Selection Bias

Selection bias arises when entry into the analysed sample depends jointly on the exposure and the outcome, or on something causing both. Sampling and volunteer bias act at recruitment, since people who enrol differ from those who do not in health and motivation. Referral bias follows when a specialist centre sees a severe subset. Healthy user bias arises when people who accept a treatment already do other healthy things. Attrition bias arises after enrolment, when dropout differs by arm or outcome.

The most instructive case remains hormone therapy. Large observational cohorts reported substantially less coronary heart disease among postmenopausal women taking oestrogen, an association robust across studies. The Women's Health Initiative randomised trial did not reproduce a coronary benefit, and the leading explanation is that women who took hormone therapy were healthier and more engaged with medical care in ways adjustment did not capture.

Randomisation removes selection bias at the point of allocation and nowhere else. It re-enters through differential dropout, which is why intention-to-treat analysis keeps everyone in the arm assigned and why a per-protocol population is partly self-selected. It is also distinct from restricted generalisability: narrow eligibility criteria give an unbiased estimate for the enrolled population that may describe nobody outside it, an external validity problem rather than an internal one.

The version met constantly outside trials is the survivorship denominator. Clinic outcome data drawn from patients who came back excludes those who stopped because it did not work or made them unwell, and online surveys of peptide users sample those still using and willing to answer. A satisfaction rate of 90 percent computed over responders is a statement about responders, and the count of those who left is the figure not shown.