We no longer allow MLM and MISSING. We recommend MLR and MISSING instead if you want a robust estimator.
Yan Li posted on Monday, December 11, 2006 - 8:38 pm
Thank you Dr. Muthen. Here is a follow up question. I want to do a model comparison for nested and comparison models. I read your guidance about "Difference Testing Using the Loglikelihood" on the website. One thing confuses me is that I got L0 and L1 for both models (2 L0 and 2 L1 values) and I don't know which L0/L1 to plug into the formula (TRd = -2*(L0 - L1)/cd).
(2) Also does the parameters p0/p1 refer to the # of free parameters in the output?
(3) At last, when we have the TRd, shall I look at the chi-square table using df difference (nested df-comparison df) and check out the p value?
WLSMV is not robust to non-normality for continuous variables. You can use MLR with a combination of continuous and categorical indicators. Factors with categorical indicators require numerical integration so we recommend not too many of these factors.
Under similar circumstances listed by RuoShui above (a model with continuous predictor variables, one of which exhibits kurtosis, and a categorical outcome variable), is MLR still appropriate if the categorical variable is in fact binary?
In regression, the model is estimated conditioned on the observed exogenous variables. No distributional assumptions are made about them. You can use WLSMV or MLR with a binary dependent variable. Distributional assumptions are made about only continuous dependent variables.