Selection bias

It enters through many doors. Volunteer samples attract atypical participants, differential dropout removes certain kinds of people from one arm, survivorship bias studies only those who made it to the endpoint, and referral patterns fill hospital-based samples with unusual cases. Each mechanism ties who is in the data to the very thing being measured.

The consequences resist statistical rescue. Unlike random noise, selection bias shifts estimates in a direction set by the selection process, and adjustment helps only when the selection mechanism is understood and measured. A large sample does not fix it, since bias and precision are different problems.

When reviewing, trace how participants entered and left the study. High or differential attrition, unclear sampling frames, and results generalized to populations that were never sampled are all warning signs. A clear participant flow diagram, as CONSORT and STROBE require, makes selection effects visible.