Strong consensus on US Supreme Court spans a century

TL;DR

This paper uses a pairwise maximum entropy model to analyze and simulate the voting behavior of the US Supreme Court over a century, revealing high consensus stability and complex dissent patterns.

Contribution

It introduces a minimal pairwise correlation model that accurately predicts Supreme Court voting, challenging the idea that high-order interactions are necessary.

Findings

Consensus decays very slowly over a century.

Dissenting blocs follow a heavy-tailed Zipf's law.

Minimal pairwise model outperforms more complex models.

Abstract

The US Supreme Court throughout the 20th century has been characterized as being divided between liberals and conservatives, suggesting that justices with similar ideologies would have voted similarly had they overlapped in tenure. What if they had? I build an empirical, quantitative model of this counterfactual hypothesis using pairwise maximum entropy. I infer how 36 justices from 1946-2016 would have all voted on a Super Supreme Court. The model is strikingly consistent with a standard voting model from political science despite using $10^5$ less parameters and fitting the observed statistics better. As with historical courts, the Super Court is dominated by consensus. The rate at which consensus decays as more justices are included is extremely slow, nearly 100 years, and indicates that the modern Supreme Court is an extremely stable institution. Beyond consensus, I discover a rich structure of dissenting blocs that are distributed along a heavy-tailed Zipf's law. The heavy tail means that dominant dissenting modes fail to capture the entire spectrum of dissent. Thus, I find that Supreme Court voting over time is not low-dimensional despite implications to the contrary in historical analysis of Supreme Court voting. Although it has been long presumed that strong higher order correlations are induced by features of the cases, the institution, and the justices, I show that such complexity can be expressed in a minimal model relying only on pairwise correlations. From the perspective of model selection, this minimal model may generalize better and thus be useful for prediction of Supreme Court voting over time.