I was having trouble finding this exact circumstance; hopefully I haven't missed an obvious previous answer.

I have a target variable, Y (discrete counts), and two potential regressors, A (discrete counts) and B (continuous). There's an obvious, necessary causal relationship A -> B from the data generating process, and B is in fact highly correlated with A. (R ~= .95)

My aim in the analysis is to determine whether B is a significant factor in addition to A, so the collinearity is problematic. My first pass (feel free to suggest principled alternatives if this seems kafka; not a stats wiz yet) is to fit a negative binomial regression essentially as Y ~ A + B, and determine whether the value fit to B has a significant value.

However, the collinearity makes the regression suspect if the ordinal variable A is treated as a continuous variable. In this circumstance, is it acceptable and useful to treat A as a categorical variable using dummy encoding?

At a glance, this seems to ensure B can't be collinear with any level of A. But that seems suspiciously like a free lunch to me - am I missing something that may be problematic here?

(I am aware of dummy trap, but unless I'm mistaken, that seems like a separate and easily avoided encoding issue...)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.