Regression-Discontinuity Designs and Popular Elections: Implications of Pro-Incumbent Bias in Close U.S. House Races

The regression-discontinuity (RD) design has experienced a resurgence of interest in many fields, including political science. Following David Lee’s pioneering work on the incumbency advantage in U.S. House elections, many scholars have begun to apply RD designs to popular elections. Under certain assumptions, the RD design yields causal estimates that approach the gold standard of randomized experiments. There are good reasons to believe, however, that outcomes of narrowly decided U.S. Congressional elections are anything but random. Consistent with this notion, we demonstrate that bare winners and bare losers in U.S. House elections exhibit dramatic differences on key pretreatment covariates. Bare winners have substantially more financial resources than candidates who just lose and are far more likely to belong to the party that won the previous election in the district. Congressional Quarterly’s pre-election race ratings predict the outcomes of even extremely close elections with a high degree of accuracy. Close House elections are so predictable that it is impossible to achieve covariate balance between matched treated and control observations. We show that post-election recounts do not change election results often enough to account for the large imbalances in pre-election covariates that we observe. We conclude that bare winners and bare losers in U.S. House elections are not exchangeable ex ante. Therefore, the RD design is not valid in the case of U.S. House elections, and its application to other kinds of electoral data requires careful justification of the identifying assumptions. We discuss the practical advantages and drawbacks of RD designs and offer advice on the implementation and interpretation of such designs. We thank David Lee for generously providing replication files for Lee (2008). We also thank Scott Adler, David Brady, Gary Jacobson, Keith Poole, and Jonathan Wand for sharing their data with us. Henry Brady, Andy Eggers, Luke Keele, Winston Lin, Walter Mebane, Jr., Eric Schickler, and Laura Stoker gave useful feedback, and Peter Ryan helped shape the project in its formative stages. Willa Caughey, Mona Fang, and Sarah Weiner provided valuable research assistance. ŽPh.D. candidate, Travers Department of Political Science, UC Berkeley, http://devincaughey.com Associate Professor, Travers Department of Political Science and Director, Center for Causal Inference and Program Evaluation, Institute of Governmental Studies, UC Berkeley, http://sekhon.berkeley.edu