Skip to content
Prev 5050 / 20628 Next

Mixed-model polytomous ordered logistic regression ?

Le lundi 03 janvier 2011 ? 19:54 +0100, Rune Haubo a ?crit :
That was my point, and Ken Kornblauch has been kind enough to send me
his code, which seems quite good. I have not yet worked on his code (and
probably won't for the next month at least), but it's probably a good
start.

The "800 degrees of freedom" problem still bother me. But that's another
problem (see below).
Thank you for the hint. I'll have a look at that. I have trouble
remembering which of the 2500+ R packages available on CRAN will solve
my problem-of-the-moment. Furthermore, I tend to trust more packages
doing something I've personally tinkered with. Hence my current
infatuation with BUGS (i. e. JAGS for the time being)...

But my point wasn't to solve a specific problem, but to point out that
ordinal data are a sufficiently frequent case to deserve "first-class
treatment". I am old enough to have been teached what was then called
"the general linear model", which was about the only case whee
multivariate model could be realistically fitted. Things tended to exa,d
a bit with censored data and the introduction of the Cox model, then to
Poisson and logistic regression. I tend to think af ordered data as
another kind of data that should be analysable the same way. That's what
polr() partially does (with a limited but wisely choosen set of
options).

Similarly, mixed models became *practically* fittable first for the
dreaded Gaussian case, then for Poisson, binomial and negative-binomial
cases. Ordered data should be analysable the same way.

Another case (not directly in my current interests, but one never
knows...) is categorical data. Log-linear models allow for a similar
treatment, and should be (too) analysable in the mixed models case.

A lot is to be learned from the Bayesian models for this kind of
problems. I'm currently exploring the (now somewhat outdated) Congdon's
textbooks on Bayesian modeling, and it raises interesting ideas bout
what should be doable from a frequentist point of view...

On the other hand, I'm more and more inclined to think that the Bayesian
point of view is extremely valid. The first difficulty is to convince
journals' editors that "p-value is crack" and that a posterior joint
distribution says all that can be learned about from an
experiment/observation, provided the priors are correctly chosen. The
second (and harder) is to find ways to express Bayesian probability
models in a more concise way than what is currently allowed by BUGS :
one should not have to type dozens of almost-identical snippets of code
for each and every intermediate parameter appearing in such  model. My
current idea is tat some sort of a macro facility should help, but is
hard to design *correctly*. The third (and hardest) difficulty is to
implement that efficiently. I'm a bit stricken to see all "public" BUGS
implementations using only one of a 4- ou 8-CPU machine, even for parts
(model updating of parallel chains) that are "embarrassingly
parallel" ((C) D. Spiegelhalter & al.).

In our present case, the Bayesian answer "feels" more reasonable than
the answer given by polr/glm, in particular because the variability
estimation of the thresholds seems to avoid the artefact induced by a
non-centered independent variable...
Ex 15.1, section 15.5 of Gelman & Hill (2007), p. 342. The printer's
info at the beginning of the book (just after the full title page, but I
don't know how this page is called in American usage) says it's "First
published 2007//Reprinted with corrections 2007//10th printing 2009"

I am not aware of any other edition.

HTH,

					Emmanuel Charpentier