I am comparing two groups with multiple measurements before and after intervention (4 measurements before and 4 measurements after intervention in both experimental and control groups).

The outcome is level of anxiety. I am planning to use generalized estimating equations to analyse the treatment effect. However, the groups are not equivalent (different hospitals) and a senior statistician colleague of mine recommended some sort of matching technique. The problem is the popular matching techniques like propensity score matching are not valid when the data are correlated and it has been also criticized. What do you suggest me of a better method of adjusting or solving this baseline differences. I would appreciate it if you give me examples with Stata commands.

  • $\begingroup$ Random effects modeling may allow for adjustment of clustering and repeated measures. Why have you chosen GEE? Can you describe the data in more detail (eg N, N clusters)? $\endgroup$ – Todd D Dec 5 '17 at 17:25
  • $\begingroup$ Thank you for your reply. The patient scores (outcome variable on the scale of 0-10) are not normally distributed. I thought GEE better suits here. What we hypothesize was the anxiety score of patients will be lower in the intervention group compared to the controls, after the intervention. We have 3 hospitals (one intervention group (n=200 patients before intervention and n=200 after intervention; 2 control groups (again each n=200 before intervention and n=200 after intervention), giving the total number of patients included in all 3 hospitals to be 1200). ... $\endgroup$ – user187574 Dec 6 '17 at 6:34
  • $\begingroup$ Each patient has four measures before the intervention and four measure after the intervention in both control and experiment groups. But, as I mentioned earlier, the groups I am comparing are not equivalent (comparable) at baseline score and other important covariates. we have assigned the hospitals to intervention and control group without randomisation(quasi-randomised). So I am kind of stuck. Thank you. $\endgroup$ – user187574 Dec 6 '17 at 6:34

I suggest a longitudinal hierarchical linear model with time invariant covariates. See Kwok et al, 2008 and Raudenbush and Bryk, 2002.

If your outcome is anxiety (continuous), I would put scores on level 1, pre-post(dichotomous) on level 2, time-invariant covariates (e.g., sex, hispanic, etc. (dichotomous)) on level 3, and then experimental/control (dichotomous) on level 4.


Your Answer

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

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