Eduovisual
Biostatistics & Population Health
Recall bias and information bias
— Cases (those with disease) tend to search their memory more thoroughly for exposures than controls, inflating exposure prevalence in cases.
— Result: spurious association or overestimation of odds ratio.
— Study asks mothers of children with birth defects vs healthy children about first-trimester exposures (medications, infections, alcohol).
— Patients with cancer queried about prior dietary habits, occupational exposures, cell-phone use.
— Parents of autistic vs neurotypical children asked about vaccines or pregnancy events.
— Any retrospective questionnaire about emotionally charged or stigmatized behaviors.
— Outcome ascertainment differs by exposure status (surveillance/detection bias).
— Interviewers know group assignment (interviewer/observer bias).
— Self-reported data on sensitive topics (alcohol, sexual behavior, adherence) — social desirability bias.
— Measurement instruments are uncalibrated or applied inconsistently between groups.
Board pearl: Recall bias is the single most-tested bias on case-control studies of congenital anomalies, cancer etiology, and adverse drug effects discovered after marketing. If the stem mentions mothers, memory, or a rare bad outcome being asked about retrospectively — pick recall bias.
Key distinction: Selection bias happens before data collection (who gets in); information bias happens during data collection (how data are measured). Recognizing this temporal split lets you eliminate distractors quickly on exam day.
— Design: case-control or retrospective cohort (memory-dependent) → think recall bias.
— Design: prospective cohort or RCT with non-blinded assessors → think observer/interviewer bias.
— Design: registry or administrative database with differential follow-up → think surveillance/detection bias.
— "Mothers were asked to recall…" → recall bias.
— "The investigator knew which group each subject was in…" → observer bias.
— "Patients on drug X had more frequent clinic visits, and more adverse events were detected…" → surveillance bias.
— "Self-reported condom use / alcohol intake / medication adherence…" → reporting/social desirability bias.
— "Outcome was abstracted from charts; coding differed between hospitals…" → misclassification.
— Non-differential misclassification (random, equal in both groups) → biases toward the null (underestimates true effect).
— Differential misclassification (unequal between groups, e.g., recall bias) → biases away from or toward null unpredictably, usually away (overestimate).
— Medications during pregnancy, vaccines, dietary habits, head trauma history, family history of cancer, occupational chemical exposure.
Step 3 management (study-design version): When designing a study to prevent recall bias, choose a prospective cohort so exposure is recorded before outcome is known. If forced to use case-control, use structured questionnaires, medical records rather than memory, and blinded interviewers.
Board pearl: Non-differential misclassification almost always pushes the measure of association toward 1.0 (null) — making a real effect look smaller. Differential misclassification can go either way, but on the boards, recall bias in case-control studies of birth defects is almost always framed as inflating the apparent association.
— Is it cross-sectional, case-control, cohort, or RCT?
— Retrospective vs prospective? Memory-dependent vs record-based?
— Who collected exposure data? Subject self-report vs chart vs biomarker vs registry?
— Was the collector blinded to outcome status?
— Was the instrument validated and applied identically to both groups?
— Who adjudicated the outcome? Blinded to exposure?
— Were both groups followed with equal intensity (visit frequency, imaging, labs)?
— Was the outcome definition identical across groups and over time?
— Is the misclassification likely differential (group-specific) or non-differential?
— If non-differential and binary exposure → bias toward null.
— If differential → bias direction depends on which group is over- or under-reported.
— Unblinded outcome assessors in an open-label trial of a subjective endpoint (pain, quality of life) → observer bias.
— Differential loss to follow-up >20% with reasons related to exposure → attrition/information bias hybrid.
— Outcome ICD codes used as gold standard without validation → misclassification.
CCS pearl: Treat critical-appraisal questions like a CCS case: order the design, examine the measurement, monitor for differential ascertainment, then diagnose the specific bias. A stepwise approach prevents you from defaulting to "recall bias" for every imperfect study.
Key distinction: Confounding is a third-variable problem (smoking confounds coffee–MI); information bias is a measurement problem. If the stem says "smokers also drank more coffee," that's confounding, not bias.
— Differential memory between cases and controls.
— Setting: case-control, especially with emotionally salient outcomes.
— Direction: usually away from null.
— Data collector's knowledge of group assignment influences measurement.
— Setting: unblinded RCTs, chart abstraction studies.
— Fix: blinding of outcome assessors; standardized instruments.
— Exposed group gets more testing → more outcomes "found."
— Classic example: HRT and endometrial cancer — women on estrogen have more bleeding, more biopsies, more cancers detected.
— Fix: equal follow-up protocols; objective outcome triggers.
— Subjects under-report stigmatized behavior (alcohol, drug use, non-adherence) or over-report virtuous behavior (exercise, vegetable intake).
— Fix: anonymous surveys, biomarkers (urine cotinine, hair toxicology, HbA1c for adherence).
— Subjects change behavior because they know they are being observed.
— Fix: unobtrusive measurement, run-in periods.
— Differential: unequal between groups → unpredictable direction.
— Non-differential: equal between groups → toward null.
Board pearl: Surveillance bias is the favorite "trap" answer when the stem describes a drug's "side effect" detected only because patients on the drug see their doctor more often. Look for differential follow-up intensity as the giveaway phrase.
Step 3 management: When a study's main weakness is recall, the next-best step is to validate exposure with an objective source — pharmacy records, birth certificates, employment logs, or stored biospecimens — and re-analyze.
— Screening detects disease earlier; survival "from diagnosis" appears longer, but date of death is unchanged.
— Classic: prostate cancer PSA screening showing longer survival without mortality benefit.
— Fix: use disease-specific mortality, not survival time.
— Screening preferentially detects slow-growing, indolent disease because fast-growing tumors present symptomatically between screens.
— Result: screen-detected cancers look more curable than they really are.
— Fix: randomized screening trials with mortality endpoints.
— Detection of disease that would never have caused symptoms (e.g., small papillary thyroid cancers).
— Cohort studies where exposure definition requires surviving long enough to receive the exposure (e.g., "patients who received statin within 1 year after MI"); those who died early can't be classified as exposed.
— Fix: time-varying exposure analysis, landmark analysis.
— Early symptoms of undiagnosed disease prompt medication use; medication then blamed for the disease.
— Example: PPI use "causes" gastric cancer — but PPIs were started for early cancer symptoms.
— Fix: lag period between exposure and outcome ascertainment.
— Improved staging reclassifies patients; both old and new stage-groups appear to have better outcomes without any real change.
Key distinction: Lead-time and length-time bias are screening biases — they appear whenever a stem evaluates a new screening test. If the answer choices include "survival was measured from diagnosis," lead-time bias is the answer.
Board pearl: Immortal time bias is the post-MI statin trick. Always ask whether exposure status could only be assigned to patients who survived a waiting period.
— Double-blind RCT with objective endpoint (mortality, lab value): minimal information bias.
— Prospective cohort with biomarker-confirmed exposure and blinded outcome adjudication.
— Open-label RCT with subjective outcome (pain VAS) → observer bias.
— Prospective cohort with self-reported exposure → reporting bias but not recall bias (because exposure recorded before outcome).
— Case-control with retrospective interview about distant exposures → recall bias.
— Registry studies with differential coding/follow-up.
— Cross-sectional surveys on sensitive behavior.
— Recall bias → use medical records, prescription databases, birth registries; blind interviewers to case/control status; use structured, identical questionnaires.
— Observer bias → blind assessors; use objective endpoints.
— Surveillance bias → equal-intensity follow-up; pre-specified outcome ascertainment.
— Reporting bias → anonymous collection, biomarker validation.
— Hawthorne → unobtrusive measurement, run-in periods.
Step 3 management: The single highest-yield "treatment" is blinding — of subjects (placebo), data collectors (interviewers), outcome assessors (endpoint committee), and analysts. Quadruple-blinding addresses most information bias mechanisms simultaneously.
Board pearl: If a question asks "what is the best way to prevent this bias?" and recall bias is the issue, the answer is almost always: obtain exposure data from records collected before the outcome occurred (e.g., prenatal records, pharmacy fills) rather than from patient/parent interviews.
— Single-blind: subjects unaware → reduces Hawthorne, placebo effects.
— Double-blind: subjects + investigators unaware → reduces observer bias, performance bias.
— Triple-blind: + outcome adjudicators → reduces ascertainment bias.
— Quadruple-blind: + data analysts → reduces analytic bias.
— All-cause mortality, lab values, imaging measurements adjudicated by blinded core lab.
— Replaces subjective scales prone to observer bias.
— Identical questionnaires, training of interviewers, calibrated instruments.
— Confirm ICD codes against chart review in a sample; report sensitivity/specificity of case definition.
— Cotinine for smoking, HbA1c or pill counts for adherence, drug levels for pharmacotherapy.
— Identify non-adherent or intolerant subjects before randomization; reduces post-randomization information bias.
— Reduces protopathic and confounding-by-indication issues better than placebo when studying real-world drug effects.
— Outcomes biologically unrelated to exposure used to detect residual bias.
Key distinction: Blinding prevents information bias; randomization prevents confounding and selection bias. A double-blind unrandomized study still has confounding; a randomized unblinded study still has observer bias. Step 3 loves to test this pairing.
Board pearl: When a stem asks about reducing bias in an RCT with a subjective outcome (pain, fatigue, quality of life), the right answer is blinding the outcome assessor, not increasing the sample size — bigger samples don't fix systematic error.
— Vary assumptions about misclassification rates; report range of effect estimates.
— Quantitative bias analysis assigns probability distributions to bias parameters.
— Compare self-report to gold standard in a subset; calculate sensitivity/specificity; correct main analysis using these.
— Adjusts effect estimates for measurement error in continuous exposures (e.g., dietary recall vs doubly-labeled water).
— Handles missing data assumed missing at random; reduces information loss but not systematic bias from non-random missingness.
— Weights subjects to address differential loss to follow-up or differential measurement.
— Uses genetic variants as proxies for exposure to bypass self-report inaccuracy and confounding.
— Outcomes/exposures that should have null association; if non-null, residual bias is present.
— Mostly for confounding, but matching can reduce differential ascertainment when paired with blinded outcome capture.
— Validate exposure or outcome in a subsample with gold standard, extrapolate.
Step 3 management: When a published study shows an association but the editorial criticizes "potential recall bias," the next-best validation step is a prospective cohort using pre-collected administrative or biomarker data, not a larger case-control study. Size doesn't fix systematic error.
CCS pearl: On critical-appraisal items, if the analytic fix matches the bias mechanism (e.g., blinded adjudication for observer bias, biomarker for reporting bias, pre-collected records for recall bias), it's the answer. Mismatch (e.g., "increase sample size" for recall bias) is the distractor.
— Cognitive decline → poorer recall of remote exposures (medications, occupational history, diet).
— May be non-differential if equal in cases and controls (biases toward null) or differential if disease itself affects memory (e.g., dementia cases vs controls asked about midlife exposures).
— Classic trap: case-control of Alzheimer's risk factors with patient self-report — exposure data should come from informants or records.
— Frequent clinic visits, more imaging → incidental findings inflate "disease" prevalence.
— Example: incidentally found renal cysts, thyroid nodules, pulmonary nodules in screened elders.
— When subjects can't self-report (dementia, deceased cases), spouses/children provide data — accuracy varies, often differential between cases (proxy) and controls (self).
— Fix: use proxies for both groups or restrict analysis.
— Often excluded from RCTs → external validity (generalizability) issue rather than internal information bias.
— When included, adherence monitoring is harder (altered drug levels, polypharmacy) → reporting bias on adherence.
— Medication recall is especially poor; pharmacy fill records are gold standard for exposure ascertainment in these populations.
Key distinction: Generalizability (external validity) is harmed when special populations are excluded; information bias (internal validity) is harmed when they are included but mismeasured. Step 3 will offer both as distractors — pick based on whether the stem describes exclusion or mismeasurement.
Board pearl: In Alzheimer's case-control studies, always suspect recall bias and always prefer informant-based or registry-based exposure ascertainment. This is a classic NBME-style stem.
— Mothers of infants with birth defects scrutinize their pregnancies for any unusual exposure (OTC meds, infections, alcohol, caffeine, hot tubs).
— Mothers of healthy infants have no reason to dwell on these details.
— Result: exposures appear over-represented in cases → spurious or inflated association.
— Historical example: Bendectin (doxylamine/pyridoxine) and birth defects — case-control studies suggested risk; later cohort and registry data refuted it. Recall bias contributed to the initial alarm.
— Use prospective pregnancy registries (e.g., antiepileptic drug pregnancy registries, vaccine safety datalink).
— Link to prenatal medical records and pharmacy databases.
— Use prescription claims rather than maternal interview.
— Parental recall about child's diet, behavior, infections is biased by current diagnosis (autism, ADHD, asthma).
— Example: vaccine-autism case-control studies plagued by recall and selection bias; refuted by large prospective registry studies (Danish cohort).
— Substance-use populations: massive social desirability bias; use urine toxicology or hair analysis.
— Sexual behavior surveys: under-reporting in face-to-face interviews vs computer-assisted self-interview.
— Refugee/immigrant populations: language and cultural differences create measurement non-equivalence across groups.
Step 3 management: For teratogenicity questions, the definitive study design is a prospective pregnancy exposure registry linked to delivery records. Case-control studies of birth defects are inherently vulnerable to recall bias and should be considered hypothesis-generating only.
Board pearl: Any vignette involving mothers, memory, and birth defects → recall bias. Any vignette involving registries and prospective enrollment → recall bias is controlled.
— Recall bias suggesting drug X causes birth defect Y → unnecessary regulatory action, patient anxiety, undertreatment of treatable maternal conditions (e.g., depression, epilepsy, hyperemesis).
— Real-world harm: pregnant women avoiding vaccines, antidepressants, antiemetics based on biased data.
— Non-differential misclassification toward null may hide real harms, delaying drug withdrawal.
— Example: under-reporting of adherence in trials may obscure efficacy differences.
— Years spent investigating spurious associations (vaccine-autism literature).
— Conflicting studies due to varying bias profiles → clinical confusion, guideline whiplash.
— Surveillance bias overestimates screening benefit → overdiagnosis, overtreatment, harms of biopsy/surgery.
— Lead-time bias falsely validates screening tests that don't reduce mortality.
— Risk-factor estimates from biased studies inform guidelines; biased estimates → mis-targeted interventions.
— Mass-tort litigation based on biased case-control studies (silicone breast implants and autoimmune disease — refuted by cohort data).
Key distinction: Information bias produces systematic error; random error (chance) produces imprecision. Bigger samples reduce random error but not bias. Confidence intervals narrow but remain centered on a wrong value. Step 3 loves this distinction.
Board pearl: When a stem says "the study found a statistically significant association (p<0.001) but the result is likely wrong because…" — the answer is almost always a bias, not chance. Tiny p-values do not protect against systematic error.
— Case-control study of rare outcome with maternal self-report of exposure, no record validation, no blinding.
— Single-center retrospective chart review with unblinded outcome adjudication.
— Observational study claiming benefit of screening using survival-from-diagnosis as endpoint.
— Registry analyses with potential surveillance bias, awaiting RCT confirmation.
— Pharmacovigilance signals from spontaneous reporting (FAERS).
— Well-conducted prospective cohort with biomarker exposure and blinded outcome adjudication.
— Pragmatic RCT with PROBE design (open-label, blinded endpoint).
— Multicenter double-blind RCT with objective endpoint, low loss to follow-up, pre-registered protocol.
— Meta-analysis of consistent high-quality RCTs.
— Biostatistician involvement at design stage prevents most analyzable bias.
— IRB oversight for ethical data collection.
— Independent Data Safety Monitoring Board (DSMB) for ongoing trials — addresses interim analysis bias.
— Pre-registration (ClinicalTrials.gov), CONSORT/STROBE adherence, transparent reporting of missing data and sensitivity analyses.
Step 3 management: When asked "should this study change practice?" and the design has obvious recall or surveillance bias, the correct answer is "await confirmation from a prospective or randomized study" — not "apply the findings cautiously" and not "perform meta-analysis of similar biased studies."
CCS pearl: Meta-analyzing biased studies amplifies bias; it does not cancel it. "Garbage in, garbage out" applies. Look for this distractor.
— Recall: failure of memory (cognitive).
— Reporting: failure of disclosure (motivational — shame, social desirability).
— A patient who can't remember alcohol intake = recall; one who lies about it = reporting.
— Recall: error originates in the subject.
— Interviewer: error originates in the data collector probing differently across groups.
— Both can co-occur; blinding interviewers addresses the latter, not the former.
— Observer: measurement of the outcome is distorted by knowledge of exposure.
— Detection: outcome is looked for more thoroughly in one group (surveillance frequency).
— Mnemonic: observer distorts, detection discovers.
— Differential: error rate differs by group → unpredictable bias direction, usually away from null.
— Non-differential: error rate equal across groups → bias toward null (for binary exposures with 2-category outcome).
— Hawthorne: behavior changes because of being observed.
— Placebo: physiologic/symptomatic response to inert intervention.
— Both addressed by blinding and control groups.
— Protopathic: subclinical disease symptoms prompt the exposure (drug); a measurement-timing problem.
— Reverse causation: outcome truly causes exposure; a causal-direction problem.
— Overlap exists; lag periods address both.
Board pearl: When the stem features stigmatized self-report (sex, drugs, adherence) in a prospective study, the bias is reporting/social desirability, not recall. Recall requires retrospection.
Key distinction: All recall bias is information bias; not all information bias is recall bias. On Step 3, "information bias" is the umbrella answer; "recall bias" is the specific answer — choose the most specific correct option.
— Systematic error in who enters or remains in the study.
— Subtypes: Berkson's (hospital-based controls), healthy worker effect, non-response, loss to follow-up.
— Cue words: "volunteers," "hospital controls," "responders to survey," "patients who completed the study."
— A third variable independently associated with both exposure and outcome distorts the apparent relationship.
— Example: coffee–MI association confounded by smoking.
— Addressed by randomization (prevention) or stratification/regression (analysis).
— Not a bias — a real biological phenomenon where effect differs by subgroup.
— Report stratum-specific estimates; do NOT "adjust away."
— Reflected in p-values and CIs; reduced by larger samples.
— Not the same as bias; bias is systematic, chance is random.
— Inferring individual-level associations from group-level data.
— Distinct from information bias — it's an inferential error.
— Extreme baseline values move toward the mean on repeat measurement regardless of treatment.
— Cue: studies enrolling subjects with very high or low baseline values, no control group.
— Selective publication of positive results; affects meta-analyses.
— Detect via funnel plots, Egger's test.
Step 3 management: Use this triage on every epidemiology stem — ask sequentially: (1) Who got in? (selection) (2) How were they measured? (information) (3) Is there a third variable? (confounding) (4) Could it be chance? (random error). The first "yes" usually identifies the bias.
Board pearl: Loss to follow-up is selection bias if it removes subjects from analysis; it becomes information bias if the remaining subjects' data are inaccurate. Often both — read carefully.
— Compare results across multiple studies using different designs with different bias structures.
— If RCTs, cohort studies, and Mendelian randomization all point the same direction, the conclusion is robust.
— If only case-control studies show the effect → suspect recall bias.
— Re-analyze excluding subjects with proxy-reported data.
— Re-analyze restricting to record-validated exposures.
— Report E-values quantifying how strong an unmeasured bias would need to be to nullify the result.
— Independent replication in a different population with different measurement methods.
— Failure to replicate is a hallmark of bias-driven findings.
— Tools: Cochrane RoB 2 for RCTs, ROBINS-I for observational studies, QUADAS-2 for diagnostic studies.
— Downgrade certainty (GRADE) when high risk of bias is present.
— Post-marketing registries, pragmatic trials, EHR-based studies with validated phenotypes correct early biased signals.
— CONSORT, STROBE, PRISMA reporting checklists force authors to disclose bias-relevant details (blinding, loss to follow-up, exposure ascertainment).
Key distinction: Internal validity (freedom from bias in this study) must be established before external validity (generalizability) is even meaningful. A biased study cannot be "generalized" — it can only be replaced.
Board pearl: When asked how to address concerns about recall bias in an already-published case-control study, the best answer is to replicate the analysis using a prospective registry or claims database — not to perform a meta-analysis, not to expand the sample, not to add covariates.
— Audit data collection forms; check inter-rater reliability (kappa statistic) periodically.
— Kappa <0.6 suggests measurement inconsistency → retraining or instrument refinement.
— Goal: <20% loss for cohort studies; <5% for RCTs.
— Differential loss by exposure or outcome → bias signal.
— Document reasons for loss; not all loss creates bias (random vs informative).
— Pill counts, electronic monitoring (MEMS caps), pharmacy refills, drug levels.
— Reduces reporting bias in self-reported adherence.
— Have a subset of records dual-coded; report agreement statistics.
— Critical for chart-abstraction studies.
— Survey participants and investigators about perceived assignment at trial end.
— If guessing rate >50%, blinding failed → observer bias possible.
— Informed consent should include realistic expectations to minimize Hawthorne and placebo amplification.
— Standardized scripts across study sites prevent interviewer drift.
— Step 3 expects physicians to read journal articles and identify bias subtypes during MOC, journal club, P&T committee work, and shared decision-making with patients.
Step 3 management: For a journal-club question asking what to monitor in an open-label trial of a subjective outcome, recommend blinded outcome adjudication committee and inter-rater reliability checks — these directly counter observer/information bias.
CCS pearl: Inter-rater reliability (kappa) is a measurement statistic; it tells you about measurement consistency, not measurement validity. High kappa with bad measurement = consistently wrong.
— Subjects must understand if outcome assessors are blinded, if biospecimens will be tested for stigmatized conditions (HIV, substance use), and how their data will be linked to records.
— Withholding methodologic limitations from consent documents undermines autonomy.
— Stronger confidentiality protections (anonymous questionnaires, certificates of confidentiality) reduce reporting bias on sensitive topics — a patient-safety AND validity win.
— Studies on intimate partner violence, child abuse, or substance use during pregnancy may trigger mandatory reporting → respondents under-report → information bias.
— Researchers must disclose reporting obligations in consent, accepting bias as cost of legal/ethical duty.
— Selective outcome reporting (changing primary endpoint after seeing data) is a form of information/analytic bias and constitutes research misconduct.
— Pre-registration on ClinicalTrials.gov is required for FDA-regulated trials and most major journals.
— Discharge decisions based on biased screening data (lead-time bias) may harm patients (e.g., aggressive workup of overdiagnosed cancer).
— QI projects using biased pre/post designs (regression to the mean, Hawthorne) may falsely claim safety improvements; rigorous design protects future patients.
— Industry-sponsored studies more likely to show favorable results; mechanism often through outcome selection, analytic choices, and reporting — categorized as information/analytic bias.
Board pearl: A pre-registered protocol with a Data Safety Monitoring Board, blinded endpoint adjudication, and intent-to-treat analysis is the gold-standard ethical/methodologic package. Stems offering this as an option in trial design are usually correct.
Step 3 management: When a clinician sees a sponsored study claiming a new drug's benefit on subjective endpoints from an unblinded trial, the appropriate action before adopting is to await independent replication with blinded endpoints — not to prescribe based on early-marketing data. This is both an evidence and a patient-safety issue.
Board pearl: Memorize the direction of bias rule: non-differential → toward null; differential → usually away. This single fact answers 30%+ of bias-direction questions.
Key distinction: Sample size fixes chance, not bias. If an answer choice is "increase the sample size" for a bias problem, eliminate it immediately.
— Stem: Mothers of children with cleft palate vs healthy children asked about first-trimester medication use; OR = 3.1.
— Answer: Recall bias. Fix: pharmacy records / prospective registry.
— Stem: Open-label trial of new analgesic with VAS pain score; investigators rate pain.
— Answer: Observer / ascertainment bias. Fix: blinded outcome adjudication.
— Stem: Patients on drug X have higher rate of cancer Y; drug X patients see oncologists 3× more often.
— Answer: Surveillance (detection) bias. Fix: equal-intensity follow-up protocol.
— Stem: New screening test increases 5-year survival from diagnosis from 50% to 80%; mortality unchanged.
— Answer: Lead-time bias. Use disease-specific mortality.
— Stem: Screen-detected tumors smaller, lower grade, better outcomes than symptomatic tumors.
— Answer: Length-time bias. Confirm with RCT mortality endpoints.
— Stem: Cohort defines exposure as "received drug within 1 year post-MI"; exposed group lives longer.
— Answer: Immortal time bias. Use time-varying exposure.
— Stem: PPI users have higher gastric cancer rate.
— Answer: Protopathic bias. Apply lag period.
— Stem: Compliance with hand hygiene rises from 40% to 90% when auditor is visible on the ward.
— Answer: Hawthorne effect. Unobtrusive measurement.
— Stem: Both groups equally misclassified on smoking status; observed OR is 1.2 but true OR is 1.8.
— Answer: Non-differential misclassification → bias toward null.
Step 3 management: On every bias stem, locate the timing (before/during/after data collection), the mechanism (memory, motivation, measurement, follow-up intensity), and the direction (toward or away from null). Three clues, one answer.
Board pearl: "Increase the sample size" is almost never the right answer to a bias question. The right answer always addresses the systematic error mechanism directly.
Information bias is systematic error in how exposure, outcome, or covariates are measured, with recall bias being its case-control archetype where differential memory between cases and controls inflates apparent associations — and the fix is always to match the methodologic remedy to the mechanism, not to enlarge the sample.
Board pearl: If the stem mentions mothers, memory, and a bad outcome — pick recall bias and recommend a prospective registry. If it mentions unblinded subjective endpoints — pick observer bias and recommend blinded adjudication. If it mentions differential follow-up intensity — pick surveillance bias and recommend equal-intensity protocols. These three pattern-recognitions cover the majority of Step 3 information-bias questions and translate directly into the methodologic recommendations you'll make on quality committees, IRBs, and journal clubs throughout your career.