Exploring Behavioral Determinants of Flood Insurance Adoption with Explainable Machine Learning in the Continental US

Abstract

<div aria-live="assertive"><span>Flood insurance is a straightforward way to provide resources</span><span> for ex-post recovery from the damages caused by floods</span><span> and</span><span>,</span><span> therefore</span><span>,</span><span> streng</span><span>t</span><span>hen household resilience</span><span> against this type of natural hazards</span><span>. The US National Flood Insurance Program is the centralized source of flood insurance in the US providing more than 5 million policies in force today. However, only less than 5% of </span><span>all </span><span>US households </span><span>are currently insured against flood damage</span><span>. Understanding the determinants of flood insurance purchase is key to support the development of future resilience strategies. Yet, the question o</span><span>f</span><span> which household characteristics and motivations le</span><span>a</span><span>d to flood insurance purchase is still not answered. </span><span>In this work we consider </span><span>flood</span><span> insurance </span><span>adoption</span><span> at the spatial scale of census tract (unit of ~ 3000 inhabitants) as an indicator for flood resilience. We test 397 candidate features to identify relevant determinants of flood resilience</span><span> in the continentall US</span><span>. Our feature set predominantly includes socio-economic variables from the American Community Survey</span><span>,</span><span> along with the flood history, </span><span>rate discounts</span><span>,</span><span> and home ownership. We appl</span><span>y</span><span> an explainable Machine Learning approach based on Light Gradient Boosting Machine (LightGBM) to predict insurance coverage and estimated the SHAP values (SHapley Additive exPlanations) for e</span><span>ach</span><span> feature. </span><span>SHAP values indicate the marginal contribution of each feature to the model output for every census tract</span><span>. </span><span>This</span><span> enables us to understand how our data-driven model deducted the predictions</span><span> and to reduce the initial set of candidate features to a subset of representative features that explain flood insurance adoption</span><span>.  </span><span>We found that insurance coverage at the whole US scale is driven by home ownership, previous flood severity and frequency</span><span>,</span><span> as well as </span><span>financial incentives</span><span>. Conversely, the impact of socio-economic background is marginal at this scale of aggregation. In other words, if a census tract experienced a very severe flood in the past, more inhabitants are insured, compared to inhabitants in census tracts with no direct experience of severe floods. The same counts for regular flooding, yet to a smaller degree. Also, people in census tracts which do not profit from their communities voluntarily implementing floodplain management strategies to acquire subsidized insurance rates are less willing to purchase private insurance. Our results overall suggest that</span><span> households will get insured irrespective of their social background</span><span>, if the community provides financial incentives by participating in the community rating system or has experienced severe flooding. Finallly, we identify areas prone to fluvial flooding (e.g., Lower Mississippi) with potential to improve flood resilience by community subsidi</span><span>z</span><span>ation. Targeted risk communication should be aimed at urban areas with high fluctuation of inhabitants that are unaware of the flooding history.</span></div><div aria-live="assertive"></div>