Hi, I need some help about logistic regression.
My participants were divided in 2 groups (control group and experimental group). They then performed a short test, with a binary outcome for each item (there were 4 items in the test, supposed to measure the same capacity to do something or not). Therafter, only the participants from the experimental group got a specific feedback.
One week later, the same participants from the two groups came back and performed the test again.
It was expected that we will observe more "1" for the experimental group (and not for the control group) in week 2.
I got the advice to perform a mixed-model logistic regression with R, with participants and items as random effects + the group (control vs. experimental) and the week (week 1 vs. week 2) as fixed effects (interacting). As I use R, I performed the test with glmer, and the prediction was well verified.
However... I recently read that I should not do this, as the observations are not independent, as the same participants were present in week 1 and week 2.
But you also have to know that the items in the week 1 test and the week 2 test were not the same (they were just supposed to all measure the same capacity), so each of the 8 items has only 1 measure per participant, and each participant has 1 measure per item. This may be very important, because I am not sure to understand whether the independance of observations is verified here or not, as the two test are not exactly the same.
I don't know what to think now... I am a bit lost about what to use in R to address this issue.
Perhaps instead of trying a logistic regression analysis should I (1) calculate a Cronbach alpha for the 8 items of my test (or better 2 alphas for each 4-items test), (2) use the total score (on 4) as my DV for each participant for each week, and (3) use an ANOVA considering scores between week 1 and week 2 as matched by participant (I am not a native English speaker, I hope "matched by participant" makes sense for you).
The advice of logistic regression was given by a collegue more used than me to statistical analysis, but I wonder now if that was really a good idea (and if it was the most simple way to anayse the data).
Thanks by advance.
Last edited by Gorgonzola; 05-09-2015 at 04:33 PM.
Tweet |