I am doing a Generalised Linear Model analysis (GEE) and need some advice. I am comparing two well-randomised groups for an outcome measured at 5 follow up points, and the only real difference between groups at baseline is the baseline value of the outcome variable. I want to know if I should include the baseline measure of this outcome as one of the 'repeated measures' or just use it as a covariate. I suppose my lack of understanding of how GEEs work is the issue here. If it works by comparing the regression lines (over time) between the two groups (and thus the comparison is simply of the gradient of the lines) then it makes perfect sense to include the baseline measure in with the 5 follow up measures. But if there is also some comparison of the absolute magnitudes of the outcome variable at each time point in each group then including the baseline in the repeated measures would surely be a source of bias: the baseline difference could potentially contribute to any conclusion of a difference at follow up (or might eradicate any true difference). So do I include the baseline measure in the 'repeated measures' or do I just include it as a covariate? I'm using SPSS by the way. Thanks in advance!
Hi Mark,
The set up for the model depends on how your data is laid out. The way you describe the model in the second post is how one would perform a difference-in-difference analysis. However, each individual in the data will have two rows of data, one for the pre period and one for the post period, indicated by a 0 or 1 on a variable called Post (in the following example):
Y = a + b1Tr + b2Post + b3Tr*Post + bkXk + e,
where Y is an interval-scaled outcome variable, Tr is the treatment indicator (0/1), Post is a post period indicator (0/1), and Tr*Post is an interaction term that represents the difference-in-differences, (bkXk simply represents additional covariates as needed)
b3 (pred) = (¯y2,Treat - ¯y1,Treat) - (¯y2,Control - ¯y1,Control)
I hope this helps
Ariel
To me, Yeah sure. In addition you can consider it as a co-variate in ANCOVA.
Thanks Manoochehr! I think I've got it sorted out now. I realised you can work out the group, time and group x time effects, and so the group x time effects won't be affected by the baseline difference, though the group effect would. However the group x time effects are what we would look at. It makes sense to include the baseline variable if the test is mainly to compare gradients, as otherwise the change up to the first follow up is lost (if you don't include zero then the changes from zero to first follow up don't get considered).
Hi Mark,
The set up for the model depends on how your data is laid out. The way you describe the model in the second post is how one would perform a difference-in-difference analysis. However, each individual in the data will have two rows of data, one for the pre period and one for the post period, indicated by a 0 or 1 on a variable called Post (in the following example):
Y = a + b1Tr + b2Post + b3Tr*Post + bkXk + e,
where Y is an interval-scaled outcome variable, Tr is the treatment indicator (0/1), Post is a post period indicator (0/1), and Tr*Post is an interaction term that represents the difference-in-differences, (bkXk simply represents additional covariates as needed)
b3 (pred) = (¯y2,Treat - ¯y1,Treat) - (¯y2,Control - ¯y1,Control)
I hope this helps
Ariel
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis