Using fixed effects with zero-truncated count models

[Anne Byrne]

Hello! I am attempting to estimate a model using a panel of data that includes between 1 and 13 years of observations for over 40,000 households (i.e. unbalanced). The data is well-suited for Poisson or negative binomial models, but given that it only includes positive values for the outcome, I think zero-truncated models would be best. I don't see an option within the xtpoisson/xtnbreg suite to account for the zero truncation. Similarly, for the ztp and ztnb, I don't see a way to do fixed effects other than to hard code them as dummy variables. Unfortunately that seems to cause a problem with my max variables because the panel is rather large. I can try bootstrapping, but I'm wondering if there's another way. Any advice is very much appreciated.

[Camille Antinori]

Hi Anne, I am also trying similar: xtnbreg, weighted, and with zero-truncation. So far, I have found nothing yet. If I do, I will let you know. How did you ever resolve your issue?

[Maxence Morlet]

Could you show the distribution of your outcome?

From what I read, you could try the user-written ppmlhdfe command, and cluster your standard errors to account for over or underdispersion (negative binomial models are not known for robustness).

So in sum, I would go for Poisson with clustered standard errors.

[Jeff Wooldridge]

Agree with Maxence -- whether or not there are zeros.

[Maxence Morlet]

Jeff Wooldridge I learn from the very best