I have a regression setting in which I would like to explore the influence of a given variable on a certain patient condition, given a level of risk for the same outcome, predicted with a more complex non-linear model.

How do I factor out this per-patient risk? should I include it as a covariate (of course it will get a huge effect size, explaining most of the risk) or as an offset (sort of prior risk)?


  • $\begingroup$ When I do this 'iterated modelling' and I want to include the first prediction as a feature for the second one then I always end up not doing it because I simply wonder: what value should I put for the first prediction in the training set? I experimented with putting the actual value plus a little (random but controled) error as a feature value. Pro: you can smoothly select how important this feature will be treated (bigger error = less important). However, it did not perform well and I think that this is the wrong way to do it... 'offset' sounds more reasonable to me. $\endgroup$ – Fabian Werner Feb 14 '18 at 10:34
  • $\begingroup$ So I understood you would suggest using it as an offset then? $\endgroup$ – Bakaburg Feb 14 '18 at 11:43
  • 1
    $\begingroup$ As stated: the only way to include it as a feature I can come up with is to disturb it artificially, so I would proceed as follows: do both and see which one performs better :-) $\endgroup$ – Fabian Werner Feb 14 '18 at 11:47

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.