Ordinal regression using R
I'd be grateful for advice from anyone who has done ordinal regression in R. The advice in Aitkin et al ('Statistical Modelling in R', chapter 5) seems extraordinarily convoluted. Is there no R package that does it more straightforwardly? If not, then it would seem that it is much easier to do it in a multilevel package such as MLwiN. Does anyone have any comments on that?
In fact, this is not really a question specifically about ordinal regression: the same set of questions would apply to any kind of multinomial regression.
I used a statistical package called Hmisc to do ordinal regression in S-plus a few years ago (I was still an undergrad, so it can't have been that complicated, though my colleague and I found it a bit of a challenge at the time). It is also available in R:
http://cran.r-project.org/web/packages/Hmisc/index.html
This is the paper to which we referred for the methodology (there may be some more recent publications):
Harrell FE, Margolis PA, Gove S, et al. (1998) Development of a clinical prediction model for an ordinal outcome: The World Health Organization Multicentre Study of clinical signs and etiological agents of pneumonia, sepsis and meningitis in young infants. Statistics in Medicine 17(8), 909-944
I hope this helps!
Quick post…
Would you like to fit the continuation ratio model or a proportional odds model?
Both of these models can be fitted in Stata (for cross-sectional data).
I understand that the random effects proportional odds models can be fitted in MLwiN (http://www.cmm.bristol.ac.uk/learning-training/multilevel-m-support/new1...).
Random effects versions of the models are available in gllamm, (although it uses quadrature and is therefore slow unless starting values are specified. I suspect that MLwiN could be put to use to estimate this model using MCMC.
The Lancaster e-Social Science SABRE project http://sabre.lancs.ac.uk/ might also be of interest – the website states that “SABRE is a program for the statistical analysis of multi-process random effect response data. These responses can take the form of binary, ordinal, count and linear recurrent events. The response sequences can be of different types. Such multi-process data is common in many research areas, e.g. the analysis of work and life histories. Sabre has been used intensively on many longitudinal datasets surveys either with recurrent information collected over time or with a clustered sampling scheme. Support for parallel processors can dramatically cut run times by many orders of magnitude. Sabre may be run as a stand-alone package or as a library for the package R or a plugin for the package Stata. The current release is 6.0. SABRE development was funded by ESRC as a pilot demonstrator project in e-Social Science. Current development is being funded by CQeSS (see also NCeSS) and Lancaster University.”
This text might also be of interest http://sabre.lancs.ac.uk/sabreR_coursebook.pdf
Good luck.
The R package "rms" is able to fit proportional odds ordinal logistic regression models. Otherwise, the package "VGAM" is very flexible in this area, although it may be overly complex for your requirements. Personally I've tended to use WinBUGS for this kind of model as I invariably require some kind of random effect structure.