Confounding, Bias, and Study Validity
This area gathers the concepts epidemiologists use to judge whether an observed exposure-outcome association reflects a real effect or an artefact. It distinguishes systematic error — confounding, selection bias, and information bias — from random error, and frames the result in terms of internal validity (is the estimate correct for the study population?) and the related notion of effect modification (does the effect differ across subgroups?).
Definition
Confounding, bias, and study validity together name the framework by which epidemiology evaluates whether a measured association is a valid estimate of a causal effect, separating systematic error (confounding, selection bias, information bias) from random error and from genuine variation in effect across subgroups (effect modification).
Scope
The area is an orienting overview of the threats to validity in epidemiologic studies and the vocabulary used to reason about them. It links the detailed topic entries on confounding, selection bias, information bias, effect modification and interaction, and internal validity. It is a methodological reference and does not give clinical or individual treatment guidance.
Sub-topics
Core questions
- Is the observed association explained by a common cause of exposure and outcome (confounding)?
- Did the way subjects entered or stayed in the study distort the association (selection bias)?
- Were exposure or outcome measured or reported differently across groups (information bias)?
- Does the effect genuinely differ across subgroups (effect modification), and is that distinct from confounding?
- Taken together, is the estimate internally valid for the population actually studied?
Key concepts
- Systematic versus random error
- Confounding
- Selection bias
- Information (measurement) bias
- Effect modification and interaction
- Internal validity
- External validity (generalisability)
- Causal diagrams (DAGs)
Mechanisms
A measured association can depart from the true causal effect for several distinct reasons. Confounding arises when a third factor is a common cause of both exposure and outcome, mixing its effect with the one under study. Selection bias arises when the procedures that bring subjects into the analysis — and keep them there — depend jointly on exposure and outcome, distorting the association in the analysed sample. Information bias arises when exposure or outcome is misclassified, and the misclassification may be non-differential (blurring the estimate toward the null) or differential (shifting it in either direction). These systematic errors are conceptually separate from random error, which reflects sampling variability and is summarised by confidence intervals. Effect modification is not an error at all: it describes real variation in the effect across levels of a third variable. Causal diagrams (directed acyclic graphs) give a common language for distinguishing confounding from selection bias and for deciding what to adjust for.
Clinical relevance
These concepts are central to appraising the evidence that underlies health knowledge. Whether a reported association between an exposure and a disease should be believed depends on how well a study controlled confounding and bias and on whether its estimate is internally valid. The area describes how evidence is judged, not what any individual should do about a diagnosis or treatment.
Epidemiology
Reasoning about confounding and bias is part of every observational study and is built into reporting standards such as the STROBE statement, which asks authors to describe their handling of these threats. The framework is applied across cohort, case-control, and cross-sectional designs and increasingly through explicit causal-diagram methods.
Evidence & guidelines
The STROBE statement (von Elm et al., 2007) is a widely adopted reporting guideline that requires observational studies to address sources of bias, confounding control, and the limitations bearing on internal and external validity.
History
The vocabulary of bias and confounding crystallised over the twentieth century as observational epidemiology matured, drawing on debates about causal inference from non-experimental data. From the late twentieth century, formal causal models — potential outcomes and directed acyclic graphs — gave precise definitions that unified previously separate notions of confounding and selection bias and clarified their distinction from effect modification.
Debates
- Are confounding and selection bias one phenomenon or two?
- Causal-diagram accounts show confounding (a common cause of exposure and outcome) and selection bias (conditioning on a common effect, or collider) as structurally distinct, even though both produce non-causal associations; some classical treatments group them more loosely.
Key figures
- Sander Greenland
- James Robins
- Judea Pearl
- Kenneth Rothman
- Miguel Hernán
Related topics
Seminal works
- greenland-pearl-robins-1999
- grimes-schulz-2002-bias
- delgado-rodriguez-2004
Frequently asked questions
- What is the difference between bias and confounding?
- Both are systematic errors, but confounding is mixing of effects from a common cause of exposure and outcome, whereas bias here refers to distortions introduced by how subjects are selected (selection bias) or how variables are measured (information bias).
- Is effect modification a type of bias?
- No. Effect modification describes genuine variation in an effect across subgroups; it is a feature of the relationship being studied, not an error to be eliminated like confounding or bias.