Choice of effect measure for epidemiological data.

Area of Science:

• Epidemiology
• Biostatistics
• Clinical Research

Background:

• Ongoing debate in scientific literature regarding the optimal effect measure for epidemiologic data.
• Disagreement persists between clinical and statistical criteria for selecting effect measures.
• Need for clear understanding of different measures' characteristics and applications.

Purpose of the Study:

• To present and compare defining characteristics of key effect measures: risk difference (RD), relative risk (RR), and odds ratio (OR) for binary data.
• To consider both clinical and statistical perspectives in evaluating these measures.
• To discuss the relationships between RD, RR, odds ratio (OR), relative risk reduction (RRR), and number needed to treat (NNT).

Main Methods:

• Comparative analysis of the theoretical properties of RD, RR, and OR.
• Numerical comparison using models with constant RD, RR, and OR to assess practical differences.
• Evaluation of potential limitations, such as predicting impossible event rates (<0% or >100%) with RD and RR models.

Main Results:

• Typically, RD, RR, and OR show small numerical differences in practice, except in cases of extreme extrapolation.
• RD and RR models have the potential to predict event rates outside the valid range of 0% to 100%.
• Each measure possesses theoretical justification, with RD and RR potentially better for risk communication and OR for data analysis.

Conclusions:

• Different effect measures may be necessary for distinct objectives: data analysis versus risk communication.
• Maintaining a clear distinction between the goals of statistical analysis and subsequent communication of findings is crucial.
• The choice of effect measure should align with the specific aims of the epidemiological study and its interpretation.