In 1948, a Chicago Tribune front-page headline trumpeted “DEWEY DEFEATS TRUMAN,” the bold capital letters incorrectly announcing the defeated challenger as the presidential victor. Many newspapers and broadcasters committed the same gaffe in 2000, when a tight race between George W. Bush and Al Gore see-sawed throughout the night of November 7.
These two incidents show the pitfalls of political prognostication. Anticipating which box voters will check on Election Day and understanding how their attitudes and preferences have been shaped are complex challenges. Political scientists Sung-youn Kim, a visiting assistant professor at the University of Iowa, and Milton Lodge and Charles Taber, professors at Stony Brook University in New York, have developed a novel approach to these questions, creating a computational model that simulates the vicissitudes of political opinion. Recent calculations using TeraGrid resources at NCSA and the San Diego Supercomputer Center (SDSC) demonstrated that their model returns results that accord well with actual polling data. Their research was presented at both the 2004 Midwest Political Science Association Conference and the 2004 American Political Science Association Conference.
How do people assess candidates? How do campaign events and new information change their views? While there are various theories that address these questions, Kim (then a doctoral student at Stony Brook), Lodge, and Taber saw gaps between existing models and empirical findings. Several theories seemed to partially explain actual political behavior and judgment, but none seemed complete on its own.
“There are two classes of empirical evidence and theories that our model is built upon: the classic cognitive paradigm and the theories of political information processing, including the on-line processing and memory-based processing models,” Kim explains. “The model is built by integrating and incorporating what these theories postulate.”
Both the on-line and memory-based processing theories–theories regarding how people evaluate political objects, including candidates–were then integrated into the model as its judgment mechanism. The on-line processing theory asserts that the affective summary evaluation (or valence) linked to every object in memory is updated continuously whenever an individual is exposed to new information; the individual maintains a running evaluative tally for each object. Although the original information that entered into the judgment process may be forgotten, the evaluative tallies are immediately accessible. In other words, on seeing or hearing the name “George W. Bush” a person will immediately know how she feels about the candidate, even though she might not be able to say why she feels as she does.
The memory-based processing theory holds that different, often conflicting, considerations and feelings that come to mind at a particular moment influence the evaluation of an object. The accessibility of these concepts in memory determines what comes to mind and thereby influences how those concepts influence evaluation of the object. When a person is asked for his evaluation of Al Gore, for example, his answer will depend in part on which of the many facts about Gore held in his long-term memory are uppermost in his mind at the time.
Kim, Lodge, and Taber developed six algorithms to represent their amalgam of theories in a computational model. They cleverly dubbed this integrated computational model John Q. Public.
Creating virtual voters
To put the model to the test, they needed to construct virtual voters, or agents, each with a unique initial mindset that could react to campaign information according to the theories integrated into the model.
The agents were provided with the information about general concepts (honesty, determination, right, wrong, etc.) from a standard data set and with information about political objects (parties, candidates, issues, etc.) gathered from the pre-election survey of the 2000 National Annenberg Election Survey (NAES), which was conducted from November 1999 to January 2001 and was at the time the largest-ever survey of the American electorate.
The tens of thousands of NAES respondents were asked to place themselves in one of five ideological categories: strong conservative (7 percent), conservative (29 percent), moderate (41 percent), liberal (19 percent), or strong liberal (4 percent). Kim examined the responses given by members of each self-identified group before the pivotal Republican National Convention, calculating the mean and standard deviation for each group's attitudes toward the candidates, parties, issues, and their perceptions of the candidates' traits and stances on the issues. From these group profiles, Kim then generated 100 knowledge structures with random variations, each one representing the knowledge and attitudes held in long-term memory by a simulated survey respondent.
These weblike associative semantic networks use nodes to represent concepts or objects (such as a candidate, an issue, etc.), with the links between nodes standing in for the associative strengths (implicational beliefs) between them. Attitudes are represented by positive or negative valences attached to each node.
Processing campaign messages
The population of agents was now ready for the simulated 2000 presidential campaign to unfold. Real-world voters, of course, are inundated with political messages from newspapers, television news broadcasts, billboards, radio and TV ads, talk radio, televised debates, and personal conversations. The researchers needed to limit the information processed by the agents, however, so all campaign information was obtained from two New York newspapers, Newsday and The New York Times. News reports on five key campaign events were boiled down to simple sentences (“Bush said Gore is dishonest” and “Bush said Bush is anti-abortion,” for example).
The computational model parses each sentence, retrieves relevant concepts from the long-term memory of each agent, and updates each agent's knowledge and attitudes accordingly. For example, an ideologically conservative agent begins with a positive evaluation of George W. Bush and of tax cuts. A news report on the GOP convention indicating that “Bush said Bush supports tax cut” would be “read” by the agent, which would pull its knowledge and evaluation of the object “George W. Bush” and the concept “tax cut” from its long-term memory. The positive evaluation of the tax-cut concept influences the agent's evaluation of Bush. The agent's knowledge map is updated with this new evaluation of Bush, guiding the agent's response to subsequent survey questions.
For the simulation, the 100 agents first responded to the NAES 2000 survey questions, providing a baseline akin to the responses recorded before the GOP convention. Campaign information was gleaned from newspaper reports after the GOP and Democratic conventions and after the three presidential debates. Model surveys were conducted after each event to gauge the agents' reactions. This simulation was repeated 100 times due to the model's stochastic components, which create randomized variations in how readily each agent accesses information, how strong associations are between objects in long-term memory, and how each agent interprets information.
Because the simulation employed multiple independent agents, with each one representing a single voter, each agent had to be loaded in a separate thread. Because of this complexity and the sheer computational intensity of the simulation, the researchers relied on the computational power of the TeraGrid, employing Itanium 2 systems at both NCSA and SDSC. Using 10 processors, the simulation took about five hours and produced 10 GB of output.
Comparing simulation to survey
The researchers averaged the results of the 100 simulations they conducted of the changing candidate evaluations of the 100 virtual voters. They compared these results to the fluctuations recorded by the 2000 NAES, finding a strong correlation of about .80. They also compared the simulated and actual evaluations across the five ideological groups to determine how well the simulation matched the actual distribution of candidate evaluations across the five groups, finding a correlation of .98.
There were some discrepancies between the simulation and reality, however. The simulated changes tended to be less pronounced than the fluctuations in the survey data, and in some cases the direction of change was different (for example, while actual moderate voters had a more positive evaluation of Gore after the GOP convention, the simulation showed that moderates had a more negative opinion of Gore at that time). Kim believes that incorporating additional information-processing theories will refine the model and improve the correlation between simulation and survey results.
Now that the simulation's accuracy has been demonstrated, it can be used as a platform to develop and test hypotheses. By incorporating new parameters, researchers can see how these changes affect candidate evaluations. As the model is refined further, it could even be used by political strategists to predict how the electorate will respond to information and how framing an issue or candidate can sway public opinion.
Source: Access Online. Reprinted with permission of National Center for Supercomputing Applications