Question about zero-inflated Poisson glmer

I have now played around more with the data an the models both using lme4
and glmmTMB.

I can report the following:

Modeling the data with a zero-inflated Poisson improves the model
significantly. However, when calling predict and simulating rpoissons, I
end up with nearly no values that are zero (in the original data there are
96% zero).

When I model the data with overdisperion by including an observation-level
random effect, I can also improve the model (not surprisingly due to the
random effect). When I predict outcomes by ignoring the observation-level
random effect (in lme4), I receive bad prediction if I compare it to the
original data. While many zeros can be captured (of course), the positive
outcomes can not be captured well.

Combining zero inflation and overdispersion further improves the model, but
I can only do that with glmmTMB and then have troubles doing predictions
ignoring the observation-level random effect.

Another side question:

In lme4, when I do:

m = glm(x~1,family="poisson")
rpois(n=len(data),lambda=predict(m, type='response',re.form=NA)

vs.

simulate(1,m,re.form=NA)

I receive different outcomes? Do I understand these function wrongly?

Would appreciate some more help/pointers!

Thanks,
Philipp

2016-06-24 15:52 GMT+02:00 Philipp Singer <killver at gmail.com>:

Question about zero-inflated Poisson glmer

Thread (29 messages)