Modelling departure from randomised treatment in randomised controlled trials with survival outcomes

Dodd, Susanna ORCID: 0000-0003-2851-3337
Modelling departure from randomised treatment in randomised controlled trials with survival outcomes. PhD thesis, University of Liverpool.

[img] Text
DoddSus_Dec2014_2006887.pdf - Unspecified
Available under License Creative Commons Attribution.

Download (3MB)


Randomised controlled trials are considered the gold standard study design, as random treatment assignment provides balance in prognosis between treatment arms and protects against selection bias. When trials are subject to departures from randomised treatment, however, simple but naïve statistical methods that purport to estimate treatment efficacy, such as per protocol or as treated analyses, fail to respect this randomisation balance and typically introduce selection bias. This bias occurs because departure from randomised treatment is often clinically indicated, resulting in systematic differences between patients who do and do not adhere to their assigned intervention. There exist more appropriate statistical methods to adjust for departure from randomised treatment but, as demonstrated by a review of published trials, these are rarely employed, primarily due to their complexity and unfamiliarity. The focus of this research has been to explore, explain, demonstrate and compare the use of causal methodologies in the analysis of trials, in order to increase the accessibility and comprehensibility by non-specialist analysts of the available, but somewhat technical, statistical methods to adjust for treatment deviations. An overview of such methods is presented, intended as an aid to researchers new to the field of causal inference, with an emphasis on practical considerations necessary to ensure appropriate implementation of techniques, and complemented by a number of guidance tools summarising the necessary clinical and statistical considerations when carrying out such analyses. Practical demonstrations of causal analysis techniques are then presented, with existing methods extended and adapted to allow for complexities arising from the trial scenarios. A particular application from epilepsy demonstrates the impact of various statistical factors when adjusting for skewed time-varying confounders and different reasons for treatment changes on a complicated time to event outcome, including choice of model (pooled logistic regression versus Cox models for inverse probability of censoring weighting methods, compared with a rank-preserving structural failure time model), time interval (for creating panel data for time-varying confounders and outcome), confidence interval estimation method (standard versus bootstrapped) and the considerations regarding use of spline variables to estimate underlying risk in pooled logistic regression. In this example, the structural failure time model is severely limited by its restriction on the types of treatment changes that can be adjusted for; as such, the majority of treatment changes are necessarily censored, introducing bias similar to that in a per protocol analysis. With inverse probability weighting adjustment, as more treatment changes and confounders are accounted for, treatment effects are observed to move further away from the null. Generally, Cox models seemed to be more susceptible to changes in modelling factors (confidence interval estimation, time interval and confounder adjustment) and displayed greater fluctuations in treatment effect than corresponding pooled logistic regression models. This apparent greater stability of logistic regression, even when subject to severe overfitting, represents a major advantage over Cox modelling in this context, countering the inherent complications relating to the fitting of spline variables. This novel application of complex methods in a complicated trial scenario provides a useful example for discussion of typical analysis issues and limitations, as it addresses challenges that are likely to be common in trials featuring problems with nonadherence. Recommendations are provided for analysts when considering which of these analysis methods should be applied in a given trial setting.

Item Type: Thesis (PhD)
Additional Information: Date: 2014-12-23 (completed)
Subjects: ?? Q1 ??
Depositing User: Symplectic Admin
Date Deposited: 18 Aug 2015 09:45
Last Modified: 17 Dec 2022 00:50
DOI: 10.17638/02006887