Randomized trial with repeated measure - which test?

#1
Hi,

I´m currently conducting a randomized placebo trial where we test the effect of one treatment on a single outcome variable. We measure the outcome variable at four time points for both groups: baseline, 3 months, 6 months and 12 months.

There is no crossover in the trial and we expect the largest part of the difference (if any) between the two treatment to occur between baseline and the first measurement at three months.

Which statistical test would be most suitable to used to show that there is a difference between the treatment and placebo?

I´m pretty used to working with STATA, but completely new to repeated measure analysis - so any thoughts and explanations are gratefully appreciated.
 
#2
I don't know STATA at all, but you want some kind of multilevel model. If your outcome variable is continuous, then the linear multilevel model is the place to start. (These models go by a variety of names - mixed models being most prominent). Essentially the regular linear model is (in matrix notation)

[math] Y = XB + e [/math]

where the e are assumed to be independent. This assumption is violated with repeated measures and MLM use instead

[math] Y = XB + ZG + e [/math]

to deal with this. (This divides things into fixed and random effects)
 

Karabiner

TS Contributor
#3
Or, as an alternative to a multilevel model, one could
use the "classical" repeated measures analysis of variance,
with an additional grouping (between-subjects) factor
"Mixed" ANOVA). Although I don't know STATA, I'd guess
that there is a well-documented procedure for this.

Just my 2pence

K.
 
#4
Or, as an alternative to a multilevel model, one could
use the "classical" repeated measures analysis of variance,
with an additional grouping (between-subjects) factor
"Mixed" ANOVA). Although I don't know STATA, I'd guess
that there is a well-documented procedure for this.

Just my 2pence

K.
That's certainly an option, however.....while the terminology for these models can be very confusing, I believe that the model that is usually called "repeated measures ANOVA" makes assumptions that the MLM model does not. See Hedeker & Gibbons, Longitudinal Data Analysis .

In particular, RM-ANOVA assumes sphericity, which is often unreasonable. There are some correction method within RM-ANOVA, but why not use a model that doesn't make that assumption in the first place?