Tests for Overdispertion {DCluster} | R Documentation |
When working with count data, the assumption of a Poisson model is common. However, sometimes the variance of the data is significantly higher that their mean which means that the assumption of that data have been drawn from a Poisson distribution is wrong.
In that case is is usually said that data are overdispersed and a better
model must be proposed. A good choice is a Negative Binomial distribution
(see, for example, negative.binomial
.
Tests for overdispersion available in this package are the Likelihood Ratio Test (LRT) and Dean's P_B and P'_B tests.
test.nb.pois(x.nb, x.glm) DeanB(x.glm, alternative="greater") DeanB2(x.glm, alternative="greater")
x.nb |
Fitted Negative Binomial. |
x.glm |
Fitted Poisson model. |
alternative |
Alternative hipothesis to be tested. It can be "less", "greater" or "two.sided", although the usual choice will often be "greater". |
The LRT is computed to compare a fitted Poisson model against a fitted Negative Binomial model.
Dean's P_B and P'_B tests are score tests. These two tests were proposed for the case in which we look for overdispersion of the form var(Y_i)=μ_i(1+tau μ_i), where E(Y_i)=μ_i. See Dean (1992) for more details.
An object of type htest with the results (p-value, etc.).
Dean, C.B. (1992), Testing for overdispersion in Poisson and binomial regression models, J. Amer. Statist. Assoc. 87, 451-457.
DCluster, achisq.stat, pottwhit.stat, negative.binomial (MASS), glm.nb (MASS)
library(spdep) library(MASS) data(nc.sids) sids<-data.frame(Observed=nc.sids$SID74) sids<-cbind(sids, Expected=nc.sids$BIR74*sum(nc.sids$SID74)/sum(nc.sids$BIR74)) sids<-cbind(sids, x=nc.sids$x, y=nc.sids$y) x.glm<-glm(Observed~1+offset(log(sids$Expected)), data=sids, family=poisson()) x.nb<-glm.nb(Observed~1+offset(log(Expected)), data=sids) print(test.nb.pois(x.nb, x.glm)) print(DeanB(x.glm)) print(DeanB2(x.glm))