Hello, We have outcome data collected at each year from ages 13 to 25 in a longitudinal design, and we would like to create a latent growth model based on these data. We have finished collecting the data for ages 13-18 and are still collecting data for the older ages, but we would like to get a picture of the emerging patterns. There are two issues we would like to ask about.
One is that, due to the nature of longitudinal data collection and the fact that we are still collecting data, there is a lot of missingness, particularly at older ages: out of 738 total cases, ages 13 through 17 have about 400 nonmissing cases each year, age 18 has 293 nonmissing, ages 19 through 23 have about 130 nonmissing, and ages 24 and 25 have fewer than 60 nonmissing cases each. This causes very low covariance coverage for many of the variables. Can you suggest a different way to organize the data to compensate for this problem, or can we run the models with what we have?
The second issue is that the outcome is problems with alcohol use, so it has many zeros. Only about 10-15% of our participants have non-zero outcome at each time point. Can this type of model handle there being many zeroes? That is, can this model detect changes with such low levels, or do we need to try another method? Thank you.
You may want to look at the Mplus UG examples regarding multiple-cohort analysis and regarding two-part growth modeling. These topics will be taught in a few weeks at our Johns Hopkins course. See also the web videos from Topics 3 and 4 on the web site.