Randomization inference with imperfect compliance in the ACE-inhibitor after anthracycline randomized trial

• 成果类型: Article
• 署名作者: Greevy, R; Silber, JH; Cnaan, A; Rosenbaum, PR
• 署名单位: University of Pennsylvania; University of Pennsylvania; University of Pennsylvania; Pennsylvania Medicine; Childrens Hospital of Philadelphia
• 刊物名称: JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
• ISSN/ISSBN: 0162-1459
• DOI: 10.1198/016214504000000025
• 发表日期: 2004
• 页码: 7-15
• 关键词: to-treat analysis; noncompliance; Intention; identification; variables; tests

摘要

Anthracyclines are quite effective at curing certain cancers of childhood, but they may damage the heart. The ACE-Inhibitor After Anthracycline (AAA) study compared enalapril to placebo in a randomized trial in an effort to determine whether treatment with enalapril would preserve or improve cardiac function among children previously treated with anthracylines. As is true in many clinical trials, patient compliance with the study protocol was imperfect: some children took less than the prescribed dose of enalapril or placebo. Most analytical procedures that acknowledge imperfect compliance do so at significant cost, abandoning the tight logic of random assignment. With noncompliance, assignment to enalapril or placebo is randomized, but the dose of enalapril actually received is not, and self-selection effects parallel to those in observational studies can exist and have been documented in some instances. Some researchers advocate adherence to the strict logic of randomization by reporting only, or else strongly emphasizing, the so-called intent-to-treat analysis, which makes no use of information about compliance. Other researchers report analyses that are not justified by random assignment and can be subject to substantial biases, such as per protocol analyses or treatment received analyses. Here we apply a recent proposal for randomization inference with an instrumental variable that uses randomization as the reasoned basis for inference in Fisher's phrase. We make no assumption that compliance is random indeed, compliance may be severely biased. Importantly, the proposed analysis will find a statistically significant effect of the treatment if and only if the intent-to-treat analysis finds a significant effect; yet, unlike intent-to-treat analysis, our analysis acknowledges that a patient assigned to a drug that he or she does not take will not receive the drug's pharmacological benefits.