Advanced statistical methods in meta-analysis with applications in preclinical neurological drug discovery data

Conventional drug development generally starts with laboratory studies in animals, which can be used prior to testing safety and efficacy in humans to justify potential costs and risks. In the neurosciences, this has been characterised by substantial efficacy observed in animal studies that do not translate to similar efficacy in the clinic. This “translational failure” may in part be due to poor design and analysis of animal studies. Treatment effects are highly homogeneous within animal studies and highly heterogeneous between animal studies. The opposite is true in clinical trials. Meta-analysis of preclinical data is used to understand the sources of heterogeneity in experimental findings and involves different statistical issues from those relating to clinical meta-analysis. In this thesis, I investigate the methodological challenges of preclinical meta-analysis for estimating and explaining heterogeneity based on neurological drug discovery data. In the first part of the thesis, assuming aggregate level data for a continuous outcome, I present a comprehensive summary and comparison of the most common methods for estimating and quantifying heterogeneity in meta-analysis of preclinical data. I focus on two topics: (1) estimation of heterogeneity using method of moments (Dersimonian-Laird, DL), maximum likelihood (REML) and a Bayesian approach; and (2) comparison of univariate versus multivariable meta-regression for adjusting heterogeneity in treatment effects between studies. My findings indicate no difference between REML and the Bayesian method, and both are recommended over DL. Moreover, I show that multiple meta-regression is worthwhile to explain heterogeneity, reducing more variability than any univariate models and increasing the explained proportion of heterogeneity. For further understanding heterogeneity in animal studies, and to evaluate the sufficiency of current preclinical evidence for translation, I empirically investigate how measures of heterogeneity from meta-analyses of animal studies change as evidence accumulates. I explore how heterogeneity measures change with the inclusion of more studies using cumulative meta-analyses and cumulative meta-regression of seven systematic review datasets of varying sizes. The preliminary findings suggest that it may be possible to identify systematic characteristics of heterogeneity within preclinical datasets which can be used alongside other measures to guide decisions to proceed with human testing. The second part of this thesis focuses on investigating the approaches for quantifying heterogeneity using individual animal data meta-analysis. Individual data meta-analysis is the gold standard for synthesising evidence in clinical trials as it allows detailed data checking and exploration of factors contributing to heterogeneity. However, it is rarely undertaken due to limited access to original data. To explore potential benefits of individual data meta-analysis in preclinical setting, under a Bayesian framework, I examine the relationship between individual and aggregate level meta-analysis in quantifying heterogeneity based on a general linear mixed-effects model methodology and consider the impact of distributional assumptions. My findings highlight that despite providing similar results as the aggregate level analysis, individual level analysis offers more flexibility to explore model assumptions, implement different distributions and explore potential effect modifiers at the animal level. Additionally, I illustrate the impact of the number of animals and studies and the magnitude of heterogeneity on the accuracy of estimates using simulations. Extending the exploration of methods in quantifying heterogeneity in meta-analysis with one endpoint, I present my work investigating the joint synthesis of multiple correlated endpoints that are repeatedly measured over a longitudinal profile, based on a multivariate longitudinal meta-analysis of individual animal data. Using linear mixed model methodology, I illustrate how the joint synthesis of multiple correlated endpoints can be applied on preclinical meta-analysis also accounting for longitudinal structure. I further discuss its advantages over separate univariate meta-analyses ignoring correlations