If I calculate the McFadden pseudo-R2 using 1- (LLf/LL0)
LLf <- logLik(Model) #full model LL0 <- logLik(Model.0) #model with intercept only.
Why do I get different results if the model is fit with polr (MASS package) and with vglm (VGAM) using family=propodds.
For polr =
LLf = -6973.556 (df=16) LL0 = -9691.214 (df=3) R2 = 0.2804249
For vglm =
LLf = -1233.347 LL0 = 3951.005 R2 = 0.6878397
I don't understand why the log likelihood would go up by fitting polr and down using vglm?
I assumed I reversed something, but it seems not.
Fitted coefficients and standard errors are equal between the two models. AIC is completely different, but I imagine this is expected. I also notice the intercepts have reversed values, which seems strange. I cannot get vglm to generate a model with the "correct" intercepts.