I have 1 cohort of patients, each and every one having suffered an outcome event Y. I split this cohort into two groups of patients, both with the same disease but different variations (so say cancer size small vs. cancer size big).

I want to know whether a set of variables increase the odds of having outcome Y, in cancer size small vs. cancer size big. I.e. does smoking (for example) in those with small cancer size increase their risk of outcome Y, vs those with big cancer size.

I'm a bit lost as to whether to use ANOVA or logistic regression as my model. I've usually used logistic regression so I don't want to get stuck thinking everything should be modeled with logistic regression.


1 Answer 1


if your outcome Y is dichotomous then use logistic regression (it seems like you want a subgroup analysis ie interaction will give separate estimates of the OR for smoking for 'big' and 'small' size). ANOVA is just regression, with Y continuous.

  • $\begingroup$ Right. What if I decide to keep the original groups instead of dichotomizing them, so cancer size 1-2cm, cancer size 2-3 cm, cancer size 3-4cm etc. Would that be ordered probit regression? I see on this website: stats.idre.ucla.edu/other/mult-pkg/whatstat that Kruskal-Wallis would be recommended, which is a form of ANOVA? $\endgroup$
    – Paze
    Commented Oct 28, 2019 at 13:43
  • $\begingroup$ then the OR for smoking depends on continuous size and you'd likely plot OR (y-axis) against size (x-axis). Your outcome determines the analysis (logistic regression v anova), and size is not your outcome, so i'd remove that confusion from your thinking. I'm not sure what your outcome is (you havent said, maybe for confidentiality etc), and i assume it's dichotomous $\endgroup$
    – pau13rown
    Commented Oct 29, 2019 at 12:27
  • $\begingroup$ Sorry I mean if the outcome is ordered categorical. So does smoking affect cancer size, with cancer size as dependent with say 4 levels. $\endgroup$
    – Paze
    Commented Oct 29, 2019 at 12:41
  • $\begingroup$ oh! i misunderstood from the outset. Yes, it would be proportional odds modelling. Kruskall-wallis won't do it because it doesn't allow for covariate adjustment and you mentioned a 'set of variables' $\endgroup$
    – pau13rown
    Commented Oct 29, 2019 at 13:21

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.