Interpretation of Regressions with Multiple Proxies

Multiple proxy variables are typically available for an unobserved explanatory variable in a regression. We provide a procedure by which the coefficient of interest can be estimated from a regression in which all the proxies are included simultaneously. This estimator is superior in large samples to the common practice of creating a summary measure of the proxy variables. We examine the relationship between parents' income and children's reading test scores in the United States, and between parents' assets and children's school enrollment in India, and demonstrate that the reduction in attenuation bias from a better use of proxy variables can be significant. Copyright by the President and Fellows of Harvard College and the Massachusetts Institute of Technology.