A Flexible Defense Against the Winner's Curse

  • 2024-11-27 18:14:08
  • Tijana Zrnic, William Fithian
  • 0

Abstract

Across science and policy, decision-makers often need to draw conclusionsabout the best candidate among competing alternatives. For instance,researchers may seek to infer the effectiveness of the most successfultreatment or determine which demographic group benefits most from a specifictreatment. Similarly, in machine learning, practitioners are often interestedin the population performance of the model that performs best empirically.However, cherry-picking the best candidate leads to the winner's curse: theobserved performance for the winner is biased upwards, rendering conclusionsbased on standard measures of uncertainty invalid. We introduce the zoomcorrection, a novel approach for valid inference on the winner. Our method isflexible: it can be employed in both parametric and nonparametric settings, canhandle arbitrary dependencies between candidates, and automatically adapts tothe level of selection bias. The method easily extends to important relatedproblems, such as inference on the top k winners, inference on the value andidentity of the population winner, and inference on "near-winners."

 

Quick Read (beta)

loading the full paper ...