This question is similar to one that was asked almost 10 years ago but never got answered. I want to perform a CFA on a data set mainly consisting of categorical data. It also contains around 2-3 percent of missing values. As explained here, I therefore carry out multiple imputations using R's mice package and then use the multiply imputed data to estimate the model using lavaan and the runMI command of the semTools package. That seems to work well.

However, it is common that one also provides some comments about the correlation matrix of the data used for the analysis when presenting the analysis results. Ideally one would even present the correlation matrix. Unfortunately, I have not find any way to carry out (polychoric) correlation analysis using multiply imputed data in R.

I therefore generally wonder what is common in instances like this. Does one use the original dataset (including missing values) to carry out the correlational analysis? This would make some sense. Imputation of missing values is also based on the correlation amongst relevant predictors. I guess one would certainly not go for one of the imputed data sets since thus would impose the challenge of justifying one's (arbitrary) choice of one of imputed data sets. Finally, another option could be to generate the multiple imputed data sets and then carry out a correlational analysis using sampling weights (e.g. svycor). I guess then one would apply equal sampling weights (assuming the syntax lets one do this). However, before going through all the pain I would like to get some feedback.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.