Sci. Aging Knowl. Environ., 6 April 2005
Vol. 2005, Issue 14, p. pe9
[DOI: 10.1126/sageke.2005.14.pe9]


Future Mortality: A Bumpy Road to Shangri-La?

Shripad Tuljapurkar

The author is in the Department of Biological Sciences at Stanford University, Stanford CA 94305, USA. E-mail: tulja{at}

Key Words: health • mortality • risk factor • obesity • life expectancy


Human lives have lengthened dramatically over the past two centuries. Fig. 1 displays the period life expectancy at birth (e0, the average age at death assuming mortality rates for a particular year apply throughout life) for females in Sweden (1, 2) from 1751 to 2003. Change was modest through the 1700s and early 1800s; in about the mid-1800s, a sustained increase began and has continued until the present. The value of e0 rose from 53.6 years in 1900 to 72.4 in 1950 and then to 82.4 in 2003, so over the past century the rate of change slowed from 0.38 years per calendar year in the first half to 0.22 years per calendar year in the second half. Similar gains have been made in all industrialized countries and have been followed in the last several decades by gains in life expectancy in most of the developing world. Long lives and extended aging are now important features in the human prospect and drive a great deal of research in health, demography, and other fields. The future course of life expectancy matters to individual and public welfare, as does the relation between health and life expectancy. In a recent paper, Olshansky et al. argue that a widespread increase in obesity in the United States implies a worsening of health and a reduction in life expectancy, which they estimate at between 0.3 and 0.75 years (3). They suggest that this obesity "epidemic" may substantially reduce future increases in e0 for the United States and even lead to a decrease in life expectancy over the next few decades. How strong are their arguments and the implications they draw from them? Here, we provide a context within which we evaluate several of the arguments they make.

View larger version (14K):
[in this window]
[in a new window]
Fig. 1. Period life expectancy for females in Sweden from 1751 to 2003. Note the absence of much change until the mid-1800s and the concavity that appears in the last half century.

First, we consider what can be learned from trends in life expectancy and mortality. There is much to learn from history, even though Olshansky et al. criticize extrapolation as a method of projecting life expectancy (in fact, they end up doing a lot of extrapolation themselves). Second, we discuss the slippery concept of health and the relation between health and mortality risks, as well as the relation between individual attributes and mortality risk. Finally, we evaluate the implications for U.S. Social Security and Medicare.

Trends in Life Expectancy and Mortality

We have pointed out the slowdown in the recent rate of increase in e0 for Sweden, which can also be seen as a concavity of the later section of the curve plotted in Fig. 1. This slowdown is a consequence of the historical pattern of decline in mortality, in which mortality at the youngest ages fell faster and earlier than at later ages. Olshansky et al. assert that this pattern necessarily limits increases in e0 over the coming decades. It is certainly true that mortality reduction at a given age affects e0 partly via a gain in years lived after that age, so reducing mortality at age 65, for example, has less numerical effect than reducing mortality at age 5. However, there is still considerable scope for mortality improvement--in 1999, the probability of dying by age 65 was 0.18 in the United States and 0.09 in Sweden and by age 75 was 0.36 and 0.21, respectively. Clearly many lives could be lengthened by reducing mortality at later ages. A related point is that in the next few decades life expectancy can expand significantly whether or not the maximum age at death increases. When Olshansky et al. say that "another quantum leap in life expectancy can only occur if the future is different from the past" [(3), p. 1139], they are decrying possible increases in e0 at the rate of the past 50 years, which was perhaps 0.22 years per calendar year. This amounts to 2.2 years per decade, which would put the U.S. female e0 by 2050 at about 90 years--hardly a quantum leap, given that e0 values for females in Japan and Sweden are at 84 years today.

Extrapolations and correlations are both potentially hazardous statistical procedures, and Olshansky et al. are right to express skepticism about the former, although their analysis rests on both. For mortality change, however, history provides us with a way of testing projection methods, and the Lee-Carter method used by several of the authors cited by Olshansky et al. does very well in such tests (4). This method is based on an empirical pattern: In industrialized countries, mortality rates at all ages have fallen at a nearly constant exponential rate over the past half century. The Lee-Carter method predicts future increases in e0 at close to historical rates, without assuming an increase in the maximum age at death. A distinguishing feature of this method, incidentally, is that it is not extrapolative by design; rather, it begins with a general procedure for extracting the dominant time trends, if there are any to be found, from a time series of multivariate data. When applied to mortality rates, the procedure yields a clear dominant trend that happens to be a steady linear decline in the logarithms of mortality rates at all ages (5).

A final point about mortality history is that the short-term past (e.g., 10 years) of mortality or life expectancy is a thoroughly unreliable guide to the near or distant future. Consider Fig. 2, which displays annual changes in female e0 for the United States. It is apparent that there has never been a decade of "steady" gains and that most decades include years (in the 1960s, the mid-1980s, and the 1990s) in which e0 actually declined or in which it barely increased. Sweden's history of change (Fig. 3) is similarly volatile; by the way, there is no apparent correlation between negative or slow periods in the two series. Despite these short-term slowdowns and even reversals, the overall change in e0 over a few decades has been resolutely positive in both countries. Thus, actuaries and demographers have advised the use of long historical periods to avoid the myopia caused by looking too closely at a particular short run (4, 5). However, Olshansky et al. are firmly myopic when they choose to extrapolate--thus, they criticize the U.S. Social Security Administration (SSA) actuaries for not being worried that the rise in U.S. life expectancy appeared to have stalled in the 1990s [(3), p. 1142].

View larger version (25K):
[in this window]
[in a new window]
Fig. 2. Annual changes in period life expectancy for U.S. females from 1950 to 1999. Note the location of the zero on the vertical axis and the volatility of change, which is negative in several years over the period. The cumulative change over the period is positive.


View larger version (26K):
[in this window]
[in a new window]
Fig. 3. Annual changes in period life expectancy for Swedish females from 1950 to 2003. Note the location of the zero on the vertical axis and that change is negative in several years over the period.

Health and Mortality Risks

Health, like art, is something we presumably know when we see it, or perhaps recognize best in its absence. We can certainly assess particular dimensions of health, such as resting metabolic rate or some aspect of blood chemistry. However, health in the large, as a unitary concept, is difficult to define operationally, so we usually assess health in terms of proxies that capture broad aspects of health. Mortality risk is one natural proxy: Healthier people should have lower mortality risks. Mortality risk is revealed only when people die, though, so we must match the attributes of those who die to living individuals and assign risk to the attributes. We are thus firmly set on the not-so-firm ground of correlation analysis. At least two difficult questions arise: Which attributes should we use, and how do we test for important but missing attributes?

The evidence for a correlation between obesity [measured by body mass index (BMI), which is defined as (mass in kg)/(height in meters)2] and mortality risk is substantial, as is the evidence for a rise in obesity in the United States (3) (see Mizuno Review). Most analyses, including that underlying the work by Olshansky et al. (6), measure how much an attribute like obesity changes the mortality risk relative to some reference (such as a population average or a designated reference value of the attribute), but there is considerable evidence that other individual attributes influence the relative mortality risk of people with different levels of obesity. For example, physical fitness and activity affect relative mortality risk at any level of obesity (7, 8), so it is not known how much variation there might be between risks for people who are equally obese but differ in fitness (9). There is also evidence that obesity affects relative mortality risk differently in older people than in younger ones, perhaps because extra body mass plays a protective role in older people with chronic ailments (10). Olshansky et al. use obesity as a single factor to determine relative risk and assign the same optimal BMI (that is, the BMI with the lowest mortality risk) at all ages. How can we assess the likely error in such an approach?

Consider two groups of individuals, "thin" and "obese," and focus on the differences between individuals in the age at death (rather than the risk of death at a particular age). The total variance in the age at death (over all individuals) decomposes into (i) an average of within-group variances, plus (ii) a between-group variance that depends on the difference between average ages at death. If the between-group variance dominates, then the difference between averages is what matters; if within-group variance dominates, then we are missing important attributes that significantly affect the result. We do not know what the result of such a decomposition would be for obesity, but for two other single factors with a strong correlation to relative risk--education and income--we find that within-group variance dominates between-group variance (11). Given this result, and what we know about the interaction of other attributes with obesity, we can reasonably suspect that within-group variances for obesity groupings could be large. Thus, the consequences of obesity estimated by Olshansky et al. are likely to be quite uncertain. We return to this question when we discuss the fiscal implications of their result.

Another way of assessing risk factors is to look at deaths by causes that are correlated to those risk factors and then use cause-specific death rates to construct a trend in overall mortality. Such an enterprise founders on correlations between causes of death, which are often attributed to their dependence on the same risk factors. Obesity is a correlative factor for a long list of causes of death, which makes cause-specific analysis particularly difficult. Olshansky et al. make reference to an earlier study by Olshansky, Carnes, and Cassel in which they made the remarkable claim that the elimination of most major causes of adult death would have a very modest effect on U.S. life expectancy (12). That result is mysterious: If life expectancy doesn't change, something is killing people; if we have eliminated what were considered to be the major causes of death, undetected causes must remain.

Long-Term Mortality Decline

It is certainly plausible that obesity measures aspects of individual mortality risk relative to a population average (even if imprecisely). So do characteristics like education, income, place of residence, and so on. We can imagine a multidimensional space of such risk factors, with individuals being distributed in this space and arranged by relative mortality risk, but history tells us that over time the absolute average mortality risk has changed, so the entire distribution rides on some trajectory of changing average risk. Olshansky et al. do not care about this secular drift (that is, drift taking place over a long period of time), focusing only on the distribution of individuals by obesity. However, shifts in absolute risk and the distribution of relative risks do interact--the question is how? One answer to this question comes from looking at the change with time in the population variability of age at death. We live in a period of low infant and child mortality, so the variability that matters is in adult death. To describe this, we consider only deaths that occur at or after an age of 20 years and compute the standard deviation S20 in ages of adult death (11). Using Swedish data on life expectancy since 1950, Fig. 4 shows that S20 has declined substantially as life expectancy has gone up. In other words, as absolute average mortality risks have declined, so have the mortality risks for all subgroups. We conclude that mortality decline is indeed a "rising tide that lifts all boats." This finding argues against the possibility that Olshansky et al. raise, that shifting risk factors might indeed come to alter the temporal increase in life expectancy.

View larger version (11K):
[in this window]
[in a new window]
Fig. 4. The variance in age at adult death measured by the standard deviation S20 in ages of adult death (that is, deaths occurring after age 20 years) plotted against life expectancy for Swedish females since 1950. Compare Fig. 1 to see that we move to the right as time passes and life expectancy rises and that S20 falls over time.

We thus come to a long-standing and unsolved question: What are the determinants of secular mortality decline? McKeown showed many years ago that secular decline is driven by factors such as public health and the standard of living, and Preston detailed the contributions of economic determinants (13, 14). In terms of aggregate factors, our understanding has not advanced much further. Only one theory ties secular change in mortality to risk factors--Fogel's argument that secular mortality decline reflects a long-term secular shift in physical attributes (the population distribution of BMI) and their relation to mortality (the optimal BMI) over time (15). In the longer historical perspective, humans have become longer lived as they have become larger and heavier. Indeed, it is only recently that any significant fraction of human populations had the opportunity to become obese (most still do not). Olshansky et al. do not mention Fogel's work, but it is instructive to think of their argument as an overshoot in which we have too much of a good thing. Set against this possibility, the rise in obesity coincides with, indeed is made possible by, a remarkable rise in living standards that is known to contribute to the longer term reduction of absolute (as opposed to relative) mortality risk. The connection between Fogel's work and Olshansky et al. is suggestive but little more.


One direct effect of changes in life expectancy and health is to change the projected costs of publicly provided retirement and health care. In the United States, the SSA pays retirement benefits and Medicare pays many health care expenses for people over 65. Olshansky et al. start with their estimate of at most 0.75 years lost in life expectancy and argue that extrapolation of recent trends [(3), p. 1141] implies that this loss could increase to as much as 5 years in coming decades as currently obese younger people age. We have noted evidence that the relative mortality risk attributed to obesity actually declines with age, which greatly weakens their argument, and their basic projection is rather uncertain in any case.

How would their projection of a slowdown in long-term mortality decline affect U.S. Social Security? The SSA actuaries project under current rules that the annual fiscal balance (income minus outgo) will become negative after 2017 and that accumulated surpluses will be exhausted by about 2040, after which there will be a deficit (16). Olshansky et al. believe that these fiscal projections are driven by large projected increases in future life expectancy and that people "may be inadvertently saving Social Security by becoming more obese." They fundamentally misunderstand the demography of the Social Security problem: The projected fiscal problems through 2050 are driven by the aging of the baby boom and have little to do with future decreases in mortality. These problems stem from the high fertility of the wartime generation that produced the babies in the boom. Unless obesity dramatically shortens life expectancy relative to today's level, not some projected future level, it will have little effect on Social Security until about 2050. Beyond that, increasing life expectancy is certainly a factor, but even at current mortality levels the system's fiscal balance will not be restored without either changes in the rules or an upturn in fertility.

Medicare is potentially a different story, because the expenses there depend on people's health. If growing obesity leads to protracted ill health, as Olshansky et al. argue, and if people seek more health care as a result, the average cost per year of life could well rise. To assess the importance of this effect, though, we must look at two major factors driving Medicare costs. One is that a substantial part of old-age spending occurs in the last years of life (17), and it is not obvious that this spending will increase as a result of changes in general health (people don't die healthy). Another is that Medicare projections reflect a large rate of increase of total costs that seems to be independent of the health status of the population (18) and that rate of increase dwarfs even the most aggressive projections by Olshansky et al.

So at the level of population mortality trends and individual risks, we have good reasons to be skeptical about the quantitative results in Olshansky et al. and their fiscal implications, but the evidence they summarize about the associations between BMI and health is extensive and deserves more analysis. They have done a useful job of placing the question of population health status front and center, and those of us with high BMIs should probably plan to lose some weight.

April 6, 2005
  1. Period life expectancy is computed from age-specific mortality rates experienced in a particular year (19).
  2. Data for Sweden and the United States are from the Human Mortality Database [University of California, Berkeley (USA) and Max Planck Institute for Demographic Research (Germany)]. Data were downloaded on various dates in March 2005.
  3. S. J. Olshansky, D. J. Passaro, R. C. Hershow, J. Layden, B. A. Carnes, J. Brody, L. Hayflick, R. N. Butler, D. B. Allison, D. S. Ludwig, A potential decline in life expectancy in the United States in the 21st century. N. Engl. J. Med. 352, 1138-1145 (2005).[CrossRef][Medline]
  4. R. Lee, T. Miller, Evaluating the performance of the Lee-Carter approach to modeling and forecasting mortality. Demography 38, 537-549 (2001).[Medline]
  5. S. Tuljapurkar, N. Li, C. Boe, A universal pattern of mortality decline in the G7 countries. Nature 405, 789-792 (2000).[CrossRef][Medline]
  6. K. R. Fontaine, D. T. Redden, C. Wang, A. O. Westfall, D. B. Allison, Years of life lost due to obesity. JAMA 289, 187-193 (2003).[CrossRef][Medline]
  7. T. R. Wessel, C. B. Arant, M. B. Olson, B. D. Johnson, S. E. Reis, B. L. Sharaf, L. J. Shaw, E. Handberg, G. Sopko, S. F. Kelsey et al., Relationship of physical fitness vs body mass index with coronary artery disease and cardiovascular events in women. JAMA 292, 1179-1187 (2004).[CrossRef][Medline]
  8. S. N. Blair, T. S. Church, The fitness, obesity, and health equation: Is physical activity the common denominator? JAMA 292, 1232-1234 (2004).[CrossRef][Medline]
  9. D. B. Allison, C. Wang, D. T. Redden, A. O. Westfall, K. R. Fontaine, Obesity and years of life lost--Reply. JAMA 289, 1777-1778 (2003).[CrossRef]
  10. D. C. Grabowski, J. E. Ellis, High body mass index does not predict mortality in older people: Analysis of the longitudinal study of aging. J. Am. Geriatr. Soc. 49, 968-979 (2001).[CrossRef][Medline]
  11. R. Edwards, S. Tuljapurkar, Inequality in life spans and mortality convergence across industrialized countries. Unpublished working paper, Morrison Institute for Population and Resource Studies, Stanford University (2005).
  12. S. J. Olshansky, B. A. Carnes, C. Cassel, In search of Methuselah: Estimating the upper limits to human longevity. Science 250, 634-640 (1990).[Abstract/Free Full Text]
  13. T. McKeown, The Modern Rise of Population (Academic Press, New York, 1976).
  14. S. H. Preston, The changing relation between mortality and level of economic development. Popul. Stud. (Camb.) 29, 231-248 (1975).
  15. R. W. Fogel, Economic growth, population theory, and physiology: The bearing of long-term processes on the making of economic policy. Amer. Econ. Rev. 84, 369-395 (1994).
  16. The 2005 Annual Report of the Board of Trustees of the Federal Old-Age and Survivors Insurance and Disability Insurance Trust Funds. U.S. Social Security Administration (2005).
  17. T. Miller, Increasing longevity and Medicare expenditures. Demography 38, 215-226 (2001).[Medline]
  18. M. E. Chernew, R. A. Hirth, D. M. Cutler, Increased spending on health care: How much can the United States afford? Health Aff. (Millwood) 22, 15-25 (2003).
  19. S. H. Preston, P. Heuveline, M. Guillot, Demography: Measuring and Modeling Population Processes (Blackwell, Malden, MA, 2001).
Citation: S. Tuljapurkar, Future Mortality: A Bumpy Road to Shangri-La? Sci. Aging Knowl. Environ. 2005 (14), pe9 (2005).

Misdirection on the Road to Shangri-La.
S. J. Olshansky, B. A. Carnes, R. Hershow, D. Passaro, J. Layden, J. Brody, L. Hayflick, R. N. Butler, D. B. Allison, and D. S. Ludwig (2005)
Sci. Aging Knowl. Environ. 2005, pe15
   Abstract »    Full Text »

Science of Aging Knowledge Environment. ISSN 1539-6150