Address correspondence to Richard Kronick, Ph.D., Department of Family and Preventive Medicine, Division of Heath Care Sciences, UCSD School of Medicine, 9500 Gilman Drive-0622, La Jolla, CA 92093-0622; e-mail: firstname.lastname@example.org.
Objective. To improve understanding of the relationship between lack of insurance and risk of subsequent mortality.
Data Sources. Adults who reported being uninsured or privately insured in the National Health Interview Survey from 1986 to 2000 were followed prospectively for mortality from initial interview through 2002. Baseline information was obtained on 672,526 respondents, age 18–64 at the time of the interview. Follow-up information on vital status was obtained for 643,001 (96 percent) of these respondents, with approximately 5.4 million person-years of follow-up.
Study Design. Relationships between insurance status and subsequent mortality are examined using Cox proportional hazard survival analysis.
Principal Findings. Adjusted for demographic, health status, and health behavior characteristics, the risk of subsequent mortality is no different for uninsured respondents than for those covered by employer-sponsored group insurance at baseline (hazard ratio 1.03, 95 percent confidence interval [CI], 0.95–1.12). Omitting health status as a control variable increases the estimated hazard ratio to 1.10 (95 percent CI, 1.03–1.19). Also omitting smoking status and body mass index increases the hazard ratio to 1.20 (95 percent CI, 1.15–1.24). The estimated association between lack of insurance and mortality is not larger among disadvantaged subgroups; when the analysis is restricted to amenable causes of death; when the follow-up period is shortened (to increase the likelihood of comparing the continuously insured and continuously uninsured); and does not change after people turn 65 and gain Medicare coverage.
Conclusions. The Institute of Medicine's estimate that lack of insurance leads to 18,000 excess deaths each year is almost certainly incorrect. It is not possible to draw firm causal inferences from the results of observational analyses, but there is little evidence to suggest that extending insurance coverage to all adults would have a large effect on the number of deaths in the United States.
In a widely cited report, the Institute of Medicine (IoM)'s Committee on the Consequences of Uninsurance estimated that the mortality rate of the uninsured is 25 percent higher than for otherwise similar people with health insurance (IoM 2002). The IoM estimated that 18,000 excess deaths occurred annually because 40 million Americans lacked insurance. Relying on the methods used by the IoM, Stan Dorn at the Urban Institute updated this work to reflect the growth in the uninsured and estimated that there were 22,000 excess deaths as a result of lack of insurance in 2006 (Dorn 2008).
At the time that the IoM published its report, only two studies had been conducted that analyzed the relationship between lack of insurance and mortality in a random sample of adults (Franks, Clancy, and Gold 1993; Sorlie et al. 1994), and the IoM relied heavily on these two studies to estimate that lack of insurance increased the mortality rate by 25 percent. The studies were similar to each other in design. Each started with a random sample of the U.S. population, and linked the survey participants to death certificate information to calculate the probabilities of survival over a follow-up period. Each study used Cox proportional hazards regression to estimate whether lack of insurance at baseline, controlling for other characteristics, was associated with elevated risk of subsequent mortality. Both studies estimated that lack of insurance was associated with approximately a 25 percent increased risk of mortality during the follow-up period. The apparent similarity of the results emboldened the IoM committee to conclude that lack of insurance was associated with a 25 percent increase in risk.
There are three reasons to be skeptical of the IoM conclusion that universal coverage would result in an annual reduction of 18,000 deaths. First, the 95 percent confidence intervals (CI) in each of the two studies were large, stretching from approximately no effect to increased mortality of 50 percent. Thus, even ignoring the next two concerns, we could be pretty sure that the right answer is somewhere between 0 and 36,000 excess deaths.
Second, although the two studies were similar in general design, they were different in a crucial detail: the analysis using National Health and Nutrition Examination Survey (NHANES) data controlled for self-reported health status and smoking behavior, while these two variables were not available in the Current Population Survey (CPS) data. As I show below, if the study using CPS data had been able to control for self-reported health status and smoking behavior, the authors would almost certainly have concluded that their best estimate was that there was no difference in the survival probabilities of otherwise similar insured and uninsured persons. Given the wide CIs in both studies, the estimated null effect in the CPS study would not have been inconsistent with the estimate of a 25 percent effect in the NHANES study, but undoubtedly would have given the IoM Committee members pause in deciding whether 0 or 25 percent or some other number was the best estimate of the excess mortality due to lack of insurance.
Third, as has been discussed by Levy and Meltzer (2004), it is difficult to draw inferences about the causal relationship between lack of insurance and mortality from the kinds of observational analyses relied upon by the IoM. The inferential difficulties inherent in observational studies lead us to prefer evidence from quasi-experimental or experimental designs. However, as discussed below, there is no experimental evidence and virtually no quasi-experimental evidence about the relationship between health insurance and mortality.
In this paper, I replicate the earlier observational studies, making three major improvements. First, I use a dataset that is substantially larger than that in the previous work, providing a much more precise estimate of the association between lack of insurance and mortality. Second, I explicitly consider the implications of alternative choices for the variables that are included on the right-hand side of the model. Third, I probe more deeply into understanding the associations between lack of insurance and mortality with a variety of subsidiary analyses.
The next section reviews the results of previous work that attempts to estimate the relationship between lack of insurance and mortality and discusses the difficulties in drawing inferences about causality from the results of observational analyses. Subsequent sections present data and methods, results, and discussion.
A large body of research analyzes the relationship between lack of insurance and morbidity (Hadley 2003). Almost all of this work shows that the uninsured have worse health outcomes than the insured, although methodological difficulties plague many of the studies, and the magnitude of the estimated effects varies greatly with the outcome, the population, and the study methods.
At the time of the IoM report in 2002, only two published studies analyzed the relationship between lack of insurance and mortality in a random sample of adults. Key features of the two studies relied upon by the IoM in reaching the conclusion that lack of insurance led to a 25 percent increase in mortality risk are summarized in Table 1. As described above, the two studies were similar in design, differing primarily in the variables included on the right-hand side of the Cox proportional hazard regression—in particular, the inclusion of self-reported health status and smoking behavior in the NHANES study, and the exclusion of these variables from the CPS study.
Table 1. Comparison of Population-Based Studies of Lack of Insurance and Mortality
BMI, body mass index; CPS, Current Population Survey; HRS, Health and Retirement Survey; NDI, National Death Index; NHANES, National Health and Nutrition Examination Survey; NHIS, National Health Interview Survey; NLMS, National Longitudinal Mortality Survey.
Risk of mortality among uninsured, compared with privately insured (95% confidence interval)
White employed men 1.3 (1.0–1.6)
White employed women 1.2 (0.8–1.7)
Black men 1.5 (1.0–2.1)
Black women 0.8 (0.5–1.4)
Subsequent to the publication of the IoM report, two studies analyzed data on 51–61-year-olds interviewed in the 1992 Health and Retirement Survey (HRS) (McWilliams et al. 2004; Baker et al. 2006). These studies used a similar design to the previous work, and controlled for self-reported health status and health behaviors in addition to basic demographic and socioeconomic factors. Results from the HRS-based studies are consistent with results from the Franks study, but different from the results that Sorlie would have produced if the CPS data had included information on self-reported health and smoking status.
Helen Levy and David Meltzer assert that observational studies are “not likely to provide much insight into the causal effects of health insurance on health” (Levy and Meltzer 2004), and this review of the Franks, Sorlie, McWilliams, and Baker studies illustrates some of the reasons for their assertion.
One vexing problem is the question of which covariates should be included on the right-hand side of the analysis. The Franks, McWilliams, and Baker studies each included health and smoking status on the right-hand side of the model, while the Sorlie study did not. These covariates were excluded from the Sorlie study because they were not available in the CPS data, but it is not clear that they should be included in the analysis even when the data are available. If lack of insurance causes people to be in poorer health, or makes it more likely that they cannot quit smoking, then including these variables in the analysis will bias downwards the estimated effect of lack of insurance on mortality. Conversely, if they are excluded from the analysis, and if uninsured and insured differ in health status or smoking behavior for reasons that are not causal effects of lack of insurance, then the estimated effect of lack of insurance will be biased upwards.
An even more vexing problem is the difficulty of drawing firm inferences from observational analyses given the likelihood of omitted variables and the problems created by reverse causality—not only might insurance affect health, but health status affects the desire for insurance and the ability to become insured. If we find that survival probabilities are lower for uninsured respondents than for similarly situated respondents with insurance, we cannot be sure whether those differences are a causal result of not having insurance, or whether there are unmeasured factors that are associated with lack of insurance and causally related to mortality.
Plausible hypotheses can be constructed regarding a variety of omitted variables. For example, characteristics such as willingness to take risks, amount of value placed on health, amount of value placed on health care, ability to pass an underwriting screen in the nongroup market, and the levels of job-related, family-related, and environmental stress are plausibly causally connected to survival probabilities. These characteristics are not well measured in any of the surveys used in the analyses reviewed here, and are likely related to lack of insurance. As a result of omitting these covariates from the multivariate analysis, any observed association between lack of insurance and mortality may not reflect a causal mechanism of being uninsured, but rather the effects of these omitted variables on mortality.
Given the difficulties in inferring causality from the results of observational studies, we would much prefer to have evidence from environments in which there is an exogenous source of variation in insurance coverage. The best evidence to assess the effects of insurance on mortality would come from a randomized experiment, in which the only difference between the insured and the uninsured is that some people were randomly assigned to insurance, and others to being uninsured. No such experiment exists. Second best would be evidence from quasi-experiments, in which the reason that some people are insured and others uninsured has nothing to do with choices that people made, but rather is the result of some exogenously made decision. Unfortunately, evidence from the results of quasi-experiments is limited.
The existence of nearly universal coverage at age 65 in the United States provides a natural experiment. Large numbers of previously uninsured people become insured through no decision of their own. If health insurance improves the probability of survival, then mortality rates should change discontinuously at age 65 as previously uninsured people become insured. However, three separate analyses, each using somewhat different approaches, demonstrate that Medicare has no effect on the survival prospects for 65-year-olds (Card, Dobkin, and Maestas 2004; Finkelstein and McKnight 2005; W. H. Dow, unpublished data). An earlier analysis (F. Lichtenberg, unpublished data) suggested that Medicare did lower the mortality of 65-year-olds, but W. H. Dow (unpublished data) demonstrates that the Lichtenberg result was an artifact of the aggregation process that SSA used to construct the life tables used in that analysis.
A recent analysis by Card and colleagues suggests that Medicare does improve survival for a subset of hospitalized patients in California (Card, Dobkin, and Maestas 2007). However, the authors conclude that, “… we suspect that the measured mortality effects arise because Medicare imposes fewer restrictions than private insurance or Medicaid, leading to more (and possibly higher-quality) services to Medicare patients.” The results are not directly useful in anticipating the effects on mortality of transitioning the under-65 uninsured to Medicaid or private insurance.
Three of the four observational studies of the association between lack of insurance and mortality suggest that mortality is higher for the uninsured than for similar insured persons, although the CIs in these studies are wide. There is no experimental evidence that addresses the question. The broadest quasiexperimental evidence shows that mortality for all those aged 65 does not decline as a result of nearly universal coverage under Medicare. Given the level of uncertainty generated by the observational and quasi-experimental results, analysis of a recently available and very large observational dataset will improve our understanding of the relationship between lack of insurance and mortality.
The analysis uses data from the National Health Interview Survey (NHIS) and the NHIS Linked Mortality Files. The NHIS is a multistage probability sample of the U.S. civilian noninstitutionalized population conducted by the National Center for Health Statistics (NCHS). NCHS statisticians have matched adult respondents in the 1986 through 2000 NHIS samples with the 2002 National Death Index (NDI; NCHS 2005).
Children were not followed for mortality and are not included in this study. This study excludes the 1987 and 1988 surveys, because questions on health insurance were not asked in those years.
The combined 1986 and 1989 through 2000 NHIS contains 771,466 respondents, age 18–64. Following the strategy of other researchers, the analysis excludes 98,940 (12.8 percent) respondents with public insurance (Medicare, Medicaid, Champus, VA, and other public coverage) or with unknown insurance status. Nonelderly adults obtain Medicare, and, in many cases, Medicaid coverage as a result of being sick or disabled, and it is difficult to disentangle the reason for obtaining public coverage from the effects of this coverage on health.
Of the 672,526 adult respondents who were either privately insured or uninsured, the analysis excludes 29,525 (4.4 percent) who were ineligible for linking to the NDI because they did not provide social security numbers or other information required for a match. The exclusions average 2–3 percent per year until 1997, and approximately 10 percent per year in 1997 and beyond. The remaining 643,001 respondents were followed for between 2 years (for 2000 interviewees) and 16 years (for 1986 interviewees).
The main analysis is a Cox proportional hazard regression predicting survival time. The analysis tests whether respondents who were uninsured when surveyed had different survival probabilities than otherwise similar privately insured respondents. Covariates include factors known to be related to survival (Table 1).
NHIS questions on insurance status and on some covariates changed slightly across years, particularly with a 1997 redesign. All variables are recoded to be as comparable as possible across years.
Questions on current and former smoking status were not asked in 1986, 1989, or 1996. In the 1990 through 1995 surveys, questions on smoking were asked of a randomly selected subset of respondents in a supplement to the NHIS. In 1997 and subsequent years smoking status, height, weight, and a range of other information were asked of a randomly selected sample adult in each household. The analyses including smoking status as a covariate are limited to 196,371 respondents and to 12 years of follow-up.
Health status is measured using the Health and Activity Limitation Index (HALex), a generic measure of health-related quality of life on a 0–1.0 continuum. HALex consists of two attributes: self-reported health status (excellent, very good, good, fair, or poor) and activity limitations (Erickson 1998; Erickson, Wilson, and Shannon 1999).
Approximately 15 percent of respondents did not provide information on family income, and are included in the analysis with an “income unknown” covariate. Before 1997, information on income was top-coded at US$50,000; as a result, the highest income category is “greater than 200 percent of the federal poverty level (FPL).” To improve measurement of economic status, covariates are included for “living in a mobile home” and for “no telephone in the household.”
All analyses adjust for the multistage, complex survey sample and use revised sampling weights that were created when NHIS was merged with NDI. Because the NHIS years span a sample redesign, it is difficult to analyze the entire dataset using SUDAAN. Within each design era, the parameter estimates using SUDAAN and SAS are virtually identical, and the parameter estimates from SUDAAN analyses are almost uniformly no more than 10 percent larger than those from SAS. The results below are from SAS, with the standard errors inflated by 10 percent as a conservative estimate of the design effect.
Approximately 20 percent of the sample reported not having health insurance at baseline (Table 2). On almost every characteristic measured, the uninsured are in higher risk groups. The uninsured are more likely to be low income, living without a telephone or in a mobile home, not in the labor force, poorly educated, in poorer health, and current smokers.
Table 2. Relationships among Health Insurance Status, Baseline Characteristics, and Unadjusted Mortality Rate
Number of Deaths
Cumulative Mortality Rate
Number of respondents
Table includes adult respondents, age 18–64, at the time of interview, who reported being privately insured or uninsured. Cumulative mortality rate is the fraction of respondents with a matching record in the 2002 NDI. For smoking and BMI, survey years are 1990–1995 and 1997–2000.
Source: Health Interview Surveys, 1986, 1989–2000, linked to the 2002 National Death Index (NDI).
BMI, body mass index; FPL, federal poverty level.
Not in labor force
Cigarette smoking, N=196,334
By the end of the follow-up period, 23,667 (3.5 percent of the weighted sample) had died. Controlling for age, the mortality rate is greater for uninsured than for insured respondents (Figure 1).1
However, after adjustment for the characteristics shown in Table 1, lacking health insurance at baseline is not independently associated with an increased risk of mortality (hazard ratio 1.03, 95 percent CI, 0.95–1.12) (Table 3).
Table 3. Proportional Hazards for Survival Time, Adjusted for Baseline Characteristics
Source: Health Interview Surveys, 1990–2000, linked to the 2002 National Death Index.
BMI, body mass index; CI, confidence interval; FPL, federal poverty level.
Race and ethnicity
Less than HS
MA degree or more
Family income and employment status
No telephone in house
Live in a mobile home
Not in the labor force
Unweighted number of respondents
Other baseline characteristics are associated with probability of survival in expected directions: respondents who are male, older, African American, not married, with less education, not in the labor force, in poorer baseline health, and who smoke have higher mortality risk than others.
It is difficult to believe that lack of insurance is not associated with a greater risk of mortality, and I conduct a variety of subsidiary analyses in an attempt to understand whether the finding is robust to alternative analytic approaches.
The null effect observed in Table 3 may result, in part, from changes in insurance status during the follow-up period—some of those who are uninsured gain coverage, while some of the insured lose coverage. Two analyses explore whether changes in insurance status might account for the lack of effect. First, the follow-up period is shortened to increase the likelihood of comparing people who were continuously insured and uninsured. There is no indication that the relationship between lack of insurance and mortality is greater when the follow-up period is shorter (Table S1). Second, in 1993 and subsequent surveys, the NHIS asked uninsured respondents how long they had been uninsured. Respondents who were uninsured for longer periods of time before the interview are more likely than others to remain uninsured subsequent to the interview. If changes in insurance status after the interview attenuate the estimated effect of lack of coverage, the estimated effect should be larger among respondents reporting longer periods of uninsurance. However, there is no indication of a dose–response relationship between time uninsured and probability of survival (Table S2).
Another possible explanation for the null effect is that many causes of death cannot be prevented by better health care. I reestimate the basic model, but limit the causes of death to causes thought to be amenable to better health care (Nolte and McKee 2003). These “amenable” causes, including pneumonia, influenza, hypertension, diabetes, and cancers of the breast, cervix, and colon, account for approximately 15 percent of the deaths in the sample. There is no indication that lack of insurance has any effect on this subset of deaths (Table S3). Nolte and McKee suggest that ischemic heart disease, which accounts for approximately 8 percent of the deaths in the sample, might potentially be included in the list of “amenable” causes. Replicating the analysis after classifying ischemic heart disease as “amenable” produces virtually identical results (data not shown).
Lack of insurance may have no effect on survival probabilities for the entire population, but might matter for those who are most vulnerable. I estimate the basic model in Table 3 on subsets of the population—respondents with low income, low levels of education, those not in the labor force, those in poor health, those who are smokers, those who are 50–64 years at baseline—and find no significant effect of insurance among any disadvantaged subset (Table S4).
I also investigate the relationship between lack of insurance and mortality after previously uninsured respondents turn 65. We might expect that the difference in mortality rates between the uninsured and insured in Figure 1 would narrow after people reach age 65, when Medicare coverage is virtually complete. However, the relationship between lack of insurance at baseline and mortality does not change after respondents turn 65 (Figure S1), although relatively small numbers of under-65 NHIS respondents who are 65 and over during the follow-up period limit the power to observe a significant effect of Medicare on mortality. To test whether including deaths after age 65 affects the estimated relationship between lack of insurance and mortality, I censor the follow-up period when respondents reach age 65 and reestimate the model in Table 3. The results are no different from the uncensored model (Table S5).
As anticipated by the discussion above, the estimated magnitude of the relationship between lack of insurance and mortality does depend on the variables that are controlled for in the analysis. When baseline health status is removed from the model, the hazard ratio for lack of insurance increases from the 1.03 estimated in Table 3 to 1.10 (95 percent CI, 1.03–1.19) (Table 4). Omitting smoking status and body mass index from the model increases the hazard ratio to 1.20 (95 percent CI, 1.15–1.24). Omitting labor force participation increases the hazard ratio to 1.25, and omitting income increases it further to 1.37. Controlling only for age and gender, the hazard ratio is 1.71 (95 percent CI, 1.65–1.76). In the discussion below I consider the implications of these results.
Table 4. Proportional Hazards Estimates of the Lack of Insurance on Survival Time, Controlling for Baseline Characteristics
Covariates Included in the Analysis
Hazard ratio is the estimated odds ratio from a Cox proportional hazards regression, controlling for the variables listed. Each model includes all of the variables in the models above—that is, the model with race and ethnicity also includes age and gender; the model with education also includes age, gender, race, and ethnicity. Smoking status and BMI are only available for a subset of respondents; estimating the model including all variables through labor force participation on this subset produces an odds ratio on being uninsured of 1.22, very similar to the 1.20 estimated on the whole sample.
Source: 1986–2000 NHIS (except 1987 and 1988) merged with the NDI. N=642,734 for regressions including all variables through labor force participation. N=196,371 for the last two regressions which include smoking status.
BMI, body mass index; CI, confidence interval; NDI, National Death Index; NHIS, National Health Interview Survey.
Race, ethnicity, immigration status, veteran status
Family income, telephone in home, mobile home
Labor force participation
Health and activities
Limitation index (HALex)
Comparison to Previous Work
The results presented here are largely consistent with results from Sorlie and colleagues. Sorlie estimated a hazard ratio of 1.3 (95 percent CI, 1.0–1.6) for white males in a model controlling for age and income; in a similar model using NHIS data, I estimate a hazard ratio of 1.43 (95 percent CI, 1.34–1.53) (data not shown).
The results here appear to be different from the point estimate in the Franks study, but the 95 percent CIs overlap. The Franks study included privately insured and uninsured respondents who were 65 and older at the time of the interview, while my work excludes all respondents who were over 65 when interviewed, because almost all of them are covered by Medicare, and the few who are not are clearly different from those who are. Although there were relatively few over-65 non-Medicare respondents included in the Franks' study, approximately one third of them were uninsured, and their mortality rate was very high, possibly accounting for some of the difference between the results of the Franks study and the results reported here (P. Franks, personal communication, September 2, 2006).
Similarly, the point estimate in this study appears to be different than the point estimates in the studies using data on 51–61-year-olds from the HRS studies. Without adjustment for covariates, the crude relative risk in the NHIS data is 1.5 (n=46,999; 95 percent CI, 1.38–1.53), compared with a crude relative risk of 1.77 (n=8,789; 95 percent CI, 1.50–2.07) in the HRS data (Table S6). Adjustment for covariates has virtually identical effects in both datasets, decreasing the point estimate of relative risk to 1.13 in the NHIS data and 1.35 in the HRS data.
The difference between the unadjusted mortality rates of the insured and uninsured is smaller in the NHIS data than in either the NHANES or the HRS data. I am unaware of any systematic reason why results would be different in the NHIS data.
The IoM concluded that lack of universal coverage in the United States led to an excess of 18,000 deaths per year. That conclusion is almost certainly incorrect. It was based on two observational studies. Each of those studies analyzed survival probabilities for a sample of respondents, some of whom were insured, and others of whom were uninsured at the time they were interviewed. The researchers asked whether respondents who were uninsured at the time of the interview were less likely to survive during a follow-up period than similar respondents who were insured at the time of the interview. Both studies estimated that the uninsured were 25 percent less likely to survive than similar respondents who were insured, but the studies differed in how they defined similar—one study, using data from NHANES, controlled for self-reported health status and smoking behavior, while the other study, using data from the CPS, did not. If the study using CPS data had been able to control for self-reported health status and smoking behavior, it would almost certainly have found no difference in survival between the uninsured and similar insured respondents. The sample sizes in each study were relatively small, and CIs in each study were wide.
The results reported here replicate the previous work, but with a much larger sample, and a more nuanced approach to the analysis. These results demonstrate that if two people are otherwise similar at baseline on characteristics controlled for in the model presented in Table 3, but one is insured and the other uninsured, their likelihood of survival over a 2–16-year follow-up period is nearly identical. Further, I show that survival probabilities for the insured and uninsured are similar even among disadvantaged subsets of the population; that there are no differences for long-term uninsured compared with short-term uninsured; that the results are no different when the length of the follow-up period is shortened; and that there are no differences when causes of death are restricted to those causes thought to be amenable to the quality of health care. These results are based on analysis of approximately 640,000 respondents with over 5 million person-years of follow-up.
This observational analysis has a number of limitations. The uninsured are disadvantaged in many ways (Table 2), and if some of those disadvantages result from being uninsured, lack of insurance has an indirect effect on mortality. For example, if lack of insurance in the period before being interviewed causes people to be sicker than they would be had they been continuously insured, that would both be an important effect in and of itself, and an avenue for an indirect effect on mortality.
Unmeasured differences between the insured and uninsured may mask the true effects of insurance on the prospects for survival. If the uninsured are healthier than the insured in ways not measured by the factors controlled for in Table 3, then these results will underestimate the magnitude of the causal relationship between lack of insurance and mortality. However, on almost all measured characteristics the uninsured are at higher risk than the insured. It seems likely that on unmeasured characteristics the uninsured are also at higher risk than the uninsured, although it is possible that the opposite is true.
There are two main conclusions from this work. First, the unadjusted difference between the morality rate of the insured and uninsured varies across datasets: the unadjusted difference in survival probabilities for the insured and uninsured is similar in the NHIS and the CPS data, and this difference is smaller than in the NHANES and in the HRS data. Given the much larger sample size in the NHIS analysis, it makes sense to put more weight on these results than on the results from other datasets, but the differences across the datasets inevitably create uncertainty. Second, the strength of the association between lack of insurance and mortality is highly dependent on the covariates included in the model—an alternate title for this paper might be “Lack of insurance and mortality: you get what you control for.” When self-reported health status is included as a covariate, there appears to be no difference in the survival probabilities of uninsured and similar insured persons. With health status excluded, the uninsured have a 10 percent higher mortality rate than similar insured persons.
Given the inherent uncertainties in inferring causality from the results of observational analyses, the results presented here are not able to provide a definitive answer to the question, “How many fewer deaths would there be in the United States if all residents were continuously covered by health insurance?” However, in the absence of any experimental results and with only fragmentary and inconclusive quasi-experimental results, the observational results reported here provide the best available evidence to formulate a tentative answer to that question. The answer suggested by the evidence presented here is that there would not be much change in the number of deaths in the United States as a result of universal coverage, although the difficulties in inferring causality from observational analyses temper the strength of this conclusion.
The results presented here are counterintuitive. The uninsured use about one-half as much health care as they would if they were insured, are less likely to have a medical home, and are much more likely to report not getting needed care because of cost (Ayanian et al. 2000; Hadley and Holahan 2003). Although not all medical care contributes directly to better health outcomes, some of the care that is not received by the uninsured (but that would be received by similar insured persons) should contribute to health and chances of survival. It is not clear why this common sense notion is at best only weakly supported by the analysis.
Part of the answer may be that the safety net catches some uninsured people before illness and restricted access to medical care lead to premature death. Each year approximately 750,000 disabled people gain Medicaid eligibility and a similar number become Medicare beneficiaries. The mortality rate of the uninsured would almost certainly be greater if the Medicaid and Medicare programs did not save some of them.2 And in some communities a viable system of public hospitals and community clinics may provide “good enough” access to care for the uninsured to keep their mortality rate similar to that of the insured.
This paper has focused on gaining a better understanding of the relationship between health insurance and mortality, and it has not analyzed a variety of other potential positive effects of continuous coverage on the currently uninsured, including effects on morbidity, financial security, mental health, and job lock. There are both good theoretical reasons as well as reasonable evidence to suggest beneficial effects of health insurance in many of these areas (Hadley 2003).
The results of this work strongly suggest that arguments in favor of universal coverage should not focus on the beneficial effects of that policy on the life expectancy of the currently uninsured. It makes more sense to turn the question on its head and ask, “What benefits are there to our economy or our society from a system that allows 45 million Americans (and growing) to be without coverage?” A set of well-worn arguments in economics considers the tradeoffs between equity and efficiency, noting that policies that increase equity often do so at the cost of efficiency. But the American health insurance conundrum stands these arguments on their head—by most standards we have the most inefficient health care system in the developed world, as well as the most inequitable (Garber and Skinner 2008). Achieving universal coverage in the context of a more sensible and equitable system of health care financing will not magically solve the problems of inefficiency or reduce the unsustainable growth rate of health care expenditures, but arguably would be a good start.
Joint Acknowledgment/Disclosure Statement: This work was supported by a National Center for Health Statistics/Academy Health Fellowship. Chris Cox at the National Center for Health Statistics provided invaluable help in using the linked National Health Interview Survey/National Death Index files. I would like to thank Jane Sisk, Sandy Decker, Hal Luft, Willard Manning, and Larry Casalino for their helpful comments on earlier drafts.
1 The mortality rates in Figure 1 are for single years of age—for example, among respondents who were insured when they were interviewed, the mortality rate at age 42 was 0.2 percent. Respondents are in the denominator of the mortality rate calculation multiple times—for example, a respondent in the 1986 NHIS who was age 35 at the time of interview and who was still alive in 2002 would be in the denominator of the calculation for each age from 35 through 51. If that respondent died, in, for example, 1995, then he would be in the numerator of the calculation for 44-year-olds.