Project

Back to overview

On the usefulness of causal observational studies to reduce bias and improve estimates in meta-analysis of rare binary events

English title On the usefulness of causal observational studies to reduce bias and improve estimates in meta-analysis of rare binary events
Applicant Taffé Patrick
Number 169504
Funding scheme Project funding (Div. I-III)
Research institution Unité de Prévention Communautaire IUMSP Université de Lausanne et CHUV
Institution of higher education University of Lausanne - LA
Main discipline Mathematics
Start/End 01.01.2017 - 31.12.2019
Approved amount 185'358.00
Show all

All Disciplines (2)

Discipline
Mathematics
Medical Statistics

Keywords (4)

propensity score methods; sparse data meta-analysis; causal observational studies; rare events

Lay Summary (French)

Lead
La méta-analyse est une procédure statistique permettant de synthétiser les résultats quantitatifs de plusieurs études publiées sur un thème particulier, comme la survenue d'événements secondaires lors d'une intervention médicale. C’est une méthodologie très utile dans le contexte de la médecine fondée sur les preuves (EBM). Cependant, lorsque les événements sont rares toutes les méthodes statistiques conventionnelles de méta-analyse produisent des résultats imprécis et biaisés. Une solution à ce problème est de considérer dans la méta-analyse, en plus des essais cliniques randomisés (RCTs), les études observationnelles causales.
Lay summary
Utilité des études observationnelle causales pour réduire le biais et améliorer les estimations dans les méta-analyses d’événements binaires rares

Lead

La méta-analyse est une procédure statistique permettant de synthétiser les résultats quantitatifs de plusieurs études publiées sur un thème particulier, comme la survenue d'événements secondaires lors d'une intervention médicale. C’est une méthodologie très utile dans le contexte de la médecine fondée sur les preuves (EBM). Cependant, lorsque les événements sont rares toutes les méthodes statistiques conventionnelles de méta-analyse produisent des résultats imprécis et biaisés. Une solution à ce problème est de considérer dans la méta-analyse, en plus des essais cliniques randomisés (RCTs), les études observationnelles causales.

Contenu et objectifs du travail de recherche

L’objectif principal de ce projet de recherche est de démontrer l’utilité et l’importance des études observationnelles causales, en particulier dans le contexte des méta-analyses d’événements rares, et de briser le dogme « seuls les RCTs fournissent des preuves fiables ». En effet, les RCTs sont généralement réalisés sur une sous-population peu représentative de la population générale, ce qui peut induire un biais de sélection important, tandis que les études observationnelles sont en général plus représentatives et de plus grandes tailles, ce qui permet d’estimer le risque relatif avec moins de biais et une plus grande précision.

 Contexte scientifique et social du projet de recherche

Notre travail permettra non seulement de clarifier les limites des méthodes statistiques conventionnelles de méta-analyse, mais surtout de mettre en évidence l’importance fondamentale des études observationnelles causales et leur rôle clé dans les méta-analyses et l’EBM.

Mots-clés

Meta-analysis, rare events, sparse data, causal methods, propensity score.

Direct link to Lay Summary Last update: 19.10.2016

Responsible applicant and co-applicants

Employees

Name Institute

Abstract

SummaryBackground: It is well recognized that standard methods of meta-analysis break down with rare binary events, that is, if the events we wish to observe occur sparsely. Not only are effect sizes and within study variances badly estimated, but heterogeneity is generally not identifiable or strongly underestimated, and the overall summary index is biased. Despite considerable efforts in the development of new methods to try to improve the estimates in sparse data meta-analysis, none of the methods developed so far have provided a convincing solution. Indeed, with rare or very rare events (i.e. 3~5 or 0~2 events, as typically arises with event rates comprised within 5%~1% or even < 1% and 50~250 patients per trial arm) parameter estimation is generally plagued by bias in sample sizes often used in randomized controlled trials (RCT), particularly when the groups are markedly unbalanced. Also, in the case of total zero-event trials the continuity correction and statistical method used may influence the inference drawn from the meta-analysis.When event rates are low and data sparse conventional RCTs often do not provide enough information to reliably estimate prevalence of incidents, whatever the statistical method used. To remedy this problem, one should seek supplementary information in the published literature, namely in causal observational studies. A causal observational study is a study where the effect size of interest has been estimated by a causal methodology, such as matching or propensity score methods, and not simply by standard regression methods. Due to their more robust nature, causal methods have proven to be very useful in observational studies and are able to recover effect sizes of the same amplitude as in RCTs. In addition, individuals enrolled in observational studies are generally more representative of the target population than people enrolled in RCTs. Thus, we may combine both RCTs and causal observational studies to achieve increased precision and validity in effect size estimation.Our goal, therefore, is to show by simulations and case studies that causal observational studies not only allow to increase precision and validity of effect size estimation, but also that they provide a rather natural solution to the problems of sparse data meta-analyses.Methods: We will use simulation studies to generate populations in which the true causal effect of an intervention aimed at reducing the rate of an adverse event is known. The simulation parameters will be set to reproduce published meta-analyses on death and safety outcomes, which are very often plagued by sparse data. In the fixed effect setting the impact of the intervention will be homogeneous, whereas in the random effects setting it will be heterogeneous and partially depend on measurable as well as unmeasurable variables. The intervention (e.g. treatment) will be allocated completely at random in the samples mimicking RCTs, whereas it will depend on individuals’ characteristics (such as confounding factors and other predictors) in the samples mimicking observational studies. Thus, each sample, depending on the mechanism of selection of individuals and allocation of the intervention adopted, will represent either an observational study or a RCT.Based on these simulated data, the performance (in terms of bias and precision) of various recently published methods of meta-analysis of rare binary averse event data both in settings with and without heterogeneity, will be assessed and compared to our proposal.
-