Performance-based measures of physical function as mortality predictors: Incremental value beyond self-reports.

BACKGROUND: Although previous studies have indicated that performance assessments strongly predict future survival, few have evaluated the incremental value in the presence of controls for self-reported activity and mobility limitations.
OBJECTIVE: We assess and compare the added value of four tests - walking speed, chair stands, grip strength, and peak expiratory flow (PEF) - for predicting all-cause mortality.
METHODS: Using population-based samples of older adults in Costa Rica (n = 2290, aged 60+) and Taiwan (n = 1219, aged 53+), we estimate proportional hazards models of mortality for an approximate five-year period. Receiver Operator Characteristic (ROC) curves are used to assess the prognostic value of each performance assessment.
RESULTS: Self-reported measures of physical limitations contribute substantial gains in mortality prediction, whereas performance-based assessments yield modest incremental gains. PEF provides the greatest added value, followed by grip strength. Our results suggest that including more than two performance assessments may provide little improvement in mortality prediction.
CONCLUSIONS: PEF and grip strength are often simpler to administer in home interview settings, impose less of a burden on some respondents, and, in the presence of self-reported limitations, appear to be better predictors of mortality than do walking speed or chair stands. COMMENTS: Being unable to perform the test is often a strong predictor of mortality, but these indicators are not well-defined. Exclusion rates vary by the specific task and are likely to depend on the underlying demographic, health, social and cultural characteristics of the sample.

1. Introduction

The process linking chronic and acute health conditions to disability has been conceptualized as a progression through four stages: (1) pathology; (2) substantial impairments; (3) functional limitations (restrictions in basic physical or mental performance); and (4) disability (difficulties in various activities of daily life, Verbrugge and Jette 1994). Measures associated with these stages have provided researchers and health professionals with valuable indicators of the current and future health status of community-dwelling populations of older adults. Population-based surveys frequently include measures of the disablement process, either from self-reports or as interviewed-administered performance tests. Although these two types of measures – often labeled “subjective” and “objective” respectively – are statistically correlated, they are thought to capture distinct constructs (Reuben et al. 2004). Nevertheless, both self-reports and performance tasks have been shown to predict subsequent deterioration in health in diverse settings (Bernard et al. 1997; Cooper et al. 2010; Ferrucci et al. 1991; Gill, Robison, and Tinetti 1998; Guralnik et al. 1995; Guralnik et al. 2000; Reuben, Siu, and Kimpau 1992; Tinetti et al. 1995).

Each method has advantages and drawbacks. The most frequently used self-reported measures identify: (1) limitations in performing basic activities of daily living (ADL) such as bathing or eating, and (2) mobility limitations such as walking or raising one’s arms. These indicators are easy to collect, inexpensive, and focus on behaviors that are clinically relevant and signal the need for caregiving (Reuben et al. 2004). Although the subjective component of self-reports may capture information of prognostic value, self-reports are likely to be biased by myriad factors (e.g., environmental conditions, cultural preferences, or attitudes) that impede comparisons across populations or subgroups (Melzer et al. 2004) and, given their focus on functional limitations and especially disability, they are unlikely to identify individuals in the early stages of impairment (Reuben et al. 2004). Still, they may be better suited than performance tests to assess individuals’ interactions with their immediate environments and the associated constraints.

Performance assessments, such as grip strength, peak expiratory flow (PEF), chair stands, and walking speed, are largely determined by physiological functions that typically decline with age and may underlie frailty (Cooper et al. 2010). Specifically, walking speed and chair stands reflect lower limb strength and mobility, grip strength is a marker of overall muscle strength, and peak expiratory flow is an indicator of lung capacity and airway obstruction. These measurements are considered to have greater face validity and perhaps reliability than self-reports, and are likely to be more sensitive to changes over time, more comparable across different contexts, and better suited to capturing variation across a wide continuum of physical function including early stages of impairment (Guralnik et al. 1989; Guralnik et al. 1994; Myers et al. 1993); some suggest that they provide “preclinical” detection of functional decline (Reuben et al. 2004). The predictive power of performance assessments for future health outcomes likely results from bi-directional pathways. For example, not only do poor performance measures appear to be causally related to increased risks of injury and the onset of disability, but a history of chronic disease and malnutrition is often the source of these physical impairments (Fried and Guralnik 1997). Despite their utility in health surveys, performance assessments come at a cost, not only a monetary cost, rather they also place a burden on the respondent and interviewer. The tests are time-consuming, require substantial effort for some older or weak respondents, need special equipment and space for administering, and may compromise response rates for the overall survey. Thus, the decision about whether to include them in home-based interviews is likely to depend on their added value beyond information captured by self-reports.

Although there are many potential uses of performance assessments in both population-based research and clinical practice for example, the identification of individuals at risk of developing disabilities or requiring institution-based care here we evaluate these measures only in terms of their prognostic value for survival. We use survival as our criterion because mortality is the health measure most frequently used to characterize the overall wellbeing of a population. Moreover, mortality is a well-defined and salient health outcome with little measurement error. Virtually all of the best quality prognostic indices for predicting survival (e.g., Lee et al. 2006; Schonberg et al. 2009) include physical function as a key indicator. Indeed, Swindell et al. (2010) found that five of the ten strongest predictors of mortality (among 377 predictors tested) were measures of physical function, including both self-reported and performance-based indicators.

Our primary objective in this paper is to ascertain the incremental improvement derived from performance tests beyond that provided by self-reported limitations in predicting mortality. Our secondary objectives are to determine whether some performance assessments are stronger predictors than others and whether their predictive power persists with controls for self-reported disease status. As biodemographic surveys increasingly incorporate performance tests into household visits, it is important to evaluate the net contribution of these assessments for predicting survival. To assess the robustness of our findings, we use data from two population-based surveys of older adults that collected similar information in Taiwan and in Costa Rica. Although we have no reason to expect the relationship to differ between these countries, replication of the results in different settings may strengthen our inferences.

2. Background

There has been little research addressing the central questions in this analysis. Few studies predicting all-cause mortality in a general population of older adults include both self-reported and performance-based assessments of physical function. Almost all existing studies are based in high income countries primarily the US (Al Snih et al. 2002; Cook et al. 1991; Guralnik et al. 1994; Hardy et al. 2007; Markides et al. 2001; Melzer, Lan, and Guralnik 2003; Reuben et al. 2004), but also Italy (Cesari et al. 2008) and Australia (Simons et al. 2011). Two exceptions are a study in China (Feng et al. 2010) and another in Costa Rica (Rosero-Bixby and Dow 2012).

Beyond this geographic restriction, the studies described above that include both self-reported and performance-based measures of physical function as predictors of total mortality have several drawbacks. First, although most include self-reported ADL limitations, fewer include self-reported mobility limitations, which afflict adults across a broader age range. Second, researchers often combine several different performance assessments into a summary measure, but they do not always evaluate the individual tests. For example, several of these studies use the Short Physical Performance Battery (SPPB), which comprises a timed walk, repeated chair stands, and balance stands (Cesari et al. 2008; Guralnik et al. 1994; Hardy et al. 2007; Markides et al. 2001; Reuben et al. 2004), and one uses the MOBLI (Index of Mobility-Related Limitations), which replaces balance stands with PEF (Melzer, Lan, and Guralnik 2003). Only two of these studies evaluate all of the underlying individual tests as predictors of mortality (Cesari et al. 2008; Markides et al. 2001).

Despite variation in the measures collected and in the lengths of follow-up, most studies find that performance tests predict mortality net of self-reported limitations. However, perhaps because of variations in study design, researchers disagree about which tests best predict mortality. For example, Cesari et al. (2008) report chair stands to be the strongest component of the SPPB in predicting mortality, whereas several earlier studies identify walking speed (Al Snih et al. 2002; Markides et al. 2001; Ostir et al. 2007). Rosero-Bixby and Dow (2012) find grip strength to be the strongest predictor of mortality in women, whereas PEF is the best predictor in men; walking speed is only marginally significant net of the other covariates. Most studies that have examined grip strength or PEF in combination with self-reports include few, if any, additional performance assessments, making it difficult to evaluate these two measures vis-à-vis the others.

Our study extends the existing literature in two important ways. First, we examine whether previous findings, based mostly on the US, pertain to two middle income countries: Taiwan and Costa Rica. Second, we consider a set of four performance tests to assess whether chair stands or timed walks as some earlier work suggests are the strongest predictors of mortality, or whether grip strength or PEF, which are often not part of the most frequently used batteries of tests, perform as well or better. We hypothesize that grip strength and PEF may perform better because they are likely to be less correlated with the lower extremity functions captured by the self-reports.

3. Materials and methods

3.1 Data

The data come from the first wave of the Costa Rican Study on Longevity and Healthy Aging (CRELES) and the second wave of the Social Environment and Biomarkers of Aging Study (SEBAS). In both surveys, an interview conducted in the respondent’s home included questions related to health and a series of performance-based assessments. SEBAS and CRELES included informed consent from all participants and received ethical approval from the human subjects committees at the institutions conducting the studies.

CRELES is a longitudinal study based on a national sample of residents of Costa Rica aged 60 and older in 2005, with oversampling of the oldest old (Rosero-Bixby, Fernández, and Dow 2013). The sample was selected randomly from the 2000 census database using a multi-stage sampling design. For this analysis, we use data from the first wave conducted between November 2004 and September 2006. Interviews were completed by 2827 respondents (85 per cent of survivors located).

The SEBAS cohort represents a random subsample of participants in the nationally representative Taiwan Longitudinal Study of Aging (TLSA); elderly persons and urban residents were oversampled. The TLSA began in 1989 and younger refresher cohorts were added in 1996 and 2003; all three cohorts were selected randomly using a multistage sampling design. The sampling frame for the 2006 SEBAS included: a) an older cohort (aged 60+) of respondents from the 1999 wave of TLSA who completed the 2000 SEBAS medical examination, and b) a younger cohort (aged 53–60) of respondents first interviewed in the 2003 wave of TLSA. Interviews were completed by 1284 respondents aged 53 and older in 2006 (87% response rate). Additional details are provided elsewhere (Chang et al. 2012).

3.2 Measures

With the exception of mortality, all measures come from the first wave of CRELES (2004–2006) and the second wave of SEBAS (2006).

3.2.1 Mortality

For SEBAS, survival status as of June 30, 2011 was ascertained through linkage with the death certificate file maintained by the Department of Health and the household registration database maintained by the Ministry of the Interior. Survival status in CRELES was established in two ways: (1) through the computer records in the National Death Registry up to December 31, 2010, and (2) during the second (2006–2008) and third (2008–2010) waves of home visits. The computer follow-up used the unique identification number (the cédula) that all Costa Ricans have. Five out of the 566 deaths found in the field were not found in the Registry, suggesting a death under-registration rate of 1 per cent. In contrast, 10 per cent of the deaths from the Registry were not found in the field, appearing in the second and third waves as loss of follow-up. For the foreigners in the sample (approx. 3 per cent), survival was established only in the field because they did not have a unique identification number with which to link them to the Registry.

3.2.2 Measures of physical function

We include two self-reported measures of physical function. Limitations with activities of daily living (ADL) are based on five ADLs (bathing; eating; toileting; moving around the house; getting out of bed). We count the number of ADLs (0–5) that the respondent reported difficulty performing. The measure of mobility limitations counts the number of mobility tasks (0–4) that the respondent reported difficulty performing. Three of these tasks were similar in both countries (i.e., walking; climbing stairs; raising his/her arms). In Costa Rica, respondents were asked to demonstrate whether they could lift their arms above their shoulders; those who did not attempt the activity are coded as missing. In Taiwan, difficulty raising both hands over the head was based on self-reports. The fourth self-reported function differed between Costa Rica (i.e., pushing or pulling a large object such as a recliner chair) and Taiwan (i.e., lifting or carrying 11–12 kg). In Costa Rica, respondents who reported that they “do not do that activity” were coded as missing; this response category was not an option in Taiwan.

Four health assessments were administered by trained interviewers: grip strength, PEF, timed walk, and chair stands. In each case, the test was coded as missing if the respondent refused (or in the case of SEBAS, was unable to understand the instructions or there was equipment failure).

Grip strength (in kg) was measured using a dynamometer (CRELES: Creative Health Products dynamometer [model T-18]; SEBAS: North Coast™ hydraulic hand dynamometer [NC70142]); we used the highest level from two trials on the dominant hand in CRELES and three trials on each hand in SEBAS. The respondent was coded as “unable” to perform the test if s/he: a) met the exclusion criteria for both hands; b) tried but was unable to do the test; c) did not attempt the test for safety reasons; d) (in CRELES only) was not tested because of disability; e) (in SEBAS only) did not attempt the test because of weakness, stroke, or frailty or stopped because of pain or discomfort. The exclusion criteria included: surgery on hand/wrist/arm in the past three months and (in SEBAS only) recent injury, worsening pain, swelling, inflammation, or severe pain in the hand/wrist.

Lung function, represented by PEF (L/min), was based on the maximum of three trials using a peak flow meter (CRELES: Mini Wright; SEBAS: TruZone®). In CRELES, the “unable” group included respondents who were unable to complete even one trial. In SEBAS, the respondent was coded as “unable” if: a) s/he met the exclusion criteria (i.e., surgery on chest/abdomen or hospitalized for a heart problem in the past six weeks; detached retina or eye surgery in past six months; hospitalized for respiratory or lung infection in past three weeks), b) s/he did not attempt the test because of stroke or illness, or c) s/he or the interviewer felt it was unsafe.

To measure walking speed (m/sec), the respondents were asked to walk three meters at their normal speed. The respondent started from a sitting position in CRELES and a standing position in SEBAS. CRELES included one trial; SEBAS conducted two trials (we used the faster of the two). The respondent was coded as “unable” to complete the test if s/he: a) tried but was unable to do the test; b) did not attempt the test for safety reasons; or c) (in CRELES only) reported any problem that would impair him/her from doing a mobility and flexibility test.

For the chair stand test, the respondent was asked to stand up and sit down again five times in a row as quickly as possible without stopping while keeping his/her arms folded across his/her chest. For those able to complete five stands, the completion time was recorded. To adjust for differences in chair height, we regressed the completion time (c) for individual i on chair height (h) controlling for the respondent’s age and height (SEBAS)/knee height (CRELES), with models fit separately by sex (Cornman et al. 2011). The adjusted completion time was calculated as , where β was the coefficient for h from the sex-specific model and h̄ was the mean chair height among the pooled sample (44.5 cm). Chair stand speed was computed as five divided by the adjusted time c̃ (Cornman et al. 2011). The respondent was coded as “unable” to perform the test if s/he: a) met the exclusion criteria (CRELES: reported any problem that would impair him/her from doing a mobility and flexibility test; SEBAS: was in a wheelchair); b) tried but was unable to complete five stands, or c) did not attempt the test for safety reasons.

3.2.3 Control variables

All models include controls for age, sex, education, and urban residence. In auxiliary analyses, we also control for several other variables that may be associated with both physical function and mortality: self-reported measures of specific health conditions (i.e., cancer, heart disease, diabetes, stroke, respiratory disease, arthritis, hypertension); smoking status (never, former, current); exercise (3+ times per week); and hospitalization in the past year.

3.3 Analytical strategy

Among those interviewed, 19.0 per cent of the CRELES sample and 5.1 per cent of the SEBAS sample are missing data for a measure of physical function, either self-reported (n=410 in CRELES; n=6 in SEBAS) or performance-based (n=164 in CRELES; n=59 in SEBAS). Exclusion of these respondents leaves an analysis sample of 2290 for Costa Rica and 1219 for Taiwan. In order to ascertain the robustness of the results to the treatment of missing data, we use multiple imputation to re-estimate the final models for the full samples.

Descriptive statistics are weighted to account for oversampling and for differential response rates by age, sex, and other covariates. Survival models are fit separately by country using unweighted data. We estimate age-specific mortality using a Gompertz proportional hazards model with time measured in terms of age. The Gompertz function assumes that the force of mortality increases exponentially with age and generally fits very well at older ages (Horiuchi and Coale 1982). In initial tests (not shown), the age slope of mortality (γ) did not differ significantly by sex in these samples.

Some of the performance tests are strongly correlated with one another (e.g., among those with measurements, the Pearson correlations between grip strength and PEF are 0.60 in CRELES and 0.65 in SEBAS, and correlations between walking speed and chair stand speed are 0.37 in CRELES and 0.51 in SEBAS). Thus, we model the performance assessments individually as well as jointly. Our models are fit in two stages. First, we estimate a model that includes only self-reported ADL and mobility limitations in addition to sociodemographic controls. Next, we fit a series of models that add a performance test, with a categorical specification: those unable to perform a given test are assigned a separate category, and the remaining responses are recoded into quartiles based on the weighted distribution of the pooled samples. These models are not designed to reflect causal processes but rather to evaluate the prognostic value of performance assessments over and above that of self-reported measures of physical function.

To evaluate the predictive ability of different measures of physical function, we compare the model-based predicted probability of dying by the end of follow-up (see Supplementary Material for details) with the observed binary outcome (death vs. survival) to estimate the area under the receiver-operating-characteristic (ROC) curve (AUC). The AUC is the most popular measure of discrimination and can be interpreted as the probability that those who actually died are assigned a higher predicted probability of death than are their counterparts who survived (Pencina and D’Agostino 2004). We use the “rocgold” procedure in Stata 12.1 to test whether the addition of performance assessments yields a significant improvement in mortality prediction. Similarly, we compare the AUC from models that include individual performance tests to determine whether some assessments are better predictors than others. We present both the absolute increase in AUC between pairs of nested models as well as the increase in AUC for a given model as a percentage of the area above the curve (i.e., (1-AUC) or the fraction of incorrect predictions) for the simpler model.

Finally, to help gauge the magnitude of the association with mortality, we calculate the predicted probability of dying between exact ages 70 and 75 (5q̂70 ) for selected levels of performance. This age interval corresponds roughly to the mean age at the time of the survey (69.3) and at the end of follow-up (74.3) for the pooled sample (weighted). These probabilities are estimated by setting the selected performance assessment at the specified value, fixing all other covariates at the weighted mean for the pooled sample, and using the model coefficients to predict the probability of dying between exact ages 70 and 75 (see Supplementary Material for details). For each performance test, we provide two sets of predicted probabilities. The first is based on a model that includes only the selected performance assessment and sociodemographic controls, thus representing the “gross effect” without adjusting for any other measures of physical function. The second is based on a model that includes all physical function measures (both self-reported and performance-based) in addition to control variables and thus, represents the “net effect” after adjusting for the effects of all other measures of physical function.

4. Results

In CRELES, there were 569 deaths by January 1, 2011; the average length of mortality follow-up was 5.2 years (max., 6.2 years). For SEBAS, there were 140 deaths by June 30, 2011; mean follow-up was 4.7 years (max., 4.9 years). The Costa Rican sample is older than the Taiwanese sample (Table 1), reflecting differences in the sampled population (60+ for CRELES, 53+ for SEBAS). A small proportion of respondents report any of the five ADL limitations (11 per cent in Costa Rica, 7 per cent in Taiwan), suggesting low rates of severe disability in these populations. Functional limitations are more frequent: a sizeable proportion report at least one of the four mobility limitations (55 per cent in Costa Rica; 33 per cent in Taiwan; not shown). The performance-based tests capture more subtle variations in physical capability. For example, even among those who report no ADL or mobility limitations, there is wide variation in performance (e.g., grip strength ranges from 5 to 58 kg in Costa Rica; 4 to 68 kg in Taiwan; not shown). Thus, although the performance tests are moderately correlated with the self-reported measures (r = −0.13 to −0.52), they are more likely to capture early stages of impairment before a limitation would be acknowledged with standard questionnaires.

Table 1

Descriptive statistics for all analysis variables, by country, weighted analyses

	Costa Rica [CRELES] (n=2290)	Taiwan [SEBAS] (n=1219)
Control variables
Age, mean (SD)a	70.5 (7.9)	65.9 (9.2)
Female, %	50.8	46.1
Urban resident, %	62.5	47.4
Years of completed education (0–17), mean (SD)	5.2 (4.2)	6.6 (4.7)
Self-reported measures of physical functiona
Number of ADL limitations (0–5), mean (SD)	0.3 (0.9)	0.2 (0.9)
Number of mobility limitations (0–4), mean (SD)	1.2 (1.3)	0.7 (1.1)
Performance-based measures of physical functiona
Unable to perform grip strength test, %	2.8	3.4
Grip strength (kg), mean (SD)b,c	27.3 (9.1)	28.2 (10.6)
Unable to perform PEF test, %	8.9	3.5
PEF (L/min), mean (SD)b,d	314.6 (121.2)	334.5 (135.7)
Unable to perform timed walk, %	8.5	4.2
Walking speed (m/sec), mean (SD)b,e	0.6 (0.2)	0.9 (0.3)
Unable to perform chair stands test, %	11.7	8.4
Chair stand speed (stand/sec), mean (SD)b	0.4 (0.1)	0.5 (0.2)
Died by the end of followupf, %	16.5	10.5

ADL, Activities of Daily Living; PEF, peak expiratory flow

Measured at the first wave of CRELES (2004–2006) and the second wave of SEBAS (2006).

Among those able to complete the test.

Maximum from trials on both hands except for n=8 in CRELES and n=46 in SEBAS who did not complete trials on both hands.

Maximum from three trials except for n=7 in CRELES and n=13 in SEBAS who completed only one or two trials.

In SEBAS, n=12 walked only 2–2.5m because of space limitations in the respondent’s home; walking speed was calculated based on the distance actually walked.

Follow-up ended on January 1, 2011 for CRELES and June 30, 2011 for SEBAS. In CRELES, a few foreigners (n=49) were censored early, at the date of last contact. The unweighted number of deaths was 579 in CRELES and 140 in SEBAS.

In the hazard models that include only self-reported measures (Model 1, Tables 2 and 3), mobility limitations are associated with higher mortality rates in both Costa Rica (HR=1.29, p<0.001) and Taiwan (HR=1.50, p<0.001), but ADL limitations are significant only in Costa Rica (HR=1.22, p<0.001). In models that add a single performance test (Models 2a–2d), joint Wald chi-square tests (based on the four parameters for a given assessment) indicate that each of the four performance tests is significantly associated with mortality in Costa Rica, but only grip strength and PEF are significantly associated with mortality in Taiwan. The fact that the sample size in Costa Rica is more than twice that in Taiwan accounts for the significance of some estimates for Costa Rica even when the magnitudes of the hazard ratios are similar to or smaller than the corresponding values in Taiwan.

Table 2

Hazard ratios from Gompertz model of age-specific mortalitya, Costa Rica (n=2290)

Model numberb	(1)	(2a)	(2b)	(2c)	(2d)	(5)
Self-reported measures
Number of ADL limitations	1.22***	1.20***	1.19***	1.19***	1.18***	1.16***
Number of mobility limitations	1.29***	1.26***	1.25***	1.23***	1.23***	1.18***
Performance-based tests
Grip strength (kg)
Unable to perform test		2.22***				1.51
Bottom quartile (2–20)		2.05***				1.52*
2nd quartile (20.1–26)		1.29				1.03
3rd quartile (26.5–34)		1.44				1.30
Top quartile (34.5–68)		(ref)				(ref)
Joint testc (df = 4)		p<0.001				p~0.016
PEF (L/min)
Unable to perform test			3.09***			2.64***
Bottom quartile (50–230)			2.49***			2.17***
2nd quartile (235–300)			1.69*			1.59*
3rd quartile (310–390)			1.72*			1.68*
Top quartile (400–800)			(ref)			(ref)
Joint testc (df = 4)			p<0.001			p<0.001
Walking speed (m/sec)
Unable to perform test				2.51**		1.55
Bottom quartile (0.06–0.52)				2.16*		1.73
2nd quartile (0.52–0.66)				1.75		1.56
3rd quartile (0.66–0.83)				1.40		1.34
Top quartile (0.83–2.14)				(ref)		(ref)
Joint testc (df = 4)				p~0.002		p~0.095
Chair stand speed (stands/sec)
Unable to perform test					2.02**	1.55
Bottom quartile (0.09–0.34)					1.60	1.25
2nd quartile (0.34–0.42)					1.21	1.06
3rd quartile (0.42–0.52)					1.30	1.24
Top quartile (0.52–1.98)					(ref)	(ref)
Joint testc (df = 4)					p~0.007	p~0.267

p<0.001,

p<0.01,

p<0.05

All models control for sex, education, and urban residence.

Model numbers correspond to those presented in Table 4.

Joint Wald chi-square test for the four parameters pertaining to specified assessment.

Table 3

Hazard ratios from Gompertz model of age-specific mortalitya, Taiwan (n=1219)

Model numberb	(1)	(2a)	(2b)	(2c)	(2d)	(5)
Self-reported measures
Number of ADL limitations	1.11	1.07	1.04	1.02	1.08	1.04
Number of mobility limitations	1.50***	1.37***	1.39***	1.46***	1.40***	1.31**
Performance-based tests
Grip strength (kg)
Unable to perform test		3.15*				1.00
Bottom quartile (2–20)		3.36**				2.50*
2nd quartile (20.1–26)		2.13*				1.61
3rd quartile (26.5–34)		1.42				1.24
Top quartile (34.5–68)		(ref)				(ref)
Joint testc (df = 4)		p~0.011				p~0.049
PEF (L/min)
Unable to perform test			5.17***			5.15**
Bottom quartile (50–230)			2.93**			2.22*
2nd quartile (235–300)			2.99**			2.32*
3rd quartile (310–390)			2.10*			1.94
Top quartile (400–800)			(ref)			(ref)
Joint testc (df = 4)			p~0.004			p~0.029
Walking speed (m/sec)
Unable to perform test				2.17		1.35
Bottom quartile (0.06–0.52)				1.48		0.97
2nd quartile (0.52–0.66)				0.79		0.56
3rd quartile (0.66–0.83)				1.51		1.18
Top quartile (0.83–2.14)				(ref)		(ref)
Joint testc (df = 4)				p~0.079		p~0.187
Chair stand speed (stands/sec)
Unable to perform test					2.14*	1.29
Bottom quartile (0.09–0.34)					1.58	1.23
2nd quartile (0.34–0.42)					1.42	1.13
3rd quartile (0.42–0.52)					1.75	1.51
Top quartile (0.52–1.98)					(ref)	(ref)
Joint testc (df = 4)					p<0.262	p~0.721

p<0.001,

p<0.01,

p<0.05

All models control for sex, education, and urban residence.

Model numbers correspond to those presented in Table 4.

Joint Wald chi-square test for the four parameters pertaining to specified assessment.

The hazard ratios are substantially attenuated when all performance tests are included in a single model (Model 5, Tables 2 and 3). Grip strength and PEF continue to be significantly related to mortality in both countries, but walking speed and chair stand speed are not. Being unable to perform the PEF test is a particularly strong predictor of mortality (HR=2.6 in Costa Rica; HR=5.2 in Taiwan relative to those in the top quartile). Poor PEF performance (bottom quartile) also strongly predicts mortality in both countries (HRs>2). In addition to the models presented in Tables 2 and 3, we estimated an auxiliary model that includes controls for specific health conditions, smoking, exercise, and hospitalization. Although the coefficients for PEF weaken slightly with the inclusion of these variables, both PEF and grip strength remain strong predictors of mortality (results not shown). We also used multiple imputation to re-estimate these same models for the full samples (n=2827 in Costa Rica, n=1284 in Taiwan). The coefficients from these models are very similar to those shown in Tables 2 and 3.

Comparisons of the AUC for different models show that self-reported measures of physical function yield a substantial and significant improvement in mortality prediction compared with a baseline model that adjusts only for sociodemographic characteristics — the increase in AUC is greater than 10 per cent of the area above the curve (1-AUC) for the baseline model in both countries (Table 4). Beyond the prognostic value of self-reported physical function, PEF provides the greatest improvement in predictive ability in both countries; the other three performance assessments produce smaller improvements that are significant only in Costa Rica.

Table 4

Comparisons of the AUC for various models, by country

		Comparison	Costa Rica			Taiwan
			AUC	Δ in AUCa	Δ as a % of unexplainedb	AUC	Δ in AUCa	Δ as a % of unexplainedb
0	Baseline model (control variables only)		0.765			0.773
1	Add self-reported physical functionc	vs. Model 0	0.801	0.036***	15%	0.800	0.027**	12%
2a	Model 1 + grip strength	vs. Model 1	0.806	0.006*	3%	0.813	0.013	6%
2b	Model 1 + PEF	vs. Model 1	0.808	0.008	4%	0.815	0.015*	7%
2c	Model 1 + walking speed	vs. Model 1	0.806	0.005*	3%	0.809	0.009	5%
2d	Model 1 + chair stand speed	vs. Model 1	0.805	0.005*	2%	0.805	0.005	3%
3a	Model 1 + (PEF, grip strength)	vs. Model 2b	0.812	0.004*	2%	0.825	0.010	5%
3b	Model 1 + (PEF, walking speed)	vs. Model 2b	0.812	0.004	2%	0.823	0.008*	4%
3c	Model 1 + (PEF, chair stand speed)	vs. Model 2b	0.812	0.004*	2%	0.819	0.004	2%
4a	Model 1 + (PEF, grip strength, walking speed)	vs. Model 3a	0.815	0.003	2%	0.832	0.007	4%
4b	Model 1 + (PEF, grip strength, chair stand speed)	vs. Model 3a	0.815	0.002	1%	0.828	0.003	1%
4c	Model 1 + (PEF, walking speed, chair stand speed)	vs. Model 3b	0.813	0.001	1%	0.825	0.002	1%
5	Model 1 + all four performance testsd	vs. Model 4a	0.816	0.001	0%	0.834	0.002	1%

p<0.001,

p<0.01,

p<0.05

The change (Δ) in the AUC and associated significance level is based on a comparison with the model indicated.

The Δ in the AUC as a percent of unexplained is computed as: , where Model X is the comparison model and Model Y is the current model.

This model is the same as Model 1 presented in Tables 2 and 3.

This model is the same as Model 5 presented in Tables 2 and 3.

A ranking of the four assessments in terms of the absolute increment in AUC results in the following order of performance measures, from largest to smallest: PEF, grip strength, walking speed, and chair stands. This ordering is used to evaluate whether successive additions of assessments provide substantial improvements in mortality prediction. Based on the criterion of 0.01 to indicate a meaningful improvement in the AUC (Pencina et al. 2008), PEF appears to be a useful measure (the gain in AUC equaled 0.015 in Taiwan and was just below 0.01 in Costa Rica). Compared with PEF alone (Model 2b), the addition of each of the other three performance assessments (Models 3a–3c) yields a similar increase in AUC in Costa Rica (0.004). In Taiwan, grip strength (Model 3a) and walking speed (Model 3b) lead to the largest increases in AUC, while chair stand speed results in a more modest improvement in AUC.

Table 5 shows the extent of variation in the predicted probability of dying between exact ages 70 and 75 (5q̂70 ) by selected levels of performance on the physical assessments. Differences are particularly large for grip strength and PEF. For example, the gross effect for PEF (i.e., before adjusting for the effects of other measures of physical function) suggests that 28 per cent of Costa Ricans who were unable to perform the test would be expected to die between ages 70 and 75 compared with only 6 per cent of those in the top quartile; corresponding figures for Taiwan are 37 per cent vs. 4 per cent. After adjustment for the contribution of other physical function measures, the net effect for PEF is smaller, but remains substantial (Costa Rica: 17 per cent of unable vs. 7 per cent in the top quartile; Taiwan: 22 per cent vs. 5 per cent, respectively). Some of the differences are large even without considering those who were unable to complete the tests. For example, in both countries, the net effect of PEF on mortality is twice as high for those in the bottom compared with the top quartile.

Table 5

Predicted probability of dying between exact ages 70 and 75 (5q̂70 ) for selected levels of performance on the physical assessmentsa, by country

	Gross effectb		Net effectc
	Costa Rica	Taiwan	Costa Rica	Taiwan
Observed 5q̂70 in the national populationd
Total	0.103	0.123
Men	0.125	0.154
Women	0.082	0.095
Performance-based tests
Grip strength
Unable	0.35	0.29	0.13	0.06
Bottom quartile (2–20 kg)	0.21	0.20	0.13	0.14
Top quartile (34.5–68 kg)	0.08	0.04	0.09	0.06
PEF
Unable	0.28	0.37	0.17	0.22
Bottom quartile (50–230 L/min)	0.18	0.15	0.14	0.10
Top quartile (400–800 L/min)	0.06	0.04	0.07	0.05
Walking speed
Unable	0.32	0.34	0.12	0.13
Bottom quartile (0.06–0.52 m/sec)	0.16	0.18	0.13	0.09
Top quartile (0.83–2.14 m/sec)	0.06	0.07	0.08	0.09
Chair stand speed
Unable	0.29	0.28	0.14	0.09
Bottom quartile (0.09–0.34 stands/sec)	0.15	0.14	0.11	0.09
Top quartile (0.52–1.98 stands/sec)	0.08	0.06	0.09	0.07

All models control for sex, age, education, and urban residence. The predicted probabilities of dying between exact ages 70 and 75 are estimated by setting the selected measure of physical function at the specified value and fixing all other covariates at the weighted mean for the pooled sample.

We fit a separate model for each performance test and adjust only for sociodemographic control variables.

We fit a model that includes all measures of physical function (both self-reported and performance-based) in addition to control variables (same as Model 2, Table 2).

The observed 5q̂70 is based on the period life table for 2007. Estimates for Taiwan come from the Human Mortality Database (2013). Estimates for Costa Rica come from the Centro Centroamericano de Población (2010).

In auxiliary models (not shown), we explored whether the assessments are more predictive for older (70+) than for younger (<70) persons and for men than women results which would be consistent with tentative findings from a meta-analysis (Cooper et al. 2010). In line with results from the EPESE study in the US (Reuben et al. 2004), we also considered whether performance tests have greater prognostic value among those who report themselves to be highly functioning (i.e., no self-reported limitations) compared with those reporting limitations. Among the many interactions tested, few were significant, close to what we would expect by chance. Thus, we found little evidence that the prognostic value of the performance tests differs by age, sex or the presence of self-reported limitations.

5. Discussion

Health interview surveys routinely ask respondents about their physical limitations, but they are much less likely to measure respondents’ physical performance with trained observers, presumably because of the expense, complex logistics, and burden for participants. Because it is easier to obtain self-reports than to administer performance tasks, it is important to ascertain the additional value derived from such tests. Of course, their value depends in large part on the particular questions being addressed. In this analysis we consider only one criterion: the prognostic value of four types of performance tests over and above self-reported limitations for five-year mortality. For two population-based samples of older individuals living in the community, we find that PEF and grip strength (and gait speed in Costa Rica) yield sizeable improvements in mortality prediction above and beyond self-reported limitations. The gains in prediction are stronger in Taiwan than in Costa Rica but, because of the smaller sample size, are less likely to be statistically significant in Taiwan.

The paucity of previous studies addressing this issue, combined with variability across data sets in the measures collected, has yielded few robust findings. One study suggested that the chair stand test provided the strongest mortality prediction (Cesari et al. 2008), while other work favored walking speed (Al Snih et al. 2002; Markides et al. 2001), and yet another study reported that grip strength had the biggest effect in women and PEF had the biggest effect in men (Rosero-Bixby and Dow 2012). None of these prior studies included all four performance tests. Our findings, which are remarkably similar in Taiwan and Costa Rica, reveal that for both countries – in the presence of controls for self-reported ADL and mobility limitations – PEF has the strongest association with mortality. In models that include all four assessments, PEF and grip strength measures significantly predict mortality, whereas chair stand and walking speed do not. The discrepancies between our results and previous findings are likely due to two factors. One relates to the batteries of tests used in the earlier studies: chair stands and walking speed were generally included, whereas grip strength and PEF were not. A second explanation is that previous studies controlled for self-reported ADL limitations, but not mobility limitations. Tabulations (not shown) indicated that, when we omitted controls for self-reported limitations, the advantage of PEF and grip strength relative to the other two performance tasks was diminished, particularly in Costa Rica. This result is not surprising, as ADL limitations, mobility limitations, chair stand speed, and walking speed all reflect lower extremity function.

Many studies have examined the links between PEF or grip strength and mortality (although only a few have done so in conjunction with other performance tests and self-reported limitations). Most have found strong and robust associations between each of these two performance tasks and all-cause mortality (Cook et al. 1991; Cooper et al. 2010; Rosero-Bixby and Dow 2012; Simons et al. 2011; Vaz Fragoso et al. 2008). Although explanations for these strong links remain unclear (Gale et al. 2007; Schrack, Simonsick, and Ferrucci 2010), multiple mechanisms are likely to be involved. Rantanen et al. (2003) argue that grip strength, which reflects overall muscle strength, predicts survival in part because of chronic diseases that result in muscle impairment via such mechanisms as nutritional deficiency, inflammation, physical inactivity, and depression. However, even in the absence of recognizable disease, weak muscle strength may increase susceptibility to injury and compromise subsequent healing (Rantanen et al. 2003). It could also reflect early life nutrition and fetal development (Gale et al. 2007; Rantanen et al. 2003) or be a marker of subclinical disease (Rantanen et al. 2003). Similarly, Vaz Fragoso et al. (2008) suggest that a diminished PEF reflects more than chronic lung disease or exposure to smoke or pollution; for example, low PEF may indicate impaired respiratory muscle strength, upper-extremity functional limitations, and poor cognitive function. Both PEF and grip strength are thought to capture a person’s overall vigor or vitality (Cook et al. 1991). Despite the cost of specialized equipment, both measures have some advantages as risk assessment tools in household surveys, particularly in developing countries: Unlike chair stands, there is no need to adjust for differences in chair size; in contrast to the timed walk, they do not require unobstructed space; and there is no need to time performance, which may increase measurement error. Nevertheless, accurate measurements of all four performance assessments necessitate careful interviewer training and calibration of instruments.

One limitation of this study is the attrition of respondents due to loss to follow-up and mortality. As with any study of older populations, the analysis is restricted to those who survived to the age range targeted by the survey. In the case of the Taiwanese survey, there was also attrition resulting from loss to follow-up during the course of the longitudinal study. A second disadvantage – as with any application of performance tasks – is that some respondents are unwilling or unable to participate or they are excluded because of health and safety concerns. This proportion is much lower for grip strength (5 per cent in SEBAS; 11 per cent in CRELES) than for chair stands (12 per cent and 27 per cent, respectively). As indicated by the hazard ratios in Tables 2 and 3, being unable to perform a test (including those excluded for health reasons) is often a strong predictor of impending death. However, these indicators are not well-defined: Exclusion rates vary by the specific task and are likely to depend on the underlying demographic, health, social, and cultural characteristics of the sample. They are also likely to differ by the degree of caution and safety concerns of the survey investigators, thereby generating variability across surveys in the strength of the association between these variables and health status. A third concern is limited statistical power, particularly in Taiwan, where numerous estimates are substantial but not statistically significant.

Few researchers question the importance of collecting self-reported data on physical limitations in older populations. In this study we demonstrate the utility of performance assessments, even when high quality self-reported measures are available. Nonetheless, our findings indicate that it may not be necessary or cost effective to collect a large battery of performance tests, in part because of substantial correlations among the tasks. For predicting mortality in community-based populations – at least within a five-year period – measurements of PEF and grip strength may suffice. Although our findings are reinforced by the consistency of estimates for Costa Rica and Taiwan, future work should examine the robustness of these results to different settings and populations, and, most importantly, to alternative health outcomes. For example, it is quite possible that performance measures that are redundant for predicting mortality are strong predictors of future disability. With the recent expansion of data collection in longitudinal household surveys, future analyses should also assess the predictive power of changes in performance between survey waves, which may enhance models of health and survival beyond what can ascertained from cross-sectional measures. They may also add prognostic value to models that include changes in self-reported health and functional ability as well as changes in physiological markers (e.g., biomarkers of the cardiovascular or immune system).