Social
Historicizing instrumental variables: Insights from an IV for racial segregation Mark Hernandez* Mark Hernandez Hernandez Drexel University
Background: Instrumental variables are increasingly sought-after for use in studies of social exposures and health, given how difficult (and unethical) social exposures are to randomize. Epidemiologists are thus looking to economics studies to find potential IVs. However, many IVs’ impacts on social exposures may be historically contingent and change across time and space. Here, we examine an instrumental variable for residential segregation used in health research as a case study of the importance of historicizing health econometrics.
Methods: We examine an instrumental variable developed by Ananat, who showed that the quasi-random degree to which cities built their railroad networks in a grid in the late 1800s/early 1900s served as an instrument for how racially segregated those cities became during the Great Migration. We focus on the first stage of IV analyses using this instrument, regressing metro area anti-Black segregation on Ananat’s railroad index. We specifically examine whether different cities’ segregation trajectories vary over time and space in response to the IV.
Results: We find that railroad configuration-induced changes in segregation are historically contingent. The IV not only caused affected cities to become more segregated initially, it also slowed the degree to which segregation declined even many decades later, well after railroads were dismantled. As such, the strength of the IV grows as time passes, requiring careful first-stage modeling.
Conclusion: Instrumental variables for social exposures in one time or place cannot be assumed to operate similarly in others, even when an IV performs well on average across study sites and years. First stage models ignoring this heterogeneity will produce less precise estimates that could have been strengthened, particularly, with better mapping between (A) detailed historical grounding and sociologic theory and (B) how first stage models are operationalized.