nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Work around for GLMM with repeated measures?

Peter_Polito

A potential customer needs GLMM with repeated measures. I understand GLMM is available as an add-on and will be native in 17 but will not handle repeated measures. Currently, Mixed Models can handle the repeated measures and random effects but not non-normal responses. Any creative workarounds to handles a) a non-normal response (Poisson), b) random effects, and c) repeated measures? The use case is a manufacturer of insecticide performing counts on plots of grass to test efficacy of their product. Thank you.

Accepted answers

eclaassen

Hi, Peter,

I know this doesn't help right now, but I can offer that we're optimistic that at least some repeated measures structures will be added for 18. We know they're necessary (along with additional distributions); the initial release in 17 had to be scoped to get what is probably ~80% of use cases. We'll continue adding things, just as GenReg added distributions and methods across several releases.

Mia pinged @russ_wolfinger and I about your post, and Russ had the following suggestion:
" I’d suggest a log2(1+x) transform of the counts and use normal theory mixed models. Unit changes in the transformed data are interpretable as a doubling of the count.

One thing to watch out for is an excess of zeros, in which case the zero-inflated models in GenReg might make sense at first, but note these do not handle random effects or repeated measures. Could make sense to pre-analyze the zero / non-zero data and then the nonzero data as above."

All comments

Chris_Kirchberg

Hi @Peter_Polito ,

Are you sure it is a Poisson distribution? Is it counts per unit time? Or is it a Binomial/Negative Binomial Distribution, just counts? Either way, both might approximate a normal distribution if lambda is larger (Poisson) or p and n are larger (Binomial). Best to check the distribution of the response first.

With this said, I think it is still OK to try and use Mixed Models if the responses are close to normal distribution and the residuals are normally distributed. Hopefully others will correct me if I am wrong.

Peter_Polito

Thanks @Chris_Kirchberg, this seemed like the most logical path forward. The customer said it is Poisson but I haven't seen the data myself. I appreciate the confirmation.

eclaassen

Hi, Peter,

Peter_Polito

A brief update. I've spoken to the customer some more and he has told me that the EPA requires them to use Generalized Linear Mixed Models (with repeated measures) in the analysis they submit for approval. I found this line in EPA documentation, "An analysis using generalized linear mixed models...is recommended to determine the LD values for each pest strain tested." https://www.regulations.gov/document/EPA-HQ-OPPT-2009-0150-0037, p. 34.

eclaassen

If they are required to use a GLMM, they will likely need to use SAS PROC GLIMMIX or an R GLMM package for now. I've not used R for a repeated measures GLMM, so I can't speak very well to that. But SAS can certainly do the job.