I'm running latent class analysis on a 12-item scale with 3-level response set (0, 1-2 times, 3+ times). However, very few individuals indicate the 3+ category for 9 of the items (and only about 10% for the other 3 items). I'm running the model as a 3-level ordered categorical response set, but wonder if that poses a problem given the distribution (i.e., should I convert the data to binary -- yes/no) responses? My preference is no (since one of my classes picks up the increased probability of frequent engagement in the 3 behaviors, but I'm worried this may be an artifact of the distribution.
I would look at the bivariate frequency tables for the 12 items. If you do not have zero or very small cells in the bivariate tables, you should be fine. If you do, you may want to collapse categories.