xtmlogit & sample selection

[Laurence Lester]

This may be more a methodology than Stata question but in the help for xtmlogit most of the examples are estimating labour market status (not in labour force (NLF), employed, unemployed (Participants). I'm puzzled about this. My understanding is that this should be a selection model (Heckman): (i) model NLF vs in LF; (ii) model Participants (ie Emp vs Unemp) with selection correction. Can someone tell me what I'm missing please.
Thanks
Laurence

[Andrew Musau]

I do not follow what the argument is here. If you are talking about examples involving the empirical model of married women's labor supply, then the outcome is "wage" in examples involving heckman. The reason here is that wage is observed only for women who are employed. On the other hand, the outcome in multinomial logit models is employment status with categories such as "not in the labor force", "work part-time" and "work full-time". We precisely know who is in each of these categories. So what is the issue with investigating factors associated with being in a particular category?

[Laurence Lester]

Thank you Andrew. I'm hesitant to continue this as it's not a Stata issue really but since you were kind enough to respond here are my thoughts on your reply.

Perhaps I do not understand completely. Yes, when wage is an outcome it is clearly a selection issue. But I thought the same reasoning applied to the model of employed vs unemployed.
I thought that in labour market theory it was assumed individuals chose to be NLF or labour force participants prior to ‘choosing’ to be employed or unemployed, so only participants are observed in the second stage.
This is a 'selection problem' because participants are systematically different to NLF, or as Rubin (1978) put it "If people with different X differ in unobserved ways we have 'non-ignorable selection'."

Or to put it another way specific to this issue, in samples where individuals' decision to participate in the labour force may not be random but are a consequence of unobservables. (E.g., characteristics such as education level, health status, etc are more likely to influence participate in the labour force, so a model for Employed vs Unemployed that ignores this non-randomness may produce biased estimates of labour force participation and employment outcomes.

For example, Christelis Georgarakos, and Jappelli (2019) in “The impact of education on labour force participation and unemployment in Greece: evidence from EU-SILC” specifically apply the Heckman selection model to exactly the case I am referring to..

Regards
Laurence

[Andrew Musau]

Two issues:

1. If you have cross-sectional data and are interested in modeling whether an individual is employed vs. unemployed based on a number of observables, then you may rightly argue that that employment is observed only if one is in the labor force. Provided that an exclusion restriction exists, you can estimate a probit model with sample selection using e.g. heckprobit. However, if one considers that not in the labor force, employed and unemployed are distinct categories, then there is nothing wrong with running a multinomial logit model. The purpose of this model is to predict the probability of an individual choosing one category over the others, based on a set of predictors. This enables one to understand the factors that influence these choices.

2.With panel data, you have repeated observations of the same individuals over time. Therefore, there are several instances where individuals switch betwen labor force participation and non-participation.

[Laurence Lester]

Dear Andrew
Thanks again for the effort. I feel we are both repeating ourselves (and/or I have not made myself clear). I am happy with my approach and as this is not a Stata issue I am happy to leave it at "thanks".
Regards
Laurence