Eduovisual
Biostatistics & Population Health
Detection and ascertainment bias
— Detection bias (a.k.a. surveillance or workup bias): exposed/treated group undergoes more testing → more disease "found."
— Ascertainment bias: the way cases or exposures are identified favors one group (e.g., chart review only captures coded diagnoses).
— Postmenopausal estrogen users appear to have ↑ endometrial cancer — but they also bleed more, prompting biopsy (true association exists, but magnitude was inflated historically by detection).
— Oral contraceptive users have more Pap smears → more cervical dysplasia "detected."
— Patients on statins get more LFTs → more "transaminitis" labeled.
— Hospitalized patients screened with CT have more incidental PEs vs outpatients (subsegmental PE epidemic).
— Cancer screening cohorts appear to have ↑ incidence in screened group purely from lead-time and length-time effects (a form of detection bias in screening studies).
— Exposure or intervention itself prompts more clinical contact (visits, labs, imaging).
— Outcome is asymptomatic or subclinical (microalbuminuria, small thyroid nodules, DVT on screening US).
— Diagnosis depends on a subjective or operator-dependent test (echo EF, pathology grading).
Board pearl: If the exposure makes patients get tested more, and the outcome is something only found by testing, suspect detection bias before accepting the association as causal.
— "A retrospective cohort using billing codes found patients on drug X had higher rates of condition Y." → Coding-based ascertainment; drug X patients see physicians more, get coded more.
— "Patients enrolled from a tertiary referral center had higher rates of complication Z than community controls." → Referral (Berkson) bias, a cousin — sicker patients funnel in.
— "Women on HRT had higher rates of breast cancer; they also had annual mammograms included in the prescribing protocol." → Detection bias inflates the RR.
— "A registry of screened smokers showed 5-year lung cancer survival of 70% vs 15% in unscreened." → Lead-time and length-time bias.
— "Cases were identified by self-report, controls by chart review." → Differential ascertainment.
— Was outcome ascertainment blinded to exposure status?
— Were identical diagnostic criteria and intensity of follow-up applied to both groups?
— Was the data source administrative (claims/ICD codes) vs prospective adjudication?
— Were screening intervals the same?
— Did the exposed group have more clinical encounters by design?
Key distinction: Recall bias = patient remembers differently. Detection bias = clinician looks differently. Ascertainment bias = the case-finding system captures differently. All three are forms of information (measurement) bias, distinct from selection and confounding.
— Retrospective + administrative database + observational → high risk.
— Outcome definition reads "any ICD-10 code for…" → ascertainment depends on coding fidelity.
— Exposed group is a clinic cohort; unexposed is the general population → differential surveillance intensity.
— Is the outcome objective and uniformly applied (death, lab cutoff)? Lower risk.
— Is it subjective, requires testing, or only diagnosed if symptomatic? Higher risk.
— Were adjudicators blinded to exposure?
— Mean number of visits, labs, or imaging studies per group should be reported. Asymmetry = detection bias signal.
— Loss to follow-up >20% or differential between arms amplifies ascertainment problems.
— Blinding of outcome assessors (yes/no).
— Standardized protocol for diagnostic workup (yes/no).
— Active vs passive surveillance (active = same testing schedule both arms; passive = wait for clinical presentation, prone to bias).
— Use of an adjudication committee masked to exposure.
Step 3 management: When a vignette asks "which feature of this study most reduces detection bias?" — the answer is almost always blinded outcome assessment with a standardized, protocol-driven evaluation applied identically to both groups, regardless of exposure status. That is the bedside maneuver that fixes the bias.
— Does the exposure alter the probability of being tested for the outcome?
— Does the case status alter how investigators search for the exposure?
— Same test? Same threshold? Same interval? Same observer? Same blinding?
— Compare testing rates between groups (e.g., mammograms/year, troponins drawn).
— A 2-fold testing differential can easily generate a spurious RR ~1.5–2.0 for asymptomatic conditions.
— Did authors restrict to symptomatic cases only? Did they require two independent confirmatory tests? Did they limit to outcomes diagnosed >1 year after exposure to avoid protopathic/surveillance effects?
— Per-arm rates of diagnostic procedures.
— Time-to-diagnosis distribution by group.
— Proportion of outcomes that were incidental vs symptomatic.
— Sensitivity/specificity of the case definition.
— Survival appears longer because diagnosis is earlier, not because death is delayed.
— Fix: report mortality rate, not 5-year survival.
— Screening preferentially captures indolent, slow-growing disease.
— Fix: randomized screening trials with disease-specific mortality as endpoint.
Board pearl: If a screening study reports only 5-year survival or stage at diagnosis rather than all-cause or disease-specific mortality, suspect lead-time/length-time bias — a screening-flavored detection bias.
— Diagnostic suspicion bias: clinician knowing exposure status orders more confirmatory tests in exposed patients. Mitigation: blinded outcome adjudication.
— Exposure suspicion bias: in case-control studies, investigators probe cases harder for exposure history. Mitigation: standardized structured interview, blinding interviewers to case status.
— Workup bias (verification bias): in diagnostic accuracy studies, only patients with positive screening tests get the gold standard. Inflates sensitivity, deflates specificity. Mitigation: apply gold standard to a random sample of negatives too.
— Family information bias: relatives of cases recall family history more thoroughly than relatives of controls.
— E-value: minimum strength an unmeasured factor (or differential measurement) would need to nullify the observed association. Small E-value → fragile finding.
— Quantitative bias analysis: model differential sensitivity/specificity of outcome detection between arms and recompute RR.
— Negative control outcomes: if exposed group also shows ↑ in an outcome biologically unrelated to exposure, surveillance bias is likely (e.g., HRT users also "have more" benign skin lesions detected).
— Consecutive enrollment, prespecified threshold, blinded interpretation, all subjects receive reference standard.
— Capture–recapture methods estimate undercounting.
— Compare registry incidence to population-based estimates.
Key distinction: Verification (workup) bias specifically corrupts sensitivity/specificity of a diagnostic test; detection bias corrupts relative risks/odds ratios in etiologic studies. Same family, different downstream metric — know which the question is asking about.
— Low risk: prospective RCT, blinded outcome assessors, identical protocol-driven testing schedule, objective hard endpoint (death, MI by central adjudication).
— Moderate risk: prospective cohort with standardized follow-up but open-label exposure; outcome requires clinical recognition.
— High risk: retrospective observational using administrative data; outcome is subclinical or test-dependent; exposed group has more medical contact built in.
— Very high risk: case-control with non-blinded interviewers, self-reported exposure, or registry-only case ascertainment from one specialty center.
— If estimated bias magnitude could shift RR across the null (1.0) or across a clinical decision threshold (e.g., NNT vs NNH crossover), the study should not drive practice.
— If the association is large (RR >3), biologically coherent, dose-dependent, and replicated, detection bias is unlikely to fully explain it (Bradford Hill thinking).
— Faced with a vignette study and a clinical decision: ask "is this finding strong enough, and clean enough, to change what I do Monday morning?"
— A single observational study with plausible detection bias → do not change management; await RCT or pooled data.
— Guidelines built on RCTs with blinded adjudication → act.
— Replicated in RCT? → accept.
— Observational only but huge effect + dose response + biologic plausibility? → likely real.
— Observational, modest effect, plausible surveillance asymmetry? → reserve judgment.
Board pearl: Detection bias typically inflates associations toward the alternative; it rarely creates an association from true null in a perfectly balanced exposure — but it absolutely can when testing rates differ several-fold.
— Outcome adjudicators do not know exposure/treatment assignment.
— Single-blind (subject), double-blind (subject + investigator), triple-blind (+ analyst).
— Eliminates diagnostic suspicion bias.
— Identical visit schedule, identical labs, identical imaging at identical intervals in both arms.
— Removes differential surveillance intensity.
— Hard endpoints (all-cause mortality, cardiovascular death by central adjudication, lab value crossing prespecified threshold).
— Composite endpoints adjudicated by independent Clinical Endpoints Committee (CEC).
— Rather than waiting for symptomatic presentation, screen all participants at set intervals (e.g., scheduled venous duplex for DVT).
— Same ICD code definition, same biopsy criteria, same imaging protocol read by central core lab.
— Balances unmeasured factors that drive testing intensity (health-seeking behavior, SES, insurance).
— ClinicalTrials.gov registration; prevents outcome-switching and selective ascertainment.
— Use incident (newly diagnosed) rather than prevalent cases to limit survivorship + ascertainment confound.
— Blind interviewers to case/control status.
— Use structured, identical questionnaires for cases and controls.
— Apply the reference standard to all subjects, independent of index test result.
Step 3 management: When asked "what single design feature best minimizes detection bias?" — blinded outcome adjudication with standardized assessment protocols is the highest-yield answer.
— Pick an outcome biologically unrelated to exposure but subject to the same surveillance. If the exposed group also shows ↑ rate of this control outcome, the original signal is contaminated by detection bias.
— Example: testing whether statins "prevent" dementia — use hip fracture as negative control; if statin users also have less hip fracture in the data, healthy-user/surveillance bias is at play.
— Compare new users of drug A vs new users of drug B (both with similar surveillance footprints) rather than users vs non-users.
— Equalizes healthcare contact and detection opportunity.
— Limit analysis to symptomatic cases, or to cases diagnosed via a single uniform pathway (e.g., ED presentation only).
— Vary assumed sensitivity/specificity of outcome ascertainment across plausible ranges and recompute RR.
— Quantitative bias analysis / probabilistic bias modeling.
— Use a variable that affects exposure but not surveillance to estimate causal effect (rare on Step 3 but worth recognizing).
— Helps balance measured confounders including markers of healthcare utilization (visits, prior labs).
— Caveat: does not fix unmeasured surveillance asymmetry.
— Include number of prior outpatient visits, prior imaging, or prior labs as covariates.
Board pearl: Propensity scores and multivariable adjustment cannot fix detection bias if the unequal surveillance is unmeasured. The cleanest fix is prospective design with blinded, symmetric ascertainment — not statistical wizardry afterward.
— Higher baseline healthcare contact → more incidental findings (thyroid nodules, lung nodules, microalbuminuria, atrial fibrillation on monitors).
— More polypharmacy → more lab monitoring → more "abnormal" labs ascertained.
— Frequent imaging for unrelated indications → incidentalomas drive apparent ↑ disease incidence.
— Studies showing higher AF incidence in patients on cardiology follow-up may largely reflect more ECGs and Holter monitors, not true biology.
— Apparent CKD progression in patients on nephrotoxic drugs may be inflated by more frequent creatinine checks.
— "Subclinical" hypothyroidism prevalence rises with testing frequency, not necessarily with age alone.
— These patients receive mandatory lab monitoring per drug labels → more abnormalities detected → spurious ↑ "drug-induced" AKI or LFT elevation rates.
— Active comparator designs (compare to another renally-monitored drug) help.
— Beware extrapolating observational findings about new-onset disease in highly-monitored populations to the general public.
— Conversely, beware underdiagnosis in less-monitored populations (uninsured, rural) — a flip-side ascertainment bias that underestimates disease burden and understates disparities.
Key distinction: In high-contact populations, detection bias overestimates disease frequency; in low-contact populations, ascertainment failure underestimates it. The same bias mechanism (unequal looking) produces opposite distortions. Step 3 may ask which direction — anchor on who is being looked at more.
— Pregnant women have intensive prenatal surveillance → routine urinalysis detects asymptomatic bacteriuria, routine glucose screen detects GDM, routine BP checks detect gestational HTN.
— Comparing pregnant to non-pregnant cohorts for these conditions is structurally biased.
— Drug safety in pregnancy: post-marketing registries enroll exposed women who are intensively followed; comparator general-population pregnancies have lower ascertainment of minor congenital anomalies → spurious teratogen signals.
— Well-child visits drive ascertainment of developmental delay, BMI categorization, and asymptomatic findings.
— Children in higher-SES families with more frequent visits may appear to have more diagnoses (ADHD, autism) — partly true biology, partly differential ascertainment. Step 3 may probe disparities literacy here.
— Lead-time bias: earlier diagnosis lengthens apparent survival without lengthening life.
— Length-time bias: screening preferentially catches indolent tumors.
— Overdiagnosis: detection of cancers that would never have caused symptoms (most clearly seen in prostate, thyroid, breast DCIS).
— Correct endpoint for screening trials: disease-specific or all-cause mortality in an intention-to-screen randomized comparison (e.g., NLST for lung CT screening).
— "Healthy worker effect" interacts with surveillance: employed cohorts have mandatory medical exams → both healthier baseline and more disease ascertained.
Step 3 management: When a screening intervention is "proven" by improved 5-year survival or earlier stage at diagnosis only, decline to recommend until mortality endpoint RCT data are available. This is a frequent counseling-question right answer (e.g., whole-body MRI marketing).
— Hormone replacement therapy for cardioprotection was supported by observational data inflated by healthy-user and surveillance effects; WHI RCT later showed harm.
— Vitamin E, beta-carotene for cancer prevention — observational signals not confirmed in RCTs.
— Thyroid cancer "epidemic" in South Korea after national ultrasound screening — incidence quintupled, mortality unchanged. Surgeries, lifelong levothyroxine, complications — all driven by detection.
— Prostate cancer overdiagnosis from PSA screening.
— Pulmonary embolism: rising incidence from CTPA sensitivity for subsegmental clots without parallel mortality benefit.
— Public health investment chasing a biased signal.
— Underdiagnosis in low-access populations makes their disease burden look lower than reality, attenuating policy attention.
— Drugs withdrawn or labeled with warnings from biased post-marketing signals.
— Headlines reverse themselves ("coffee causes/prevents cancer"), partially driven by detection bias in observational nutrition epidemiology.
— Incidentaloma on imaging ordered because of an exposure → biopsy → complication → finding was benign.
— This is the patient-safety face of detection bias.
CCS pearl: If a CCS-style stem describes ordering "whole-body CT for screening" or "tumor markers without indication," the right move is to not order, counsel the patient on overdiagnosis and false-positive cascades, and document shared decision-making — directly applying detection-bias literacy at the bedside.
— Single-center observational study driving practice change.
— Effect size implausibly large for the proposed mechanism, or implausibly modest given strong biologic plausibility.
— Outcome ascertainment differs between arms.
— Industry-sponsored with post-hoc subgroup as headline finding.
— Surrogate endpoint substitution (e.g., PSA, HbA1c) without mortality data.
— Conflict between RCT and observational findings.
— Hospital QI dashboards comparing readmissions or HAI rates between units must adjust for surveillance intensity, or the unit that audits more carefully looks worse — punishing diligence.
— Hospitals with active surveillance for CLABSI report higher rates than those with passive surveillance; benchmarking without method standardization is invalid. Escalate to infection prevention committee.
— Pharmacovigilance signal in registry → request post-authorization safety study with active-comparator new-user design, not user vs non-user.
— Recommendations graded on observational evidence (level C / expert opinion) should be applied with humility; escalate to specialty consult or shared decision-making.
— Discordant test results, unexplained incidentalomas, or pressure to over-investigate → specialty consult, multidisciplinary review.
Step 3 management: On exam, when offered "act on this observational signal" vs "recommend confirmatory RCT / continue current standard pending better data" — usually pick the latter, especially when detection bias is plausible and effect size is modest.
— Differential accuracy of self-reported exposure between cases and controls.
— Classic in case-control studies of birth defects (mothers of affected infants recall exposures more vividly).
— Fix: medical-record-based exposure ascertainment; use of incident cases; biomarker confirmation.
— Interviewer aware of case status probes harder.
— Fix: blinding interviewers; structured questionnaires.
— Non-differential: errors equal between groups → biases toward the null.
— Differential: errors unequal between groups → can bias in either direction. Detection bias is a form of differential outcome misclassification.
— Selective reporting of significant outcomes; positive trials more likely published. A literature-level ascertainment problem.
— Subjects change behavior because observed. A measurement-modification cousin, not strictly detection bias but related.
— Drug given for early symptoms of an undiagnosed disease, then "appears" to cause the disease. Distinct mechanism (reverse causation in time) but often co-occurs with surveillance asymmetry.
Key distinction: Non-differential misclassification → bias toward the null (attenuates true effects). Differential misclassification (including detection bias) → bias in unpredictable direction, often away from the null. Step 3 loves this directionality question.
— Error in who enters or remains in the study.
— Berkson bias: hospital-based case-control studies — both cases and controls are sicker than community.
— Healthy worker effect: employed cohorts healthier than general population.
— Loss to follow-up / attrition bias: differential dropout.
— Self-selection (volunteer) bias: screening trial volunteers are healthier.
— Immortal time bias: misclassified person-time in pharmacoepi.
— Fix: random sampling, population-based controls, intention-to-treat analysis, defined cohort entry.
— A third variable independently associated with both exposure and outcome distorts the association.
— Classic: coffee–lung cancer association confounded by smoking.
— Fix: randomization (best), restriction, matching, stratification, multivariable adjustment, propensity scores, instrumental variables.
— "More cases were detected in the exposed group because they had more frequent screening" → detection bias.
— "More exposed individuals enrolled because they were referred from a specialty clinic" → selection bias.
— "The association disappeared after adjusting for age" → confounding.
— Not a bias — a real biological phenomenon where exposure effect varies by a third variable. Do not "correct"; instead, report stratified estimates.
— Random shrinks with sample size; systematic (including detection bias) does not — larger N just gives a precise wrong answer.
Board pearl: A bigger study does not fix detection bias — it makes the biased estimate more precise. Only better design (blinding, symmetric ascertainment, randomization) fixes it.
— Read the Methods section first, not the abstract.
— Ask: who was assessed, how often, with what test, by whom, blinded to what?
— Default to RCT-grade evidence for changes in management; treat observational signals as hypothesis-generating.
— Apply Bradford Hill viewpoints (strength, consistency, temporality, dose-response, plausibility, coherence, experiment, analogy, specificity) when interpreting observational data.
— Standardize surveillance protocols across units before comparing rates (CLABSI, CAUTI, readmissions).
— Risk-adjust outcomes using validated methods, not raw counts.
— Use run charts and control charts that account for surveillance intensity changes.
— Maintain population-based registries with standardized case definitions.
— Mortality endpoints in screening evaluations; resist 5-year survival as primary metric.
— Discuss overdiagnosis risk when ordering screening tests (USPSTF shared decision-making framework — PSA, lung CT, mammography in 40s).
— Avoid low-value testing in asymptomatic patients (Choosing Wisely).
— Use GRADE framework; downgrade observational evidence for risk of bias.
— Track replication: single observational finding ≠ established truth.
Step 3 management: For a patient asking about a "new study showing drug X prevents Alzheimer's" — counsel that observational findings often reverse in RCTs (e.g., hormone therapy, vitamin E), recommend continued standard care, and document shared decision-making. This is high-yield outpatient EBM behavior.
— Was outcome assessment blinded? Y/N.
— Were diagnostic criteria standardized and applied identically? Y/N.
— Was follow-up symmetric in frequency and intensity? Y/N.
— Was loss to follow-up <20% and balanced between arms? Y/N.
— Were endpoints adjudicated by an independent committee? Y/N.
— Was the trial pre-registered with prespecified outcomes? Y/N.
— Any "No" answers → downgrade confidence accordingly.
— CONSORT (RCTs), STROBE (observational), STARD (diagnostic accuracy), PRISMA (systematic reviews), TRIPOD (prediction models).
— Each has explicit items addressing ascertainment symmetry.
— Risk of bias (incl. detection bias), inconsistency, indirectness, imprecision, publication bias.
— Subscribe to journals with structured methodologic critique (NEJM Journal Watch, ACP Journal Club).
— Annual review of major guideline updates (USPSTF, AHA/ACC, ADA).
— When deviating from standard care because of media-driven study, slow down — schedule a follow-up visit, review primary literature, discuss with patient.
— Document shared decision-making, especially for screening tests with known overdiagnosis risk.
— Review your own ordering patterns for low-value screening; ascertainment bias also lives in your own clinic.
Board pearl: USPSTF grades reflect net benefit considering overdiagnosis and bias risk. Grade D ("don't do it") often exists because earlier observational evidence was undone by bias-aware analyses (e.g., PSA screening for low-risk men, ovarian cancer screening).
— Patients have a right to know overdiagnosis and false-positive cascade risks, not just sensitivity/specificity.
— Example: Before ordering PSA in a 55-year-old, document discussion of mortality benefit (modest), overdiagnosis (substantial), and treatment side effects (incontinence, ED). Shared decision-making is the standard.
— Whole-body MRI marketed direct-to-consumer — duty to counsel against in asymptomatic average-risk patients due to inevitable incidentalomas and cascade harm.
— Public health reporting (e.g., communicable diseases) creates ascertainment differentials across jurisdictions. Clinicians should report uniformly per state law regardless of patient characteristics to avoid amplifying disparity-driven ascertainment bias.
— Patients discharged with incidental findings (e.g., lung nodule on trauma CT) must have explicit follow-up documented and communicated to PCP and patient.
— Failure to communicate incidentalomas is a leading cause of diagnostic-error malpractice claims. Use closed-loop communication and tracking systems.
— Differential ascertainment across race, insurance, and geography produces biased registries and biased AI/risk-prediction tools trained on those data.
— Algorithms trained on under-ascertained populations systematically under-predict disease risk in those groups — a patient safety hazard.
— IRBs require equitable subject selection; differential ascertainment can constitute injustice under Belmont principles.
— Industry-sponsored studies with non-blinded outcome assessment are particular hazards; disclosure mandated.
CCS pearl: A patient with an incidental 6-mm lung nodule on ED imaging needs explicit follow-up plan (Fleischner Society criteria), PCP communication, and documented patient counseling before discharge — closing the ascertainment loop is a patient-safety act, not optional.
— HRT ↔ endometrial cancer detection (bleeding prompts biopsy).
— OCPs ↔ cervical dysplasia detection (more Pap smears).
— Statins ↔ transaminitis label (more LFTs drawn).
— Cardiology follow-up ↔ apparent ↑ AF incidence (more ECG/Holter).
— Screening cohort ↔ inflated 5-year survival (lead-time + length-time).
— South Korean thyroid cancer "epidemic" ↔ population ultrasound screening (classic overdiagnosis case).
— PSA screening ↔ prostate cancer overdiagnosis; USPSTF grade C for ages 55–69, D for ≥70.
— Tertiary referral center series ↔ Berkson bias.
— Hospital-based case-control study ↔ Berkson bias.
— Retrospective claims database study ↔ coding-based ascertainment risk.
— Self-reported exposure in case-control ↔ recall bias.
— Unblinded outcome adjudication ↔ diagnostic suspicion bias.
— Workup bias ↔ diagnostic test accuracy studies (verification bias).
— Non-differential misclassification ↔ bias toward the null.
— Differential misclassification ↔ bias in any direction.
— Negative control outcome ↔ tool to detect surveillance bias.
— Active comparator new-user design ↔ pharmacoepi fix for healthy-user/surveillance bias.
— Loss to follow-up threshold of concern: >20% or differential >5% between arms.
— Bradford Hill criteria: 9 viewpoints, not strict rules.
— Quantitative bias analysis is increasingly expected in observational pharmacoepi.
— HRT cardioprotection (observational → WHI RCT harm).
— Class I antiarrhythmics post-MI (surrogate endpoint → CAST mortality).
— Routine PSA screening (observational survival → RCT modest/no mortality benefit).
Board pearl: When in doubt, the answer involving blinded outcome adjudication, prospective randomization, or mortality endpoint is usually the right "fix" choice.
— Stem: A cohort study finds patients on drug X have higher rates of microalbuminuria. Drug X requires quarterly urine monitoring per FDA label; controls have urine checked only when symptomatic.
— Question: Which bias most likely explains this association?
— Answer: Detection (surveillance) bias.
— Stem: Non-differential misclassification of outcome status.
— Question: Effect on observed relative risk?
— Answer: Bias toward the null (1.0).
— Stem: Open-label trial of a new antihypertensive; investigators measure BP at follow-up visits.
— Question: Which most reduces detection bias?
— Answer: Blinded outcome assessment (or automated home BP monitoring with prespecified protocol).
— Stem: Screened smokers have 5-year lung cancer survival of 70% vs 15% in unscreened.
— Question: Best explanation?
— Answer: Lead-time bias (and possibly length-time / overdiagnosis).
— Stem: Mothers of infants with birth defects recall medication use more thoroughly than controls.
— Answer: Recall bias, not detection bias.
— Stem: Two ICUs report different CLABSI rates; one uses active blood-culture surveillance.
— Question: Best interpretation?
— Answer: Ascertainment differs; standardize surveillance before comparing.
— Stem: Patient asks about whole-body MRI screening after seeing an ad.
— Question: Best response?
— Answer: Counsel against; discuss overdiagnosis and incidentaloma cascade harms.
— Stem: Investigators compare new users of SGLT2 inhibitors to non-users of any diabetes drug.
— Question: Improvement?
— Answer: Active comparator new-user design (e.g., vs DPP-4 inhibitors).
Step 3 management: When stem provides a study and asks "what would you tell the patient?" — usually continue evidence-based standard care; do not change practice based on a biased observational signal.
Detection and ascertainment bias is differential outcome (or exposure) misclassification arising when one group is looked at more carefully, more often, or with a more sensitive method than the other — inflating or distorting associations in ways no amount of sample size or post-hoc adjustment can fix, and curable only by prospective, blinded, symmetric, protocol-driven ascertainment.
Board pearl: If the exam asks how to fix it, prospective randomization with blinded outcome adjudication and symmetric protocol-driven assessment is the canonical right answer; if it asks what to tell the patient, shared decision-making about overdiagnosis and adherence to USPSTF/guideline-grade evidence is the canonical right answer — those two reflexes will carry you through every Step 3 detection-bias question you will ever see.