EVENTS

Date:	Friday, April 8, 2005
Time:	9:00 AM -- 11:15 AM
Location:	Wohlstetter Conference Center, Twelfth Floor, AEI 1150 Seventeenth Street, N.W., Washington, D.C. 20036

April 2005

9 Million Fewer Uninsured?

In its latest report, the Census Bureau estimated that there were 45 million Americans without health insurance in 2003. Two analyses sponsored by the U.S. Department of Health and Human Services (HHS) conclude that between 4 and 9 million people reported as uninsured actually have coverage through state Medicaid programs. How good are these estimates? How do they affect federal policy? An April 8 AEI seminar explored methods of improving the estimates of the uninsured.

Michael O’Grady
Department of Health and Human Services

The administration has various initiatives underway that aim to increase access to health care. While the Medicare Modernization Act of 2003 is best known for delivering a drug benefit to seniors, it also established health savings accounts (HSAs), from which tax-advantaged savings can be withdrawn to pay for medical expenses. President George W. Bush has also expanded community health centers, which are an important source of care for many, including immigrants. Since 2002, Trade Adjustment Assistance has provided tax credits for health insurance to workers who have lost their jobs due to international trade or whose pensions have gone bankrupt. The State Children’s Health Insurance Program (SCHIP) continues to ensure widespread access to comprehensive care for children.

In his 2006 budget, the president outlines several new policies to promote access to health care. As in previous years, he proposes tax credits for the purchase of health insurance for individuals who are either self-employed or employed by a small business that does not offer coverage. President Bush also proposes tax rebates for small businesses that contribute to their employees’ HSAs. He is committed to further expansion of community health centers and greater outreach to children who are eligible for Medicaid or SCHIP but not enrolled.

The administration recognizes that the uninsured consists of various diverse subpopulations. They differ in demographics such as income, employment status, race, and age, as well as in their encounters with non-insurance. Some are uninsured for a few months, others for several years. A college graduate without insurance for a few months before beginning a job is very different from an uninsured fifty-five-year-old with a chronic illness. Priorities and policies must be set with an understanding of the diverse profiles among the uninsured.

Four major federal surveys measure the uninsured. Two are sponsored by the Department of Health and Human Services, and two are conducted by the Census Bureau. Disagreement in their estimates creates a tenuous foundation on which to develop policies to help the uninsured. The Current Population Survey (CPS) is the most widely cited of the four surveys. In 2003, the CPS estimated that 45 million Americans were uninsured for the full year, but that figure is much larger than all the other surveys’ estimates. For that reason, some argue that the CPS measures the uninsured at a single point in time, not throughout the year. The administration, on the other hand, believes that CPS measures the full-year uninsured but undercounts Medicaid enrollees.

HHS has partnered with the Actuarial Research Corporation (ARC) and the Urban Institute to devise two separate models to address CPS’s Medicaid undercount. For 2003, the ARC simulation estimated that more than 9 million people enrolled in Medicaid failed to report it, for a total uninsured population of 36 million. It found close to 9 million non-citizens without insurance, which may make ethnic and racial patterns of insurance particularly useful in developing policy. Another 3.5 million of the uninsured were eligible for Medicaid or SCHIP, but not enrolled, which reinforces the need for better program outreach. Almost 10 million people without insurance have incomes greater than 300 percent of poverty. More than 5 million are childless adults, some of whom may be uninsured by choice, but others of whom may be ineligible for public coverage because they have no spouse or children, despite otherwise qualifying need. The Urban Institute model estimated that only 3.6 million of the Medicaid undercount failed to report any coverage and were counted among the uninsured. Other segments of the uninsured population were roughly comparable to ARC’s estimates.

Accurate data on the size and characteristics of the uninsured population must inform policy development. The ARC and Urban Institute models were bipartisan efforts that, while differing on the magnitude of CPS’s Medicaid undercount and the effect of this undercount on the uninsured population, agree on its existence. Even if the uninsured number is actually millions fewer, it remains a significant problem that deserves serious attention.

Cathi Callahan
Actuarial Research Corporation

The CPS estimate of the population “never insured” in calendar year 2001 is double that of the Census Bureau’s Survey of Income and Program Participation (SIPP). Another survey, the Medical Expenditure Panel Survey (MEPS), reports a figure in between, revealing broad disagreement for this measure.

CPS collects a wealth of demographic data for each calendar year. It asks insurance status for any time during the preceding year (not a single point in time, but rather “ever covered”), which is useful for aligning insurance status with other demographic measures.

MEPS and SIPP report similar estimates of the number of Medicaid enrollees for 2001, but the CPS figure is almost 15 million lower than SIPP and approximately 20 million lower than Medicaid program data. CPS is not a health insurance survey, but its reports are released in a timely manner each fall and capture much media attention. Its large sample size also allows detailed subgroup analysis.

In adjusting CPS data, ARC imputed Medicaid coverage to some of the uninsured and to some individuals with other coverage. The ARC model used Medicaid program statistics to construct targets for CPS adjustments. To make the datasets comparable, the data were converted to calendar years, institutionalized beneficiaries were excluded, individuals receiving only partial benefits were omitted, SCHIP was moved to its own category, and the remainder was divided between full-year and part-year Medicaid enrollees.

Demographic data on individuals that CPS reported as Medicaid enrollees were used to profile the undercounted population. Within certain demographic brackets based on insurance status, income, family structure, and work status, the adjusted CPS data were compared to Medicaid targets. For 2003, ARC’s adjustments moved the CPS’s uninsured rate down from 16 percent to 12 percent.

The Medicaid undercount has grown over time, causing increased divergence between census rates of Medicaid coverage and ARC’s adjusted rates. The gap between CPS and ARC uninsured rates, on the other hand, has remained roughly stable over time.

In future analysis, ARC expects to improve the Medicaid targets it constructs through the analysis of Medicaid program micro-data, which would provide enhancements in their estimates of the institutionalized population and enrollees receiving partial benefits, as well as provide state-specific targets.  In addition, enhancements in the ARC methodology, including better baseline profiles and use of new and upcoming research on survey responses, will also improve assessments of the Medicaid and uninsured populations.

Linda Giannarelli
Urban Institute

The Urban Institute’s estimate is derived from the Transfer Income Model, version 3 (TRIM3). TRIM3 is a comprehensive CPS-based microsimulation model of taxes, health insurance, and transfer programs. TRIM3’s Medicaid and SCHIP simulation identified who was eligible for enrollment in those programs and selected a likely caseload from among the eligible individuals, attempting to match the size and characteristics of the actual enrolled population. A reduction in the estimated number of full-year uninsured people is a byproduct of the correction for Medicaid/SCHIP underreporting.

TRIM3 relied on Medicaid program data for a count of non-institutionalized beneficiaries enrolled at any time in calendar year 2001—40.5 million fully covered Medicaid beneficiaries, and 4.6 million SCHIP enrollees. For 2001, CPS found 23.3 million positively reported enrollees in the two programs, and 30.3 million enrollees once corrections were made by the Census Bureau for incomplete responses.

Medicaid and SCHIP enrollment may be underreported because respondents are confused. They may report the wrong type of coverage if, for instance, they are covered by a health maintenance organization (HMO) through Medicaid. They may forget about a few months of coverage or fail to realize that someone in their household is insured. Other respondents may perceive a stigma attached to Medicaid coverage and be reluctant to report it. Low-income people may be under-weighted because they are under-identified by the census. On the other hand, the administrative data to which CPS counts are compared may be biased upward if individuals are kept on Medicaid’s rolls past their actual enrollment, which would appear to increase the CPS undercount.

In the absence of definitive information on the causes of Medicaid/SCHIP underreporting, adjustments to correct for underreporting must make assumptions.  TRIM3 used administrative data to develop average monthly and annual targets for Medicaid enrollment, broken down by state and user group (children, adults, disabled, and the elderly). Census Bureau weights were unchanged, and additional Medicaid/SCHIP enrollees were chosen from anyone meeting state and federal eligibility criteria.  TRIM3’s eligibility simulation is extremely detailed, capturing state-specific variations and capturing all the major avenues to eligibility (including Section 1931, state options, percent-of-poverty, medically needy, and so on).

A total of 58.4 million individuals were identified as eligible for either Medicaid or SCHIP in TRIM3’s 2001 simulation (March 2002 CPS data). Of those, 20.8 million self-reported coverage in the CPS, 24.6 million had other coverage but reported none, and 13 million had no insurance and did not report any.  Of the 58.4 million, eligible individuals who also qualified for welfare or cash assistance were automatically simulated to be enrolled, as were eligible respondents who self-reported (or reported by proxy) Medicaid/SCHIP enrollment. Other eligible individuals were selected at random to meet targets within each state and user group.

Replacing the Census Bureau’s imputation with TRIM3’s and removing Medicaid or SCHIP from apparently ineligible people docked almost 7 million people from the estimated rolls, but other corrections added almost 18 million people to Medicaid and SCHIP. TRIM3’s manipulation of 2001 CPS data revealed 41.2 million Medicaid or SCHIP enrollees (including part-year enrollees), and 37.6 million full-year uninsured Americans. TRIM3’s adjustments revealed a modest drop in the uninsured rate—13 percent, compared to CPS’s 14 percent.

Future TRIM3 simulations should match CPS data to Medicaid program data by more measures than state and user group. The analysis would also benefit from greater consideration of other insurance coverage, and from research on population weighting and Medicaid underreporting. Different methods of adjusting for the Medicaid undercount inevitably produce different estimates of the uninsured. A better understanding of who fails to report Medicaid or SCHIP coverage in the CPS is critical to developing these estimates.

John L. Czajka
Mathematica Policy Research

ARC and the Urban Institute have given us useful enhancements to the CPS data on insurance coverage.  ARC has generated a methodologically consistent CPS time series of health insurance coverage. The ARC and TRIM3 estimates of Medicaid enrollment in the CPS universe are much improved because they provide a better match to program administrative data. ARC and TRIM3’s estimates of the uninsured in 2001 differ in part because the Medicaid enrollment targets they used are discrepant and because TRIM3 falls short of its enrollment target by 2.5 million. Moreover, the two models draw a different share of their imputed Medicaid population from the group that CPS identifies as uninsured. More than half of ARC’s Medicaid population is classified as uninsured by the CPS, compared to less than one-third for TRIM3.

ARC’s and TRIM3’s estimates of Medicaid enrollees who were reported as uninsured should be interpreted with caution: they cannot simply be subtracted from the CPS’s estimate of the uninsured if that estimate continues to be interpreted as point-in-time. Also, comparison with other surveys’ estimates of the full-year uninsured suggests that unreported private coverage may prove to be the larger problem. While the adjusted CPS estimates may identify several million fewer uninsured, they still overstate the total uninsured population. SIPP, which is believed to avoid the bias of Medicaid undercount, still reports 14 million fewer uninsured than the lower, ARC-adjusted CPS figure. This suggests that both the ARC and TRIM3 simulations capture many privately covered individuals in their estimates of the uninsured. It is also important to recognize the various limitations of simulations in general.

The ARC and TRIM3 teams provide analysts with information on the size and characteristics of Medicaid/SCHIP eligible and ineligible populations, as well as useful data on the size and age distribution of the population that is eligible for Medicaid or SCHIP but not enrolled. They also make available estimates of average monthly and annual-ever Medicaid enrollees. The models should not be relied on, however, for characteristics of simulated but not reported Medicaid/SCHIP enrollees or of eligible but not enrolled individuals. Again, their work did not produce an estimate of the total number of the uninsured.

Future enhancements should use available research to develop a model of Medicaid non-reporting. It should consider ways to adjust private coverage estimates using MEPS or SIPP and investigate the recent divergence in the children’s uninsured rates reported by CPS and the National Health Interview Survey.

Michael Davern
State Health Access Data Assistance Center, University of Minnesota

Medicaid beneficiaries do sometimes report being uninsured in surveys, but they do so infrequently. Five recent studies showed that only between 0.6 and 5.8 percent of Medicaid enrollees report being uninsured. The TRIM3 and ARC simulations, however, assume in their models that 9 and 19 percent of Medicaid enrollees report no insurance.

About 85 percent of Medicaid enrollees report Medicaid coverage, so most of the undercount comes from their reports of other coverage, not of a total absence of insurance. People not enrolled in Medicaid can report Medicaid coverage as well. Slightly more than 2 percent of under-sixty-five commercial plan enrollees identify their insurance as Medicaid. Since those individuals comprise the largest group sampled, they account for one-quarter of the self-reported survey Medicaid count. Because of survey design, researchers may often count the uninsured as insured as well. In the CPS, for example, there are nine opportunities to report insurance, which multiplies the possibility of keystroke error or of response bias.  This is multiplied in longitudinal surveys such as the SIPP, where people have three interviews a year to falsely report being insured and/or falsely report having Medicaid coverage.

Because surveys of health insurance coverage are generally good at coding insured people as insured, researchers should be cautious not to overly adjust the uninsured reports collected in surveys. One effective way to impute Medicaid coverage would be for the Census Bureau to match CPS respondents and Medicaid enrollment data.

If Medicaid enrollees fairly reliably report their public coverage, the undercount can perhaps be better explained by sample coverage error, survey non-response, survey weighting, administrative data problems, and measurement errors leading to misclassification of Medicaid enrollees as having other public program or commercial coverage.

Chris L. Peterson
Congressional Research Service

Reliable survey data is critical to developing sound public policy. Most surveys undercount the Medicaid population when compared to administrative data. In 2000, SIPP’s estimate of the total Medicaid population was closest to program rolls, but its age distribution was very different from official statistics. It is important not only to measure more people but also to identify the right people.

A huge range of numbers can satisfy a generic question on the number of uninsured. There is much variation among estimates of non-coverage for an entire year but better agreement among point-in-time estimates. The CPS is fairly in line with other surveys’ point-in-time estimates. With various measures of the uninsured ranging from 20 to 80 million, analysts face three options. They can consider the CPS a point-in-time estimate (though it is not), call for new surveys of the uninsured, or use CPS data adjusted by smart people. While the last option seems most promising, different assumptions and manipulations can produce sharply different results, and excessive imputation simply replaces one bias with another. One-quarter of the Census Bureau’s official uninsured statistic is based not on actual survey responses, but on imputed coverage, and only one-half of ARC and TRIM3’s estimates is derived from actual survey data. Despite the many limitations of the CPS and other surveys, analysts should pursue ways to ground public policy in smart interpretations of the data.

Charles Nelson
Census Bureau

The CPS asks nine questions about insurance to determine survey respondents’ coverage status, including one to confirm non-coverage. When the final verification question was added in 2000, 3.3 million people dropped out of the pool of uninsured. Most of them identified private coverage, and only 300,000 people reported being on Medicaid. Medicaid enrollees, then, may more often misreport their insurance than fail to report it at all.

CPS data are publicly released in a timely manner and are easy to use. The survey captures many years of data and many useful economic indicators, such as types of income and labor force status. However, it does not focus on health insurance, and the long reference period may cause a degree of inaccuracy in reporting. There is little flexibility for adding new content to the CPS, and its uninsured figures are consistently higher than other annual surveys. In 2001, CPS reported that 85.4 percent of respondents had been covered at some point in the previous year, while SIPP reported a significantly greater share, 93.2 percent. Most of the discrepancy came from accounts of private coverage.

In assessing the ARC and TRIM3 undercount adjustments, it is important to consider the adequacy of the controls, misreporting versus not reporting coverage, and the impact of underestimating private coverage. The TRIM3 model is consistent in its use of eligibility rules to both assign and take away coverage, though asset tests may be unreliable because assets are themselves underreported in the CPS. Another advantage of the TRIM3 model is its state-specific imputation of coverage. ARC’s method is useful because it remains interpretable as a time series. Its SCHIP estimate, however, may be overly adjusted.

Future CPS improvements will target the questionnaire and methods of imputation. Match studies may prove particularly useful for the latter.

AEI research assistant Ximena Pinell prepared this summary.