Identifying and adjusting for bias due to missing outcome values in randomised controlled trials

Abstract

Randomised controlled trials are considered the gold standard for assessing the causal effect of an intervention on an outcome, but remain vulnerable to bias, for example due to dropout. When outcomes are missing not at random (MNAR), the estimated treatment effect will be biased when missingness is ignored, as in a complete case analysis (CCA), but also when using methods that assume outcomes are missing at random (MAR), such as multiple imputation.

The National Research Council and European Medical Agency recommend assessing the robust- ness of the main analysis by performing sensitivity analyses under a range of plausible alternative MNAR assumptions, using pattern mixture models (PMMs) and selection models (SMs). The Cochrane Collaboration advise assessing the risk of MNAR bias by checking for differential dropout, while other sources additionally recommend examining the distribution of baseline co- variates across trial arms in the observed data. However, these guidelines are rarely implemented in practice. Within this thesis, I aimed to develop novel markers of MNAR dropout and sensitivity analyses that could be adopted to assess and account for the risk of bias. In all chapters, I describe the methods and apply these to multiple trials.

In Chapter 2, I propose using observed differences in outcome variance between trial arms as a marker of MNAR dropout. In Chapter 3, I investigate additional potential markers of MNAR dropout across various dropout scenarios, showing that existing risk-of-bias tools are not conclusive. I propose additional markers that can be used in concert to assess the risk of MNAR dropout. In Chapter 4, I consider the trimmed means (TM) estimator, which estimates the treatment effect using data remaining after trimming equal fractions of both groups. I propose using this estimator as simpler alternative to PMM and SM sensitivity analyses and derive formulae for establishing the direction of bias of CCA and TM estimates.

Date of Award	9 May 2023
Original language	English
Awarding Institution	University of Bristol
Supervisor	Kate M Tilling (Supervisor), Tom M Palmer (Supervisor) & Kaitlin H Wade (Supervisor)