Using Selection Models to Deal with Publication Bias

• Activity Number: 63
• Type: Invited
• Date/Time: Monday, August 9, 2004 : 8:30 AM to 10:20 AM
• Sponsor: Section on Health Policy Statistics
• Author(s): Norma Terrin*+ and Christopher H. Schmid and Michael F. Dowd
• Companies: Tufts-New England Medical Center and Tufts-New England Medical Center and Tufts-New England Medical Center
• Keywords: publication bias ; selection model ; trim and fill ; funnel plot

Abstract

Publication and related biases occur when statistically significant results are more accessible than nonsignificant results. The funnel plot, a popular tool for detecting publication bias, is a scatterplot of studies in a meta-analysis, with a measure of effect on the horizontal axis, and a measure of precision on the vertical axis. Asymmetry in the funnel plot is interpreted as evidence of publication bias. The "Trim and Fill" method imputes studies to make the funnel plot appear symmetric, then pools actual and imputed studies to obtain an overall effect estimate. However, asymmetry may have causes other than publication bias, including study heterogeneity and chance. Selection modeling is a method of bias adjustment that does not use the funnel plot. Study effects are modeled using random effects, and the selection process is modeled by assigning a weight to the estimated effect from each study. The form of the weight function may be parametric or nonparametric, and the estimation method may be maximum likelihood or Bayesian. We describe and compare the performance of several selection models, applied to a wide range of simulated and actual meta-analyses.