I am doing a study on developing a new instrument:
This new instrument had 73 continuous (5-point likert, though) items, but using Principal Component Analysis in SPSS, I removed bad items and finally got 53 items which had 15 factors. I have a concern if this factor structure is biased due to the many missing data. Since SPSS uses listwise deletion, only 474 cases out of the total 644 were used for this PCA analysis. I thought that ML method in Mplus EFA may solve this missing data problem, and I wanted to compare the factor structures between using the PCA in SPSS and ML in Mplus. So I used the following command:
Analysis: estimator = ml; Type = efa 1 15 missing;
But, the program is running for hours without ending. I guess there is something wrong in the model, but I am stuck here. Please help!
ML involves heavy computations with 53 variables and 15 factors, and you are doing 15 analyses with 1-15 factors. I would use a simpler approach such as ULS to hone in on a more limited range of factors for which you do ML.
Oh, I tried "estimator=uls" and the program finished with convergence. Did you mean that? I see only RMSEA as goodnees of fix index in the output. Is there a way to get chi-square values and others? Thanks
Sorry about posting separately. It seems that I cannot edit my previous message... I wish so.
I have only RMSEA info in the output, but anyways based on that, 15 factor model gives bad RMSEA like 0.5. I tried even 20 factors, but RMSEA goes down to 0.2. Does this mean that my principal component analysis in SPSS with listwise deletion of missing values resulted in a wrong factor structure?
What would you suggested as a next step? Thanks so much!
PCA is not a great approach to doing factor analysis. Listwise deletion may hurt as well. But the most likely source of your misfit is that you are exploring a "new instrument" as you say. Not all of the items may follow a neat factor model no matter how many factors you add. This is the time to explore by deleting, modifying, and adding items until what you hypothesize to measure is well measured. In terms of doing better EFA, you may also want to take a look at
Fabrigar, L.R., Wegener, D.T., MacCallum, R.C. & Strahan, E.J. (1999). Evaluating the use of exploratory factor analysis in psychological research. Psychological Methods, 4, 272-299.
Student posted on Tuesday, July 15, 2008 - 9:28 am
Thank you so much for your help and references. I will try more and get back to you. Thanks.
Ml parameter estimates are robust to multivariate non-normality as are MLR standard errors
toby hopp posted on Monday, October 10, 2016 - 11:45 am
I'm using a fairly simple set of models model to explore a large dataset (n = ~83,000). I have ten variables, all of which are count measures. I'm looking at 3,4, and 5 factor models. If I treat these measures as continuous, the analyses complete very quickly. However, if I declare the variables as count measures and use MLR estimation, the analyses take an extraordinary time to run (8 hours and counting). My computational resources seem adequate, so I'm wondering if the large n is simply overwhelming Mplus.