confoundvis for Visualizing and Reporting Sensitivity to Unmeasured Confounding in R

Abstract

Sensitivity analysis asks how strong unmeasured confounding would have to be before an observational conclusion changes. Established methods answer this question using different quantities, including partial $R^2$ values, correlation products, and risk ratios. In this article, we present confoundvis, an R package that organizes these results through a common object, plotting functions, and reports. It calculates partial $R^2$ and impact-threshold paths from fitted linear models, imports results from sensemakr, konfound, and EValue, and provides measured-covariate benchmarks for the OLS frameworks. We demonstrate the workflow with 5,735 patients from the right heart catheterization study. The adjusted linear probability coefficient is 0.0634, the point-null robustness value is 0.0608, and the conditional significance impact threshold is 0.0376. The largest adverse leave-one-term-out measured-covariate product is 0.0090. Because the benchmark and threshold use different conditioning sets, their ratio is descriptive. A separate modified Poisson model gives a risk ratio of 1.103 and an E-value of 1.439. The contribution is a software workflow that keeps each method's units and decision rule clear. The package is available from CRAN.