Message/Author 

anon9210 posted on Saturday, September 11, 2010  7:38 am



Hi, I am running EFAs and CFAs on some nominal survey data and have a couple of questions about that: 1.) I have specified TYPE=COMPLEX in my analyses, and have specified clustering, stratification and weight variables. Based on the output, MPlus does seem to be taking these into account  the output lists these under "VARIABLES WITH SPECIAL FUNCTIONS". However, despite this, the output seems to be printing out only the "unweighted" N. For example, the unweighted N of my sample is 3197, and the weighted N is approximately 2980. However, I see only 3197 in the output. Is this because Mplus is scaling the weights in some way? 2.) While I get RMSEA, WRMR, CFI and TLI, as part of my output, I don't get AIC and BIC for the CFAs that I have been running. Do I need to add some special command for these? Thanks in advance! 


1. Yes, Mplus rescales the weights so that they sum to the total number of observations. 2. You must be using weighted least squares. AIC and BIC are available for maximum likelihood estimation. 

anon9210 posted on Saturday, September 11, 2010  9:40 am



1. Is there any way to turn the scaling off? 2. Indeed, I am using WLSMV as an estimator. Would it be possible to calculate AIC and BIC somehow? 


1. No. 2. AIC and BIC are based on the loglikelihood. Weighted least squares does not use a loglikelihood. 

anon9210 posted on Saturday, September 11, 2010  4:24 pm



Could I calculate AIC/BIC using the model chi square instead? I have seen it reported in some other articles. Here's an example where the author did use WLS, but has reported AIC and BIC. I am assuming he used the chisquare value for doing so...? See table 2 in the following article  http://archpsyc.amaassn.org/cgi/content/full/56/10/921 


I don't know of any methods investigation regarding computing BIC from chisquare with weighted least squares regression. 

Corey Savage posted on Saturday, January 30, 2016  11:35 am



If I have AIC and BIC values in the tens of thousands (e.g. 64,659) what would be an appropriate amount to differentiate between values? Both indices continue to decrease through the 8 class model. The decrease, however, is only about 80100 units once comparing the 6 vs 7 for example. I have not yet found where the BIC begins to decrease. I am using a number of Rasch scales and count variables as indiciators, some of which have a fair amount of skewness and some are multimodal. 


The more nonnormal your outcomes the more latent classes you tend to need. But perhaps you need some withinclass correlations to improve BIC, for instance by adding a factor. 

Back to top 