Machine learning improves our understanding of how laws are made and who deserves credit for them
In the early years of his long congressional tenure, Senator Warren Magnuson (D-WA) was known as an effective purveyor of pork. Following a tough reelection campaign in 1962 and in light of changing times, his staff encouraged him to expand his portfolio. As Commerce Committee Chair, Magnuson and his staff searched for an issue to champion. A bill (S. 2231) introduced by a junior senator on the committee, Abraham Ribicoff (D-CT), proposed significantly strengthening the federal government’s authority to regulate motor vehicle safety. Ribicoff’s bill went nowhere in the committee (despite 12 Senate cosponsors). However, early the next year Magnuson introduced his own traffic safety bill (S. 3005) that incorporated much of the language of Ribicoff’s proposal, and used his agenda setting powers to shepherd it into law (the National Traffic and Motor Vehicle Act of 1966).
This vignette highlights a well understood fact about lawmaking. A proposal that originates with one bill can become law as a provision of another bill. However, scholars and journalists typically study effectiveness by equating the progress of bills with the progress of policy ideas. This approach has two potential flaws. The first is that bills sometimes advance for reasons that have little to do with the sponsor. For example, Congress must pass defense authorization bills annually. By convention, these ‘must pass’ bills are sponsored by the leaders of the relevant committees or subcommittees. The second potential flaw is that the lawmaking contributions of other legislators are completely overlooked. Defense authorizations, for example, frequently incorporate the substance of many other bills. By current standards, the sponsors of these incorporated bills are less effective.
We begin to address these limitations of current approaches by using “text reuse” methods to study the progress of policy ideas. The basic intuition behind the method is plagiarism detection. In some cases, legislators do borrow other legislators’ ideas with or without their consent. But bill text reuse also reflects institutional realities and strategic considerations. Opportunities to pass stand-alone bills are limited by agenda space and strategic partisan considerations. Out of 10,000 bills introduced in a typical Congress, only about 300 will become law. Moreover, the legislators who set the agenda (majority party members and committee leaders) tend to hoard these valuable credit claiming opportunities. Another constraint is the Constitutional requirement that all bills raising revenue originate in the House. Practically speaking, this means that no Senate bill with revenue implications can become law.
The only way to enact a Senate revenue bill is to incorporate it as a “hitchhiker” into a House bill. For many other legislators, the best option for enacting their policy proposals is to include them as hitchhikers on majority party or leader-sponsored bills. Importantly, because hitchhiker bills are much less visible, they are probably more likely to be judged on their policy merits than on their partisan political implications.
The Patient Protection and Affordable Care Act (ACA) is 900 pages long. As introduced, the bill that eventually became law (H.R. 3590) was just seven pages long and proposed mortgage assistance for service personnel. It would clearly be a mistake to give the bill’s original sponsor (Charles Rangel (D-NY)) credit for the passage of this landmark legislation. In “Tracing the Flow of Policy Ideas: A Text Reuse Approach,” we employ text reuse methods to study the lineage of the ACA’s many provisions. We do this by comparing the textual similarity of each section of the final law to each section of every other bill that did not become law in the same Congress. We detect hundreds of such connections. We also discover that the ACA includes many provisions first proposed by Republicans even though the Act itself did not attract a single Republican vote. One of its major provisions (Nursing Home Transparency) that we traced to a Republican-sponsored bill was a product of years of bipartisan collaboration at the committee level. The ACA presented a window of opportunity to finally enact this legislation.
Hitchhiking appears to be a normal, if largely overlooked, lawmaking practice. In our forthcoming AJPS article, we identify when complete bills become law as insertions in other bills across two decades of lawmaking. Here we confirm that ‘hitchhiker’ bills are not occasional exceptions to the general pattern. About half of all bills that become law become law as hitchhikers. Sponsors of successful hitchhikers are also different from sponsors of successful stand-alone bills. Twice as many members can claim at least one legislative success, and they are much more likely to be non-leaders and members of the minority party. Not surprisingly, Senate bills are much more likely to become law as hitchhikers. Indeed, more Senate bills become law as hitchhikers on House bills than become law on their own.
Scholars and journalists acknowledge that bill success is an imperfect measure of member effectiveness. But the repeated emphasis on bill progress still leaves the impression that advancing stand-alone bills is how things get done. Our research to date only considers complete bill hitchhikers from the same Congress and we find that they account for about half of all enactments. We have yet to consider cases of cross-Congress hitchhikers and cases where only parts of bills are enacted as a hitchhikers. Clearly, they are an important part of lawmaking that deserve more attention.
None of these preliminary findings should be surprising to legislative scholars and congressional insiders. What is different and exciting is how machine learning methods create opportunities to study longstanding subjects like effectiveness and coalition building in new and improved ways. And as we think we have shown, new approaches can lead to new understandings.
Andreu Casas is a Moore Sloan research fellow in the Center for Data Science at New York University, Matthew J. Denny is a Ph.D. candidate in Political Science and Social Data Analytics at Penn State, and John Wilkerson is a professor of Political Science at the University of Washington.