I am analyzing physical activity (PA) data as part of a larger RCT study and I want to see if there has been an improvement in PA levels from baseline to post intervention between the intervention and control groups. I am able to use the tabulate command (STATA) to see the n(%). My challenge in getting around this is, I have three variables, 1. PA level at baseline 2. PA level at post intervention 3. Randomized group and all three are categorized.
My Dear you can use Logistic analysis , Robust Logistic ,.. I am published artical in this subject at my count (researchgate) you can download this paper... I hope all success for your work in all fileds.
PA is your dependent variable. Can it be treated as an ordinal variable?
Your other variables are Group and Time. I presume you also have Subject identified, which is just an identification for each person.
What approach is best to use depends on 1) Can PA be treated as an ordinal variable? and 2) Do you have Subject identified at both Times?
Assuming you can, and want to, treat PA as an ordinal variable, and you have Subject identified for each Time, the appropriate approach would be ordinal regression, with a mixed-effects (multi-level) approach. This isn't the simplest thing, and I don't use Stata at all, so I can't tell you how friendly Stata is to this approach. (For example, will Stata give you post-separations of Groups?) Using the example here (https://www.stata.com/features/overview/multilevel-ordered-logistic-models/), you would want to a model something like PA Group Time || Subject: , and probably adding a Group#Time interaction.
A similar approach could be followed treating the dependent variable as categorical.
With this setup, it isn't easy to find a simpler approach. If you could convert your dependent variable to the Change-in-PA, that would help, but it doesn't sound like this is possible. You could also treat each group as a different experiment, and just compare PA across Time within each Group. This isn't ideal, and may not tell you what you want.
The following code template may help:
// make some artifical data, assuming outcome is normally distributed
clear
set obs 100
gen id = _n
set seed 2805
gen treat = rbinomial(1,0.5)
gen time = 0
gen ybaseline = rnormal(100,10)
// treatment effect ("true" effect) increases over time (only for treatment group)
gen ytime1 = ybaseline + 10*treat + rnormal(0,10)
gen ytime2 = ybaseline + 20*treat + rnormal(0,10)
// make data long format
expand 3
bysort id: replace time = _n - 1
gen y = ybaseline if time == 0
replace y = ytime1 if time == 1
replace y = ytime2 if time == 2
drop ybaseline ytime1 ytime2
// random intercept model with patients as "clusters" and treatment x time interaction
// estimated treatment effects over time are 12.25 and 20.6 respectively (vs "true" effects)
xtmixed y i.treat##i.time || id:,mle cov(unstructured)
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics