checking the number of matches in psmatch2

[Tracy Lam]

Hi Statalisters,

I am using the user written command psmatch2 for propensity score matching. I run the

psmatch2 tx zmath2, outcome(zmath3) caliper(0.01) noreplacement neighbor(1) common logit ate

but my output is gives me 66 untreated off support, 137 untreated off support, 295 untreated on support, and 137 treated on support. Since this is nearest neighbor matching without replacement, should the number of matches either be 295 control to 295 treatment or 281 control to 281 treatment because for every treated subject there is a match for one treated subject?

Any advice is much appreciated. I've attached a screenshot. Thanks!

[Melissa Garrido]

Hi Tracy,

Common support refers to the range of propensity scores that overlap across the treatment and comparison groups. If observations lie outside of that range, they'll be dropped from your matched sample (regardless of the type of matching you do). In your current specification, over a quarter of your sample is outside of the range of common support (suggesting that there are some individuals who always receive a treatment and some who would never receive a treatment). When you go to interpret your results, it's important to keep in mind that you cannot generalize your results to individuals with characteristics similar to those outside of the range of common support.

Also, your number of matches within the range of common support is sensitive to the specification of the caliper size. For each treated individual with propensity score x, -psmatch2- looks for the nearest neighbor within the comparison group, as long as that neighbor is within the range x-.01 to x + .01. A caliper that is .2 * the standard deviation of the logit of the propensity score is often small enough and will give you more matches. If you didn't specify a caliper in the example you've given, you should end up with 281 treated and 281 untreated individuals in your sample (one match for each treated individual within the range of common support).

Hope this helps,
Melissa

[Melissa Garrido]

I received a private follow-up question on this topic - re-posting here so everyone can see it:

"Is it ever possible to do 1:1 matching and get an uneven number of subjects in the treatment and control? Could it be that there are many controls available so you can get more than one match for each treated subject?
Also, do you have any literature suggestions about this kind of matching?"


Here is my response:

By definition, 1:1 matching with no replacement should give you the same number of treated and comparison subjects (assuming all observations are in the range of common support). If you do 1:1 matching with replacement (where a comparison individual can be used as a match more than one time), you will probably end up with fewer comparison than treated individuals. If you have many more controls than treated individuals, you may wish to do 1:n matching, where each treated individual is matched to n (a popular choice is 3) different comparison individuals.

Elizabeth Stuart's article "Matching Methods for Causal Inference: A Review and a Look Forward" (Stat Sci 2010; 25(1): 1-21) is a good reference on matching choices.

Hope this helps!

Melissa

[Putut Purwandono]

Dear Melissa,

I am trying to combine PSM with the DID. When I use the nearest neighbor matching with 2 nearest, the psmatch2 produces 2 new variables (among others), the _n1 & _n2, which indicate the id of control matched samples. My step is to drop all observations whose the _n1 and _n2 is equal to "." or missing. The next step is to combine the original id and the id listed in _n1 & _n2 as observation that I am going to use in the DID estimation. Of course, I dropped duplicated id in my new id list. Did I do the correct steps in doing PSM+DID?

Another question is : is it okey to incorporate only 1 covariates plus a cross-term variable which is the DID estimator? Because I have only 1 covariates available in both period (my panel is two years). Other covariates are available only during the pre-treatment period. Thank you.

[Steve Samuels]

Dear Putut: You have responded to a topic that was closed two years ago and your question is not related to that topic. Start a new one on the main forum page.