Bias and confounding, in plain words

Bias is a fault in how you got your data. Confounding is a third variable explaining your result.

Selection bias — the people in your study are not like the people you want to draw conclusions about. Measurement bias — your instrument or observer is systematically wrong, often in one direction. Recall bias — people with the outcome remember exposures better than people without.

Confounding is different. It is a variable related to both your exposure and your outcome, which makes them look connected when they are not, or hides a connection that is real.

Why it matters

Every examiner asks about these, and 'we did not consider it' is a worse answer than 'we could not control for it, and here is how that limits the finding'. Confounding in particular can be dealt with in the analysis — by stratifying, or by regression — but only if you recorded the confounder.

An example

You find that students who attend more classes score higher. Before concluding that attendance raises marks, consider prior ability: a stronger student both attends more and scores higher. Prior ability is the confounder, and unless you measured it you cannot separate the two.

Common mistakes

• Listing 'bias' in the limitations without saying which bias, in which direction, and how large it might be.
• Not recording the obvious confounders — age, sex, severity — so they cannot be adjusted for later.

Read next

• Types of variable, and why the type decides the test — Numerical or categorical; and if categorical, ordered or not.
• Which study design you are actually running — Observational or interventional; and if observational, which direction time runs in.