Dear colleagues,
I am running an ordinal regression to find associations between the built environment and the frequency of using transport modes. For several variables and modes the parallel lines assumption is violated. I could use other methods, but for key variables the ordinal model and plots are quite telling and the POL assumotion holds.
I noticed that in most cases grouping the bottom two frequencies eliminates the problem. In other, limiting the analysis to the upper three categories (out of six) allows for POL assumption to hold.
Are both ways acceptable ways to mitigate the issue? Of course this will affect the odd ratios, but as long as it is done consitently for each affected travel mode this should be no issue.
In the first instance, if grouping the variables does not alter the expected range or nature of the variable in question and the sample size is far greater than the expected coefficient.
The same is applied to the second option. However, I am suspecting an error in your data measurement. The error has injected a problem or violated one of the OLS assumptions. CHECK YOUR DATA AGAIN
In the first instance, if grouping the variables does not alter the expected range or nature of the variable in question and the sample size is far greater than the expected coefficient.
The same is applied to the second option. However, I am suspecting an error in your data measurement. The error has injected a problem or violated one of the OLS assumptions. CHECK YOUR DATA AGAIN
Dear Sunday,
thank you so much for the answer. I am fairly sure the data set is correct. The one issue may be several extreme values at the upper end of the data set (but they are "real).
I ran a generalized model (gologit), which solved the issue (obviously), but resulted in issues with Hauck-Donner effect for the same categories. I may just end up regrouping (or transforming the predictor variable in question).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression