DiD and matching using repeated cross sections

[Pablo Evia]

Hi everyone, I am dealing with two separe household surveys of consecutive years, and trying to analyze a policy change in the middle of those two years. I am trying a DiD model with covariates, and I was wondering how I could use a DiD model with previous matching to account for plausible differences among treatment and control groups. I am aware of the "diff" command in Stata, but DiD + matching with it needs panel id's, which of course I don't have. Any suggestion? Thanks in advance!!!

Pablo

[Clyde Schechter]

Can you explain what you mean by "previous matching" in a context where you have two consecutive cross sections?

Generically, if the policy change in the middle of those two years applied to everybody, you can't do a DID analysis. A DID analysis requires a group of people who are initially unexposed and then become exposed to the policy, and another group who are unexposed throughout the same observation period. Your data don't sound like they fit that design, although perhaps I am missing something here.

[Pablo Evia]

Dear Clyde, many thanks for your response. Sorry for not answering before, but I was not aware of how the Statalist sends me notifications when I receive a reply in any of my posts. Anyway, what I mean by "previous matching" is that I don't take the whole control group, but only those that are very similar in their observables to the treatment group, and for that I use a matching procedure. I apply the procedure previously to the policy change (strictly speaking, there are not treatment and control groups before the treatment, but I know who will be treated after the policy change) and after the policy change. My data consists in household surveys (2 rounds) for years 2007 and 2008. Policy change was in the first day of 2008. Thanks and warm regards.

[Clyde Schechter]

OK, now I understand. I don't myself use the -diff- command. I just do it from the basic official Stata commands, but I imagine that the approach using -diff- would be similar.

You need a matched pair identifier: some variable that takes on a the same value for both members of the pair (or all members of the tuple if it's n:1 matching) you have matched, and distinct values for different pairs (tuples). Then you use that just as if it were a panel variable. We'll call that one pair_id. I assume you also have a household id, hh_id. I assume you also have a variable that indicates membership in the treatment or control group. Let's call that study_group. And you have a variable that marks the pre- and post-policy years: let's call that era. FInally there is some outcome whose response to the policy change you are trying to estimate, call it outcome.So it looks something like this:

mixed outcome i.study_group##i.era || pair_id: || hhid:

If your outcome is not continuous, you can use some other suitable -me- type command. You can include other covariates, of course, and if you want robust variance estimation that is available as well.

[Pablo Evia]

Thanks for your answer Clyde. Warm regards,

Pablo