Search

In this doctoral thesis, I reflect on methods to leverage non-representative samples to measure and forecast public opinion in the run-up to an election. The thesis is made up of three separate papers, dealing with a number of issues around the topic of convenience samples; these issues include a) how and where to collect non-representative data, and how to enhance them with external, auxiliary sources; b) how to construct reliable and deep stratification weights, when these are not immediately available from the census, and when no single dataset contains an exhaustive set of the desired post-stratification information; c) how to optimise learning models to automate variable selection, produce accurate out-of-sample predictions, and improve fitting speed. Findings from this thesis suggest unrepresentative samples can be used to obtain real-time readings of public opinion, at a fraction of the cost of traditional survey methods, provided a number of non-trivial and computationally demanding model adjustments are implemented to smooth and re-weight the samples.