Estimating pooled survival benefit using individual patient-level data meta-analyses in cost-effectiveness analysis of health care technologies: case and simulation study.
Béranger Lueza Msc, Department of Biostatistics and Epidemiology, Gustave Roussy, Villejuif, France CESP, INSERM U1018, Université Paris-Sud, Villejuif, France PhD student, Health Economist/Biostatistician
Monday, 23 March 2015, 12pm to 1pm
Richard Doll Lecture Theatre, Richard Doll Building, University of Oxford Old Road Campus, OX3 7LF
Hosted by HERC
The aim of a meta-analysis is to estimate a pooled treatment effect from data of all randomised controlled trials that investigated the same clinical question. Individual patient data (IPD) meta-analysis (MA) has become the gold standard for obtaining the best evidence to inform treatment effects. In economic evaluation, a commonly used outcome measure for the treatment effect is survival benefit (SB). A challenge when using IPD from a MA is taking into account the hierarchical structure of the data (data clustering with patients within trials) to estimate the pooled SB. The impact of using different methods to estimate treatment effect using IPD-MA has not been extensively explored in the literature. In order to contribute to fill this gap, our objective was to compare survival analysis methods (both parametric and non-parametric) to estimate the pooled SB from IPD-MA. First, we show a case study in lung cancer that suggests that the method chosen to estimate the pooled SB is likely to impact cost-effectiveness results. Then, we present results from a simulation study in which we introduced between-trial heterogeneity (treatment effect and baseline hazard) through discrete random effects. It allowed us to analytically derive the “true” pooled SB. We considered scenarios varying key parameters such as treatment effect size, between-trial heterogeneity, the number of trials, and the number of patients. Over 1,000 replicates per scenario, we compared the survival analysis methods in terms of bias, empirical standard error, and average standard error. Preliminary results suggest that all methods perform well even for high treatment effect and high heterogeneity. However, with high treatment effect heterogeneity, the variance of pooled SB is under-estimated.