Does more medical care improve population health? New evidence for an old controversy

Abstract

This article investigates the aggregate relationship between medical care and health for the US population. I use annual state level panel data for the period 1983 to 2000 to estimate static and dynamic health production function models. I find no compelling evidence that greater aggregate utilization of medical care from application of existing technology improves population health by lowering mortality in the short run or long run. My results suggest that development of new medical technologies that diffuse rapidly throughout the nation and at different rates across states may well explain much of the decline in the age-adjusted death rate over the past several decades, as well as persistent differences in mortality across geographic regions. Overall, my findings suggest that the US may be experiencing ‘flat of the curve medicine’ with future improvements in mortality from medical care coming from new and better technologies rather than greater intensity of services.

Notes

¹ Historical data on medical care expenditures in the US are from the Centers for Medicare and Medicaid Services (CMS) website, http://www.cms.hhs.gov.

²An often used approach to describe technical change in the econometric literature on production and cost is either the inclusion of a time trend or a set of time specific dummy variables. For example, see Baltagi and Griffin (1988).

³ The medical care utilization variable has several unavoidable sources of measurement error. The data available on state health care spending for the time period covered in this study measures the value of health care services produced by providers, not the amount consumed by residents in a state. Border crossing by state residents for health care services results in overstating spending in net exporter states and understating spending in net importer states. In addition, deviations from the national average price of medical care services among states, to the extent which they exist, distort measures of real medical care spending. These sources of measurement error may well lead to attenuation bias, which I attempt to mitigate by using an IV estimator.

⁴ Past studies that analyse the effect of nonmedical variables include Auster et al. (1969), Hadley (1982), Wolf (1986), Smith (1999), Grossman (2000), Ruhm (2003), Deaton (2002), Thornton (2002), Fuchs (2004a, b), and Marmot (2004).

⁵ See Appendix for years for which data is available for interpolated variables.

⁶ The F-statistics for instrument relevance for 2SLS specifications that include time effects and fixed effects range between 15.69 and 38.30 providing supporting evidence that age distribution variables are strong instruments for all 2SLS specifications reported in Table 2.

⁷ These studies generally report 2SLS estimates that are about two to three times as large as the OLS estimates, but do not assess the strength of the instruments used. Estimates provided by all previous state-level studies are imprecise and fail to achieve statistical significance at any reasonable level. The relatively more precise estimate obtained in this study may result from more observations and less collinearity from the use of panel data, as well as stronger instruments, and coefficient and SE estimates that account for possible heteroscedasticity.

⁸ When educational attainment variables are included in this specification, the estimate of the coefficient of income changes from −0.241 to −0.247 suggesting that income is not serving as a proxy for omitted education.

⁹ Estimates of the year effects, state effects and state-specific trends are available upon request from the author.

¹⁰ Regression results for all variables are available upon request from the author.

¹¹ Data are available for medical care utilization and age distribution variables for years 1980–2000, and therefore the distributed lag models reported in Table 4 are estimated using 850 observations.