| Message/Author |
|
|
|
|
| I have repeated measures data on 12 mood items that were assessed every evening for 14 consecutive days in over 400 adults. The response scale for each item ranges from 0 to 4. I had originally run a TYPE=COMPLEX EFA using Mplus, which showed a well-fitting 3-factor solution with 6 of the 12 mood items loading on a "positive affect" factor, and the other 6 items loading on a "negative affect" factor (there were also two items that cross-loaded on a third factor, but the loadings for these two items were weaker on this factor than on the positive and negative affect factors). However, I received a review on this manuscript that took issue with my approach because the 3-factor solution resulted in an eigenvalue that dipped just below 1 (0.925), despite the 3-factor solution fitting the data markedly better than the 2-factor solution obtained based on retaining only factors with eigenvalues > 1. The reviewer wants me to use parallel analysis to decide how many factors to retain. As best as I can tell though, the routine PARALLEL= is not available in Mplus (I'm using version 8.1) for either TYPE=COMPLEX or TYPE=TWOLEVEL EFA. Is there any other way to use Mplus to conduct an EFA that accounts for the nesting in my data while using the PARALLEL= option requested by the reviewer to determine an optimum number of factors? |
|
|
|
|
| No Parallel is not available for twolevel. Why not use BIC instead? |
|
|
|
|
Thank you for your prompt reply! By pretty much every definition (other than having an eigenvalue < 1), our 3-factor solution fit better than the 2-factor solution (fit indices are terrible for the 2-factor solution, but range from good to excellent for the 3-factor solution). However, the reviewer asked us to run a parallel analysis because he or she felt that having a solution with an eigenvalue < 1 was untenable, regardless of all the other information we provided regarding model fit.
So, I guess my follow-up question then is, is PARALLEL not available for nested data for any particular conceptual reason (i.e., the technique would not be expected to behave well with nested data)? Or, would it be theoretically plausible to do, but just hasn't been implemented into the Mplus code (if this is the case, would it be possible to "manually" perform a similar analysis using the Monte Carlo protocols built into Mplus)? |
|
|
|
|
| I don't agree that "having a solution with an eigenvalue <1 is untenable" - that's obsolete advice. I believe Parallel could be extended to multilevel but that's not in Mplus (yet). I think it would be hard/cumbersome to do this manually. Instead, convince the reviewer that going by BIC is good enough. |
|
|
|
|
| I will certainly try - thank you! |
|
|
| Danming An posted on Thursday, June 13, 2019 - 11:35 am
|
|
|
| Hi Dr. Muthen, I am thinking about doing parallel analysis manually for a project using type = complex. Could you elaborate a bit on how to do it? I know I can simulate random data sets but do I need to account for the dependency within each cluster when I simulate the data? Thank you! |
|
|
|
|
| That's a research topic. You would simulate data from a twolevel model if you have clustering. |
|
|
| Back to top |
