Message/Author 


I have a large correlation matrix in excel that has multiple missing values. First off, can you please clarify what type of file I need to put this in for Mplus to read it (I've read different things .txt. prn, etc.) Secondly, what if anything do I need to put in the cells that are the missing values for it to be read? Do I need to use "correlation" for type in this case and is an ML estimator the best/necessary option? 


Is it the correlation matrix that has missing correlations in it, or is it the raw data that has missing data for certain subjects on certain variables? 


It is the correlation matrix that has missing correlations in it. It is actually a correlation matrix of metaanalytically derived correlations wherein some correlations are not found in the literature. 


What kind of model are you fitting to this correlation matrix? 


I'm hoping to fit multiple models, both EFA and CFAs. More so CFA's where I want to compare various models with 14 latent factors (regarding psychopathy) and also to correlate those latent factors with other factors regarding the five factor model of personality. 


I don't see how this can be handled. Unless the missing correlation values are in a certain pattern so the matrix can be broken up into groups with fewer variables per group. 


As it is, I don't see a way to break up the matrix into smaller groupings, and I am aiming to do models that necessarily require all of the information. To clarify is it required that if one wants to use a correlation matrix as the sole input nothing can be missing? I've read other studies with the same problem that state "Missing correlations were estimated with composite maximum likelihood (ML) methods. In this approach, an exploratory factor model was first fit to the available correlations of the matrix; this model was used to generate modelpredicted correlations, which were used as imputed values for the missing correlations in the matrix." Is this not possible with Mplus? 


I don't know about this composite ML approach, but if it is sound I would think there is a way to do it in Mplus. Please send me an article on it. 


This is the article in which I found the above quote Sharma, L., Markon, K. E., & Clark, L. A. (2014). Toward a theory of distinct types of 'impulsive' behaviors: A metaanalysis of selfreport and behavioral measures. Psychological Bulletin, 140(2), 374408. doi:10.1037/a0034418. The authors talk only briefly about the technique on pg. 379 in the paragraph that begins "This methodology resulted in a 58X58 correlation matrix..." (I have emailed the corresponding author of this article to clarify what program they used as well). There are a few other resources they cite that I have either not yet had access to (Markon, 2011; Lindsay, 1988) or am having trouble following because of the complexity (Cudeck, 2000). Continuing thanks for your responses. 


what is the Cudeck, 2000 ref? 


Cudek, R. (2000) An Estimate of the Covariance Between Variables Which are not Jointly Observed. Psychometrika, 65(4), 539546. 


Will have a look at it. 

Back to top 