Genetic, Physiological, and Lifestyle Predictors of Mortality in the General Population

METHODS

The Rotterdam Study is a population-based prospective cohort initiated in 1990. It was designed to investigate risk factors for diseases in 7983 participants aged 55 years or older in the Ommoord district of Rotterdam, The Netherlands.^8,9 In the initial and subsequent investigation waves, trained research assistants collected data on health, medication use, medical and family history, and lifestyle factors in extensive home interviews. Participants subsequently visited the research center for clinical examinations, with a special emphasis on imaging and on collecting and storing biospecimens to facilitate in-depth molecular and genetic analyses. Data analyzed in this study concern 5974 participants with genetic information available from the first wave of the Rotterdam Study.

RESULTS

At baseline, participants were on average 69 years of age (range = 55–99 years; Table 1) and 59% were female. Of the 5974 participants, 3174 died (mortality rate = 4.2 per 100 person-years) during the follow-up period. From 162 a priori identified risk factors of mortality (supplemental Table C), backward regression retained 108 variables independently in the final model. Of these, 36 were significantly related to mortality (P < .05) as independent risk factors (Table 1). Age (hazard ratio [HR] = 1.09; 95% confidence interval [CI] = 1.08, 1.10) and female gender (HR = 0.71; 95% CI = 0.62, 0.81) were strongly associated with mortality (Table 1).

Of the candidate genes, a priori identified in the literature, APOE, insulin-like growth factor 1 receptor (IGF1R), and Werner syndrome, RecQ helicase-like (WRN) were significant and independent predictors of mortality. From the 50 independent loci, identified from the meta-analysis of 8 discovery genome-wide association studies on time to death, 2 SNPs in the neighborhood of tripartite motif-containing 32 (TRIM32) and methionine adenosyltransferase II, β (MAT2B) were associated with mortality.

Social class and living in serviced housing were independently associated with risk of death. Smoking status and pack-years as well as energy intake were also associated with mortality. The physiological measures blood pressure, body mass index, and waist circumference, and, in particular, the risk indicators assessed in blood (such as erythrocyte sedimentation rate, leukocytes, creatinine, C-reactive protein, and total cholesterol) or with imaging (such as bone mineral density of the femoral neck and aortic calcification) were all independently related to mortality. Diabetes, cardiac diseases, Parkinson’s disease, cancer, and cognitive function remained independently associated with death. Self-perceived comparative health was a good indicator of mortality risk, as were unintentional weight loss and serious illness in the past 5 years.

The predictive power of the variable groups is best explained in 2 ways. First, Figures 1 and 2 show the development of predictive quality over time. For each point during follow-up, the graphs depict the respective time-dependent area under the curve of a given variable group. Next, Table 2 quantifies the predictive quality for 5 specific prediction intervals (1, 3, 5, 10, and 15 years).

Figure 1 shows that over time, all variable groups except genetic risk markers exhibited decreasing ability to predict death. Figure 1 and Table 2 also demonstrate that prediction based solely on age and gender consistently outperformed all other groups of variables (C-statistic for 15-year follow-up [C_15y] = 0.76; 95% CI = 0.75, 0.77; Table 2). Physiological risk and socioeconomic characteristics each predicted mortality equally well over 15 years (each C_15y = 0.72; 95% CI = 0.71, 0.73). Although of significantly less predictive quality, the C-index of genetic risk markers was still better than chance (C_15y = 0.56; 95% CI = 0.55, 0.57).

Figure 2 shows the performance of the age- and gender-adjusted model compared with the fully adjusted model (C_15y = 0.80; 95% CI = 0.79, 0.81). Whereas adding socioeconomic and lifestyle information to age and gender (C_15y = 0.77; 95% CI = 0.77, 0.78) only slightly improved the predictive quality, the combination of age, gender, general health, disease, and physiology (C_15y = 0.79; 95% CI = 0.78, 0.80) predicted mortality almost as well as all covariates that remained after backward regression.

To allow comparison with other studies of cause-specific mortality and different population health status, we report the associations of the final model stratified by prevalent, baseline disease status and for cardiovascular disease mortality in Table D (available as a supplement to the online version of this article at http://www.ajph.org).

DISCUSSION

From a set of 162 established risk factors and candidate SNPs for mortality, we identified 36 (31 nongenetic, 5 genetic) independent and significant predictors of mortality. Specific genetic factors were independently associated with mortality and jointly predicted mortality better than chance. However, genetic information added little to age, gender, and other traditional predictors of mortality.

This analysis confirms prior findings that multiple diseases, as well as socioeconomics and lifestyle, jointly influence mortality in the aging adult population.^3,6 Numerous predictors remained independently and significantly associated with mortality. Although several markers of prevalent disease remained associated with mortality, their risk ratios were attenuated. Others, such as prevalent dementia, cerebrovascular accidents, and transient ischemic attacks, were not more effective at predicting mortality than indicators of disease severity and subclinical disease such as the Mini-Mental State Examination and serum C-reactive protein.

Several specific SNPs were independently associated with mortality in our analysis. In accordance with the literature, each additional copy of the APOE ε4 allele increased mortality in our study cohort.^1,36,37 Similarly, the IGF1R and WRN genes have been described before as being associated with longevity and aging, respectively, via improving stress resistance, innate immunity, metabolic maintenance and repair of DNA.¹ This study also confirmed 2 novel loci in the vicinity of the TRIM32 and MAT2B genes as being associated with death. These 2 loci were identified from the pool of SNPs identified by the meta-analysis of the discovery cohorts in the genome-wide association studies. These genes have recently been associated with cancer proliferation.^38,39

Interestingly, unlike the predictive ability of all other domains, the ability of genetic markers to predict death remained constant during the course of follow-up. The stability of the predictive power of the SNPs observed in this study is probably due to the permanent nature of the genetic makeup. Other variables showed decreasing predictive power over time that can be explained by changes in the value of these variables during the course of follow-up.

Our results support the view that specific SNPs can be identified that are associated with mortality and might be used for risk prediction.^40,41 The results also show, however, that these common SNPs have very limited predictive power and that, especially when used in combination with traditional risk factors, they contribute very little, if anything at all, to improve the prediction of death in the general population aged 55 years and older.

Another finding relates to age and gender as predictors of mortality. In our study, the relative risk for mortality per year of age in the univariate model was only reduced by 25% in the fully adjusted model. This is compatible with the idea that aging is not merely the clinical manifestation of disease but an underlying, disease-independent accumulation of pathophysiological changes that favor mortality over time.^42–44

The gender differences in this study were not due to differences in prevalent diseases at the onset of the study. Females exhibited even stronger reduced risks after adjustment for other risk factors. This strongly suggests that the gender difference in survival cannot be explained by differences in health behavior and disease at baseline.⁴⁵ We can only speculate that the survival benefits of females can be found in gender effects or different genetic origins not accounted for in this analysis. One of the potential genetic candidates that could contribute to the female survival advantage is the X chromosome.⁴⁶ Another genetic candidate is mitochondrial DNA (mtDNA). It has been suggested that, because mtDNA is inherited from the maternal line, a possible intergenomic conflict between mtDNA and nuclear DNA favors female survival.⁴⁷ We could not include markers on the X chromosome and mtDNA because the X chromosome is not commonly analyzed in all the discovery cohorts and mtDNA is also not available on all genotyping platforms.

Prediction of mortality by age and gender improved by only 5% upon inclusion of all independent mortality predictors over the entire 15 years of follow-up. This is particularly interesting considering the multitude of independent risk factors identified in this study, and because all groups contributed significant variables. From the time-dependent area-under-the-curve analyses, 2 observations are particularly noteworthy: first, most of the additional information beyond age and gender stemmed from indicators of physiology, disease, and general health; second, although socioeconomic factors were equally good at predicting mortality as indicators of disease, the combination of socioeconomics, lifestyle, and genetic markers contributed visibly to the explanation of mortality risk only after 10 years of follow-up and beyond. Thus, although socioeconomics and lifestyle were associated to mortality, they seemingly exerted their effects on mortality through physiological risk indicators and disease rather than by acting independently on mortality.⁴⁸ This underscores the importance of socioeconomics and living conditions for public policy aimed at reducing health inequalities.

To summarize the findings reported here and in other studies, one can insinuate a cascade from gene to individual health to death, in which every step is accompanied by environmental influences, some of which are controlled by the individual (such as physical activity and obesity) whereas others are defined by the individual’s living circumstances, cultural heritage and surroundings. Figure A (available as a supplement to the online version of this article at http://www.ajph.org) illustrates at which stage during the course of aging different interventions (e.g., improvements in living circumstances or the introduction of a new therapeutic drug) can feasibly act and which health gains could be expected.