heckman selection model

[Olivia Johns]

heckman lwage educ exper expersq nwifeinc age, select( inlf= educ exper expersq age faminc )
heckman lwage educ exper expersq nwifeinc age, select( inlf= educ exper expersq age faminc ) twostep

(inlf= takes a binary value, either you are in labor force or not)

I ran the Heckman selection model, using both ML and twostep to estimate the returns to education,
but i get completely different results.

For ML method, my t-stat is stat.sig.different from zero, but
for twostep method, my t-stat on educ became highly insignificant!

How can they give out such different results?
My dependent variables and exclusionary variables for both methods are identical.

Help please?

[Alfonso Sánchez-Peñalver]

Hi Olivia,

the difference in estimation methods is that maximum likelihood assumes a bivariate normality, while the two-step method is more flexible, since it uses OLS in the second step, and thus only assumes univariate normality. If any of the two residuals are not normal, ML will provide inconsistent estimates, whereas the two-step method will only provide inconsistent estimates if the selection errors are non-normal.

I hope this helps.