The rhetoric in Washington, D.C., of late has been rather monotonic. The political landscape seems dominated by the economy, from stimulus bills and consumer protection agencies to tea parties and anti-tax rallies. Yet despite Capitol Hill’s singular focus, new research suggests the left and the right are still separated by a cultural divide.
Political scientists have long questioned the influence of culture on politics in America. In their eyes, the difference between liberals and conservatives boiled down to one thing—the economy. That prevailing theory has been driven by analyses of voting behavior, a relatively new approach that has revolutionized the field. But even counting votes can fall short. In votes, experts hear only a whisper of a congressman’s ideology. His or her true orientation is often muddled by party strategies, ornery constituencies, and other political necessities.
So political scientists have turned to an old tool: speech analysis. This time, though, they wield it with a modern grip. Previous analyses of congressional speeches relied on manually classifying the content, a time-consuming and laborious process. The approach used in the new paper supplants man-hours with processor time, using intelligent computer programs to distill hundreds of hours of transcribed speeches into clarified political ideologies.
Perhaps more interesting than guessing a senator’s ideology were the words the algorithm used to distinguish between liberal and conservative members.
The Cultural Divide
The discovery that cultural issues drive politics in the United States may not come as a surprise to the casual observer, “but within the area of people that study it systematically, it’s not the mainstream view,” says Daniel Diermeier, a professor of managerial economics and decision sciences at the Kellogg School of Management. “I think the popular press has been arguing for a while that value politics has been important, but the research in political science to a large extent didn’t support that.”
The added depth provided by speeches—Senate speeches in particular—over vote tallies is partly responsible for the study’s findings. Senate speeches are longer and less encumbered by rules than speeches in the House of Representatives, allowing its members to express their ideologies more freely. “What’s nice about Senate speeches is one day they talk about foreign policy, the next day they talk about social security, the next day they talk about stem cell research,” Diermeier says. “So if we find similarities across those or consistency, that’s pretty strong evidence that there’s an underlying ideological dimension.”
Diermeier and his colleagues Jean-François Godbout, an assistant professor at the Université de Montréal; Bei Yu, a computer scientist at Syracuse University; and Stefan Kaufmann, a computational linguist at Northwestern University, used a form of machine learning called support vector machines to identify the frequency with which words appeared in transcribed Senate speeches. The algorithm was trained on the speeches of the twenty-five most liberal and twenty-five most conservative senators from the 101st through 107th Congresses, as determined through voting behavior. Then, it was applied to the speeches of the twenty-five most liberal and twenty-five most conservative senators of the 108th Congress, forty-five of whom were the same as in the 107th Congress. The algorithm correctly identified the positions of the fifty extreme senators of the 108th Congress with 92 percent accuracy.
When the algorithm was pointed at the speeches of the fifty moderate senators, its accuracy plummeted, which ironically was what the researchers had hoped would happen. Ideologies, they posited, should be clearly demarcated at the extremes but fuzzy in the middle. The algorithm’s uncertainty appeared to confirm this result.
Perhaps more interesting than guessing a senator’s ideology—a difficult task for a computer program—were the words the algorithm used to distinguish between liberal and conservative members. Conservatives frequently used words with cultural significance, such as marriage, cloning, unborn, homosexual, and abortion. Liberals, on the other hand, focused mostly on local environmental and economic issues, using terms such as ethanol, hydrogen, and arctic in reference to the Arctic National Wildlife Refuge. The cultural terms liberals used only related to gun control, such as handgun and gun. Words relating to economic issues appeared in both parties’ speeches, but they were neither prevalent nor frequently used.
Culturally significant words popped up again and again, despite the variation in the Senate’s legislative agenda. “Even though the topics that Senators talk about vary a lot, there’s still enough similarity that you can pick up,” Diermeier says. He likens the algorithm to Netflix’s movie recommendation engine. With enough speeches to analyze, “we get a glimpse into a Senator’s ideology.”
The research also confirmed a well-known finding in the political science community, that the continuum of political ideologies can be mapped along a single line. Political ideologies in the U.S. Senate are “low dimensional,” in other words. For example, imagine a conservative senator who supports low taxes and the death penalty. It is unlikely that they support a woman’s right to choose, but highly likely that they are against gay marriage. A senator who supports both low taxes and gay marriage is “a completely plausible position,” Diermeier says, “but there are very few people who have it.”
“The question is, is that an accident? Is that a consequence of institutions?” Diermeier asks. Some experts have suggested the two-party system is to blame, that it forces politicians to pick one side or another to distinguish themselves from their opponent. Diermeier suspects otherwise, that there is an historical or even cognitive force behind America’s rigid political boundaries. Still, he cautions that he and his colleagues do not have enough information to confirm that claim. Their study was limited by the availability of digitized data. As a result, they were not able to observe substantial changes in ideologies over time.
Getting a handle on if and how ideology has changed over time in the U.S. Congress is on the researchers’ to-do list. They also intend to run candidates’ campaign websites through the algorithm. Eventually, they hope their quantitative approach will allow political scientists to objectively compare the prevailing ideologies of different countries, not just political parties within one nation.
Related reading on Kellogg Insight: