Poisson FE Model Question

[Trent Claybaugh]

Hi All,

I have a panel data set with counts as the dependent variable, block groups as the geography of observation, and year as the time of observation.

When I just use year FE, my IV of interest is very significant and the coefficient looks as expected. However, when I use two-way FE (block groups & year), my IV of interest becomes insignificant.

Could this be due to a lack of within variation? There are 190 groups, each with 7 observations.

I then tried using year FE, but included an indicator variable for census tracts (i.tract) in the equation to mimic a FE and increase the amount of within variation that I'm able to exploit. When I tried this, my IV of interest was very significant again (not so much as with just time FE, but still good).

Does this final approach make sense or is the insignificance of the two-way FE model (block groups & year) trying to tell me something?

Thanks!

[Trent Claybaugh]

Or would it make sense to treat it as a pooled cross-section and use indicator variables for both (i.tract, i.year)? Joao Santos Silva

[Joao Santos Silva]

Dear Trent Claybaugh,

If I understand it correctly, you should get the same estimates when you use the FE for groups and include the year dummies, when you use FE for years and include the group dummies, or when you run a simple Poisson regression with dummies for years and groups. I assume that the difference in significance is because you are using robust standard errors and different ways of specifying the FE imply different assumptions about clustering. It is up to you to decide what is the best way to cluster the standard errors.

Best wishes,

Joao