I am using an ordinal logistic regression to answer my research question on factors that predict helpfulness ratings of hospitalization experience. My outcome variable has three levels ranging from helpful to unhelpful. I have multiple predictor variables, consisting of both categorical and continuous, I am testing at the univariate level before I enter them into my full regression model. A few variables are significant and pass the test of parallel lines. However, one of my significant variables does not pass the assumption of proportional odds.
Do I just exclude this one variable from my model or should I use a different test? I am using SPSS.
Thank you in advance
That test is known to be anti- conservative and finds non parallel slopes when in substantive terms the differences are not that great. Howeve, a significant result does not mean that the covariate is unimportant but potentially that it is having a differential effect on the ordinal categories of the response.
I do not know if spss has a facility to allow the extra terms associated with non parallel relations ; conceptually this like a dummy (for each of the ordinal categories, less 1) interacting with the covariate. Practically you could fit a nominal (unordered) multinomial logit which spss does have to see the estimated slopes for each response category; are they substantively different? Of course, you are not then modelling the cumulative logit. Some software allows you to contain slopes to be the same for some covariates in the ordinal model but to be freely estimated for others.
Another way of evaluating the assumption is to fit a set of binary logistic models , see
https://www.restore.ac.uk/srme/www/fac/soc/wie/research-new/srme/modules/mod5/9/index.html
You have need of a little more complex model with different linear regressions on the same explanatory variables. The major complication is one of interpretation of the results that may become a little cumbersome.
What is the sample size? I ask because statistical tests of the assumptions of other statistical tests typically have too little power when n is small and too much power when n is large. I am wondering if the test of parallel lines is over-powered in your case.
Dear Alix Aboussouan ,
As a rule, I would not recommend just excluding one group from a model just because it fails a test. Perhaps its failure has a valid reason.
That test is known to be anti- conservative and finds non parallel slopes when in substantive terms the differences are not that great. Howeve, a significant result does not mean that the covariate is unimportant but potentially that it is having a differential effect on the ordinal categories of the response.
I do not know if spss has a facility to allow the extra terms associated with non parallel relations ; conceptually this like a dummy (for each of the ordinal categories, less 1) interacting with the covariate. Practically you could fit a nominal (unordered) multinomial logit which spss does have to see the estimated slopes for each response category; are they substantively different? Of course, you are not then modelling the cumulative logit. Some software allows you to contain slopes to be the same for some covariates in the ordinal model but to be freely estimated for others.
Another way of evaluating the assumption is to fit a set of binary logistic models , see
https://www.restore.ac.uk/srme/www/fac/soc/wie/research-new/srme/modules/mod5/9/index.html
Dear Kelvyn
I think that a test of likelihood ratio between the original model and other with non parallel regression could be more reliable to detect what model is more plausible that a standar wald test. Do you think the same?.
Guillermo Enrique Ramos ; I think that would be the case. The problem here is the ability of commonly-used software to include the extra terms for differential slopes. I know that MLwiN has this capacity:
http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#discrete
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS