Heckman maximum likelihood method non-convergence issue

[Gauss Cheung]

Hello experts,

I am new to this forum and would really appreciate any guidance on the captioned question.

Specifically, I am using the Heckman package (heckman) with the maximum likelihood method (default method, not the two step method). Unfortuntately I encountered the problems with an aparent non-concave function and non-convergence (see the attached picture). May I know what does it mean and how to solve this issue?

I read the heckman documentation and followed the suggestions in attempts to mitigate these issues but they seem to be not working.

My specification is:

heckman y x1 x2 i.id i.time, select(x1 x2 x3 i.id i.time)

where x3 is the exclusion restriction variable in the first stage.

Other alternative specifications I tried include:

heckman y x1 x2 i.id i.time, select(x1 x2 x3 i.id i.time) rhosigma vce(oim) difficult nolog iterate(200) technique(bhhh)

Thank you in advance!

[Maarten Buis]

i.id i.time sounds to me like an awful lot of parameters to estimate, maybe too much for your data. Maybe you get more luck with twostep? Regardless, you are probably at the edge (or over it) of what is feasible with your data.

[Jeff Wooldridge]

It’s not just a computational issue. This suffers from the incidental parameters problem with N large and T not so large. I proposed an alternative based on correlated random effects in my 1995 Journal of Econometrics paper. I think there’s a community contributed command for it, too.

[Jeff Wooldridge]

The command is xtheckmanfe, written by Fernando Rios-Avila.