| Package | Ordinal | Multinomial | Partial proportional odds | Scale Effects | Random Effects |
|---|---|---|---|---|---|
| nnet | no | multinom |
N/A | N/A | No |
| MASS | polr |
No | No | No | No |
| ordinal | clm/clm2 |
all X's in nominal (may not converge) |
offending X's in nominal |
via scale |
clmm/clmm2 |
| VGAM | yes | ? | ? | ? | No |
| MCMCglmm | yes | ? | ? | ? | ? |
| Mixcat | yes | ? | ? | npmlt |
|
| mlogit | ? | mlogit |
N/A | N/A | ? |
| arm | bayespolr |
? | ? | ? | bglmer* |
| rms | lrm/orm |
? | ? | ? | No |
| MCMCpack | MCMCoprobit |
MCMCmnl |
? | ? | Yes |
| nparLD | yes** | No | N/A | N/A | (repeated measures) |
*Apparently this can be done using multiple logits(a more detailed post here); the ordinal package intro vignette offers a related insight into that, there appears to be a related paper, too: http://www.sciencedirect.com/science/article/pii/S0895435605003537
**limited number of factors
- Package ordinal has some nice, clear vignettes
- Package MCMCglmm has also a very complete vignette, actually course notes. A shorter tutorial for multinomial random effects can be found here http://hlplab.wordpress.com/2009/05/07/multinomial-random-effects-models-in-r/
- UCLA stats has a well documented mlogit tutorial using
multinom, and an ologit tutorial usingpolr. - Wiekvoet shoes how to analyse ordinal data using different classic methods and a JAGS model
Jeff Sauro: Should You Care If Your Rating Scale Data Is Interval Or Ordinal? http://www.measuringusability.com/blog/interval-ordinal.php
Martin, Karen. The analysis factor: Can Likert Scale Data ever be Continuous?
- A summary of the two factions, with some clear recommendations http://www.theanalysisfactor.com/can-likert-scale-data-ever-be-continuous/
Gelman: "The other thing you could try if you have multilevel ordered outcomes is to just model them as continuous. I bet that would work just fine, and then you could check things using some binary splits. And then you could use glmer." http://andrewgelman.com/2010/03/03/fitting_a_mulit/
Gelman: "You can typically treat a discrete outcome (for example, responses on a 1-5 scale) as numeric. Don’t worry about ordered logit/probit/etc,, just run your regression already." http://andrewgelman.com/2010/12/05/what_do_practit/
By Alan Agresti (http://www.stat.ufl.edu/~aa/):
- An Introduction to Categorical Data Analysis (2007)
- Categorical Data Analysis (2002, 2013) (I've read only 2002)
- Analysis of Ordinal Categorical Data (1984, 2010) (I've read only 1984)
Hedeker and Gibbons, Longitudinal Data Analysis (2006) (http://tigger.uic.edu/~hedeker/long.html)
- Some resources on his course site: http://www.uic.edu/classes/bstt/bstt513/index.html
- Books: Agresti and Hedeker (see above)
- Tutorials:
- UCLA's ologit tutorial explains how to analyze this graphically
- In the vignettes for the ordinal package two ways of testing/relaxing this assumption are explained, via
nominalandscale
- B. Jones gives a detailed paper and presentation on why it's important
- Apparently not a crucial assumption (Hedeker and Gibbons, 2006, p.?)
- Based on the "latent variable" approach to ologits/probits according to the ordinal package intro (p. ?).
The link
http://www.ats.ucla.edu/stat/r/dae/mlogit.htm might be wrong now, it should probably be
https://stats.idre.ucla.edu/r/dae/multinomial-logistic-regression/ now? Cheers, Felix