If condition with xtreg appropriate?

[Anna Jongma]

Dear all,

For my master's thesis, I'm trying to look at the effect of a policy change on women's labour force participation, using a diff-in-diff analysis with individual fixed effects. For my main analysis, I use two waves of data (2009 and 2014, with the change happening in 2011). Having found no effect on female participation overall, I would now like to look at differences by subgroups, in particular by marital status.

My question is whether using the if condition after the xtreg command is appropriate in this setting; since I am looking at a sample of rather young individuals, many women get married between 2009 and 2014. If I understand correctly, using "if married==1" would include women in the 2014 analysis who were not married yet in 2009, unbalancing the panel. Whilst I know that xtreg can deal with unbalanced panels, I wonder whether it makes sense to include women who are married in 2014, but not in 2009, given that marriage and labour force participation are likely to be correlated. Should I just focused on the subsample of women who are married in both 2009 and 2014?

Many thanks in advance and apologies if this is not a 'proper' Stata question but rather a more scientific question!

Kind regards,

Anna

[Clyde Schechter]

Only you can answer this question. You need to clarify in your own mind what your research question is.

Are you asking about the effect of getting married on labor force participation? Or about the effect of being married?

If the former, then you need to select only those women who were unmarried in 2009 (along with the follow-up of these same women in 2014) and contrast the labor force 2014 labor force participation among those who married by 2014 with those who did not.

If the latter, then it would seem you would want to include all the women in your analysis, with marital status as a predictor variable in your model.

You also need to clarify whether you are asking whether marital status itself affects labor force participation, or whether it modifies the policy effect, or, perhaps both.

When you have a clear statement of your question, it is then just a matter of translating that into a statistical hypothesis and into code. If you want assistance with these latter steps, do post back once you have arrived at a clear statement of your research question.

[Anna Jongma]

Thank you very much for your quick reply! I realise that perhaps my statement was a bit unclear.

I am interested in the effect of being married on labour force participation. In my initial analysis, I indeed included all women (in fact, I included both men and women), with marital status as a predictor variable. However, in this model, my coefficient of interest (treatment * female * post dummy), was not significant.

As I believe that marital status affects labour force participation and alters the effect of the treatment for women, I thought a next step would be to run my same regression on the subsample of married people and see if the triple interaction term (treatment * female * post dummy) is significant here. However, I am unsure whether I should run this regression on the subsample of people who were married in both 2009 and 2014, or simply add the condition if married==1 to my initial regression (thus including a number of people who were married in 2014, but not in 2009).
I hope this clarifies things a bit and apologies if I misunderstood your explanation!

[Clyde Schechter]

If you believe that marital status both affects female labor force participation and alters the treatment effect, then you need a model that includes married##treatment##post, and I would do this analysis among females only (-if female == 1). The coefficient of this three-way interaction variable will tell you how much the treatment effect differs between married and unmarried women.

[Anna Jongma]

Thank you! I will try this!