Diagnostic Cutoff Scores and Classification Accuracy

Core Idea

Clinical cutoff scores define pass/fail or disorder decisions. Optimal cutoffs balance sensitivity (true positive rate) and specificity (true negative rate), using receiver operating characteristic curves. A single cutoff reflects a chosen trade-off; raising cutoffs increases specificity but decreases sensitivity. Reporting confidence intervals and misclassification rates improves ethical use.

How It's Best Learned

Generate ROC curves for actual test data, calculate sensitivity/specificity at various cutoffs, and discuss practical implications of different choices for stakeholders.

Explainer

You already know from the standard error of measurement that no test score is a perfect reflection of true ability — every score is a sample from a distribution of possible scores, with measurement error around it. That measurement error matters enormously when you need to make a yes/no decision: does this person have a disorder? Do they qualify for services? A diagnostic cutoff score converts a continuous scale into a binary classification, and the central question is: at what point do you draw the line, and what are the consequences of being wrong?

There are two types of classification error. A false positive occurs when someone without the condition is classified as having it. A false negative occurs when someone with the condition is missed. Neither error is neutral: false positives may lead to unnecessary treatment and stigmatization; false negatives leave real conditions undetected and untreated. Sensitivity measures how well the test identifies true cases — formally, the proportion of actual positives correctly classified (high sensitivity means few false negatives). Specificity measures how well it excludes non-cases — the proportion of actual negatives correctly classified (high specificity means few false positives). These two quantities are structurally in tension: lowering the cutoff score lets more people through (higher sensitivity, lower specificity); raising it excludes more people (lower sensitivity, higher specificity).

The ROC (Receiver Operating Characteristic) curve visualizes this trade-off across all possible cutoff values. For each candidate cutoff, you plot sensitivity on the y-axis against the false positive rate (1 − specificity) on the x-axis. A test with no diagnostic value falls along the diagonal — at any sensitivity level, you achieve the same false positive rate by chance. A useful test curves toward the upper-left corner. The area under the ROC curve (AUC) summarizes overall diagnostic accuracy in a single number: 0.5 is chance performance, 1.0 is perfect discrimination, and values above 0.70 are generally considered clinically useful.

Choosing the optimal cutoff is a values judgment, not a purely statistical one. For a screening test aimed at a serious, treatable condition — suicidality, early-stage cancer — you prioritize sensitivity. You'd rather have false alarms than miss real cases, especially when false positives can be filtered by follow-up assessment. For a test used to allocate a scarce benefit, you might prioritize specificity. These priorities should be explicit and transparent — not hidden inside a number that appears neutral. Because measurement error creates a band of uncertainty around any score, cutoffs should never be applied mechanically; reporting confidence intervals around classification decisions and acknowledging the standard error of measurement are basic requirements for ethical diagnostic practice.