heckman sample selection

[ashkan faramarzi]

Hi,
I am trying to use Heckman two-step model in order to take care of sample selection bias. I have a question regarding the second stage. here is my model:

heckman SC rmshi mc1 lptd lpbc iros fss dltts B2B res ci yd* id*, select(lpd= rmshi fss mc1 lptd iros fss dltts res ci yd* id*) twostep mills(mymills)

lpd is a bianry variable (1= main group, 0=the matched group)

my question is after running this model, should I take mymills (the inverse mills ratio) and use it to estimate below model or not:

reg SC rmshi mc1 lptd iros fss dltts B2B res ci yd* id* mymills if lpd==1

[FernandoRios]

Hi Ashkan,
I don t see why would you like to do that. After all, the outcome from:

reg SC rmshi mc1 lptd iros fss dltts B2B res ci yd* id* mymills if lpd==1

should be the same as what yo get from

heckman SC rmshi mc1 lptd lpbc iros fss dltts B2B res ci yd* id*, select(lpd= rmshi fss mc1 lptd iros fss dltts res ci yd* id*) twostep mills(mymills)

Perhaps there is something else that you want to do?
Fernando

[ashkan faramarzi]

Hi Fernando,
I read that first to calculate the inverse mills ratio and then include the inverse mills ratio into the second stage, thats why I want to do that. Also, the outcome is very different, with (reg SC rmshi mc1 lptd iros fss dltts B2B res ci yd* id* mymills if lpd==1) I get very significant results, but the heckman gave me nothing.

[FernandoRios]

Well, the command heckman does exactly what you describe.
1st Estimates IMR,
2nd. Estimates the outcome model including the IMR
Now, when you say the outcomes are very different. Keep in mind that the point estimates should be the same. However, because IMR is a predicted variable, the standard errors need to be corrected. This is internally done with heckman,twostep.
So, your regressions may be giving you significant results because they ignore the standard error correction.
HTH
Fernando