ROC curves - conditional logistic regression

[Peter Baade]

I am running a conditional logistic regression in Stata 15.1, with cases and controls matched by the variable id_cases

The syntax for the model is:
clogit casecontrol i.thick i.level i.ulceration ib2.morp ib4.subsite, group(id_cases) or

I was then able to generate a ROC curve from this model using the following commands:
predict p
roctab casecontrol p
roctab casecontrol p, graph

I would value guidance on
1. Is generating an ROC curve for a conditional logistic regression model statistically valid?;
2. If so, is the above method appropriate?; and
3. I am assuming the above method would not calculate appropriate confidence intervals for the ROC curve estimates. Is this correct?

Thanks for any advice.

Peter.

[Clyde Schechter]

No it's not valid. Here's why. The "p" that you can predict after -clogit- is not the probability of a positive response. Rather it is, among the group of observations that this observation belongs to, what is the probability that this particular one will have a positive response. It is not the unconditional probability of a positive response that an ROC curve would be based on.

[Peter Baade]

Thank you, Clyde, for your prompt and helpful response.