Propensity score matching

[Abdallah Alidu]

Hello,
My understanding of propensity score matching technique is that the approach balances the differences in covariates between two groups so as to allow for comparison of the mean outcomes of these groups. Base on this idea, is it possible to use the method in finding the differences in mean outcome of males and females? If yes, how can model be explained if for instance age is positive and significant?
Abdallah.

[Matt Warkentin]

Hi Abdallah,

Yes, the goal of propensity scores (whether matching, covariate adjustment, IPTW, quantile-stratification) is to balance covariates between treatment groups (in the binary case) so that you can estimate the unconfounded/adjusted estimate of the outcome conditional on the treatment. In your described example, if your outcome is a continuous outcome, we could apply the following approach (where x1, x2...xj are covariates for balance (e.g. age); y is the continuous outcome):

logistic sex x1 x2 x3 x4
predict ps , pr
gen iptw = 1/ps

*IPTW-weighted method
regress y i.sex [pw=iptw]

*Covariate-adjustment method
regress y i.sex ps

*Covariate-adjustment method with non-linear PS
mfp: regress y sex ps

This could be extended to PS matching where you match individuals based on propensity scores (and some calliper width), and then run the proper model. I think you'll find that IPTW and covariate-adjustment approaches will provide similar coefficients. I am sitting in class so I haven't ran through the matching code to ensure it is correct.

[Abdallah Alidu]

Dear Matt,
Thank you very much for the quick response. But if I measured sex as 1=male and 0=female, how would I interprete the significance of a variable like x1 (e. g. age) in the logistic regression model? Will it be appropriate to say that age is more likely to increase the probability of being a male? That is, if age is positively signed in the model.
Abdallah.

[Matt Warkentin]

Hi Abdallah,

In your propensity model (e.g. logistic regression), a positive coefficient indicates that age increases the probability of the outcome (sex=male). In a more intuitive explanation, with the way your sex variable is coded (0=female, 1=male), a positive logit coefficient indicates, in your sample, men are older than women, on average.

[Abdallah Alidu]

Dear Matt,
Thank you very much. I appreciate your help which couldn't have come at a best time than this.