Standardization and Prediction to Control Confounding: Estimating Risk Differences and Ratios for Clinical Interpretations and Decision Making.

Abstract

What is the added risk for death from smoking? Logistic regression has become the most common statistical method to answer such questions in the biomedical literature. However, the typical analyses estimate odds ratios, a metric too often misunderstood and misinterpreted. Although estimates of risks, and their differences and ratios, offer transparent clinical interpretations, commonly used statistical models have known methodological shortcomings. "Standardization" through modeling, weighting, or matching offers a solution. The goals of this article are to review classical concepts of standardization and to link them to regression modeling for causal inference. The authors also describe approaches based on weighting and matching compared with regression-based standardization. Using an example of smoking from the ARIC (Atherosclerosis Risk in Communities) study, they explain the value of standardization, long used in medicine and public health, to estimate risks and their differences and ratios for binary outcomes. The authors demonstrate how standard statistical software using models that best fit the data and respect underlying biological or clinical processes can reexpress results in clinically meaningful metrics. The Supplement offers examples with common software packages.