I calculated a spectrum of 2 to 7 latent classes from data on behaviour, using poLCA in R.
The model is based on 14 ordinal variables with 20 grades each. The number of respondents is 1.000.
Parameters used: 1.000 reps, maximum of 10.000 iterations, include missings (virtually no missing values in the dataset),
Ínterestingly AIC is quite constant without a significant minimum and BIC is increasing from the 2 class solution on.
The exact values (without fractional digits):
AIC: 19926 - 19912 - 19950 - 20024 - 20113 - 20210
BIC: 20915 - 21398 - 21933 - 22504 - 23091 - 23685
Maximum LL: -9762; -9654; -9572; -9508; -9451; -9399
Contentwise the 5-class solution is the most plausible one and the identified groups differ significantly in their profiles, but there is little support for that choice by information criteria etc.
I do agree that a meaningful model is to be preferred over one that is only suggested by quantitative criteria, but the fact that the numbers do not provide any orientation in this case makes me doubt.
Does anyone know what could cause that phenomenon?
Thanks in advance!
Stephan
BIC isn't helpful here because the number of distinct response patterns is so large I suspect. AIC is more useful and I would subtract 19926 from everything and look at the delta AIC from the simplest model as the absolute value of AIC is uninteresting.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R Statistical Package