0

I am trying to perform an ordinal logistic regression in R and have recently attempted to use clmm2() function from the ordinal package. Unfortunately I've come to a roadblock.

  1. How do I assess model fit?

  2. How do I test multiple hypotheses by comparing AIC values of multiple models if I also need to assess the proportional odds assumption?

The data I'm dealing with: The ordered response variable is the consistency of bird food availability. We asked respondents to give us the percent days food was available for each season. I then bucketed it into none, pulsed, constant. So I have four rows for each respondent, making respondent a random variable in the model.

I have multiple predictors: season, age, etc.

Rachael
  • 81
  • 1
  • 9
  • Have you seen this recent thread? https://stats.stackexchange.com/questions/461273/how-to-build-regression-model-in-r-for-nested-data-and-ordinal-dependent-variabl/461439#461439 The non-proportional odds models are nested within the proportional odds modes, so you can use likelihood ratio testing (`anova`) to compare models, assuming the predictors are the same between the two models. – Erik Ruzek Apr 21 '20 at 13:08
  • @ErikRuzek I did see that thread. I had trouble getting the mixor() function to run and when I did, I didn't understand the output. I also couldn't find anything to use to assess model fit. Plus, this doesn't get at the hurdle I have -- compare hypothesis models by AIC first or assess prop-odds assumption first? – Rachael Apr 21 '20 at 18:48
  • I would probably figure out the model you want to run based on testing your hypotheses using `clmm2` and then use `mixor` to test the non-proportional odds assumption. Can you post your `clmm2` and `mixor` code so we can help with diagnosis? – Erik Ruzek Apr 21 '20 at 19:57
  • @ErikRuzek thank you for the suggestion. I've had trouble understanding the mixor output and better understand the clmm2 output. Is there a reason you suggest mixor over clmm2? I think clmm2 can be used to test the assumption as well. – Rachael Apr 24 '20 at 15:46
  • It appears it is possible to test for non-proportional odds with `clmm2` using the `nominal = ~ predictor_name` option. Then you can run a likelihood ratio test using `anova` to determine whether the non-proportional odds model provides a better fit to the data. It is not clear to me if you have to do this for each predictor separately (I think that is the case). In `mixor` you can allow for multiple predictors to have interactions with the thresholds, although estimation may take a while. – Erik Ruzek Apr 24 '20 at 19:33
  • Okay, I think I'll try that. Do you have a preference for clmm2 or mixor? – Rachael Apr 28 '20 at 12:31
  • Either will work. If you feel more comfortable with `clmm2`, then use that. – Erik Ruzek Apr 28 '20 at 15:39

0 Answers0