observer bias

**Definition:** Observer bias (also called detection bias, ascertainment bias, or interviewer/observer-expectancy bias) is a systematic error in data collection or outcome assessment caused by the observer's preconceived notions or knowledge of a subject's exposure, group assignment, or hypothesis. It falls under *measurement/information bias*.

**Mechanism:** When the assessor knows which group a participant belongs to (e.g., treatment vs. placebo, exposed vs. unexposed), they may unconsciously probe more thoroughly, interpret ambiguous findings in the expected direction, or record data differently between groups. This produces differential misclassification, which can bias results toward or away from the null.

**Related concepts (commonly tested distinctions):** - **Pygmalion effect (observer-expectancy effect):** The researcher's belief in the intervention subtly influences participant performance or outcomes (a form of observer bias). - **Hawthorne effect:** Participants change their behavior because they know they are being observed. This is a *subject* effect, not an observer effect. - **Recall bias:** Participants with disease recall past exposures differently than controls (typical of case-control studies). - **Interviewer bias:** The interviewer elicits information differently from cases and controls. - **Placebo effect:** Subjects' expectations alter their reported or actual outcomes. - **Detection/surveillance bias:** One group is monitored more closely, so outcomes are found more often (e.g., more screening in exposed patients). - **Lead-time bias and length-time bias:** Screening biases, not observer biases. - **Selection bias:** Error in how participants are chosen or retained (e.g., Berkson, attrition, nonresponse). It is distinct from measurement bias.

**Clinical/study design significance:** - Most problematic with subjective outcomes (pain scales, radiograph interpretation, physical exam findings, symptom severity ratings) and in unblinded studies. - **Prevention/mitigation:** - **Blinding (masking):** Single-blind (participants unaware), double-blind (participants and investigators/assessors unaware), and triple-blind (also the data analysts). Double-blinding is the key fix for observer bias. - Use objective, standardized outcome measures (e.g., lab values, automated readings). - Standardize protocols and train observers. - Use independent, blinded outcome adjudicators. - Use multiple observers and assess inter-rater reliability (kappa). - **Randomization** addresses confounding and selection bias but does *not* by itself prevent observer bias. Blinding does.

**USMLE high-yield pearl:** A question stem where the evaluating physician knows which patients received the drug and rates improvement more favorably in that group → **observer bias** (fix: double-blinding). If the *patients* change behavior because they know they're being studied → **Hawthorne effect**. If the *researcher's expectations* alter the participants' performance → **Pygmalion effect**.