Propensity score matching - binary outcome variable

[Felix Stein]

Hi all,

I have a panel data for a treatment and control group and want to compare whether they differ with regard to a binary outcome variable. I have used propensity score matching to identify pairs. I want to determine the effect of treatment.

My question is which procedure is the most appropriate. I have been using the user written-command psmatch2 because it allows me to enforce matching for the same year.

Thanks in advance,
Ben

[naveed ahmed]

This link is useful for understanding the treatment effects command of stata but may not answer your question totally. https://www.ssc.wisc.edu/sscc/pubs/stata_psmatch.htm

[Alistair Windsor]

If you are interested in using matching then full matching is generally better than propensity score matching. I use propensity score reweighting instead of matching. It does not discard information in your pseudo-control population unnecessarily and behaves much better with regard to bootstrapping. There are two issues to be aware of when matching: the issue of common support (observations in your treatment group that do not have reasonable matches in the pseudo-control population), and the issue of how to compute the standard error of the difference of means. Most of the literature completely ignores the fact that the control group is itself the result of an estimation and hence is a source of variation. Much of the rest use bootstrapping which is not on sound theoretical footing with discontinuous estimators like matching. There is a formula by Ambie and Imbens (I think) but my recollection is ut is asymptotic.

What at version of Stata are you using? This is apparently core functionality in Stata 14 but I have a package for doing ps reweighting in Stata 13. If you are interest I can see if they will take it on the SSC.

[Felix Stein]

Alistair Windsor Thanks for your fast response. I am using Stata 13. There is common support for my treatment group.

Disregarding the issues of propensity score matching. How do I state whether the treatment and control group differ with regards to the binary outcome variable and what is the best way to obtain it in stata?

Thanks a lot.

[Marcos Almeida]

Hello Felix,

The main command is - teffects - and you may get useful information by taking a close look at this: http://http://www.stata.com/manuals13/te.pdf

teffects estimates potential-outcome means (POMs), average treatment effects (ATEs), and average treatment effects on the treated (ATETs) using observational data. Regression-adjustment, inverseprobability-weighting, and matching estimators are provided, as are doubly robust methods that combine regression adjustment and inverse-probability weighting. teffects overlap plots the estimated densities of the probability of getting each treatment level. The outcome models can be continuous, binary, count, or nonnegative. The treatment model can be binary, or it can be multinomial, allowing for multivalued treatments.

Best,

Marcos

[Felix Stein]

Marcos Almeida Thanks a lot for the advice. I knew about the command but did not know that it had this featur and it seems to be perfect for my needs. However, does the ATE output and the relevant test produced by teffects provide me with the solution or do I need further calculations?

[Marcos Almeida]

I gather there is no right or wrong here. But I recommend you take a look at ATE versus ATET, and see which one is more appropriate to you study.

[Felix Stein]

Marcos Almeida Is propensity score matching a valid approach if my "outcome" variable of interest is pre-treatment? My task at hand is to compare whether treatment and control group differ with respect to a variable which is already in place before treatment. I am not interested in comparing the variable pre and post treatment. Thanks.

[Marcos Almeida]

Maybe you should also think about using instrumental-variables. Under a binary-outcome model, though, thinks get more complicated and you may consider studying hard how to perform a couple of specific estimations... But that may become quite complex, perhaps overwhelming to some, and I strongly recommend to discuss the model in full detail with your research group, ideally, with statisticians whose expertise relates to these matters. Unfortunately, at this very moment, I have not enough expertise to guide you through these realms.

[Felix Stein]

I think that is way to complicated for the purpose of my research. I just want to find comparable companies for my treatment and control group... I just want to understand whether they differ with respect to one variable which is prior to the treatment. I am replicating parts of this paper https://papers.ssrn.com/sol3/papers....act_id=1705707. The author does compare characteristics of treatment and control group prior to treatment by matching them according to certain variables. I was wondering whether procedures such as propensity score matching is appropriate to do is or whether I should use another procedure...