ijsmr logo-pdf 1349088093

Meta-Analysis of Incidence Rate Data in the Presence of Zero-Event and Single-Arm Studies Pages 57-66

Romain Piaget-Rossel and Patrick Taffé


Published: 10 October 2019

Abstract: Unlike the classical two-stage DerSimonian and Laird meta-analysis method, the one-stage random-effects Poisson and Negative-binomial models have the great advantage of including the information contained in studies reporting zero event in one or both arms and in studies with one missing arm. Since the Negative-binomial distribution relaxes the assumption of equi-dispersion made by the Poisson, it should perform better when data exhibit over-dispersion. However, the superiority of the Negative-binomial model with rare events and single-arm studies is unclear and needs to be investigated. Moreover, to the best of our knowledge, this model has never been investigated in the context of a meta-analysis of incidence rate data with heterogeneous intervention effect. Therefore, we assessed the performance of the univariate and bivariate random-effects Poison and Negative-binomial models using simulations calibrated on a real dataset from a study on the surgical management of phyllodes tumors. Results suggested that the bivariate random-effects Negative-binomial model should be favored for the meta-analysis of incidence rate data exhibiting over-dispersion, even in the presence of zero-event and single-arm studies.

Keywords: Incidence rate, Meta-analysis, Negative-binomial model, Poisson model, Rare events, Random effects.


Submit to FacebookSubmit to TwitterSubmit to LinkedIn