Multiple imputation methods for valid causal treatment effects estimation after departures from protocol

Randomised controlled trials are considered the gold standard in determining the effect of an intervention. Two issues that often undermine the credibility of effectiveness measures are non-compliance with treatment allocation (participants who do not adhere to the protocol, e.g. by not receiving the intended treatment) and missing data (e.g. loss to follow-up).

Both these issues are relevant to the adoption of the Intention-to-treat (ITT) analysis, which aims to measure treatment effectiveness. The ITT principle states that all individuals randomised in a clinical trial should be included in the analysis, in the groups to which they were randomised, regardless of any departures from the randomised treatment. Following this principle preserves the benefits of randomisation, i.e. having treatment groups that do not differ systematically on any factors except those assigned in the trial. But such pragmatic estimates may not be the only estimates of interest, and given the considerable costs and high number of patients involved in a typical confirmatory clinical trial, efforts should be made to obtain valid explanatory treatment effects, i.e. the "real" treatment effects that a patient taking the treatment 100% as prescribed can expect on average.

Ad hoc methods which ignore treatment allocation may lead to incorrect estimates of treatment effect. Existing statistical methods can handle some complex longitudinal settings, e.g. when compliance with treatment varies with time and is associated with the clinical outcome of interest (assumed to be a continuous measure). However, they rely on assumptions which are difficult to understand and verify, and involve sophisticated numerical iterative procedures to obtain the estimates of interest. Even the most sophisticated method cannot currently handle the estimation of causal treatment effects on binary outcomes (outcomes that take only two values, e.g. mortality).
It is therefore important that practical and efficient methods are available for handling non-compliance and missing data, especially when the data structure is complex, in a unified, transparent, and systematic manner.

The proposed research aims to develop new statistical methods to deal with these two issues within a single unifying framework and under transparent assumptions.
This novel method is based on multiple imputation, which is a practical and flexible method already widely used to address missing data problems. The methods developed will enable researchers to obtain precise and unbiased estimates of the causal effect of treatment, thereby answering important clinical and public health questions about the average treatment effects at different levels of compliance. This may in turn help individual clinicians and patients take more informed decisions about which treatment to follow.

The methods developed will be published in scientific journals. In addition, new methods will be implemented into statistical software packages to encourage uptake.
The work will be conducted by the Fellow.

The aim is to develop methods based on multiple imputation (MI) for handling deviations from protocol, defined as non-compliance and drop-outs, in time-varying complex randomised clinical trial data where current methods are not applicable (hierarchical data, binary outcomes with continuous non-compliance and bivariate outcomes).
This will be achieved by extending a novel method (called CRK henceforth), based on pattern-mixture models. It uses MI to handle drop-outs and non-compliance in a unified manner, and estimates efficacy (called de jure), by imputing post-deviation data from a distribution whose parameters have been estimated using the observed data from those subjects who comply in their corresponding allocated treatment arm.

Firstly, I will relate this approach to the existing causal inference methods. I will establish relationships between the different assumptions used by each method for parameter identification.

Secondly, a Principal Stratification (PS) approach that uses MI estimation for the complier-average causal effect (CACE) will be extended to time-varying binary non-compliance settings. PS divides the trial population into compliance classes. As the number of possible compliance histories grows exponentially with the number of follow-up visits, CACE estimation becomes high-dimensional. A latent-class representing the time-invariant super-compliance class is used as a pragmatic solution.

Thirdly, I will extend the CRK approach to handle time-varying non-compliance of bivariate outcomes (e.g. cost-effectiveness) and hierarchical data (e.g. data from a cluster randomised trial).

Finally, I will extend the method to binary outcomes in the time-varying non-compliance setting and develop a framework for sensitivity analysis to study the robustness of results when there are departures from assumptions.

Throughout, newly-developed methods will be implemented in statistical packages for R and STATA.

The main beneficiaries of the research output will be researchers interested in testing new interventions through randomised clinical trials and in finding unbiased efficacy estimates. These include trial statisticians and health economists as well as clinicians, in academia and pharmaceutical companies, and those involved in the NIHR Health Technology Appraisal programme.

The research output will also help inform future clinical trial design, by highlighting the importance of good quality data collection of levels of compliance and reasons for non-compliance, in particular if related to adverse effects or to the clinical outcome of interest.

Other beneficiaries include public health policymakers. Reimbursement authorities require cost-effectiveness analysis to determine if a new treatment is cost-effective. Frequently, this is done using clinical trial data, which often fail to adjust for non-compliance, even though the levels of compliance in the trial population are likely to differ from those in general practice. Thus providing a tool for sensitivity of cost-effectiveness estimates to changes in compliance would help medical decision makers, such as NICE, in deciding the robustness of the results.

It will also have an impact on clinical decision making. Individual patients and clinicians may make more informed choices as to: (a) what benefit to expect from following a certain treatment to the letter, or (b) when a particular individual's compliance with treatment is predicted to be less than the "effective" threshold, the clinician and the individual may opt for a different treatment, whose rigours and side-effects are such, that the patient is expected to adhere sufficiently in order to benefit from it. Therefore, I envisage my research could go some way towards personalised medicine, an area of growing interest in the research medical community.