I look for a alternative regression to LPM and logit use in Stata with binary dependent variable. So far I tried to read about nonparametric or semi-parametric methods, but I cant find a matching one. The sample is quite complex, and the error term seems to be heteroskedastic.
Mplus does very well with complex survey data and finite mixture data. Both can be combined with a binary dependent variable.
J Owens posted on Thursday, March 28, 2013 - 1:14 pm
First, a clarification: It is my understanding that if I am estimating a path model with a combination of categorical and continuous dependent variables but do not declare the categorical (here, binary) variable as being categorical, then Mplus uses the MLR estimator, treats the categorical dependent variable as continuous, and I can think of this as a linear probability model (LPM). Is this correct?
Second, if I declare the categorical dependent variable as categorical, how would I get Mplus to use the polychoric correlation matrix with the ML/MLR estimator *instead of switching to the WLSMV estimator with theta parameterization* (note: I have covariates in my model but can leave them out if necessary to see how the use of polychorics changes the coefficient estimates)?
If you treat a dependent variable as categorical and use maximum likelihood estimation, you will obtain a logistic regression as the default. Using LINK-PROBIT will get you a probit regression. Using a polychoric correlation matrix as data would not be correct.
J Owens posted on Thursday, March 28, 2013 - 3:49 pm