I am using a longitudinal dataset (NHANES) where the baseline ages range from 25 to 75 and there are three followups, after 10, 15 and 20 years. I need to estimate a GCM, starting with a univariate GC for a 'health' variable. I wanted to simply estimate a univariate GC with slope indicators based on the years to followup (0,10,15,20) and using baseline age as a predictor. However, because of the wide baseline age range I was advised to use the ages at interviews for the slope indicators: ages are grouped into 5-year intervals and a person gets a missing value if no interview was conducted at that age range and the appropriate value otherwise. This way, however, there are so many missing values that the model doesn't get estimated. I lowered the covariance coverage and the convergence criterion but to no avail. Q1: Do I really need to do the ages as slope indicators or is the way I started adequate? Q2: If the second way is better, are there any tricks to making the model work? Thank you very much for your suggestions!
bmuthen posted on Sunday, October 23, 2005 - 10:52 pm
It sounds like you should view this as a multiple-cohort situation and handle it via multiple-group growth model analysis (one group for each cohort). The cohorts (groups) are the age groups - I guess 10 of them. I would assume NHANES has a large enough sample to give sufficients numbers in each of these age cohorts. Each cohort is measured 4 times. In this way, low coverage is not an issue unless there is very strong attrition. The multiple-group approach allows you to test hypotheses about varying degrees of invariance of parameters of the growth model across the age cohorts.
Thanks so much for your answer! May I ask two more quick question? I intended to examine how the trajectory of excess weight affected changes in health. If I use the multiple-cohort setup, won't the whole model become too unwieldy? Would it really be wrong to use the times of measurement as the time variables? (I mainly ask the second question because the listwise deletion of individuals who die during the followup, which I understand to be the default in multiple cohort setup, would be a problem.) I really appreciate your help!
bmuthen posted on Monday, October 24, 2005 - 5:25 pm
The multiple-cohort/mutliple-group approach is not complicated. Using times of measurement is not as good of an approach. Listwise deletion is only the default in the automatic data rearrangement Mplus does using MCOHORT. I am not referring to that. Instead, you would rearrange your data yourself. With this approach, attrition is not a problem.
Dear Dr. Muthen, I have rearranged the data myself and tried to estimate the model (not as a multiple-cohort setup) but the covariance coverage is too low. Is there a way to use the MC setup with the data already rearranged? I hate to ask you this question since it may have an obvious answer but I couldn't find it in the manual. Perhaps you could point me to source(s) I could research? thank you very much for your answer, Anna