Maternal Smoking, Misclassification, and Infant Health

Tanguy Brachet, University of Pennsylvania

Abstract

Identifying the causal effect of prenatal maternal smoking on infant health is complicated by unobservable maternal characteristics and behaviors which are plausibly related both to birth outcomes and to a mother's decisions to smoke. Previous economic studies have addressed this omitted variables problem with the use of instrumental variables (IV). However, with (noisy) self-reported data on maternal tobacco use, misreporting can enduce more severe biases in the IV estimates than those resulting from the endogeneity problem which instrumental variables were originally invoked to eliminate. In this paper, I propose a method of estimating the effect of maternal smoking on birth outcomes that addresses both the endogeneity and measurement error problems.

When a binary variable is misclassified, the measurement error is necessarily negatively correlated with the truth. This observation has important implications for an IV framework in which the endogenous variable is a potentially mismeasured binary variable. Ignoring misclassification leads to attenuation in the first stage coefficients and, by extension, to inflated second stage estimates of the causal effects of interest. The GMM approach I propose is based on recently developed parametric methods for misclassified binary dependent variables that allow me to recover consistent estimates of the second stage coefficients, as well as of the misclassification probabilities.

Using this method, I then re-analyze the relationship between infant health and maternal smoking. When state cigarette excise taxes are used as an instrument for tobacco use without accounting for measurement error in self-reported smoking, the conventional IV estimate of the birth weight cost of smoking is only slightly smaller in magnitude than its OLS counterpart for whites, but substantially larger for African Americans: of the order of one third of the average infant's birth weight. Accounting for misclassification yields causal estimates that are considerably smaller in absolute value and more consistent with experimental estimates. The results also suggests that the effect is heterogeneous across races: the birth weight loss due to smoking is around 200 grams for whites and roughly twice that amount for African Americans.