A random-effects meta-regression model for studying nonlinear dose-response relationship

INTRODUCTION. A fundamental challenge in meta-analysis of published epidemiological dose-response data is the estimation of the function describing how the risk of disease varies across different levels of a given exposure. The usual approach consists of estimating the linear change in the natural logarithm of the relative risk estimate per unit of exposure within each study, and then combining these estimates across studies [1]. Three major statistical issues to deal with when using this approach have been reported in literature: (i) the correlation among reported dose-specific logRRs estimates due to the common reference group within the same study (ii) the heterogeneity between studies and (iii) the nonlinear trend components of the dose response relationship. AIMS. The aim of our work is to develop a method that addresses simultaneously the three statistical issues cited above by implementing a random-effects meta-regression model in a nonlinear dose-response relationship framework. To illustrate the proposed methodology, the results of a meta-analysis to study the effect of alcohol on the risk of esophageal cancer are showed. CONCLUSION. In this work, we have discussed a flexible curve-regressing method to perform random effects meta-analysis of epidemiological dose- response data. The proposed model has been found to perform satisfactorily in estimating the dose-response relationship between alcohol consumption and esophageal cancer risk, taking into account both the within-study and between-studies variances.