Hello there,
I am performing repeated measures analysis where I need to look at changes in blood parameters over a period of time (lets say week). So, should I use week as a factor or continuous variable in the repeated mixed model. It would really be a great help and I would appreciate!
Both is possible. If your time variable is good structured and does not have many levels then you could consider it as a factor. But if you want to find the effect of a one unit increment in that variable or it is not good structured or has many levels, then you could consider it continuous.
Dear Mehmet,
Thanks for your answer! Would you please further clarify what do you mean by "good structured" time variable and how many are actually too many levels?
By good structured I mean that you measured regularly like per a week and this pattern repeats. There is no clear cutoff for too many, what I meant you should not have too many unregularly repeated measurements for regarding it as a factor.
Thanks Mehmet! Well, I have at least two different data-sets. In the first one, I need to look at for 10 weeks and in another, measurements were taken for every 20 min for 48 hours and this was repeated for 4 weeks!!! In both cases, in every time points, there is an equal space! Another thing, I am really interested to find if there is any difference within each time points! No sure how to proceed!!
Treat it as continuous. You may need to consider nonlinear transforms of the time variable (e.g. Log) if the changes aren't linear as a function of time.
Regarding tests at each time point, you usually don't bother. You really want to know if the intercepts or slopes differ (although centering means the intercept is a test on the middle of the time range). However, if you need to test specific points, your software may be able to even if time is continuous: multicomp in R allows it.
For your 4-week data set, it's unclear. Was it two days of a series of measures followed by a few days of no measures and then another two days of tests, etc.? If so, use two time-based predictors: week and 20 min period within a week with the latter number restarted with each new week. That will capture molar (across week) and molecular (within 48 hour period) changes.
Good luck!
Dear Michael E Young,
Thanks for your response. When I treat time (week 0 to 10) as continuous variable and ask R to provide lsmeans and Tukey adjusted pairwise post-hoc comparisons within each time points (for all week from 0 to 10), then it provides this only for week 5. But same syntax works perfectly when time is treated as a factor and gives comparisons with each weeks! Do not know why!!
Regarding your question, 20 min measurements (for 48 hours) were repeated for each week i.e. from week 1 to 4, I think it is - minutes nested within week (week/mins).
I've never used lsmeans for continuous, but the multicomp package can be used for this purpose. You just have to understand the way things are coded. However, if you focus on trends, the an lstrends command may be of use.
You are used to thinking categorically. You need to rethink your research questions in continuous terms.
Mike
There are advantages to either coding. The difference between the two models is in effect a test of nonlinearity of the continuous effect. There is also the option of ordered categorical coding.
One can also code time differently for fixed and random parts of the model. For example continuously as a fixed effect and categorically as random effect. The latter can be useful in modelling complex covariation (e.g., to handle violations of multi-sample sphericity).
Agreed, treating time as continuous opens up a whole tool box of continuous time/shape functions you can try. Treating time as a factor leaves that unspecified, so it's the method that is the most flexible and imposes the fewest constraints on the shape of the time effect. That could be a good thing (complex relationships could be revealed) or a bad thing (trivial/sample dependent week-to-week or day-to-day effects get misperceived as important). Maybe this depends on how much guidance from the literature is available to specify a function for the time effect. If it's well known, treat it as continuous and model it as others have.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis