Ordered logit - interpretation of fitted vs. actual

[adrian_d]

Hi,
I tried for the first time an ordered logit estimation in Eviews, and I don't understand the fitted vs. actual ("resids") graph. I'm using Eviews7, you can find the file attached. There, in page Untitled1 I have equation EQ01_MOD. On the resids graph it seems like the fit is quite ok, fitting and even overestimating most of the peaks. However looking at prediction evaluation (or at probabilities extracted from the equation) there is very poor fit of high levels of dependend variable. Can you explain me what is the mechanism behind this? Should I look at the graph or only at predictions?

[EViews Glenn]

I'd look at the fit probabilities. The Actual/Fit/Residual looks at deviations from the fitted conditional mean, and in my view, isn't particularly useful in this context. In newer version of EViews we only produce the residuals plot since the former can be a bit misleading.

It's not surprising that probability fits aren't that high in a ordered model as you're trying to fit a bunch of regions simultaneously. My general feel is that your model isn't bad,
with large increases in prediction for the 0 case, and pretty good increases for the 2 case (albeit for smallish samples).