I´m trying to find variables predicting a disease by using first logistic regression for each variable on the disease and then entering the significant variables into a multiple logistic regression model. However, one of the variables in the multivariate model is a clinical score, which contains some of the variables adjusted for (among others bmi<18). My question is: does it make sense to have both in the multivariate model? And if yes, what is the interpretation of an insignificant clinical score in the multivariate model? Is it insignificant because I controlled for some of its variables?


1 Answer 1


Two comments:

First, I would explore other methods of variable selection. Looking at a set of unadjusted regressions and choosing the "significant" variables to then include in a final model is not an ideal approach. Your options are plentiful - search around here for topics on model selection to get you started. There are experts on this topic floating around, so perhaps some of them will add more here.

Second, think about the interpretation of the regression coefficient in a multiple regression. An increase in your clinical score is associated with X change in Y, holding covariates constant. Theoretically, if your clinical score increases, and you hold some component of that score constant, than the estimated association must be due to the other aspects of the clinical scale.

What to do really depends on what your goals are. Are you more interested in predicting things, or making any causal inferences? Can you discard the clinical scale entirely and focus on its components,or are you stuck using it?

  • $\begingroup$ The score has been developed some yers ago and has been validated when using it to monitor patients. Now we would like to see if it has other areas it can e used in, fx in screening a group of patients for this disease. I am also looking at other, by previous research suggested, variables to see what of the current knowledge works best. Apart from the log reg I also use ROC stat of the significant variables. So to answer the last question; I am stuck with the score. $\endgroup$
    – fra
    Commented Jan 9, 2013 at 15:02
  • 1
    $\begingroup$ Then I definitely wouldn't include components of the score as covariates, for the reason stated above. This older but very useful paper might help further expand your horizons a bit (Harrell's 1996 Stats Med paper on multivariable prognostic models - unt.edu/rss/class/Jon/MiscDocs/Harrell_1996.pdf) $\endgroup$
    – D L Dahly
    Commented Jan 11, 2013 at 13:17
  • $\begingroup$ To add to D L's good responses: You never want to have a situation in which, because of multivariate control, you end up controlling part of a relationship right out of itself. Statistical cannibalism, you might call it. What meaning would there be to say the relationship between X1 and Y was ... when part of X1 (embodied in X2 and X3) was being controlled? It would be nonsensical. $\endgroup$
    – rolando2
    Commented Jan 12, 2013 at 1:25
  • $\begingroup$ Thank you both for very useful comments! I read the paper. However, now I am thinking to leave out the log reg completely and focus only the ROC analysis. Would that be enough? Or is it better to do af univariable log reg of all possible variables and then choose the significant ones to assess further in a ROC statistic? $\endgroup$
    – fra
    Commented Jan 13, 2013 at 12:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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