Partially Identifying the Prevalence of Health Insurance Given Contaminated Sampling Response Error

This paper derives simple closed-form identification regions for the U.S. nonelderly population's prevalence of health insurance coverage in the presence of household reporting errors. The methods extend Horowitz and Manski's (1995) nonparametric analysis of contaminated samples for the case that the outcome is binary. In this case, draws from the alternative distribution (i.e., not the distribution of interest) might naturally be defined as response errors. The derived identification regions can dramatically reduce the degree of uncertainty about the outcome distribution compared with the contaminated sampling bounds. These regions are estimated using data from the Medical Expenditure Panel Survey (MEPS) combined with health insurance validation data available for a nonrandom portion of the sample.