For both count and incidence rate data, it is complicated to provide reliable inference of a treatment effect when the number of observed events is too low. Therefore, the idea of regrouping several studies to increase the amount of available information seems particularly appealing in such settings. Unfortunately, standard meta-analysis methods break down with rare events. This thesis aimed at studying the challenge of combining the lack of information. Throughout four articles, we assessed, via simulations, the performance of several alternative meta-analysis methods that better accommodate rare events. Not only did we consider existing methods, but we also designed innovative methods for both count and incidence rate data. Based on the results obtained in these different papers, we were able to draw several recommendations for applied researchers. With count data, and under the assumption of a homogeneous treatment effect, the Mantel-Haenszel method can be used safely, no matter the scarcity level considered. A newly designed pseudo-likelihood approach performed as well as the Mantel-Haenszel method and allowed a gain of precision when the meta-analysis included studies with missing treatment arms. Moreover, unlike Mantel-Haenszel, this pseudo-likelihood approach could be extended to settings with treatment effect heterogeneity and was shown to provide good estimates of the mean treatment effect and informative prediction intervals, even in extremely rare event settings. As for the meta-analysis of incidence rate data, we found that accounting for over-dispersion using a negative-binomial model allowed for improvements in the performance of the classical Poisson model, even in the presence of studies reporting zero event and/or only one treatment arm.