Project

Discipline

adverse events; correlated data; meta-analysis; rare outcomes

Lead
Methoden zur Meta-Analyse von Sicherheitsdaten
Lay summary
Wissenschaftler fassen systematisch die Ergebnisse aus mehreren Studien zusammen um z.B. zu entscheiden, ob ein Medikament wirksamer oder sicherer ist als ein anderes. Diese quantitative Zusammenfassung der Ergebnisse von mehreren Studien wird Meta-Analyse genannt. Diese Form der Analyse wird heutzutage von nationalen und internationalen Institutionen im Gesundheitswesen, wie beispielsweise der Weltgesundheitsorganisation WHO, routinemässig eingesetzt bei der Erstellung von Richtlinien und um ganz allgemein die Entscheidungsfindung im Gesundheitswesen zu unterstützen. Meta-Analysen über die Sicherheit von Interventionen bringen spezifische technische Herausforderungen mit sich. Die notwendigen statistischen Methoden für solche Analysen sind noch nicht voll entwickelt insbesondere in Situationen in denen mehr als zwei Interventionen miteinander verglichen werden sollen oder verschiedene Arten von Nebenwirkungen in einer Analyse berücksichtigt werden. Das Ziel dieses Projektes ist es, die aktuell verfügbaren Methoden zur Meta-Analyse von Sicherheitsdaten von medizinischen Interventionen weiterzuentwickeln und neue statistische Ansätze zu entwickeln. Die Methoden werden ausgehend von einer Vielzahl realer Datensätze aus den Bereichen  Orthopädie, Psychiatrie, Kardiologie etc. entwickelt. Das Projekt wird dazu beitragen, dass zukünftig geeignetere Methoden zur Beurteilung der Sicherheit von medizinischen Interventionen zur Verfügung stehen und dadurch letztendlich einen Beitrag zur Risikominimierung für Patienten und Patientinnen leisten.

Direct link to Lay Summary

Last update: 31.03.2016

Active participation

Number	Title	Start	Funding scheme

Background: A synthesis of 42 trials on the safety of rosiglitazone found an increased risk for myocardial infarction and cardiovascular death. This highly influential and publicized meta-analysis resulted in the withdrawal of rosiglitazone from the European market. However, the statistical methodology employed was debated and in subsequent analyses of the same data a son-significant association was suggested casting doubts about the validity of the conclusions drawn in the original analysis. Such conflicting evidence arises because of important methodological gaps in the synthesis of data about the safety of medical interventions. Treatment adverse events (AEs) are typically rare and this poses technical constraints rendering standard statistical methodology impertinent. Although meta-analytic methods for sparse data are available, their applicability is limited in practice because they pertain to the case of a single outcome (univariate meta-analysis) and a single treatment comparison (pairwise meta-analysis). Studies typically report on multiple AEs measured on the same set of patients data so that data is often highly correlated; correlations might also arise as the result of study design, e.g. in cross-over trials. Also, in most medical areas there are multiple available treatments for the same disease, but the currently available methods for the synthesis of sparse data cannot be readily used for network meta-analysis, a statistical technique compiling evidence about multiple competing interventions. Aim: The aim of this project is to advance the methods for synthesizing evidence from randomized trials on the safety of interventions by developing and exploring meta-analytical models for correlated rare events and network meta-analysis of AEs. Methods: Developments will be stimulated by a range of datasets in a variety of clinical research areas. Our starting methodological point will be previous work performed in the area of rare outcomes meta-analysis. We will consider methods that do not assume a normal approximation to dichotomous or count data and do not need the much-criticized continuity correction for studies bearing arms with zero events. These include, but are not limited to, logistic regression, using the arcsine difference, Poisson models and a Beta-Binomial regression model. We will extend such methods for the case of two correlated outcomes by employing their multivariate versions. Most multivariate approaches (e.g. using the multinomial distribution) need detailed input data such as the number of patients with events for both considered outcomes while other multivariate models (e.g. the multivariate Poisson model) need input about the sample correlations. As such information is rarely reported in articles, we will employ a bivariate-Binomial model and we will consider previous work that utilizes external information to bypass the problem of unreported correlations. We will undertake a simulations study to evaluate the properties of the multivariate methods and their advantages compared to univariate approaches. All developments will be programmed in open-source software and routines will be made available. Significance: Multivariate approaches are expected to increase precision in the estimation of the treatment effects. Such an increase in precision is highly desirable in the area of safety where events are rare. Our project is expected to enrich the meta-analytic arsenal with useful tools to assist decision makers in establishing the risk-benefit profile of many competing interventions.