Poission regression using fixed effect while clustering

[Tariq Abdullah]

Hello,

My dependent variable is number of incidents in a region over different years. My region is unit and I'm trying to cluster also at unit.

To run xtpoission regression , is the following command correct where I wanted to include fixed effects of unit and year while clustering at unit?

xi: xtpoisson yvar xvar controlvar policy i.year i.unit, robust i(unit) fe

Here the policy variable either takes the value 0 or 1. I'm trying to see how the policy affects the dependent variable y.

[Erik Ruzek]

You can omit i.unit from the main portion of the xtpoisson statement. The i(unit) option takes care of that for you. Also, unless you are using a much older version of Stata, you do not need the xi: prefix. As of Stata 11, the program natively handles indicator variables correctly as long as they have i. in front of them.

[Tariq Abdullah]

But, when I use the following command I do not see the unit fixed effect part in my regression at unit level on the log result.

xtpoisson yvar xvar controlvar policy i.year, robust i(unit) fe

when I use ols regression the regression takes at least 30 min to perform since it's clustering at 4436 units. Also on the log result i can see getting the fixed effects of 6000 units.

But the xtpoission result comes within two seconds without any evidence of fixed effects of those 4436 units. I just see the year fixed effects. However this part is mentioned (Std. err. adjusted for clustering on unit)

This is OLS results log

Linear regression                               Number of obs     =     25,926
                                                F(19, 4435)       =          .
                                                Prob > F          =          .
                                                R-squared         =     0.8636
                                                Root MSE          =     10.428

                                   (Std. err. adjusted for 4,436 clusters in unit)

This is my xtpoission regression's log

note: 80 groups (80 obs) dropped because of only one obs per group
note: 2435 groups (13772 obs) dropped because of all zero outcomes

Iteration 0:   log pseudolikelihood = -38287.836  
Iteration 1:   log pseudolikelihood = -37988.815  
Iteration 2:   log pseudolikelihood = -37988.714  
Iteration 3:   log pseudolikelihood = -37988.714  

Conditional fixed-effects Poisson regression         Number of obs    = 12,074
Group variable: unit                              Number of groups =  1,921

                                                     Obs per group:
                                                                  min =      2
                                                                  avg =    6.3
                                                                  max =      7

                                                     Wald chi2(20)    =  38.22
Log pseudolikelihood = -37988.714                    Prob > chi2      = 0.0083

                                       (Std. err. adjusted for clustering on unit)