To check for overdispersion in GLM with a Poisson distribution one can compare the residual deviance with the residual degrees of freedom. If they are equal the Poisson error assumption is appropriate (The R Book, Crawley, 2007). If the residual deviance is bigger than the residual degrees of freedom overdispersion is indicated. Crawley indicates that overdispersion for Poisson distributions can be corrected with the quasi-Poisson distribution.

My question is how can I check for overdispersion with the Gaussian distribution and how can I correct for it?


1 Answer 1


how can I check for overdispersion with the Gaussian distribution and how can I correct for it?

The Poisson and the binomial have a variance that's a fixed function of the mean. e.g. for a Poisson, $\text{Var}(X)=\mu$, so it's possible to have some count data which has $\text{Var}(X)>\mu$, i.e. more dispersed than would be expected for the Poisson. There's no corresponding situation for the Gaussian. [If variance were some fixed value, like $1$, then a sample with larger variance would be overdispersed, but in the Gaussian family it's just another Gaussian.]

Since the Gaussian has a variance parameter, more dispersion will just be a larger variance parameter... so you don't have overdispersion with the Gaussian.

So there's nothing to correct. (On the other hand, changing dispersion would be an issue to deal with)

  • $\begingroup$ To this can be added the fact that the variance and mean of a Gaussian typically don't even have the same dimensions or units of measurement. So comparing variance and mean is usually nonsensical. In a way, this is the same point from another perspective: not only is there no relationship; there could not be. (The exceptions would be when the variable being considered has no units.) $\endgroup$
    – Nick Cox
    Apr 2, 2015 at 12:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.