I am doing a latent curve model (i.e. growth model) with data including MZ, DZ, full- and half-siblings. Thus, there is nonidependence within sib-clusters and the degree of clustering will vary between the 4 sib-types. I want one set of estimates though, not 4 (i.e. one for each sib-type, as in a multiple group analysis). Essentially, I just want the standard errors to be adjusted for the complex clustering, but otherwise (i.e. one set of parameters) standard output. Also, for the command that does this, could you direct me to the underlying math. thanks much.
Perhaps you want to do Type= Complex where cluster = family. The cluster members are then the MZ...sibs.
This is using a "sandwich" estimator for the SEs as described in the Version 2 Tech appendix 8, eqn (170). Here it is only assumed that you have independent observations across families, not within families.
I am confused by your statement that you want one set of estimates, not 4. You also say that you want one for each sib-type - isn't that 4?
Sorry i was not clear. I have four types of sibpairs in the data (MZ twin...half sibs). I want to estimate 1 set of parameters for the whole sample, as opposed to one set of parameters for each sib-type (as in a multiple group analysis). I was concerned that using the standard: "Type= Complex where cluster = family" would not be sufficient to account for the differing degree of clustering between sib-types. I don't know though. What do you think?
The sandwich formula assumes nothing about degree or varying degree of within-family correlatedness, so I think it should be fine.
One could alternatively analyze in a multivariate fashion (2 sibs measured on 1 variable gives 2-variate observation vector), where the correlatedness is instead modeled. You say that you don't want multiple-group analysis, but there is the new alternative used in the Version 4 User's Guide QTL example, ex5.23, where the sib type correlation variation can be read in as data (1., 0.5, etc).