We consider the availability of new harmonized data sources and novel machine learning methodologies in the construction of a social vulnerability index (SoVI), a multidimensional measure that defines how individuals’ and communities may respond to hazards including natural disasters, economic changes, and global health crises. The factors underpinning social vulnerability—namely, economic status, age, disability, language, ethnicity, and location—are well understood from a theoretical perspective, and existing indices are generally constructed based on specific data chosen to represent these factors. Further, the indices’ construction methods generally assume structured, linear relationships among input variables and may not capture subtle nonlinear patterns more reflective of the multidimensionality of social vulnerability. We compare a procedure which considers an increased number of variables to describe the SoVI factors with existing approaches that choose specific variables based on consensus within the social science community. Reproducing the analysis across eight countries, as well as leveraging deep learning methods which in recent years have been found to be powerful for finding structure in data, demonstrate that wealth-related factors consistently explain the largest variance and are the most common element in social vulnerability.