I'm trying to fit a multinomial response mixed effects model on a largish (~150,000 row) data set and this package seems to be the only that can do it in reasonable time. But now I'm struggling to compare the model outputs with outputs from other mixed effects modelling packages, specifically lme4.
With an lme4 fitted model I can do ranef(m) and get the random effects, with the conditional variances matrix too. This can be fed into dotplot to get a plot of "confidence intervals" for the random effects. See help(lme4::ranef) for an example.
What I want to do is make these plots from mblogit output. It seems the $random.effects attribute has the random effects values, but I'm struggling to find or compute the equivalent conditional variances matrix.
My current working assumption is that the following two binomial models have the same model specification:
and if I can replicate ranef(gl) from mb then I'll be most of the way there. There are differences in these fitted models but I'm not sure if that's down to a completely different model specification or just a different optimiser stopping in essentially the same flat point of likelihood space.
I'm trying to fit a multinomial response mixed effects model on a largish (~150,000 row) data set and this package seems to be the only that can do it in reasonable time. But now I'm struggling to compare the model outputs with outputs from other mixed effects modelling packages, specifically lme4.
With an lme4 fitted model I can do
ranef(m)
and get the random effects, with the conditional variances matrix too. This can be fed intodotplot
to get a plot of "confidence intervals" for the random effects. Seehelp(lme4::ranef)
for an example.What I want to do is make these plots from
mblogit
output. It seems the$random.effects
attribute has the random effects values, but I'm struggling to find or compute the equivalent conditional variances matrix.My current working assumption is that the following two binomial models have the same model specification:
and if I can replicate
ranef(gl)
frommb
then I'll be most of the way there. There are differences in these fitted models but I'm not sure if that's down to a completely different model specification or just a different optimiser stopping in essentially the same flat point of likelihood space.