Model selection/comparing model fit in a quasi likelihood model (i.e., Poisson w/robust standard errors)

[Matthew Barnett]

Dear all,

I am modelling household water use data (i.e., total annual household water use). The distribution of this variable is log-normal -- but rather than log-transforming this outcome variable, I have elected to use a Poisson model w/robust standard errors (see: https://blog.stata.com/2011/08/22/us...tell-a-friend/).

My question here is concerning model fit/comparison across models. I am wondering if AIC and BIC (as estimated by Stata's estat ic command) are appropriate as a measurement for model fit, given that this modelling approach is based on log psuedolikelihood? And if not...which criterion/statistics may be used to assess/compare model fit for such a model?

Thanks very much for your time,

Matt

[Matthew Barnett]

Could I just use the log pseudolikelihood itself to compare fit across models (i.e., adding variables and interactions across a nesting structure)? Or perhaps a “pseudo-deviance” (i.e, the log psuedolikelihood multiplied by -2)?

[Joao Santos Silva]

Dear Matthew Barnett

I am afraid none of those methods are valid in this context but you can simply use t-statistics and F-tests.

Best wishes,

Joao

[Matthew Barnett]

Joao Santos Silva thank you very much! That is very helpful. I was trying to do lrtest which of course wasn’t working because of the quasi likelihood estimation. Test and testparm are working wonders though.

Your help is greatly appreciated!