The need to combine the results of clinical studies to yield new insights and to avoid the additional time, effort, and expense of collecting new data is an important area of growing interest. This need is especially acute when new studies either contradict or modify the conclusions of older studies provoking disagreement in the medical community and causing confusion among consumers of health care.
The aim of this study is to further enhance and evaluate three major approaches to combining results from clinical studies: meta-analysis, which uses only summary data from several trials to assess the impact of a treatment; """"""""super-analysis"""""""", which directly combines the primary individual-patient data and then constructs accurate biologically- plausible models of patient risk; and meta-regression, where between- study heterogeneity is modelled with summary data and models of patient risk are developed. Despite their success, many meta-analyses have demonstrated that treatment effects differ significantly across studies (heterogeneity) and over subgroups of patients, raising questions about the appropriateness of only reporting overall effects. Modelling directly on the combined individual-patient data, a technique we have called """"""""super-analysis"""""""", allows a more detailed analysis of the treatment effect. One of the few examples of this method is the Thrombolytic Predictive Instrument (TPI) project conducted by this group of investigators, where a 5000-patient database incorporating individual patient data from 12 major trials and registries of thrombolytic therapy was created, and five logistic regression models were constructed to predict important clinical outcomes relating to the use of thrombolytic therapy. Meta-regression, where models of patient risk may be developed without requiring individual patient data, offers a promising compromise between these approaches. These approaches to combining clinical trial data will be applied to three important clinical areas; the impact of the use of thrombolytic therapy, and the treatment of hypertension and of congestive heart failure. The constructed models will be assessed with respect to their form of presentation, variable selection, estimated treatment effect, model goodness-of-fit, predictive performance, their clinical applicability, and applicability to health care policy. We seek to determine the clinical and health care policy questions relating to medical treatment that each technique can answer and especially to determine what further information derives from using increasingly detailed data.
