Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form

Causal effect estimation in randomized controlled trials with imperfect compliance

Lyu, Lingyun (2018) Causal effect estimation in randomized controlled trials with imperfect compliance. Master's Thesis, University of Pittsburgh. (Unpublished)

[img] PDF
Submitted Version
Restricted to University of Pittsburgh users only until April 2021.

Download (2MB) | Request a Copy

Abstract

Randomized controlled trials (RCT) are widely considered as the gold standard in generating evidence about the efficacy and safety of an experimental treatment. In practice, however, RCTs often suffer from non-compliance to the assigned treatment threatening the validity of the study results. Intent-to-treat (ITT) has been widely adopted as the standard analyses for such trials. However, under imperfect compliance, ITT validly estimates the treatment effectiveness instead of the treatment effect as-received (efficacy). Under the potential outcomes framework and certain assumptions, the treatment effect as-received may be represented by the Complier Average Causal Effect (CACE), the average treatment effects in the subgroups of compliers. Common methods used to estimate the CACE are As-Treated and Per-Protocol, both of which may introduce confounded comparisons between treatment arms due to the inherent differences between compliers and non-compliers. To provide valid estimates of CACE, causal inference methods such as propensity score (PS) and instrumental variables (IV)-based approaches have been proposed in the literature. As long as an instrument exists, IV-based methods could provide inferences that are less model dependent. They do not necessarily require adjusting for covariates and avoids model selection and specification issues that PS-based methods face for the propensity-to-comply model. Due to random allocation, the randomization assignment often meets the assumptions imposed by an instrument and is widely accepted as a valid instrument in many situations. The most common IV-based estimation method is 2-Stage-Least-Squares (2SLS). For binary outcome and binary treatment groups, estimating risk ratios or odds ratios have been the subject of many studies in the literature. When interest lies in estimating the risk difference (RD) as the CACE, a linear probability model in the second stage is commonly used. However, there is lack of consensus about what is the most suitable in the first stage where the observed treatment received is regressed to the treatment assignment (instrument).
The goal of this study is to empirically investigate the different IV-based approaches to estimate the risk difference as CACE in RCTs with binary outcome and binary treatment group. We compared the performance of these methods with respect to bias, efficiency, and power and compare these to PP as the standard approach to estimate CACE. We also examined how their performance is affected by varying levels of compliance, effect size, sample size. In addition, we evaluated their statistical properties when measured confounders exist.
We found that all the IV-based methods generally provide valid and very similar estimates, efficiency and power in the setting where there are no measured confounders, while the PP shows large bias in the presence of unmeasured confounders. However, when we can account for measured confounders, a 3-stage approach may provide more efficient estimates and yield higher power. As the compliance probability goes to 1 or as the sample size increases, the differences between the different IV-based methods become negligible.
Public health significance: Results of RCTs are commonly used to implement policies or recommend guidelines to improve public health and patient care. Non-compliance however is common in RCTs and threatens the validity of its results. This study compares different strategies in providing correct estimates of treatment effect under imperfect compliance. This is critical in assessing the utility of an experimental treatment for adoption in clinical practice.


Share

Citation/Export:
Social Networking:
Share |

Details

Item Type: University of Pittsburgh ETD
Status: Unpublished
Creators/Authors:
CreatorsEmailPitt UsernameORCID
Lyu, Lingyunlil114@pitt.edulil114
Contributors:
ContributionContributors NameEmailPitt UsernameORCID
Committee ChairYabes, Jonathanjgy2@pitt.edujgy2UNSPECIFIED
Committee MemberYouk, Adaayouk@pitt.eduayoukUNSPECIFIED
Committee MemberJacobs, Bruceblj27@pitt.edubllj27UNSPECIFIED
ETD Committee:
TitleMemberEmail AddressPitt UsernameORCID
Committee ChairYabes, Jonathanjgy2@pitt.edu
Committee MemberYouk,, Adaayouk@pitt.edu
Committee MemberJacobs, Bruceblj27@pitt.edu
Date: 5 April 2018
Date Type: Submission
Defense Date: 18 April 2018
Approval Date: 28 June 2018
Access Restriction: 3 year -- Restrict access to University of Pittsburgh for a period of 3 years.
Number of Pages: 75
Institution: University of Pittsburgh
Schools and Programs: Graduate School of Public Health > Biostatistics
Degree: MS - Master of Science
Thesis Type: Master's Thesis
Refereed: No
Uncontrolled Keywords: Randomized controlled trials (RCT); Complier Average Causal Effect (CACE); instrumental variables (IV); Per-Protocol(PP); binary outcome; risk difference (RD)
Date Deposited: 28 Jun 2018 20:11
Last Modified: 02 Jul 2018 21:30
URI: http://d-scholarship.pitt.edu/id/eprint/34159

Metrics

Monthly Views for the past 3 years

Plum Analytics


Actions (login required)

View Item View Item