0
$\begingroup$

I would have a question on multilevel models, which is related to a particular case that I am considering. The question is related to the number of my groups and the number of measures in each of them.

I have almost 4.000 of subjects that I would like to consider as my groups. For each of them I have a certain number of observations which cover a period of a year. The point is that the number of measures for each subjects can vary a lot (from a minimum of 19 measures to a maximum of 355 measures). My design is unbalanced and I cannot set a definite number of measures for each subjects (equal for all of them) since the predictive variables vary rapidly with the time, and therefore I cannot consider a "mean value" for them. For example, one of the variables is the external temperature on the day of each observations and therefore I cannot take a mean value of the temperature for each months, because the temperature varies and this variability is exactly one of the effects that I want to take into account.

I was wondering if such a model is adequate for a big number of observations and groups (like the ones that I have) or instead if you could suggest me some other models that take into account the correlations that I have between the various observations (wthin each group) and at the same time is more adequate for the analysis that I am trying to do.

Using SAS I tried to apply a PROC MIXED for this model and I saw, in the log window, that the software did not consider some observations because of missing values. I think that this message is due to the fact that my observations are unbalanced. What should I do to solve this problem? Moreover I have a high value of AIC (292315.8). As far as I know this value is useful to compare different models in order to find the best one but I do not know if this value is meaningful also as an absolute value or just as a comparison.

Thank you very much for any help you will give me!

$\endgroup$

1 Answer 1

0
$\begingroup$

First, multilevel models should be fine for this sort of problem.

Second, I am not absolutely positive but I am pretty sure the SAS log message is fine too. It's not about unbalanced design, it's just noting that you have missing observations (which you knew). In any case, questions about software are off-topic here.

Third, an AIC by itself is pretty meaningless.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.