Josh Hudder at Rule 22 claims that polarization in Congress is mostly a measurement phenomenon caused by the mid-1970s rule changes that led to more recorded procedural votes in the House and Senate. The argument goes that procedural votes are more partisan so if we observe more of them, measured partisan polarization would go up.

I have several reasons for being skeptical of this claim, but let me just share one. In the technical language of ideal point estimation, Hudder’s procedural argument is that the frequency of “cutpoints” that divide the parties went up increasing the measured polarization. The problem with Hudder’s argument is that scaling procedures like NOMINATE and Bayesian IRT are fairly insensitive the the distribution of cutpoints. What really matters is that there are at least some cutpoints that separate each pair of adjacent legislators. In other words, suppose we estimated ideal points from a set of roll calls that have cutpoints between each pair of members. If we then added a zillion party line votes, estimated polarization would not change that much (it might reduce some overlap but it would not necessarily move the party means or medians apart).

The right experiment would be estimate polarization holding the distribution of cutpoints constant or vote margins. This experiment is conducted in chapter 2 of my book with Poole and Rosenthal. We found that the counterfactual polarization measures (those holding the margins distribution constant) correlated with the observed measures at greater than .95.

I also refer readers to Alex Hirsch’s recent paper in Political Analysis which also shows the robustness of ideal point estimates to changes in the distribution of cutpoints.