I am analyzing proportion data with GLMM in which the number of occurrence a behaviour of interest has been displayed and has not been displayed are concatenated and fitted as a single response variable.

Fitting a conventional bimodial model with glmer and computing the dispersion statistic as the sum of square of the pearson's residuals/residual degree of freedom returns a value of 4.9, suggesting a high overdispersion.

I then fitted a similar model in glmmADMB but specified a beta-binomial error structure instead, which I thought would be appropriate to account for the large variance in my data. The model runs smoothly and estimated parameters, standard errors and p-value look more realistic than the binomial model. However at my surprise, the dispersion statistic calculated as described above returned a value very similar to the one obtained when fitting the binomial model. I was wondering whether this is something I have to worry about and indicate that the model is inappropriate or that the overdispersion statistic is of no relevance anymore when using beta-binomial error structure


Your Answer

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

Browse other questions tagged or ask your own question.