Thread by @chrishanretty

Tonight I'll be presenting my #APSA2020 paper, "The voting power of demographic groups". You can find the paper at drive.google.com/file/d/1G-SHiy… and a video presentation at (1/n)

The idea behind the paper is simple: the power of a voter group is the number of seats where the result would have been different had that voter group not voted—or alternately, the number of seats in which that voter group was pivotal (2/n)

The implementation is also simple, in a sense, and relies on multilevel regression and post-stratification (MRP) plus some post-hoc adjustments (3/n)

The first stage in doing MRP is to estimate a model which links voter (and area) characteristics to predicted probabilities of voting in different ways. Once you've got that, all that's left is to add up rows of data in a post-stratification frame (4/n)

So why can't we add up those rows in different ways to get counterfactual outcomes? That's what my paper does (5/n)

Here are pivotality scores for level 1 voter groups (grps defined by a single demographic attr) in
#GE2010. Ethnic minority voters (f.ex.) were pivotal in 25 seats, & more pivotal on a pop'n adj. basis than white voters, largely because they're fairly compact behind Labour (6/n)

Pivotality is different from being a swing voter. There's no relationship between how pivotal a group is and how "available" it is. F.ex., older voter grps can be pivotal, but they're generally not "available" because they've built up a habit of voting in a particular way. (7/n)

Surprisingly, group turnout isn't a predictor of how pivotal a group will be. It's how homogeneous the group is, and whether or not it supports parties which win seats ("expensive tastes", in my language) (8/n)

More details in the paper, and thoughts and comments welcome. The idea seems obvious in retrospect, so I'm worried someone else has already done it (tho I realize ppl have already produced estimates of pivotality of particular grps) (9/n)

Thanks go to the @LeverhulmeTrust for giving me money to do research like this (10/10)