Supervised Learning of Candidate Positions Across Issues with Itemized Donations

• In Event: Elections and Campaign Spending in Legislative Politics

Abstract

Even though a series of Supreme Court decisions has been undoing major components of campaign finance regulation in the United States for the past decade, the Court has consistently upheld disclosure of itemized donations even through watershed moments such as Citizens United v. FEC. A key assumption underlying the Court’s support for disclosure, which has not been systematically examined to our best knowledge, is that itemized donations could be an important source of information for voters to ascertain whether candidates are beholden to special interests, and to help voters place candidates on the political spectrum “more precisely than is often possible solely on the basis of party labels and campaign speeches” (Buckley v. Valeo). Rigorously assessing this informational assumption is important not only for informing ongoing legal debates regarding disclosure (e.g., CREW v. FEC), but also for enhancing our understanding of the informational role that campaign finance plays in democratic accountability in the United States.

To this end, we seek to quantify the extent to which itemized donations predict candidates’ positions across policy domains (e.g., immigration). Specifically, we will train supervised machine learning models on all itemized donations made to candidates--including the identity of the donors as well as the amounts given--to predict how candidates vote on a given policy domain as measured by interest group ratings and legislative scorecards.

Our approach allows for three contributions. First, by examining how well our model predicts congressional voting across issues relative to predictions based solely on party labels or single-dimension ideology, we can directly evaluate the Court’s assumption on the informational value of campaign finance disclosure. For example, consider Justin Amash and Jason Chaffetz. Both are current or former Republican members of the House, and have identical DW-NOMINATE scores for the last legislative session in which Chaffetz was in office. However, Amash is a vocal critic of the intelligence and surveillance industry while Chaffetz chaired the Subcommittee on National Security. Even though neither party label nor single-dimensional ideal point distinguishes Amash from Chaffetz, we can detect their disagreement over this particular policy domain by the fact that defense industry donors such as Northrop Grumman contributed only to Chaffetz and not to Amash.

By quantifying the gains in predictive accuracy based on our model relative to party or ideology-based benchmarks, we also engage with two noted patterns of representation in the United States. The extent to which such predictive gains have changed since the 1980s (the earliest period for which we observe itemized donations), we could shed light on how much intraparty variation in candidates positions across different issues may have narrowed as American politics became increasingly polarized (e.g., McCarty, Poole, and Rosenthal 2006). In addition, since voters need not hold ideologically consistent views across different issues (e.g., Broockman 2016), voters may care especially about whether a given candidate departs from the standard position adopted by her party or ideological camp on specific issues of interest--information that voters may glean from itemized donations.

Our second contribution is methodological: due to the dense network linking donors and recipients of itemized contributions, our model can predict issue-specific positions for a much wider set of candidates than is currently feasible. Existing measures of positions across policy domains, such as interest group ratings or issue-specific NOMINATE scores, largely depend on roll call records which are not available for candidates who haven’t served in office. But because even first-time challengers often share itemized donors with incumbents, our donations-based supervised model can predict these challengers’ positions across issues before they accumulate any legislative records. In other words, our donations-based supervised model could help to fill the informational gap voters face in assessing new faces in politics.

Last but not least, our methodological approach allows us to identify specific individual or institutional donors (e.g., the NRA) that strongly signal candidates’ positions on a particular issue (e.g., gun rights), as well as the minimum number of itemized donations needed to place a candidate on a given policy domain with reasonable accuracy. Following Kosinski, Stillwell, and Graepel (2013), these metrics could allow for more interpretable and accessible voter guides using our itemized donations-based model of candidates’ issue-specific positions.

Authors

• Adam Bonica, Stanford University

• Zhao Li, Princeton University