Machine learning applications to survey nonresponse

Collins, John

Vorschau

PDF
Machine_Learning_Applications_to_Survey_Nonresponse (3).pdf - Veröffentlichte Version
Download (14MB)

URN:	urn:nbn:de:bsz:180-madoc-711190
Dokumenttyp:	Dissertation
Erscheinungsjahr:	2025
Ort der Veröffentlichung:	Mannheim
Hochschule:	University of Mannheim
Gutachter:	Florian Keusch, Frauke Kreuter
Datum der mündl. Prüfung:	2025
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Sozialwissenschaften > Social Data Science and Methodology (Keusch 2022-)
Lizenz:	Creative Commons Namensnennung 4.0 International (CC BY 4.0)
Fachgebiet:	300 Sozialwissenschaften, Soziologie, Anthropologie
Freie Schlagwörter (Englisch):	machine learning , nonresponse bias , adaptive survey design
Abstract:	This dissertation explores how Machine Learning (ML) can help researchers avoid biased inferences due to low response rates in general population surveys. Low response rates can, though do not always, cause these survey studies to fail in one of their primary objectives: making accurate inferences about a population based on a sample. When survey respondents differ systematically from nonrespondents in regards to the topic of the study, the resulting sample will be biased toward the characteristics of those who participate. This effect is called ’nonresponse bias’. Machine learning, a subfield of computer science, focuses on developing algorithms that predict outcomes based on historical data. Given this principle, ML algorithms are a natural choice to learn patterns in survey data and predict individual tendencies to participate, which, as I shall explain, can in turn be leveraged to address nonresponse bias in various ways. The contributions of this dissertation, while varied, follow a common approach: applying ML techniques in novel ways to the challenge of survey nonresponse and demonstrating how survey practitioners can benefit from adopting these innovative methods. Specifically, this work provides survey practitioners with new methods for evaluating the role of past behavior in predicting future nonresponse behavior (Chapter 2), making earlier predictions in newly commenced panel surveys (Chapter 3), enhancing response rates with model-based incentive targeting (Chapter 4), and improving election predictions (which are often confounded by nonresponse bias) by augmenting poll-based models with ML (Chapter 5). Chapters 2-4 are about techniques to ameliorate nonresponse bias, Chapter 5 is about a technique for correcting inferences despite the presence of nonresponse bias.