Dan, If you have access to the v-c matrix of point estimates, the method of Li, Raghunathan and Rubin (JASA 1991) can be used. If you have access to the code that computes the test statistics, the method of Meng and Rubin (Biometrika 1992), which is asymptotiocally the same as the former, can be used.
If you only have the p-values/chi-square statistics, you can use the less accurate method of Li, Meng, Raghunathan and Rubin (Statistica Sinica 1991). All methods are described and summarized in a couple of pages in Rubin and Schenker (Statistics in Medicine 1991, pp. 589-590). Don On Wed, 12 Dec 2001, Dan Russell wrote: > On the basis of Schafer's multiple imputation program, I have computed five > data "complete" sets from a sample of 790 cases, and in turn have tested a > structural equation model with each data set. (You may be wondering why I > did not use the FIML option for estimating SEM models with missing data, > that is available in a number of programs. The problem is that I have a > dichotomous endogenous variable [mortality] that will not work with the > missing data options currently available in these programs.) I now have > five chi-square values reflecting the fit of the model to each data > set. My question is, how should one combine these chi-square values across > data sets to get an "average" value, that reflects the fit of the model > across data sets? I believe there may be some work by Rubin and his > colleagues on this issue; any references that anyone could provide me with > to this or other relevant literature would be greatly appreciated. > > Thanks, > > Dan > > > Daniel W. Russell > Professor, Department of Psychology and > Institute for Social and Behavioral Research > Iowa State University > 2625 N. Loop Drive, Suite 500 > Ames, IA 50010-8296 > USA > (515) 294-7081 Fax: (515) 294-3613 > Homepage: http://psych-server.iastate.edu/faculty/drussell/homepage.htm >
