Heckprob estimation : missing for Std. Err., z, P>|z| and IC in selection equation

[Flayols Alexandra]

Hello Stata users,

I have a problem with my estimation for my equation for men (with 3659 observations). I don't have this problem for other estimation with the same model.
I estimate an heckprob model : the probability to be employed or unemployed. With a selection equation about the participation on labor market. The problem appear on the selection equation for the country of birth (please see picture).
I search an explication in the stata help. I find a similar problem (maybe...) in the documentation [R] logit but isn't for an heckprob model.
Have you ever seen a similar problem ?

Thank you.

[FernandoRios]

Hi Alexandra,
My bet is that you have very little variation between employed unemployed by country of birth. For instance, if you do something like:
tab paysnaiss2 employed
My guess is that you will see a very small frequency for that particular country.
Solution, perhaps, would be to collapse that country with some other (neighbor country perhaps).

HTH
Fernando

[Flayols Alexandra]

Hi Fernando,
Thank for yor reply.

First I precise my stata command :
heckprob actifocc $xgenre if sexe==0 [pweight=poidsi], select(participe=$xgenre $xselect)

With actifocc==1 if employed, 0 if unemployed. And participe==1 if participe on labor market, 0 otherwise.

I have look :
tab paysnaiss2 actifocc
tab paysnaiss2 participe

Without weight I have only 9 observations for unemployed for this country of birth (total : 209) and 15 observation for the participation on labor market (total 144).
I think you have right but I would to avoid to collapse with other country... My problem concerns the South of European Union and I would to avoid to collapse with the the other country of European Union because generally distinct them because the effect is not the same. However, I am aware of the possible limitations of the data...

[FernandoRios]

Hi Alexandra,
I understand you prefer not to "collapse" the data, as the country characteristics might be different. One option you could try is to use the two step heckman, instead of the maximum likelihood (you are using now). The two step is often more flexible with near exact prediction, and the assumptions are also weaker (marginal distributions to be normal rather than joint to be normal).
You could also try excluding this country (which might entail arguing why not to include it)
Hope this helps.
Fernando

[Flayols Alexandra]

My model is a probit model that is why I use the heckprob command. If I understand right, you suggest used the heckman command with a twostep option ? But this command is for a linear model. So I cannot used it.

Thank for your suggestions.