Skip to content

model selection methodology

2 messages · Alessandra Bielli, David Duffy

#
Dear List,

I am writing this here because I am using glmmTMB to run a zero inflated
model, but the question is possibly more general, so please feel free to
redirect me to another list/help page.

I want to test whether a treatment (a) has an effect on a
dependent variable (y), so I built a full model :

m1 <- glmmTMB(y ~ a + offset(log(b)) + (1|ID), data=x, ziformula =  ~ a,
              family="poisson")

I used the dredge function to generate a model selection table and the top
model did not include the treatment (a) in the conditional model nor in the
ZI model.

My usual way to proceed is to run diagnostics for the selected model and
conclude that the effect of treatment on my dependent variable is not
statistically significant.

My questions are:
1- is this the right way to proceed or should I check the diagnostics plots
BEFORE model selection?
2- it has happened to me that, while the full model converged, the reduced
model gave a convergence warning message. Considering that my only goal was
to test whether the effect of treatment was significant or not, how would
the convergence issue influence my conclusions?

I feel like these are very basic questions but very important because I do
not want to draw wrong conclusions. Thanks,

Alessandra
1 day later
#
Hi.

I suspect we would need more information. A real or simulated dataset of the correct form would be useful. In fact, if you could generate multiple such 
datasets, you could answer your own question about testing.
So. are there multiple records for each ID, or are you looking at both extra-poisson variation
and zero-inflation?
Personally, I think these are kind of cutting edge type questions. I would looking at
diagnostics from even simpler models first (say, poisson and quasi-poisson GLM, or hurdle model), and
the same model using a different package (say in STAN with uninformative priors).