OHDSI Home | Forums | Wiki | Github

How is PS matching incorporated into a model?


Propensity Score (PS) matching can be done using the the nearest neighbour method: one unexposed subject with one (or more) exposed subject with similar PS.

Once I have calculated the PS and I have matched the subjects…

How do I introduce this information into the model? (For example with the package lme4).
Or it can only be done with special packages such as Cyclops and MatchIt?

Do I just need to add the PS term in the model?

logit(Outcome) ~ Treatment + PS

Do I also need to add the pair or cluster as a random effect?

logit(Outcome) ~  Treatment + PS + (1|cluster)

Or all I need to do is to subsample the original dataset keeping just matched pairs and discard unmatched people?

Please see the CohortMethod vignette for details on how to use the PS matching.

I had already read that vignette.
There the PS is used internally by the OHDSI tools.
It seems it creates a subpopulation by matching people, and discards the remaining ones. Isn’t it?

But I’m asking how to get the same results with common packages such as glm, lme4, survival… How do I introduce the PS matching in models created with those packages?

If it was easy to do with existing packages we wouldn’t have created our own.