서브메뉴
검색
Precise Randomized Experiments Through Design and Estimation
Precise Randomized Experiments Through Design and Estimation
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104743
- ISBN
- 9798290651828
- DDC
- 362.72
- 저자명
- Li, Harrison H.
- 서명/저자
- Precise Randomized Experiments Through Design and Estimation
- 발행사항
- [Sl] : Stanford University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 224 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Owen, Art.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2025.
- 초록/해제
- 요약A properly conducted randomized experiment is invaluable for causal inference, since treatment randomization ensures by design that there are no confounders that can explain systematic differences in observed outcomes between treatment groups. In practice, however, the scale of randomized experiments is often limited by economic or ethical considerations, meaning that the resulting causal estimates are often noisy. We propose and study methods for design and point estimation to mitigate this in several experimental settings of interest. Here, design refers to choosing a propensity score--- the probability a subject receives a binary treatment as a function of observed covariates --- to maximize the precision (i.e., minimize the variance) of a downstream estimator.First, we consider a "tie-breaker design" where an investigator aims to respect external, non-statistical preferences for who should receive treatment. We formulate these preferences as a constraint on the correlation between the treatment indicator and an observed scalar "running variable." For any such constraint, we can construct a propensity score that is piecewise constant in the running variable with few discontinuities that provably maximizes any continuous precision objective based on the covariance matrix of an ordinary least squares estimator.Next, we turn our attention to a "batched experiment" where experimental subjects enter in multiple waves, as in many social experiments. We show how plugging a pooled propensity score that averages across all batches into certain semiparametric estimators asymptotically dominates popular alternatives that aggregate estimators computed separately from each batch. Then, we use observations from earlier batches to choose a propensity score for the next batch that minimizes an estimate of the asymptotic variance of our pooled estimator, using careful cross-fitting to carefully ensure the adaptive procedure does not affect the targeted asymptotic variance of the final estimator.Finally, we consider settings where an experimenter has an auxiliary observational dataset that may be biased for estimating treatment effects due to unobserved confounders or selection. We provide a novel framework based on semiparametric estimation theory that unifies a growing literature leveraging various assumptions about this bias enabling the observational data to be used to increase estimator precision. The framework shows how to systematically derive an estimator with asymptotic variance equal to the semiparametric efficiency bound subject to these assumptions, thus maximally leveraging them for precision gains and improving upon some existing ideas.
- 일반주제명
- Head Start project
- 일반주제명
- Statistics
- 키워드
- Propensity score
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358724
■00520260202104743
■006m o d
■007cr#unu||||||||
■020 ▼a9798290651828
■035 ▼a(MiAaPQ)AAI32149721
■035 ▼a(MiAaPQ)Stanfordsn782gc9812
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a362.72
■1001 ▼aLi, Harrison H.
■24510▼aPrecise Randomized Experiments Through Design and Estimation
■260 ▼a[Sl]▼bStanford University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a224 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Owen, Art.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2025.
■520 ▼aA properly conducted randomized experiment is invaluable for causal inference, since treatment randomization ensures by design that there are no confounders that can explain systematic differences in observed outcomes between treatment groups. In practice, however, the scale of randomized experiments is often limited by economic or ethical considerations, meaning that the resulting causal estimates are often noisy. We propose and study methods for design and point estimation to mitigate this in several experimental settings of interest. Here, design refers to choosing a propensity score--- the probability a subject receives a binary treatment as a function of observed covariates --- to maximize the precision (i.e., minimize the variance) of a downstream estimator.First, we consider a "tie-breaker design" where an investigator aims to respect external, non-statistical preferences for who should receive treatment. We formulate these preferences as a constraint on the correlation between the treatment indicator and an observed scalar "running variable." For any such constraint, we can construct a propensity score that is piecewise constant in the running variable with few discontinuities that provably maximizes any continuous precision objective based on the covariance matrix of an ordinary least squares estimator.Next, we turn our attention to a "batched experiment" where experimental subjects enter in multiple waves, as in many social experiments. We show how plugging a pooled propensity score that averages across all batches into certain semiparametric estimators asymptotically dominates popular alternatives that aggregate estimators computed separately from each batch. Then, we use observations from earlier batches to choose a propensity score for the next batch that minimizes an estimate of the asymptotic variance of our pooled estimator, using careful cross-fitting to carefully ensure the adaptive procedure does not affect the targeted asymptotic variance of the final estimator.Finally, we consider settings where an experimenter has an auxiliary observational dataset that may be biased for estimating treatment effects due to unobserved confounders or selection. We provide a novel framework based on semiparametric estimation theory that unifies a growing literature leveraging various assumptions about this bias enabling the observational data to be used to increase estimator precision. The framework shows how to systematically derive an estimator with asymptotic variance equal to the semiparametric efficiency bound subject to these assumptions, thus maximally leveraging them for precision gains and improving upon some existing ideas.
■590 ▼aSchool code: 0212.
■650 4▼aHead Start project
■650 4▼aStatistics
■653 ▼aPropensity score
■653 ▼aSemiparametric estimation theory
■690 ▼a0463
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358724▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
