[R-sig-ME] Ordinal Package and post hoc test

[David Duffy]

On Tue, 29 May 2012, fernando barbero wrote:

>  I understand the LRT approach, but I think that the
> reduced model (fm2) should have 3 populations (pop2, pop3 and pop4), so I
> can asses if there are differences between pop1 and the other three (then
> with another reduced model, with pop1,pop3 and pop4 I could asses
> differences between pop2 and the other ones, and like that I would test all
> the populations), what do you think about this?

Sure, you can construct all possible subsets/contrasts "by hand" if you 
wish, and correct for multiple comparisons.

You might speak to a local statistician, as having thresholds and the 
populations as fixed and random effects complicates things (well I find it 
complicated!).  For example, differences in RE variances with equal means 
(in the middle category, say) for the populations appear as "spreading" of 
the thresholds ie proportions in Cats 1 and 3 are expected to increase or 
decrease in a nice way -- you should check whether the model is 
sensible.