In a simulation study evaluating strategies for estimating differences between proportions after multiple imputation, researchers found that Agresti-Caffo confidence intervals performed similarly to Newcombe-Wilson intervals, whereas Wald intervals suffered from undercoverage in smaller sample sizes.
Published in BMC Medical Research Methodology by MW Heymans and JL Brand, the investigation addressed critical statistical challenges in observational studies and randomized trials. Differences between proportions and their corresponding confidence intervals are widely used to assess group differences in these designs, but missing data can introduce severe bias if not handled carefully.
To evaluate how interval estimation behaves alongside multiple imputation, the authors compared three specific construction procedures applied across multiply imputed datasets: the Wald method, the Agresti-Caffo method, and the Newcombe-Wilson method. Prior to this investigation, the behavior of these techniques when combined with multiple imputation had remained unclear.
The researchers conducted an extensive simulation study to measure how each strategy performed across varied conditions. Their experimental setup tested missing completely at random (MCAR) and missing at random (MAR) mechanisms at missingness levels of 10% and 30%, evaluating both small and large sample sizes across a wide range of proportion values.
Across the simulation scenarios, the team compared coverage probability at nominal coverage levels of 0.95 and 0.90, as well as interval lengths, statistical power, and type 1 error rates. In all large sample size scenarios, the three methods under study performed equally well, closely matching the target 0.95 and 0.90 coverage probabilities while producing similar interval lengths, type 1 error rates, and power curves.
Performance differed notably in smaller sample sizes. While the Newcombe-Wilson and Agresti-Caffo methods performed similarly overall, the Newcombe-Wilson approach became more conservative at low and high proportion values. Under the same conditions, the Wald method proved too liberal at extreme proportions and produced confidence intervals that were too narrow, leading the authors to conclude that Agresti-Caffo intervals represent a strong alternative to the Newcombe-Wilson method.
Subject of Research: Confidence interval methods for difference between proportions after multiple imputation
Article Title: Confidence intervals for the difference between proportions after multiple imputation: a simulation study comparing different strategies
Article References: Heymans, M., & Brand, J. (2026). Confidence intervals for the difference between proportions after multiple imputation: a simulation study comparing different strategies. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03072-z
Image Credits: AI Generated
DOI: 10.1186/s12874-026-03072-z
Keywords: multiple imputation, confidence intervals, difference between proportions, simulation study, Agresti-Caffo method, Newcombe-Wilson method, Wald interval, missing data, missing completely at random, missing at random, coverage probability, type 1 error, statistical power, interval length, observational studies, randomized trials
News Source: Ophelia Keating. (November 12, 2026). Agresti-Caffo Intervals Match Newcombe-Wilson Performance After Multiple Imputation in Simulations. Scienmag.



