Controlling for baseline variables in parallel group RCT design

[Sam Honer]

Hello,

I have data from a 2-arm parallel group RCT, where patients were randomised on 1:1 basis into treatment vs. control. Randomisation was stratified based on age group. Baseline data was collected prior to randomisation and then follow up data was collected at 4 weeks post-randomisation.

Missing values were only present in baseline variables (bmi) and mean imputation was used in this case (as suggested by White & Thomson, 2005). Data was reshaped to long format, below is a snippet of the data:

id age occasion bmi sex trtmnt score
1 23 0 24 f 1 18
1 23 1 24 f 1 23
2 26 0 22.2 m 0 23
2 26 1 22.2 m 0 22
3 31 0 23.7 m 1 20
3 31 1 23.7 m 1 24
.
.
.

I intend to run a mixed effects model adjusting for baseline scores, sex, age, and body mass index (BMI) as indicated by the protocol in the following way:

gen treatXocc = treat * occasion
xtmixed score i.treat age bmi i.sex treatXocc || pid:

However, I was unsure of a few things:

- Am I correct in thinking that in using the mixed effects model as such then baseline scores are being accounted for through the allowance of random intercepts?
- How do I interpret coefficients of the interaction term?
- Should I also include a dummy variable for occasion here?

Thank you,
Sam

[Sam Honer]

Apologies, just realised the data is not legible in the original post.

id	age	occasion	trtmnt	bmi	sex
1	23	0	1	24	f
1	23	1	1	24	f
2	26	0	0	22.2	m
2	26	1	0	22.2	m
3	31	0	1	23.7	m
3	31	1	1	23.7	m

[Joseph Coveney]

- Am I correct in thinking that in using the mixed effects model as such then baseline scores are being accounted for through the allowance of random intercepts?

In the model that you show, you've explicitly included baseline characteristics as explanatory variables (predictors).

- How do I interpret coefficients of the interaction term?
- Should I also include a dummy variable for occasion here?

Have you seen this thread? The discussion and literature cited there directly relate to these two questions. (Is there a reason that you don't use Stata's factor variable notation for the interaction term?)