On the Use and Misuse of Child Height-for-Age Z-score in the Demographic and Health Surveys

This paper addresses the problem of model misspecification bias when estimating cohort-level determinants of child height-for-age zscore (HAZ) using data from the Demographic and Health Surveys (DHS). I show that the combination of DHS survey design and the biological realities of child health in developing countries create an artifact that can strongly bias regression estimates when identification relies on seasonal, annual or spatio-temporal variation associated with a subject’s birth cohort. I formalize the econometric problem and show that flexible specifications of the HAZ-age profile can greatly mitigate the bias. When regression models can exploit within-cohort variation in the covariate of interest, appropriate fixed-effects models can effectively purge the bias. I also provide Monte Carlo evidence that DHS recommended inference strategies produce standard errors that are too small when estimating birth cohort determinants of HAZ. ∗I would like to thank Hilary Hoynes, Douglas Miller, and Steve Vosti for helpful comments on previous drafts. I would also like to thank Kathryn Dewey, Edward Miguel, Manisha Shah, Tania Barham, Jorge Agüero, Adrian Barnett, Bruno Schoumaker and Tom Vogl for helpful conversations and encouragement, and seminar participants at UC Davis, UC Berkeley, and the Pacific Development Conference. All mistakes, conceptual shortcomings, and methodological imperfections are my own.