Stage 2 model issue (invalid iteration)

[sweeptime]

Dear Don,

I would like to add more information about running the models. The stage 1 analysis ran normally but there are some issues in stage 2 analysis. The message showed "invalid iteration" and the incomplete outputs are as attached.

I need to use "stage2only.exe" and "stage2only.def" to get the stage 2 analysis run.

Stage2.zip

Thanks!

Best,
Wei-Lin

Hi Wei-Lin. That's REALLY weird. I have replicated the problem using
my executable so I will work on fixing it--the problem is in the
second stage program. I will upload a new version of the program and
let you know once I've fixed it.

Rachel

[hedeker]

ok, there definitely is a problem with the 3 way interactions (covariate by location by scale) in the stage 2 analysis. The models don't run because these interactions are not being created correctly. something is going wrong in the creation of these interactions. However, i would prefer if the stage 1 analysis is redone with a centered and scaled version of the timing variable. Here, day is used as the timing variable in three ways: mean model, random slope, and WS variance model. It is coded from 1 to 31, so the intercepts in all submodels are kind of meaningless - they represent what is going on when day=0, but day never equals 0. This is particularly problematic for the random location intercept, which represents a subjects deviation from the mean trend when day=0. Also, the scaling of day (from 1 to 31) induces very small values for the slope of day, since these represent changes per day. This is particularly troublesome for the variance of the random slope, as the slopes are all very small values, resulting in near-zero variance for the random slope. So, the stage one analysis should be redone using a transformed version of the day variable. I would suggest:

week_c = (day-16.5)/7;

Here, the zero value of week_c will be in the middle of time, and the slope will represent changes per week. This will improve the estimation of the stage 1 model, and also make the results much more interpretable. I added this variable to the dataset, which is what i used in the youtube presentation - you can get it here: https://voices.uchicago.edu/hedeker/mixwild_mixregls/