An Automatic Finite-Sample Robustness Check: Can Dropping a Little Data Change Conclusions?

Watch a distinguished seminar from MIT's Tamara Broderick exploring the critical question of data robustness in statistical analysis. Learn about an innovative method to assess how removing small fractions of data can potentially change research conclusions. The talk examines how practitioners often analyze samples with the intention of applying findings to broader populations, but may not consider how sensitive their conclusions are to small subsets of influential data points. Broderick presents an approximation technique based on the classical influence function that automatically evaluates this sensitivity for common estimators, complete with error bounds and a low-cost exact lower bound. Discover why sensitivity is determined by signal-to-noise ratios rather than sample size or misspecification, and how several influential economics papers' conclusions can be altered by removing less than 1% of their data. This presentation offers valuable insights for researchers and data scientists concerned with the robustness and generalizability of their analytical findings.