Increasing Uptake of Social Distancing during COVID-19: Machine Learning Strategies for Targeted Interventions

Social distancing emerged as one of the early critical non-pharmaceutical interventions to fighting the spread of COVID-19. However, in the US, this behavior was not universally adopted. In late March of 2020, we surveyed 2,500 US respondents to better understand the drivers of social distancing behavior. Our survey measured demographics, social distancing awareness, beliefs, barriers, and behaviors. We first used predictive modeling to identify a broad set of factors correlated with social distancing. However, it could not reveal which variables were the critical causal drivers of this behavior. We used Bayesian network (BN) to map the causal relationships between variables. Our BN pinpointed which variables were causal drivers of social distancing intentions and behavior: higher financial security, higher information seeking, and higher worry about the coronavirus. The BN cast doubt on the effectiveness of potential interventions that would have been suggested by the predictive model alone, such as interventions on community norms perceptions, as well as factors that have previously received attention in the media, such as religion and political affiliation. Finally, to more easily identify target groups for policy recommendations, we performed K-means clustering that distinguished population segments based on social distancing beliefs and behavior. We identified four segments ranging from a 'worried social distancers' (55.3% always social distanced), to 'uninformed skeptics' (25.9% always practiced). Taken together, our results demonstrate how a precision public health approach can help policymakers design more targeted and efficient public health interventions for social distancing. This approach can help prioritize messages most effective for matched population targets, increasing desirable outcomes while potentially saving resources.