Influence measures after -meqrlogit-

[Chris Byrne]

Hello,

I'm using Stata BE17 to run an analysis for a recent study. One binary predictor with random effects for location and store at that location.

 meqrlogit [primary outcome] i.predictor ||location:||store:, or variance laplace

The regression ran fine and it was significant with good AUC.

I'm wanting to check influence measures (e.g. Cook's D/DfBeta etc.) to check for outliers/influential cases, but can't work out how to do so using any of the post-estimation commands. I'm led to believe these measures weren't designed for MLE/are more appropriate to OLS regression, and are unavailable for meqrlogit in Stata.

Could anyone point me towards a workaround or a relevant tutorial at all please which could help me identify outliers/influential cases?

Many thanks
Chris

[Nick Cox]

Please tell us about cross-posting https://www.statalist.org/forums/help#crossposting

https://www.reddit.com/r/stata/comme...ter_meqrlogit/

[Leonardo Guizzetti]

Cook's D and the link are more difficult to define when working with multilevel/hierarchical models because of the random components. One meaningful generalization of leverage (Cook's D) is to run the regression repeatedly, each time leaving out a whole unit at level 2 or level 3, and observing how the coefficients of interest change. There is a package -mlt- (SSC) that implements those ideas suggested by Raudenbush and Bryk and Bosker and Snijders. From the description, they use the old style commands -xtlogit-, which might present problems with the new -me- commands. I also don't know if it will hand 3-level models, but you can try it for yourself.

[Chris Byrne]

Nick Cox I apologise for the duplication. I'll know for future reference to link posts made in other forums. I had (obviously incorrectly) presumed it would reach different groups and increase the odds of some insightful answers.

Leonardo Guizzetti Thanks for your response, I will look into -mlt- and try your suggestion on leaving out units.

Thanks

Chris

[Nick Cox]

Chris Byrne That's OK, but the logic is faulty. First, you're an example of someone who posts in various places, so the existence proof is trivial: membership does overlap. Second, even if groups were distinct, except for yourself, people should still want to know what has been said elsewhere, either because answers would be interesting to them too, or because they have an answer yet don't want to duplicate what has already been said somewhere.

[Chris Byrne]

Nick Cox fair enough, Nick. Point taken