Estimate Dependency in Meta-Analysis: A Generalized-Weights Solution to Sample Overlap
2017/08/22
研討會日期 : 2017-08-22
時間 : 14:30
主講人 : Professor Heiko Rachinger
主持人 : Professor Yu-Chin Hsu
地點 : Conference Room B110
演講者簡介 : Professor Rachinger received his Ph.D. from Charles III University of Madrid in 2012. He is currently an Assistant Professor of Economics at the University of Vienna. His research interests include econometrics, macroeconometrics, and finance.
演講摘要 : A common feature of meta-analyses in economics, especially in macroeconomics and related subfields, is that the samples underlying the reported effect sizes overlap. The resulting positive correlation between effect sizes decreases the efficiency of standard meta-estimation methods. Ignoring sample overlap generates downward-biased standard errors and, thus, invalid meta-inference. This paper argues that the variance-covariance matrix describing the structure of dependency between primary estimates can be feasibly specified as function of information that is typically reported in the primary studies. Meta-estimation efficiency can then be enhanced by using the resulting matrix in a Generalized Least Squares fashion. Our simulations illustrate efficiency losses and size distortions potentially arising from empirically-relevant sample overlap scenarios.
