I’m working on a survey study and my data do not satisfy the proportional odds assumption (checked as Ordinal Logistic Regression | R Data Analysis Examples).
I know that there are more complex models which can allow for partial proportionality or non proportionality but I’d like to keep the presentation of my results clear and straightforward the most. What do you think would be the best way to analyze and/or represent my data (I’m ok also with plots)?
I saw some Authors do multiple logistic regressions for every cutpoint in the likert scale.
I have a series of 5 points likert scale questions with 4 demographic characteristic I want to use as covariates.
There are many ways to look at this excellent question, and thanks for the pointer to the excellent UCLA article. The most cohensive approach is to specify a prior distribution for the amount of non-proportionality through the use of a partial proportional odds model. The R rmsb package blrm function allows you to do this. Short of that, you can either ignore non-proportional odds, knowing that the odds ratio may still be very meaningful, or fit a partial proportional odds model or constrained partial PO model using either Bayesian modeling with blrm or using a frequentist procedure with the VGAM package vglm function.
Analyzing all possible cutpoints would give you the worst of all possible worlds unless the sample size is very large, in which case polytomous logistic regression is the way to put all those together without making any proportionality assumption.
Has there been simulations work done to assess the decrease in power for a proportional odds model when the constant proportional odds ratio assumption is not satisfied? Thanks.
In the unadjusted case the test statistic for OR=1 in the PO model is virtually identical (after a monotonic transformation) to the Wilcoxon test statistic. So you can study either one to determine how much power loss occurs when a certain amount of non-PO. Note that in the PO model you can also add a parameter to model non-PO (partial PO model) if you are willing to devote 2 or more d.f. to treatment.
Hi, I was looking for discussion on an alternative analysis method for ordinal logistic regression (when proportional odds assumption is not met) under multilevel modeling framework when I saw this thread. I was wondering if it is safe to go with mixed effect multinomial regression (gsem) instead, assuming there is no natural order among the categories (e.g., abstinent, low risk drinking and high risk drinking)? I work with Stata, I am not sure if the gologit2 package or its equivalent is available under MLM framework.
thank you!
Wossen