Two Approaches to Classifying and Quantifying Physical Resilience in Longitudinal Data

Abstract

Background

Approaches for quantifying physical resilience in older adults have not been described.

Methods

We apply two conceptual approaches to defining physical resilience to existing longitudinal data sets in which outcomes are measured after an acute physical stressor. A “recovery phenotype” approach uses statistical methods to describe how quickly and completely a patient recovers. Statistical methods using a recovery phenotype approach can consider multiple outcomes simultaneously in a composite score (eg, factor analysis and principal components analysis) or identify groups of patients with similar recovery trajectories across multiple outcomes (eg, latent class profile analysis). An “expected recovery differential” approach quantifies how patients’ actual outcomes are compared to their predicted outcome based on a population-derived model and their individual clinical characteristics at the time of the stressor.

Results

Application of the approaches identified different participants as being the most or least physically resilient. In the viral respiratory cohort (n = 186) weighted kappa for agreement across resilience quartiles was 0.37 (0.27–0.47). The expected recovery differential approach identified a group with more comorbidities and lower baseline function as highly resilient. In the hip fracture cohort (n = 541), comparison of the expected recovery differentials across 10 outcome measures within individuals provided preliminary support for the hypothesis that there is a latent resilience trait at the whole-person level.

Conclusions

We posit that recovery phenotypes may be useful in clinical applications such as prediction models because they summarize the observed outcomes across multiple measures. Expected recovery differentials offer insight into mechanisms behind physical resilience not captured by age and other comorbidities.

Physical resilience, defined as one’s ability to resist decline or recover function following a physical stressor, has received recent attention as a key factor in successful aging (1). Although psychological and social stressors such as discrimination and bereavement are also associated with functional decline, recent work has focused on physical stressors to facilitate identification of underlying mechanisms of recovery (2). Physical resilience is posited to be a whole-person level characteristic that cuts across organ systems; if proven, then interventions that enhance physical resilience have the potential to improve multiple outcomes in response to a variety of different stressors (2). Further, physical resilience is hypothesized to be associated with healthy biology in the “pillars of aging,” a set of seven intertwined processes that have been proposed as drivers of age-related change (3). Determining whether biomarkers of these “pillars” are associated with resilient outcomes may elucidate mechanisms and identify new interventional targets.

A conceptual model for physical resilience has been proposed further identifying stressor characteristics, psychological characteristics, social support, and environmental factors as critical components to recovery after a health stressor (Figure 1) (1). Because healthy biology in the pillars of aging affects multiple tissues and organ systems, measures of physical resilience will ideally include multiple outcomes within each individual. Statistical approaches that can incorporate multiple outcomes are, therefore, needed.

Before research to understand mechanisms or improve physical resilience can proceed, we must first define resilient phenotypes in a variety of stressor types. Although it is hypothesized that mechanisms of physical resilience can be identified at the cellular level, individual stressor types frequently affect some organ systems more than others, resulting in distinct and measurable “resiliencies” (eg, immune recovery after chemotherapy or musculoskeletal recovery after an injury). Defining resilience requires identification of a stressor followed by multiple functional observations over time; therefore, using existing longitudinal data sets is a logical place to start. Different approaches for defining resilience may result in individuals being variably defined as more or less resilient and have important implications for subsequent studies. The objective of this article was to describe two different conceptual approaches that could be used to define levels of resilience, the statistical methods that can be used within each approach, and their potential applications in gerontological research.

The Recovery Phenotype Approach

At one level, researchers and clinicians simply wish to describe how rapidly and completely a patient “bounces back” from a physical stressor. We refer to this approach for identifying physical resilience as the “recovery phenotype.” To date, most longitudinal studies have reported change in a single outcome at a single time point (eg, “recovery of prior ambulatory status 1 year after hip fracture”). However, multiple parameters can be used to provide more nuanced descriptions of outcome measures over time, as depicted in Figure 2A. Some individuals may not experience any decline at all following a physical stressor. Such individuals could be separately classified as “resistant,” or be considered as having the highest level of physical resilience based on their lack of initial decline; the approaches described here could be operationalized in either way. Newer statistical techniques such as latent class profile analysis allow us to consider multiple outcomes simultaneously, defining groups of patients with similar recovery patterns across multiple domains (eg, ambulation, function, and cognition) and predictors of group membership. Those with the greatest recovery across multiple domains are defined as the most highly resilient.

The recovery phenotype approach is grounded in clinical thinking. One starts with actual observed clinical events and then uses statistical techniques to summarize or classify the recovery patterns. This method is, therefore, most appropriate for prognostic models in clinical practice and in identifying high and low-risk groups for health services research.

The Expected Recovery Differential Approach

At another level, we are also interested in identifying patients who recover faster or more completely than we expect them to given their demographic characteristics, comorbidities, environmental supports, and other factors. In other words, there may be value in identifying patients whose actual outcome is better than their predicted outcome, regardless of the ultimate functional level achieved. We refer to this approach for quantifying physical resilience as the “expected recovery differential,” or the difference between an individual’s actual and predicted outcomes (Figure 2B).

The expected recovery differential approach is grounded in statistical thinking. Beginning with a cohort containing clinical characteristics and outcomes, one derives a statistical model that estimates the impact of clinical characteristics on outcome trajectories after a particular stressor. Next, this model uses individual patient’s clinical characteristics to estimate expected recovery for each patient. Finally, each patient’s expected recovery is compared with the actual observed outcome or slope of recovery. This approach may be the most appropriate for identifying biological mechanisms underlying physical resilience beyond those determined by age, demographic factors, and comorbidities.

In some cases, these two approaches will assign different levels of physical resilience to the same patient. For example, consider a frail 80-year-old man who ambulates independently at baseline but has severe cognitive impairment and multiple comorbidities and is therefore not expected to regain ambulatory ability after hip fracture. If this patient is eventually able to walk with assistance, he would be classified as “resilient” by expected recovery differential because he did better than anticipated but not by recovery phenotype because he did not reach his baseline ambulatory level. Conversely, a healthy active 65-year-old patient who rapidly recovers to baseline would be classified as highly resilient by recovery phenotype but not by expected recovery differential because her recovery was expected. Therefore, the most resilient group in studies using the recovery phenotype approach is likely to be comprised primarily of the youngest and healthiest individuals, whereas studies using the expected recovery differential will include a greater proportion of older but robust individuals, as well as those with substantial comorbidities or frailty who nonetheless recover well.

In the sections that follow, we apply these approaches for quantifying physical resilience to two cohorts of older adults experiencing two different physical stressors: hip fracture and hospitalization for viral respiratory infection. We describe the statistical methods used to define/classify level of physical resilience to illustrate how methods should be modified to accommodate different clinical situations and types of data. We report initial “proof of concept” evidence that the expected recovery differential approach may be identifying an underlying characteristic that cuts across multiple functional domains. This would support our hypothesis that the expected recovery differential may be useful in isolating underlying biological mechanisms that support resilience—beyond age, comorbidity, and other known risk factors. Modeling approaches are described at a conceptual level so that their clinical and research applications are more apparent.

Methods

Description of the Cohorts

Baltimore hip studies

To define phenotypes of musculoskeletal recovery after hip fracture, we used data from three cohorts of the Baltimore Hip Studies (BHS, n = 541) (4–6), a series of prospective studies enrolling community-dwelling patients aged 65 years and older admitted for surgical repair of a nonpathological hip fracture to one of eight hospitals in Baltimore. Patients with moderate or severe cognitive impairment (Mini-Mental Status Exam score ≤ 20) were excluded from two of the cohorts but were included in the third. Two cohorts included only women and one cohort included men and women.

Informed consent was obtained within 15 days of the hip fracture or hospital admission, depending on the study. After informed consent, subjects and/or their surrogates provided demographic information and self-reported prefracture medical, cognitive, affective, and functional status using validated tools, including activities of daily living (ADLs) (7), physical activities of daily living (PADLs) (7), and Yale Physical Activity Scale (8). Trained chart abstractors recorded information about hip fracture type, anesthesia and surgical approach, and hospital events, including incident delirium and rehabilitation received. Patients were then evaluated in their home or current place of residence at 2, 6, and 12 months post-fracture/admission. Self-reported functional measures were repeated, and physical performance tests administered including gait speed, balance (short physical performance battery (9) or Tinetti (10)), grip strength, chair stand test (11), and Lower Extremity Gain Scale (12). Subjects were asked to wear an activity monitor to assess physical activity. Two of the cohorts randomized subjects to a physical activity intervention (5,13). Subject characteristics and primary functional outcomes are shown in Table 1.

Acute respiratory infection cohort

To define phenotypes of recovery after presentation with acute lower respiratory infection, we used data and samples selected from the Community Acquired Pneumonia and Sepsis Outcome Diagnostics study (n = 186), comprised adults admitted to one of four hospitals from 2003 to 2009, or to the Duke Emergency Department lower respiratory infection cohort (14, 15). Patients were selected for the current analysis if they were aged 60 years and older with a clinical presentation consistent with an acute lower respiratory infection, defined by the presence of at least two qualifying symptoms or one qualifying symptom and at least one qualifying vital sign abnormality. Qualifying symptoms included headache, rhinorrhea, nasal congestion, sneezing, sore throat, itchy or watery eyes, conjunctivitis, cough, shortness of breath, sputum production, chest pain, or wheezing. Qualifying vital signs included tachycardia with heart rate more than 90 bpm, tachypnea with respiratory rate more than 20 breaths per minute and temperature less than or equal to 36°C or at least 38°C. Patients were excluded for known or suspected coinfection at any other anatomic site (14–18). After informed consent, patient’s demographics, past medical history, and physical examination were recorded. Microbiological workup and treatment were managed by the health care provider as deemed clinically appropriate. Retrospective clinical adjudication of patient outcomes was performed at least 28 days following enrollment per study protocol (14,15).

Sputum was obtained for gram stain, culture, and polymerase chain reaction of common viral pathogens. A panel of three infectious disease specialists independently classified each infection as bacterial, known viral, probable viral, or undetermined etiology using established criteria; disagreements were resolved by consensus. For this analysis, we also restricted the cohort to patients with known or probable viral infections to reduce the heterogeneity of stressor type. Length of hospital stay, intensive care unit (ICU) admission, death within 28 days of presentation, and discharge location were recorded as outcome measures. Subject characteristics are shown in Table 1.

Methods for Defining Recovery Phenotype

The two cohorts described earlier have very different outcome measure types, frequencies, degrees of missing data, and follow-up durations. Thus, optimal methods for defining the recovery phenotype differ between the two. We illustrate three different methods here. Statistical methods are described at a conceptual level here, with additional details in the statistical appendix.

A simple approach to define recovery phenotype is to develop a clinical classification system across multiple outcome variables. For example, in the acute lower respiratory infection cohort we sorted (ordered) subjects by 28 day mortality, then ICU admission, then ICU length of stay, and finally hospital length of stay. This data set did not include robust information on other potential indicators of recovery, such as performance measures or self-reported function; hence, the phenotype of recovery was based on objective available data. Logical cut points that divided subjects into four approximately equal groups with different outcomes were identified; the group with the least resilient recovery phenotypes resulted in a group of subjects who died or had any ICU stay, whereas the group with the most resilient recovery phenotypes included those with 1–2 day hospital stays and no mortality.

A more quantitative approach can be applied using the same data with a principal components analysis. This method describes the variability of observed, correlated variables (eg, length of stay, mortality, and ICU admission) and searches for joint variation that may be due to unobserved “latent variables” (the recovery phenotype). Observed variables are modeled as a linear combination of potential factors plus error terms. Using the interdependencies between observed variables allows us to reduce the set of variables into a single numeric score describing the degree of recovery. We performed a principal components analysis on the following variables: hospital length of stay, ICU length of stay, and discharge to a higher level of care. Mortality was excluded from this metric because of its competing risk with length of stay outcomes (ie, early death results in short length of stay). This analysis generated a weighted multivariable “outcome score” for each individual as a metric of his or her recovery phenotype. This score is a continuous variable, rather than a class variable, which provides the advantage of greater power for association studies. Scores can also be divided into quartiles or quintiles as we have done here to allow comparison of subjects identified as highly resilient by the different approaches.

The rich, longitudinal data and larger sample size in the BHS lends itself to latent class profile analysis (19,20), allowing examination of more complex patterns of recovery across multiple domains. Also referred to as latent growth mixture modeling or latent class trajectory analysis, this method allows us to incorporate the slope (recovery rate) and intercept (either the baseline or final recovery level, adjusted for the other) of multiple measures simultaneously (eg, gait speed, balance, PADLs, and activity level). Groups of individuals (classes) who have similar patterns of recovery are identified. The trajectories of each measure need not be in the same direction; for example, there may be a group of patients with rapid improvement in gait speed who exhibit poor recovery or decline in balance and PADLs. After classes are defined statistically, patterns are described clinically (eg, “rapid ambulatory recovery without functional or activity recovery”). Predictors of class membership such as demographics, comorbidities, environmental factors, and stressor characteristics can be identified using regression modeling. Results from these analyses are beyond the scope of this paper and will be presented elsewhere.

Methods for Defining Expected Recovery Differential

A similar general approach for quantifying the expected recovery differential was applied to both cohorts but with important process differences due to the type of outcome data available. For the acute respiratory infection cohort, a principal component analysis regression model was built to predict the “outcome score” as described earlier. The principal component analysis included 19 clinical predictor variables, including demographics, major comorbidities, and laboratory findings. On the basis of correlations between these variables, principal component analysis was used to reduce the dimensionality of 19 predictors into a smaller number of components, each comprised “weights” of correlated predictor clinical variables. This method is useful with relatively small data sets to avoid overfitting. For each subject, the predicted outcome was estimated based on the principal component analysis model weights and their individual characteristics. Then the difference between this value and their actual outcome as described earlier was calculated. In this case, positive values indicate that the subject recovered better than predicted, whereas negative values indicate that the subject did worse than predicted.

In the BHS, we have 10 individual outcomes rather than a single outcome score to predict. These include “Activity” variables (step counts by actigraphy, Yale survey of physical activity); “Strength” variables (chair stands, grip strength); “Balance” as measured by the Tinetti balance scale; “Ambulation” variables (gait speed, independence category); and “Disability” variables (Lower Extremity Gain Scale, ADL, PADL). A predictive model was developed for each outcome separately, using a mixed model with random and fixed effects. Predictive variables included age, body mass index, education, gender, race, depression, self-reported health, number of co-morbidities, fracture site, delirium, dementia, and discharge destination. From this model, the 12 month predicted value was derived for each outcome. The expected recovery differential value for each outcome was calculated as the difference between actual and predicted 12 month values.

Agreement of Expected Recovery Differentials Across Variables

Our next objective was to explore whether the expected recovery differential approach is measuring an underlying resilience characteristic at the whole-person level that influences recovery across domains. To do this, we compared the expected recovery differentials across all 10 outcomes within each individual. Quartiles of expected recovery differential were compared across all 10 outcomes using weighted kappas. We then plotted each subject’s expected recovery differential result for each pair of outcomes. We evaluated the linearity of the resulting plot (indicating correlation between the expected recovery differential in the two outcome measures). In addition, we derived a score that would reflect overall recovery across all 10 domains. To do this, we calculated a total standardized score (total expected recovery differential) for each person. When we plotted each individual’s recovery along with two separate domains, we were also interested in whether we observed clustering of individuals by total expected recovery differential score. Clustering of individuals with the highest and lowest total expected recovery differential scores would provide evidence of consistency in the degree of an individual’s expected recovery differential (relative to other cohort members), across “total recovery” and the two domains of function represented in a given plot.

Results

Selected results illustrating the type of output obtained and the agreement of the approaches for identifying physical resilience are presented here. A complete description of the resilience phenotypes obtained from both cohorts is beyond the scope of this paper.

Agreement Between Approaches—Example of Viral Respiratory Infection Cohort

The approaches identified different sets of participants as being the most or least physically resilient. To illustrate this, in the viral respiratory infection cohort, we used two different methods to rank individuals into quartiles of recovery phenotype: a simple clinical classification and a principal components analysis. The agreement between these two methods as defined by weighted kappa was 0.58 (95% CI = 0.50–0.66), indicating moderate agreement beyond chance. As anticipated due to its different underlying conceptual approach, the expected recovery differential approach showed less agreement with the two recovery phenotype approaches. The weighted kappa was 0.52 (0.43–0.60) between the expected recovery differential and factor analysis quartiles and 0.37 (0.27–0.47) between the expected recovery differential and the simple clinical classification quartiles. The clinical characteristics of individuals ranked in the highest and lowest physical groups in each approach are compared in Table 2. On average, the expected recovery differential approach defines an apparently sicker group with more comorbidities as being highly resilient compared with the two recovery phenotype methods. Conversely, the expected recovery differential method defines a “healthier” group as having low levels of resilience.

Initial Evidence of an Underlying Resilience Characteristic Across Multiple Domains—Example of BHS

Overall, 12.4% participants scored in the highest quartile of expected resilience differentials for 50% or more of the 10 outcomes; in other words, they recovered substantially better than expected in more than half of their outcome measures. Conversely, 13.4% of the cohort scored in the lowest quartile for half or more of the outcomes, suggesting lower than expected recovery in multiple domains. Examining weighted kappas for agreement in expected recovery differential quartiles for all possible pairs of outcome measures, we observed good agreement (proportion of expected agreement beyond chance ≥ 0.4) among 20 of the 90 possible combinations (range: 0.4–0.55), moderate agreement beyond chance (0.2–0.39) in 33 combinations, and minimal agreement beyond chance in 37 (<0.2). In general, pairs of outcomes defined a priori as measuring a similar domain (Figure 3) had greater agreement than those measuring different domains.

Figure 4 displays exemplar plots of each subject’s expected resilience differential for two outcomes, with their total expected recovery differential score averaged across all 10 outcomes reflected by dot color. Individuals in the highest total expected recovery differential quartile are red, middle quartiles are blue, and the lowest are green. If there is a common underlying resilience characteristic that determines outcome in both variables, then highly resilient individuals (red dots) should group together at the right side of the figure, whereas individuals with low total expected recovery differentials (green dots) should group together at the left. In Figure 4A, subjects’ expected recovery differential score in instrumental activities of daily living and lower extremity physical ADLs appear to be correlated, with most individuals with low total recovery differential rated as low on both individual measures, and those with high differentials are high on both. In Figure 4B, subjects’ expected recovery differential score in grip strength and lower extremity physical ADL score are not correlated, but again, there is grouping of total expected recovery differential, suggesting an underlying resilience characteristic associated with recovery in different domains. In Figure 4C, subjects’ expected recovery differential score in self-reported activity hours and actigraphy data are correlated, but there is no apparent grouping by total expected recovery differential, perhaps because there was a randomized physical activity intervention for two of the cohorts, which may have influenced these two functional outcomes more than the others.

Discussion

In the emerging field of physical resilience, there are currently no standard approaches to describing resilience phenotypes or quantifying levels of physical resilience. The two approaches we have described were developed de novo using existing data sets to address different scientific questions and hypotheses. In the acute lower respiratory infection cohort, we found moderate agreement between two different methods to define a recovery phenotype and less agreement between methods that define the recovery phenotype and the expected recovery differential. In the BHS, we found evidence that the expected recovery differential is correlated across multiple outcome measures within individuals, suggesting that it is measuring an underlying construct (latent variable) associated with recovery across functional domains. The terminology in this field is still evolving; however, we believe that it is important to distinguish conceptually between approaches that (i) represent descriptive phenotypes of actual recovery patterns and (ii) consider whether or not the recovery is different than expected. Certainly, other statistical methods can be used within these two categories. Next, we consider the potential applications of each general approach. These two approaches likely both have utility for different clinical scenarios and scientific questions.

The recovery phenotype approach is currently being used to refine descriptions of outcomes following stressors in older adults, integrating multiple outcomes simultaneously and describing more complex patterns of recovery of function in different domains. By building clinical prediction tools to predict recovery across multiple domains or outcomes, we can enhance communication of prognosis and better target patients for enrollment in clinical programs and health service interventions. Clinical tests that further improve prediction of recovery can then be identified; examples of clinical tests under investigation include dual-tasking and other “stress” tests before or immediately after the stressor. Further insights into mediators of recovery, and potential interventions to address them, may be gained by identifying phenotypes in which the recovery trajectory differs across different outcome domains. For example, a phenotype pattern in which lower extremity function recovers rapidly after hip fracture whereas activity and function lag might prompt an investigation into affective or environmental mediators and suggest the need for occupational therapy or activity promotion interventions.

On the other hand, there are several limitations to the recovery phenotype approach. Latent class profile analyses require larger sample sizes and are more sensitive to missing data, especially because they may not have resulted from a missing at random or missing completely at random process. It is difficult to incorporate death—the ultimate “low resilience” outcome—because it either leads to missing functional measures or directly competes with them (eg, length of stay). Finally, the latent class approach results in more complex descriptions of recovery, which may be hard to describe and apply clinically.

We hypothesize that the expected recovery differential approach may be useful as a way to uncover potential biological mechanisms underpinning physical resilience that are not simply related to age or other demographic and comorbid factors. If there are conserved pathways that are common across individuals with a similarly expected recovery differential independent of comorbid conditions, then interventions that affect these pathways might bolster resilience across multiple stressors and outcomes. In pathway discovery research, a typical first step is to compare biomarker levels among individuals with the highest and lowest levels of the phenotype of interest. If the recovery phenotype is used, differences in biomarkers are likely to be driven by the known drivers of recovery such as age, gender, race, and common comorbidities such as diabetes. By comparing individuals who recover much better or worse than expected after accounting for the impact of these characteristics, we may be more likely to identify conserved, comorbidity-independent biomarkers of resilience. This is especially useful for studies with relatively small sample sizes as there are limited degrees of freedom available to develop fully adjusted models for significant biomarker associations after the fact. Another advantage of the expected recovery differential approach is that it is less sensitive to missing data, although the final outcome measure must be present (or imputed based on prior measures) for subjects to be included.

However, the expected recovery differential approach confers several drawbacks. First, this approach may require investigators to choose a single time point for comparing expected versus actual recovery, thereby losing some of the information contained in a recovery trajectory such as the rate and timing of recovery. In this analysis, 12 months were chosen as the latest available time point that was both clinically relevant and made maximal use of the available data. In larger data sets with more time points measured, the expected recovery slope could be compared with the actual recovery slope. Second, the choice of covariates becomes particularly important. If the goal is to identify biological pathways, we do not want to include characteristics that may mediate those biological effects. Ideally, this choice would be governed by a conceptual framework, and assuring model fit is important. Further, if covariates are used in the model predicting expected outcome, we cannot measure their association with the resulting expected recovery differential. Therefore, this approach is not as useful for quantifying the impact of environmental, psychological, or other clinical characteristics associated with physical resilience. Third, as stated previously, this approach will tend to exclude younger, healthier subjects whose recovery is expected, and therefore may be particularly susceptible to ceiling and floor effects. Therefore, it may be most appropriate for persons in a middle range of expected recovery; until these issues are clarified, using both approaches in biomarker association studies may be wise. Finally, the expected recovery differential is not easy to describe clinically.

Both approaches require repeated measures of clinically important outcomes over time to optimally quantify resilience; otherwise, the resulting phenotypes are dominated by binary outcomes, such as death or ICU admission, with less ability to distinguish between degrees of functional recovery. Limitations notable in our analyses include the relatively small sample sizes and limited outcome measures available in the acute respiratory infection cohort. However, our use of real-world data mirrors many experimental situations and highlights the relative strengths and drawbacks of each approach. Missing data is a common challenge in resilience research, and the factor analysis method used here requires an assumption of missing-at-random, which is likely untrue for frail populations. We mitigated this by imputing logical values when the reason for missingness was known; for example, when a physical performance test was coded as “participant unable to do,” the lowest possible score was imputed. However, additional work in managing missing data is needed. Other research needed in this area includes validation of these approaches in other stressor/recovery paradigms, as well as refining methods to calculate them under different conditions of data type and availability.

Conclusions

We describe two approaches to categorize and quantify physical resilience in older adults: a recovery phenotype and an expected recovery differential approach. Importantly, these methods identify overlapping, but distinct, populations. Recovery phenotypes are likely to be most useful in clinical applications such as clinical prediction models or outcome classification for intervention research. On the other hand, we propose that expected recovery differentials may be better suited to identify mechanisms underlying physical resilience and serve as targets for interventions designed to improve physical resilience. Exploratory analyses with the expected recovery differential in which hip fracture was the stressor provide preliminary evidence that physical resilience, or the tendency to retain or regain function after a health stressor, may reflect an underlying characteristic at the whole-person level that influences recovery across multiple functional domains.

Funding

This work was supported by National Institute on Aging (NIA) (AG056925-01, AG049077-01A1, 2P30AG028716-11, and P30-AG02133). Dr Whitson’s contributions were supported by the Duke Claude D. Pepper Older American Independence Center (P30AG028716), the Physical Resilience Indicators and Mechanisms in the Elderly (PRIME) Collaborative (UH2AG056925), and the National Center for Advancing Translational Sciences of the National Institutes of Health (UL1TR002553).

Acknowledgements

C.C.E. and H.W. designed and performed these studies and drafted the manuscript. C.F.P. and R.S. performed statistical modeling and analysis. K.E.S. and K.M.H. assisted with design, results interpretation, and critical revisions of the manuscript. A.B. and M.M. assisted with preparation, analysis, and results reporting for the viral respiratory infection cohort. J.M. and D.O. assisted with preparation, analysis, and results reporting for the Baltimore Hip Studies. D.M.C. assisted with drafting, editing, and finalizing the manuscript.

Conflict of Interest

Colón-Emeric—During the past year, I consulted or served on advisory boards for: Novartis, Amgen, and Biscardia. There are no relevant conflicts of interest with this work.

Schmader—During the past year, I received research grant funding from GSK. There are no relevant conflicts of interest with this work.

Magaziner—During the past year, I have consulted or served on advisory boards for the following entities: American Orthopaedic Association; Ammonett; Novartis; Pluristem; Viking

The following investigators have no conflicts and nothing to report: Whitson, Pieper, Crabtree, Huffman, Orwig, Sloane, Bloom, McClain.

References