Heckman 2 Step Estimator

[Solomon Lin]

Hi,

I am using panel data to estimate the effect of drug use on wages and employment. I currently have a 3 equation simultaneous equation model with Log Wages, Employment and Drug Use as the 3 dependent variables. Unfortunately, many of the respondents answered saying their wages were 0 (and so I cannot compute Log Wages for them).

Question 1) How should I deal with those individuals whose wages are 0? Should I drop them? Or should I set Log Wage = 0 whenever Wage = 0?
Question 2) If I drop those individuals who have 0 wages, should I run a Heckman 2 Step Estimator?

Thanks!

[Phil Bromiley]

You might look at Gould's 22 August 2011 blog posting titled "Use poisson rather than regress: tell a friend", both the posting and the comments including related citations. One major issue is whether zero wages is just a continuation of wages or is something very different (like being unemployed). In the former, you might need to look at Tobit like estimators. The latter might lead you to a zero inflated estimator.

You may find that GSEM is good for such complex models (or CMP, a user written program).

Dropping unemployed/zero wage individuals would create serious selection issues you'd need to address. You probably don't want to set log wage=0 when wage ==0. This means you'd be recording higher logged wages for zero wage individuals than for individuals with wages below 1 (that have negative values when logged). I believe there is a Stata Journal entry (reproduced in the Stata Tips book) on such transformations as well.