Message/Author 

Sywang posted on Friday, July 13, 2012  4:46 pm



Hello, I'm running a CFA with multiple imputed data (I generated outside Mplus), the estimator is WLSMV because I have nonnormal likert scale. All the fit index values in the mplus output are showed as means over multiple datasets and there is no pvalue for chisquare. How should I interpret them? Thank you! 


The values given are averages over the imputations. The only fit statistic that has been developed for multiple imputation is chisquare for maximum likelihood estimation for continuous variables. 

Sywang posted on Friday, July 13, 2012  5:14 pm



Thank you for your reply. Do you have any suggestion? How do I know my model fit? 


Unless you are using maximum likelihood with continuous outcomes, there are no fit statistics developed for multiple imputation. 


Hi, I have a related question to this thread. I have a data set N=107 in which about 25 observations are missing data on all outcome variables (4 outcome). I am trying to run a simple path analysis with a few predictors and these four outcomes. Since I have cases in which data is missing on all variables except xvariables, FIML does not estimate that missingness. I have resorted to multiple imputation. Is there anything else I can do, other than multiple imputation? I noticed that when I use MLR estimator (I have skewed data), the only fit indices reported are AIC, BIC, and SRMR. Should I interpret the SRMR normally? It seems a lot higher than it should be, at least it's a lot higher than in the nonimputed analysis I ran. I have seen you say before that fit indices for multiple imputation do not have strong theory behind them. Thanks, Eric 


I would use FIML. You don't gain anything by first doing multiple imputation of the missing values on those 4 outcomes. FIML and MI use exactly the same information from the data. 


Thanks for the answer. I'm a bit confused however, when I don't impute, 25 observations are excluded because of missing on all variables except for xvariables. This reduces my total observations to 82 when using FIML via MLR estimation. When I use MI, I have my full sample size of 107. Does it still give you the same approximate information, despite excluding that many observations? 


The MI run works with subjects that have missing on all y's because it treats all variables equally (essentially as y variables). You can get the same FIML behavior as MI if you bring in the x's into the model by mentioning their variances. 


Thanks. I'm still running into problems. Let me tell you exactly what my data is, and what I have done thus far. I have 107 observations. Four continuos outcome variables: 1. peer competence 2. academic outcome 3. conduct problems 4. hyperactivity I have three predictors: 1. Gender 2. IQ 3. Executive functioning (EF) Initially my EF variable had a variance of 2311 while both gender and IQ had variances around .5. I rescaled the variances so they would all be between 0 and 10 using the define command. In my path analysis I regressed all four outcomes onto each of the three predictors. I mentioned the variances: EF@2.568; IQ@2.409; GEN@1.546; Now my output is telling me that the chisquare is negative, any reason why? If I don't specify the exact variances: EF; IQ; GEN; I get the warning message that of NONPOSITIVE DEFINITE, and it mentions there being a problem with my gender variable. I'm pretty sure this is not identified anyways. THANKS! 


Please send the output and your license number to support@statmodel.com. 

Back to top 