Using partial identification methods to estimate the effect of violence against women on their children’s health outcomes

ABSTRACT

A growing literature in economics seeks to estimate the costs of violence against women by examining, for example, its impact on the health outcomes of their children. However, it is difficult to assign a causal interpretation to these nonexperimental studies due to the presence of unobservable characteristics affecting violence and health outcomes simultaneously. The lack of credible instrumental variables applicable in several countries further limits our knowledge. I address this gap by using new partial identification methods to estimate the relative size of the unobservables needed to eliminate the estimated effects in nonexperimental studies. I also expand the external validity of the analysis by using data from five standardized nationally representative household surveys in Latin America. Consistent with previous studies, cross-sectional estimates show large negative associations between violence against women and an array of child health outcomes. However, when accounting for omitted variable bias, at best, two-thirds of the estimates remain robust and they are concentrated on the outcomes with the largest cross-sectional estimated impacts.

KEYWORDS:

• Partial identification
• domestic violence
• children’s health
• Latin America

JEL CLASSIFICATION:

• I10
• I12
• J16

Acknowledgement

I thank Catalina Herrera and Michele Back for their helpful comments and suggestions.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

¹ See Yount et al.’s (2011) for a meta-analysis and mechanisms.

² Aizer (2011) uses county and time variation in the enforcement of California’s domestic violence laws. Nuhu (2016) uses proxies for legislation change in Ghana. However, the lack of comparable variation in enforcement and a different cultural context limits the use of similar instruments for most developing countries, including those in Latin America.

³ See Agüero (2013) for more details about these questions and alternative measures for these countries.

⁴ Altonji et al.’s and Oster’s methods are designed for linear models; extensions to nonlinear models are left for future research.

⁵ For some regressions, there is even a sign difference between the role of the observables and unobservables denoted by δ < 0.

⁶ R² is obtained from the regression with all the (observable) controls. See Table 2, column 1.