Phil Wood posted on Monday, November 06, 2006 - 5:16 pm
Hi folks, This post does not involve any error that needs fixing- My purpose is just to outline what I did and see if I missed something, Data: NLSY publicly available records. The variable of interest is past week alcohol consumption measured annually in a population of about 2,600 young adults aged 17 to 25. Description: Data show predominantly large cohort effects: Individuals initially assessed at ages 20-22 show large initial consumption rates which decrease dramatically. Simple correlations based on the raw data reveal modest correlations over time (around .2-.3). Analysis: I fit a simple, but alternate two latent group GMM to these data. Specifically, one group has a free factor mean, factor variance, and loadings. Identification is accomplished by constraining the sum of the squared loadings to be one. The second group has zero factor loadings, factor mean and factor variance but free Error variances. Given the age heterogeniety, I used the knownclass statement to model the 7 age categories. I then specified c#1 on agec#1; (etc.) for each of the ages, and free means for each of the age classes on the factor: %agec#1% (etc.) [f1*]; Although this setup converges, the means within each of the classes is the same, which is puzzling. Any advice/reactions would be appreciated- thanks! Phil
I don't see anything wrong with the approach. I assume that the model turns out identified, e.g. that the intercepts of the outcomes are fixed at zero. And that you haven't gotten a solution with 50%/50% in the 2 classes. Seems like this analysis could also be done using 6 age dummies as x's and doing a c on x regression with f on x as well. And both approaches should give different [f] means.
Other than that, Support would need to see the material to say more.