[R-sig-ME] Cook's distance: glmmADMB ZIP / influence.ME not compatible?
bbolker at gmail.com
Wed Mar 18 14:26:16 CET 2015
Ashley Asmus <ashley.asmus at ...> writes:
> Dear r-sig-ME,
> I am working with a small (zero-inflated) dataset of counts. The design is
> repeated-measures, with several measures at each time point made in
> experimental blocks. I have fit the model with glmmADMB using the
> zero-inflated poisson options:
> zip<-glmmadmb(wasp~ treatment + (1|sample.period)
> + (1|block), data=data,
> family="poisson", zeroInflation=TRUE)
> I am suspicious of 1-3 observations (the dataset is only 24 observations
> big, so I promise to be cautious), and wish to evaluate their influence on
> the model fit. Usually I work in lme4, then use the influence.ME package
> and eliminate points based on cook's distance/leverage plots. Applying
> influence.ME to the glmmADMB model produces a strange error message:
> infl <- influence(zip, obs = TRUE)
> Error in eval(expr, envir, enclos) : object 'sample.period' not found
> For some reason influence.ME is not finding my random factor
> "sample.period." Are glmmADMB and influence.ME incompatible when
> zero-inflation models are involved? If anyone knows of an alternative
> approach I would love to hear about it.
No, glmmADMB isn't compatible with influence.ME; the guts of
glmmADMB are very different.
It sounds like you've identified suspicious observations already,
but if you haven't then I would try identifying by eye from fitted
vs. residual or predictor vs. residual plots, then drop them and
see what happens to the fitted model.
More information about the R-sig-mixed-models