Message/Author 

Armen posted on Wednesday, July 26, 2006  4:05 pm



Dear Dr. Muthen, I am using WLSMV to fit a Multivariate ordered probit model. I have some missing observations on Y as well as X's. If I use TYPE=MISSING in the ANALYSIS command would it be compatable with WLSMV estimator? Could you please explain how does Mplus handle missingness in this case? Thank you 


If you use TYPE=MISSING; with the WLSMV estimator, a pairwise present method is used when there are no covariates in the model. When there are covariates in the model, missingness is a function of the covariates observed covariates. 

Sally Czaja posted on Monday, August 21, 2006  1:04 pm



Hello, I am attempting to test a model, with missing data, that includes four binary dependent variables (three are mediators). WLSMV provides probit regression coefficients, which are not typically used in my field and I fear will be confusing to reviewers and readers. Is it possible to obtain logistic coefficients/odds ratios with WLSMV? Or some other technique that accounts for missing data? As an alternative, I attempted to use ML, MLR, and MLF with the sample that has complete data but neither chisquare nor other fit indices were provided in the outome. Am I doing something wrong?? Thank you! 


You can obtain logisitc regression coefficients using maximum likelihood estimation along with the CATEGORICAL option of the VARIABLE command. You will not, however, obtain the regular fit statistics because means, variances, and covariances are not sufficient statistics for model estimation. In this case, you can use the loglikelihood to test nested models. 


Hi Bengt and Linda, I am using "type= H1 missing" and WLSMV estimation for a CFA of ordinal data (six categories) and found that the output does not provide the SRMR value (the WRMR is provided instead). Is there a mathematic reason why the SRMR is not provided when "type = h1 missing" is used? Rick. 


SRMR is given only when there are all categorical variables, no thresholds in the model, and no covariates in the model Thresholds are automatically included with TYPE=MISSING;. 

dlsp posted on Thursday, August 27, 2009  2:02 am



Dear Bengt and Linda, I am running a LGM with categorical data, 5 measurement points and predictors for the intercepts and slopes. I compare the results of the MLR and the WLSMV estimator using the missing option for both. Somehow, with both estimators a lot of cases (same number) are dropped from the estimation because of missngs on xvariables. I thought this problem does not occur with MLR and missingness? Do you have an explanation why the same number of cases is dropped with MLR and WLSMV? 


Missing data theory does not apply to observed exogenous variables. The model is estimated conditioned on these variables. 


Greetings, I understand that with WLSMV estimation, the default method that is used to deal with missing data is pairwise present when no covariates are in the model. This gets slightly closer to FIML when covariates are present in the model. Ok. However, recent versions of Mplus include the Auxilliary (m) function to specify lists of auxilliary variables for use in the missing data handling process. I am curious about how does this combines with, and modify, the default method through which WLSMV estimation handles missing data as implemented in Mplus ? As a quick follow up: What would be the impact of using as "auxilliary (m)" variables that are equivalent to the variables already in the model (lets say after using minor transformations, such as adding 1). Thanks 


I would not say the WLSMV with covariates approaches FIML in handling missing data. This would be the case only if the covariates predict the missingness. AUXILIARY (m) is not available with categorical outcomes. We found that it did not work well in this setting. It was originally developed for continuous outcomes. I don't think what you suggest would work. 

Daniel Lee posted on Monday, November 19, 2018  11:27 am



Just to get some clarity on how missing data is handled when using WLSMV, are cases automatically dropped if missing on covariate? For cases with no missingness on the exogenous variables, but missing on endogenous variables (mediators, outcomes), is pairwise deletion used for such cases? Thank you! 


Q1: Yes. Q2: Yes, when computing the sample statistics to which the model is fitted. 

Back to top 