Regression to the Mean

Any measurement is a stable component plus a fluctuating one: biological variation, measurement error, the state someone was in that morning. Select people because their measured value was extreme and you have selected disproportionately for those whose fluctuating component was extreme too. On remeasurement the stable part persists and the fluctuation is redrawn, so the group average moves back towards the population mean with no intervention at all. The effect grows as the measurement becomes less reliable.

Galton described it in the heights of parents and children in the 1880s, and clinical medicine meets it daily: enrol on a single elevated blood pressure reading and the second reading is lower on average without a tablet being taken. Symptom-driven conditions amplify it, because people present at their worst: tendinopathy flares, low back pain and migraine frequency all fluctuate widely, and the symptomatic peak is when someone books an appointment.

This is much of why a concurrent control group exists. A single-arm before-and-after study cannot separate treatment effect from regression, natural history and placebo response, because all four push the second measurement the same way. Randomisation solves it not by preventing regression but by letting both arms regress equally, so the between-arm difference stays interpretable even when both improve from baseline.

The subtle version analyses response by baseline severity and concludes the worst-affected benefit most. Change scores are mathematically coupled to baseline values through shared measurement error, so the correlation appears where the treatment did nothing. The blunter version is the clinic case series: patients arrive at their worst, receive an injection, improve, and the improvement is credited to the compound rather than to the reason they walked in.