A solid camp with flowing soldiers: heterogeneous public engagement with science communication on Twitter

The study investigates how heterogeneity in public engagement shapes fluctuations in science communication on social media, focusing on COVID-19-related research dissemination on Twitter. Grounded in the Public Engagement with Science (PES) model, which emphasizes two-way interactions among stakeholders, the authors argue that engagement varies widely: few users engage persistently while many participate briefly around hot topics. They posit that external shocks (e.g., competing events, misinformation) and heterogeneous engagement jointly drive volatility and potential inequities in science communication. The research questions center on whether users with different engagement levels move differently with overall fluctuation, how their positions in communication networks vary, and how their topical preferences shift under changing conditions. The study combines empirical analyses of Altmetric Twitter data with a network-based SIRS model incorporating individual information acceptance thresholds and external shocks to explain observed patterns, particularly the heightened sensitivity of low-engagement users to fluctuations.

The paper situates its work within PES, contrasting it with the deficit model and highlighting critiques that PES often lacks clear outcome measures. Prior studies show scientific information dissemination aids disease prevention and policy compliance, but public engagement is hindered by low interest, trust, and understanding. Attention dynamics are influenced by individual preferences, social influence, and external shocks (Crane & Sornette, 2008). External shocks such as competing events, conspiracy theories, and misinformation shape public attention (Scheufele & Krause, 2019; West & Bergstrom, 2021), with heterogeneous audiences responding differently (Byrne & Hart, 2009) and potentially amplifying disparities (Hart & Nisbet, 2014; Gray et al., 2020; Bonaccorsi et al., 2020). The authors connect this to attention flow and scarcity theories, suggesting science communication outcomes hinge on allocation of limited attention (Davenport & Beck, 2001; Kahneman, 1973; McCombs & Zhu, 1995).

Data: The authors used Altmetric records of retweets of COVID-19-related scientific papers on Twitter filtered by the keyword "COVID-19," covering March 2020–May 2022 (27 months) with 13,806,356 retweets. Engagement measure: Users were grouped into 27 subgroups based on the number of months they engaged (1–27) during the period, capturing persistence/sustainability of engagement. Analyses: 1) Subgroup mobility: For each month, they computed the retweet proportion per subgroup (monthly subgroup retweets divided by total monthly retweets) to assess heterogeneous flow. They examined associations between subgroup proportions and the overall retweet trend using Pearson correlation and Maximum Information Coefficient (MIC). 2) Network centrality: Because Altmetric lacks explicit retweet edges, they matched tweet @usernames to Twitter IDs (85.2% matched) and constructed 27 monthly directed, weighted retweet networks (nodes: users; edges: retweets from i to j; weights: retweet frequency). They computed relative average centrality for each subgroup (subgroup average centrality divided by network average) using in-degree and in-closeness (inward information) and out-degree and reverse-network closeness (outward information). Associations with overall trends were assessed via Pearson and MIC. 3) Topic preference: They mapped retweeted papers to Fields of Research (FoR), focusing on 11 Medicine and Health Sciences, 16 Studies in Human Society, and 14 Economics. Given topical stabilization after September 2020, they analyzed September 2020–May 2022, correlating subgroup field shares with overall subgroup retweet trends. Model and simulations: They developed a network SIRS model with heterogeneous individual information acceptance thresholds. States: Susceptible (S), Infected (I), Recovered (R). Transmission probability depends on an individual's base acceptance and threshold; external shocks modulate acceptance when shock popularity exceeds the threshold (β increases by a function of I − x_i; otherwise unchanged). Recovery rate μ governs leaving I to R; a re-susceptible rate allows cycling. Simulation network: Barabási–Albert scale-free network with 5000 nodes, average degree 6. Initial conditions: 5% nodes infected. Thresholds: 10 levels from 0 to 0.9 (step 0.1), distributed by a power law (exponent 1.5) to reflect many low-threshold and fewer high-threshold users. External shocks: negative shock during t∈[15,20] with intensity I=0.5; positive shock during t∈[35,50] with intensity I=0.8. They computed subgroup infection proportions over time, correlations (Pearson, MIC) with overall infection trends, and built diffusion networks per time step to assess relative centralities (out-degree, reverse-closeness, in-degree, closeness). They also varied the threshold distribution heterogeneity via power-law exponent α (including α=0 uniform) to test effects on propagation effectiveness and resilience.

- Empirical flow: Low-engagement users exhibited larger fluctuations in retweet proportion, with rapid initial surges and subsequent declines; dips for low-engagement users aligned with overall downturns (e.g., June 2020, Dec 2020, Mar 2021, June 2021, Oct 2021), while high-engagement users showed minor peaks. - Association with overall trend: Significant positive Pearson correlations for low-engagement users and significant negative correlations for high-engagement users; MIC scores were high at both extremes. There was a significant linear relationship between engagement level and Pearson correlation coefficient (R2=0.876, p<0.001). - Network position: For outward information, subgroups with engagement 6–20 had significant positive correlations between their relative average centralities (out-degree and reverse-network closeness) and the overall propagation trend; other groups showed no significant association. For inward information, high-engagement subgroups had stronger associations, indicating these users often serve as vital intermediaries whose prominence may weaken during declines. - Topic preferences: After Sep 2020, low-engagement users' engagement with fields covaried with their overall retweet trends: positive correlation (~0.5) for Medicine and Health Sciences; negative correlations (about −0.6 for Studies in Human Society and about −0.5 for Economics). As their spread decreases, their attention shifts from health to socio-economic topics and vice versa. High-engagement users' preferences were generally not significantly associated, with exceptions (e.g., engagement 25 and 27 positively correlated with medical topics). - Modeling alignment: Simulations reproduced empirical patterns: low-threshold groups showed significant positive correlations with overall infection trends, high-threshold groups significant negative correlations; linear relation between threshold and correlation (R2≈0.887–0.888, p<0.001). Under negative shocks, low-threshold users’ proportions decreased while high-threshold increased; the opposite occurred under positive shocks. - Heterogeneity effects: Making threshold distributions uniform (α=0) increased overall propagation effectiveness by 23.9% over long-tailed distributions (α=1.5) and improved resilience to negative shocks (11% vs 24.3% decrease in infected counts before/after interference). - Centrality in simulations: Outward centralities of groups with thresholds <0.4 correlated negatively with overall trends, while thresholds >0.4 correlated positively; inward centralities were positively associated for higher-threshold groups, mirroring empirical results. The model could not fully reproduce the declining Pearson coefficients observed empirically for engagement >20.

The findings support the hypothesis that engagement heterogeneity and external shocks jointly shape fluctuations in science communication. Low-engagement users are highly sensitive to overall volatility, entering quickly during surges and disengaging during declines, while high-engagement users provide stability and structural support in networks. This asymmetry explains uneven information flow during instability and helps interpret shifts in topical attention, wherein low-engagement users pivot between health and socio-economic content depending on broader trends. The model, incorporating individual acceptance thresholds and shock-modulated transmission, offers a mechanistic account consistent with empirical correlations, centrality evolution, and subgroup flow patterns. Conceptually, the study links PES to attention flow dynamics, reframing science communication outcomes as a function of limited, zero-sum public attention and highlighting how external shocks can exacerbate disparities and implicit biases among user subpopulations.

The paper contributes to PES by operationalizing public engagement heterogeneity and connecting it to macroscopic fluctuations in science communication. Empirically, it shows low-engagement users drive much of the volatility and topic shifts, while high-engagement users sustain network structure. Theoretical modeling with heterogeneous acceptance thresholds and external shocks replicates these patterns and clarifies mechanisms behind subgroup centrality evolution. Practically, reducing the proportion of very low-threshold (bursty) participants or tailoring targeted, customized messaging can improve propagation efficiency (e.g., a 23.9% improvement under a uniform threshold distribution) and resilience to shocks. Future research should: 1) further integrate simulation with observed data to capture complex behaviors of highly engaged users; 2) refine subgrouping by sustainability for PES measurement; 3) treat communication fluctuation as an explicit outcome of PES; and 4) examine institutional roles and cross-cultural contexts to generalize findings beyond COVID-19 and Twitter.

- Data reconstruction: Original Altmetric data lacked explicit retweet edges; retweet networks were reconstructed from Altmetric search results with 85.2% username-to-ID matching, potentially omitting or misclassifying some ties. - Platform and topic scope: Analyses focus on Twitter and COVID-19-related research during March 2020–May 2022; generalizability to other platforms, topics, and periods may be limited. - Modeling simplifications: Simulations used a static BA network and stylized SIRS dynamics with threshold-modulated transmission and simplified external shocks; real-world dynamics (e.g., evolving networks, content features, institutional interventions) are more complex. - Partial model fit: The model could not fully explain the declining Pearson coefficients for users with very high engagement (>20 months), indicating additional mechanisms may govern highly engaged participants’ behavior.