Combining treatment effects from clinical trial data and routinely collected healthcare data to answer important research questions

Background Randomised clinical trials (RCTs) can provide an unbiased estimate of the average treatment effect for the population represented by the study participants.

However, they require a significant amount of time, funding and resources to address their research question adequately.

When RCTs terminate with insufficient power, the information collected may be discarded, implying a waste of patient and health professionals time, resources and data collected, and eroding public and patient trust in research.

Observational studies from routinely collected data offer opportunities to provide supplementary evidence to RCTs, but a major concern is inherent biases such as confounding by indication.

Target trial emulations have the potential to address some of these challenges by implementing design principles similar to RCTs. Their use has recently increased in clinical research with the development of the target trial framework.

In addition, there is still much uncertainty about which statistical methods are most appropriate for combining historical data with RCTs.

The Bayesian approach enables results from previous studies to directly contribute to the analysis of the data from the new study through the prior distribution, and can provide a natural and principled way to combine the information.

Aim The overall aim is to define a Bayesian framework for combining treatment effects from RCT data and target trial emulations from routinely collected data, and to exemplify with two case studies.

Methods (1) The FIRST-line support for Assistance in Breathing in Children (FIRST-ABC) study was a master protocol of two pragmatic, non-inferiority RCTs.

We will use the step-up RCT as case study 1, and following the target trial framework emulate it using routinely collected data from the PICANet clinical audit.

The target trial emulation and RCT estimates will be compared, providing an empirical bias assessment. (2) We will conduct a simulation study to compare different Bayesian methods for combining multiple sources of data.

Modelling strategies will include building a joint model through linking multiple sub-models and a multi-stage modelling approach that incorporates information from the previous stage through the prior distribution.

We will compare performance and ease of implementation, and explore different types of priors. (3) We will assess the proposed methodology using case study 1. An early closure of FIRST-ABC will be mimicked, so it no longer answers its research question.

This partial FIRST-ABC data will be combined with the target trial emulation generated from PICANet. (4) Using case study 2 (Immune Modulation domain from REMAP-CAP), we will further test and refine our proposed methods.

Impact and Dissemination The main project output will be a Bayesian framework for conducting a target trial emulation, and combining the resulting data with RCT data to better estimate treatment effects.

Other outputs will include the R and Stan model code, which will be made publicly available, documented case studies, and a simulation exercise.

The novel statistical methodology will enable more effective use of routine data and RCT data from under-powered trials.

This will positively impact the evaluation of treatment strategies and hence healthcare services, benefitting patients and the wider public.