I am working on a project where I am using observational data from patients and trying to find a causal relationship on how Treatment dose affects Patient recovery. Since the data is observational, I am using Propensity Score Matching to match Control Subjects to the Treated subjects, to make the experiment as close to a Randomized experiment as possible.

I have 120 Treated subjects and 175 Control subjects. To my understanding, if I am picking the best match for each Treated subject, I should have 120 subjects in the matched Control pool.

When using the "Match" function from the "Matching" package in R, one of the parameters I have to specify is whether the estimand is calculated using ATT or ATE. From what I understand, ATE (Average Treatment Effect) and ATT(Average Treatment effect on the Treated) are parameters not used for finding matches, but for estimating the causal effect after matching.

When I use "ATT", I find that my Treatment group and matched Control group have 120 subjects each, as expected. But when I use "ATE", when I investigate the result of my matching model, I notice that both my Treated and Control groups have 295 subjects each, with 120 unique values (of all treated subjects) in the treated group, and 175 unique values (of all control subjects) in the control group. I am not entirely sure what is happening here, and I was expecting to see 120 values in the Treated and the matched Control (similar to what I get when using ATE). Could someone please explain what is happening?

  • 1
    $\begingroup$ This feature is poorly explained in the documentation. I recommend you email the package author and ask him what this option does. Typically, matching is only used to estimate the ATT; it's unusual to estimate the ATE with it, and the option to do so in Matching deserves more explanation. My guess is that all treated units are matched with a control unit, all control units are matched with a treated unit, and some weights are estimated to account for this unusual matching. In this way, all units would be retained but with weights applied. $\endgroup$ – Noah Feb 4 at 5:26
  • $\begingroup$ That guess does make sense if true. I will email the professor who authored the package and ask what it means. I am a little glad I am not the only one struggling to make sense of this, because I was worried I was missing something very simple $\endgroup$ – stats_nerd Feb 4 at 6:06
  • $\begingroup$ If you get an answer, it would be great if you could post an answer to your own question. $\endgroup$ – Noah Feb 11 at 2:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.