回复 第1楼 的 zhanglei：
1. Yes, you can do prediction. Actually, it's pretty good. You can check literature on BLUP and its equivalence to penalized fitting in RKHS. But this could be somewhat mathematically involved if your background is not math/stat.
2. Random effect selection turns out to be a rather tricky issue and I don't think statisticians have a universal solution. From a practical point of view, trial-and-error is probably unavoidable. I'd suggest clarifying your specific goal and concentrating on that target using the available data.
3. Just like random effect selection, AIC/BICs in the mixed model context is non-regular. Considerable care has to be given to their validity. Generally, mixed models are used when there are multiple sources of randomness and relatedness among observations, e.g., your longitudinal data. The simplest case is paired two-sample comparison, which can be thought as a mixed model, where the observations within a pair cannot be assumed to be independent.
I understand that the above comments might be overly vague for you. But given the limited time, this is what I can say for now. For detailed analysis of your problem/data, I'd suggest finding a professional statistician to collaborate with you.
