Introduction

Purpose in life—the perception that one’s life has meaning, direction, and goals—is a psychological dimension of wellbeing typically measured via the Ryff and Keyes 7-question instrument [1]. Longevity, measured here as all-cause mortality over a 12-year follow-up period, represents the most fundamental health outcome of interest. Physical health—including chronic conditions, functional limitations, and objective measures such as lung function and grip strength—is both a determinant of longevity and a potential source of confounding in studies of psychological predictors of mortality. Purpose, longevity, and health are deeply intertwined: physical health directly affects longevity, physical health shapes psychological states including one’s sense of purpose, and purpose may influence biological processes or health behaviors that affect longevity. Understanding the nature of these interrelationships has significant implications for public health—if life purpose causally improves longevity, it represents a potentially important and low-cost lever for intervention.

Motivated by the potential public health implications of a causal purpose-longevity relation, previous work documents a remarkably strong association between life purpose and longevity [2–19]. A 2016 meta-analysis, for instance, reports that higher purpose is associated with a relative risk of 0.83 for all-cause mortality (p < 0.001) [20]. The purpose-longevity association might arise, however, because (i) high purpose causes lower future mortality risk, or (ii) a decline in baseline health (that, by definition, causes higher mortality risk) either confounds the purpose-longevity relation by causing individuals to respond more negatively to questions about purpose (e.g., “I enjoy making plans for the future and working to make them a reality”) or mediates it by serving as the pathway through which purpose affects longevity. Inconsistent with the confounding or mediation explanations, the literature concludes that purpose has a strong independent effect on longevity beyond baseline health. That is, baseline health confounding or mediation appears to play little role in explaining the purpose-mortality association, as extant work finds purpose predicts longevity after controlling for “baseline health,” “baseline physical health,” and “medical status…and health behaviors” [2,4,8]. For example, a recent study estimates that a two standard deviation shift in purpose is associated with more than 8 years of additional life and that this relation is “… substantially unchanged …” when adjusting for “… health behaviors, activity limitations, and health conditions …” [18].

A concern with the interpretation that purpose causes longevity beyond baseline health, however, is that this conclusion only holds if baseline health is adequately measured. If baseline health is coarsely measured, the correlation between life purpose and longevity may arise because purpose proxies for additional variation in baseline health. That is, if baseline health either confounds or mediates the purpose-mortality relation, then mismeasurement in baseline health will inflate the purpose-longevity relation [21]. This study’s aim is to investigate the role of baseline health mismeasurement in driving the purpose-longevity association by adopting two standard approaches for dealing with a confounding variable: reducing baseline health measurement error and restricting the sample to individuals who begin with more homogeneous baseline health levels [21,22]. If, as the literature suggests, purpose predicts longevity when controlling for baseline health, then the hazard ratios associated with purpose should largely be invariant to variation in measurement error in baseline health or the exclusion of already ill individuals.

Materials and methods

Results

Table 1 reports the results. Baseline health mismeasurement is reduced moving from Model 1 to Model 4. Results for the initial 4-year period (2006–2010, years 1–4), the subsequent 4-year period (2010–2014, years 5–8) which, by definition, is limited to individuals who survive the initial 4-year period, and final 4-year period (2014–2018, years 9–12) which is limited to individuals who survive at least 8 years post baseline, are reported, respectively, in the top, middle, and bottom portions of Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. HRs and 95% CIs of all-cause mortality.

https://doi.org/10.1371/journal.pone.0349401.t001

The patterns in Table 1—both across models and over horizons—are striking. First, for individuals in the lowest purpose category, there is a monotonic decline in hazard ratios associated with purpose when baseline health mismeasurement is mitigated (i.e., moving from Model 1 to Model 4) holding period constant, and over horizon holding model constant (moving from years 1−4 to years 5−8 to years 9−12). Second, the magnitude of the reduction is substantial. For instance, moving from Model 1 to Model 3, the magnitude of the effect size associated with the lowest purpose category falls 53% when examining survival in the first 4 years (i.e., ln(1.77)/ln(3.38)-1) and 68% for years 5−8 (i.e., ln(1.20)/ln(1.78)-1). Third, holding the model constant, the early mortality exclusion also results in a precipitous fall in the magnitude of the hazard ratio associated with low life purpose. For example, in Model 1, the magnitude of the low life purpose hazard ratio falls 53% (ln(1.78)/ln(3.38)-1) when simply limiting the sample to those who survive at least 4 years post baseline and falls more than 100% (i.e., the hazard ratio drops to < 1) when limiting the sample to those who survive at least 8 years post baseline.

Fourth, consistent with mismeasurement of baseline health, functional score predicts longevity only when excluding the more robust objective baseline health metrics. That is, functional score is statistically meaningful only in Model 2 for years 1–4 (HR = 1.23; 95% CI = 1.16, 1.30) and 5–8 (HR = 1.18; 95% CI = 1.11, 1.25). Once incorporating broad limitations, lung function, and grip strength (i.e., Models 3 and 4), functional score is no longer meaningfully related to survival in years 1–4 (HR = 0.99; 95% CI 0.90, 1.09) or years 5–8 (HR = 0.98; 95% CI = 0.89, 1.09). In contrast, broad limitations, lung function, and grip strength continue to meaningfully predict mortality risk. For instance, even after conditioning on survival for at least 4 years post baseline, higher broad limitations (HR = 1.23; 95% CI = 1.09, 1.39), lower lung function (HR = 0.77; 95% CI = 0.70, 0.84), and lower grip strength (HR = 0.85; 95% CI = 0.78, 0.93) all predict greater mortality risk in years 5–8.

To better gauge the potential clinical relevance, we estimate absolute risk estimates for the typical woman and man in our data (see S16 Text 2 and S17 Fig 3 in S1 File). Consistent with the pattern in hazard ratios reported in Table 1, the difference in absolute risks is greatly attenuated when either including the additional health metrics or introducing an early mortality exclusion. For instance, absent any health measures (Model 1), the difference in absolute risks for the typical woman in our sample over the first four years (years 1–4) in the lowest and highest life purpose categories is 4.9% (7.1% for low purpose versus 2.2% for high purpose). The gap between the purpose groups, however, systematically declines as additional health metrics are added to the analysis. Moving from Model 1 to Model 4, for example, the gap falls from 4.9% to 1.1%. Further consistent with the results in Table 1, the patterns in absolute risk greatly attenuate once conditioning on surviving at least 4 or 8 years post baseline. For instance, once conditioning the sample to those that survive at least 4 years post-baseline, the point estimates suggest the typical woman in the lowest life purpose category has lower absolute risk than the typical woman in the second highest life purpose category (based on Models 3 or 4).

We also examine the sensitivity of the baseline health estimates to the inclusion of purpose. Specifically, S11 Table 8, S12 Table 9, and S13 Table 10 in S1 File report hazard ratios for the baseline health metrics for Models 2, 3, and 4, respectively, both with and without the life purpose covariates. Two findings are noteworthy. First, baseline health strongly and independently predicts longevity across all model specifications and time periods. For example, when excluding purpose from Model 3, lung function (HR = 0.76; 95% CI = 0.69, 0.83), grip strength (HR = 0.81; 95% CI = 0.74, 0.90), and broad limitations (HR = 1.27; 95% CI = 1.11, 1.44) all strongly predict mortality risk in years 1–4. Second, the baseline health metric hazard ratios are essentially unchanged whether or not life purpose is included in the model. For instance, the health metric coefficients in Model 3 shift by less than 0.01 in almost all cases across all three periods. The results confirm that although adding baseline health greatly attenuates the purpose-longevity association, adding purpose has essentially no effect on the health-longevity relation.

Our tests in Table 1 follow the approach in Alimujiang et al. [2], which uses life purpose categories rather than a continuous variable to avoid imposing a specific functional form on the relation between life purpose and mortality risk. In the supporting information [see S18, S19 in S1 File], we report results after replacing the purpose categorical variable with a continuous purpose variable. These tests lead to the same conclusions—the relation between purpose and longevity is greatly attenuated when better controlling for baseline health and when employing an early mortality exclusion.

Discussion

The purpose-mortality association systematically attenuates as measurement error in baseline health is reduced by (i) employing more robust measures of baseline health or (ii) excluding individuals who are already seriously ill. Importantly, the strategies we employ reduce, but do not eliminate, residual confounding. For example, because many seriously ill individuals survive for more than 4 or 8 years, the early mortality exclusion will not eliminate residual confounding. Similarly, although moving from Model 1 to Model 4 reduces residual confounding, the more robust baseline health metrics still contain measurement error. For instance, even with our most extended model, the framework does not (a) differentiate testicular cancer (with a 5-year survival rate in excess of 90%) from pancreatic cancer (with a 5-year survival rate less than 10%), (b) differentiate stage I lung cancer (with 5-year survival rate greater than 80%) from stage IV lung cancer (with 5-year survival rate less than 25%), or (c) classify chronic kidney disease as a chronic condition (the 5-year survival rate for a person on dialysis is 35–40%). For instance, purpose will still capture differences in health between an individual given a diagnosis with a 10% 5-year survival rate and an individual given a diagnosis with a 90% 5-year survival rate (assuming such diagnoses result in different answers to questions such as “I enjoy making plans for the future and working to make them a reality”) unless the difference in diagnoses are fully captured by broad limitations, lung function, grip strength, and self-rated health.

In short, if health impacts how respondents answer questions about purpose, then unless one can perfectly measure baseline health, residual confounding remains a potential issue. Thus, for example, an inclination to view the fact that, although the relation between the lowest purpose category and longevity is no longer statistically significant in Models 3 (HR = 1.77; 95% CI = 0.97, 3.22) and 4 (HR = 1.54; 95% CI = 0.85, 2.82) when there is no early mortality exclusion (i.e., the 2006–2010 period), the point estimates have the predicted pattern and the confidence intervals are wide, must be tempered with the knowledge that all 4 models still contain residual baseline health confounding.

Beyond the issue of residual confounding, our prospective cohort study framework does not differentiate baseline health confounding from baseline health mediation of the purpose-longevity relation, as both produce observationally identical patterns of attenuation [30]. An alternative interpretation is that baseline health mediates rather than confounds the relation between purpose and longevity. There are, however, three important considerations associated with the mediation interpretation. First, if baseline health mediates the relation, then any control for baseline health is an overadjustment and resulting purpose hazard ratios underestimate the total effect size of purpose on longevity [31,32]. Thus, under the mediation interpretation, Model 1 purpose estimates are less biased than those in Models 2, 3, and 4, and Model 1 estimates are still too conservative if the remaining covariates (e.g., BMI) also capture baseline health (S20 Table 13 in S1 File reports the hazard ratios associated with purpose excluding all control variables). Even if Model 1 estimates are too conservative, the magnitude of the life purpose causal effect is astounding under the mediation interpretation. For example, based on the Model 1 estimates (and as shown in S17 Fig 3 in S1 File), holding other factors constant, a man in the highest life purpose category exhibits an absolute mortality risk of 4.1% in years 1–4, while a man in the lowest life purpose category exhibits an absolute mortality risk triple (12.3%) that value.

Second, the mediation interpretation requires a reinterpretation of extant evidence—if baseline health fully mediates the purpose-longevity relation then, by definition, purpose does not predict longevity once controlling for baseline health as claimed in the literature. Moreover, if baseline health fully mediates the relation, then baseline health appears to be the causal channel. Thus, evidence to support other (e.g., future behavioral, biological, or stress-buffering) causal channels would require either (i) evidence that these variables mediate the purpose-longevity relation when perfectly (or at least better) controlling for baseline health (i.e., these factors are additional mediators), or (ii) evidence that these factors mediate the purpose-baseline health-longevity relation (e.g., causation runs from purpose to baseline health to future behavior to longevity).

Third, an arguably more reasonable interpretation is that baseline health both confounds and mediates the purpose-longevity relation. That is, the relation between purpose and baseline health is bidirectional [8]—purpose causes better baseline health which leads to longer lives (i.e., baseline health mediates the purpose-longevity relation) and better baseline health causes an individual to respond more favorably to questions about purpose (i.e., baseline health confounds the purpose-longevity relation). Because this is just a combination of the two explanations discussed above, this interpretation is also inconsistent with the literature’s conclusion that purpose predicts longevity after accounting for baseline health.

Perhaps more puzzling is that if the relation is bidirectional, it is so in a peculiar fashion as the evidence in Table 1 suggests that, regardless of the extent of controlling for baseline health, the purpose categories are monotonically related to longevity only in the 4 years immediately following baseline. In contrast, the Model 3 lung function hazard ratio is 0.76 (95% CI = 0.69, 0.84) for years 1–4, 0.77 (95% CI = 0.70, 0.84) for years 5–8, and 0.81 (95% CI = 0.74, 0.88) for years 9–12. Thus, under the mediation explanation, however purpose affects baseline health, that portion of baseline health caused by purpose only predicts near-term longevity. For instance, if baseline health mediates the relation between purpose and longevity because higher purpose is associated with the behavior of more frequent checkups, then the benefits of these more frequent checkups only appear in the short-term. Such an interpretation appears to contrast with evidence that the benefits of checkups primarily accrue over longer horizons [33].

Related, the early mortality exclusion may result in attenuation because purpose (either independently or through baseline health) has the strongest protective effects for individuals who do not survive at least 4 or 8 years post baseline. Nonetheless, the early mortality exclusion is a well-recognized method of controlling for baseline health in both the life purpose and broader literatures. Specifically, although several purpose-longevity studies [2,6,16] include a 1- or 2-year early mortality exclusion in their sensitivity analysis, extant work suggests that a substantially longer exclusion period is needed to better control for baseline health as most serious illnesses have survival rates beyond 1–2 years [34,35]. For example, a recent meta-analysis [36] of the relation between BMI and mortality limited the sample to 189 studies that had a minimum of 5-year early mortality exclusion to minimize baseline health confounding and a recent study [21] investigating the relation between physical activity and longevity considers a 4-year early mortality exclusion.

We also consider a wide range of sensitivity tests that demonstrate the robustness of our conclusions including a detailed investigation of the role of multicollinearity (S21, S22, S23, S24, S25, and S26 in S1 File), updating only health or purpose (S27 and S28 in S1 File), updating both health and purpose (S27 and S29 in S1 File), estimating Model 2 for the sample that does not require the additional health variables (S30 in S1 File), adding the psychological status variables used in the Alimujiang [2] extended model (S31 in S1 File), comparing the relations over years 1–2 versus 3–4 (S32 and S33 in S1 File), partitioning the sample into those with and those without a chronic condition at baseline (S34, S35, and S36 in S1 File), partitioning the sample by respondent age (S34, S37, and S38 in S1 File), and using raw (rather than age- and gender-adjusted) lung function and grip strength (S39 in S1 File). As detailed in the supporting information, regardless of how we frame the test, the results uniformly support the hypothesis that the purpose-longevity relation is greatly attenuated when minimizing measurement error in baseline health.

Our results demonstrate that the relation between purpose and longevity greatly attenuates when better capturing heterogeneity in baseline health, both through more robust baseline health measures and through an early mortality exclusion. At the same time, because our study is based on the data, methodology, and structure of a single previous study [2], it shares that study’s limitations. It is possible that alternative research designs or data may find evidence that purpose predicts longevity even when better controlling for baseline health. That is, the “absence of evidence is not evidence of absence” [37]. Nonetheless, given only coarse metrics of baseline health have been used in this literature, to claim a causal relation beyond that captured by baseline health, future work would need to demonstrate that the relation persists when better controlling for baseline health [20].

Conclusions

Our evidence suggests the well-documented association between purpose and longevity arises from either (i) residual baseline health confounding, (ii) baseline health fully (or nearly fully) mediating the purpose-longevity relation, or (iii) some combination of (i) and (ii). All three interpretations are inconsistent with the widely held view that purpose causes longevity when controlling for baseline health.

The literature’s consensus that life purpose causes longevity appears to be so strong that recent reviews not only propose that future work should focus on understanding how purpose causes longevity (i.e., through behavioral, biological, and stress-buffering channels) rather than if purpose causes longevity, but also advise practicing health care professionals to focus on purpose to mitigate their patients’ mortality risk [5,7,8,38–43]. Moreover, governments and non-profit organizations are investing in life purpose intervention programs based on the perceived causal relation between purpose levels and subsequent morbidity. For instance, based on the interpretation in the extant literature, the American Heart Association identifies purpose (and well-being) as a new “Noteworthy” focus of their 2030 Impact Goals. Thus, understanding the causal relation between purpose, baseline health, and longevity is critical as, at best, this guidance is premature. At worst, if these findings replicate across other datasets and research designs, resources are being poorly allocated, and life purpose interventions provide potentially false guidance to patients, health care professionals, and those who influence public policy.

Supporting information

Acknowledgments

RS and HT contributed equally to this work.