Anonymous posted on Tuesday, May 31, 2005 - 4:12 pm
I'm attempting to run growth models (a total of 7 time points) with siblings and the siblings are clustered within families. There are 389 clusters and the cluster sizes range from 1 to 5. Over half of the clusters have a cluster size of 1. Is this contributing to high standard errors? If so, would it be better to randomly select one sibling per family?
So you are saying that of the 389 families, more than half have only 1 sibling in them? If so, no this should not cause problem - it does not cause high SEs as far as I can see. The 1-sibling families might, however, be somewhat different from the multi-sibling families, but that is a matter of whether or not to assume the same model for all families.
Anonymous posted on Wednesday, June 01, 2005 - 7:51 am
Thanks for the quick thought.
I guess there are additional layers that could play a role and it appears that the sample weight is playing a role. When the model is run with the sampling weight, the SEs for the between level intercept and slope variances are 469.5 and 394.6; however, when the model is run without the sampling weight, the SEs are 5.903 and 2.562.
Would there be an impact of sampling weight and cluster size?
I am conducting analyses using data from a survey that utilises a sampling weight. I have read several articles on complex survey designs on the MPlus website. I conclude from these papers that when a study involves a weighting variable, one should use the pseudo maximum likelihood (PML) estimator. Am I correct? Furthermore, I was wondering if you could explain to me how to compute this command in MPlus Version 4.1? I have been unable to find this command from the MPlus manual.