dispmod (version 1.0.1)

glm.poisson.disp: Overdispersed Poisson log-linear models.

Description

This function estimates overdispersed Poisson log-linear models using the approach discussed by Breslow N.E. (1984).

Usage

glm.poisson.disp(object, maxit = 30, verbose = TRUE)

Arguments

object

an object of class `"glm"' providing a fitted binomial logistic regression model.

maxit

integer giving the maximal number of iterations for the model fitting procedure.

verbose

logical, if TRUE information are printed during each step of the algorithm.

Value

• The function returns an object of class `"glm"' with the usual information (see help(glm)) and the added components:
• dispersion
  the estimated dispersion parameter.
• disp.weights
  the final weights used to fit the model.

Details

Breslow (1984) proposed an iterative algorithm for fitting overdispersed Poisson log-linear models. The method is similar to that proposed by Williams (1982) for handling overdispersion in logistic regression models (glm.binomial.disp).

Suppose we observe $n$ independent responses such that $$y_i|\lambda_i \sim \mathrm{Poisson}(\lambda_in_i)$$ for $i=1\ldots,n$. The response variable $y_i$ may be an event counts variable observed over a period of time (or in the space) of length $n_i$, whereas $\lambda_i$ is the rate parameter. Then, $$E(y_i|\lambda_i) = \mu_i = \lambda_in_i=\exp(\log(n_i) + \log(\lambda_i))$$ where $\log(n_i)$ is an offset and $\log(\lambda_i)=\beta'x_i$ expresses the dependence of the Poisson rate parameter on a set of, say $p$, predictors. If the periods of time are all of the same length, we can set $n_i=1$ for all $i$ so the offset is zero.

The Poisson distribution has $E(y_i|\lambda_i)=Var(y_i|\lambda_i)$, but it may happen that the actual variance exceeds the nominal variance under the assumed probability model. Suppose now that $\theta_i=\lambda_i n_i$ is a random variable distributed according to $$\theta_i \sim \mathrm{Gamma} (\mu_i, 1/\phi)$$ where $E(\theta_i)=\mu_i$ and $Var(\theta_i)=\mu_i^2\phi$. Thus, it can be shown that the unconditional mean and variance of $y_i$ are given by $$E(y_i) = \mu_i$$ and $$Var(y_i) = \mu_i + \mu_i^2\phi = \mu_i(1+\mu_i\phi)$$ Hence, for $\phi>0$ we have overdispersion. It is interesting to note that the same mean and variance arise also if we assume a negative binomial distribution for the response variable.

The method proposed by Breslow uses an iterative algorithm for estimating the dispersion parameter $\phi$ and hence the necessary weights $1/(1+\mu_i\hat{\phi})$ (for details see Breslow, 1984).

References

Breslow, N.E. (1984), Extra-Poisson variation in log-linear models, Applied Statistics, 33, 38--44.

See Also

lm, glm, lm.disp, glm.binomial.disp

Examples

##-- Salmonella TA98 data

data(salmonellaTA98)
attach(salmonellaTA98)
log.x10 <- log(x+10)
mod <- glm(y ~ log.x10 + x, family=poisson(log)) 
summary(mod)

mod.disp <- glm.poisson.disp(mod)
summary(mod.disp)
mod.disp$dispersion

# compute predictions on a grid of x-values...
x0 <- seq(min(x), max(x), length=50) 
eta0 <- predict(mod, newdata=data.frame(log.x10=log(x0+10), x=x0), se=TRUE)
eta0.disp <- predict(mod.disp, newdata=data.frame(log.x10=log(x0+10), x=x0), se=TRUE)
# ... and plot the mean functions with variability bands
plot(x, y)
lines(x0, exp(eta0$fit))
lines(x0, exp(eta0$fit+2*eta0$se), lty=2)
lines(x0, exp(eta0$fit-2*eta0$se), lty=2)
lines(x0, exp(eta0.disp$fit), col=2)
lines(x0, exp(eta0.disp$fit+2*eta0.disp$se), lty=2, col=2)
lines(x0, exp(eta0.disp$fit-2*eta0.disp$se), lty=2, col=2)

##--  Holford's data

data(holford)
attach(holford)

mod <- glm(incid ~ offset(log(pop)) + Age + Cohort, family=poisson(log)) 
summary(mod)

mod.disp <- glm.poisson.disp(mod)
summary(mod.disp)
mod.disp$dispersion