A new paper by UC-Berkeley’s Tim Brown adds to the growing evidence about the health effects attributable to public health investments. A key innovation in Brown’s approach is the use of Koyck distributed lag models that allow exploration of the time paths that determine how spending in one period influences health outcomes in subsequent periods. This innovation is important given that many public health programs and policies target chronic diseases and risk factors with relatively long incubation and progression periods.
Brown’s study uses annual data on the expenditures of California’s county public health departments during 2001-2008, linked with county-level estimates of all-cause mortality. By exploiting both cross-sectional and longitudinal variation in county public health agency spending per capita, he finds that a $10 increase in agency spending per capita reduces the all-cause mortality rate by 9.1 deaths per 100,000 over this 8 year period. At current levels of spending in California, these results suggest that local public health agencies have averted a total of 27,000 deaths per year in California during this period, generating a total economic value of $212 billion – more than $100 in benefit for every $1 invested.
Another notable feature of this study is its approach for addressing endogeneity bias in estimates of how spending influences mortality. Addressing the endogeneity of spending is absolutely essential for supporting causal inferences about the health effects of public health spending, as my own work has shown. Eliminating this bias generally requires using an instrumental-variables (IV) approach that hinges on finding instruments that induce exogenous variation in spending but that have no direct effect on the health outcomes of interest. These IVs function as a stand-in for randomization, allowing researchers to approximate results that would be obtained if it were possible to randomly assign communities and agencies to different levels of spending (with certain caveats of course).
The Brown study employs a method recently proposed by Lewbel to reduce endogeneity bias when traditional instrumental variables are not available. The “Lewbel IV approach” relies purely on heteroskedasticity in the model to identify unbiased estimates. By relying on distributional assumptions about the error term, this approach is likely to be less reliable than traditional IV methods, and the results must be interpreted with caution. However, in the absence of traditional IVs, this approach may be the only feasible strategy available to address endogeneity bias.
This new work, supported through the Robert Wood Johnson Foundation’s PHSSR research portfolio, is an important contribution to the growing evidence base about the value of public health programs and policies. For my fellow travelers in empirical public health research, I would urge caution in employing some of the methods illustrated here – particularly the Lewbel approach, as it is not a shortcut to a quicker and easier IV strategy. There is no substitute for using knowledge of the underlying data generating process to find the strongest possible research design and analytic approach for an observational study, and traditional IV methods usually offer the best alternative to an experimental study. But in certain circumstances, the methods demonstrated in this new study may offer a next-best alternative.