I need to study which variables influence the change in a score after a treatment.

The problem is that the score (from 0 to 18) is actually the sum of six 4-levels (0 to 3) ordinal variables which represent the severity of a group of symptoms.

Which regression can help me model this outcome? Given the non-continuous nature of such dependent variable, an ordinal regression should be appropriate. But since we have 19 possible levels for the outcome maybe it's not the best choice (it's more a feeling, I would be glad if someone could formally explain me why). So maybe OLS linear regression is ok in this setting, or even better quasi-poisson, since the value can only be positive.

By the way, I know that summing up ordinal scores doesn't make really sense, but I need the have just one model, not one for each fo the six symptoms variables. If you have alternative designs to suggest for this kind of problem please do suggest it.


If my understanding is right, you have six measurements for each subject/patient. Then these outcomes from the six measurements may be highly correlated.

On the top of my head, you may want to try multivariable regression.

See here

  • $\begingroup$ Uhm, the effect sizes I would extract from a multivariate analysis would be the same as doing six separate linear regression, isn'it? $\endgroup$ – Bakaburg Jun 1 '16 at 15:30
  • $\begingroup$ Probably not, I would expect using multivariate analysis will incorporate the correlation among the six dependent variables, thus, given different estimates of the effect size from fitting six model separately. $\endgroup$ – Runfirst Jun 3 '16 at 3:33

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.