Despite scientific consensus on vaccines and anthropogenic climate change, public opinion remains divided, leading to polarized debates often reflected in online social media like Twitter. Twitter's popularity and ease of information sharing make it ideal for studying opinion spread and sentiment in social networks. Echo chambers are common, and studies show Twitter sentiment correlates with real-world behavior. Analyzing Twitter debates on these topics could help understand real-world decision-making, although such research is still developing. The study focuses on comparing these two seemingly disparate topics, noting a shared feature: both herd immunity (vaccines) and reduced greenhouse gas emissions are common pool resources subject to free-rider problems and both represent coupled human-environment systems. This study hypothesizes that both conversations exhibit similar sentiment distributions and network structures.

Existing research highlights the use of Twitter to study sentiment and opinions related to vaccines and climate change. Studies show correlations between Twitter sentiment and real-world vaccination behavior, as well as the use of Twitter to predict and monitor infectious disease outbreaks and market trends. However, there's limited comparative research between these two topics. This study aims to bridge this gap by analyzing large datasets of tweets to compare user sentiment and the structure of user and community networks related to both climate change and vaccines.

The study utilized two datasets purchased from GNIP (now Twitter): one on vaccines (April 1, 2007 – October 15, 2016) and one on climate change (April 4, 2015 – October 15, 2016). Additional datasets were collected using the Twitter API (October 16, 2016 – August 1, 2018). Tweet sentiment was classified using machine learning: a support vector machine (SVM) and an ensemble averaging method comprising RNN-GRU, RNN-LSTM, logistic regression, and GBDT. Training sets were created by human annotation of randomly sampled tweets, ensuring inter-rater reliability. Users were classified as pro, anti, or neutral based on the plurality of sentiments in their tweets. User networks were constructed based on retweets and mentions (directed, undirected, and mutual). Network statistics (connectivity, shortest path length, degree, entropy, assortativity) were computed using NetworkX. Community detection was performed using the Louvain algorithm. Analyses were conducted at increasing scales: tweets, users, user networks, and community networks. Additional analyses included subsampling the vaccine dataset and using an 80% threshold for user sentiment classification to ensure robustness.

The study revealed striking differences between the vaccine and climate change conversations. The majority of climate change tweets were pro-climate, while vaccine tweets were predominantly neutral. This pattern held across different datasets and classification methods. User sentiment mirrored the tweet sentiment distributions. A significant asymmetry existed in participation: more users participated in both the climate and vaccine conversations than vice-versa. Pro-vaccine users overwhelmingly supported anthropogenic climate change, while pro-climate users showed mixed opinions about vaccines. Network analysis revealed higher assortativity in the vaccine conversation: users tended to interact with others sharing their sentiment more often than in the climate change conversation. This higher assortativity was consistent across various network types and user activity levels. Vaccine community networks were larger, more fragmented, with many isolated neutral communities, while climate change communities were more interconnected. The vaccine conversation displayed more isolated communities.

The findings contradict the initial hypothesis that the social networks would be similar. The observed differences may be attributed to the spatial scale of the underlying environmental systems. Infectious diseases, relevant to vaccination, have localized impacts, leading to more isolated online communities. Climate change, on the other hand, has broad spatial effects, forcing individuals with differing sentiments to interact more, potentially facilitating social learning and norm shifts. The higher proportion of neutral sentiment in vaccine tweets might stem from the more immediate and personal nature of infectious disease experiences. The asymmetry in conversation participation could reflect the relative banality of some vaccination-related conversations. The observation that pro-vaccine users were predominantly pro-climate change lends support to hypotheses of a liberal bias against vaccines.

This study highlights unexpected differences in online social network structures surrounding vaccines and climate change. The higher assortativity and fragmented nature of vaccine conversations, contrasted with the more interconnected climate change conversations, suggest the influence of spatial scale in environmental systems on human interaction patterns. Comparative analyses across different fields can offer valuable insights for future research on coupled human-environment systems. Further research could explore the relationship between network structure and social norm shifts, using other coupled human-environment systems.

The study acknowledges limitations such as machine learning classification errors, the simplicity of a three-way sentiment classification, and the fact that online sentiment might not fully represent the general population's views. The lack of socio-demographic data in Twitter limits the ability to fully explore those influences. Future work could address these limitations by integrating data on human mobility and socio-demographic characteristics.