Estimating mass sentiment from noisy, biased signals: recent Australian elections and the Voice referendum

Abstract

I deploy a class of hidden Markov models to recover the trajectory of latent public opinion from a series of noisy and possibly biased signals: public opinion polls. Example applications include recent Australian federal elections and the current Voice referendum campaign. In addition to recovering both level and trend in mass sentiment, incorporating known election results ex post identifies pollster biases. Pollster biases are a durable characteristic of Australian election polling; calibrating model outputs to adjust for these biases produced accurate forecasts of the 2022 Australian Federal election result. Recent Australian federal elections also reveal a persistent pattern of voter sentiment trending towards the Coalition over the campaign. We also discuss extensions of the model, including switching between volatility regimes and step discontinuities in response to shocks such as leadership switches, scandals and gaffes.