Double machine learning and design in batch adaptive experiments

We consider an experiment with at least two stages or batches and O(N)O\left(N) subjects per batch. First, we propose a semiparametric treatment effect estimator that efficiently pools information across the batches, and we show that it asymptotically dominates alternatives that aggregate single bat...

Full description

Saved in:
Bibliographic Details
Main Authors: Li Harrison H., Owen Art B.
Format: Article
Language:English
Published: De Gruyter 2024-11-01
Series:Journal of Causal Inference
Subjects:
Online Access:https://doi.org/10.1515/jci-2023-0068
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1846171065249693696
author Li Harrison H.
Owen Art B.
author_facet Li Harrison H.
Owen Art B.
author_sort Li Harrison H.
collection DOAJ
description We consider an experiment with at least two stages or batches and O(N)O\left(N) subjects per batch. First, we propose a semiparametric treatment effect estimator that efficiently pools information across the batches, and we show that it asymptotically dominates alternatives that aggregate single batch estimates. Then, we consider the design problem of learning propensity scores for assigning treatment in the later batches of the experiment to maximize the asymptotic precision of this estimator. For two common causal estimands, we estimate this precision using observations from previous batches, and then solve a finite-dimensional concave maximization problem to adaptively learn flexible propensity scores that converge to suitably defined optima in each batch at rate Op(N−1⁄4){O}_{p}\left({N}^{-1/4}). By extending the framework of double machine learning, we show this rate suffices for our pooled estimator to attain the targeted precision after each batch, as long as nuisance function estimates converge at rate op(N−1⁄4){o}_{p}\left({N}^{-1/4}). These relatively weak rate requirements enable the investigator to avoid the common practice of discretizing the covariate space for design and estimation in batch adaptive experiments while maintaining the advantages of pooling. Our numerical study shows that such discretization often leads to substantial asymptotic and finite sample precision losses outweighing any gains from design.
format Article
id doaj-art-b84429a08c2b45e781021313f8f90fa5
institution Kabale University
issn 2193-3685
language English
publishDate 2024-11-01
publisher De Gruyter
record_format Article
series Journal of Causal Inference
spelling doaj-art-b84429a08c2b45e781021313f8f90fa52024-11-11T08:36:32ZengDe GruyterJournal of Causal Inference2193-36852024-11-0112155810810.1515/jci-2023-0068Double machine learning and design in batch adaptive experimentsLi Harrison H.0Owen Art B.1 Department of Statistics, Stanford University, Stanford, California, United States Department of Statistics, Stanford University, Stanford, California, United StatesWe consider an experiment with at least two stages or batches and O(N)O\left(N) subjects per batch. First, we propose a semiparametric treatment effect estimator that efficiently pools information across the batches, and we show that it asymptotically dominates alternatives that aggregate single batch estimates. Then, we consider the design problem of learning propensity scores for assigning treatment in the later batches of the experiment to maximize the asymptotic precision of this estimator. For two common causal estimands, we estimate this precision using observations from previous batches, and then solve a finite-dimensional concave maximization problem to adaptively learn flexible propensity scores that converge to suitably defined optima in each batch at rate Op(N−1⁄4){O}_{p}\left({N}^{-1/4}). By extending the framework of double machine learning, we show this rate suffices for our pooled estimator to attain the targeted precision after each batch, as long as nuisance function estimates converge at rate op(N−1⁄4){o}_{p}\left({N}^{-1/4}). These relatively weak rate requirements enable the investigator to avoid the common practice of discretizing the covariate space for design and estimation in batch adaptive experiments while maintaining the advantages of pooling. Our numerical study shows that such discretization often leads to substantial asymptotic and finite sample precision losses outweighing any gains from design.https://doi.org/10.1515/jci-2023-0068adaptive designpropensity scorepooled estimationpartially linear modelconvex optimization62k05
spellingShingle Li Harrison H.
Owen Art B.
Double machine learning and design in batch adaptive experiments
Journal of Causal Inference
adaptive design
propensity score
pooled estimation
partially linear model
convex optimization
62k05
title Double machine learning and design in batch adaptive experiments
title_full Double machine learning and design in batch adaptive experiments
title_fullStr Double machine learning and design in batch adaptive experiments
title_full_unstemmed Double machine learning and design in batch adaptive experiments
title_short Double machine learning and design in batch adaptive experiments
title_sort double machine learning and design in batch adaptive experiments
topic adaptive design
propensity score
pooled estimation
partially linear model
convex optimization
62k05
url https://doi.org/10.1515/jci-2023-0068
work_keys_str_mv AT liharrisonh doublemachinelearninganddesigninbatchadaptiveexperiments
AT owenartb doublemachinelearninganddesigninbatchadaptiveexperiments