0

I'd like to fit an ordered logit model in R with just linear terms.

I am using the polr package to get the ordered logit (first time using this package), with an ordinal dependent and three binary (hum2-4) and one 4-level (it1) predictors. it1 shows up with what looks like quadratic and cubic terms in the fit--see below.

I don't think I want higher-order terms.

  • Is that what these are, or is it something peculiar to do with contrasts?
  • If they're higher-order terms, how can I suppress them?
  • If they're something else, I'd love an explanation (or a link to one).

Thanks in advance for any enlightenment you might offer.

Call:
polr(formula = hum1 ~ hum2 + hum3 + hum4 + it1, data = dataOfInterest_na, 
na.action = na.exclude)

Coefficients:
          Value Std. Error t value
hum2No  1.23618     0.1444  8.5594
hum3No  0.81917     0.1280  6.4008
hum4No -0.54814     0.2248 -2.4383
it1.L   0.31895     0.1394  2.2876
it1.Q   0.28690     0.1230  2.3326
it1.C   0.02601     0.1044  0.2493
maitagorri
  • 41
  • 5
  • What are the data types of the variables involved (is `it1` a factor? or an ordered factor)? It's much easier to help with a [reproducible example](https://stackoverflow.com/questions/5963269/how-to-make-a-great-r-reproducible-example). – MrFlick Oct 10 '18 at 19:51
  • it1 is an ordered factor. I'm afraid I can't share the data, but should be able to make up some data that behaves the same - I'll try and throw something together :) thanks for checking. – maitagorri Oct 11 '18 at 21:07
  • How did you want to code the ordered factor? By default R will use a contrast that uses orthogonal polynomials [ref](https://stat.ethz.ch/pipermail/r-help/2007-January/123268.html). You can change that with functions like `contrasts()`. But it depends on exactly what behavior you want. – MrFlick Oct 11 '18 at 21:15
  • I thought I wanted some kind of pseudo-linear/ latent variable type treatment of the predictors -- not sure anymore though; this is the first time I'm dealing with this type of model. Ideal would be a coding that would allow for an as-simple-as-possible interpretation of coefficients for a Likert-scale factor. Perhaps my question should be about stats/choice of coding rather than r specific. – maitagorri Oct 13 '18 at 09:50

0 Answers0