Abstract

Although background evidence always informs clinical study design, individual trial data are often first analysed in isolation and then meta-analysis is used to synthesise evidence accrued across multiple studies. However, in “small population” settings such as rare diseases, paediatric populations, or personalised/stratified medicine, the feasible sample size that can be recruited in each individual study typically falls short of that needed for a conventionally-powered trial. To address these limitations, we assess two avenues for integrated analysis of clinical data. First, we discuss the use of Bayesian informative priors based on relevant external data to increase the power and precision of such trials and propose a range of operating characteristics that can be useful to evaluate and compare designs. Here we present various prior and posterior summaries of the available historical, current and total evidence which can be used to help sponsors and regulators assess the strength of the prior assumptions and the extent to which the current trial data can influence the final posterior inference. We illustrate these methods through some recent case studies covering both paediatric settings borrowing efficacy data from adults, and in confirmatory trials in adults borrowing historical controls. Second, we explore the use of genetic annotation for design and analysis of biomarker-rich “large p, small n” early phase oncology clinical trials. While the primary goal of these trials is to assess safety and preliminary clinical efficacy in a possibly heterogeneous patient population, a key secondary objective is to identify prognostic and predictive markers for design adaptation and for developing companion diagnostics. Here we focus on the analysis of associations between baseline gene expression and response to therapy, comparing the operational characteristics of bottom-up and top-down methods. While bottom-up methods use gene annotations for interpretation of the analysis results, top-down methods incorporate these background data within the structure of a hierarchical model likelihood for estimating pathway-level associations to clinical response. We show that, while top-down methods benefit from the advantages of parsimonious parametric modelling, they are not less exposed than bottom-up methods to a reliance on a knowledge basis that is both incomplete and in constant evolution.