[R] Models for Discrete Choice in R

[Frank E Harrell Jr]

Iuri Gavronski wrote:
> Frank,
> 
> I certainly can't speak for Emmanuel. I don't know his reasons.
> 
> The reason I've posted this question is the fact that (as far as I
> understood), ordinal regression is based on logistic regression (or
> probit), and logistic regression expects a formula like dichotomous ~
> ratio1 + ratio2 + ... + ration. However, most examples I've found in
> Design, MASS and VGAM test models like ordinal ~ categorical1 +
> categorical2 + ... + categoricaln.
> 
> I wonder if it is just coincidence or I have just found the wrong functions.

Logistic regression includes binary, ordinal, and polytomous 
(multinomial) cases.  Binary logistic regression needs a binary 
response.  Ordinal logistic regression (usually proportional odds but 
can be other flavors such as continuation ratio model) needs an ordinal 
response.  The polr and lrm functions work this way.

Frank

> 
> Best,
> 
> Iuri.
> 
> On Mon, Nov 9, 2009 at 11:43 AM, Frank E Harrell Jr
> <f.harrell at vanderbilt.edu> wrote:
>> Emmanuel Charpentier wrote:
>>> Le dimanche 08 novembre 2009 à 19:05 -0600, Frank E Harrell Jr a écrit :
>>>> Emmanuel Charpentier wrote:
>>>>> Le dimanche 08 novembre 2009 à 17:07 -0200, Iuri Gavronski a écrit :
>>>>>> Hi,
>>>>>>
>>>>>> I would like to fit Logit models for ordered data, such as those
>>>>>> suggested by Greene (2003), p. 736.
>>>>>>
>>>>>> Does anyone suggests any package in R for that?
>>>>> look up the polr function in package MASS (and read the relevant pages
>>>>> in V&R4 and some quoted references...) or the slightly more
>>>>> sophisticated (larger range of models) lrm function in F. Harrell's
>>>>> Design (now rms) packge (but be aware that Design is a huge beast witch
>>>>> carries its own "computing universe", based on (strong) Harrell's view
>>>>> of what a regression analysis should be : reading his book is, IMHO,
>>>>> necessary to understand his choices and agree (or disgree) with them).
>>>>>
>>>>> If you have a multilevel model (a. k. a. one "random effect" grouping),
>>>>> the "repolr" packge aims at that, but I've been unable to use it
>>>>> recently (numerical exceptions).
>>>>>
>>>>>> By the way, my dependent variable is ordinal and my independent
>>>>>> variables are ratio/intervalar.
>>>>> Numeric ? Then maybe some recoding/transformation is in order ... in
>>>>> which case Design/rms might or might not be useful.
>>>> I'm not clear on what recoding or transformation is needed for an ordinal
>>>> dependent variable and ratio/interval independent variables, nor why
>>>> rms/Design would not be useful.
>>> I was thinking about transformations/recoding of the *independent*
>>> variables...
>>>                                        Emmanuel Charpentier
>> I realize that; still unclear.
>> Frank