Is it possible to use propensity score matching (or something similar) to create a control group if you do not yet have outcome data? I am looking into using PSM to help determine the effects that an academic intervention has on student dropout rates, and I have a question about how to use it.
Currently, I have information on whether a student was in the treatment, and some demographic information that I want to use as covariates. I do not yet have information on whether or not students dropped out.
I am trying to do as much of my analyses upfront as I can. I would like to create a control group that I can compare to my treatment group in the future, when I have drop out data. Is that possible with PSM? Or do I need to wait to get the dropout data before I can do PSM?
 A: Not only can you do that, this this one of the main motivations for using matching methods (such as PSM) over other methods. You can perform the matching to select a subset of individuals that you need to follow up on instead of collecting outcome data on the full sample. This is especially useful when getting outcome data is expensive or challenging. It also ensures there is a separation between the design and analysis stages, since if you pre-publish your matched sample, you can't try a bunch of different matching methods after you've collected the outcome to try to find one that gives you the largest treatment effect.
You do need to know which outcome variable(s) you will collect, because to decide what variables you need to match on is to decide which variables are confounders, which are variables that cause both selection into treatment and the outcome. A confounder of one outcome may not be a confounder for another, so it is important to have a causal theory for the outcomes you plan on collecting data for in order to know which covariates you need to match on.
A: It's not strictly pertinent to your question, but without outcome data you can also use unsupervised machine learning techniques (eg clustering) to get better insights on your study population, and potentially inform patient characterization.
