Relative Measures of Association for Binary Outcomes: Challenges and Recommendations for the Global Health Researcher

Abstract

Background: Binary outcomes—which have two distinct levels (e.g., disease yes/no)—are commonly collected in global health research. The relative association of an exposure (e.g., a treatment) and such an outcome can be quantified using a ratio measure such as a risk ratio or an odds ratio. Although the odds ratio is more frequently reported than the risk ratio, many researchers, policymakers, and the general public frequently interpret it as a risk ratio. This is particularly problematic when the outcome is common because the magnitude of association is larger on the odds ratio scale than the risk ratio scale. Some recently published global health studies included misinterpretation of the odds ratio, which we hypothesize is because statistical methods for risk ratio estimation are not well known in the global health research community. Objectives: To compare and contrast available statistical methods to estimate relative measures of association for binary outcomes and to provide recommendations regarding their use. Methods: Logistic regression for odds ratios and four approaches for risk ratios: two direct regression approaches (modified log-Poisson and log-binomial) and two indirect methods (standardization and substitution) based on logistic regression. Findings: Illustrative examples demonstrate that misinterpretation of the odds ratio remains a common issue in global health research. Among the four methods presented for estimation of risk ratios, the modified log-Poisson approach is generally preferred because it has the best numerical performance and it is as easy to implement as is logistic regression for odds ratio estimation. Conclusions: We conclude that, when study design allows, studies with binary outcomes should preferably report risk ratios to measure relative association.