Using digital trace data to study public sentiment toward the police. A demonstration case on the George Floyd killing

In this paper, we assess a measurement approach to study sentiment toward the police using large-scale digital trace data and natural language processing (NLP). Existing research on the consequences of high-profile cases of police violence relies almost exclusively on surveys, which are often constrained by fieldwork timing, selective nonresponse, and limited temporal resolution. We test an alternative, easy-to apply, fully observational measurement strategy that uses YouTube Data Tools to collect millions of unsolicited public comments and applies a refined dictionary-based sentiment algorithm incorporating valence shifters to estimate attitudes expressed in police-related discourse. Using the killing of George Floyd as a demonstration case, we show how this approach yields high-frequency attitudinal indicators that closely mirror known temporal patterns from survey-based studies, including the immediate but temporary downturn in sentiment toward the police. Our analysis suggests that process-generated text from YouTube video comments can meaningfully broaden the methodological toolkit available to policing and criminology researchers. However, we stress that survey items and NLP-based sentiment scores capture fundamentally different constructs, and while the latter may be useful in some cases (e.g. facilitating rapid assessments of attitudinal dynamics), they are not a substitute for survey data.