Model selection in dose-response meta-analysis of summarized data

A linear mixed-effects model for the synthesis of multiple tables of summarized dose-response data has been recently proposed and implemented in the drmeta command. One of the main advantages offered by this framework is the possibility to fit complex models avoiding exclusion of studies contrasting a limited number of doses. The aim of this presentation is to evaluate the ability of Akaike's information criterion (AIC) to suggest the true dose-response relationship. Statistical experiments are conducted under the assumption of either a linear (Shape 1) or nonlinear (Shape 2) relationship between a quantitative dose and mean outcome. Tables of summarized data are generated upon categorization of the dose into quantiles. Every simulated dose-response meta-analysis is analyzed with a linear-mixed effects model using two commonly used strategies: linear function and splines. Accuracy of the AIC is assessed by calculating the proportion of times in a large number of experiments the Shape 1 and Shape 2 are correctly identified by choosing the lowest AIC among the two modeling strategies. I also explore how this accuracy may vary according to the distribution of the dose and the way it has been categorized.