I would like to fit a Nonlinear Mixed Model with random effect using a NLMIXED in SAS but I am having problems getting the starting values.
My model for longitudinal data only has the outcome which is continuous and the covariate that is time range from 1 to 52. I observed SASonality in the exploration that and also CURVE EXPERT sugested a SINUSOIDAL model. The model is below:
y=(A+ai) +(B+bi)t+[(C+ci)+(D+di)t]*sin(w*t+delta)
ai , bi, ci, and di they are random effect for each subject.
So I need to get starting values for fixed effects (A, B, C, D, delta) w=2Pi/T .. T=1,2,3,4
Try an independent model (i.e. no random effects) and use the result of that model with all random effects are set to very small value (e.g. 0.01)
First unforetunately you need to test your data before trying to model it. The two most important tests are whether (a) it is truly continuous data and (b) whether it fits a normal curve (i.e. gausian distributions). If it is not a guasian distribution then you will either need to use a different model or to transform your data so that it is "normalized". I don't know whether this is a biological question you are asking or you are trying to do pure mathmatical modeling for a proposed phenomenon. But biological variability is famous for causing data to be uninterpretable. Make sure if you are using biological data, that you consider at first a few simple ANOVA to determine whether you need to get rid of any data points from this particular analysis because they are irrelevant to the question which you ask. I can give examples if needed, but unforetunately I have to run now...a book that might help if you are trying to ask a biostats question is http://www.amazon.com/Biostatistics-Essentials-Geoffrey-Ph-D-Norman/dp/1607951630/ref=sr_1_16?ie=UTF8&qid=1338899955&sr=8-16...it is a funny book, but full of helpful information as well.
Go with the advice of Reza but also try ranges i.e. 0 to 10 by .1. This will take up a lot of computing time but well worth the wait if successful. Also try the FIRO method if appropriate (see SAS documentation), it is a little more forgiving. If you run into negative variances (a possibility) try Choleskey root parameterization.
Best of luck!
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software