Dear FreeSurfer community,
I have a few questions regarding the usage of the LME matlab tools for longitudinal data analyses, and how to set up the model. We have patients and control data that we want to compare with respect to cortical changes over time (two time points: baseline and follow-up). We are also dealing with several cases (100 out of 300) with only one time point. The model is set up with “time” between scans, group, group*time, sex, and age as regressors. But, dealing with only two time points raised some questions:
1) I have read mixed opinions about fitting random slopes when dealing with two time points. Even though it seems to make sense to model individual random slopes, a fit through two time points may actually cause a problem. It also seems redundant to include 2 random factors (intercept and slope) in that case, as fitting random intercepts seems to be sufficient for the covariance estimation. So my question: Is it recommended to model only one random factor (intercept), or even wrong to model two (intercept and slope), when dealing with 2 TPs (especially when dealing with drop outs)?
2) This question is regarding how to set up the time variable. We have a large variance in that measure, meaning the follow-up scans were done between 5-7 years after baseline. As this contributes to the variance of the dependent variable, I would like to include the exact time interval as continuous number in the model (thus a slightly different number for each participant), rather than using time as a two-level factor (e.g. “0” for baseline and “6” for all follow-up cases). I am wondering which approach is recommended and whether there are any significant benefits of using one method (time as two-level factor) over the other (time as continuous) in a LME model with two time points.
3) We have several cases with data at one single time point (ca 100 out of 300). In the design matrix, I include single rows for drop out cases, with time=0. I assume that their number of repeated measures (1) will be handled by the ni-vector generated by the sorting procedure (sortData in matlab). However, we also have cases who were scanned only during clinical follow-up intervention (6 years after the baseline data was collected). Ideally, I would like to avoid assigning those to the baseline time point, for several reasons (e.g. treatment effects, potential scanner drifts over 6 years, etc). Instead, I want to assign them to the follow-up time point. If the better choice in point 2) is to use the actual (continuous) numbers for “time”, I am thinking about assigning the population median of the time difference between scans to those cases. Is this a reasonable approach? Does not seem to be the cleanest way of doing it, as we introduce some error, but may be a reasonable trade off to avoid dropping data of ca. 50 individuals. What is your opinion on that matter?
Thanks for any help and opinions. Best, Christoph