1
$\begingroup$

I have covariates X, Y, and response Z.

X is discrete ("A", "B", "C"), Y is continuous (from 0 - 100).

The issue is that for X = A, Y is usually in the range 0-50. From X = C, Y is usually in the range 50-100. Only for X = B does Y cross the entire 0-100 range.

What does this mean for my analysis? How do I handle this situation? I want to use X and Y as covariates, yet above problem almost seems to indicate that I should do a regression on X and Y (intuitively, if there IS a correlation, then it should probably be X which causes Y, i.e. E(Y) = F(X)).

$\endgroup$
1
$\begingroup$

What you are describing does not seem to indicate correlation of covariates, i.e. multicollinearity. Rather, $X$ indicates a grouping in your data, restricting $Y$.

How to proceed depends on what type of model you are after and what kind of statistical inference you want to make. For the "A, B, C grouping" you may want to model separate slopes for the impact on $Z$, i.e. specify an interaction effect between $X$ and $Y$.

In addition, your study may indicate that there may be missing data problems due to selection bias (A and C are not observed for all values of Y). In the context of treatment evaluation, if A, B and C indicate treatment, you might want to consider the counterfactuals of each of the groups's impact on $Z$ being in another group, based on the observable variable $Y$.

| cite | improve this answer | |
$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.