Iterative not-concave problem random panel Poisson regression

[Sishi Wu]

Dear Statalisters,
I am currently doing various xtpoisson regressions on my master's thesis data. Most of my regressions have gone without any problems (I've made several models with IVs and interactions). However, now in my joint regression of the two IVs, I get the not concave message on stata. Strangely enough, I don't get this message when I add the interaction term IV to it and study the interaction.
To illustrate this, here is the code, not significant only ①, does anyone know how this is possible? What steps are being taken to address this problem?
①xtpoisson sales IV1 IV2 $control ,re vce(r)
②xtpoisson sales IV1 $control , re vce(r)
③xtpoisson sales IV2 $control ,re vce(r)
④xtpoisson sales IV1 IV2 IV1*IV2 $control , re vce(r)
⑤xtpoisson sales IV1 IV2 MV IV1*MV IV2*MV $control , re vce(r)
And is it necessary to add IV1*IV2 to the study ⑤?
Any help would be greatly appreciated

[Carlo Lazzaro]

Sishi,
welcome to this forum.
You should give a true and fair view of the data generating process you're investigating.
Statistical significance is an optional.

[Sishi Wu]

The Hausmann test says I should use the random effect, but the random effect appears notconcave, I don't understand why, or do I need to choose to use the fe effect

[Joseph Coveney]

The Hausmann test says I should use the random effect, but the random effect appears notconcave,

I take it that you conducted the Hausman test test on the less parsimonious model, the one that, as you mention in #1, includes includes the interaction term. (It's a little confusing in that your screenshot of the nonconverging example shows a predictor called interaction.)

I don't understand why, or do I need to choose to use the fe effect

Well, you have more than 800 groups and more than 10 000 observations. If your concern in conducing the Hausman test is about the relative efficiency of the fixed-effects estimator, then with that sample size and fewer than a dozen predictors, perhaps you can relax at little.

With the default gamma distribution your random effect seems to have collapsed to zero. Have you considered trying the normal option as a Hail Mary pass?

[Joao Santos Silva]

Dear Sishi Wu,

The Poisson RE estimator makes a distributional assumption about the individual effects that is difficult to justify. This contrasts with the FE estimator that is a very robust estimator. So, I would forget about the RE estimator and just use FE. On why your estimator does not converge, my guess is that you have conditional under-dispersion, but that does not affect the suitability of the FE estimator.

Best wishes,

Joao