Fixed Effects Poisson Cluster at Higher Level

[Jin Zhang]

Dear Community,

I need to run a Poisson regression, control for firm fixed effects, year fixed effects, and cluster standard errors at state level. The problem is that firms may relocate to other states during our sample period, so firms are not nested within clusters. Is there any command that can be used in this specific situation?

Thank you.

[Carlo Lazzaro]

Jin:
welcome to this forum.
As companies can relocate during your sample period (by the way: from your description you're seemingly dealing with panel data. Assumed that you would go -fe-, I guess that, being aware that -xtpoisson- allows conditional -fe- only, you are challenging yourself with a pooled poisson regression considering -firm- as panel indicator) why don't clustering your SEs at firm level?

. use http://www.stata-press.com/data/r15/ships

. poisson accident i.ship op_75_79 co_65_69 co_70_74 co_75_79, exp(service) irr cluster(ship)

Iteration 0:   log pseudolikelihood = -120.33891
Iteration 1:   log pseudolikelihood = -68.554441
Iteration 2:   log pseudolikelihood = -68.280965
Iteration 3:   log pseudolikelihood = -68.280771
Iteration 4:   log pseudolikelihood = -68.280771

Poisson regression                              Number of obs     =         34
                                                Wald chi2(3)      =          .
Log pseudolikelihood = -68.280771               Prob > chi2       =          .

                                   (Std. Err. adjusted for 5 clusters in ship)
------------------------------------------------------------------------------
             |               Robust
    accident |        IRR   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        ship |
          2  |   .5808026   .0342346    -9.22   0.000     .5174349    .6519306
          3  |    .502881   .0147863   -23.38   0.000     .4747197     .532713
          4  |    .926852   .0419759    -1.68   0.093     .8481265    1.012885
          5  |   1.384833    .034818    12.95   0.000     1.318245    1.454784
             |
    op_75_79 |   1.468831   .1245519     4.53   0.000     1.243921    1.734406
    co_65_69 |   2.008002   .1401996     9.98   0.000     1.751189    2.302478
    co_70_74 |    2.26693   .2617247     7.09   0.000     1.807857    2.842577
    co_75_79 |   1.573695   .2789904     2.56   0.011     1.111774    2.227536
       _cons |   .0016518   .0001088   -97.22   0.000     .0014517    .0018795
 ln(service) |          1  (exposure)
------------------------------------------------------------------------------
Note: _cons estimates baseline incidence rate.

NB: Wald test unreported due to clustered standard errors (see help j_robustsingular).

.

[Jin Zhang]

Hi Carlo,

Thank you so much for your prompt reply and detailed example. I am studying the impact of a state-level regulation (a regulation passed in several states of the United States) on firm's innovation efficiency. So I have to cluster my standard errors at the state levl.

Best,
Jin

[Carlo Lazzaro]

Jin:
thanks for providing further details (by the way: it's far better to explain everything right from the start and report what you typed and what Stata gave you back via CODE delimiters, as recommended by FAQ).
If you go pooled -poisson- (and you do not have to -xtset- your data beforehand), you can try to cluster your SEs at state level and see what happens to your coefficents.