Selection Bias

Selection bias is a systematic error that arises when the way participants enter or stay in a study distorts its estimate of an association or effect.^[2][4] It occurs through flaws in how participants are selected, recruited, assigned to groups, retained, or analyzed, producing study groups that are not comparable and compromising the internal validity of the findings. For an association or treatment effect, the distortion arises when selection into the study or the analysis is related to both the exposure and the outcome, or when randomized groups differ because allocation was not concealed; for a descriptive estimate such as a prevalence, selection related to the outcome alone is enough. A sample that differs from the target population in other ways limits external validity rather than biasing the comparison.

Selection bias is one of the most pervasive threats to the validity of research findings. In randomized trials, it can occur if allocation is not properly concealed. In observational studies, it can arise when the way participants enter the study, or stay in it, is related to both the exposure and the outcome, as in some hospital based samples. In surveys, it can result from nonrandom sampling or low response rates. Selection bias cannot be corrected by increasing sample size, as it is a systematic rather than random error.^[2] Identifying and mitigating selection bias is essential for producing trustworthy evidence.

Selection bias takes several forms depending on the study design. In randomized trials, inadequate allocation concealment allows enrolling clinicians to influence which patients receive which treatment, undermining randomization. In observational studies, a closely related problem is confounding by indication, which occurs when the factors that determine treatment assignment also affect the outcome; epidemiology classifies it as confounding rather than selection bias, although some texts group the two.^[4] In cohort studies and trials, loss to follow up creates attrition bias when dropout is related to both the exposure (or treatment) and the outcome, so that those who remain are no longer comparable.^[4][2] In case control studies, differential selection of cases and controls can bias the odds ratio estimate.^[2][4]

Berkson's bias (admission rate bias) occurs when hospital based samples are used to study disease associations and the probability of being hospitalized depends on both the disease and the exposure (often another condition), which can create or distort an association that does not exist in the population.^[2] Volunteer bias arises when study participants who self select differ from the general population. Survivorship bias occurs when the sample includes only individuals who survived to the point of observation, excluding those who died earlier.

Strategies for preventing or mitigating selection bias include randomization with proper allocation concealment, drawing cases and controls from the same source population, minimizing loss to follow up, statistical weighting (such as inverse probability weighting) to adjust for known differences between the sample and the population or between those retained and those lost, and sensitivity analyses for selection that depends on unmeasured factors.^[2][4] Propensity score methods and sensitivity analyses for unmeasured confounding address confounding, which is often discussed alongside selection bias.

Consider selection bias as a potential threat in every study design, and take steps to minimize it during the design phase.^[1][2] In randomized trials, ensure rigorous allocation concealment and keep loss to follow up low. In observational studies, draw comparison groups from the same source population, use weighting where entry or dropout depends on measured factors, and address the closely related problem of confounding with propensity scores or multivariable adjustment for measured confounders, instrumental variables where a valid instrument exists, and sensitivity analyses for unmeasured confounding. In systematic reviews, assess selection bias as part of the risk of bias evaluation for each included study.

An observational study compares surgical outcomes between two hospitals. Hospital A, a tertiary referral center, receives more complex cases than Hospital B, a community hospital. A naive comparison shows higher mortality at Hospital A. After adjusting for case complexity using propensity score matching on 22 baseline variables, Hospital A actually has lower risk adjusted mortality (OR = 0.76, 95% CI: 0.62 to 0.93), suggesting that the unadjusted comparison was distorted by the referral of sicker patients to the more capable hospital, a case mix difference that epidemiology classifies as confounding (some texts call it referral or selection bias); because matching balances only measured variables, unmeasured differences in severity could remain.