The utility of prediction models to oversample the long-term uninsured.
ABSTRACT To evaluate the performance of prediction models in identifying the long-term uninsured and their utility for oversampling purposes in national health care surveys.
Nationally representative data from the Medical Expenditure Panel Survey (MEPS) were used to examine national estimates of nonelderly adults without health insurance coverage for 2 consecutive years and to identify the factors that distinguished them from the short-term uninsured and those who are continually insured. The MEPS data were also used in the development of the prediction models to identify individuals most likely to experience long-term spells without coverage in the future. The prediction models were developed using data from the MEPS panel covering 2004-2005 and evaluated with an independent MEPS panel.
Study findings revealed these prediction models to be markedly effective statistical tools in facilitating an efficient over-sample of individuals likely to be uninsured for long periods of duration in the future. Use of these models for oversampling purposes, to support a 50% increase in sample yield over a self-weighting design, permits the selection of the target sample of individuals who are continuously uninsured for 2 consecutive years in the most cost-efficient manner. This methodology allows for an overall sample size specification for nonelderly adults that is at least 25% lower than a design without access to the predictor variables from a screening interview or without application of oversampling techniques.
This examination of the performance of probabilistic models, to both identify and facilitate an oversample of the long-term uninsured, demonstrates the viability of these model-based sampling methodologies for adoption in national health care surveys.
- [Show abstract] [Hide abstract]
ABSTRACT: There is a growing demand for timely, high quality and precise estimates of health care parameters at the national and sub-national levels, and associated readily accessible data resources to inform health care policy and practice. Existing sentinel health care databases that provide nationally representative population based data on measures of health care access, cost, use, health insurance coverage, health status and health care quality, provide the necessary foundation to support descriptive and behavioral analyses of the U.S. health care system. Such studies help inform assessments of the availability and costs of private health insurance in the employment-related and non-group markets; the population enrolled in public health insurance coverage and those without health care coverage; and the role of health status in health care use, expenditures, and household decision making, and in health insurance and employment choices. To complement these assessments of the "current state" of health care, policymakers also depend on model-based estimates of the "future state" under alternative demographic, economic and technological assumptions, which are subject to greater levels of uncertainty traditionally associated with sampling and nonsampling error. Such modeling efforts directed to predicting the "future state" include economic models projecting health care expenditures and utilization, estimating the impact of changes in financing, coverage, and reimbursement policy, and determining who benefits and who bears the cost of a change in policy. Government and non-governmental entities rely upon these data to evaluate health reform policies, the effect of tax code changes on health expenditures and tax revenue, and proposed changes in government health programs such as Medicare. Comparable standards of data quality and statistical integrity for these types of modeling and microsimulation efforts are needed to ensure policymakers have a sound understanding of the level of uncertainty associated with these model-based estimates. This presentation will focus on several of these ongoing health care modeling and microsimulation efforts to characterize sources of uncertainty in the resultant estimates and methodologies that can be employed to better quantify their error bounds.
- Medical care 07/2009; 47(7 Suppl 1):S1-6. DOI:10.1097/MLR.0b013e3181a7e401 · 2.94 Impact Factor