Published online by Cambridge University Press: 06 November 2020
Cross-national statistical research based on “all country” data sets involves no deliberate selection and hence ignores the potential for endogenous selection bias. We show that these designs are prone to selection bias if existing units are subject to differential survival rates induced, in part, by treatment. Using rudimentary graph theory, we present survivorship bias as a form of collider bias, which is related to but distinct from selection on the dependent variable. Because collider bias is always relative to a specific causal model, we present a causal model of post-colonial sovereignty on the Arabian Peninsula, show that it implies survivorship bias in the form of false positives with respect to the political resource curse, and provide historical evidence confirming that the model correctly depicts the creation of sovereign countries on the Arabian Peninsula but not elsewhere. When we correct for endogenous selection bias, the effect of oil on autocratic survival is shown to be negligible. The study motivates the need to think more broadly about the nature of the data-generating process when making causal inferences with observational data and to construct statistical models that are sensitive to treatment heterogeneity and rooted in context-specific knowledge and qualitative inferences.
A list of permanent links to Supplemental Materials provided by the authors precedes the References section.
Data replication sets are available in Harvard Dataverse at: https://doi.org/10.7910/DVN/ZU3K7H
Versions of this article were presented at the August 2018 meeting of the American Political Science Association and at the Workshop on Oil and the Changing Rentier State convened by the Project on Middle East Political Science at George Washington University. They are grateful for the comments of participants on these occasions. They also received invaluable comments from Gregory Gause, Steffen Hertog, Jack Paine, Michael Ross, Jan Vogler, and the editor and anonymous reviewers of this journal. Simonas Cepenas, Christopher Dictus, and Hye Ryeon Jang provided excellent research assistance.