In a very general sense, what are the different ways in which a retrospective observational study can be made to be robust?

Cross-validation? Well-defined exposures? etc.


1 Answer 1


The weakness of observational studies is the lack of randomization. The best way to make these studies valid (not biased), is to do them by matching cases to controls through methods such as propensity scoring. This makes each experimental case very similar to its control case in terms of the various covariates that are expected to affect the response (e.g. age, gender, status etc). My assumption is that you are using the term robust to mean close to unbiased which a randomized control study would be.

  • $\begingroup$ But propensity score matching only works if there isn't selection on unobservable factors. $\endgroup$
    – Charlie
    Sep 12, 2012 at 20:21
  • $\begingroup$ Can propensity score matching be used in cohort studies as well as case-controls? As Charlie points out, there will be many unobservable factors in a cohort study, especially one of a retrospective nature. $\endgroup$ Sep 13, 2012 at 11:06
  • $\begingroup$ @Charlie Of course you can't match on factors you don't have information for matching. But if they are not observed you do not know if they are imbalanced and can't directly assess their potential to bias results. $\endgroup$ Sep 13, 2012 at 11:36
  • $\begingroup$ @user1202664 I guess I would ask why you can't randomize. Is it because you are doing the study retrospectively? As I mentioned you can't match on unobservable and you probably don't know how they affect the analysis. You could not do covariate adjustment for them either (which you could do through regression modeling if they were observed). $\endgroup$ Sep 13, 2012 at 11:40
  • $\begingroup$ @MichaelChernick That's right. I'm doing it retrospectively. What alternatives do I have to PSM with the covariates I do have? You mention covariate adjustment: Could I do that as well as PSM or is it overkill? Also there are any ways of trying to account for unknown covariates? Thanks $\endgroup$ Sep 13, 2012 at 11:46

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.