General Medicine · Medical Literature
Logistic Regression.
Clinical relevance
Logistic regression is one of the most widely used statistical methods in clinical research, underpinning risk prediction models, diagnostic studies, and multivariable analyses across virtually every medical specialty. Publication in NEJM Evidence signals a methodological educational focus aimed at practicing clinicians who consume, appraise, and apply research findings in patient care. Understanding its assumptions, outputs, and limitations is essential for critically interpreting the growing body of observational and trial-based literature that informs modern clinical decision-making.
Key findings
1. NOTE — No abstract was available; all observations below are grounded in established statistical and clinical-research literature rather than article-specific data, and should be verified against the full text. 2. Logistic regression models a binary outcome (e.g., disease present/absent, mortality yes/no) by estimating the log-odds of the outcome as a linear combination of predictor variables, yielding odds ratios (ORs) with 95% confidence intervals for each covariate. 3. The method allows simultaneous adjustment for multiple confounders, making it the standard approach for multivariable analysis in retrospective cohorts, case-control studies, and secondary analyses of randomized trials. 4. Key assumptions include independence of observations, absence of perfect multicollinearity among predictors, and an adequate events-per-variable (EPV) ratio — commonly cited as ≥10 events per predictor — to avoid overfitting and biased coefficient estimates. 5. Model discrimination is typically reported as the c-statistic (equivalent to the area under the ROC curve; AUC), where 0.5 indicates no discrimination and 1.0 indicates perfect discrimination; calibration is assessed via tools such as the Hosmer–Lemeshow test or calibration plots. 6. Penalized variants (e.g., LASSO, ridge regression) and extensions (e.g., multinomial or ordinal logistic regression for non-binary outcomes) address limitations of standard logistic regression in high-dimensional or small-sample settings.
Practice implications
Clinicians appraising studies that use logistic regression should scrutinize the EPV ratio, the completeness of confounder adjustment, and both discrimination and calibration metrics before applying reported ORs to patient care — an OR from a poorly calibrated model can be statistically significant yet clinically misleading. When interpreting ORs, practitioners must recognize that ORs overestimate relative risk when the outcome is common (prevalence >10%), and risk ratios or absolute risk differences may be more clinically meaningful for patient counseling and shared decision-making. Institutions developing or adopting clinical risk-prediction tools built on logistic regression should require external validation in populations similar to their own patient demographics before embedding such models into clinical workflows or electronic health record decision-support systems.