0
$\begingroup$

I factor analyzing a measure with 55 categorical items (3 categories each). I am use CFA to test a 7 factor model. I have a very large sample (>10,000), but approximately 20% of the sample is missing data on at least one item.

I am using MPlus and have considered both the WLSMV and ML estimators. WLSMV is appropriate for categorical data and in Mplus it provides multiple fit statistics. However, this estimator uses either pairwise or list wise deletion. Given, the amount of missing data it would seem that this approach would warrant some sort of a priori item level imputation. On the other hand, ML is a full information approach that works well when the data are MAR or MCAR and the standard errors and chi-square test statistic are robust to non-normality. However, given the large number of factors this takes a VERY long time to run and given the number of items a chi-square statistic cannot be computed.

I am looking for any guidance on the best approach to take in this situation. Specifically, my questions are: 1) Can anyone point to references would argue for the use of one approach over the other?, 2) Can anyone point to references that examined item level imputation?, and 3) Does anyone have a different strategy that would be more amenable to this situation?

Thank you in advance for any thoughts/guidance.

$\endgroup$
1
  • $\begingroup$ This question seems to be more about MPlus, its options and best practices, than about factor analysis in general. $\endgroup$
    – ttnphns
    Commented Oct 31, 2014 at 19:41

2 Answers 2

1
$\begingroup$

To answer #3 - have thought of using Multiple Correspondence Analysis instead? With that many categorical variables Factor Analysis is unadviseable. https://en.m.wikipedia.org/wiki/Multiple_correspondence_analysis

$\endgroup$
0
$\begingroup$

I found this to be a helpful discussion of ML vs MI with additional references therein (eg to Van Buuren work) http://support.sas.com/resources/papers/proceedings12/312-2012.pdf

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.