effect of treatment in a three arm study

[Christian Bernard]

Dear Statalist.

I have a dataset from a three arms study (one of the arms is control, and the other two are two different treatments) where a participant was observed for occurrence of binary outcome at three time points (baseline midline and endline) six months apart. I would like to find out which treatment had the largest effect. kindly point me to the right comand to use.

Presently I am doing melogit outcome i.arm || timepoints:

Is this correct

Regards,
Christian

[Mike Lacy]

I'm thinking you want the outcome X time interaction, since the question is presumably "Were there differences across arm in individuals' outcome paths?" And, you'd want individual id as the subject effect. So, I'd think you'd want:

meologit outcome i.arm##i.timepoints || id:

From what you describe here, though, I'd say that the -melogit- model might be fancier than what you need, since an ordinary comparison of proportions might be more straightforward, although less flexible, I admit.

// Simulate data
clear
set seed 20332
set obs 300
gen int id = _n
gen byte arm = mod(id, 3)
expand 3
bysort id: gen timepoints = _n
gen outcome = runiform() < 0.2 + 0.05 * arm * time
// end simulate data
//
// Calculate outcome change for time 2 vs. 1, for example.
bysort id (timepoints) : gen byte change12= outcome[2] - outcome[1]
//
// CI and Test.
gen byte better12 = (change12 > 0)   // better vs. (no change or worse)
prtest better12 if inlist(arm, 1,2), by(arm)
// Quasi-McNemar approach, ignore individuals who did not change.
prtest change12 if inlist(arm, 1,2) & (change12 !=0) , by(arm) // better vs. no change or worse
// Make other arm and time comparisons as desired
// ...
//

[Christian Bernard]

Thank you very much

[Winfried Koch]

From my experience in the analysis of clinical trials, I would like encouraging the use of the melogit model as recommended by Mike but with the following change:
- Model the outcome for midline and endline only and use outcome at baseline as a covariate.