I'm trying to construct a medical risk score. I was given some advice by a statistician and they said that one of the stages after the variable selection stage is to take the regression coefficients of those variables in the model and then divide the variables by the minimum absolute value of all the regression coefficients. Does anyone know what this is about?

  • 1
    $\begingroup$ It depends on what you do next--what would that be? $\endgroup$ – whuber Dec 5 '19 at 17:58
  • $\begingroup$ @whuber From there you are supposed to get the "score" for each variable and then you sum those values together to get someone's risk score $\endgroup$ – 762 Dec 5 '19 at 18:54
  • 1
    $\begingroup$ The details matter: how do we "get" the score from these normalized values? What do you do when the coefficient with smallest absolute value is zero? Do you include the value of the intercept term (if any) in that calculation? Are you aiming for a risk score that lies between determined limits like 0 and 10, for instance? $\endgroup$ – whuber Dec 5 '19 at 18:59
  • $\begingroup$ The model is a logistic regression model with dichotomous predictors. I think they use log odds here because they mentioned to use absolute value. It sounds like they want to divide all the coefficients by the smallest coefficient and round up the numbers to get scores, but that seems odd to me $\endgroup$ – 762 Dec 6 '19 at 23:25
  • $\begingroup$ I agree--it's hard to imagine even a plausible misunderstanding that would lead to such a recipe. The best I can come up with is that after dividing by the smallest absolute value and then rounding everything, all coefficients are integral. But there are much better ways to create integral coefficients to approximate a regression! $\endgroup$ – whuber Dec 6 '19 at 23:38

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.