| Message/Author |
|
|
|
|
Hi all:
I know this question has been asked in many different shades at one point or another on the boards but I still am a bit unclear. I ran the exact same EFA in Mplus and SPSS (also used LISTWISE=ON in Mplus to mimic the SPSS run) and got somewhat different loading patterns and values (overall fit about the same). Specifically, the way items loaded in the Mplus run were a bit more favorable in terms of conceptually understanding the factors (the run was oblique Promax). For example, some of items that SPSS had showing up on a separate factor loaded on one and only factor in the Mplus run. Is there any known reason for the discrepancy? Could the Kaiser normalization of SPSS be a source of this? Or perhaps the number of iterations specified for a model to converge?
Thanks in advance for any insight. |
|
|
|
|
| This may be related to the Promax parameter that governs how correlated one allows the factors to be (see, e.g. the Lawley-Maxwell FA book). It is a setting that different software may choose differently. If I remember correctly, this SPSS-Mplus difference was discovered in a earlier interchange. |
|
|
|
|
I had temporarily forgotten my password but was just ready to reply with this before I saw your response, Dr. Muthen.
Just read a bit on Kappa value for the Promax rotation. I read that the default in Mplus is 3 though this was an old post relevant to an older version of Mplus. Given that SPSS uses 4 as the default, this could be one source of the discrepancy. That said, the Mplus solution is more conceptually interpretable...so I am wondering if it is indeed the case that the lower value of kappa Mplus uses is somewhat more favorable as:
"Higher values of Kappa lead to higher correlations among factors and simpler structure of the loadings. The optimum Kappa is that value which gives the simplest structure with the lowest correlations among factors. The default value of 4 was recommended by its developers Hendrickson and White (1964) as generally providing a good solution. For SPSS Factor, Kappa must be a real value that is greater than or equal to 1.0."
--SPSS tech note |
|
|
|
|
| It could depend on the application which value gets the simplest pattern. I'd be curious to hear what you get when using our default rotation Geomin where you have the possibility of different settings. We are viewing Promax as a somewhat outdated rotation method per the writings of Jennrich and Browne. |
|
|
|
|
I ran with the Geomin and it basically matched the Mplus Promax run (which was more conceptually favorable than the parallel SPSS run). I did not mess with the epsilon value as I really have no compelling or justifiable reason to do so. What does puzzle me is that even when I adjusted the kappa value for the Promax run in SPSS down to 3, it still produced a different loading pattern than the Mplus run (essentially the SPSS run creates a second factor which is redundant with another factor--the Mplus run doesn't).
I read in the previous thread that there was some thought given to allowing user-adjustment to the kappa value for Promax rotation in Mplus but I did not see any mention in the manual so I assume that it was not implemented? |
|
|
|
|
Then I would think the Kaiser normalization that you mentioned earlier is the explanation for the difference.
Right, it wasn't implemented. We decided to not continue developing the Promax track given that Quartimin and Geomin got better press. |
|
|
| Back to top |
