Introduction

“The impetus to foretell death is both fundamental and ancient”

(p. 23) [1]

Laymen, social scientists, and health researchers have long shared a fascination with predicting a person’s longevity or future age at death. In a modern day version of fortune telling, the internet is replete with websites and “apps,” with such evocative names as “Death Clock,” “Death Timer,” and “Deadline,” designed to predict age at death or time remaining before death on the basis of a battery of simple questions.

Clinicians have developed more scientifically based prognostic tools to predict how long a person will survive, most often a patient who is severely ill. Such estimates help physicians assess the costs and benefits of interventions and assist patients and families in making healthcare decisions. Nevertheless, clinicians often avoid making prognoses—not only because of the unpleasantness, anxiety, and sadness that they experience when communicating such information to patients, but also because they are aware of the enormous uncertainty inherent in their predictions [1].

Social scientists and epidemiologists have largely relied on statistical models to predict mortality from a diverse set of information. Based on these models, researchers have identified variables that show strong associations with the probability of dying in a follow-up period, typically without making explicit forecasts. The explanatory variables in these models comprise both proximate determinants of survival (e.g., disease and disability) and more distant demographic, social, economic, and environmental factors. Information about these predictors is typically obtained from responses to questions in household surveys of middle-aged and older adults, with follow-up data on survival generally derived from registration systems (see, for example, [2–4]).

The criteria used to develop these statistical models and prognostic tools differ substantially across researchers. One explanation for this variability is disciplinary myopia: scholars often focus on measures most prominent in their discipline, thereby failing to consider the relative importance of a comprehensive set of predictors. It is also the case that, until the recent proliferation of biosocial surveys [5, 6], very few datasets incorporated a sufficiently broad set of demographic, social, health, and clinical variables along with linkages to subsequent mortality. In this paper, we use the extensive information collected in four recent biosocial surveys—in England, the US, Costa Rica and Taiwan—each of which is based on a large, national sample. We identify the best predictors of survival and assess robustness of these findings with multiple measures of discrimination. We address two specific questions. First, what factors matter most in determining whether an older adult will die in the next five years? Second, are the principal predictors the same across countries with similar levels of life expectancy but distinct social environments, levels of affluence, and access to health care?

Materials and Methods

Analytical Strategy

To minimize selection biases resulting from missing data, we followed standard practices of multiple imputation to handle missing data (see S2 Appendix) [7, 8].

To account for differential response rates and oversampling, we weighted descriptive statistics (S5 Table). For ELSA, given limited information regarding the timing of deaths, we used a logit model (i.e., with death by the end of 2009 as the outcome). For NHANES, SEBAS, and CRELES, we estimated a Cox hazard model using age as the “clock” to predict age-specific death rates for the five-year period post-survey, thereby allowing the underlying baseline hazard over age to assume whatever functional form best fits the data. All models adjusted for age and sex and were fit separately by country using unweighted data. The use of sampling weights in multivariate regression analysis remains a matter of debate among statisticians (see, for example, [9]). For the rankings presented here, we added each of the remaining predictors one at a time to this model. In the first stage of the analysis (Figs 1 & 2), we examined only variables available for all four surveys; in the second stage (Fig 3), we considered additional variables collected in only some of the surveys.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Top Ten Predictors (out of 25 variables available in all four datasets) of Five-Year Mortality After Adjustment for Age and Sex, Ranked by the Gain in AUC, A) Costa Rica, B) England, C) Taiwan, and D) the U.S.

Abbreviations: ADL, Activities of daily living; AUC, Area under the receiver operating characteristic curve; CRP, C-reactive protein; Depress, Depressive Symptoms; Educ, Educational attainment; HbA1c, Glycosylated hemoglobin; Heart dis., Heart disease; IADL, Instrumental Activities of Daily Living; SAH, Self-assessed health status, Social Int., Social integration. ^a The values for diabetes and depressive symptoms in Taiwan are too close to be distinguishable.

https://doi.org/10.1371/journal.pone.0159273.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Environmental Predictors of Five-Year Mortality After Adjustment for Age and Sex, Ranked by the Gain in AUC, A) Costa Rica, B) England, C) Taiwan, and D) the U.S.

Abbreviations: Educ, Educational attainment; Social Int., Social integration.

https://doi.org/10.1371/journal.pone.0159273.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Additional Predictors of Five-Year Mortality that Yield a Gain in AUC≥0.01 After Adjustment for Age and Sex, A) Costa Rica, B) England, C) Taiwan, and D) the U.S.

The unlabeled markers represent variables displayed in Fig 1. Abbreviations: 5+ Meds, 5+ Medications; AUC, Area under the receiver operating characteristic curve; Chair std., Chair stand speed; Church Attend., Church attendance; Fbg, Fibrinogen; Gait sp., Gait Speed; Grip, Grip strength; Hcy, Homocysteine; Hosp stays, Number of hospitalizations; IAH, Interviewer-assessed health status; IL-6, Interleukin-6; PEF, Peak expiratory flow; sE-selectin, Soluble E-selectin; sICAM-1, Soluble intercellular adhesion molecule 1.

https://doi.org/10.1371/journal.pone.0159273.g003

We tested interaction terms between each predictor and age (linear) to allow for variation in the effect of the predictor across age (i.e., non-proportional hazards). If p<0.05 (two-tailed test), we included the interaction along with the main effect for that predictor (see S6 Table, which lists the predictors with non-proportional hazards).

We present the hazard ratios (odds ratios for ELSA) for the top 10 predictors in each country in S7 Table. However, because our objective is to assess the predictive ability of variables rather than the magnitude of the associations, we rely on measures of discrimination to determine the rankings of predictors. Results are based on the most frequently used discrimination measure, the area under the receiver operating characteristic curve (AUC). The receiver operating characteristic (ROC) curve is a graph of sensitivity (i.e., true positive rate) against 1-specificity (false positive rate), calculated for each possible cutoff point for distinguishing between two groups of individuals (e.g., alive or dead). The area under that curve summarizes how well the model discriminates between decedents and survivors. Although there is no scientific basis for evaluating the magnitude of the AUC, Pencina et al. [10] suggest that increases of at least 0.01 denote meaningful improvements. We identify this threshold of 0.01 in our graphical results (Figs 1 & 2).

To assess the robustness of our conclusions, we present the rankings of top predictors based on the AUC as well as on two additional measures of discrimination: a category-free version of the Net Reclassification Improvement (NRI(>0)) and the Integrated Discrimination Improvement (IDI) (see S1–S4 Figs, which show the top ten predictors ranked by the AUC, NRI(>0), and IDI) [10, 11]. See S3 Appendix, which describes how the measures of discrimination are calculated.

Results

To assess the improvement in the AUC, all models are compared with a basic model—one that uses age as the time metric and controls only for sex. The AUC values for this model are 0.766 for Costa Rica, 0.776 for Taiwan, 0.788 for the U.S., and 0.808 for England.

Discussion

To the best of our knowledge, this is the first multinational study to evaluate a wide spectrum of mortality predictors. The findings are consistent for four countries with similar levels of life expectancy (between 78.8 and 80.6 years in 2010) [12, 13], but with different cultural traditions, levels of economic development, and epidemiologic transitions over recent decades. Although only future work can establish the reproducibility of our results, the consistency of findings across countries is very encouraging, particularly in light of increasing concern about the inaccuracy and exaggeration of published results in science and medicine [14–16].

Among the variables available in all datasets, three self-reported measures of health are among the strongest predictors in all countries: IADL limitations, mobility limitations, and the simple question on overall health status (SAH). These findings, based on the AUC, are robust to two alternative measures of discrimination (S1–S4 Figs).

CRP and other inflammatory markers are notable mortality predictors, which supports the insight from Glare and colleagues that “novel prognostic factors,” including CRP and cytokines, may prove useful in recalibrating physicians’ survival predictions [17]. Two other biomarkers—serum albumin and homocysteine—are important predictors in the two countries for which they are available.

Results from Taiwan suggest that the interviewer’s rating is a better mortality predictor than the rating of overall health provided by the respondent. Interviewers may base their evaluations not only on information directly obtained in the survey, but also on their observations of the respondent’s appearance, responsiveness, cognition, and disposition [18]. Although the physicians’ ratings are weak predictors, these physicians were seeing respondents for the first time and basing their judgments on only a routine physical examination and a medical history completed by the respondent. Our findings suggest that inexpensive, easily obtained health evaluations by external observers—including lay people, caregivers, and medical personnel—may enhance prediction.

The dominance of health measures in the set of leading predictors is not altogether surprising: researchers would anticipate that measures of health—being more proximate to death—should yield stronger predictions than behaviors and social and demographic factors. Nevertheless, exercise frequency appears on the top 10 lists for three countries, sometimes above measures of disease prevalence and biomarkers. This finding raises an important caveat, namely that our analysis cannot be used to infer causal linkages between the predictors and survival. For example, exercise may discriminate well partly because of reverse causality (poor health leads to reduced exercise). However, such endogeneity is not a shortcoming when our goal is to identify warning signals (i.e., prognosis) rather than potentially modifiable risk factors.

In recognition of the importance of integrating prognosis into clinical decisions, Yourman and colleagues [19] identified 16 validated non-disease-specific prognostic indexes that forecast mortality risk for older adults over a period between six months and five years into the future (available at eprognosis.ucsf.edu). Concerns have been raised regarding bias, quality and generalizability of these instruments [19]. Our findings reveal additional limitations: (1) only one prognostic index incorporates SAH, despite its simplicity and predictive ability; (2) the indexes typically include several IADL or mobility limitations, but rarely the full battery typically collected in household surveys; (3) the indexes often include physicians’ assessments, rather than patients’ reports, of functional limitations, even though physicians vastly underreport functional limitations [20]; (4) no index explicitly includes information obtained from lay persons (e.g., relatives or interviewers); and (5) no index incorporates the biomarkers that we identified as strong predictors (i.e., CRP and other inflammatory markers, serum albumin, and homocysteine).

We recognize that the leading predictors of survival are likely to differ between samples of seriously ill patients—a prime concern in the clinical literature—and general population-based samples of older adults analyzed by epidemiologists and social scientists. For example, indexes on eprognosis.ucsf.edu intended for patients in medical facilities encompass a shorter time frame of prediction and more clinical variables (e.g., symptoms of disease) than those for community residents. These differences highlight the need to develop instruments for the appropriate population and time frame. Still, non-clinical variables (e.g., social contact and support, economic status, perceptions) are likely to predict survival among all individuals, seriously ill or not, underscoring the need for self-reported information in all settings [1].

A major strength of this analysis is that we bridge the division between clinicians and researchers interested in mortality prediction and employ a more diverse set of predictors than previous studies. Our findings are based on large nationally representative samples that span age ranges capturing the vast majority of deaths and are robust across countries and measures of discrimination. Future work is needed to determine the generalizability of our findings to younger ages and to assess variation in the leading predictors by other demographic characteristics, most notably sex. The objectives of this analysis have not required model building but this is an important next step, especially for applications that require explicit forecasts. In particular, given potential correlations between predictors and the desire for concise instruments, future research is needed to incorporate these strong predictors into parsimonious indexes that are validated with external data.

Identification of the most powerful predictors of survival is an important undertaking for social scientists, epidemiologists and clinicians. Death is a commonly used, and well-measured, endpoint in studies of environmental influences on health. Such studies place increasing emphasis on understanding the behavioral and biological pathways that underlie social stratification in health and survival and the impact of stressful experience. From a clinical perspective, long periods of illness—which characterize high life expectancy populations—underscore the need to identify individuals at high risk of death. Prognoses made by physicians and other health professionals for seriously ill patients have generally been overly optimistic [17, 21, 22]. Given advances in medical technology, growing acknowledgment of iatrogenic consequences of treatment, and a desire for cost-effective health care, optimal decision-making requires accurate prognosis of survival. Better prognostic tools may encourage physicians to have conversations with their patients about their survival prospects. Despite the inherent difficulties, such discussions may empower individuals to have more control over their final years of life and reduce unnecessary treatment that may compromise end-of-life quality.

Supporting Information

Acknowledgments

We are grateful to Germán Rodríguez for statistical advice and programming assistance; the many investigators, staff, and fieldworkers at the Universidad de Costa Rica, Centro Centroamericano de Población, and Instituto de Investigaciones en Salud who made the CRELES study possible; and the staff at the Surveillance and Research Division (Health Promotion Administration at the Ministry of Health and Welfare in Taiwan) who were instrumental in the design and implementation of the SEBAS and supervised all aspects of the fieldwork and data processing.