Bias, Confounding & Causal Reasoning
Identify systematic distortions and distinguish a confounder from a mediator or an effect modifier.
Quick revision ↓ Medical Dictionary
01Bias versus random error
Bias is a systematic departure from the quantity the study intends to estimate. It can arise in selection, measurement, missing outcomes or selective reporting. Random error is sampling variation and is often reduced by more independent observations. A large sample does not automatically reduce bias. Describe the specific mechanism and likely consequences rather than merely calling a study biased because its result is unexpected.
02Confounding in a clinical comparison
Confounding occurs when the exposure-outcome comparison mixes the exposure effect with differences in other causes of the outcome. In an observational comparison, sicker patients may preferentially receive a treatment and also have a worse prognosis. This confounding by indication can make a useful treatment look harmful. Measure relevant baseline factors and use design and analysis methods justified by a causal understanding.
03Design and adjustment
Randomisation, restriction and matching can help address confounding at design stage; stratification and regression are analysis approaches. Adjustment cannot guarantee removal of unmeasured or poorly measured confounding. Matching must be accounted for in the analysis and is not a universal cure. Choose adjustment variables for their causal role, not simply because they have a small univariable p value in the dataset.
04Mediators, colliders and modification
A mediator lies on a causal pathway from exposure to outcome; adjusting for it can remove part of the total effect you want to estimate. A collider is a common effect of two variables; conditioning on it can introduce an association. Effect modification means an effect differs across levels of another factor on a specified scale. It is a feature to understand, not necessarily an error to eliminate.
05Build a cautious causal argument
Temporal ordering, plausible mechanisms, consistent findings and appropriate comparisons can strengthen a causal interpretation, but no single observational association proves causation. Ask what would have happened to comparable people without the exposure. Consider alternative explanations and analyses that test their plausibility. Use a causal diagram before modelling when possible, and distinguish estimating a total causal effect from predicting which patients will have an outcome.
High-Yield Revision
- Confounding by indication is common in treatment comparisons.
- Adjusting for more variables is not always better.
- Effect modification depends on the chosen effect scale.
Common Pitfalls
- Do not adjust automatically for a mediator when estimating a total effect.
- A larger sample can produce a more precise biased estimate.
References & Further Reading
Use the full methods and relevant chapters when appraising a paper. These sources support further study; the examples and questions above are original teaching exercises.
- Oxford CEBM — Critical appraisal tools
- Cochrane Handbook — Methods for evidence synthesis
- CDC — Analysing and interpreting epidemiological data
- ASA — Statement on p-values and statistical significance
- Oxford CEBM — Likelihood ratios
- Cochrane — GRADE and summary of findings
- WMA — Declaration of Helsinki, 2024 revision
- EQUATOR — Reporting guideline library
- NIH OBSSR — Mixed-methods research resources
- Institute for Healthcare Improvement — Model for Improvement