Poisson pseudo maximum-likelihood in a Diff-in-Diff analysis

[Ali Taleb]

Hi Everyone,

I am performing a Diff-in-diff analysis using an outcome with many zeros in the panel data. For this reason, I need to use the Poisson pseudo-maximum-likelihood estimation. I have searched online, especially on stata forum in order to know how it works on in stata but I didn't find anything regarding this issue.

Would anyone help maybe by sending a good source or by sending the stata code?

This is the regression code I am using at the moment:

xi: areg MW treatment i.year Grad Bach  business Age Unemp Republic Democrats   Electricity income, absorb(county) vce(cluster county);

Thanks,
Ali

[Joao Santos Silva]

Dear Ali Taleb,

DiD with PPML works almost as in a linear model, just keep in mind that the dependent variable should not be logged and that the regressors will be inside an exponential function (so you may want to log some of them). Notice, however, that the interpretation is a bit different: you get more a ratio of ratios than a dif in dif.

Best wishes,

Joao

[Ali Taleb]

Dear Joao,

Thank you for your reply. That's right, there are some papers that used this method (with regressors inside an exponential function). However, I would like to know what type of variables should I log? In fact, zeros count for 17% of the data but I am not sure if there is a threshold regarding the number ( or percentage) of zeros accepted in an OLS regression.

I used the "ppmlhdfe" code instead of the "areg". It gave me different results actually in terms of significance level, R square, and even coefficient.

Best regards
Ali

[Joao Santos Silva]

Dear Ali Taleb,

If you have non-negative data, it is better to avoid OLS even if the percentage of zeros is low. The fact that ppmlhdfe leads to different results just confirms that it is important to account for the nature of the data.

About which variables to log, that is up to the analyst.

Best wishes,

Joao

[Ali Taleb]

Thank you Joao!