Statistics Every Clinician Must Understand to Read Research Confidently

Clinical practice is increasingly driven by numerical evidence: risk ratios, number needed to treat, hazard ratios, area under the ROC curve, minimum clinically important differences. These metrics appear in papers, guidelines, and industry presentations. Understanding them at a working level — sufficient to evaluate their validity and relevance — is not optional for a clinician who takes evidence-based medicine seriously. It is a clinical competency with direct patient safety implications.

Relative versus absolute risk: the most misused distinction in medicine

A treatment that reduces the relative risk of a complication by 50% sounds impressive and important. If the baseline event rate is 2%, the absolute risk reduction is 1% — meaning 100 patients must be treated to prevent one event. The number needed to treat of 100 contextualises the clinical and resource implications of the intervention in a way that relative risk reduction alone obscures. Industry presentations and promotional materials almost invariably use relative risk measures. Clinical guidelines — and clinicians making evidence-based decisions — should demand both.

The p-value: what it tells you and what it does not

A p-value below 0.05 tells you that the result is unlikely to have occurred by chance if the null hypothesis is true. It does not tell you that the effect is clinically meaningful, that the study was well-designed, or that the result will replicate. A well-powered study of a clinically irrelevant effect will produce a highly significant p-value. Effect sizes and confidence intervals provide the information that p-values alone cannot — and should be the primary statistical focus when reading any clinical trial report.

The minimum clinically important difference: the threshold that matters

Every patient-reported outcome measure has an established minimum clinically important difference — the smallest change that patients actually notice and consider meaningful. Studies should be powered to detect effects at or above the MCID. A statistically significant improvement in Oxford Shoulder Score of 2 points that falls below the established MCID of approximately 11 points is a real result that reflects a real change in the measured score — but it is not a clinically meaningful improvement for patients. That distinction is critical for interpreting whether a treatment actually benefits the people it is given to.

Calibration and discrimination in prediction models — the difference between a model that correctly ranks patients by risk and one that accurately predicts their absolute probability of an event — matter for different clinical purposes and are frequently conflated. A model with good discrimination but poor calibration may reliably identify high-risk patients without accurately predicting what proportion of them will experience the event. Understanding which property matters for your clinical application is essential for using prediction tools appropriately.