Propensity Score Matching on Panel Data

[Duke Alabanza]

Hi all,

I'm currently looking to perform a propensity score matching (PSM) estimator on panel data. My study consists of 39 countries over a 23 year period (1990 - 2012), and I'm trying to ascertain the impact of my treatment variable, which is a particular policy. To give you an idea of the treatment variable data, if the Czech Republic implemented this policy in 2002, it would be assigned a dummy value of 0 before 2002 and a 1 from 2002 to 2012. I was wondering if it would be appropriate to conduct PSM on this dataset as it is? In this case, for example, the Czech Republic in 2002 would be matched with several countries that are most similar to it (based on my control variables), with the only difference being that other countries did not implement this policy. Could I perhaps specify my matches, so that the Czech Republic in 2002 is only able to match with observations in 2002 (this way I could account for heterogeneity across time)? I would really appreciate your help. Thank you very much.

Duke

[Marcos Almeida]

Hello Duke,

Welcome to the Stata Forum. This seems to be an interesting query. Hopefully you'll get great insights on the matter. That said, as you remarked, PSM is based on this principle: you have a "treatment" variable (usually binary), the dependent variable and a whole set of covariates. In short, treatment 1 is compared to treatment 2, but both treatments should ideally be applied to a fraction from the same sample and followed for a similar period. When a treatment has a gap of, say, 10 years, you may not have the same sample anymore. What is more, several historical factors may have played an important role, to the extent that measuring the treatment effect is eventually spoiled.

Best,

Marcos

[Charlie Joyez]

Duke,

To complete Marcos answer, I'd say that PSM is probably not adapted to your issue, because of the temporal lag.
The proper estimation strategy to your question would rather be a difference in difference, which will in addition to treatment effect, control for the time trend of none-treated group (other countries). This way you capture the difference between the treated and non-treated group, cleaned from temporal trend (assumed the temporal trend is similar between treated and non treated groups)

Some user written program run diff-in-diff models (type search differences in differences).

Hope this helps,
Charlie

[Duke Alabanza]

Dear Marcos and Charlie,

Thank you very much for your insights--much appreciated. I've looked into the difference in differences model in the past, but had a little difficulty grasping its application in the context of my model due to the staggered entry of my treatment variables. Anyhow, I'll have another look into it right now to see how this model can address trends across time. I'm also using a fixed effects model to control for heterogeneity across countries and years, so hopefully that works out. Thanks, again.

Duke