Our primary data source for insurance coverage and individual level demographic variables is the Current Population Survey (CPS). The CPS asks about insurance coverage in the previous year. To increase sample size, we pool data from the 1989–1991 samples for the early time period and from the 1998–2000 samples for the later time period.
Our primary measure of the change in health insurance premiums is based on data from the Health Insurance Association of America and the Kaiser Family Foundation/Health Research and Educational Trust Survey of Employer-Sponsored Health Benefits (KPMG Survey 1988, 1989, 1998
) and Kaiser Family Foundation (1999)
. These surveys obtain information about the type of policies offered and their premiums. We pool surveys from 1988 and 1989 for the early years and 1998 and 1999 for the later years.
We include metropolitan statistical areas in our sample if there are at least 10 premium observations in both the early and late time periods, to minimize sampling error. This yields a total of 64 metropolitan statistical areas. There are 2,111 premium observations in the early years and 4,006 observations in the later years. The MSAs in our sample account for about half of the total United States nonelderly population.
Premiums are for an individual policy offered in a group setting. To account for the differing nature of health insurance coverage, we adjust premiums using a regression model relating plan premium to type of plan (HMO, PPO, POS plan, indemnity plan) and several plan benefit characteristics. Specifically, we regressed the premiums on: plan type, whether an employer offers multiple plans, interactions of plan type and employer offering multiple plans, and interactions of plan type and time period; plan deductible, coinsurance, and copay amounts; whether plans offer prescription drugs, outpatient mental health benefits, inpatient mental health benefits, and maternity benefits; firm size, industry, and a set of MSA dummies and MSA dummies interacted with time period to measure the plan type and benefit adjusted premiums in an MSA. The coefficients from this regression were generally reasonable, indicating that more generous benefits implied greater premiums, and the R2 from this regression was 0.26.
To check the reasonableness of this premium measure, we correlated it with Medicare Part B spending at the MSA level (ρ=0.47) and per capita personal health care spending in the state for the nonelderly population (ρ=0.62). Medicare spending data are from CMS, Office of the Actuary. State spending is calculated from CMS state health accounts data and excludes Medicare spending. Our premium measure may be a more accurate reflection of costs for the nonelderly than Medicare spending because it is less affected by changing costs for services such as home care. It is likely a better measure than state health care spending because insurance markets are likely smaller than the state. State spending blends cost trends across multiple urban and rural areas. Undoubtedly, measurement error in the premium measure remains. If the associated bias is large, our IV models, discussed below, will provide a better measure of the relationship between rising premiums and coverage.
In our models we include two sets of controls (taken from the CPS). Demographic controls are intended to absorb some of the unobserved factors that could confound the analysis. The second set of explanatory variables is designed to control for competing explanations for declining coverage that have been proposed in the literature. We have tried to adopt the measurement approaches used in all the other work to better control for competing explanations.
We include the following demographic variables for each individual and the head of their HIU: age, gender, race/ethnicity, education, marital status, industry, occupation, full/part-time work status, government versus private employer, and firm size. We also include indicators of whether there are no workers or more than one worker in the HIU; interactions of being a spouse or a child in a family with multiple workers; binary variables for the income decile the HIU falls into, calculated separately for singles and married people, and interactions of income decile and marital status of the HIU head. We include interaction terms between these variables and a binary variable capturing observations in the later period to allow for the possibility that their effect changes over time.
Several metropolitan area-level demographic factors are included based on CPS data. These capture market-level effects and competing explanations for the decline in coverage. The MSA-level covariates include the share of the population that is foreign born, the share of the population in the metropolitan area that is nonwhite, the share that is elderly, average HIU income, and the share of women that are working. We also include the local unemployment rate, which is from BLS data available on the Area Resource File (ARF). Unless otherwise indicated, we do not interact the MSA-level or policy covariates with a time period dummy.
Two of the important potential explanations for declining coverage that have been explored in the literature are rising Medicaid eligibility and falling tax subsidies. We control for these explanations using the approaches followed by studies focusing on these explanations. Specifically, we generate measures of the generosity of Medicaid coverage of children following the approach of Cutler and Gruber (1996a)
, using information from the Intergovernmental Health Policy Project (1988
and the National Governors' Association (1990
. They measure Medicaid eligibility by the fraction of HIU health spending eligible for Medicaid, based on family composition, which captures the role of Medicaid eligibility in the context of family health insurance decisions. This is calculated by applying state regulations to CPS data to assess generosity at the state level and adding controls for the fraction of family health spending attributable to each child age.4
To control for changing taxes, we follow the methods of Gruber (2001)
. Specifically, we match average tax rates to individuals based on their state of residence, income decile, marital status, and year (Gruber 2001
). Data on federal, state, and local income tax rates are obtained for 1990 and 1999 from the TAXSIM program of the National Bureau of Economic Research (Feenberg and Coutts 1993
). Gruber (2001)
suggests that this IVs measure of taxes is preferable to calculating tax rates at the individual level to avoid endogeneity associated with the relationship between coverage, income, and taxes.
We measure the availability of charity care in each metropolitan area as the number of public and teaching hospital beds per thousand residents, obtained from the American Hospital Association (AHA) and ARF (American Hospital Association 1991
; Area Resource File 2001
). We measure these variables only in 1990 to minimize reverse causality issues that may arise because these measures could be affected by insurance, rather than the reverse. We include this variable as an interaction with time period, to examine whether the effect of charity care availability changes over time.
State regulations in the 1990s established many restrictions on insurance pricing that could affect coverage (Simon 2000
). We indicate with a dummy variable whether the state has passed rating reforms, which limits the variability in prices across groups, or enacted guaranteed issue, which requires insurers to sell at least some policies to all groups, for the small-group insurance market.
In our base specification we estimate probit regression models for the probability that an individual has health insurance coverage from any source. The regressions are weighted using the Current Population Survey sampling weights to reflect the national sample. Our standard errors are clustered by MSA-time period, which accounts for unobserved factors that are correlated within MSA/time period cells. Our predictions about the impact of insurance premiums and other factors on coverage are obtained from these probit regression estimates. We control for general time trends and time invariant area traits using a dummy variable equal to 1 for the late sample, and dummy variables for each metropolitan area. To project forward, we approximate the effect of rising premiums over GDP growth by inflating premiums by 1, 2, and 3 percent average annual growth for 10 years, holding income and all other variables constant.
The nature of the insurance questions in the Current Population Survey changed somewhat over time (Swartz 1997
; Fronstin 2001
; Mills 2002
). Time trends in insurance coverage are thus subject to some uncertainty. These changes in question wording should not affect our analysis, however, because we control for national trends and rely on differential changes across areas to identify the effects of rising premiums and other covariates.
We were concerned about two potential sources of bias in our probit estimates. First, our premium data is measured with noise, both because of the small sample in the survey of employers and because of imprecision in our ability to adjust for all benefit traits. This measurement error will lead to an underestimate of the effects of premiums on coverage. Second, there is the potential for reverse causality because of selection, if providers in markets with more uninsured individuals raise prices or shift costs to private insurers. This bias would overstate the effects of premiums on coverage.
To address these dual concerns we estimated linear probability models analogous to the probit model using IV techniques (McClellan, Newhouse, and McNeil 1994
). As an instrument for private premiums we used Medicare Part B and state level per capita spending. The joint significance test on the two spending variables from the first stage regression gives an F
-statistic of 52.66, indicating that the two spending measures are highly correlated with premiums, which is a condition for IV analysis (Staiger and Stock 1997
). In addition, the partial R2
is 0.77, suggesting that the instruments have sufficient explanatory power to identify the system.5