Publications

Detailed Information

Difference in restricted mean survival time in observational studies: A review of estimation methods and a development of sensitivity analysis : 제한된 평균 생존시간 추정 방법 고찰 및 새로운 민감도 분석 방법 개발

DC Field Value Language
dc.contributor.advisor박지훈-
dc.contributor.author이승재-
dc.date.accessioned2023-06-29T02:27:11Z-
dc.date.available2023-06-29T02:27:11Z-
dc.date.issued2023-
dc.identifier.other000000174847-
dc.identifier.urihttps://hdl.handle.net/10371/194090-
dc.identifier.urihttps://dcollection.snu.ac.kr/common/orgView/000000174847ko_KR
dc.description학위논문(박사) -- 서울대학교대학원 : 융합과학기술대학원 응용바이오공학과, 2023. 2. 박지훈.-
dc.description.abstractThe difference in restricted mean survival time (RMST) has been increasingly used as an alternative measure to hazard ratio in survival analysis. Unlike relative effect measure such as hazard ratio, RMST difference provides information about an intuitively interpretable absolute risk and is known to be robust regardless of the proportional hazards assumption.

In experimental studies such as a randomized controlled trial, the RMST is calculated by integrating the area under the Kaplan-Meier curve up to a specific time point, and the difference in RMST between the two groups is used as a causal effect of exposure. However, in observational studies, the standard Kaplan-Meier estimator cannot be directly used for calculating the RMST because of confounding bias due to non-random exposure assignment. The difference in RMST adjusted for potential confounders can be estimated using methods such as direct RMST regression, inverse probability weighting, G-computation, etc. Through multiple simulation studies in which all the models were correctly specified, we confirmed that all the methods being considered provided the unbiased estimates with the percentile bootstrap confidence intervals achieving near nominal coverage probability.

Although several methods have been developed for evaluating the difference in RMST adjusted for potential confounders in the observational study, there is no study on the sensitivity analysis of unmeasured confounding. Therefore, we propose a novel sensitivity analysis method that considers unmeasured confounding for evaluating the estimate of the difference in adjusted RMST. Given a user-specified sensitivity parameter, one can obtain the sensitivity range and confidence interval of bias-adjusted difference in RMST. It is necessary to solve a complex optimization problem to obtain the sensitivity range and confidence interval, but there is no analytic solution except in special cases. While the optimization problem can be directly solved by using an optimization algorithm such as L-BFGS-B (hereafter referred to this method as the direct optimization method), it takes considerable computational time. Therefore, we propose an approximate optimization method comparable to the direct optimization method in terms of bias, achieving substantial reduction in the computational time. Through intensive Monte Carlo simulation studies, we showed that the proposed approximate optimization method can be a practical alternative. When applying our sensitivity analysis method in practice, we recommend using the approximate optimization method in case that the censoring rate is less than 0.7. Otherwise, one may use the direct optimization method using an optimization algorithm.
-
dc.description.abstract생존 분석에서 제한된 평균 생존 시간(restricted mean survival time; RMST)의 차이는 위험 비율(hazard ratio)에 대한 대안 척도로 점점 더 많이 사용되고 있다. 위험 비율과 같은 상대적 효과 측도(relative effect measure)와 달리, RMST 차이는 직관적으로 해석 가능한 절대 위험(absolute risk)에 대한 정보를 제공하며 비례 위험 가정에 관계없이 로버스트한 것으로 알려져 있다.

무작위대조시험에서는 Kaplan-Meier 곡선 아래의 면적을 특정 시점까지 적분하여 RMST를 계산하고, 두 그룹 간의 RMST 차이를 노출(exposure)의 인과효과로 사용한다. 이에 반해, 관찰 연구에서는 비무작위 노출 할당으로 인한 교란 편향 때문에 표준적인 Kaplan-Meier 추정량을 RMST 계산에 직접 사용할 수 없다. 이러한 교란 편향을 보정한 RMST의 차이를 계산하는 방법으로는 직접 RMST 회귀, 역 확률 가중치 (inverse probability weighting), G-computation 등이 있다. 모든 모델이 올바르게 지정된 복수의 시뮬레이션을 통해 우리는 고려한 모든 방법이 비편향추정값(unbiased estimate)을 제공하고 백분위수 붓스트랩 (percentile bootstrap) 신뢰구간이 명목표함확률(nominal coverage probability)을 달성함을 확인했다.

관찰 연구에서 교란에 대해 보정된 RMST의 차이를 평가하기 위한 몇 가지 방법이 개발되었지만, 측정되지 않은 교란의 민감도 분석(sensitivity analysis)에 대한 연구는 아직까지 없다. 따라서, 우리는 보정된 RMST의 차이를 평가하기 위해 측정되지 않은 교란을 고려하는 새로운 민감도 분석 방법을 제안한다. 사용자 지정 민감도 매개변수가 주어지면, 편향 조정된 RMST 차이(bias-adjusted difference in RMST)의 추정치에 대한 민감도 범위(sensitivity range)와 신뢰구간을 얻을 수 있다. 민감도 범위와 신뢰구간을 얻기 위해서는 복잡한 최적화 문제를 풀어야 하지만, 특별한 경우를 제외하고는 분석적 해가 존재하지 않는다. 최적화 문제의 해를 L-BFGS-B와 같은 최적화 알고리즘을 사용하여 구할 수 있지만 (직접 최적화 방법), 이 경우 해를 구하기 위해 상당한 계산 시간이 소요된다. 따라서, 우리는 편향과 계산시간 모두에서 직접 최적화 방법보다 열등하지 않은 근사 최적화 방법을 제안했고, 집약적인 Monte Carlo 시뮬레이션을 통해 제안한 근사 최적화 방법이 실용적인 대안책이 될 수 있음을 보였다. 민감도 분석을 실제 문제에 적용할 때, 우리는 중도절단률(censoring rate)이 0.7 미만인 경우 근사 최적화 방법을 사용하고, 중도절단률이 0.7 이상인 경우 직접 최적화 방법을 사용하는 것을 권고한다.
-
dc.description.tableofcontents1 Introduction 1
2 Restricted Mean Survival Time 6
2.1 Notation and assumptions 6
2.2 Difference in RMST 7
3 Methods for Estimation of Difference in RMST 9
3.1 Difference in RMST in randomized controlled trial 9
3.2 Difference in RMST in observational study 11
3.2.1 Direct regression 11
3.2.1.1 Pseudo-observation 11
3.2.1.2 ANCOVA-type model 13
3.2.2 Inverse probability weighting 14
3.2.2.1 IP weighted Cox model 14
3.2.2.2 Adjusted Kaplan–Meier estimator 15
3.2.3 G-computation 17
3.3 Simulation study 1 18
3.3.1 Simulation settings 18
3.3.2 True value of difference in RMST 19
3.3.3 Simulation study 1 results 21
3.4 Real data analysis 1: Colon cancer data 25
4 Sensitivity Analysis 28
4.1 Background 28
4.2 Sensitivity model 29
4.3 Estimate of difference in bias-adjusted RMST 31
4.4 Sensitivity range 32
4.5 Analytic solution to bias-adjusted RMST in special case 35
4.6 Methods for solution of optimization problem in general case 38
4.7 Confidence interval for partially identified region 40
4.8 Simulation study 2 41
4.8.1 Simulation study 2.1: Bias and computational time 42
4.8.2 Simulation study 2.2: Sensitivity range and coverage rate 46
4.9 Real data analysis 2 47
4.9.1 Real data analysis 2.1: GBSG data 49
4.9.2 Real data analysis 2.2: NSCLC data 52
5 Discussion 56
Bibliography 61
Appendices 70
Appendix A Appendix for Chapter 3 70
A.1 Example R codes 70
A.1.1 R code for Kaplan-Meier estimator 72
A.1.2 R code for pseudo-observation 73
A.1.3 R code for ANCOVA-type model 74
A.1.4 R code for IP weighted Cox model 75
A.1.5 R code for adjusted Kaplan-Meier estimator 76
A.1.6 R code for G-computation 77
A.2 Proof of true value for difference in RMST 77
A.3 Simulation study 1 results for sample size 1,000 79
A.4 Pooled logistic regression model 83
Appendix B Appendix for Chapter 4 86
B.1 Proof of reducing (4.8) to linear fractional programming in special case 86
B.2 Proof of reducing (4.8) to linear fractional programming in alternative setting 89
B.3 Proof of non-convergence to boundary values 90
B.4 Details for simulation study 2.1 (β0 = −1.9) 94
B.5 Details for simulation study 2.1 (β0 = −0.425) 99
Appendix C Appendix for Chapter 5 105
C.1 Alternative sensitivity analysis method 105
C.2 Limitation of alternative method 108
Abstract (in Korean) 1
-
dc.format.extent120-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subjectrestricted mean survival time-
dc.subjectcausal inference-
dc.subjectsurvival analysis-
dc.subjectobservational study-
dc.subjectsensitivity analysis-
dc.subjectunmeasured confounding-
dc.subject.ddc660.6-
dc.titleDifference in restricted mean survival time in observational studies: A review of estimation methods and a development of sensitivity analysis-
dc.title.alternative제한된 평균 생존시간 추정 방법 고찰 및 새로운 민감도 분석 방법 개발-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorSeungjae Lee-
dc.contributor.department융합과학기술대학원 응용바이오공학과-
dc.description.degree박사-
dc.date.awarded2023-02-
dc.identifier.uciI804:11032-000000174847-
dc.identifier.holdings000000000049▲000000000056▲000000174847▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share