Deep exploratory regression modelling of survey data. With applications to electoral survey data of the 2014 elections in Belgium.

• Main Author: Henry, Lionel
• Other Authors: Van Haute, Emilie
• Format: Doctoral Thesis
• Language: en
• Published: Universite Libre de Bruxelles 2020
• Subjects: Science politique générale; Econométrie et méthodes statistiques :théorie et applications; regression modelling; research methods; survey data; exploratory; bayesian
• Online Access: https://dipot.ulb.ac.be/dspace/bitstream/2013/308280/5/Contrat_Henry_Lionel.pdf; https://dipot.ulb.ac.be/dspace/bitstream/2013/308280/3/manuscript.pdf; https://dipot.ulb.ac.be/dspace/bitstream/2013/308280/4/TOC.pdf; http://hdl.handle.net/2013/ULB-DIPOT:oai:dipot.ulb.ac.be:2013/308280

This thesis contributes practical and conceptual tools for discovering and understanding the variation of quantitative patterns in social and political survey data. It uses regression modelling as an exploratory method with a focus on deep rather than wide model specifications, i.e. on interaction terms rather than control variables. Our main research question is how can we learn from survey data with an exploratory approach of regression modelling. We also seek to answer two more specific questions, what sort of quantitative variations should an exploratory approach seek to model, and how do we deal with statistical uncertainty within an exploratory approach. Our work shows how to use regression modelling for exploratory purposes by interpreting the results descriptively, and connecting these summaries to theory through an act of interpretation. Using data from the Partirep electoral survey of the 2014 elections in Belgium, we illustrate how the emphasis on group variations and interactions has both empirical and theoretical value. We propose to summarise the results of exploratory modelling in a notebook containing a series of increasingly disaggregated prediction graphs. These notebooks help researchers to increase their domain numeracy, i.e. develop a quantitative understanding of the patterns in the data. Regarding statistical uncertainty, we mitigate the risks of modelling sampling noise by using standard errors of binned averages as precision hints that serve as an indication of excessive disaggregation. We also lay out the path for regularising the estimates of the final results with Bayesian models by exploring methods of including the sampling weights in these models. === Doctorat en Sciences politiques et sociales === info:eu-repo/semantics/nonPublished