Making sure something is kosher
This is looking at patient outcomes with doctor as random effect Some patient demographics are missing I figured this wasn't going to be as easy as pool glmer imputations from mice, then use se.ranef from arm to get standard errors on random effects I was hoping to get away with 1000 bootstrapped samples, 5-10 imputations each bootstrap Too optimistic?
On Wednesday, August 12, 2015, Paul Johnson <pauljohn32 at gmail.com> wrote:
I think you mean to ask, how does one pool bootstrapped confidence intervals constructed from 20 imputations? Same question about variance estimates, or ICC. Appears now, each of those is a separate battle. Similarly, any anova() test done separately. Rubin's rules great for MLEs. Not so much for all that other stuff that we do. VanBuren worked out the R^2 and some papers exist on model chisquare. if you find one on CI, let me know. I'm wondering if you could simply aggregate runs and build a CI. if not, appears you need better math stat training than I have. One project here explored that. If you are in the "I need 100 imputations" crew and you want 5000 bootstraps, well... get a cluster. I'll ask that team what they know. pj Paul Johnson http://pj.freefaculty.org On Aug 11, 2015 7:06 PM, "Mitchell Maltenfort" <mmalten at gmail.com <javascript:_e(%7B%7D,'cvml','mmalten at gmail.com');>> wrote:
I realize combining multiple imputation with random effect estimation is
dodgy.
Supposing I use mice with glmer to do the model fits, and then bootstrap
to
get confidence intervals on the random effect estimates.
Is that a clean way to do it? If not, what's recommended?
Thanks!
--
Sent from Gmail Mobile
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org <javascript:_e(%7B%7D,'cvml','R-sig-mixed-models at r-project.org');> mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
Sent from Gmail Mobile [[alternative HTML version deleted]]