Dear colleagues,
I'm analysing a randomized controlled trial in Nepal. I built a linear multilevel model with a score as the outcome, time as a dummy variable (i.time) and the interaction term i.intervention#i.time and id as second level in Stata:
xtmixed outcome i.time i.intervention#i.time || id:, var
According to Twisk 2018, excluding the treatment variable intervention should adjust for baseline score ( 10.1016/j.conctc.2018.03.008 ).
I then used the fitted predicted values
predict predri, fitted level(id)
to make a graph of the predicted scores over time. Predicted scores at baseline are different between control and intervention group, even though if they were adjusted for baseline they should be the same.
Am I doing something wrong?
Many thanks for your help!
Sebastián
You can either time-vlcenter the values or make baseline values t=0.
Hope this helps
Matt
Hi Matt,
My doubt is why, if supposedly I'm adjusting for baseline values, baselines at time 0 are not the same for the control and intervention group. And their difference 0
Sebastián
I am not really understanding the question]
but this papers show the importance of including baseline
Article Statistical analysis and optimal design for cluster randomized trials
Thanks Kelvyn Jones! My question is a practical one. I am adjusting for baseline values, but the predicted means at time 0 are not equal for control and intervention group (as I would expect if baseline values are adjusted for)
I think that this is to do with what predictions your are making - do them for exactly same values of the predictors for those on the intervention and those on the control (and not just average values). It is also worth putting baseline as the response and seeing the difference between the two groups so you will know how good your randomization has been.
Have a look at this too
https://gking.harvard.edu/files/matchp.pdf
Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference
You should trust your randomization, even if it's not perfect. If you plan to control for baseline values, it should preferably have been pre-specified in your protocol.
Best
John-Kåre
- Badges
- Science topic
- Similar topics
- Psychology