Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | "Ariel Linden, DrPH" <ariel.linden@gmail.com> |
To | <statalist@hsphsun2.harvard.edu> |
Subject | Re: st: Propensity score matching: Must all treated samples have a counterfactual? |
Date | Sat, 17 Aug 2013 09:06:39 -0400 |
I feel like a broken record, however I will suggest (as I do to most posters about propensity score matching) that you read some basic literature on propensity score matching. In particular Stuart (2010) and Caliendo & Kopeinig (2008). First off, when using a matching strategy, if a treated individual does not have a matched control, it means that there is no counterfactual for that case. This could be a situation where they are at the extreme end of the propensity score range, or if your matching algorithm matched too many controls to a given treated individual, leaving no controls for the next treated individual. I suggest you consider using common support, so that you won't have "missing counterfactuals" at the tails, and I suggest you ask yourself why are you choosing this particular approach? You certainly have not given us any insights as to why you have chosen radius matching and nearest neighbor matching (which are two different algorithms - unless you mean to say caliper instead of radius?)... As for your statement " nearest 3 within +- x score", I assume that you're referring to the caliper, but it is not clear. The general rule of thumb is to use a caliper of 0.20 to 0.25 of the standard deviation of the propensity score (see references below for brief review and further references). However, you'll have to decide for yourself what is a reasonable caliper that optimizes balance on covariates. In other words, too large of a caliper may result in reducing balance, while too narrow of a caliper will limit matched sample size... Ariel References : Stuart, E.A. (2010) Matching methods for causal inference: a review and a look forward. Statistical Science, 25(1), 1-21. Caliendo, M. Kopeinig, S. (2008) Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys, 22, 31-72. Date: Fri, 16 Aug 2013 14:57:09 +0100 From: Ricky Lim <ricky.lim12345@gmail.com> Subject: st: Propensity score matching: Must all treated samples have a counterfactual? Dear Statalisters, I ran running propensity score matching using -pscore- and would like to do radius & nearest neighbour matching (nearest 3 within +- x score). 15 out of 20 of my treated samples have more than 3 counterfactuals within 0.05 scores, whereas the remaining 5 have 1 or none. Their propensity scores are at the extreme and I would need to increase the range to +-0.5 before they will have 3 counterfactuals each. My questions are: 1. Must every treated sample have a counterfactual / control? 2. Is a radius of 0.5 score too wide? Is it be acceptable to use the nearest 3 within 0.5 score? Any advice is deeply appreciated. Thank you very much in advanced. Regards, Ricky * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/faqs/resources/statalist-faq/ * http://www.ats.ucla.edu/stat/stata/