I need to estimate a probit with selection, and one of my regressors, income, is simulated. I am using MI to estimate the model, following the suggestion provided here:
that says that it is possible to treat simulated data as if multiply imputed.
The --heckprob-- command is not supported by --mi estimate--, and I'm forcing it with the --cmdok-- option (it should be OK, according to this post:
)
My data are at the individual level, and it is a repeated cross-section dataset. What I'm running is essentially:
where x1 is income, x2 another continuous regressor, x3 a continuous variable and x4 a variable that acts as state-year FE.
The estimation converges, and I get coefficients and standard errors. However, the second parameter of the F-statistics is missing, and of course also the value of the F statistic and its p-value:
I tried to remove the cluster option, and substitute heckprob with heckman, but the F statistic is still missing. The correlation between the regressors is not high, and I don't have any dummy variable among the regressors (x4 is not a dummy, but it acts as a FE.)
If I run --heckprob-- on one of the simulations only, i.e. I get rid of --mi estimate-- and only use --heckprob-- I get values for the Wald chi2.
What could be the cause for my problem? Could it be that heckprob is actually not suitable for MI? And, even if my F statistic is missing, can I trust the coefficients and the standard errors that I get from the estimations?
Thanks,
Rossella
HTML Code:
http://www.stata.com/statalist/archive/2012-11/msg00757.html
The --heckprob-- command is not supported by --mi estimate--, and I'm forcing it with the --cmdok-- option (it should be OK, according to this post:
HTML Code:
http://stats.stackexchange.com/questions/65678/using-heckman-in-combination-with-mi-estimate-stata
My data are at the individual level, and it is a repeated cross-section dataset. What I'm running is essentially:
Code:
mi estimate, cmdok post: heckprob y c.x1##c.x2 [pweight=weight], select(z = x3 x4 c.x1> 0##c.x2) vce(cluster psu)
The estimation converges, and I get coefficients and standard errors. However, the second parameter of the F-statistics is missing, and of course also the value of the F statistic and its p-value:
Code:
F( 4, .) = . Prob > F = .
If I run --heckprob-- on one of the simulations only, i.e. I get rid of --mi estimate-- and only use --heckprob-- I get values for the Wald chi2.
What could be the cause for my problem? Could it be that heckprob is actually not suitable for MI? And, even if my F statistic is missing, can I trust the coefficients and the standard errors that I get from the estimations?
Thanks,
Rossella