Bias and Confounding

#Systematic problems differ from chance

Bias is a systematic distortion arising from how a study is designed, conducted, analysed or reported. It differs from random variation: increasing the number of participants may improve precision without fixing the underlying problem. Bias can make an effect appear larger, smaller or even point in the wrong direction.

Selection bias can occur when the people included or retained create a misleading comparison. Measurement bias can arise when information is collected differently or inaccurately across groups. For example, more intensive monitoring in one group can make an outcome appear more common even when underlying occurrence is similar.

#Understanding confounding

Confounding occurs when another factor influences the comparison between an exposure and an outcome. In a hypothetical treatment study, people with more severe illness might be more likely to receive a particular treatment. Their illness severity could also increase their risk of a poor outcome.

A simple comparison might then wrongly attribute that higher risk to the treatment. Confounding can also make an ineffective intervention appear beneficial. The challenge is to separate the effect of the exposure from differences that already existed or developed through processes relevant to the study question.

#How studies reduce distortion

Randomisation helps balance confounding factors on average. Other safeguards include appropriate comparison groups, consistent measurements, blinded assessment and efforts to reduce missing data. Observational analyses may use matching, stratification or statistical adjustment, but these approaches depend on relevant factors being identified and measured adequately.

#Look for remaining weaknesses

Adjustment is not a universal repair. Unmeasured factors can remain, and adjusting for an inappropriate variable can introduce new problems. Sensitivity analyses examine how conclusions respond to different assumptions, but cannot prove that every source of distortion has been removed.

Selective reporting also matters: published findings may differ from planned analyses, or negative studies may remain unavailable. Assessing bias therefore means examining the whole research process, not searching for a single reassuring label or assuming that a large sample guarantees a trustworthy answer.

#Common misunderstandings

A large study is not automatically a reliable study. More participants can reduce uncertainty from chance, but they do not necessarily correct systematic problems in how participants were selected, exposures were measured, or outcomes were recorded.

Confounding is also not the same as dishonesty or poor intentions. It can arise because groups differ in ways that affect the outcome, even when researchers work carefully. A statistically significant result does not rule out confounding or bias.

Random assignment and random sampling serve different purposes. Random assignment helps make treatment groups comparable; random sampling helps a sample represent a wider population. Neither guarantees that every aspect of a study is sound.

Finally, “adjusted” does not mean “fully corrected.” Statistical adjustment depends on which factors were measured, how accurately they were measured, and whether the analysis handled them appropriately. Adjusting for the wrong factors can sometimes introduce distortion rather than remove it.

#Questions worth asking a clinician

• Could the way participants were selected or followed up systematically distort this study’s findings?
• Were outcomes measured in the same way across groups, and could knowing the treatment influence those measurements?
• Which factors, such as age or illness severity, could explain group differences rather than the treatment itself?
• What design safeguards and statistical adjustments addressed bias and confounding, and what limitations remain?
• Even with a large sample, could systematic bias still make this study’s results misleading?