Bias and Confounding

#Bias is different from chance

Bias is systematic distortion in a research result. Random error is variation caused by chance. More observations can make an estimate more precise without removing a systematic problem in how the comparison was made.

Cochrane evaluates bias for a particular result, not as a single label for every finding in a study. It also distinguishes bias from limited applicability: a result may be sound for the people studied yet not answer the same question for a different population.

Evidence: STROBE: observational reporting explanation and elaboration / Cochrane Handbook, Chapter 25: bias in non-randomized intervention studies

#Confounding changes the comparison

In non-randomised intervention research, confounding occurs when common causes influence both treatment choice and the outcome. For example, pre-existing illness severity may affect which treatment a person receives and their later health. The observed association can then differ from the treatment effect.

Randomisation prevents prognostic factors from determining assignment and helps avoid this problem at the start of a trial. In observational analyses, adjustment depends on relevant factors being identified, measured and modelled adequately. Residual confounding can remain; adjusting for a variable affected by treatment can itself introduce bias.

Evidence: Cochrane Handbook, Chapter 25: bias in non-randomized intervention studies / Cochrane Handbook, Chapter 8: risk of bias in a randomized trial

#Selection and measurement also matter

Selection can distort a result when inclusion, exclusion or missing follow-up creates a misleading comparison. Read who entered the analysis, who was left out and whether the missing information could relate to both the exposure and outcome.

Measurement problems can arise when exposure or outcome information is misclassified. Different intensity of observation between groups can change how often an outcome is detected. Blinded assessment may help, but read the actual measurement method rather than assuming that a design label removes the problem.

Evidence: Cochrane Handbook, Chapter 25: bias in non-randomized intervention studies / STROBE: observational reporting explanation and elaboration

#Assess the process, not a reassuring word

Planned outcomes, available measurements and alternative analyses can produce different estimates. Selecting a reported result because of its size, direction or P value can bias what readers see. Cochrane distinguishes this from the separate problem of whole studies or results not being available.

Ask how the study reduced confounding, selection problems, measurement error, missing data and selective reporting. A large sample or the word "adjusted" cannot answer those questions alone. A formal assessment needs the study details and relevant methodological and subject expertise; this page is an introduction, not an expert rating.

Evidence: Cochrane Handbook, Chapter 25: bias in non-randomized intervention studies / CONSORT 2025 explanation and elaboration

Source note

The sections above were checked against the linked sources. No clinical review has been performed. This is general research education, not a clinical guideline.