I’m fitting a path analysis model that has a single dichotomous outcome, along with several dichotomous ‘endogenous’ mediators.
The data come from a complex multistage sampling design and utilize the balanced repeated replicate (BRR) for variance estimation. At AERA 2005, Dr. Stapleton presented a method for using MPLUS to obtain adjusted standard errors for BRR, and adjust the chi-square stat. For this procedure, I must fit the model as many times as there are replicate weights (in my case, 16) and save the parameter estimates and chi square from each run.
Because I had dichotomous outcomes/endogenous variables, I started out using WLS (or WLSV). Several of the runs (at least 6 out of 16) did not converge--even with a very high number of iterations I switched to MLR, and was able to proceed to adjsuting the standard errors. NOTE that I do NOT use TYPE=COMPLEX for BRR variance estimation. (However, I cannot obtain the chi-square statistic using MLR)
What kinds of issues do you think I should consider in using MLR vs. WLS ? Are there any fit indices I could report using MLR?