In an intervention study with subjects randomly allocated to two treatments
(treat A and B) and three time points (time) plus an additional baseline
measurement (dv_base), I've set up the following model to test for
differences in temporal courses of treatments for the outcome (dv), thereby
allowing for individual intercepts and slopes:
lmer(dv ~ dv_base+treat*time+(1+time|subject))
Fixed effects:
Estimate Std. Error DF t value
(Intercept) -1.080041 0.126665 58 -8.527
dv_base -0.888656 0.090617 53 -9.807
treatB 0.645455 0.190541 53 3.387
time -0.001726 0.163044 58 -0.011
treatB:time 0.377888 0.271972 58 1.389
I'm interested now in comparing estimated treatment means for let's say
the last time point and I've centered 'time' accordingly. The term
'treatB' shows the difference which is relatively high and significantly
different from 0 (t(53)=3.387). Now I want to compute the effect size of this
difference, based on the model, not on the observed values. From my
understanding I could obtain raw effect sizes by simply reporting the value of
the term 'treatB' (=0.645). When it comes to standardized effect sizes
(comparable to Cohen's d) I could simply take the t-value and the degrees of
freedom and use the formula 2*t/sqrt(DF)=0.930.
My question: Is this a correct procedure? I'm somewhat unsure as I've
never encountered it in the literature.
And help on this is greatly appreciated.
Andrea
[[alternative HTML version deleted]]