My friend taught me to use Relative Risk as a guide to check if my coefficients make sense. For example, I have a propensity to default model, where the variable fl_default is equal to 1 if the individual defaulted and equal to zero if he didn't default. I also have the region where individuals belong: A, B, C, D and E - and other variables. In the figure below, I have the number of people that defaulted in each region and their relative risk, and what I understand is that people in region A, B and D are good customers, they have more propensity to pay.

Figure 1: Relative Risk

However, if I run a logistic regression and the coefficient for the predictor REGION is the opposite of this Relative Risk, should I treat this variable (make some modification) or it is possible that the coefficient gives a different direction than the relative risk? What could be happening with my model?

  • $\begingroup$ A more straightforward diagnostic for the sign of the model coefficient is to compare it to the sign of a pairwise, spearman correlation. If they aren't the same, then that can suggest an underlying issue with collinearity. $\endgroup$ – Mike Hunter Oct 12 '16 at 15:46

This is not a direct answer to your question but I suspect it might solve some issues for you.

A relative risk compares two groups so in your situation you could compare the risk of defaulting in group A versus group D. The risk in group A is $\frac{86}{86 + 179}$ and in group D is $\frac{240}{240 + 268}$. The risk ratio is thus the ratio of these two quantities $$\frac{\frac{86}{86 + 179}}{\frac{240}{240 + 268}}$$ If when you work them out correctly you still get an issue then perhaps better to post a new question.


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.