Propensity Score Weights - Differ by Period

[Korstiaan Wapenaar]

Dear all,

Apologies if my query format is incorrect as this is my first post.

I am currently conducting a difference-in-difference analysis on panel data (2 pre and 2 post treatment periods). The goal is a doubly robust DiD regression, weighting for a propensity score/inverse probability of treatment. The analysis regards the provision of piped water to households and the subsequent effects on child health status.

Having perused statlist for a number of hours this evening, it is apparent that employing an xtreg, fe is superior to the standard reg command, due to the panel data format of the data (subsequently individual unobservables are differenced out, and not simply aggregate unobservables).

A number of methods to estimate the propensity score have been made use, such as the popular psmatch2 as well as standard logistic regression on the probability of treatment.

The issue concerns the weighting requirements of the xtreg, fe model - it must be constant within the group variable through time. Both the Psmatch2 package and the logistic regression produce a propensity score for treatment reflective of the characteristics specific to that period (i.e. the propensity scores between periods 1 and 2 differ as the characteristics used to compute them change through time).

How might one calculate a singular pre-treatment propensity score on the basis of pre-treatment characteristics, regardless of period. If rephrased, how might one generate a propensity score that, for each category, does not differ between periods 1 and 2. I have considered averaging the two, though this seems archaic.

All support is much appreciated

Thanks and regards

Thanks and Regards

[David Radwin]

I can't speak to the xtreg aspect, but if you think the two sets of pretreatment covariates are meaningfully different, and are related to selection into treatment, why not use both sets in estimating the propensity scores in the same equation?

Conceptually, the model would be like

treatment = b₀ + b₁*covariate1_t1 + b₂*covariate1_t2 + b₃*covariate2_t1 + b₄*covariate2_t2 + . . . + error term.

You might have to reconfigure your dataset to calculate the pscores and reconfigure back again for the rest of the analysis.

David

[Wang Sonne]

Why don't you go for the Diff user written command of Villa(2012) in Stata where you have the options of applying a Diff In Diff +PSM estimation which seems like what you'd like to do. (I have attached some documentation to the command here). I am also using a PSM with baseline and endline data but struggling now with performing an heterogeneity analysis by trying to interact my treatment variable with a baseline covariate to be able to perform an heterogeneity analayis.