Professional Documents
Culture Documents
Approximating Point Process Likelihoods With GLIM
Approximating Point Process Likelihoods With GLIM
Approximating Point Process Likelihoods With GLIM
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
data=spruces
. n <- data$n
xx <- quadscheme(unmark(data),nd=100,method="grid")
Xn <- as.ppp(xx,win=data$window)
Xn
w <- dirichlet.weights(Xn)
length(w)
sum(w)-area.owin(data$windo)
#fitting homogeneous Poisson process
log.lik.poiss <- function(par){
n*log(par[1])- area.owin(data$window)* par[1]
}
optim(0.5,log.lik.poiss, control=list(fnscale=-1),hessian=TRUE)
#fitting inhomogeneous Poisson process see page 97 of spatstat
log.lik.i.poiss <- function(par){
sum(par[1]+par[2]*data$x+par[3]*data$y)-sum(w*exp( par[1]+par[2]*Xn$x+par[3]*Xn$y ))
}
opt <- optim(c(0,0,0),log.lik.i.poiss, control=list(fnscale=-1),hessian=T)
eigen(opt$hessian)
Wiley and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to
Journal of the Royal Statistical Society. Series C (Applied Statistics).
http://www.jstor.org
CSIRO, Sydney,Australia
SUMMARY
This paper shows how approximatemaximumlikelihoodestimationforfairlygeneralpointprocesses
on the line can be performedwith GLIM. The approximation is based on a weighted sum
approximationto an integralin the likelihood.Various weightingschemes are brieflyexamined. The
methodology is illustratedwith an example, and its extension to Poisson processes in higher
dimensions is brieflydescribed.
Keywords: GLIM; Likelihood;Pointprocess; Poisson process; Weights
1. Introduction
Althoughmodelsforpointprocesseson thelineand theirstatisticalanalysisabound
(Cox and Lewis, 1966;Cox and Isham, 1980),therehas been,untilrecently, relatively
littlemodellingof the dependenceof pointprocesseson covariatesand theirasso-
ciatedstatisticalanalysis.Probablytheearliestmajor modelsof thistypewereCox's
(1972) modulatedPoisson processesand theirgeneralization,and Berman's (1981)
modulated gamma processes. Both papers only seriouslyaddressed questions of
hypothesistesting(via the likelihoodratiotest),but did not satisfactorily
solve the
problemof parameterestimation.More recently,therehave been several papers
devotedto seismologicalapplications,e.g. Ogata and Akaike (1982), Ogata (1983,
1988), Ogata and Katsura(1986) and Vere-Jonesand Ozaki (1982). In thesepapers,
therelationshipbetweenthepointprocessand thecovariates,whilebeingnumerically
convenient,can giveriseto likelihoodswhichcan be negativeforcertainparameter
values. To overcomethisproblem,most of the above-mentioned researchersuse a
penalizedlikelihoodapproachto estimatetheparameters.As a result,the statistical
propertiesof theseestimatorsare not clear.
Anotherrecentdevelopmentin this fieldis the use of generalizedlinearmodels
(GLMs). Becker(1983, 1986) and Brillinger and Preisler(1986) use theGLM frame-
workforspecificinhomogeneousPoisson processes.However,theydo notexploitthe
fullpotentialofthisframework. In thispaper,we shallshowhowa fairlylargeclass of
point processes on the line, whichincorporatedependenceon covariates,can be
modelledas GLMs; thisclass includesmodels(or slightvariantsof them)used in all
3. Approximating
LikelihoodswithGLIM
There are two closelyrelatedways of using GLIM to maximizeequation (2.3)
approximately.Both involvea discretization
of thetimeaxis. One is a probabilistic
approximation,whiletheotheris numerical.In theprobabilisticapproximation,the
wherethewjareweightsdetermined
bysomequadratureruleand 0 s, < s2 < . . .
SM= T. Choose thesj so thatX- {xl,. ..., x} C {s1,S2, . ., SM} S. We shall call
pointsin S, X and S - X the designpoints,the data pointsand the dummypoints
respectively. Let Nj = 1 if sj E X and Nj = 0 otherwise.Then substitution
of equation
(3.1) into equation (2.3) gives
M
L(O) _ Wj{(Nj/wj) log Xo(sj FsJ)- Xo(sj
I1)}. (3.2)
j=1
The right-hand side of equation (3.2) is (modulo a constantnot dependingon 0) the
weightedlog-likelihoodof a familyof independentPoisson variateswithparameters
X0(sj IFsj). The maximizationof equation(3.2) can thusbe accomplishedin GLIM by
declaring{Nj/wj} as theYVARIATE vectorwithPoisson ERRORS, WEIGHT given
by { wj}, and theLINK functionto be g (LOG is thedefaultforPoisson errors).
Note the followingpoints:
GLIM does not object to the factthatthe 'Poisson' variablesNj/wjand the
(a)
weightswjare non-integer;
(b) GLIM does objectto the wjbeingnegative;
(c) theparameterestimates,likelihoodvaluesand covariancematricesproduced
by GLIM when maximizingthe approximatelog-likelihood(3.2) are valid
approximationsto analogous functionsof thetruelog-likelihood(2.3) in the
sensethat,ifM --+oo in such a waythatapproximation(3.1) becomesexact,
theapproximating functionsof equation(3.2) willconvergeto theanalogous
functionsof equation (2.3).
Often,anydummypointswillbe chosento be equispaced,thenumberbeingchosen
accordingto the(numerical)accuracyof estimationrequired.The necessityof using
thedata pointswillimpose a non-uniform spacingof thedesignpoints,however.
Simpleschemesforweightsare obtainedeitherby
(a) a linear approximation to X(x) on [sj, sj? j,i]= 1, . ., M- 1 or
(b) a quadratic approximationto Xo(x) on [s2j-1, S2j 1], j= 1, * *, I (M- 1)
We tried both the IDENTITY link g(y) =y (i.e. Ogata's model but without any
penalty function) and the LOG link g(y) = exp y. Because of its complexity,we have
only considered versions of equation (3.3) withthe value of d (1.07 x 10-4) chosen by
Ogata. For a range of (J, K) values, we used GLIM to maximize log-likelihood (3.2)
withX0(x IFx) given by equation (3.3). We used both linear and quadratic weights. The
total number M of design points used varied between the minimumpossible, 60, and
1093. With the IDENTITY link, we found that GLIM stopped for some (J, K, M)
combinations because X (sj IFj) was negative for somej. However, no problems were
encountered using the LOG link. With this link function,we calculated the AIC for
each (J, K) combination (i.e. equation (2.6) with Lp replaced by L). Often there are
several models with AIC values very close to the minimum. It is important in such
cases to estimate the log-likelihood accurately; however, if our model is fixed and we
are only interested in parameter estimates and their standard errors, accurate
estimation of the log-likelihood may be less important. A CIF such as equation (3.3)
will tend to oscillate more as J or K increases, by analogy with ordinarypolynomials.
Hence, as a general rule, the larger J or K is, the more design points will be needed to
estimateL(O) accurately. This is borne out by the example considered here. In Table 1,
we give the AIC values forvarious (J, K) combinations using the LOG link and linear
weightswith M= 60 and M= 1061.
Table 1 shows clearly how the two approximations to the AIC for each (J, K)
combination become more discrepant broadly speaking as J or K increases. For all
TABLE 1
AIC valuesfor a model withCIF givenbyequation (3.3) usingg(x) = exp x, d= 1.07x 10 4and linear
weightswithM= 60 and M= 1061
0 743.9 743.9 727.6 727.4 730.4 729.5 730.5 723.6 735.5 727.9
1 725.7 725.3 729.5 729.1 732.1 731.2 732.0 724.5 737.4 729.6
2 727.6 727.1 731.5 731.1 730.6 729.3 733.9 725.8 738.7 731.6
0
o
C)
W0- t
l I
days
scheme,approximateXG(x)byXG(sj)on theDirichlettilesurrounding
sj. In thesecond
scheme,approximateX0(x)by a plane on the Delaunay trianglecontainingx. Both
schemesare guaranteedto producenon-negativeweights.
Our experiencehas been thatDirichletor Delaunay-basedquadraturerulesgive
answerscomparable withthose fromproductrules using far fewerdesign points
withconsequentsubstantialsavingsin CPU time. Details are available fromthe
authors.
5. SummarizingRemarks
In thispaper, we have shownhow approximateML estimationforfairlygeneral
point processeson the line and Poisson processesin space can be performedwith
GLIM. In addition,we have demonstrated how theuse of penaltyfunctionscan be
avoided,and how, forspatialdata, storageand CPU timeproblemscan be overcome
by usingDirichlettessellationsor Delaunay triangulations
insteadof productrules.
We hopethatourpaperwillencouragethegreateruse ofpointprocessmodellingwith
theaid of GLIM.
References
Ahuja, N. and Schachter,B. J. (1983) PatternModels, sect. 1.3. New York: Wiley.
Becker,N. G. (1983) Analysisof data froma singleepidemic.Aust. J. Statist.,25, 191-197.
(1986) A generalizedlinearmodellingapproachto theanalysisof data froma singleepidemic.In
Proc. Pacific StatisticalCongr.,Auckland,pp. 464-467. Amsterdam:North-Holland.
Berman,M. (1981) Inhomogeneousand modulatedgammaprocesses.Biometrika,68, 143-152.
Brillinger,D. R. and Preisler,H. K. (1986) Two examplesof quantaldata analysis:(a) multivariate
point
process,(b) puredeathprocessin an experimental design.13thInt. BiometricsConf., Seattle.
Clayton,D. G. (1983) Fittinga generalfamilyof failure-time distributions
usingGLIM. Appl. Statist.,
32, 102-109.
Cox, D. R. (1972) The statisticalanalysis of dependenciesin point processes. In StochasticPoint
Processes(ed. P. A. W. Lewis), pp. 55-66. New York: Wiley.
Cox, D. R. and Isham, V. (1980) Point Processes. London: Chapman and Hall.
Cox, D. R. and Lewis,P. A. W. (1966) The StatisticalAnalysisof Seriesof Events.London: Methuen.
Davis, P. J. and Rabinowitz,P. (1975) Methods of NumericalIntegration,sect. 5.6. New York:
AcademicPress.
Green, P. J. and Sibson, R. (1978) ComputingDirichlettessellationsin the plane. Comput. J., 21,
168-173.
Hurley,M. A. (1990) Fittingsome inhomogeneouspointprocesses.To be published.
Ogata, Y. (1978) Asymptoticbehaviourof themaximumlikelihoodestimatorsforthestationarypoint
processes.Ann. Inst. Statist.Math., 30, 243-261.
(1983) Likelihoodanalysisof pointprocessesand itsapplicationsto seismologicaldata. Bull. Int.
Statist.Inst., 50, 943-961.
(1988) Statisticalmodelsforearthquakeoccurrencesand residualanalysisforpointprocesses.J.
Am. Statist.Ass., 83, 9-27.
Ogata, Y. and Akaike, H. (1982) On linearintensity models formixeddoublystochasticPoisson and
self-excitingpointprocesses.J. R. Statist.Soc. B, 44, 102-107.
Ogata, Y. and Katsura, K. (1986) Point-processmodels with linearlyparameterizedintensityfor
applicationto earthquakedata. J. Appl. Probab. 23A, 291-310.
Payne, C. D. (1985) The GLIM SystemRelease 3.77 Manual. Oxford:NumericalAlgorithmsGroup.
Vere-Jones,D. and Musmeci,F. (1986) Orthogonalpolynomialestimatesforpointprocessdensities.
TechnicalReport.Instituteof Statisticsand OperationsResearch,VictoriaUniversity of Wellington.
Vere-Jones, D. and Ozaki, T. (1982) Some examplesof statisticalinferenceappliedto earthquakedata.
Ann. Inst. Statist.Math., 34, 189-207.