Cutoff

What a cutoff is

A cutoff is the specific value used to turn a continuous measurement, such as a biomarker level or gene copy number, into a decision rule: values at or above the threshold may be called test positive, and values below it test negative. In ROC (Receiver Operating Characteristic) analysis, the purpose of the cutoff is to help distinguish diseased from non-diseased groups or, more generally, cases from controls.

[1] How reliable are ROC cut-offs? Evidence from simulation and ... [2] Guidelines for setting cut-off scores in AUC (AUC-GUIDE) - PMC

How cutoffs are chosen

Cutoff selection is usually a trade-off between sensitivity and specificity. Sensitivity, or true positive rate, is the proportion of actual positives correctly identified; specificity is the proportion of actual negatives correctly identified. A common choice is Youden’s index, which is sensitivity plus specificity minus one; the cutoff that maximizes it is often treated as the best balance when false positives and false negatives are considered similarly important. Other rules exist, including accuracy-based, agreement-based, distance-based, and ratio-based methods, but they can give different answers.

[1] How reliable are ROC cut-offs? Evidence from simulation and ... [2] Guidelines for setting cut-off scores in AUC (AUC-GUIDE) - PMC

What AUC adds and what it does not

AUC, the Area Under the Curve, summarizes the whole ROC curve into one number from 0.5 to 1.0. It describes overall discriminative ability, where 0.5 is no better than random classification and values closer to 1.0 indicate better separation. However, AUC does not directly tell you the best single threshold, so a cutoff still has to be chosen for classification.

[2] Guidelines for setting cut-off scores in AUC (AUC-GUIDE) - PMC

Limits and context

A cutoff is context dependent. In screening, missing cases may be more harmful, so higher sensitivity may be preferred; in settings with costly follow-up tests or interventions, higher specificity may be preferred. Prevalence, meaning how common the condition is in the tested population, can affect predictive values and the practical usefulness of some cutoff methods. Simulation and real-data comparisons found Youden’s index and the product of sensitivity and specificity to be among the more stable choices, while some other methods varied more across settings. Retrospective studies with limited sample sizes, platform differences, and lack of external validation also limit how confidently a proposed cutoff can be used.

[1] How reliable are ROC cut-offs? Evidence from simulation and ... [2] Guidelines for setting cut-off scores in AUC (AUC-GUIDE) - PMC [3] Exploratory Identification of Gene Copy Number Cut-Off ... - PMC

Sources and further reading

This general lesson was generated from public source material. It explains a concept; it does not interpret a private paper or provide medical advice.

1. How reliable are ROC cut-offs? Evidence from simulation and ...
2. Guidelines for setting cut-off scores in AUC (AUC-GUIDE) - PMC
3. Exploratory Identification of Gene Copy Number Cut-Off ... - PMC