Fig 1.
Isolating behaviors associated with null risk perception in the MIT COVID-19 beliefs, behaviors & norms survey.
a: UMAP embedding and HDBSCAN clustering. Using UMAP for dimensionality reduction, projecting the multidimensional manifold where the survey data lie to two dimensions and the clustering algorithms HDBSCAN to identify questions in the survey where globally answers were similar, we isolate in the survey a set of non-compliant behaviors (red circles) associated with the lack of perception of danger associated with COVID-19. Other clusters include answers indicating higher level of perceived danger (gray pentagons, yellow triangles and orange diamonds, see Materials and methods) or with a mismatch between the importance for the respondents of taking action against the epidemics and the perceived importance for the community (brown squares). b: Fraction of respondents who declare non-compliant behaviors in the 64 countries considered. Besides the community risk (blue), all answers of the “No risk” cluster of panel (a) are associated with three questions isolating three non-compliant behaviors: unfamiliarity with social distancing (red), disbelief in face mask effectiveness (green) and the lack of preventive actions taken (yellow). Together with the lack of actions taken, also other ineffective actions are included in the “No risk” cluster, but they are here not aggregated together with the “None” answer since the answers were non-exclusives. c: World map of the fraction of respondents doubting the effectiveness of masks. Higher mistrust in mask usage is observed in Africa while Asian countries appear to be more accustomed to the public health control measure. Map dataset from Natural Earth website (https://www.naturalearthdata.com/). See also S2 Fig for maps associated with the other three questions.
Fig 2.
Population fraction of risk-deniers affects epidemics in synthetic networks models.
a: Peak of hospitalized population as a function of the fraction of population of infectious risk-deniers α, for, respectively, uniformly random (ER), Barabási-Albert (BA), Watts and Strogatz (WS) and stochastic block model (SBM4). Each dot represents the average measure across 50 dynamical samples of a single network realization, box-plots show quartiles of distributions across 50 network realizations. We perform a Kruskal-Wallis H-test to test the null hypothesis that the population medians at fixed α are equal. Post hoc pairwise comparisons between groups at fixed α are required to determine which distributions are different. To this aim, for those α values with Kruskal-Wallis H-test p value p ≤ 0.05, we use post-hoc pairwise comparisons between distributions. In particular we use a Mann-Whitney test, because groups are independent, corrected for multiple comparisons. Significance results are reported as: ****: p ≤ 10−4, ***: 10−4 < p ≤ 10−3, **: 10−3 < p ≤ 10−2, *: 10−2 < p ≤ 0.05. Results show that the presence of communities, as in SBM4, significantly decreases the hospitalized peak for all α fractions, except for α = 0.9. b: Peak of hospitalized population evaluated with respect to the one estimated at α = 0.0, as percentage increase. At high fractions of risk-deniers (α ≥ 0.9), community structures give rise to higher increase in hospitalizations peak with respect models with different topology.
Fig 3.
Population fraction of risk-deniers affects epidemics in real networks.
a: Scatter plot describing number of contacts (edges, E) as a function of the number of individuals (nodes, N) for real-world social network data set (n = 187 networks) considered in this study. Each dot is a single network, its color identifies the database source and its size is proportional to the connectivity heterogeneity measured by the Gini coefficient (see Synthetic and real-world network data for further details). Inset: distributions of modularity show that the considered data sets of real social networks spans from structures characterized by the presence of communities (high modularity) to structures with loosely connected communities (low modularity). b: Temporal evolution of the population of hospitalized patients in real networks. Model dynamics is evaluated at a fixed fraction (α = 0.5) of risk-deniers. Color identifies the database source as shown in panel (a). Values of the population at its maximum height correspond to the peak. c: Peak of hospitalized patients as a function of the population fraction of infectious risk-deniers α for data sets with multiple network instances. Each dot represents results in a single network, box-plots show quartiles of distributions in each data set. Statistical tests and significances as in Fig 2. Results show that networks characterized by high values of degree heterogeneity and modularity, like the ones in Sociopatterns data set, show a peak of hospitalized patients that is significantly lower than peak evaluated in other real networks. d: Peak of hospitalized patients evaluated with respect to the one estimated at α = 0.0, as percentage increase. At high fractions of risk-deniers (α ≥ 0.7), data sets characterized by community structures and high heterogeneity values show an higher increase in hospitalizations peak with respect data set with low modularity and heterogeneity values.
Fig 4.
Degree heterogeneity and community structure affect hospitalized peak in real networks.
Each dot represents degree heterogeneity as a function of modularity for each real network considered in this study. Model dynamics is evaluated at fixed population fraction of risk-deniers (α = 0.5) a: Peak of hospitalized patients. Results show that the lower are values of modularity and heterogeneity, the higher is hospitalized peak. b: Peak of hospitalized patients evaluated with respect to the one estimated at α = 0.0, as percentage increase. Networks with high modularity and heterogeneity show higher percentage peak increase.
Fig 5.
Graphical scheme of the epidemic model.
Compartments included in the model are: susceptible S, uninfected individuals, exposed E, individuals who are not yet contagious, infectious with compliant behaviors IC, who mitigate the spreading of the disease by adopting NPIs, infectious with non-compliant behaviors, or risk-denier infectious ID, individuals who are not compliant with mitigation rules, hospitalized H, people who require hospitalization, self-isolating or quarantined Q, infectious individuals who stay at home and are compliant with NPIs, recovered R, people who are no more infectious. Model parameters are: force of infection λ, incubation rate σ, recovery rate γ, recovery rate for hospitalized patients γhosp, hospitalization rate δ, quarantine rate η, fraction of risk-deniers α, fraction of population that needs hospitalization a and fraction of population that are quarantined at home b.