Development and Assessment of a Social Media–Based Construct of Firearm Ownership: Computational Derivation and Benchmark Comparison

Abstract

Background: Gun violence research is characterized by a dearth of data available for measuring key constructs. Social media data may offer a potential opportunity to significantly reduce that gap, but developing methods for deriving firearms-related constructs from social media data and understanding the measurement properties of such constructs are critical precursors to their broader use. Objective: This study aimed to develop a machine learning model of individual-level firearm ownership from social media data and assess the criterion validity of a state-level construct of ownership. Methods: We used survey responses to questions on firearm ownership linked with Twitter data to construct different machine learning models of firearm ownership. We externally validated these models using a set of firearm-related tweets hand-curated from the Twitter Streaming application programming interface and created state-level ownership estimates using a sample of users collected from the Twitter Decahose application programming interface. We assessed the criterion validity of state-level estimates by comparing their geographic variance to benchmark measures from the RAND State-Level Firearm Ownership Database. Results: We found that the logistic regression classifier for gun ownership performs the best with an accuracy of 0.7 and an F1-score of 0.69. We also found a strong positive correlation between Twitter-based estimates of gun ownership and benchmark ownership estimates. For states meeting a threshold requirement of a minimum of 100 labeled Twitter users, the Pearson and Spearman correlation coefficients are 0.63 (P