LibraryPublic Health Sciences· 32 of 35
Public Health Sciences

Statistical Measures and Bias

~6 min read7 sections
relative riskodds ratioNNTselection biasconfoundingp-value
⭐ High-yield🎯 Drill Public Health Sciences
Contents (7)

Statistical measures and bias are fundamental concepts in epidemiology and clinical research that determine the validity and applicability of study findings to clinical practice. Bias refers to systematic errors that distort research results in a consistent direction, affecting both internal and external validity of studies. Understanding these concepts is essential for critically appraising literature, interpreting study results, and making evidence-based clinical decisions. Clinicians must recognize how bias can artificially inflate or deflate treatment effects, disease associations, and risk estimates, potentially leading to inappropriate clinical decisions.

Bias operates through systematic deviations in study design, data collection, analysis, or interpretation rather than random chance:

  • Selection bias occurs when the method of selecting study participants creates systematic differences between groups being compared, resulting in non-representative samples that don't reflect true population effects (e.g., Berkson's bias where hospitalized patients have different exposure-disease associations than the general population)
  • Information bias (misclassification bias) results from inaccurate measurement or classification of exposures or outcomes, creating systematic errors in data collection that can mask or exaggerate true associations (includes differential misclassification affecting groups unequally and non-differential misclassification affecting all groups equally)
  • Confounding bias occurs when a third variable (confounder) is associated with both the exposure and outcome independently, distorting the apparent relationship between them and potentially explaining an observed association entirely
  • Surveillance bias arises when differential monitoring or detection of disease occurs across exposure groups (e.g., more intensive screening in intervention groups creates apparent disease excess unrelated to true incidence)
  • Hawthorne effect describes behavioral changes when study participants know they are being observed, altering their exposure or outcome patterns artificially
  • Recall bias introduces systematic error when study participants differentially remember past exposures or events based on current disease status (cases recall exposures more thoroughly than controls)
  • Lead time bias creates apparent survival improvements from screening without actual mortality benefit by advancing diagnosis date rather than extending true survival time

Statistical bias manifests through distorted study results and misleading associations rather than clinical symptoms:

  • Apparent increased association between exposure and outcome when selection bias preferentially recruits exposed cases or unexposed controls, creating false-positive associations (e.g., studying only hospitalized cases may overestimate disease severity)
  • Apparent decreased association or masked true relationships when non-differential misclassification biases results toward the null hypothesis, obscuring genuine causal relationships
  • Spurious causality attribution to exposures that merely coincide with outcomes due to confounding variables (classic example: studying coffee consumption and myocardial infarction without controlling for smoking creates spurious associations since smokers drink more coffee)
  • Temporal relationship confusion in cross-sectional studies creating apparent reverse causality (weight gain appearing to cause disease when disease actually causes weight loss)
  • Critical clinical pearl: Bias is distinct from confounding—while confounding is a real phenomenon that can be controlled for analytically, bias is a systematic error in study design that cannot be corrected after data collection

Recognizing bias requires systematic critical appraisal during study evaluation:

  • Selection bias detection: Evaluate participant recruitment methods, inclusion/exclusion criteria, response rates, and whether enrolled participants differ systematically from non-participants (response rate <70% raises selection bias concern); examine if method of ascertainment differs between comparison groups
  • Information bias detection: Assess measurement validity and reliability, evaluate whether data collection methods were standardized across groups, determine if outcome/exposure assessment was blinded to comparison group status; non-differential misclassification typically biases toward null
  • Confounding assessment: Identify variables associated with both exposure and outcome that aren't in the causal pathway; examine whether confounders were measured and balanced between groups (Table 1 comparison); recognize that randomization protects against unmeasured confounding unlike observational studies
  • Important diagnostic considerations:
  • Relative Risk (RR) and Odds Ratios (OR) can be distorted by any bias type
  • Confidence intervals that exclude 1.0 suggest statistical significance but don't exclude bias
  • P-values reflect random error only, not systematic error from bias
  • Study design choice (RCT vs. cohort vs. case-control) determines bias susceptibility

Management strategies depend on bias type and study stage:

  • Prevention during design phase (most effective): Use randomization to prevent selection and confounding bias; employ blinding (double-blind optimal) to prevent information and surveillance bias; standardize measurement protocols to minimize misclassification; use prospective data collection to reduce recall bias; clearly establish temporal relationships before exposure assessment
  • Restriction strategy: Limit study population to homogeneous subgroup without confounder variation (reduces confounding by removing confounder from population but limits generalizability)
  • Matching in case-control studies: Match cases and controls on known confounders (requires conditional analysis in matched studies); note that matching on confounders requires stratified analysis
  • Stratification (post-hoc analysis): Examine associations separately within strata of potential confounders; calculate Mantel-Haenszel estimate for adjusted effect measures across strata; assess for effect modification (interaction) between exposure and confounder
  • Statistical adjustment/multivariable regression: Adjust for multiple confounders simultaneously using logistic regression or other models; provides adjusted effect estimates and confidence intervals; only controls for measured confounders (unmeasured confounding cannot be addressed)
  • Sensitivity analysis: Test how results change with different assumptions about unmeasured confounding magnitude; assess bias impact on conclusion robustness
  • Special populations/situations:
  • Observational studies: Require more rigorous confounding assessment than RCTs; propensity score matching can help balance confounders
  • Secondary data analysis: Limited ability to prevent bias (information already collected); focus on recognizing and accounting for bias

Unrecognized bias creates serious clinical and research consequences:

  • Inappropriate clinical practice changes resulting from false-positive associations (e.g., recommending treatments based on biased studies leads to unnecessary medication, iatrogenic harm, and wasted healthcare resources; conversely, missing true associations delays beneficial treatment implementation)
  • Public health policy errors when biased population-level data inform screening recommendations, vaccination campaigns, or preventive strategies (lead time bias in cancer screening creates apparent mortality benefits without true lives saved, leading to overdiagnosis and overtreatment)
  • Drug approval/withdrawal decisions based on biased efficacy or safety data, potentially approving ineffective or unsafe medications while rejecting beneficial treatments
  • Perpetuation of healthcare disparities when selection bias systematically excludes minority populations from research, making treatment recommendations non-representative and clinically inappropriate for diverse populations
  • Loss of clinical equipoise and informed consent when biased literature exaggerates treatment benefits, preventing honest discussions with patients about true risks/benefits

  • Non-differential misclassification biases toward the null (typically underestimates true effect), while differential misclassification can bias in either direction—this distinction determines whether negative studies truly exclude an association or simply failed to detect it due to measurement error
  • Randomization prevents selection and confounding bias but doesn't prevent information bias—this is why blinding remains essential in RCTs and why some RCT results remain biased despite random allocation
  • Relative Risk (RR) from cohort studies cannot have RR <0 or >infinity (bound by zero), while Odds Ratios (OR) can range freely—knowing effect measure properties helps identify impossible results suggesting calculation errors or severe bias
  • "Adjust for confounding" doesn't mean include all variables in regression—only include true confounders (associated with exposure AND outcome, not in causal pathway); including mediators or colliders creates bias rather than preventing it (collider bias)
  • Berkson's bias is a USMLE favorite: In hospital populations, two independent diseases become negatively associated (patients hospitalized for disease A are less likely to have disease B since each alone can cause hospitalization)—this is selection bias in action
  • Classic board trap: Confusing confounding with bias; confounding is addressable through design/analysis, but bias is systematic error that cannot be corrected post-hoc—recognizing this distinction separates high-scoring students
  • Surveillance/detection bias explains many "discoveries" of rare diseases after new diagnostic tests; increased detection ≠ increased incidence—critical for interpreting epidemiologic trends

Related topics

← Back to library