This package provides tools to conduct sensitivity analysis for the identification assumptions of linear regression and instrumental variable models. Users provide the data and specify direct or comparative bounds on the influence of an unmeasured confounder on observed variables. The package estimates the resulting partially identified region (PIR) and provides uncertainty quantification. Moreover, the results can be visualized via 2 types of contour plots (R- and b-contours).
This is accomplished by casting sensitivity analysis as a constraint stochastic optimization problem; limosa.beta uses a tailored grid-search algorithm to solve the resulting optimization problem.
limosa.beta is based on the article Optimization-based Sensitivity Analysis for Unmeasured Confounding using Partial Correlations, available on arXiv. The code to create the plots and simulation studies in the paper can be found in this Github repository.
You can install the development version of limosa.beta from GitHub with:
# install.packages("devtools")
devtools::install_github("tobias-freidling/limosa.beta")