|
|
 |
| Who (what) should I believe? |
 
|
| Message/Author |
|
|
|
|
Looking for suggestions once again re GMM. I'm having trouble getting at least two of the indicators to agree for selecting the "correct" number of classes (N=396, variable scale 0-10, 7 assessments, lin & quad). As I add classes, the BIC keeps going down (by a lot). However, the VLMR yells STOP! at two classes (p-value above zero for the K vs K-1 class test for the 3-class analysis; Nylund et al. said that the VLMR was liberal in selecting more classes than it should, but if it was nonsignificant, then that was good evidence that too many classes had been extracted). The BLRT says "keep going" with p-values below .0000 for every model. I have been unsuccessful getting the VLMR or BLRT to find the same LL for 3 vs 2 class test as the model that I fit for 2 classes, but the 2 vs 1 class LL was identical. The diff for 3 vs 2 is sizable, not just decimals.
For these data, the large class (>280) has sig i s & q variances, but the i s & q var must all be fixed at 0 for the other(s) to obtain a solution.
Usually, BIC and BLRT are consistent, or BIC and VLMR, but none of these are consistent with these models. Any suggestions about what I should believe? The BIC, the VLMR, or the BLRT? Thanks! - bac |
|
|
|
|
When the fit statistics are in disagreement, it may suggest that the model is not correct for the data.
What does theory suggest? |
|
|
|
|
Aw, shucks, Linda! There you go again, bringing theory into it!
Thanks for the reminder, and also thanks for confirming that one indicator does not a model make! - bac |
|
|
| Back to top |
|
|
