Hamad and Rehkopf Respond to “Income and Health: Financial Credits as Instruments”

We appreciate the commentary by Pega (1) in response to our study, in which we used the Earned Income Tax Credit (EITC) as an instrumental variable (IV) for income to estimate the causal effects on child health (2). We agree with Pega's central argument regarding the necessary assumptions of IV analyses. There is nothing magical about calling a covariate an instrument. As Pega clearly describes, instruments can be subject to the same confounding as other variables. We believe, however, that there are approaches that can be used to address some of the inference concerns that Pega describes, many of which we used. In responding, we hope to highlight that specific subject matter knowledge about the instrument itself is critical in any IV study.

We believe the most pertinent concern raised by Pega is the possibility that other policy changes occurred simultaneously with the EITC at the federal or state level, leading us to incorrectly ascribe health effects of other programs to the EITC itself. We disagree, however, with Pega's statement that it is necessary that investigators comprehensively measure and account for relevant confounders to ensure valid inference. Doing so would potentially lead to an insurmountably complex directed acyclic graph, with reasonable arguments for an even greater number of potential confounders extending back in time to many historical governmental policies that set in place a cascade of events leading to the current US political structure. It is in fact the case that we only need to break links between confounders that lead to our exposure or outcome of interest. In our study, we addressed this in several ways. For example, we included in our model fixed effects for year—the time scale of EITC policy implementation—to alleviate bias due to secular changes in the political climate that may have led to the implementation of other potentially confounding policy changes.

Pega's second concern is that individuals may be motivated to maximize the size of their EITC payment, thus confounding the amount of benefit received. Those who have the ability to maximize their credit by changing their amount of earned income may differ in ways that also lead to better health. We agree that this is true, and we addressed this in our analysis by imputing the EITC payment size exposure using demographic variables (including earned income) from 2 years prior. In other studies, investigators have taken similar and even more conservative approaches by using level of education to predict EITC benefits (3). It is critical that qualified benefits be used as the exposure, rather than actual money received as suggested by Pega, because the latter would create the differential selection that we are working to avoid.

A third concern raised by Pega is that the “financial credits may directly improve health through increasing a psychological sense of social security” (1, p. 787) rather than through income itself. This concern can be in part addressed using detailed knowledge of the policy itself. To us it seems unlikely that a tax credit would provide a sense of security without an actual monetary transfer involved. This policy in particular is different from the promise of a more general social safety net. In fact, qualitative studies on how households view the credit and spend money are informative for evaluating this criticism (4).

These critiques have broader implications for interdisciplinary collaborations between epidemiologists and investigators in other fields. A more thorough understanding of the political climate or of historical conditions surrounding a potential policy instrument will allow investigators to apply it with better validity. For social epidemiology, this will mean collaborating with sociologists, political scientists, historians, and others who may be able to inform the selection of instruments that more closely adhere to the underlying assumptions. For example, using the placement of railroad tracks in the 19th century as an instrument for US residential segregation requires specific knowledge of railroad engineering practices in the mid-1800s, not a typical subject of epidemiologic coursework (5). The successful use of IVs requires substantive content knowledge in addition to statistical expertise.

Along with Pega, we are excited about the utilization of IV methods and other quasi-experimental approaches to address challenging questions in social epidemiology. By more rigorously providing estimates of the causal effects of adverse socioeconomic contexts, epidemiologists can contribute to the development of effective policies and other interventions to address these fundamental causes of health disparities (6). Additionally, in many cases it may not be appropriate or necessary to use an IV approach rather than to directly examine the impact of a policy change itself. An IV approach is called for if the intent is inference about a general concept (such as income), whereas an analysis of a policy change using ordinary least squares methods is most appropriate for determining the effects of the policy itself. In many cases, the latter may be more beneficial for creating a practicable social epidemiology, in particular for income receipt, for which effects on health may differ depending on the details of the policy.