Confidence intervals for policy evaluation in adaptive experiments
Abstract
Randomized controlled trials are central to the scientific process, but they can be costly. For example, a clinical trial may assign patients to treatments that are detrimental to them. Adaptive experimental designs, such as multiarmed bandit algorithms, reduce costs by increasing the probability of assigning promising treatments over the course of the experiment. However, because observations collected by these methods are dependent and their distribution is nonstationary, statistical inference can be challenging. We propose a treatment-effect estimator that has an asymptotically unbiased and normal test statistic under straightforward, relatively weak conditions on the adaptive design. This estimator generalizes for a variety of parameters of interest.
Document Details
- Document Type
- Pub Defense Publication
- Publication Date
- Apr 05, 2021
- Source ID
- 10.1073/pnas.2014602118
Entities
People
- David A. Hirshberg
- Ruohan Zhan
- Stefan Wager
- Susan Athey
- Vitor Hadad
Organizations
- Alfred P. Sloan Foundation
- National Science Foundation
- Office of Naval Research
- Stanford University