I have been researching propensity score matching to create a control group to a few schools that have received an intervention (N=25). I have a lot of covariates I could match on but have not found any research about what would be a maximum given the size of the sample pool I could use. I figured a 1-to-1 matching technique would be useful to get a control group of 25 schools but I am not sure how many variablesI could use to get the closest match of 25 schools not receiving intervention. I realize theory is important; for example, including poverty index or school population. But I want to know how many variables could maximize the propensity score given the potential reservoir. Is there a formula that you can add a covariate for every 100 samples available in the reservoir? So, if you had 200 schools available as potential controls, you could only use 2 variables within your propensity score matching regression.
Hello Jenn,
It looks as if you haven't yet had much response to your query, so let me put in my two-cents' worth:
1. I don't know of specific, empirically-derived rules for how many matching variables might functionally work with a potential pool of 25. The answer couldn't possibly be always correct, since the characteristics of the pool (e.g., how homogeneous cases were) would have a huge impact.
2. In general, more matching variables makes finding matches more difficult, even if you give wide tolerance to the degree of similarity comprising a "match."
3. In general, small pools make finding matches more difficult.
I'd try the variable having the greatest dispersion in your primary data set as a potentially important one to be matched on first. Then, perhaps, continue to add until you found that: (a) matches were no longer possible; or (b) the tolerance for a match was so large as to be difficult to defend. Other than running a simulation study with specifications comparable to your data set/variables, that's the best I can suggest.
Good luck with your work!
Thank you for the tips! I did not think to look at the range/dispersion of the potential variables within the treatment sample. I do have theory-driven variables but it would be nice to add as many variables as possible to the matching model.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size