Partially Identifying Treatment Effects in the Presence of Unobserved Treatments: Covering the Uninsured
- Presenter:
Chair: Joel Hay; Discussant: TBA Tue June 6, 2006 8:00-9:30 Room 213
Policymakers have long been interested in identifying the number of people lacking health insurance, the consequences of uninsurance for access to health care, and the potential cost of covering the uninsured. Household surveys provide the primary source of information about the uninsured, but many validation studies find some respondents misreport insurance status. Reporting errors may lead to inaccurate estimates of the number of uninsured and bias estimates of the effects of insurance on access, use, and costs. Using data from the 1996 Medical Expenditure Panel Survey (MEPS), we investigate what can be learned in the presence of arbitrary health insurance reporting errors about (a) the gap between the insured and uninsured in the use of health services and (b) the impact of universal insurance coverage on the use of services. We exploit information from insurance cards, policy booklets, and follow-back interviews with employers and insurance companies to construct validation data for a nonrandom portion of the sample. There are 18,851 nonelderly in the 1996 MEPS, and about two-thirds of them have evidence from at least one source validating their reported insurance status. Extending the theoretical literature on nonparametric bounds and treatment effects, we formally characterize the identification problem and assess the identifying power of a variety of verification, monotonicity, and independence assumptions. Molinari’s (2002) “missing treatments” bounds are tightened when there is information about the potential degree of reporting error within the subpopulation without validation data. For some results, we estimate rates of false negatives and false positives on a small subsample whose employers responded to the follow-back interviews and use them to extrapolate potential rates of reporting errors to the third of the sample that does not have validation data. Under a set of relatively strong nonparametric assumptions, we preliminarily estimate that coverage the uninsured would increase the proportion of the population using health care in a month by no more than 2.5 percentage points for adults (an 11% increase) and by no more than 0.8 percentage points for children (a 4.8% increase).
