Now, the Taller Die Earlier: The Curse of Cancer

Abstract

This study estimates the relationship between height and mortality. Individuals in the National Health Interview Survey 1986, a nationally representative U.S. sample, are linked to death certificate data until December 31, 2006. We analyze this relationship in 14,440 men and 16,390 women aged 25+. We employ the Cox proportional hazards model, controlling for birthday and education. An additional inch increase in height is related to a hazard ratio of death from all causes that is 2.2% higher for men and 2.5% higher for women. The findings are robust to changing survival distributions, and further analyses indicate that the figures are lower bounds. This relationship is mainly driven by the positive relationship between height and development of cancer. An additional inch increase in height is related to a hazard ratio of death from malignant neoplasms that is 7.1% higher for men and 5.7% higher for women. In contrast to the negative relationship between height and mortality in the past, this relationship is now positive. This demonstrates the success and accessibility of medical technology in treating patients with many acute and chronic diseases other than cancer.

There is growing interest in the long-term effects of early life conditions on later life outcomes (1). To relate early life conditions to later life outcomes, researchers have used many measures. Among them, height has often been adopted because it is relatively inexpensive and easy to measure, objective, readily available, comparable across time and space, and most of all, sensitive to early life conditions; given genetic factors, environments that provide sufficient nutrition and are free of diseases and extreme weather conditions favor the increase in height (2).

Based on this idea, much research has been done to understand the relationship between height and mortality in many countries (3). Initially, empirical methods were crude (eg, unconditional comparisons or comparisons by some observable characteristics) but later became more sophisticated (eg, Cox proportional hazards ratio model). As the methods allowed controlling for confounding factors more effectively, a variety of covariates were taken into account. In addition, researchers first paid attention to all causes of death but later considered more specific causes of death. Thus, the results have varied depending on populations, gender, sample size, number and cause of death, empirical models, and covariates. In general, however, it has been found that tall individuals lived longer than short individuals. According to a meta-analysis, the risk of all-cause mortality is 3% lower per 6.5cm (2.56 inches) increase in height (3). Some researchers have challenged this argument (4,5), but their evidence is not convincing because it is based on selective samples. For example, it is not difficult to find some populations who are short but longer-lived (eg, Okinawans in Japan), considering that many factors influence mortality. Height is a single, albeit important, factor. Regardless, such populations are outliers when we consider broader patterns.

The negative relationship between height and mortality has been explained mainly by the long-term effects of early life conditions. That is, tall individuals experience better living conditions in childhood than short individuals and suffer less from acute and chronic diseases over time and, therefore, would live longer. However, tall individuals, as compared to short individuals, are susceptible to one leading cause of death—cancer (3,6). When life expectancy was relatively short in the past, most people died before developing cancer. During this period, death was mainly caused by infectious and noncancer chronic diseases (7); tall individuals could better overcome these diseases than short ones. However, in the developed world, as people began to live longer, cancer emerged as a leading cause of death, while mortality from other diseases declined. For example, in the United States in 2011, cancer was the first leading cause of death among individuals aged 45–64 (32% of all deaths) and the second leading cause of death among individuals aged 65+ (22% of all deaths) (8). As a result, the death rates from cancer started to decline much later than did those from other diseases. In addition, medical technology is developed and accessible enough to sustain the lives of patients with many infectious and chronic diseases. Thus, the benefits that tall individuals enjoy over short ones are likely to be fewer than before. At the same time, tall individuals are more likely to develop cancer, but the treatment of cancer remains elusive, despite the progress achieved until date (9).

When all these facts are considered together, one can hypothesize that the negative relationship between height and mortality is weaker than before or even turns positive. Herein, we test this hypothesis by analyzing a nationally representative U.S. sample, which was followed in 1986–2006.

Data and Methods

We combine two datasets: the National Health Interview Survey (NHIS) and NHIS public-use linked mortality files. The NHIS—that aimed at monitoring the health of the U.S. population—has been conducted continuously since 1957. It covers the civilian noninstitutionalized population residing in the United States at the time of the interview. It is a cross-sectional household interview survey, and the sampling plan follows a multistage area probability design, which permits the representative sampling of households and noninstitutional group quarters. The variables include basic health and demographic variables. The National Center for Health Statistics has updated the mortality linkage of the NHIS years 1986–2004 to death certificate data found in the National Death Index. The most up-to-date follow-up period was dated December 31, 2006. We linked the NHIS year 1986 to the latest mortality file to obtain the longest follow-up period.

Because the number of nonwhites is not large enough for precise estimation, only whites are considered in this study. Age is restricted to 25+ to allow time for completing a college education. Height is restricted to 50–100 inches because values outside of this range are probably recording errors. Height was self-reported in the NHIS, and researchers found nonrandom discrepancies between self-reported and measured height. Cawley (10) proposed a strategy to address this concern, and we implement it. Specifically, he drew on the Third National Health and Nutrition Examination Survey (NHANES III), which contained both measured and self-reported height. Then, they regressed measured height on self-reported height, self-reported height squared, age, and age squared, separately by race and sex. We insert self-reported height and age in our data in this estimated specification to correct reporting bias. As the NHANES III was nationally representative and conducted in 1988–1994, this method is suitable for the 1986 NHIS. Once done, height ranges were 53.3–82.8 inches for men and 50.5–78.5 inches for women. Regardless, an independent study suggested that potential bias from self-reported height was not concerning (11).

Age-related shrinkage is also of potential concern. We report the results estimated using shrinkage-uncorrected height because correcting both reporting bias and shrinkage would instead cause substantial measurement error. Nevertheless, we correct shrinkage as proposed by Cline and colleagues (12) and rerun the specifications; the substance of the results remains the same (not shown). This is expected because the correlation coefficient between shrinkage-corrected and -uncorrected height is 0.96.

In addition to all causes of death, we separately consider three major causes of death. The NHIS comprises 113 causes of death listed in the International Statistical Classification of Diseases and Related Health Problems, 10th Revision (ICD-10). The cardiovascular diseases in this study correspond to I00–I99 in the ICD-10, malignant neoplasms to C00–C97, and respiratory diseases to J10–J98. For a robustness check, we also exclude nondisease-related causes of death, which correspond to V01–Y88.

The literature has typically employed the Cox proportional hazards model, and therefore, we adopt this method. In addition to (reporting bias) corrected height, we consider birthday and completed years of schooling as covariates. It is necessary to control for birthday because older individuals are shorter due to economic development and older individuals die earlier by natural processes. Otherwise, height simply captures the effect of age on mortality. In addition, it is not necessary but important to control for education because family background exerts a strong influence on education [more than school resources do (13)], and therefore, education reflects early life developmental events. Furthermore, education can directly affect mortality through the acquisition of more health-related knowledge, better understanding of such knowledge, and healthier lifestyle. Because education is correlated positively with height (Supplementary Table A1) and negatively with mortality (Table 2), controlling for education allows us to remove the protective effect of education and early life developmental events from the relationship between height and mortality; otherwise, this relationship would be underestimated if height is positively related to mortality. In addition, by comparing the results in which education is included with those in which education is not included, we can roughly assess the size of the protective effect.

Table 2.

Relative Risk of Death: Cox Proportional Hazards Model

	1	2	3
Panel A: men
Height (inches)	0.884** (0.874, 0.895)	1.010 (0.997, 1.023)	1.022** (1.008, 1.035)
Birthday (in quarter)		0.978** (0.977, 0.978)	0.978** (0.978, 0.979)
Education			0.956** (0.947, 0.965)
	4	5	6
Panel B: women
Height (inches)	0.798** (0.786, 810)	1.019* (1.005, 1.034)	1.025** (1.010, 1.040)
Birthday (in quarter)		0.978** (0.977, 0.978)	0.978** (0.977, 0.979)
Education			0.970** (0.960, 0.981)

Notes: Sample sizes are 14,421 for men and 16,383 for women. A total of 26 men and women are excluded from the analysis because they died in the same quarter when they were interviewed in 1986. Sampling weights are applied. 95% confidence intervals are in parentheses.

*p < .05; **p < .01.

Because death is recorded in quarters in the data, to be consistent, we use birth in quarters. Specifically, 0 is assigned to individuals born in the first quarter of 1960, and 1 to those born in the second quarter of 1960, and so on. Similarly, −1 is assigned to those born in the fourth quarter of 1959, and −2 to those born in the third quarter of 1959, and so on. The set of covariates is intentionally parsimonious because we want to estimate the total relationship between height and mortality, only excluding those resulting from largely predetermined variables (ie, birthday and education). Sampling weights are applied to make estimations nationally representative. The sample size is 14,440 for men and 16,390 for women. However, a total of 26 men and women are excluded because they died in the same quarter as the period of interview. Table 1 presents descriptive statistics.

Table 1.

Descriptive Statistics

	Men	Women
	Mean (SD)	Mean (SD)
Continuous variable
Height (inches)	69.4 (2.7)	63.9 (2.6)
Birthday in quarter since January 1, 1960	−85.1 (63.7)	−92.1 (68.3)
Education	12.6 (3.3)	12.2 (3.0)
Age	47.2 (15.9)	48.9 (17.1)
Discrete variable	%	%
Alive	70.2	73.2
Dead of all causes	29.8	26.8
Dead of nondiseases	1.6	0.8
Dead of malignant neoplasms	7.5	6.4
Dead of cardiovascular diseases	12.7	11.7
Dead of respiratory diseases	3.0	2.6
Not healthy at the beginning	41.3	45.8
Healthy at the beginning	58.7	54.2
Age < 45	50.7	48.1
Age ≥ 45	49.3	52.0
N	14,440	16,390

Results

Supplementary Table A1 presents the correlation coefficients between height, birthday, and education. Although the correlation coefficients are not high, all of them are positively correlated for men and women. Thus, tall men are born later and more educated than are short men. This fact is important because the relationship between height and mortality is confounded by birthday and education.

Figures 1 and 2 illustrate this point. Individuals in each gender are divided into quintiles by height, and then, individuals in the first and fifth quintiles are compared for mortality by gender. Birthday is unadjusted in Figure 1 and adjusted in Figure 2. Birthday-adjusted height is represented by the residual of the regression of height on a constant and birthday. When birthday is unadjusted, the survival rate of the tall group is higher than that of the short group. Once birthday is adjusted, however, this is no longer the case; the tall male group is more likely to die than the short male group, and there is no statistically significant difference between the tall and short female groups.

Figure 1.

Survivor functions without birthday adjustment. Note: The two survival curves for each gender are statistically significantly different.

Figure 2.

Survivor functions with birthday adjustment. Note: The two survival curves are statistically significantly different for men, but not for women.

The Cox proportional hazards model can verify this more formally. Table 2 presents the hazard ratios of the three variables. Regarding men, when only height is controlled for (column 1), an additional inch increase in height is related to an 11.6% lower hazard ratio. However, when birthday is added to the specification (column 2), the hazard ratio for height loses its statistical significance. Of course, later born individuals face a lower hazard ratio. When education is added, as expected, the hazard ratio for height regains statistical significance, indicating that an additional inch increase in height is related to a 2.2% higher hazard ratio. Similar patterns are observed for women (columns 4–6). When birthday and education are controlled for, an additional inch increase in height is related to a 2.5% higher hazard ratio, which is close to that of men.

These findings are not driven by model selection. Supplementary Table A2 presents the results estimated by other popular models with a natural proportional hazards parameterization. Whether the survival distribution is exponential, Gompertz, or Weibull, the hazard ratio of height is almost identical to that estimated by the Cox model for both genders.

Table 3 investigates whether these results are biased. Men are examined first in Panel A. If height, representing early life conditions, is more related to death due to disease than death due to nondisease, the hazard ratio for height for the pooled sample would be biased downward. Column 1 confirms this, showing that when deaths due to nondisease are excluded, the hazard ratio for height increases slightly for both genders. In addition, if short individuals are likely to be unhealthy at the time of the interview, they would subsequently die earlier than tall individuals. In this case, the hazard ratio for height for the pooled sample would also be biased downward. To check this, we restrict our analysis to individuals who reported that their health status was excellent or very good and that they were not limited in any activity. Column 2 confirms this; an additional inch increase in height is related to a 2.7% higher hazard ratio. If short and unhealthy individuals had already died young and only exceptionally healthy short individuals survived to an old age, the hazard ratio for height for the pooled sample would be biased upward. We restrict our analysis to individuals aged 45+, which is approximately the median ages for both genders. Column 3 indicates that the upward bias stemming from this is small; the hazard ratio for height is almost the same between the pooled and restricted samples. The same patterns are observed for women (columns 4–6). Thus, the hazard ratio for height for the pooled sample is more likely to be biased downward than upward, which reinforces our argument that the taller die earlier.

Table 3.

Relative Risk of Death With Varying Conditions

	Disease-Related Mortality	Healthy at the Beginning	Age ≥ 45
	1	2	3
Panel A: men
Height (inches)	1.025** (1.011, 1.039)	1.027* (1.005, 1.049)	1.021** (1.007, 1.036)
N	14,197	8,471	7,105
	4	5	6
Panel B: women
Height (inches)	1.026** (1.011, 1.042)	1.031* (1.005, 1.058)	1.023** (1.007, 1.039)
N	16,257	8,876	8,509

Notes: Birthday and education are controlled for, but not listed. Sampling weights are applied. 95% confidence intervals are in parentheses.

*p < .05; **p < .01.

Supplementary Table A3 investigates whether the relationship between height and mortality is nonlinear. When the square of height is entered into the specification, the hazard ratios for height and its squared term are not statistically significant, suggesting that the relationship is more likely to be linear. We also relax the linear parameterization by replacing the linear term of height with dummies for height quintiles. The hazard ratio generally increases with height quintiles (column 2), so the hazard ratio for individuals in the fifth quintile is 13.6% greater than that in the first quintile. The results for women are similar (columns 3–4). For both genders, some quintile dummies are not statistically significant, but all quintile dummies are jointly significantly different from zero. Overall, it appears that the linear relationship between height and mortality is not far from the truth.

Thus far, the case of all causes of death is considered. These results help one understand the general relationship between height and mortality, but it is difficult to find what disease drives this relationship. Therefore, three major diseases—malignant neoplasms, cardiovascular diseases, and respiratory diseases—are considered to evaluate the relationship with height. Table 4 presents the hazard ratio for height when age and education are controlled for but not listed for brevity. The results indicate that height is statistically significantly related to only malignant neoplasms among men; an additional increase in height is related to a 7.1% higher hazard ratio among men. Among women, the relationship is statistically significant for malignant neoplasms and cardiovascular diseases. However, the relationship for the former disease is twice as great as that for the latter: 5.7% versus 2.8%. For both genders, the magnitude for malignant neoplasms is much greater than that for all causes of death. Because more men died of cardiovascular diseases than of malignant neoplasms (12.7% vs 7.5%), the statistically nonsignificant relationship between height and cardiovascular diseases among men does not result from a lack of variation in deaths due to cardiovascular diseases. However, this possibility remains for respiratory diseases because only 3.0% of men and 2.6% of women died due to these diseases. Even if the relationship for respiratory diseases were statistically significant, it would be opposite to that of all causes of death (for men) or nil (for women). Thus, the positive relationship between height and mortality in the pooled sample is largely driven by mortality due to malignant neoplasms.

Table 4.

Relative Risk of Death by Specific Diseases

	Malignant Neoplasms	Cardiovascular Diseases	Respiratory Diseases
Panel A: men
Height (inches)	1.071** (1.044, 1.098)	1.011 (0.991, 1.031)	0.992 (0.954, 1.032)
Panel B: women
Height (inches)	1.057** (1.028, 1.086)	1.028* (1.005, 1.051)	1.001 (0.959, 1.046)

Notes: Birthday and education are controlled for, but not listed. Sample sizes are 14,421 for men and 16,383 for women. Sampling weights are applied. 95% confidence intervals are in parentheses.

*p < .05; **p < .01.

Discussion

It has been widely observed that tall individuals live longer or die later than short ones even when age and other socioeconomic conditions are controlled for. Some researchers challenged this position, but their evidence was largely based on selective samples. For example, Salaris and colleagues (5) argued that short conscripts born in a village in Italy during 1866–1915 lived longer than their tall counterparts. However, they restricted the sample to those aged 70+. If short individuals tend to be weak and die early (eg, before age 70), short individuals aged 70+ would likely be exceptionally healthy and live longer than their tall counterparts. Other examples provided by Samaras and colleagues (4) are subject to similar criticisms.

It is of great interest that our data are nationally representative and we do not attempt to restrict the sample. Nevertheless, the results consistently indicate that tall men and women die earlier than their short counterparts when a parsimonious set of variables (ie, birthday and education) are controlled for. Thus, an additional inch increase in height is related to a 2.2% higher hazard ratio for men and a 2.5% higher hazard ratio for women. Further evidence suggests that that these magnitudes are likely to be lower bounds.

Recall that the hazard ratio increases when education is controlled for. This is because education is positively related to height but negatively related to mortality. Considering the powerful role of family background on education, one can deduce from these results that early life developmental events generate two conflicting effects. On the one hand, better developmental events provide children with better growth environments (eg, better nutrition and less disease burden), but growth increases mortality. On the other hand, better developmental events help children acquire more education, which in turn, reduces mortality. It is worth highlighting these conflicting effects because little attention has been paid to the former channel while, as our results indicate, the former has already dominated the latter. Furthermore, when the results in which education is included are compared to those in which education is not included, the protective effect of education is roughly a hazard ratio of 0.012 for men and 0.006 for women. However, one should interpret these figures with caution because the coefficients on height with and without education are not statistically significantly different.

It is instructive to compare our results with those of previous studies for the United States. Costa examined a group of over 300 white Union army veterans of the American Civil War (14). She employed a proportional hazards model and found a negative relationship between height and mortality. This was true even when age and socioeconomic conditions were controlled for. On the other hand, Murray (15) investigated nearly 2,500 students at Amherst College in Amherst, Massachusetts, in the years 1861 through 1900. He employed the Cox proportional hazards model and found that height was not related to mortality whether or not other covariates were controlled for. This anomaly could stem from the highly selective nature of his sample. While Costa and Murray relied on men born in the 19th century, Liao and colleagues (16) analyzed a nationally representative sample of men and women born in the 20th century; their sample size was large at 13,031 individuals. The baseline survey was conducted during 1971–1975 and followed up to 1987. They also employed the Cox proportional hazards model and found that height was negatively related to mortality from all causes of death—with no covariate. However, once age and education were controlled for, the relationship for both genders was no longer statistically significantly. Cook and colleagues (17) also reported a U.S. case, but their sample consisted of individuals aged 65+ living in East Boston, Massachusetts. Because Liao and colleagues’ sample covered the entire country for a similar observation period, Cook and colleagues’ study is not elaborated here.

When these results are considered together with ours, some patterns emerge. The relationship between height and mortality (with birthday and education being controlled for) switched from negative to positive over time. It appears that the advantages offered by tallness have been eroded by its disadvantages. Specifically, in the past, nutrition, disease, and workload varied greatly among the U.S. population, but the medical technology was neither developed nor accessible enough to selectively treat the tall (healthy, rich) and the short (unhealthy, poor). Therefore, early life conditions were a primary determinant of death, and the taller were healthier and lived longer. As time passed, early life conditions based on nutrition, disease, and workload were no longer seriously adverse for the short (poor) in the United States, while medical technology was developed and accessible enough for both the short and the tall. However, the treatment of cancer is neither sufficiently developed nor accessible for the short and the tall, but the tall are more likely to develop cancer than the short.

In the United States, infectious and acute diseases no longer pose a serious public health threat. Death rates for infectious diseases have continued to decline since the U.S. federal government began publishing mortality statistics on an annual basis in 1900, except for the years 1918 (the influenza pandemic) and 1982–1995. More specifically, during 1900–1937, the crude infectious disease mortality rate decreased by approximately 2.8% per year, which was facilitated indirectly by the germ theory of disease and directly by public health interventions such as clean water supply and improved sanitation (the First Mortality Revolution). This downward trend was accelerated by a wave of medical innovations in the 1940s (the Second Mortality Revolution); the rate declined by 8.2% per year in the subsequent 15 years (7). At present, chronic diseases are a major threat to life. However, the lives of patients with major chronic diseases can be sustained by the current medical technology, although the quality of life is another issue. Notably, the association between obesity (a major health threat) and mortality has weakened over time (18), and now, obesity is more likely to shorten disability-free life expectancy than overall life expectancy at middle and older ages (19); succinctly, obesity does not kill, but disables. Furthermore, death rates from cardiovascular diseases already started to decline in 1950 thanks to pharmaceutical innovations, the increased effectiveness of invasive medical treatments, and behavioral changes (20).

However, death rates from cancer started to decline only in 1991 (21). In addition, at least during the follow-up period of our sample, it was very difficult to extend the lives of patients with malignant neoplasms, irrespective of their past and current socioeconomic status (9). This argument is further supported by the Emerging Risk Factors Collaboration’s finding that the positive relationship between height and cancer mortality was the same regardless of birth years (3). That is, as far as mortality is concerned (screening, diagnosis, and morbidity are different issues), younger generations do not enjoy more benefits from cancer-related medical technology than older generations. An additional piece of evidence is that, while the relationship between (Class I) obesity and death due to cardiovascular disease has decreased over time, the relationship between obesity and death due to cancer has remained the same (18).

At the same time, the World Cancer Research Fund and the American Institute of Cancer Research surveyed the literature and concluded that height was positively related to the risk of cancers of a number of sites (6). The reason appears to be related to growth hormone (GH), which affects the regulation of cellular growth via its mediator peptide insulin-like growth factor-1 (IGF-1). It is obvious that more GH promotes taller height; acromegaly is an extreme case of this. Although not as obvious as this, evidence of the causal role of more GH in higher mortality is provided by in vitro studies, animal studies, epidemiological observations within the general population and patients with excess and deficiency of GH, and therapeutic manipulation of GH and IGF-1 actions (22). IGF-1 exerts a strong effect on each of the key stages of cancer development and behavior, namely, cellular proliferation and apoptosis, angiogenesis, metastasis, and development of resistance to chemotherapeutic agents. Specifically, the proliferation effects increase epithelial cell turnover within tissues, whereas the antiapoptotic effects loosen the tight control between proliferation and cell death. This imbalance causes hyperproliferation and thereby sets the first stage in cancer development. Some stem cells undergo early genetic “hits,” and as the pool of such damaged cells increases, more of them are available to subsequent hits. High levels of IGF-1 make programmed cell death of damaged cells slightly less probable, and this malfunction can accelerate carcinogenesis; however, it remains to be confirmed whether IGF-1 initiates cancer development per se. That said, at present, the advantages offered by tallness are substantially undercut by its disadvantages.

If our explanation is correct and the medical technology is developed just enough to treat only the tall (rich, due to the treatment costs), then the negative relationship between height and mortality will return, to the extent that height is positively related to the past and current socioeconomic status. In the further future, if the technology becomes accessible to all cancer patients, the relationship between height and mortality will be nil—unless new diseases related to height appear.

This study considers the U.S. population, but our results can be generalized to other populations because the mechanism is mainly biological. In fact, the positive relationship between height and mortality may be stronger in other developed countries than in the United States because while cancer is a leading cause of death in developed countries (23), populations in other developed countries are taller (24) and expect to live longer at birth than the U.S. population (25). Moreover, as some developing countries experience economic growth, their populations grow taller (26) and live longer (25). Therefore, we may also see a positive relationship in developing countries in the future. However, we hastily add that changes in effectiveness and availability of cancer treatment mediate this relationship, and therefore, more research is needed to confirm these speculations.

We acknowledge some limitations. First, the follow-up period is 20 years, so the majority of death cases have not yet been observed. If all death cases were observed, our results could be different. However, this follow-up period is long compared with those of other studies. Furthermore, if all death cases were observed, that would make this study a historical study, thereby losing relevance to the current period. This is certainly a tradeoff, and we believe that timeliness is more important than completeness. Second, only three major causes of death are examined. This is related to the first limitation because there are not many death cases to analyze for each cause of death. Third, only the white race is considered because of the first limitation. Blacks and Hispanics may exhibit a different relationship between height and mortality than do whites, and this would enrich the understanding of this relationship. For the second and third limitations, we offer the same response provided for the first limitation.

As the follow-up period extends, future research can address these limitations. In addition, it can verify our predictions about the relationship, as medical technology and costs change over time. Furthermore, it would be of interest to check whether the mechanisms suggested by the World Cancer Research Fund and the American Institute of Cancer Research and the Emerging Risk Factors Collaboration are correct and to find other mechanisms that predispose the tall to morbidity and mortality; possible topics include the role of genetic and environment factors in the relationship and the critical periods for the relationship. Last but not least, we mostly focus on cancer. However, we detect some weak sign of a positive relationship between height and mortality from cardiovascular diseases. Although not as unidirectional in causality as for cancer, evidence suggests that GH and IGF-1 are implicated in cardiac development (27). Given this, it is worth investigating this relationship.

Supplementary Material

Please visit the article online at http://gerontologist.oxfordjournals.org/ to view supplementary material.