Methods of Estimation in Random Effects Meta-Regression

• Activity Number: 465
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Government Statistics
• Author(s): Abera Wouhib*+ and Myron J. Katzoff
• Companies: National Center for Health Statistics and Centers for Disease Control and Prevention/NCHS
• Keywords: Meta-regression model ; Random effects ; Heterogeneity parameter

Abstract

The term meta-regression often refers to a linear models of effect size with covariates. Unlike common versions of regression analyses, which include available record-level data on outcomes and covariates, meta-regressions may consist study-level covariates in addition to record-level covariates. Random effects meta-regression models account for non-systematic differences among study outcomes which cannot be explained by sampling variability alone. We demonstrate a new estimation method using a random effect meta-regression model along with other methods of meta-analysis to estimate effect size of interest and the heterogeneity component of the variance using a simulation study. In the simulation study, we draw samples from a stochastic model that closely emulates the design of health surveys. We select the estimators of effect size which are best in terms of minimizing mean-square-error and bias. We then apply these estimators to data from health surveys and explore the causes of heterogeneity in effect sizes by including covariates at the study-level, record-level or both. Finally, we compare statistical methods, and discuss the limitations and drawbacks of these methods.