Abstract
OBJECTIVES: This study aimed to compare the Delta, Greenland, and Monte Carlo methods for estimating 95% confidence intervals (CIs) of the population-attributable fraction (PAF). The objectives were to identify the optimal method and to determine the influence of primary parameters on PAF calculations. METHODS: A dataset was simulated using hypothetical values for primary parameters (population, relative risk [RR], prevalence, and variance of the beta estimator ) involved in PAF calculations. Three methods (Delta, Greenland, and Monte Carlo) were used to estimate the 95% CIs of the PAFs. Perturbation analysis was performed to assess the sensitivity of the PAF to changes in these parameters. An R Shiny application, the "GDM-PAF CI Explorer," was developed to facilitate the analysis and visualization of these computations. RESULTS: No significant differences were observed among the 3 methods when both the RR and p-value were low. The Delta method performed well under conditions of low prevalence or minimal RR, while Greenland's method was effective in scenarios with high prevalence. Meanwhile, the Monte Carlo method calculated 95% CIs of PAFs that were stable overall, though it required intensive computational resources. In a novel approach that utilized perturbation for sensitivity analysis, was identified as the most influential parameter in the estimation of CIs. CONCLUSIONS: This study emphasizes the necessity of a careful approach for comparing 95% CI estimation methods for PAFs and selecting the method that best suits the context. It provides practical guidelines to researchers to increase the reliability and accuracy of epidemiological studies.