Reassessing Models for the Study of Split-ticket Voting

Burden and Kimball (1998) report that, by using the King estimation procedure for inferring individual-level behavior from aggregate data, they are the first to produce accurate estimates of split-ticket voting rates in congressional districts. There are, however, several reasons to doubt that the King estimation procedure does, in fact, produce accurate cross-level inferences. We show that the estimation procedure is highly suspect in general and especially unhelpful with these data. Hence, the results that Burden and Kimball produce are very likely to have been spurious, and their substantive conclusions about split-ticket voting should be regarded as dubious. In a recent issue of this journal, Burden and Kimball purport to produce, for the first time, accurate estimates of ticket-splitting at the district level. They employ an ecological inference method proposed by King (1997), and their analysis of the estimates obtained from this method leads them to conclude that, contrary to some previous findings in the split-ticket voting literature, “voters are not intentionally splitting their tickets to produce divided government and moderate politics” (Burden and Kimball 1998). Instead, they claim, ticket-splitting is solely the result of lopsided congressional campaigns in which well-funded, high-quality candidates run against unknown, underfunded challengers. In this article, we catalog six reasons to doubt the generality and veracity of their claims. First, they greatly simplify the ticket-splitting phenomenon: by limiting their focus to only three offices (U.S. Representatives, Senators, and President); by ignoring all votes not cast for major-party candidates; and by imposing questionable assumptions for the purposes of tractability. Second, even if these assumptions could all be justified, King’s own checklist for determining which data are amenable to his estimation procedure suggest that these election data are particularly ill-suited in the sense that they are fundamentally uninformative. Third, were the assumptions justified and the data well-suited for this estimator, there would remain severe problems with the King estimation procedure (hereafter, for convenience, referred to as “EI”) used by Burden and Kimball. In particular, the EI program is sufficiently unstable to produce significantly different results on separate runs with the same data. Fourth, even if the EI program operated flawlessly, the underlying model is inappropriate for analysis of the Burden and Kimball data set because its assumptions are violated by the data. Fifth, the principal substantive claim in their article follows from an OLS model