Skip to content
Bio and MLData

Predicting propensity to vote

A machine-learning study of voter turnout. arXiv preprint.

Independent · 2021

Table of classifier metrics: sensitivity 0.4737, specificity 0.8372, precision 0.7896, negative predictive value 0.5524, false positive rate 0.1628, false discovery rate 0.2104, false negative rate 0.5263, accuracy 0.6325, F1 score 0.5922 and a Matthews correlation coefficient of 0.3261, each with its derivation.
The full metric set for one of the three experiments. Its Matthews correlation coefficient is 0.3261; the paper's headline figure, across all three, is 0.39.

What it is

An early project, and my first piece of published writing. We trained supervised models to predict whether an individual would turn out to vote, and reported the result using the Matthews correlation coefficient rather than accuracy, because turnout data is imbalanced enough that accuracy flatters a model that has learned very little.

0.39
Matthews correlation coefficient, best of three experiments
3
Experiments run on voting data from 2004 to 2018
2021
Posted as arXiv preprint 2102.01535

Read the preprint on arXiv