Nonlinear mixed effects models for meta-analysis.

Meta-analysis is sometimes defined narrowly as the combination of summary statistics from different studies. Here we consider cases where the raw data from each study are available, and explore hierarchical modeling approaches to meta-analysis. Any statistical analysis may present hazards---such as model misspecification and overly influential observations---and meta-analysis is particularly susceptible. One approach to this problem is through the use of graphical methods for diagnosing deviations from our assumptions. In this work we develop several new displays for meta-analysis, including the raindrop plot for displaying information from many studies simultaneously. Approaches to the estimation of linear and nonlinear mixed effects model are reviewed and generalizations of a segmented regression model are developed for use in mixed model analyses. A different way of dealing with the pitfalls of meta-analysis is through the use of robust statistics. Huggins and Richardson previously developed methods for robust estimation of linear mixed effects models. Nonlinear mixed effects models present additional challenges. Based on the approach of Huggins, we propose modifications to robustify two algorithms for estimating nonlinear mixed effects models. Our methods are illustrated using data on ulcer studies, wolf populations, and coho salmon populations.