Tyler VanderWeele, of the Departments of Epidemiology and Biostatistics at Harvard T.H. Chan School of Public Health, will present:
“Simplifying and Unifying Sensitivity Analysis”
Abstract: Potential biases often threaten conclusions in observational research in both the biomedical and social sciences. Because of this, it is important to examine the robustness of results to potential violations in the assumptions made. Sensitivity analysis provides a set of tools to examine the relative sensitivity or robustness to biases such as unmeasured confounding, measurement error, and selection bias. Unfortunately, these tools are not often used in practice. One of the impediments to the routine use of these techniques has been the complexity of their implementation, interpretation, and reporting. A relatively new metric, the E-value, provides a straightforward way to evaluate, report, and interpret the sensitivity of conclusions to potential unmeasured confounding. In many contexts, the E-value is a relatively conservative metric and its uses and limitations must be properly understood, and often a more extensive sensitivity analysis will be desired. The E-value approach, however, leaves researchers without an excuse for not performing at least a simple sensitivity analysis. Similar straightforward approaches are now also available for measurement error and selection bias, along with others covering all three of these biases simultaneously. More routine use of these techniques would help with the evaluation of evidence for causality throughout the biomedical and social sciences.
All are Welcome.