Margins after mi estimate

[Gail McAvay]

I am using the mi estimate statement with the glm procedure and would like to use the margins command. But I get a message last estimates not found. Does anyone know if it is possible to get the results of margins somehow?

[daniel klein]

Have a look at http://www.ats.ucla.edu/stat/stata/f...arginsplot.htm and the canned implementation mimrgns from the SSC.

There are some issues, though.

First off, it is not clear (at least to me) whether the described routine produces appropriate results. Marginal effects calculated by margins are based on the predicted values. Combining the marginal effects from multiple datasets into one estimate (Rubin's rules) assumes that they are normally distributed, which might be true for linear predictions (i.e. xb option), but might not be true otherwise. White et al. (2011) even suggest to obtain predicted values at the observational level with mi data - which is what mi predict does, but not margins and therefore mimrgns.

You should also be careful with marginsplot as the plotted CIs are based on inappropriate degrees of freedom and are therefor wrong (the differences are probably so tiny, you won't notice in a graph).

This said, mimrgns does not yet support pairwise comparisons. I am working on an update (for about three month or so) that will support pairwise comparisons (but not predictions on individual level) and also discuss the caveats mentioned above in the help file.

Best
Daniel


White, I. R., P. Royston, and A. M. Wood. 2011. Multiple imputation using chained equations: Issues and guidance for practice. Statistics in Medicine 30: 377-399.