Poisson Regression with non integer dependent variables

[Stefan Greim]

Dear Forum,

Does any fellow forum member happen to know if it is possible in Stata to process a poisson regression with non-integer dependent variables, or would a two step model (Heckman) potentially be more advisable?

Background: I am in a process to test the causality between institutional set-up and venture capital (VC) activities in 54 African countries for a time series of 10 years. The dependent variable is number of VC deals in a particular country in a particular year. As there were no VC activities at all in various countries and several years, the value of the dependent variable is zero in more than 50%. To avoid a bias towards zero, OLS regression seems not appropriate. A poisson regression, however, seems to work for integer values only. But if I just use the number of deals without denominating it by the number of inhabitants in a country, the outcome will favor size of a country inappropriately. If I use the dependent variable number of deals/inhabitants, the dependent variable is more accurate, but not an integer figure anymore.

Many thanks indeed,
Stefan

[Rich Goldstein]

yes; you can try yourself to see that it works; but also see the following blog: https://blog.stata.com/2011/08/22/us...tell-a-friend/

[Stefan Greim]

Thank you. This is truly helpful.

[John Mullahy]

For what it's worth...

sysuse auto
poisson gear_ratio weight mpg, vce(robust)

Results:

. poisson gear_ratio weight mpg, vce(robust)
note: noncount dependent variable encountered; results correspond to an exponential-mean
      model rather than a poisson model.

Iteration 0:  Log pseudolikelihood = -111.54809  
Iteration 1:  Log pseudolikelihood = -111.54809  

Poisson regression                                      Number of obs =     74
                                                        Wald chi2(2)  = 120.06
                                                        Prob > chi2   = 0.0000
Log pseudolikelihood = -111.54809                       Pseudo R2     = 0.0129

------------------------------------------------------------------------------
             |               Robust
  gear_ratio | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
      weight |  -.0001508   .0000208    -7.26   0.000    -.0001915   -.0001101
         mpg |  -.0002266   .0032491    -0.07   0.944    -.0065947    .0061416
       _cons |   1.557136   .1248935    12.47   0.000     1.312349    1.801923
------------------------------------------------------------------------------

While I think the alert (in red) is clearer than the one that was used in earlier Stata versions I think it could be clearer still were it to read something like:

note: noncount dependent variable encountered; results correspond to an
exponential-mean model rather than a Poisson-probability model.

or

note: noncount dependent variable encountered; results correspond to an
exponential-mean model rather than a Poisson model of a discrete probability distribution.

or

note: noncount dependent variable encountered; results correspond to an
exponential model of a conditional mean rather than a Poisson model of a discrete probability distribution.

Including in the warning note a hyperlink to the Stata Blog entry in #2 might also be valuable.

[Stefan Greim]

Very interesting. Thank you