I'm trying to do an ordinal regression in SPSS (to look at what health behaviors are related to an ordinal quality-of-life outcome). I have 6 predictors of interest and 5 covariates for control.

However, about half of them are continuous/scale variables that can't easily be dichotomized (e.g. age, BMI, and a few scores for physical fitness and a sleep questionnaire).

So my questions are:

  1. How should I interpret the odds ratios for the scale variables? These all have quite different ranges (e.g. 0-6, 0-300, and Age and BMI). I've heard that you can use the z-scores to standardize them, but am wondering if that's acceptable for a paper to be published.

  2. What else besides ORs and CIs are of value to report? Are there other tests to do that can tell me how good a fit my model is?


1 Answer 1


The scale variables are interpreted compared to their reference group. i.e: given category 3 is your reference group (unlike logistic regression, we do not have the option to directly specify the reference category and ordinal regression always chooses the last group as the reference group): one unit of increase in category one will increase the odds of your quality of life for x percentage compared to category 3.

for your second question, ofcourse the p-values for significant predictors and goodness of fit chi-Square and AIC and BIC are important indicators of your model fit to report on.

update:I don't quite understand your point 1, but if it is 1-15 scale, why don't you use it as a covariate (numeric predictor) as opposed to categorical or dichotomous? for your second point normally the best model is chosen where your AIC and BIC is the smallest- you can choose the step-wise function within spss to give all the competing models and then you can choose the model with smallest AIC and BIC and then choose the one which makes the best sense in your context, in sense of significant predictors etc- Sometimes the model with the lowest error is not really the best one- but you do not have to necessarily have all the predictors to be significant in your best model- You would preferably choose the model with the lowest error.

  • $\begingroup$ That's helpful, I appreciate the response. To follow-up on part 1, I suppose what I'm asking is, is it appropriate to use z-scores for a psychosocial score that's not easily dichotomized? (e.g. a scalar variable from a PTSD questionnaire that goes from 1-15, where it seems a little arbitrary to report an OR based on a single "unit," but that doesn't necessarily have a defined cut-off) $\endgroup$
    – Sgolenbo
    Oct 3, 2016 at 15:00
  • $\begingroup$ And for part 2, what I'm also trying to figure out is how to decide upon a "final model"--e.g. a variable for Alcohol Intake was non-significant, so I removed that; but do I keep other predictors in the model that are weak, or that do not especially contribute to the model's fit? $\endgroup$
    – Sgolenbo
    Oct 3, 2016 at 15:04
  • $\begingroup$ see above update $\endgroup$
    – RomRom
    Oct 5, 2016 at 0:18
  • $\begingroup$ Paragraph 1 is a bit confusing because it does not explain in what sense and how a scale predictor has categories, reference category. $\endgroup$
    – ttnphns
    Oct 5, 2016 at 0:46
  • $\begingroup$ Re. point #2, that's great - AIC and BIC I think should do. We do have a couple of non-significant predictors that I think need to be left in, so that's also helpful. $\endgroup$
    – Sgolenbo
    Oct 6, 2016 at 18:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.