Estimating vote-specific preferences from roll-call data using conditional autoregressive priors

Ideal point estimation in political science usually aims to reduce a matrix of votes to a small number of preference dimensions. We argue that taking a non-parametric perspective can yield measures that are more useful for some subsequent analyses. We propose a conditional autoregressive preference measurement model, which we use to generate case-specific preference estimates for US Supreme Court justices from 1946 to 2005. We show that the varying relative legal positions taken by justices across areas of law condition the opinion assignment strategy of the Chief Justice and the decisions of all justices as to whether to join the majority opinion. Unlike previous analyses that have made similar claims, using case-specific preference estimates enables us to hold constant the justices involved, providing stronger evidence that justices are strategically responsive to each others' relative positions on a case-by-case basis rather than simply their identities or average relative preferences.