Greg Roberts posted on Wednesday, January 30, 2008 - 1:46 pm
Good morning. I am fitting glm of reading achievement data from k, 1,3,and 5 grades. Data are stratifed, sampling parameters known, including probabilities of selection. Trend is curvilinear. Trying to fit an unconditional model. Several questions: 1. If I model clustering in a single level model with case-level sampling weights, are the standard errors for the growth paramters adjusted for effects of group? 2. Single level model fits well, but some difficulty with theta&psi, esp. the residual variance for t1 and t4 scores (significant negative). Tried piecewise, correlating adjacent errors, autoregressive. Only solution seems to be fixing residual t1 t4 at 0, which gives fit almost as good as the model with warning messages, yields same parameter estimates/standard errors, eliminates warning messages. Model fits data, but I am not sure how far from "unconditional" I strayed imposing constraints. Thoughts? 3. I will add between level, for questions re: school-level effects on student growth. What are implications of not modeling the grwoth at between level, instead using it to model school-level effects on case level growth. I assume that the standard errors are already adjusted (per question 1) and that my 'solution' in number 2 is the best available, which if so, is creating problems with the modeling of school-level longitudinal growth. Thoughts? I appreciate your time.
1. Use TYPE=COMPLEX; with the CLUSTER and WEIGHT options to obtain correct standard errors. 2. Hold the residual variances equal. 3. Modeling both the within and between parts using TYPE=TWOLEVEL provides a fuller set of parameters and a richer analysis. See the description at the beginning of Chapter 9 where the two methods are compared.