E-values for effect heterogeneity and approximations for causal interaction.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Estimates of effect heterogeneity can be biased by uncontrolled confounding that varies across strata of a second exposure.
• Confounding can distort the true causal effect of an exposure when its strength differs across subgroups.

Purpose of the Study:

• To propose novel methods, analogous to the E-value, for assessing the sensitivity of effect heterogeneity estimates to uncontrolled confounding.
• To quantify the strength of confounding required to alter conclusions about effect heterogeneity.

Main Methods:

• Developed E-value analogues to measure the confounding needed to shift effect heterogeneity estimates to the null or include the null in the confidence interval.
• Utilized the R package EValue for calculating these sensitivity analyses.
• Applied methods to real-world examples concerning educational attainment, dementia, obesity, and mortality.

Main Results:

• Demonstrated the application of E-value analogues to assess effect heterogeneity in studies of educational attainment and dementia incidence by sex.
• Illustrated the use of E-value analogues for evaluating effect heterogeneity in studies of obesity and all-cause mortality by age.

Conclusions:

• The proposed E-value analogues provide a quantitative measure to assess the robustness of effect heterogeneity estimates.
• Reporting these values can enhance the transparency and reliability of research findings concerning varying causal effects across populations.