Suppose I want to look at the relationship between smoking and lung cancer, and want to use propensity score matching to select a similar group of people to study.

Knowing that age, and gender are possible confounding variables, I include them in my propensity model. However, I'm a bit confused what exactly the probability should be predicting.

Wikipedia says I should model P(treatment=1 | covariates), but what is treatment?

Is it P(lung cancer | age, gender) or P(smoking | age, gender)?



1 Answer 1


By "treatment," the article means the causal variable, i.e., the variable that might cause the outcome. In your smoking example, smoking is the "treatment" and lung cancer is the outcome. So the propensity score is the probability of smoking given the confounding variables, i.e., $P(\text{smoking}|\text{age, gender})$.

The word "treatment" in this context comes from the use of propensity score for estimating the effects of medical treatments (e.g., new drugs or medicines), but the word "exposure" is commonly used in epidemiology to have the same meaning. "Intervention" and "policy" are also sometimes used depending on the context.

  • $\begingroup$ thank you very much! That cleared up so much confusion around the wide array of terminology I came across researching this topic! $\endgroup$ Commented Oct 6, 2020 at 18:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.