I the clinical medical imaging literature I often read that certain imaging parameters "predict" certain outcomes better than other imaging parameters.
These conclusions are often drawn from Area under the curve (AUC) calculation using Receiver operating characteristic (ROC) analysis with a single continuous parameter (e.g., length of a tumour on a computed tomography (CT) scan) and a binary outcome (e.g., disease A present / disease A not present), without performing/reporting a model developement, internal or external validation.
A typical conclusion would be: "Parameter A predicts disease C better (AUC 0.7) than parameter B (AUC 0.6)".
I recently read To Explain or To Predict? and Practical thoughts on explanatory vs. predictive modeling. I also read about how to quantify added predictive value of new measurements
However, I am still confused.
- Is the depicted conclusion above correct?
- If it is not correct, which wording would be correct? (I think it should be: parameter A has a highter discrimination ability than parameter B regarding disease C. However, I am not sure if we can call it a predictive discrimination ability)
- I understand that the use of AUC might not be a good measure to quantify predictive information of a single parameter. Which approach would be better? In clinical practice, a Radiologist would rate a CT scan for presence of disease C using an ordinal scale (e.g, presence of disease C is unlikely, unclear, likely). Analyses as depicted above are performed to find parameters that could improve such a rating.