Positive relationship between education level and risk perception and behavioral response: A machine learning approach

PLOS ONE | https://doi.org/10.1371/journal.pone.0321153 April 3, 2025 7 / 10 PLOS ONE Positive relationship between education level and risk perception and behavioral response among the top 20 ranked features, risk perception (xtotscore), place of residence (Q24), and education level (Q12) are particularly significant. To explore how these features influence preparatory behavior, Fig 1(b) displays a scatter plot of SHAP values for each sample and feature. Each point in the plot conveys three pieces of information: the vertical axis represents the feature, the color indicates the feature value (with red indicating higher values and blue values), and the horizontal axis shows the direction of influence on the predicted value.From the feature importance ranking and the influence of each feature on preparatory behavior, we draw the following conclusion. Education level (Q12) plays a significant role in shaping preparatory behavior (ytotscore), serving as both a core predictive indicator and a potential contextual factor. Firstly, within the multivariate explanatory variables for preparatory behavior, education level ranks third, contributing 6.87% to the overall model, highlighting its strong explanatory power. Secondly, unlike other variables, which show concentrated SHAP values around zero, education level exhibits distinct dispersion. Specifically, individuals with higher educational attainment (red points) show a strong positive correlation with preparatory behavior, while those with lower

educational attainment (blue points) display a negative correlation. This further emphasizes the importance of education level in the model and warrants further exploration of its influ- encing mechanisms [24]. Additionally, as noted by Woodridge, the interaction of significant variables during evidence generation can notably affect the effect size, suggesting that the role of education level as a control variable can directly impact the outcome, making it a critical factor in shaping general evidence. As shown in Fig 2, with behavioral intention as the output variable, Fig 2(a) presents the global feature importance ranking based on the average absolute Shapley values for each feature. Among the top 20 features, risk perception (xtotscore), place of residence (Q24), and transparency of research progress on earthquake relief and disaster prevention in the residential area (Q28_Row4) are notably significant. To explore how these features influence preparatory behavior, Fig 2(b) displays a scatter plot of SHAP values for each sample and Fig 2. Feature Impact Summary Based on SHAP Values. Left: (a) Feature Importance Ranking; Right: (b) Influence of Features on Behavioral Intention Effectiveness. https://doi.org/10.1371/journal.pone.0321153.g002