Abstract

Meta-analysis was developed to provide a quantitative summary of published results from several studies, avoiding the need of the patient-level data. Pooling directly the data is not a correct solution (Simpson's paradox) and meta-analysis requires specific statistical tools. The meta-analytic approach is especially appreciated by clinicians because of the huge number of scientist papers. In this talk, we present and discuss some techniques of meta-analysis when the outcome is binary and time-dependant (event:yes/no, and time-to-event) and two arms (or treatments) are compared. The most common statistical method is to obtain a weighted mean of the appropriate measure of association between the treatment and the outcome. The weights are the inverse of the variance and can take the inter-study variability into account (random effects model). The odds ratio and the risk ratio are not an appropriate measures of association because their estimates depend on the duration of follow-up. This may result in a high heterogeneity caused by the statistical method. We have proposed the relative log-survival as a useful alternative: it is an estimate of the hazard ratios assessable from a 2-by-2 table. Another possibility is to use a statistical method for the assessment of summary survival curves from published survival probabilities collected at several time-points. This parametric method assumes a Weibull, log-logistic or log-normal survival functions. The main disadvantage is the heterogeneity between survival curves, which may be caused by differences in the characteristics of the study populations, but does not necessarily influence the treatment effect. Our approach is to model directly the treatment effect assuming the same parametric family of the survival functions but without assessing the summary survival curves. It is a generalization of the relative log-survival which allows testing the assumption of a constant treatment effect over time, and which requires fewer parameters than the previous approach.