An overview of statistical methods for handling nonadherence to intervention protocol in randomized control trials: a methodological review

Abstract

Objective: To undertake a methodological review of statistical methods used in randomized controlled trials (RCTs) for handling intervention nonadherence.

Study design and setting: Bibliographic databases were searched using predefined search terms.

Results: A substantive number of identified studies (56%) were excluded as they only used naive per protocol analysis for handling nonadherence. Our review included 58 articles published between 1991 and 2015. A total of 88 methodological applications were made by these studies. The two most used methods were complier average causal effect (56%) and instrumental variable (23%) predominantly with the use of maximum likelihood (ML) estimators. These alternative applications typically produced treatment effects greater than the intention-to-treat effect but as their standard errors were larger there was no statistical difference between the methods.

Conclusion: A substantive proportion of RCTs rely on naive per protocol for handling nonadherence. Recent years have seen an increasing number of applications of more appropriate statistical methods, in particular complier average causal effect and instrumental variable methods. However, these later methods rely on strong underlying assumptions that may be vulnerable to violation. More empirical studies are needed that directly compare the usability and performance of different statistical methods for nonadherence in RCTs.

Keywords: Causal effect modeling; Methodological review; Nonadherence; Noncompliance; Randomized controlled trial; Statistical methods.