Observed and predicted 9-year mortality by decile of predicted risk.
Tice JA, Kanaya A, Hue T, Rubin S, Buist DSM, LaCroix A, Lacey JV, Cauley JA, Litwack S, Brinton LA, Bauer DC. Risk Factors for Mortality in Middle-aged Women. Arch Intern Med. 2006;166(22):2469-2477. doi:10.1001/archinte.166.22.2469
Copyright 2006 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.2006
Many factors contribute to mortality in older women, but their relative importance and independent contribution have been poorly characterized.
From 1990 to 1992, we assessed demographics, lifestyle measures, prevalent disease, medication use, anthropometrics, vital signs, and physical function in 17 748 postmenopausal women. We used proportional hazards modeling to evaluate their association with mortality.
During 9 years of follow-up, 1886 women (10.6%) died. The relative hazard (RH) of death was approximately 1.5 (95% confidence interval [CI], 1.5-1.6) per 5 years of age, 1.4 (95% CI, 1.2-1.6) for a history of heart disease, and 1.9 (95% CI, 1.6-2.3) for a history of breast cancer. Modifiable risk factors associated with mortality included smoking (RH, 3.7 [95% CI, 3.1-4.5] for current smokers with a ≥50 pack-year history) and systolic blood pressure (RH, 1.3 [95% CI, 1.1-1.5], fifth vs first quintile). Elevated waist-hip ratio was associated with higher mortality (RH, 1.3 [95% CI, 1.1-1.5], fifth vs first quintile), but obesity was associated with lower mortality (RH, 0.7 [95% CI, 0.6-0.9] for body mass index [calculated as weight in kilograms divided by the square of height in meters] of >35.0 vs 18.5-25.0). Poor results on the timed Up and Go Test, a measure of physical function, were also strongly associated with mortality (RH, 1.7 [95% CI, 1.4-2.0], fifth vs first quintile).
Simple measures are sufficient to stratify postmenopausal women into groups at high and low risk of dying. Smoking, central obesity, blood pressure, and physical function are potentially modifiable risk factors, although clinical trials are required to demonstrate that change in these factors affects mortality.
Cohort studies have demonstrated significant associations between many individual markers of disease severity,1,2 physical function,3- 7 or blood test results8- 14 and all-cause mortality. However, few studies have assessed the joint contributions of disease and disability on mortality. There are many validated risk indices for mortality, but they often evaluate a limited set of risk factors such as comorbidity lists,15- 18 use in-hospital or short-term mortality as the outcome,19- 23 focus on limited populations such as patients hospitalized with heart failure or the elderly,19,21,22,24 or have extreme parsimony as a primary goal.25
Our goal was to assess the relative strength and joint contribution of factors drawn from multiple domains on the risk of death in community-dwelling postmenopausal women. We were particularly interested in assessing the contribution of potentially modifiable risk factors.
The Breast and Bone Follow-up Study of the Fracture Intervention Trial (B-FIT) was initiated to investigate whether bone mineral density (BMD) is associated with cancer in postmenopausal women. Participants were women screened for the Fracture Intervention Trial, a randomized clinical trial of alendronate sodium for the prevention of osteoporotic fractures.26 The cohort was recruited from a large, geographically diverse population of older women enrolled from 1990 to 1992 in 11 metropolitan areas of the United States (listed in the Acknowledgments section). Women were recruited through advertisements in print and electronic media and direct mailings using population-based listings.26,27 Postmenopausal women aged 55 to 80 years were eligible for screening visits. All women provided written informed consent. The institutional review board at each clinical site, the coordinating center, and the National Cancer Institute approved the study protocol.
One of the original 11 sites declined participation in the present study. We excluded women with missing data (n = 4927) on any of the measures planned for analysis. Thus, 17 748 participants were included in this analysis. The mortality rate was identical (11%) for women excluded and for women included in the analysis, although bias could have been introduced if higher mortality due to some factors balanced lower mortality due to other factors.
All participants received a mailed questionnaire that collected data on participant demographics, health habits, medical history, detailed reproductive history, the 20-Item Short-Form Health Survey,28 a depression scale (Center for Epidemiological Studies–Depression Scale),29,30 the Framingham physical activity scale,31 and a modified food frequency questionnaire.32
Blood pressure was measured in the right brachial artery according to a standard protocol.33 Height and weight were measured, and body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters. Waist and hip girths were measured with a steel tape, and the waist-hip ratio (WHR) was used as an indicator of body fat distribution. The timed Up and Go Test was performed according to the method described by Podsiadlo and Richardson.34 Grip strength was measured in the dominant hand using a dynamometer. Bone mineral density of the hip was measured using dual x-ray absorbtiometry (QDR-2000, Hologic, Inc, Waltham, Mass).
We assessed the date and cause of death by linking to the National Death Index, which is accurate for the ascertainment of mortality, although classification of the underlying cause of death is limited by the use of death certificates.35- 37 The National Death Index linkage was performed in 2003 and provided data on deaths through December 31, 2001.
We assessed 62 variables for their association with mortality based on documented associations of these variables with mortality in the literature and biological plausibility (a list of variables used is available from the authors on request). We organized them into 9 related groups of variables, including demographic factors, anthropometric measures, lifestyle factors, vital signs, BMD factors, physical function, effects of disease, self-reported medical history, and reproductive factors. We used standard categories for age at baseline (per 5-year intervals) and BMI (<18.5, 18.5-24.9, 25.0-29.9, 30.0-34.9, and ≥35.0).38 We created a composite variable for smoking that included lifetime pack-years of smoking and use at enrollment (never, past with <50 pack-years, past with ≥50 pack-years, current with <50 pack-years, or current with ≥50 pack-years). We assessed all other continuous predictors as quintiles to allow for nonlinear relationships with total mortality.
We used Cox proportional hazards regression to calculate unadjusted and age-adjusted hazard ratios for total mortality for each risk factor. All predictors met the proportional hazards assumption. We built sequential models adding 1 risk factor group at a time to the model and retained variables significant at P<.05 until all variables were assessed for inclusion. Highly correlated variables such as systolic blood pressure and pulse pressure were evaluated independently, and the variable with the larger change in log-likelihood was included in the final model. After completion of the final model, interactions of age, smoking status, and prevalent coronary heart disease at baseline with other predictors of mortality were tested. None of the interaction terms were statistically significant. Given the large sample size and the number of risk factors considered, a conservative value for statistical significance (P<.001 by the likelihood ratio test) was required for inclusion in the final model. Where appropriate, a test for linear trend was performed to assess statistical significance across risk factor categories.
We calculated a risk score for each woman using coefficients from the final model. The discriminatory accuracy of the model was assessed with the concordance index (c-index).39 Because the risk score was model specific and at risk for overfitting, cross-validation of the final model was performed by recalculating the model coefficients for the final set of risk factors 1000 times using sequential random samples of 90% of the participants and calculating the c-index in the 10% of participants not used in model development.39
At inception, the mean (SD) age of participants was 68 (6) years and 95.3% were white (Table 1). There were 1886 deaths (10.6%) during 9 years of follow-up. The primary causes were cardiovascular disease (35.8%) and cancer (37.4%) (Table 2).
Factors positively associated with death in the final multivariate model included age, hypertension, diabetes mellitus, heart disease, stroke, breast cancer, no use of postmenopausal hormone therapy, recent weight loss, worse self-reported health status, current smoking, pack-years of smoking, lower BMI, higher WHR, higher systolic blood pressure, higher heart rate, longer Up and Go Test times, and weaker grip strength (Table 3). The relationship of alcohol consumption with mortality was U-shaped. Mortality was highest in women reporting no alcohol intake and in those drinking more than 60 alcoholic beverages per month.
After multivariate adjustment, the relative hazard (RH) of death was 5.2 (95% confidence interval [CI], 3.7-7.4) for women 80 years and older compared with women aged 55 to 59 years. There was at least a 33% increase in mortality per 5 years across the age range in the study. The multivariate-adjusted RHs of death were 1.4 (95% CI, 1.2-1.6) for women with a history of heart disease, 1.6 (95% CI, 1.3-1.9) for women with a history of stroke, and 1.9 (95% CI, 1.6-2.3) among women with a history of breast cancer.
Potentially modifiable risk factors associated with mortality included smoking (RH, 3.7 [95% CI, 3.1-4.5] for current smokers with ≥50 pack-year history; RH, 2.2 [95% CI, 1.9-2.6] for current smokers with <50 pack-year history), WHR (RH, 1.3 [95% CI, 1.1-1.5], fifth vs first quintile), and systolic blood pressure (RH, 1.3 [95% CI, 1.1-1.5], fifth vs first quintile). Modest alcohol intake was associated with lower mortality (RH, 0.8 [95% CI, 0.7-0.9], 1-29 drinks per month vs none), but this benefit was no longer present for women consuming more than 60 drinks per month. The timed Up and Go Test, a measure of physical function, was also strongly associated with mortality (RH, 1.7 [95% CI, 1.4-2.0], fifth vs first quintile). Other measures of physical function, such as lower scores on the physical function scale of the 20-Item Short-Form Health Survey and weaker grip strength, were also associated with mortality in the final model.
Underweight women were at an increased risk of death (Table 4). High BMI was positively associated with mortality in the age-adjusted model, but this association disappeared after further adjustment for hypertension and diabetes (Table 4). The inclusion of WHR in the model resulted in an inverse relationship between high BMI and mortality. This inverse relationship remained highly significant in the final model (RH, 0.7 [95% CI, 0.6-0.9; P<.001], BMI of ≥35.0 vs 18.5-25.0).
Unadjusted, age-adjusted, and multivariate-adjusted BMD assessed at the femoral neck, greater trochanter, and total hip were not associated with mortality (RH, 1.0 [95% CI, 0.95-1.03], fifth vs first quintile for total hip BMD). Education was inversely associated with mortality in the age-adjusted model (RH, 0.7 [95% CI, 0.6-0.9; P<.001] for women with at least a college education compared with women not completing high school). The association was no longer significant after adjustment for smoking and alcohol consumption, and the RH approached 1.0 after further adjustment for the timed Up and Go Test, hypertension, and diabetes (Table 5). Physical activity as assessed by the Framingham Activity Scale and blocks walked per day was not a significant risk factor after adjusting for other measures of physical function. Similarly, participation in the clinical Fracture Intervention Trial was associated with lower mortality in the age-adjusted model, but not in the fully adjusted model. Measures of depression and reproductive factors, such as age at menarche, age at menopause, and parity, were not significant risk factors in the final model (data not shown).
The predicted 9-year mortality for women in the highest decile of risk (35%) was almost 18 times that for women in the lowest decile (2%) and closely matched the observed mortality (Figure). The discriminatory accuracy of the model assessed by the c-index was 0.76 and was stable in cross-validation (mean c-index, 0.76 [interquartile range, 0.74-0.77]).
Among the 62 factors in 9 domains that we considered, 19 factors independently predicted mortality. The strongest associations were with age and smoking status at baseline. Many of the risk factors in the final model are potentially modifiable, including smoking, alcohol use, central obesity, systolic blood pressure, heart rate, and physical function as reflected in the Up and Go Test time, the physical function subscale of the 20-Item Short-Form Health Survey, and grip strength.
The strong association of smoking with mortality is a critical reminder that smoking is the most important modifiable risk factor that physicians and society should address, even in older women. Smoking cessation efforts remain among the most cost-effective interventions provided by clinicians.40- 44 In our study, lung cancer accounted for 25.2% of all cancer deaths; chronic obstructive pulmonary disease accounted for 61.4% of all deaths due to respiratory disease; and ischemic heart disease accounted for 35.9% of all cardiovascular deaths. Current smoking at baseline and total pack-years smoked were strongly associated with total mortality. Almost half (45.5%) of women in the study had smoked at some time in their lives, which was comparable to estimates for older women in the United States during 1990 to 1992.45
Our results and those of other studies46- 48 suggest that exercise interventions aimed at improving strength and cardiovascular fitness might improve longevity. Randomized clinical trials have demonstrated that exercise programs in older women49- 58 can prevent diabetes, lower blood pressure and heart rate, increase grip strength, improve the timed Up-and-Go Test results, and decrease central adiposity. To our knowledge, no published studies have had adequate power to assess the effect of exercise programs on total mortality, and no studies have demonstrated that change in physical function correlates with decreased mortality.
The association of BMI with mortality was intriguing. In the age-adjusted model, women in the highest obesity category (BMI, ≥35.0) had a 20% increased risk of death compared with women with a healthy BMI. Women with a modestly elevated BMI (25.0-34.0) had no increased risk of death. However, after adjusting for potential confounders and risk factors on the same causal pathway (hypertension, diabetes, cardiovascular disease, WHR, and physical function), even the highest category of obesity was associated with a 30% reduction in the risk of death. Measures of central obesity have been more consistently related to heart disease events and total mortality than overall obesity.59- 63 There was an independent linear association of elevated WHR with mortality after controlling for BMI. After adjusting for central obesity with WHR, BMI may primarily represent the effect of lean body mass on mortality. This has been observed in several other cohorts of older men and women60,63,64 and may have been missed in previous studies that did not adjust for measures of abdominal obesity.65- 68
As observed in earlier studies,67,69- 72 women with below-normal weight (BMI, <18.5) were at a very high risk of death, even after adjustment for potential confounders like smoking, recent weight loss, and comorbid illness. Some have argued that residual confounding from smoking and concurrent illness explains the increased risk. In one study, limiting the analysis to those who never smoked who had stable weight gave a monotonically increasing risk of death across the full BMI range.67 However, in our study, imposing the same limits did not change our results, nor did eliminating women who died during the first 2 years of follow-up.
We replicated the consistent finding in observational studies that women receiving hormone therapy have a lower risk of dying.73- 80 The reasons for the contradictory results of observational studies73- 80 and randomized trials81- 83 of hormone therapy remain controversial.84,85 It has been hypothesized that hormone therapy users are healthier than nonusers, but careful adjustment for differences in baseline risk factors has not fully explained the difference.76- 80 Residual confounding and adjustment for the length of time receiving hormone therapy may explain much of the discrepancy.85 However, randomized clinical trials have demonstrated that hormone therapy does not reduce mortality.81- 83
The lack of an association of BMD with mortality was unexpected. Previous studies have reported that low BMD is associated with increased mortality.86- 90 Furthermore, women with osteoporotic fractures are at increased risk of death.91- 95 More detailed analyses focusing on cause-specific mortality may help understand this perplexing finding.
The only measure of socioeconomic status that we assessed, education, was inversely associated with mortality in the age-adjusted model. However, the association completely disappeared after adjustment for lifestyle factors including smoking, alcohol use, and physical function. This confounding by lifestyle factors suggests that differences in health outcomes associated with fewer years of education might be attenuated by aggressive public health campaigns focused on smoking prevention, moderating alcohol intake, and exercise. Unfortunately, educational level attained may not the best measure of socioeconomic status in older women. Household net worth and home ownership have been suggested as better measures of socioeconomic status in the elderly because they may better represent cumulative lifetime exposure.96,97
The c-index is a measure of the ability of the model to discriminate between women who died during follow-up and those who remained alive. Among all of the pairs of women with different outcomes, the women who died had a higher risk score than did the surviving women 76% of the time. By random chance, this would occur 50% of the time. A c-index of 0.76 is good for a prognostic model. The Gail model for assessing breast cancer risk had a c-index of 0.58 in the Nurses' Health Study,98 and the Framingham model for heart disease risk assessment had a c-index of 0.63 to 0.83 when validated in 6 cohort studies.99 However, our current model is too complex to apply in daily practice. A simpler model25 with only 12 risk factors also had good discriminatory accuracy (c-index, 0.82) and may be more appropriate in the clinic.
Several studies2,100- 102 have reported on the association between cardiovascular risk factors and cardiovascular and all-cause mortality in women. However, only the Cardiovascular Health Study analyzed the combined effects of risk factors from multiple domains on total mortality in both men and women.70 The Cardiovascular Health Study analyzed data from a smaller cohort of women (2962 vs 17 748) with shorter follow-up time (5 vs 9 years) than did the B-FIT study. Despite differences between the study populations, the findings are remarkably consistent. Both analyses included approximately 20 risk factors in the final model, and the strength of the association between risk factors in common between the 2 models was similar. Except for the rare findings of aortic stenosis and occlusion of the internal carotid artery, age and smoking were the only risk factors with RHs greater than 2.0 (or <0.5) in both studies. The ability of the models to separate their respective cohorts into high- and low-risk groups reflects the cumulative contribution of many small risk factors.
There are several important limitations to our study. The women were volunteering to participate in a clinical trial and thus were likely to be healthier than an age-matched sample of the general population. They were also more highly educated than the overall US population. The women were predominantly white, which limits the ability to generalize these findings to other racial and ethnic groups. Most of the risk factors were measured by self-report; thus, there was likely some degree of misclassification, hence residual confounding. We did not have laboratory measures available for all women, so many factors with known associations with mortality could not be evaluated. Finally, although the National Death Index is estimated to have a sensitivity of about 97%,37 there was likely some degree of underascertainment of death, the primary end point.
Simple measures available on most patients visiting a primary care physician's office were sufficient to stratify postmenopausal women into groups at high and low risk of dying. Smoking, central obesity, blood pressure, and physical function were modifiable risk factors associated with mortality. Interventions targeted to improve these factors have the potential to decrease mortality in older women. Clinical trials are required to demonstrate that modification of these risk factors improves longevity and quality of life.
Correspondence: Jeffrey A. Tice, MD, Division of General Internal Medicine, Department of Medicine, University of California–San Francisco, 1701 Divisadero St, Suite 554, San Francisco, CA 94143 (firstname.lastname@example.org).
Accepted for Publication: August 14, 2006.
Author Contributions:Study concept and design: Tice, Rubin, and Bauer. Acquisition of data: Hue, Rubin, Buist, LaCroix, Lacey, Cauley, Brinton, and Bauer. Analysis and interpretation of data: Tice, Kanaya, Buist, LaCroix, Lacey, Litwack, Brinton, and Bauer. Drafting of the manuscript: Tice and Litwack. Critical revision of the manuscript for important intellectual content: Tice, Kanaya, Hue, Rubin, Buist, LaCroix, Lacey, Cauley, Brinton, and Bauer. Statistical analysis: Tice, Lacey, and Litwack. Obtained funding: Brinton and Bauer. Administrative, technical, and material support: Hue, Rubin, Buist, and Cauley. Study supervision: Brinton and Bauer.
Financial Disclosure: None reported.
Funding/Support: This study was supported by contract N02-CP-01019 from the National Cancer Institute, by a series of contracts from the National Cancer Institute to the clinical centers, and by faculty development grant K12 AR47659 from Building Interdisciplinary Research Careers in Women's Health.
Additional Information: The following centers participated in the B-FIT Study: University of California–San Francisco (coordinating center; principal investigator [PI], Douglas C. Bauer, MD); University of California–San Diego, La Jolla/Rancho Bernardo (PI, Elizabeth Barrett-Connor, MD); Center for Health Studies Group Health Cooperative, Seattle, Wash (PIs, Diana S. M. Buist, PhD, and Andrea LaCroix, PhD); University of Pittsburgh, Monongahela Valley, Pa (PI, Jane A. Cauley, DrPH); Kaiser Permanente Center for Health Research, Portland, Ore (PI, Emily Harris, PhD); Stanford University, Palo Alto, Calif (PI, William Haskell, PhD); University of Maryland, Baltimore (PI, Marc Hochberg, MD, MPH); University of Miami, Miami, Fla (PI, Silvina Levis, MD); Wake Forest University, Winston-Salem/Greensboro, NC (PI, Sara Quandt, PhD); University of Iowa, Iowa City/Bettendorf (PI, James Torner, PhD); and University of Tennessee, Memphis (PI, Suzanne Satterfield, MD, MPH). Project officers at the National Cancer Institute were Drs Lacey and Brinton.
Acknowledgment: We thank the clinical managers and study team at each of the clinical sites and the B-FIT PIs for their hard work. We also thank Catherine Ann Grundmayer, Linda Kaufman, MSN, Shelley Niwa, MA, Anita Soni, PhD, and the members of the B-FIT staff at Westat, Inc, Rockville, Md, for assistance with the field efforts of this study.