서브메뉴
검색
Essays in Identification and Inference
Essays in Identification and Inference
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103526
- ISBN
- 9798315798255
- DDC
- 610
- 서명/저자
- Essays in Identification and Inference
- 발행사항
- [Sl] : Northwestern University, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 305 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Manski, Charles F.;Canay, Ivan A.
- 학위논문주기
- Thesis (Ph.D.)--Northwestern University, 2025.
- 초록/해제
- 요약This dissertation consists of three essays studying identification and inference in settings pertaining to economics and medicine. The emphasis is on developing and utilizing frameworks that incorporate relevant features of the underlying empirical context which may invalidate frequently imposed assumptions.The first chapter studies identification of long-term treatment effects. Since long-term experimentation is frequently infeasible, a large body of recent work proposes appending a short-term experimental dataset to longitudinal observational data to provide more credible alternatives to observational studies. As the first contribution, I show that justifiable modeling assumptions remain central for plausible inference despite the addition of the experiment, just as in conventional observational studies. Experimental data bring no identifying power on their own; they serve only to amplify the identifying power of assumptions restricting conditional means of long-term potential outcomes conditional on short-term potential outcomes---temporal link functions. However, existing research argues that previously proposed modeling assumptions may frequently be challenging to justify. This motivates the second contribution. I introduce the use of treatment response assumptions that only restrict the shape of temporal link functions and may thus be defensible based on economic intuition or theory. As the third contribution, I introduce a novel identification framework which yields the smallest possible, i.e., sharp, bounds on the long-term average treatment effect under a broad class of restrictions on temporal link functions and imperfect compliance in the experiment. The framework produces the bounds via solutions to generalized bilinear problems. It thus: 1) enables the use of the proposed treatment response assumptions; 2) facilitates the development of new justifiable assumptions by removing the need to prove sharpness; 3) extends existing methods to account for imperfect compliance.The second chapter studies the measurement of misclassification rates of diagnostic tests and general binary classifiers. The rates are of great interest to regulators and clinicians. However, their identification requires knowledge of the underlying ground truth, which is often measured by an imperfect reference test or classifier. The common practice is thus to report misclassification rates with respect to the reference---"apparent'' misclassification rates---which do not measure true performance. The first contribution are the sharp bounds on the measures of true performance---sensitivity (true positive rate) and specificity (true negative rate), or equivalently false positive and negative rates, under standard assumptions in performance studies. The second contribution is the construction of uniformly consistent confidence sets in level over a relevant family of data distributions. This allows researchers to account for statistical imprecision, which is typically recommended by relevant regulatory guidelines. As the third contribution, I revisit the performance of the ubiquitous BinaxNOW COVID-19 antigen test, based on Emergency Use Authorization and independent study data. The analysis reveals that the estimated false negative rates for symptomatic and asymptomatic patients are, respectively, up to 3.17 and 4.59 times higher than the frequently cited "apparent'' false negative rate. This finding brings into question whether the test would have met the contemporaneous threshold for Emergency Use Authorization once the imperfections of the reference test are taken into account.The third chapter (joint work with Gabriel Ziegler) shows that dilation is a real-world phenomenon that may be induced by diagnostic tests under established clinical practice. Clinicians often seek to determine the probability that a patient has a suspected illness conditional on a test result. Dilation entails that conditioning on any test result only introduces uncertainty about the patient's health status and has been mostly considered a theoretical curiosity. As the first contribution, we show that dilation may be induced by conditioning on diagnostic tests whose misclassification rates are evaluated with respect to an imperfect reference, which is widespread. Moreover, dilation may occur even when tests are approved or recommended based on satisfactory "apparent'' misclassification rates, often used as the primary criterion for evaluation. This motivates the second contribution. We enable decision-makers to identify such diagnostic tests by equivalently characterizing when dilation is induced, and providing a statistical testing procedure that is uniformly consistent in level for a large family of relevant data-generating processes. For the third contribution, we study computed tomography (CT) chest scans for detecting COVID-19 infection that were recommended as a primary detection tool in epidemic areas based on conventional "apparent'' performance measures. We find that they induced dilation and thus only introduced uncertainty about the patient's health status.
- 일반주제명
- Medicine
- 기타저자
- Northwestern University Economics
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357536
■00520260202103526
■006m o d
■007cr#unu||||||||
■020 ▼a9798315798255
■035 ▼a(MiAaPQ)AAI32039291
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a610
■1001 ▼aObradovic, Filip.▼0(orcid)0000-0003-4202-9867
■24510▼aEssays in Identification and Inference
■260 ▼a[Sl]▼bNorthwestern University▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a305 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Manski, Charles F.;Canay, Ivan A.
■5021 ▼aThesis (Ph.D.)--Northwestern University, 2025.
■520 ▼aThis dissertation consists of three essays studying identification and inference in settings pertaining to economics and medicine. The emphasis is on developing and utilizing frameworks that incorporate relevant features of the underlying empirical context which may invalidate frequently imposed assumptions.The first chapter studies identification of long-term treatment effects. Since long-term experimentation is frequently infeasible, a large body of recent work proposes appending a short-term experimental dataset to longitudinal observational data to provide more credible alternatives to observational studies. As the first contribution, I show that justifiable modeling assumptions remain central for plausible inference despite the addition of the experiment, just as in conventional observational studies. Experimental data bring no identifying power on their own; they serve only to amplify the identifying power of assumptions restricting conditional means of long-term potential outcomes conditional on short-term potential outcomes---temporal link functions. However, existing research argues that previously proposed modeling assumptions may frequently be challenging to justify. This motivates the second contribution. I introduce the use of treatment response assumptions that only restrict the shape of temporal link functions and may thus be defensible based on economic intuition or theory. As the third contribution, I introduce a novel identification framework which yields the smallest possible, i.e., sharp, bounds on the long-term average treatment effect under a broad class of restrictions on temporal link functions and imperfect compliance in the experiment. The framework produces the bounds via solutions to generalized bilinear problems. It thus: 1) enables the use of the proposed treatment response assumptions; 2) facilitates the development of new justifiable assumptions by removing the need to prove sharpness; 3) extends existing methods to account for imperfect compliance.The second chapter studies the measurement of misclassification rates of diagnostic tests and general binary classifiers. The rates are of great interest to regulators and clinicians. However, their identification requires knowledge of the underlying ground truth, which is often measured by an imperfect reference test or classifier. The common practice is thus to report misclassification rates with respect to the reference---"apparent'' misclassification rates---which do not measure true performance. The first contribution are the sharp bounds on the measures of true performance---sensitivity (true positive rate) and specificity (true negative rate), or equivalently false positive and negative rates, under standard assumptions in performance studies. The second contribution is the construction of uniformly consistent confidence sets in level over a relevant family of data distributions. This allows researchers to account for statistical imprecision, which is typically recommended by relevant regulatory guidelines. As the third contribution, I revisit the performance of the ubiquitous BinaxNOW COVID-19 antigen test, based on Emergency Use Authorization and independent study data. The analysis reveals that the estimated false negative rates for symptomatic and asymptomatic patients are, respectively, up to 3.17 and 4.59 times higher than the frequently cited "apparent'' false negative rate. This finding brings into question whether the test would have met the contemporaneous threshold for Emergency Use Authorization once the imperfections of the reference test are taken into account.The third chapter (joint work with Gabriel Ziegler) shows that dilation is a real-world phenomenon that may be induced by diagnostic tests under established clinical practice. Clinicians often seek to determine the probability that a patient has a suspected illness conditional on a test result. Dilation entails that conditioning on any test result only introduces uncertainty about the patient's health status and has been mostly considered a theoretical curiosity. As the first contribution, we show that dilation may be induced by conditioning on diagnostic tests whose misclassification rates are evaluated with respect to an imperfect reference, which is widespread. Moreover, dilation may occur even when tests are approved or recommended based on satisfactory "apparent'' misclassification rates, often used as the primary criterion for evaluation. This motivates the second contribution. We enable decision-makers to identify such diagnostic tests by equivalently characterizing when dilation is induced, and providing a statistical testing procedure that is uniformly consistent in level for a large family of relevant data-generating processes. For the third contribution, we study computed tomography (CT) chest scans for detecting COVID-19 infection that were recommended as a primary detection tool in epidemic areas based on conventional "apparent'' performance measures. We find that they induced dilation and thus only introduced uncertainty about the patient's health status.
■590 ▼aSchool code: 0163.
■650 4▼aMedicine
■653 ▼aLong-term experimentation
■653 ▼aObservational data
■653 ▼aEconomic intuition
■653 ▼aComputed tomography
■690 ▼a0501
■690 ▼a0511
■690 ▼a0564
■71020▼aNorthwestern University▼bEconomics.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a0163
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357536▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


