Hi, I am running a latent class analysis with six ordered categorical indicators. I am using type = complex to account for cluster sampling. However, I find that there is disagreement between BIC and LMR LRT when judging the appropriate number of classes. BIC suggests an 8 class solution (which is interpretable), while LMR suggests only two classes. Prior to adding the type = complex statement, LMR suggested six classes. So I am wondering if it is possible that LMR under some conditions may underextract classes? I have heard that it may overextract classes, but with the current data two classes makes little sense, while the larger number of classes makes much more intuitive sense.
I am aware that theoretical considerations and interpretability must always be involved, but I have also heard that one should look no further when encountering a non-significant LMR.
Simulations show that LMR and BIC perform similarly, but real data is another story. Often LMR overextracts in my experience. Given your discrepant findings, I would inspect the profiles for 2-8 classes and see if going beyond 2 classes gives you new, substantively different and meaningful profiles. And you can also check Tech10 for absolute model fit of bivariate relations.
Thank you so much for the helpful suggestions. It appears that BIC is completely unaffected by the addition of TYPE=COMPLEX whereas LMR is greatly affected by it. I am no statistician, so I dont know why this would be. My fear is that the extracted classes are somehow an artifact of the clustering of the data and that this is what is reflected by the LMR non-significance.
Mike Todd posted on Thursday, November 14, 2013 - 5:58 pm
We are encountering a similar issue with our data. The SABIC and LMR test yield very discrepant results once potential geographic clustering is accounted for (via TYPE=COMPLEX) in our latent profile analyses.
With clustering accounted for, LMR tests point to very restricted, uninterpretable solutions.
For example, "non-COMPLEX" findings using the current set of continuous indicators and related indicators from the same sample point to interpretable 4-class solutions. Once clustering is accounted for, however, the LMR (but not the SABIC) points to a 2-class solution (i.e., the 3-class solution is not statistically better than the 2-class solution).
We are puzzled as to which fit criterion/a we should consider and/or weight more heavily.