Psacalc to estimate unobservable ability in regression capturing effect of formal vocational training

[Sayoree Gooptu]

I am regressing the binary variable NEET(not in education or training) on past vocational training including controls and state fixed effects. Since the vocational training variable can be affected by ability, I want to estimate the selection on unobservables, the criterion used by Oster (2019).

regress neet_t365_july i.urban i.male mpce i.unmarried i.religion i.social_group i.state i.edulevel##i.pastvoc365 [pweight = pop_weight] if majorstate==1, vce(cluster fsu)

psacalc beta 1.pastvoc365, delta(1) rmax(`rmax_val') * 4. Calculate the breakdown delta (the selection required to drive the treatment effect to zero) psacalc delta 1.pastvoc365, rmax(`rmax_val')

IS the process correct? How do I interpret the results? I am getting a negative delta

[Carlo Lazzaro]

Sayoree:
as yours seem to be an example of latent variable-led endogeneity, why not considering -ivregress-?

[Sayoree Gooptu]

I have checked for a suitable instrument. I thought of taking the district level availability of training centres. But, the correlation is insignificant. I am thinking of doing the Oster measurement of unobservables or an AIPW. Would that be fine?

[Carlo Lazzaro]

Sayoree:
what does the literature in your research field suggest in these occurrences?

[Sayoree Gooptu]

There are papers that have used PSM or doubly robust treatment effect. I want to apply for cross section data. I have done a multinomial logit regression for different activity status on vocational training. I am not able to figure out which robustness method to use - psm/aipw?

[Carlo Lazzaro]

Sayoree:
I would go -PSM- then.