04-02-2014 12:34 PM
If all subjects are measured at all time points, you could get a very good longitudinal analysis by using a spline on the time effect. This could be accomplished through the use of the EFFECT statement. See Example 77.4 Nonparametric Quantile Regression for Oxone Levels in the PROC QUANTREG documentation.
04-02-2014 01:08 PM
Thanks, Steve. My data is not balanced, subjects are not measured at all time points. Also, between subjects, the timepoints measured may be different. Time 1 for subject X is at a different point than Time 1 for subject Y. To be more detailed, I'm looking at markers of disease among controls and trying to plot deciles over time. I want to use these as possible cutoffs to compare to diseased patients.
04-02-2014 02:20 PM
Mmm. Well, you could "fluff" the data so that all time points are represented for all subjects, but with missing dependent values, and still get this to work. At least it should work in principle. I think that is the only way you are going to get to the quantile estimates.
Even if you were working for the mean estimate, PROC MIXED or GLIMMIX really would not like semi-sparse longitudinal data. Can the times be binned, say by week or somesuch, to get around the unevenness?
04-02-2014 02:52 PM
It is simply repeated measures so proc mixed handles it fine when I am looking for the mean estimate and use the repeated statement. Right now time is in days, I could bin to weeks, but it doesn't really solve the issue. Subject X could have the first measure at Week 2 while Subject Y has the first measure at Week 3.