Jen posted on Wednesday, February 29, 2012 - 12:07 pm
I am modeling 8 time points of binary categorical data. Looking at the data, there is no mean-level change over time (the proportions in order are .092, .131, .098, .101, .091, .072, .094, .114).
When I use an MLR estimator, the slope (in a model with only an intercept and a slope; the addition of a quadratic did not improve model fit) is non-significant as I would expect (s=.10, p=.28). However, when I use the WLSMV estimator (which would seem to be the better option since it is the default and since fit indices are available), the slope is significant and positive (s=.11, p<.001). The "estimated probabilities" graph with either estimator looks like a flat line (with a slight quadratic curve, which I figure must result somehow from the use of probits/logits, but no increase over time).
When I add any covariates to the WLSMV model, the slope is no longer different from 0 (s=.07, p=.29).
I am interested in looking at predictors of change (there is variance in the slope), but I would like to report a model without covariates first. The significant positive slope seems very suspicious given the data and the graph.
Jen posted on Wednesday, February 29, 2012 - 12:12 pm
Also, I know these estimators treat missing data differently; there is not much missing data and the differing results continue even if I only use complete cases.