Abstract
In this article we conduct a large quantitative survey of the literature on horizontal and vertical spillovers from foreign direct investment (FDI). We create a unique database of spillover estimates for each country examined in the literature. Next, we estimate the average effect corrected for publication selection bias (the preferential selection of positive and significant estimates for publication). Our results suggest that an average reported estimate of backward spillovers is statistically significant. Publication selection is evident only among studies published in peer-reviewed journals, and only among the estimates that authors consider most important. Authors with small data sets engage in more publication selection. The intensity of selection in the literature decreases over time, which supports the economics-research-cycle hypothesis.
Acknowledgements
We are grateful to Joze Damijan, Ziliang L. Deng, Adam Gersl, Galina Hale, Chidambaran Iyer, Molly Lesher, Marcella Nicolini, Pavel Vacek, and Katja Zajc-Kejzar for sending additional data, or explaining the details of their methodology, or both. The article benefited from discussions at the Meta-Analysis of Economics Research Colloquium, Conway (Arkansas), 2010; and the Global Development Conference, Bogotá (Colombia), 2011. Tomas Havranek acknowledges support from the Czech Science Foundation (grant #P402/12/G097). Zuzana Irsova acknowledges support from the Czech Science Foundation (grant #P402/11/0948) and the Grant Agency of Charles University (grant #76810). The views expressed here are ours and not necessarily those of our institutions. All remaining errors are solely our responsibility.
Notes
1. Some confusion exists in the meta-analysis literature concerning the term ‘fixed-effects estimation.’ Sometimes a simple OLS estimation is labelled ‘fixed effects' to emphasise that no random-effect component is present; the traditional fixed-effects estimation, on the other hand, is rarely applied in meta-analysis since studies reporting only one estimate are dropped in the procedure (see, for instance, Nelson and Kennedy, Citation2009). In our case, though, the large data set of spillover estimates allows us to use the traditional fixed-effects estimation, which is less restrictive than random effects.