Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: group size needed for mixed models (binary response)


From   Susan Lingle <susan.lingle@uleth.ca>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: group size needed for mixed models (binary response)
Date   Mon, 26 Nov 2007 19:29:45 -0700

Dear Jeph

Your response helps tremendously - I have been struggling with this for **quite** some time. I wanted to make the change to a mixed model, because it was in response to reviewers' comments (and it can be wise to heed such advice). But I am not familiar with mixed models, and in some ways it made sense but not in other ways. The initial plan when I captured fawns was to use data for the first sibling that was captured (most fawns have twins, but we did not always capture the twins, so the number of siblings is not a factor that I will include). That approach would be consistent with your "instinct" to select one sibling based on a certain criterion.

I am pretty sure survival of twins is correlated and can easily check that (in JMP, I'm not that familiar with Stata yet). Plus twins live in the same area and their area has a large influence on their survival. So another alternative is to include area as the random factor and not the mother's identity (mean of 36 fawns per 4 sub-areas within the study area). Previously I had planned to include area as a fixed factor.

This evening, I was considering comparing AIC (or BIC) values to select a model with or without the mother's identity included as a random factor. Would that be reasonable? Again, I am getting into procedures I have not used, and would prefer to stay on solid ground in this publication. I have not used AIC or BIC previously, but a quick check in Stata suggested that the model without the random factor of mother's identity has a slightly lower AIC & BIC and a slightly higher LL than the model that included the random factor. I could use the "unmixed" model and mention in passing that inclusion of the mother's identity did not improve the model.

I wanted to test a specific hypothesis (WT are more likely to die during summer; MD during winter) so did not plan to undertake a model selection procedure. Anyway, I think your response gives me enough food for thought now - checking the intraclass correlation is a sensible place to start before making the next decision.

Susan


*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/




© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index