Anaes · Measurement & monitoring physics

Bias and confounding in clinical research

Also known as Bias · Confounding · Selection bias · Performance bias · Detection bias · Attrition bias · Reporting bias · Effect modification · Propensity scoring

Every anaesthetic reads the literature, and every paper is threatened by two families of error that can make a useless intervention look effective or a harmful one look safe — bias and confounding. The model rests on eleven exam-critical ideas. First, BIAS is a SYSTEMATIC error that distorts the study findings in a particular direction, unlike RANDOM error (chance) which is unsystematic. Second, SELECTION BIAS is a systematic difference between those selected for study and those not — minimised by randomisation with allocation concealment, and exposed when concealment is subverted. Third, PERFORMANCE BIAS is a systematic difference in the care provided apart from the intervention, minimised by blinding participants and clinicians. Fourth, DETECTION BIAS is a systematic difference in how outcomes are assessed, minimised by blinded outcome assessors and standardised measurement. Fifth, ATTRITION BIAS is systematic loss during follow-up that differs between groups, handled by intention-to-treat analysis alongside per-protocol analysis. Sixth, REPORTING BIAS is the selective publication of favourable results — screened by funnel-plot asymmetry, with prevention by prospective registration and CONSORT reporting as the best defence. Seventh, a CONFOUNDER is a third variable associated with BOTH the exposure and the outcome but NOT on the causal pathway. Eighth, CONFOUNDING is a nuisance to be eliminated whereas EFFECT MODIFICATION is a real interaction to be reported. Ninth, in the DESIGN phase, RANDOMISATION is the only method that controls for both known and UNKNOWN confounders. Tenth, in the ANALYSIS phase, stratified analysis, multivariable regression and PROPENSITY-SCORE methods estimate effects from non-randomised data within their assumptions and limitations. Eleventh, RESIDUAL CONFOUNDING always limits observational studies, which are best read as attempts to emulate the randomised trial that was not done. Built on the Cochrane risk-of-bias tool (Higgins 2011), the allocation-concealment and blinding assessment (Schulz 2001), the allocation-subversion case study (Kennedy 2017), the intention-to-treat versus per-protocol account (Tripepi 2020), the trim-and-fill funnel method (Duval 2000), the funnel-plot education review (Afonso 2024), the mediation-analysis review (Richiardi 2013), the propensity-score author guide (Andrew 2023), the target-trial emulation framework (Hernán 2016), and the CONSORT 2010 reporting guideline (Schulz 2010).