Heckpoisson error - could not be evaluated

[Kostas Dellis]

Dear Stata Users,

I am trying to fit a Poisson model correcting for Sample selection using the heckpoisson command in Stata 17.0. I have a cross-sectional dataset for EU firms participating in meetings with the European Commission.

My dependent variable is the number of meetings held (meet) and the selection dependent variable is a dummy variable capturing whether any meeting was held for the specific firm (meet_d).

My explanatory variables include a dummy variable indicating whether a firm is in the top 2500 R&D firms (toprd_firm), a dummy variable indicating whether a firm has received an EU grant (grant) and the log of mean turnover approximating firm size (lmean_turn_m)

* Example generated by -dataex-. For more info, type help dataex
clear
input float i_lobby int meet float(meet_d toprd_firm grant lmean_turn_m)
2034  1 1 0 0         .
 501  0 0 1 0         .
1402  3 1 0 0         .
2389  0 0 0 0         .
1552 14 1 0 0 15.396323
 206  2 1 0 1 13.301535
 600  0 0 0 0  10.76539
2409  . 0 0 0         .
2251  0 0 0 0  9.528091
1349  0 0 1 0 10.567322
2307  0 0 0 1 13.585784
2154  0 0 0 0  12.94011
1242  0 0 0 1 11.215186
1751 25 1 0 0         .
 565  0 0 0 0         .
end
label values grant g
label def g 0 "No Grant", modify
label def g 1 "Grant Receiver", modify

Typing the command:

# delimit ;
heckpoisson meet toprd_firm grant  lmean_turn_m i.cnt i.nace2, iter(100)
select(meet_d = toprd_firm lmean_turn_m);

I get

initial:       log likelihood =     -<inf>  (could not be evaluated)
could not find feasible values

Changing the maximizing method or using the difficult option does not help.

Fitting a Poisson model works fine and so does a Heckman model with the same variables, however given that my dependent variable is count it is not the appropriate way to go.

[Carlo Lazzaro]

Kostas:
from your example, the most complete viable code is:

. heckpoisson meet toprd_firm grant  lmean_turn_m  , iter(100) select(meet_d = toprd_firm lmean_turn_m)

that gives me back, after -convergence not achieved- warning:

Poisson regression with endogenous selection    Number of obs     =          8
(25 quadrature points)                                Selected    =          2
                                                      Nonselected =          6

                                                Wald chi2(2)      =      14.69
Log likelihood = -8.336869                      Prob > chi2       =     0.0006

------------------------------------------------------------------------------
        meet | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
meet         |
  toprd_firm |          0  (omitted)
       grant |  -1.754673   .7451533    -2.35   0.019    -3.215147   -.2941995
lmean_turn_m |   .0912881   .0272873     3.35   0.001     .0378058    .1447703
       _cons |   1.233498          .        .       .            .           .
-------------+----------------------------------------------------------------
meet_d       |
  toprd_firm |  -.0018127          .        .       .            .           .
lmean_turn_m |  -.0002323          .        .       .            .           .
       _cons |   -.002639          .        .       .            .           .
-------------+----------------------------------------------------------------
     /athrho |   9.235314          .        .       .            .           .
    /lnsigma |  -9.713146   5221.761    -0.00   0.999    -10244.18    10224.75
-------------+----------------------------------------------------------------
         rho |          1          .                            -1           1
       sigma |   .0000605   .3158284                             0           .
------------------------------------------------------------------------------
Wald test of indep. eqns. (rho = 0): chi2(1) =        .   Prob > chi2 =      .
Warning: Convergence not achieved.

..

[Kostas Dellis]

Thank you for your prompt reply Carlo

The fact that I am not getting the output has to do with adjusting Stata memory?

[Carlo Lazzaro]

Kostas:
my gut-feeling tells me that -intpoints(#)- option should be considered.

[Kostas Dellis]

Thank you Carlo

Playing around with the inpoints(#) option does not yield results

However, removing industry fixed effects (i.nace2) does the trick

It's not ideal, however the world is almost always second best