What Can We Learn About the Effect of Mental Health on Labor Market Outcomes Under Weak Assumptions? Evidence from the NLSY79 (with Vikesh Amin, Carlos A. Flores, and Alfonso Flores-Lagunes). Labour Economics, 79, Article 102258 (2022).
We employ a nonparametric partial identification approach to bound the causal effect of poor mental health on employment and earnings using the National Longitudinal Study of Youth 1979. Our approach allows us to provide bounds on the population average treatment effect based on relatively weak, credible assumptions. We find that being categorized as depressed decreases employment by 10% and earnings by 27% at most, but we cannot statistically rule out a zero effect. We also provide insights into the heterogeneity of the effects on labor market outcomes at different levels of adverse mental health experienced (no, little, mild, moderate, and severe depressive symptoms). We find that going from having no (little) to severe depressive symptoms reduces employment by 3–18% (3–16%) and earnings by 11–44% (12–36%). The estimated bounds statistically rule out null effects for earnings but not for employment.
Housing quality, both in terms of the physical structure and neighborhood characteristics, is positively associated with health outcomes. I examine the Housing Choice Voucher (HCV) program as a possible means of improving the health of recipients who would otherwise participate in place-based assistance programs such as public housing, or rent without a subsidy in the private market. Under three weak and plausible assumptions—voucher holders have weakly worse average potential health outcomes than non-recipients (non-positive monotone treatment selection); receiving a voucher does not result in strictly poorer health (monotone treatment response); on average, potential health outcomes are positively related to reported income (monotone instrumental variable)—I obtain nonparametric bounds on the average treatment effect of HCV on the health of individuals in recipient households using nationally-representative data from the 2018 Survey of Income and Program Participation (SIPP). My preferred estimates indicate that the causal effect is positive and statistically significant, with the likelihood of good or better self-reported health status increased by at least 4.8 percentage points and at most 21.3 percentage points. I also estimate that the probability of not having been hospitalized over the previous year is increased between 0.1 and 18.4 percentage points, though the effect is not statistically distinguishable from zero at conventional levels. Among Black members of voucher households, I estimate a larger lower bound for the effect on self-reported health of 11.7 percentage points, hinting that the effect on health is potentially larger for this subpopulation.
Work in Progress
The Use of Polygenic Scores as Monotone Instrumental Variables (with Vikesh Amin, Carlos A. Flores, and Alfonso Flores-Lagunes)
Researchers are increasingly using polygenic scores (PGSs)—summary measures of genetic predisposition—as instruments to identify causal effects, but they are unlikely to satisfy the exclusion restriction. We propose the use of PGSs as monotone instrumental variables (MIV) within the nonparametric partial identification framework of Manski and Pepper (2000). Under the MIV assumption—that PGSs have a monotone (weakly increasing or decreasing) relationship with the outcome—they can help bound the causal effect of interest together with weak monotonicity assumptions regarding treatment selection and response. We provide an exposition of the Manski and Pepper (2000) framework and assumptions. We illustrate the use of PGSs in this setting with three empirical applications.
mpclr: A Command for Bias-Corrected Estimation and Valid Inference for Nonparametric Bounds on Treatment Effects (with Carlos A. Flores and Alfonso Flores-Lagunes)
Fuzzy Spatial RDD with Bounded Measurement Error in the Running Variable