic <- sapply(models, AIC)
> names(aic) <- 1:4
> aic
1 2 3 4
5191.539 5193.541 5745.554 6042.534
It would appear that K = 1 is the minimum AIC model of the four.
We can further examine the goodness-of-fit of the K = 1 model via residual analysis.
> fit <- models[[1]]
> set.seed(900)
> r <- residuals(fit, "ordinary", m = params(fit)[1])
There are a number of diagnostics one can use to assess the fit of a 1-dimensional residual process. One
example is a log-survivor plot of the interevent times. If the model fits well, then the residual process should
be homogeneous with rate m (where m is defined in (4) and specified in the call to residuals) and the
interevent times should appear as i.i.d. exponential with mean 1/m. The log-survivor plot of the interevent
times of the residual process can be constructed with the log.surv function (included in the package) and
is shown in Figure 6(a).
One can also check the stationarity of the residual process with the stationarity function. This function
divides the domain into bins of a given (user-specified) length and counts the number of points falling into
each bin is counted. The number of points in each bin is standardized by the theoretical mean and standard
deviation and the standardized counts are plotted against the left endpoints of the bins. This plot is shown
in Figure 6(b). The process generally stays within the bounds of a homogeneous Poisson process, but in
one instance the count jumps far beyond three standard deviations of the expected count. This jump could
indicate a region of non-stationarity in the residual process and a feature of the data which is not taken into
account by the model.

21
 Log Survivor Plot
of Interevent Times
●
●
●
●●
●
●
●●
●
Cumulative number

●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
20 100

●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●
●●
●●
●
●●
●●
●●
●●●
●
●●●
●●
●
●●● ●
●●●●
●
●●●
●
● ●
●
● ● ●
●●●
●●
● ●
5

●
●
●
●
1

0 50 100 150

Interevent time

Stationarity Plot
Standardized # of events per interval

h = 300
●
8

●
6

●
4

● ●
●
2

● ● ● ● ●
● ● ●
● ● ● ●
● ●
● ● ● ● ● ●
−2

● ● ● ● ● ● ●
● ● ● ● ● ●
●

0 2000 4000 6000 8000 10000 12000

Transformed time

Figure 6: Diagnostics for 1-dimensional cluster model.

22
7 Future Work
In the future we hope to develop more functions for conducting model evaluation of multi-dimensional
models. We would like to add more diagnostic plots and tools to aid in residual analysis. Also, simulation
based methods of model evaluation and prediction can be very useful and have not been discussed here. We
hope to increase the number of simulation based tools in future releases of the package.
In addition to adding features to the ptproc package, R itself is changing and the package will likely
change with R . For example, version 1.7.0 of R introduced support for packages namespaces which allows
package writers to export a subset of functions and hide internal functions. Also, we will eventually move
to the S4 style class system (Chambers, 1998), which provides a more rigorous object oriented framework.

8 Bug Reports
Please send any bug reports or suggestions to [email protected].

9 Acknowledgments
This work is part of the author’s Ph.D. dissertation from the University of California, Los Angeles. An
anonymous reviewer provided many constructive comments which contributed to the revised form of this
article. The author also thanks Rick Paik Schoenberg, Frauke Kreuter, and Jane Chung for useful comments
on the manuscript and many interesting discussions, and James Woods for generously providing the wildfire
dataset.

References
Baddeley, A. and Turner, R. (2000), “Practical Maximum Pseudolikelihood for Spatial Point Patterns,”
Australian and New Zealand Journal of Statistics, 42, 283–322.
Belisle, C. J. P. (1992), “Convergence theorems for a class of simulated annealing algorithms on Rd .” Journal
of Applied Probability, 29, 885–895.
Berman, M. and Turner, T. R. (1992), “Approximating point process likelihoods with GLIM,” Applied
Statistics, 41, 31–38.
Brillinger, D. R., Preisler, H. K., and Benoit, J. (2003), “Risk Assessment: A Forest Fire Example,” in
Science and Statistics, IMS, vol. 40 of Lecture Notes in Statistics.
Chambers, J. (1998), Programming with Data: A Guide to the S Language, Springer, New York.
Daley, D. J. and Vere-Jones, D. (1988), An Introduction to the Theory of Point Processes, Springer, NY.
Diggle, P. J. (1983), Statistical Analysis of Spatial Point Patterns, Academic Press, NY, London.
Diggle, P. J., Chetwynd, A. G., Häggkvist, R., and Morris, S. E. (1995), “Second-order Analysis of Space-time
Clustering,” Statistical Methods in Medical Research, 4, 124–136.

23
Gentleman, R. and Ihaka, R. (2000), “Lexical scope and statistical computing,” Journal of Computational
and Graphical Statistics, 9, 491–508.
Harte, D. (1998), “Documentation for the Statistical Seismology Library,” Tech. Rep. 98-10, School of
Mathematical and Computing Sciences, Victoria University of Wellington.
Ihaka, R. and Gentleman, R. (1996), “R: A language for data analysis and graphics,” Journal of Computa-
tional and Graphical Statistics, 5, 299–314.
Lewis, P. A. W. and Shedler, G. S. (1979), “Simulation of nonhomogeneous Poisson processes by thinning,”
Naval Research Logistics Quarterly, 26, 403–413.
Nelder, J. A. and Mead, R. (1965), “A simplex algorithm for function minimization,” Computer Journal, 7,
308–313.
Nocedal, J. and Wright, S. J. (1999), Numerical Optimization, Springer.
Ogata, Y. (1978), “The asymptotic behavior of maximum likelihood estimators for stationary point pro-
cesses,” Annals of the Institute of Statistical Mathematics, 30, 243–261.
— (1981), “On Lewis’ simulation method for point processes,” IEEE Transactions on Information Theory,
27, 23–31.
— (1983), “Likelihood analysis of point processes and its applications to seismological data,” Bull. Int.
Statist. Inst., 50, 943–961.
— (1988), “Statistical models for earthquake occurrences and residual analysis for point processes,” Journal
of the American Statistical Association, 83, 9–27.
— (1998), “Space-time point process models for earthquake occurrences,” Annals of the Institute of Statistical
Mathematics, 50, 379–402.
— (1999), “Seismicity analysis through point-process modeling: a review,” Pure and Applied Geophysics,
155, 471–507.
Ogata, Y. and Tanemura, M. (1984), “Likelihood analysis of spatial point patterns,” Journal of the Royal
Statistical Society, Series B, 46, 496–518.
Peng, R. D., Schoenberg, F. P., and Woods, J. (2003), “Multi-dimensional Point Process Models for Evalu-
ating a Wildfire Hazard Index,” Tech. Rep. 350, UCLA Department of Statistics.
Preisler, H. K. and Weise, D. (2002), “Forest Fires Models,” in Encyclopedia of Environmetrics, eds. El-
Shaarawi, A. and Piegorsch, W., Wiley, Chichester, pp. 808–810.
Rathbun, S. L. (1996), “Asymptotic properties of the maximum likelihood estimator for spatio-temporal
point processes,” Journal of Statistical Planning and Inference, 51, 55–74.
Rathbun, S. L. and Cressie, N. (1994a), “Asymptotic properties of estimators for the parameters of spatial
inhomogeneous Poisson point processes,” Advances in Applied Probability, 26, 122–154.
— (1994b), “A space-time survival point process for a longleaf pine forest in southern Georgia,” Journal of
the American Statistical Association, 89, 1164–1174.

24
Ripley, B. (1976), “The second-order analysis of stationary point processes,” Journal of Applied Probability,
13, 255–266.
— (1979), “Tests of ‘randomness’ for spatial point patterns,” Journal of the Royal Statistical Society, Series
B, 41, 368–374.
Rowlingson, B. and Diggle, P. (1993), “Splancs: spatial point pattern analysis code in S-Plus,” Computers
and Geosciences, 19, 627–655.
Schoenberg, F. (1999), “Transforming spatial point processes into Poisson processes,” Stochastic Processes
and their Applications, 81(2), 155–164.
Schoenberg, F. P. (2003), “Multi-dimensional Residual Analysis of Point Process Models for Earthquake
Occurrences,” Tech. Rep. 347, UCLA Department of Statistics.
Venables, W. N. and Ripley, B. D. (2002), Modern Applied Statistics with S, Springer, NY, 4th ed.
Vere-Jones, D. (1970), “Stochastic models for earthquake occurrence,” Journal of the Royal Statistical Soci-
ety, Series B, 32, 1–62.
Vere-Jones, D. and Ozaki, T. (1982), “Some examples of statistical estimation applied to earthquake data,”
Annals of the Institute of Statistical Mathematics, 34, 189–207.

25
A Appendix: Code
A.1 Simple Linear Model
Below is the code for the simple linear model used in Section 6.1. The conditional intensity is of the form

λ(t, x, y) = µ + β1 t + β2 x + β3 y.

linear.cond.int <- function(params, eval.pts, pts = NULL, data = NULL, TT = NULL) {

mu <- params[1]
beta <- params[-1]

if(is.null(TT)) {
## Evaluate
ci <- mu + eval.pts %*% beta
ci <- as.vector(ci)
}
else {
## Integrate
total.vol <- prod(apply(TT, 2, diff))
m.vol <- sapply(1:ncol(TT), function(i)
{
z <- TT[, -i, drop=FALSE]
prod(apply(z, 2, diff))
})
d <- apply(TT^2 / 2, 2, diff)
ci <- mu * total.vol + (beta * d) %*% m.vol
}
ci
}

A.2 Hawkes-type Cluster Model

The version of Hawkes’ self-exciting model used in Section 6.2 is
K
XX
λ(t) = µ + ak (t − ti )k−1 e−c(t−ti ) .
ti <t k=1

The R code is as follows:

hawkes.cond.int <- function(params, eval.pts, pts = NULL, data = NULL, TT = NULL) {

mu <- params[1]
C <- params[2]
ak <- params[-(1:2)]
K <- length(ak)

if(K < 1)

26
 stop("K must be >= 1")
if(is.null(TT)) {
S <- sapply(as.vector(eval.pts), function(x, times, ak, C)
{
use <- times < x
if(!is.na(use) && any(use)) {
d <- x - times[use]
k <- 0:(length(ak)-1)
lxk <- outer(log(d), k) + (-C * d)
sum(exp(lxk) %*% ak)
}
else 0
}, times = as.vector(pts), ak = ak, C = C)
ci <- mu + S
}
else {
Rfunc <- function(x, L, c) {
k <- 1:L
g <- gamma(k) / c^k
o <- outer(x, k, pgamma, scale = 1/c)
r <- o * rep(g, rep(nrow(o), ncol(o)))
r
}
times <- as.vector(pts)
ci <- mu * (TT[2,1] - TT[1,1])
S <- double(2)

for(i in 1:2) {
use <- times < TT[i,1]

if(any(use)) {
r <- Rfunc(TT[i,1] - times[use], K, C)
S[i] <- sum(r %*% ak)
}
}
ci <- ci + (S[2] - S[1])
}
ci
}

27