Understanding Public Attitudes Towards Immigration Through Online Social Media

Abstract

Immigration is one of the most politically salient and emotionally charged issues in contemporary public life, shaping debates about national identity, belonging, economic change and social order. Social media platforms have become central arenas in which these debates unfold, allowing immigration-related claims, fears, support and opposition to circulate rapidly and at scale. At the same time, they offer new opportunities to observe public attitudes towards immigration with a level of temporal and spatial detail that traditional methods such as surveys cannot achieve. Existing research has produced important insights into immigration attitudes, yet evidence remains fragmented and a series of deficiencies persist across three distinct areas. First, the structure of online immigration debate is poorly understood, including the extent of polarisation, the role of key users and the speed at which pro-immigration and anti-immigration content spreads. Second, although large language models (LLMs) offer powerful tools for analysing text at scale, their ability to identify immigration content and stance accurately, efficiently and across multiple languages remains insufficiently tested. Third, existing work has rarely examined how multiple local conditions combine to shape anti-immigration sentiment and immigration salience at a fine spatial scale and over long periods of time. This thesis aims to systematically address these deficiencies and to advance our understanding of how attitudes towards immigration are expressed, measured and geographically distributed online and which local factors are associated with them. To achieve this, I developed an integrated computational framework implemented across three empirical analyses. The first applied network analysis and diffusion measures to 220,870 immigration tweets in the United Kingdom between December 2019 and April 2020, examining polarisation, key producers and spreaders, bot activity and the speed of circulation. The second developed open-source fine-tuned LLMs to classify immigration topic and stance across 13 languages, testing whether they generalise to languages unseen during fine-tuning while remaining efficient, low-cost and reproducible. The third applied this pipeline to 2.6 billion geolocated tweets from the United States between 2013 and 2023, producing county-year measures of anti-immigration sentiment and immigration salience and modelling them against demographic, economic, social, political, health, spatial and historical conditions. Collectively, the findings make contributions to substantive knowledge, theory, methods, policy and data in the study of immigration attitudes. They show that online debate is polarised and uneven, that a small number of users produce and amplify a disproportionate share of anti-immigration content, which spreads faster than pro-immigration content, and that anti-immigration sentiment in the United States is geographically heterogeneous and emerges from the way multiple local conditions combine, including the size and country of origin of migrant groups, economic contraction, population decline, social capital, institutional trust and historical legacies, rather than from migrant numbers alone. The central theoretical contribution is ‘Conjunctural Immigration Theory’, which holds that anti-immigration sentiment emerges from the combination of material and non-material conditions present in a place at a given moment, so that the same material and non-material change generates anti-immigration sentiment in one local context while leaving another untouched. Methodologically, the thesis delivers an open-source and lightweight multilingual LLM for topic and stance classification at scale, and it provides new county-year estimates of immigration salience and stance for the United States. For policy, it points to the value of monitoring how rapidly anti-immigration content spreads online and of place-based responses that act on the material and non-material conditions of local areas together.