I'm conducting an analysis of a non-randomised trial in which I have measures of Y at 3 timepoints: baseline, week 2 and week 48. Subjects are divided into 3 groups and lets assume that all were measured at each timepoint.

I am interested in whether there are significant differences in the change from baseline between the 3 groups. I.e. I want to identify differences between the slopes of the 3 groups for Y.

I am using this model:

Y = Group + Time + Group*Time

Where the Group*Time parameter is of primary interest.

Since this is a non-randomised study there are baseline differences which I need to account or adjust for. I have spent some time searching for advice / help on how to do this but have only found examples where people have adjusted ANCOVA type models in a manner similar to the following:

Post-test = Pretest + Group + Confounder

Since my primary variable of interest is an interaction variable, I am concerned that simply adding in the confounding variable to the model will not suffice:

Y = Group + Time + Group*Time + Confounder

Because of this I was thinking that I need to add the confounder to the interaction term in something like:

Y = Group + Time + Group*Time + Group*Time*Confounder

But this produces a prohibitively complicated model.

I haven't seen this done elsewhere, can anyone advise as to whether I'm going about this the right way?

  • $\begingroup$ Just a quick follow up to this. I discussed the matter with a local statistician and he said that the inclusion of the confounding numerical variables as only an additive effect should be sufficient to control for baseline differences. This is then equivalent to the model: Y = Group + Time+Group*Time+Confounder I've subsequently implemented this model in LIMMA and in a mixed model approach in SPSS. $\endgroup$ – Loodramon Nov 7 '15 at 14:05

Your Answer

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

Browse other questions tagged or ask your own question.