xtheckman -could not calculate initial values; try specifying from()

[Benedetta Nussio]

Good morning all,

I am trying to use the command xtheckman as following:

xtset pidp wave
xtheckman overed4_cat_mod i.children_16_simp i.work_dimensions, select(jobstat = i.children_16_simp age)


and I get the error:" 237016 observations incorrectly specified as noncensored in select()".

In other similar cases, the problem stemmed from not having a variable in the second equation and not in the first one. But i added AGE just in the JOBSTAT equation. It should be fine.

Thanks a lot for the help,

PS. sorry for wrongly specify the title

[Jeff Wooldridge]

I see one problem right way: you don't have a variable in the selection equation that isn't also in the outcome equation. You need something that predicts section but not the outcome variable, and this should change over time.

A second issue is xtheckman command has several limitations in practice. It is a random effects estimator, which means it assumes all heterogeneity in the outcome and selection equations is independent of the explanatory variables. Moreover, it maintains no serial correlation in the idiosyncratic errors. I recommend adding the time averages of all exogenous variables in both equations, using the usual heckman command, and vce(cluster id). See my 1995 Journal of Econometrics paper on selection corrections with panel data. Oh, and you won't have estimation problems provided you find an outside variable to include in the selection equation.