Zeros, Tobit, Poisson and Fixed effects

[junlei luo]

Hi,
I have export data with about 50 observations from 3000 (300 firms and 10 years) that are zero and so are missing in log form. The zeros are important and also I think the only option in my model is to keep the log form. The independent variables have percentages and other logs of large numbers, and year and firm fixed effects, and the regular fe regression is:

xtreg  ln_y perc_1 ln_x i.year, fe

But does it make sense to simply increment these zero exports to 1 (log(1)) if they end up influencing the results too much as outliers? I was thinking to use the smallest nonzero export number as the cutoff (14.2 in log), and change all the zeros to this cutoff. Does the Tobit with lower cutoff make more sense for a dependent variable like exports, and can I include dummy variables (i.year i.firm) for the fixed effects like this？

tobit ln_y perc_1 ln_x i.year i.firm, ll(14.2)

How would interpretation of the coefficients be compared to the fixed effect regression? Is Poisson regression something I should consider or does Poisson regression not make sense since exports is not usually a poisson distribution?

Thank you for your time!

[Joao Santos Silva]

Dear junlei luo,

I would say that Poisson regression with robust (clustered) standard errors is the only thing that makes sense in this case. The data does not have to be Poisson for the estimator to be consistent. If you do a quick search you will find a lot of literature on that, including this paper.

Best wishes,

Joao