Balancing Covariates Online
Abstract
The primary goal of this project is to design an algorithm for assigning treatments in online randomized trials that balances covariates as well as the best algorithm for the offline case. In particular, it will serve as a replacement for the offline Gram-SchmidtWalk Design of Harshaw, Savje, Spielman, and Zhang. The key to the covariate balancing algorithm will be the design of an online 1-subgaussian sampler. This will be followed by the development of central limit theorems and confidence intervals for the estimates of average treatment effects computed using this design.The secondary goal of the project is to incorporate the online covariate balancing algorithm into the design of adaptive trials. These are trials in which the outcomes for some subjects are observed before others are assigned treatment. This objective will entail combining the online covariate balancing algorithm with recently developed algorithms for selecting treatment probabilities in adaptive trials that do not incorporate covariate information.The third goal of the project is to develop tools for the retrospective analysis of trials that were conducted using different methodologies. These tools will enable an analyst to estimate the balance and robustness of a study after it has been conducted, if its methodology was sufficiently well-specified, and thereby estimate confidence intervals for average treatment effects.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412611
Entities
People
- Dan Spielman
Organizations
- Office of Naval Research
- United States Navy
- Yale University