Should We Use Bayesian Methods in Meta-Analysis?.

Bayesian statistical methods are becoming increasingly popular in agronomy but, so far, most meta-analyses were performed using classic (frequentist) statistical methods in this area of science. In this study, I review several published meta-analyses in agronomy and in plant pathology where classic and Bayesian methods were both used to analyze the same datasets. I compare the results obtained with these two types of statistical methods in order to determine whether conclusions derived from classic methods differed from those obtained with Bayesian methods and, also, to identify the advantages and limits of both approaches. The meta-analyses considered here cover a large range of topics related to yield estimation, environmental impacts of farmers’ practices, or crop protection against diseases. In these studies, mean effect sizes estimated with both types of methods were very close and led to similar conclusions in most cases. As Bayesian methods require more programming effort and longer computation times, these results show that there is no clear practical advantage in using a Bayesian method compared to a classic method. However, in some applications, Bayesian methods provide a more complete description of uncertainty in estimated effect sizes. Another potential interest of Bayesian methods is that they could take prior expert knowledge (independent from the available data) into account, but research is still needed to determine what kind of prior knowledge could be used when performing meta-analysis in agronomy.