12-05-2014 05:51 PM
Hello, I hope that someone may help me
Well, I am analyzing a longitudinal data set about hearing loss which has 226 subjects with repeated measures each one (with a maximum of 12 observations per subject) over a follow up time of 22.2 years. I don't have fixed points along time, so I am dealing with my time variable as a continuous covariate. The data set is highly unbalanced, but I considered to work under the assumption of missing completely at random. Now, the problem I guess is because of how my response variable is trichotomized. I am using as cut-off points two values given by an international recognized scale for measuring hearing loss, which is giving me the problem of no convergence (SAS shows the message 'Did not converge'). I have tried all the possible options for the covariance structure, but any had worked.
The categories which are "normal", "mild" and "moderate" have 836, 31 and 6 observations respectively. When I change the cut-off points (i.e. for those given by a k-mean clustering), the following code works well:
proc glimmix data=hearing1 method=RSPL;
model ansi = age age*time/ dist=multinomial link=cumlogit solution;
random intercept age/ subject=id type=CS;
So my question is if the procedure does not converge, is it because of the frequencies in my classification? Should I split more equally balanced the observations?
I will wait for your advice
Thanks in advance
12-08-2014 11:02 AM
With that sort of imbalance, and two continuous variables that are likely to have a lot of collinearity (even if age is age at enrollment), I would expect a lot of problems with convergence. I would say that your approach is more valid for your data than applying the international cutpoints.
One other approach would be two dichotomous analyses: normal vs non-normal, and mild vs moderate (could be normal+mild vs moderate, if you want to keep the same subjects in the two analyses.)