The lifetime tenure of each of the U.S. Supreme Court’s nine justices means they have long-lasting influences on the country’s affairs. How a justice votes in a case is a reflection of his or her judicial temperament, personal philosophy and political ideology.
Political scientists, lawyers, constitutional scholars and laymen alike closely watch the court. They often attempt to divine what happens during the justices’ secret deliberation sessions and during the process of drafting the opinions that will ultimately be released to the public.
As public attention focuses on the court in the wake of the February death of Associate Justice Antonin Scalia, we at the Discovery Analytics Center at Virginia Tech offer a way for artificial intelligence to provide some insight. In a recently published paper, we propose a way to model the Supreme Court using computer-based machine learning.
We devised a data-driven framework that learns justices’ judicial preferences and voting behavior; it can be used to answer questions about the justices’ voting behavior. For example: how will the justices vote in a case about a particular topic? Which coalitions are likely to become relevant? Who is likely to be the swing justice in a given case or term?
Our model was able to create a reasonably accurate assessment of justices’ views on issues, predict their alignments on cases and identify who might be a swing vote.
Figuring out where the justices stand
Justices take initial stances toward a case by reading prehearing filings. Then they explore issues during a hearing, make decisions in a posthearing meeting among the justices and write their opinions thereafter.
These opinions represent different groups – a unanimous court, a majority of the court, dissents from with the majority decision and even opinions that concur in part and dissent in part. Each opinion can be written by one or more justices, or written by one and joined by others.
As a result, it is challenging to determine the justices’ actual positions on the issues that come before the court. In the past, methods have been devised to assign each justice a numeric score called an “ideal point,” indicating his or her hypothetical position on a spectrum of views on an issue, using news coverage of their nominations or their voting records on cases. However, these approaches do not use the rich text resources available in the decisions.
Our model, called the Supreme Court Ideal Point Miner (SCIPM), augments the existing research with information on judicial preferences gleaned from opinion texts.
Analyzing the text
The model assumes that each case involves a mixture of several issues or topics, on which justices have distinct views. For example, a case may involve a search that uncovered an illegally owned firearm; any individual justice’s views on the Fourth and Second Amendments may come into conflict with each other. The final opinion that justice writes or joins will reflect aspects of both positions.
Our system looks at the text of the opinions and counts the number of words related to each issue that factors into the decision. It assigns a relative weight to each such issue based on the share of relevant words. For example, words like “search” and “warrant” can suggest a connection to the Fourth Amendment.
By aggregating the analyses of multiple opinions – our research looked at those issued from 2010 to 2014 – the model can tell how strongly each justice feels about an issue.
But it is not capable of determining which direction those feelings go. For that we need to look at the justices’ voting records directly.
Our analysis shows the range of views held by justices on selected issues, as shown below. They range across a spectrum from very liberal (far left) to very conservative (far right).
Finding swing justices
Our model can also identify the swing justices of the court by looking at how much each justice’s ideal point varies across multiple issues. A justice who is consistently far to one side is less likely to swing than a justice whose votes range across the spectrum.
The model identifies Associate Justice Anthony Kennedy as the justice most likely to swing, followed by Chief Justice John Roberts Jr., because their positions vary more widely across issues than their colleagues. Roberts’ ranking as the second swing justice is also explainable given his (surprising) votes to affirm the Patient Protection and Affordable Care Act (“Obamacare”).
This can be validated by observing all cases decided by a margin of 5-4 – the cases where swing justices come most into play. We look at which justices have grouped themselves into those five-vote majorities. The three most frequent coalitions are shown below.
Looking at those coalitions, it is clear that, apart from top swing voter Kennedy, justices appointed by Democratic and Republican presidents tend to cluster together in split decisions.
Looking at individual justices and cases
An interesting finding of our experiment is the role of Associate Justice Sonia Sotomayor in cases related to the Fourth Amendment, which protects citizens against unreasonable search and seizure. She has a very different ideology on this particular issue compared to her colleagues. This observation can be validated with other articles that report her disagreement with her colleagues on this particular issue.
What we learn from the model can be used both to describe the decision process for an individual case and also to make predictions about future cases. Its accuracy is 79.46 percent, which we calculate by dividing into five sections the case history from 2010 to 2014 – where decisions and opinion arguments are available. We teach the model from four of the five parts, and evaluate its predictions on the fifth part against what actually happened.
Our model also reflects the reality that some justices are more frequent collaborators than others, even within their respective political wings. For example, Justice Roberts and Justice Alito generally collaborate more frequently than others.
Future research we have planned includes evaluating public response to decisions, such as commentary on social media, to learn what types of cases people are most attracted to; expanding the data in the model to cover more decisions over more time; and using text transcripts of the oral arguments to attempt to predict outcomes of cases that have not yet been decided.