BiostatisticsTier 1Medical Sciences concept

ROC/threshold selection - trade-off between sensitivity and specificity

Core concept

ROC curve plots sensitivity (true positive rate, y-axis) vs (1 - specificity) (false positive rate, x-axis) across every possible test cut-off
Moving the threshold always trades one for the other: lower the cut-off -> inc sensitivity, dec specificity; raise the cut-off -> dec sensitivity, inc specificity -- you cannot improve both by moving the threshold alone
Area under the curve (AUC) summarises overall discriminative ability: 0.5 = no better than chance, 1.0 = perfect discrimination -- independent of any one threshold choice

Youden's index (sensitivity + specificity - 1), maximised, identifies the point closest to the top-left corner -- the statistically "optimal" balanced threshold, but not necessarily the right clinical choice
Where to actually set the threshold depends on the relative cost of a false negative vs a false positive for that specific clinical use, not the geometric optimum
- Missing the diagnosis is dangerous/costly (e.g. ruling out PE, sepsis screening) -> choose a lower threshold favouring sensitivity (accept more false positives)
- False positive triggers harmful/invasive follow-up (e.g. confirmatory biopsy, unnecessary treatment) -> choose a higher threshold favouring specificity
Trap: assuming the mathematically "optimal" (Youden) point on the ROC curve is always the correct clinical cut-off -- it ignores the asymmetric consequences of each error type

Screening tests (aim: don't miss disease) are deliberately set toward high sensitivity -- accepting more false positives that get resolved by a confirmatory second test
Confirmatory/diagnostic tests (aim: don't falsely label someone with a life-changing diagnosis) are set toward high specificity
Comparing two tests: the one with the larger AUC discriminates better overall, but a smaller-AUC test can still be preferred if its curve is superior in the specific sensitivity/specificity region relevant to the clinical question

Threshold choice directly changes the LR+ and LR- at that operating point (see likelihood ratio note) -- moving along the ROC curve moves both LRs simultaneously
AUC is the same "c-statistic" reported for clinical risk prediction models (e.g. cardiovascular risk scores)
SpPin/SnNout logic (a highly specific test's positive result rules in; a highly sensitive test's negative result rules out) is the qualitative version of choosing a threshold along this curve

3 more sections, plus exam facts

Premium unlocks every note across every specialty, and the full exam fact library behind it.