We have just finished collection of behavioural data of a group of samples. We have data recorded every 10 minutes for an hour per day for 5 days (so 6 readings per day per sample for 5 days to give a total of 30 readings per sample) between a control group (5 samples = 150 readings) and test group (5 samples = 150 readings).
The question is how do we analyze the data? Is it better to pool all 150 readings together and then do a t-test with n = 150 per group?
Or alternatively is it better to average the 30 readings per sample out and do a t-test between the groups with n = 5 per group?
Or is there a better way to analyse the data, which I may have missed? I hope I am clear.
depends on the question (ie whether the temporal aspect of the repeated measures is of interest): mixed-effects model or GLM with generalised estimating equations; but do not pool
Averaging will result in loss of information. If your data are on ratio scale it is advisable t o work out canonical correlation between controlled and test groups.If data are in ranks then first calculate correlation matrix using rank correlations and then proceed Other as usual to calculate canonical correlation. Other, methods are also possible depending on the nature of hypothesis and research question to be dealt with.
The time of recording is the same across all five days, temporal aspect of repeated measures DO NOT matter in this study.
The data is a measure of time (eg. how much time the sample roamed about, stood still etc...).
I agree with Dr. Khan where averaging does result in a loss of data, though have not come accross much of a canonical correlation matrix. I am looking at Dr. Lusseau's suggestion of a GLM with GEE also.
Other suggestions are also welcome.
if you do not have to deal with a correlation structure (no temporal effect), I would simply treat the repeated measures as a random effect in a mixed effect model, no need for GEEs. depending on how far from zero your dependent variable is (a lot of roaming time), you may be fine with an assumption of normality for the residuals hence go for a linear mixed effects model, if you are close to zero, you may need more exotic distribution family. this may be helpful: http://stats.stackexchange.com/questions/58745/using-lmer-for-repeated-measures-linear-mixed-effect-model
David
Dr. Lusseau,
I have been looking at the mixed effects model for some time, and have a question ...
You propose to use the repeated measures as a random effect. something along the lines of this ?
Observation ~ Treatment + (1 | Subject) + (1 | Trial)
Just to note that this is not my forte...
KR
PH
- Badges
- Science topic