Xu, Man posted on Friday, January 13, 2012 - 3:14 am
Dear Dr. Muthen,
I am fitting a second order growth model using the "|" function.
In my waves 1 and 2, the data were collected at ages 43 and 53. At wave 3, it was all 60+, but the actual collections took place over a very large time span.
So far I took the average age (64.5) at wave 3 to set the growth parameter. And I have been using actual age at wave 3 data collection (age3) as a control variable to adjust latent trait on wave 3 (see below).
a11 would be age at data collection at wave 1, a12 being 53 and a13 being age 60+. In this case the value is the same for everyone at a11 and a12.
I actually tried to fit this model. It converged but gave a warning (below) I think at least the last sentence probably has to do with the a12 and a12 having no variance in them?
WARNING: THE MLR STANDARD ERRORS COULD NOT BE COMPUTED. THE MLF STANDARD ERRORS WERE COMPUTED INSTEAD. THE MLR CONDITION NUMBER IS -0.174D+00. PROBLEM INVOLVING PARAMETER 10. THIS MAY BE DUE TO NEAR SINGULARITY OF THE RANDOM EFFECT VARIANCE/COVARIANCE OR INCOMPLETE CONVERGENCE.
Xu, Man posted on Friday, January 13, 2012 - 12:05 pm
and do you think this warning above can be ignored? If so, should the mean of slope variance in the growth model be interpreted in the scale of a11, a12 and a12 (all being age in years), e.g. on average, each year the outcome variable changes in the amount of mean of "s"? I am quite interested in mean of slope in terms of cognitive decline after mid adulthood.
As long as you obtain MLF standard errors, you can ignore the error message.
The mean of the slope growth factor, s, is expressed in terms of a one unit change in age.
Xu, Man posted on Sunday, January 15, 2012 - 6:20 pm
Thanks a lot! Actually, I found something interesting. In the most recently model fitted to my own data, the time was measured in months (a very large value, like 300 months+). I wondered if the warning were due to the scales in the model being very different, so I change the scale to 0, 1, at time 1 and time 2, and rescaled time 3 as well accordingly. This time, the model converged without any complaint! Kate