{"entity":{"id":"population-versus-individual-risk","kind":"term","name":"Population versus individual risk","aka":["population risk","individual risk","personal risk"],"tldr":"A group event frequency does not determine one person's future. A personal probability estimate needs a suitable model, relevant inputs and evaluation in the setting where it will be used.","summary":"An observed group frequency describes a specified outcome, population, treatment context and time horizon. A probability estimate for an individual adds a prediction model and relevant inputs, with uncertainty. Ranking people well does not establish accurate probabilities, and performance in one hospital, subgroup or treatment era does not establish performance elsewhere. A model predicting outcomes under existing care also does not automatically estimate the benefit of changing that care.","asOf":"2026-11-12","links":[{"label":"NCI: understanding cancer prognosis","url":"https://www.cancer.gov/about-cancer/diagnosis-staging/prognosis"},{"label":"TRIPOD-Cluster: model evaluation across settings","url":"https://www.bmj.com/content/380/bmj-2022-071058"},{"label":"Hernan and Robins: Causal Inference: What If","url":"https://miguelhernan.org/whatifbook"},{"label":"Collins et al.: TRIPOD+AI prediction-model reporting guidance (2024)","url":"https://doi.org/10.1136/bmj-2023-078378"},{"label":"ICH E9(R1): estimands, intercurrent events and sensitivity analysis (May 2021)","url":"https://www.fda.gov/media/148473/download"}],"tags":[],"related":["confidence-interval","hazard-ratio","kaplan-meier-curve"],"cancers":[],"sections":["drug-discovery"],"technologies":[],"targets":[],"drugs":[],"companies":[],"institutions":[],"pathways":[],"terms":[],"trials":[],"people":[],"bottlenecks":[],"keyPapers":[],"journals":[],"dependsOn":[],"notes":["Population risk describes a defined group's outcomes, such as first recurrence by five years after surgery among people receiving specified care. An individual prediction applies a model and assumptions to one person's characteristics. Neither precise printing nor more biomarkers establishes that the estimate fits the situation. The model may omit relevant biology, use imperfect measurements or reflect another treatment era. Calibration asks whether groups assigned similar probabilities experience approximately the corresponding outcome frequencies; discrimination asks whether the model ranks outcomes well. Good ranking does not ensure accurate probabilities. Evaluation outside the development data is important, but performance can still vary by hospital, treatment context or subgroup. Model name and version, required inputs, missing-value rules, population, treatment, endpoint and horizon determine what the probability means, with uncertainty around performance measures.","In a fictional fully followed cohort of 100 people, twenty events by five years give a group proportion of 20%. That does not mean each person experiences one fifth of an event or identify who will have it. Suppose a model assigns 10% risk to a particular profile. About ten events among 100 comparable people in an informative independent evaluation would support calibration near that prediction, without certifying every subgroup or a different treatment sequence. Multiplying the cohort's 20% by a published marker hazard ratio and a treatment hazard ratio does not produce a validated personal probability. The ratios may come from different populations, adjusted models, endpoints and clocks, and a hazard ratio is not a fixed-time risk ratio. These examples are invented and supply no personal prognosis.","A model that ranks outcomes well at Hospital A may need evaluation at Hospital B before its printed probabilities can be trusted there, particularly if treatment and follow-up differ. Relevant calibration, outcome definitions, horizon, inputs and uncertainty matter. A post-treatment test cannot simply replace a model's required pretreatment input, or vice versa. Risk under current care is also different from benefit from changing care: outcome prediction alone does not estimate a treatment effect. An uncertain but appropriately evaluated estimate can help a discussion without promising an outcome. A confidence interval for a group parameter is not a range of individual futures, and a narrow subgroup may contain too little information for precise personalisation.","Before using a calculator, record its exact name, version and recalibration; required input definitions, units, specimen dates and missing-value rules; development population, care era and received treatments; and predicted endpoint, origin and horizon. Ask how independent evaluation checked discrimination, calibration, selected cutpoints and uncertainty in the intended setting. TRIPOD+AI covers reporting of regression and machine-learning prediction studies; it is not a quality appraisal tool or proof that acting on a score helps. If a fictional calculator requires a residual postoperative marker but only a pretreatment result exists, do not substitute it or invent a correction with unrelated hazard ratios. An unknown input is not negative, and a printed 85% remains a model estimate rather than a guarantee.","In a fictional study, a negative blood test is associated with favourable five-year outcomes while everyone receives standard treatment. Those data do not establish that stopping treatment after a negative result is safe. The missing question compares the proposed result-guided strategy with the relevant alternative, with matched sampling time, population, event, uncertainty and follow-up. Pretreatment, on-treatment and surgical samples describe different states. A later landmark includes only people eligible at that time and excludes earlier events; the last sample available during year one is not automatically an early postoperative sample. Analytical reliability, prognosis, treatment-effect prediction and benefit from a test-guided action require distinct evidence."],"provenance":{"editedBy":"OnCo education audit","editedOn":"2026-11-10"},"category":"Methods and models"},"route":"/terms/population-versus-individual-risk/","neighbours":{"term":[{"id":"confidence-interval","kind":"term","name":"Confidence interval","route":"/terms/confidence-interval/"},{"id":"hazard-ratio","kind":"term","name":"Hazard ratio (HR)","route":"/terms/hazard-ratio/"},{"id":"kaplan-meier-curve","kind":"term","name":"Kaplan-Meier curve, censoring and proportional hazards","route":"/terms/kaplan-meier-curve/"}],"section":[{"id":"drug-discovery","kind":"section","name":"Drug Discovery Platforms","route":"/fronts/drug-discovery/"}]}}