Daniel posted on Tuesday, March 16, 2004 - 4:22 am
I am working on revising a manuscript. We conducted a latent class growht model, with an ordered categorical dependent variable. We regressed the latent categorical variable on covariates in order to characterize our four trajectories. After reviewing our intial results, we changed our comparison groups to make different between trajectory characterizations. An editor of the journal questions us by saying that we are inflating the probability of a type I error by making additional comparisons. How does the issue of type I error work in this case? Can you discuss the issue of type I error in latent class growth models, particularly when additional comparisons are made, and when baselines are recentered (e.g., moving the level from time 1 to time 2)?
bmuthen posted on Tuesday, March 16, 2004 - 6:16 pm
I think Bonferroni matters are the same here as elsewhere, so to be strict you may want to look for significance at a smaller p value when you do many significance tests. What you did seems fairly harmless - one could argue that both the set of tests with different reference class and the set of tests with different centering was planned ahead of time. But still, using a smaller p might be prudent.