Agreement Between 35 Published Frailty Scores in the General Population.

In elderly populations, frailty is associated with higher mortality risk. Although many frailty scores (FS) have been proposed, no single score is considered the gold standard. We aimed to evaluate the agreement between a wide range of FS in the English Longitudinal Study of Ageing (ELSA). Through a literature search, we identified 35 FS that could be calculated in ELSA wave 2 (2004-2005). We examined agreement between each frailty score and the mean of 35 FS, using a modified Bland-Altman model and Cohen's kappa (κ). Missing data were imputed. Data from 5,377 participants (ages ≥60 years) were analyzed (44.7% men, 55.3% women). FS showed widely differing degrees of agreement with the mean of all scores and between each pair of scores. Frailty classification also showed a very wide range of agreement (Cohen's κ = 0.10-0.83). Agreement was highest among "accumulation of deficits"-type FS, while accuracy was highest for multidimensional FS. There is marked heterogeneity in the degree to which various FS estimate frailty and in the identification of particular individuals as frail. Different FS are based on different concepts of frailty, and most pairs cannot be assumed to be interchangeable. Research results based on different FS cannot be compared or pooled.

Vulnerable elderly individuals are increasingly described in the literature as being frail—that is, having a decreased ability to recover from an adverse event (1). Three main approaches have been used to conceptually define frailty.

The first approach is the “phenotype of frailty” (2), which is a physiological model focused mainly on physical frailty and which describes frailty as a phenomenon of “weakness, decreased endurance and slow performance” (2, p. 154). This approach regards frailty as separate from disability and comorbidity. The operational definition of this approach defines frailty as the presence of at least 3 out of 5 criteria (prefrailty: 1 or 2 criteria). The second approach is “accumulation of deficits” (3), which is based on the accumulation of conditions or disabilities, emphasizing the number rather than the nature of deficits. The operational definition of this approach defines frailty with at least 30 variables (4) and includes disability and comorbidity (5). The third approach is the “multidimensional model” (6), which defines frailty as a dynamic state of loss affecting 1 or more areas of functioning, such as the cognitive, physical, and social domains. Finally, some frailty scores (FS) have been operationalized mainly as the presence of disability.

Frailty is associated with a higher risk of mortality, disability, falls, fractures, hospitalization, and institutionalization (7, 8). Some evidence indicates that exercise, caloric and protein support, vitamin D supplementation, and reduction of polypharmacy can be effective in preventing the progression of frailty and the occurrence of its adverse outcomes (9). Consequently, it is important to identify frail individuals and persons at risk at an early stage (7). However, it remains unclear which tool is best suited for this purpose.

The ability of FS to accurately produce stable and reproducible results has been partially studied (10). In a systematic review of FS, Bouillon et al. (11) found that 7 out of 27 scores had been assessed for both reliability and concurrent or predictive validity. In a recent study that assessed the validity and reproducibility of 8 commonly used FS in an elderly European general population, Theou et al. (12) found that the prevalence of frailty varied from 6.1% to 43.9%; across all 8 scales, 49.3% of participants were classified as nonfrail and 2.4% were classified as frail. The authors concluded that FS have significant differences regarding validity, feasibility, and predictive ability (12).

The absence of consensus on how to conceptually define frailty and the resulting plethora of scales and scores currently hampers both research in the field and implementation of frailty assessment in clinical practice. In order to enable comparison of studies of frailty performed with different FS and to facilitate the choice of FS for future studies, it is essential to quantify the degree of agreement between scores and to understand the sources of disagreement.

Based on the hypothesis that different FS may classify different subsets of a population as frail, we set out to study the agreement between a wide set of FS in a well-characterized cohort of elderly people: participants in the English Longitudinal Study of Ageing (ELSA).

METHODS

Study population/design

ELSA is an ongoing cohort study based on a large, nationally representative sample of the middle-aged and elderly (ages ≥50 years) English population. Information about participants is gathered at 2-year intervals (waves). All waves include administration of questionnaires concerning health determinants and physical and mental health. In addition, in waves 2, 4, and 6, participants underwent a clinical examination.

Ethical approval was obtained from the National Health Service Research Ethics Committee. Participants gave signed informed consent (13). ELSA data were accessed via the UK Data Service.

We carried out a cross-sectional analysis of data from wave 2 (2004–2005) of ELSA, as this was the first wave in which a comprehensive assessment of frailty indicators was performed. Since not all frailty-related variables were measured in participants under age 60 years, we restricted our analyses to those aged 60 years or more.

Identification and selection of FS

A PubMed search of the literature was performed (date range: January 1, 1970–August 31, 2015) with the following query: “((frailty [Title/Abstract]) AND score [Title/Abstract]).” Abstracts were checked for the publication of an original frailty score. Furthermore, FS were identified on the basis of references from recent review articles (11, 14–16). Published FS were selected for inclusion if at least 80% of the component variables were available in ELSA wave 2. If information on 1 or more underlying variables (maximally 20%) comprising a score was unobtainable from the data, the frailty score was calculated based on the available variables and the total score and the cutoff were refitted to the actual number of variables (17). Variables for which data were unavailable due to the ELSA study design were not imputed.

FS were calculated while trying to be faithful to the original scores. However, it was necessary to tailor some variables to the data. For some FS, this adaptation was based on previous publications (12). FS vary in yielding continuous, categorical, or binary outputs, each with different ranges. Each score was rescaled to the interval 0–1 by dividing the original score output by the highest possible value for each score. Some scores were additionally inverted ((rescaled score × −1) + 1) to conform to our definition of 0 representing the absence of frailty and 1 representing its presence.

Missing data

If data from an available underlying variable in ELSA were missing for some participants, multiple imputation was applied (18). The amount of missing data varied from 0.04% to 24.7%.

The maximum percentage of missing data was used to decide how many imputations to perform (19). Therefore, we imputed 30 times, using chained equations (the “mice” package in R (20)). To obtain optimally plausible values for the scores, we applied imputation to the original underlying variables and calculated FS a posteriori using imputed values.

All statistical analyses were performed on the 30 imputed data sets, and resulting estimates were pooled according to the Rubin rules (18, 21). All results presented in this paper were obtained on the basis of the multiple imputation procedure described above.

Statistical analyses

The prevalence of frailty was calculated for each frailty score prior to rescaling using the original, published cutoff points. To enable comparisons between scores, we calculated the mean value, median value, and standard deviation and all further analyses on the rescaled scores in the total population and also stratified by sex, age, and smoking status.

Agreement was analyzed using 3 parallel methods:

The modified Bland-Altman model (22). In the absence of an external gold standard for frailty, we chose the mean of the 35 analyzed FS (M35FS) as a global estimate of “true frailty.” The error (the difference between each score and the M35FS) was plotted on the y-axis against the M35FS on the x-axis. Linear regression was used to calculate the dependence of each score's error (over- or underestimation) on the severity of frailty, as well as to calculate its limits of agreement. The degree of under-/overestimation was estimated at the median of the M35FS (model A).

Traditional pairwise Bland-Altman models. Traditional pairwise Bland-Altman models were built comparing all 595 possible pairs of FS. The error (the difference between each score and the mean of the 2 compared FS) was plotted on the y-axis against a rough estimate of the “true frailty,” defined as the mean of the 2 FS, on the x-axis. The width of prediction intervals and the absolute error (calculated at the median point on the x-axis) were analyzed (model B).

Cohen's kappa (κ). In order to enable comparisons across all 595 possible pairs of 35 FS in spite of different underlying concepts of frailty, different cutoff points, and the absence of a published cutoff point in some cases, we also calculated κ applying an arbitrary cutoff across all scores (defining the highest 20% of scores as “frail”). In cases where a score category straddled the 20% cutoff level, κ was calculated using a 20-bootstrap resample procedure, which classified participants from the straddling category randomly as frail/nonfrail in the proportion necessary to achieve an overall 20% prevalence of frailty. We calculated 95% confidence intervals for κ on the basis of Rubin's method for covariance and confidence interval calculation in imputed data. The mean within-imputation variance, the between-imputations variance, the total variance, and finally the confidence intervals were calculated (21).

FS were grouped into 4 main operationalization models: “phenotype of frailty” for scores based mainly on physical functioning variables; “accumulation of deficits” for scores based on various domains and at least 30 variables; “multidimensional model” for scores that analyzed at least 3 domains of functioning and included fewer than 30 variables; and “disability model” for scores based mainly on disability variables. FS were also grouped according to the stated target population: community-dwelling or clinical setting.

In addition, to assess agreement with regard to a binary or categorical definition of frailty, κ was calculated for pairs of FS with a published cutoff level (29 out of 35 FS).

RESULTS

We analyzed data from all 5,377 participants aged 60 years or over (44.7% men and 55.3% women) who attended the ELSA wave 2 clinical examination.

We identified 67 original FS through the literature search. Thirty-five of the 67 scores (52.2%) could be calculated with ELSA wave 2 data. Web Table 1 (available at https://academic.oup.com/aje) shows the list of included and excluded FS. Web Table 2 shows details of all variables for the 35 FS and their adapted version in the ELSA data set.

Table 1 presents the general characteristics of the study population by sex.

Table 1.

General Characteristics of Participants (n = 5,377) in Wave 2 of the English Longitudinal Study of Ageing, 2004–2005

Characteristic	Men (n = 2,401)		Women (n = 2,976)
	Mean (SD)	%	Mean (SD)	%
Age, years[a]	70.8 (7.6)		71.5 (8.2)[b]
Body mass index[c,d,e]	27.7 (4.1)		28.0 (5.2)
Total cognitive score (per point)[d,e,f]	26.7 (6.4)		27.6 (6.7)[b]
Marital status (currently married)[d,g]		75.0		52.8[b]
Education (no high school qualification)[d,g]		37.3		48.7[b]
Smoking status[d,g]
Current smoker		12.8		11.9[b]
Former smoker		61.9		43.5[b]
Never smoker		25.4		44.6[b]
Physical activity[d,g,h]
None (sedentary)		6.6		7.9[b]
Mild		22.7		30.9[b]
Moderate		50.9		48.7[b]
Vigorous		19.8		12.5[b]
Chronic disease[i,j]
Diabetes		11.4		8.4[b]
Hypertension		46.3		49.6[b]
Myocardial infarction		11.6		4.9[b]
Stroke		6.9		5.8
Cancer		8.4		9.8
Lung disease		8.5		8.4
Arthritis		33.6		50.7[b]
Depression symptoms (CES-D score ≥4)[d,g,k]		25.7		37.1[b]

Abbreviations: CES-D, Center for Epidemiologic Studies Depression Scale; SD, standard deviation.

aP value was derived from an unpaired t test.

bP < 0.05.

c Weight (kg)/height (m)2.

d Imputed data.

e Linear regression model.

f Sum of memory and executive indices; values range from 0 (worst) to 50 (best).

g Logistic regression.

h Self-reported frequency of mild, moderate, or vigorous activity at least once per week.

iP value was derived from a χ2 test.

j Self-reported ever diagnosis of the condition.

k Assessed by means of the 8-item CES-D.

Table 2 presents the characteristics of the 35 FS that were analyzed in this study (2, 4, 23–53). The frailty score with the highest proportion of individual-level missing values was the 70-item Frailty Index (40.5%), while the lowest proportions of missing values (0.1%) were observed for “WHOAFC and self-reported health” and the Vulnerable Elders Survey. Most of the scores (29 of 35) had published cutoffs to define frailty.

Table 2.

Characteristics of Frailty Scores Calculated Among Participants in Wave 2 of the English Longitudinal Study of Ageing, 2004–2005

First Author, Year (Reference No.)	Frailty Measure	Country	Model	Aim	Definition of Score	No. of Variables Defined[a]	No. of Variables Calculated[b]	% of Missing Data[c]
Klein, 2003 (23)	Beaver Dam Eye Study Index	United States	POF	CD	Continuous	4	4	26.5
Gill, 2002 (24)	Physical Frailty Index	United States	POF	CS	Categorical	2	2	19.1
Cesari, 2014 (25)	FiND Questionnaire	France	POF	CD	Binary	5	5	1.3
Abellan van Kan, 2008 (26)	Frail Scale	France	POF	CS	Categorical	5	5	1.3
Fried, 2001 (2)	Phenotype of Frailty	United States	POF	CD	Categorical	5	5	13.4
Rothman, 2008 (27)	Modified Phenotype of Frailty	United States	POF	CD	Continuous	7	7	15.8
Ensrud, 2007 (28)	Study of Osteoporotic Fractures	United States	POF	CS	Categorical	3	3	14.3
Guralnik, 1994 (29)	Short Physical Performance Battery	United States	POF	CD	Binary	3	3	21.8
Chin, 1999 (30)	ZED1 (Physical Activity and Low Energy)	Netherlands	POF	CD	Binary	2	2	0.5
Chin, 1999 (30)	ZED2 (Physical Activity and Weight Loss)	Netherlands	POF	CD	Binary	2	2	0.8
Chin, 1999 (30)	ZED3 (Physical Activity and Low BMI)	Netherlands	POF	CD	Binary	2	2	4.7
Freiheit, 2010 (31)	Brief Frailty Index	Canada	MD	CS	Binary	5	5	17.3
Hubbard, 2009 (32)	Modified Frailty Score	United Kingdom	MD	CS	Categorical	5	5	21.6
Balducci, 2000 (33)	CGAST	United States	MD	CS	Categorical	9	9	10.4
Ravaglia, 2008 (34)	Conselice Study of Brain Aging Score	Italy	MD	CD	Binary	9	9	23.4
Rolfson, 2006 (35)	Edmonton Frail Scale	Canada	MD	CS	Binary	9	9	12.6
Cacciatore, 2005 (36)	Frailty Staging System	Italy	MD	CS	Categorical	7	7	3.5
Bellera, 2012 (37)	G-8 geriatric screening tool	France	MD	CS	Categorical	8	7	4.2
Steverink, 2001 (38)	Groningen Frailty Indicator	Netherlands	MD	CS	Binary	11	11	14.3
Brody, 1997 (39)	Health Status Form	United States	MD	CD	Continuous	4	4	11.9
Puts, 2005 (40)	Static/Dynamic Frailty Index	Netherlands	MD	CD	Binary	9	9	25.3
Maly, 1997 (41)	Screening Instrument	United States	MD	CD	Binary	6	6	11.3
Hábert, 1996 (42)	Sherbrooke Postal Questionnaire	Canada	MD	CD	Binary	6	6	14.4
Di Bari, 2014 (43)	Inter-Frail Questionnaire	Italy	MD	CD	Binary	10	8	14.8
Gobbens, 2010 (44)	Tilburg Frailty Indicator	Netherlands	MD	CD	Binary	15	15	22.1
Jones, 2004 (45)	Comprehensive Geriatric Assessment	Canada	AOD	CD	Categorical	44	41	35.1
de Vries, 2013 (46)	Evaluative Frailty Index for Physical Activity	Netherlands	AOD	CD	Continuous	50	42	22.8
Searle, 2008 (4)	40-item Frailty Index	Canada	AOD	CD	Binary	40	37	23.7
Theou, 2013 (47)	70-item Frailty Index (SHARE)	Canada	AOD	CD	Binary	70	62	40.5
Fang, 2012 (48)	Frailty Index (BLSA)	China	AOD	CD	Continuous	35	29	17.5
Kulminski, 2007 (49)	Long Term Care Survey Frailty Index	United States	AOD	CD	Continuous	32	26	0.9
Dayhoff, 1998 (50)	WHOAFC and self-reported health	United States	DA	CD	Binary	15	14	0.1
Morris, 1984 (51)	HRCA Vulnerability Index	United States	DA	CD	Binary	10	10	18.9
Rockwood, 2005 (52)	CSHA Clinical Frailty Scale	Canada	DA	CS	Categorical	8	8	0.2
Saliba, 2001 (53)	Vulnerable Elders Survey	United States	DA	CD	Binary	13	12	0.1

Abbreviations: BLSA, Beijing Longitudinal Study of Ageing; BMI, body mass index; CS, clinical setting; CD, community-dwelling; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSHA, Canadian Study of Health and Aging; DA, disability; AOD, accumulation of deficits; ELSA, English Longitudinal Study of Ageing; FiND, Frail Non-Disabled; HRCA, Hebrew Rehabilitation Center for Aged; MD, multidimensional; POF, phenotype of frailty; SHARE, Survey of Health, Ageing and Retirement in Europe; WHOAFC, World Health Organization Assessment of Functional Capacity; ZED, Zutphen Elderly Study.

a Number of variables defined in the score.

b Number of variables calculated with the ELSA data set.

c Percentage of missing data within ELSA variables.

Prevalence as defined by the published cutoffs varied considerably. The mean prevalence of frailty was 23.1% (standard deviation, 19.7) for men (range, 0.8–65.0) and 28.9% (standard deviation, 21.9) for women (range, 1.0–72.4) (Table 3).

Table 3.

Prevalence of Frailty Among Participants in Wave 2 of the English Longitudinal Study of Ageing, 2004–2005

Frailty Model and Measure	Published Cutoff	Cutoff Used	Men			Women
			% Not Frail	% Prefrail	% Frail	% Not Frail	% Prefrail	% Frail
“Phenotype of frailty” model
Beaver Dam Eye Study Index	None	NA	NA	NA	NA	NA	NA	NA
Physical Frailty Index	>1 and >0	>1 and >0	81.9	14.9	3.2	73.8	20.4	5.8
FiND Questionnaire[a]	≥1	≥1	60.3	NA	39.7	52.3	NA	47.7
Frail Scale	>2; >0	>2; >0	61.5	31.2	7.4	53.1	38.0	8.9
Phenotype of Frailty	≥3; ≥1	≥3; ≥1	11.7	75.8	12.5	6.6	78.6	14.8
Modified Phenotype of Frailty	None	NA	NA	NA	NA	NA	NA	NA
Study of Osteoporotic Fractures	≥2; ≥1	≥2; ≥1	67.6	25.6	6.8	60.4	30.2	9.4
Short Physical Performance Battery	≤9	≤9	35.0	NA	65.0	29.0	NA	71.0
ZED1 (Physical Activity and Low Energy)	>1	>1	96.3	NA	3.7	95.5	NA	4.5
ZED2 (Physical Activity and Weight Loss)	>1	>1	98.1	NA	1.9	97.0	NA	3.0
ZED3 (Physical Activity and Low BMI)	>1	>1	99.2	NA	0.8	99.0	NA	1.0
Multidimensional model
Brief Frailty Index	≥3	≥3	91.0	NA	9.0	82.8	NA	17.2
Modified Frailty Score	>2; >0	>2; >0	20.0	16.3	63.7	12.9	18.6	68.5
CGAST	>2; >0	>2; >0	17.6	49.0	33.4	12.4	45.9	41.7
Conselice Study of Brain Aging Score	≥3	≥3	47.3	NA	52.7	65.2	NA	34.8
Edmonton Frail Scale	≥8	≥8	96.7	NA	3.3	95.2	NA	4.8
Frailty Staging System	≥4; ≥2	≥4; ≥2	58.2	30.6	11.2	50.6	34.4	15.0
G-8 geriatric screening tool	≤14	≤12	37.9	NA	62.1	27.6	NA	72.4
Groningen Frailty Indicator	≥4	≥4	64.8	NA	35.2	55.9	NA	44.1
Health Status Form	None	NA	NA	NA	NA	NA	NA	NA
Static/Dynamic Frailty Index	≥3	≥3	65.7	NA	34.3	45.7	NA	54.3
Screening Instrument	≥3	≥3	95.6	NA	4.4	92.1	NA	7.9
Sherbrooke Postal Questionnaire	≥2	≥2	81.3	NA	18.7	71.7	NA	28.3
Inter-Frail Questionnaire	≥5	≥5	98.4	NA	1.6	97.2	NA	2.8
Tilburg Frailty Indicator	≥5	≥5	64.1	NA	35.9	53.3	NA	46.7
“Accumulation of deficits” model
Comprehensive Geriatric Assessment	≥7; ≥13	≥6; ≥12	57.8	30.7	11.5	42.1	40.2	17.7
Evaluative Frailty Index for Physical Activity	None	NA	NA	NA	NA	NA	NA	NA
40-item Frailty Index	>0.2	>0.2	68.8	NA	31.2	56.7	NA	43.3
70-item Frailty Index (SHARE)	≥0.25	≥0.25	74.7	NA	25.3	62.7	NA	37.3
Frailty Index (BLSA)	None	NA	NA	NA	NA	NA	NA	NA
Long Term Care Survey Frailty Index	None	NA	NA	NA	NA	NA	NA	NA
Disability model
WHOAFC and self-reported health	≥21 or SRH = poor[b]	≥4 or SRH = poor[c]	81.5	NA	18.5	74.1	NA	25.9
HRCA Vulnerability Index	“A” box ≥1 or “A” box = 0 and “B” box ≥0[d]	“A” box ≥1 or “A” box = 0 and “B” box ≥0[d]	73.8	NA	26.2	57.6	NA	42.4
CSHA Clinical Frailty Scale	1–7	≥5	86.8	NA	13.2	83.0	NA	17.0
Vulnerable Elders Survey	≥3	≥2.8	73.9	NA	26.1	63.5	NA	36.5

Abbreviations: BLSA, Beijing Longitudinal Study of Ageing; BMI, body mass index; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSHA, Canadian Study of Health and Aging; FiND, Frail Non-Disabled; HRCA, Hebrew Rehabilitation Center for Aged; NA, not applicable; SHARE, Survey of Health, Ageing and Retirement in Europe; WHOAFC, World Health Organization Assessment of Functional Capacity; ZED, Zutphen Elderly Study.

a Frailty calculated as frail + disabled.

b Cutoff calculated dividing each variable by 5.

c Self-rated health.

d “A” box: list of 7 questions; yes = 1, no = 0. “B” box: 2 additional questions; yes = 1, no = 0.

Table 4 shows the mean frailty score values after rescaling to the 0–1 range in the whole population globally, as well as stratified by sex, age, and smoking status. Across FS, women, older participants, and smokers/former smokers were frailer than men, younger participants, and never smokers, respectively.

Table 4.

Rescaled (0–1) Mean (Standard Deviation)[a] Frailty Scores and P Values for Differences[b] Between Frailty Scores, Overall and by Sex, Age, and Smoking Status, Among Participants in Wave 2 of the English Longitudinal Study of Ageing, 2004–2005

Frailty Model and Measure	Total	Sex		Age		Smoking Status
		Male[c]	Female	≤70 Years[c]	>70 Years	Never Smoker[c]	Ex-Smoker	Current Smoker
“Phenotype of frailty” model
Beaver Dam Eye Study Index	0.37 (0.27)	0.37 (0.26)	0.38 (0.28)	0.26 (0.23)	0.49 (0.26)[d]	0.35 (0.27)	0.37 (0.27)[d]	0.42 (0.25)[d]
Physical Frailty Index	0.14 (0.27)	0.11 (0.24)	0.16 (0.29)[d]	0.07 (0.19)	0.21 (0.32)[d]	0.13 (0.27)	0.14 (0.27)	0.15 (0.28)
FiND Questionnaire	0.23 (0.23)	0.20 (0.22)	0.25 (0.24)[d]	0.18 (0.20)	0.28 (0.25)[d]	0.21 (0.22)	0.23 (0.23)[d]	0.28 (0.24)[d]
Frail Scale	0.14 (0.20)	0.13 (0.19)	0.16 (0.21)[d]	0.11 (0.17)	0.19 (0.22)[d]	0.13 (0.20)	0.15 (0.20)[d]	0.19 (0.22)[d]
Phenotype of Frailty	0.30 (0.24)	0.27 (0.23)	0.32 (0.24)[d]	0.23 (0.20)	0.38 (0.25)[d]	0.28 (0.23)	0.30 (0.23)[d]	0.34 (0.25)[d]
Modified Phenotype of Frailty	0.33 (0.23)	0.27 (0.23)	0.35 (0.23)[d]	0.25 (0.20)	0.41 (0.24)[d]	0.31 (0.23)	0.33 (0.23)[d]	0.38 (0.24)[d]
Study of Osteoporotic Fractures	0.15 (0.22)	0.13 (0.21)	0.17 (0.23)[d]	0.11 (0.19)	0.20 (0.25)[d]	0.14 (0.22)	0.15 (0.22)	0.19 (0.24)[d]
Short Physical Performance Battery[e]	0.34 (0.19)	0.32 (0.17)	0.35 (0.20)[d]	0.26 (0.15)	0.42 (0.20)[d]	0.33 (0.19)	0.34 (0.19)	0.34 (0.19)
ZED1 (Physical Activity and Low Energy)	0.32 (0.28)	0.29 (0.27)	0.35 (0.28)[d]	0.27 (0.25)	0.38 (0.29)[d]	0.30 (0.27)	0.32 (0.28)[d]	0.39 (0.30)[d]
ZED2 (Physical Activity and Weight Loss)	0.29 (0.25)	0.27 (0.24)	0.31 (0.25)[d]	0.24 (0.21)	0.35 (0.27)[d]	0.29 (0.25)	0.29 (0.24)	0.33 (0.25)[d]
ZED3 (Physical Activity and Low BMI)	0.27 (0.21)	0.24 (0.20)	0.30 (0.22)[d]	0.23 (0.19)	0.31 (0.22)[d]	0.26 (0.21)	0.26 (0.20)	0.33 (0.24)[d]
Multidimensional model
Brief Frailty Index	0.31 (0.22)	0.27 (0.20)	0.34 (0.23)[d]	0.25 (0.20)	0.37 (0.22)[d]	0.29 (0.22)	0.31 (0.22)	0.36 (0.22)[d]
Modified Frailty Score	0.34 (0.22)	0.32 (0.22)	0.37 (0.22)[d]	0.25 (0.19)	0.45 (0.21)[d]	0.34 (0.22)	0.34 (0.22)	0.38 (0.21)[d]
CGAST	0.22 (0.16)	0.20 (0.12)	0.24 (0.16)[d]	0.19 (0.15)	0.26 (0.16)[d]	0.21 (0.16)	0.23 (0.16)[d]	0.25 (0.17)[d]
Conselice Study of Brain Aging Score	0.32 (0.15)	0.37 (0.15)	0.29 (0.15)[d]	0.26 (0.13)	0.39 (0.15)[d]	0.30 (0.15)	0.34 (0.16)[d]	0.34 (0.15)[d]
Edmonton Frail Scale	0.16 (0.14)	0.15 (0.13)	0.17 (0.14)[d]	0.13 (0.12)	0.20 (0.14)[d]	0.15 (0.13)	0.17 (0.13)[d]	0.19 (0.15)[d]
Frailty Staging System	0.30 (0.22)	0.28 (0.21)	0.16 (0.21)[d]	0.22 (0.19)	0.38 (0.23)[d]	0.28 (0.22)	0.31 (0.22)[d]	0.32 (0.22)[d]
G-8 geriatric screening tool[e]	0.21 (0.14)	0.19 (0.14)	0.23 (0.15)[d]	0.16 (0.12)	0.26 (0.15)[d]	0.20 (0.15)	0.21 (0.14)[d]	0.24 (0.15)[d]
Groningen Frailty Indicator	0.25 (0.16)	0.23 (0.15)	0.26 (0.16)[d]	0.20 (0.14)	0.26 (0.15)[d]	0.23 (0.15)	0.25 (0.15)[d]	0.28 (0.16)[d]
Health Status Form	0.16 (0.20)	0.14 (0.19)	0.17 (0.21)[d]	0.08 (0.14)	0.24 (0.23)[d]	0.14 (0.20)	0.17 (0.21)[d]	0.16 (0.19)
Static/Dynamic Frailty Index	0.34 (0.17)	0.30 (0.23)	0.37 (0.17)[d]	0.27 (0.14)	0.41 (0.17)[d]	0.33 (0.17)	0.33 (0.17)	0.38 (0.17)[d]
Screening Instrument	0.17 (0.17)	0.13 (0.16)	0.19 (0.18)[d]	0.14 (0.16)	0.20 (0.18)[d]	0.16 (0.17)	0.17 (0.17)[d]	0.19 (0.19)[d]
Sherbrooke Postal Questionnaire	0.24 (0.17)	0.22 (0.17)	0.26 (0.18)[d]	0.18 (0.14)	0.30 (0.18)[d]	0.23 (0.17)	0.24 (0.17)	0.25 (0.17)[d]
Inter-Frail Questionnaire	0.22 (0.17)	0.22 (0.16)	0.23 (0.17)[d]	0.16 (0.14)	0.27 (0.17)[d]	0.21 (0.16)	0.22 (0.17)	0.25 (0.17)[d]
Tilburg Frailty Indicator	0.33 (0.18)	0.30 (0.16)	0.35 (0.18)[d]	0.27 (0.15)	0.39 (0.18)[d]	0.32 (0.17)	0.33 (0.17)[d]	0.37 (0.18)[d]
“Accumulation of deficits” model
Comprehensive Geriatric Assessment	0.18 (0.11)	0.16 (0.11)	0.19 (0.12)[d]	0.15 (0.10)	0.21 (0.12)[d]	0.17 (0.11)	0.18 (0.11)[d]	0.20 (0.12)[d]
Evaluative Frailty Index Physical Activity	0.21 (0.14)	0.19 (0.14)	0.22 (0.15)[d]	0.17 (0.13)	0.26 (0.15)[d]	0.19 (0.14)	0.21 (0.14)[d]	0.24 (0.15)[d]
40-item Frailty Index	0.20 (0.14)	0.17 (0.13)	0.21 (0.14)[d]	0.16 (0.12)	0.23 (0.14)[d]	0.18 (0.14)	0.20 (0.14)[d]	0.22 (0.15)[d]
70-item Frailty Index (SHARE)	0.21 (0.14)	0.18 (0.13)	0.23 (0.15)[d]	0.17 (0.13)	0.25 (0.15)[d]	0.19 (0.14)	0.21 (0.14)[d]	0.24 (0.15)[d]
Disability model
Frailty Index (BLSA)	0.17 (0.13)	0.15 (0.12)	0.18 (0.14)[d]	0.13 (0.11)	0.21 (0.14)[d]	0.16 (0.13)	0.17 (0.13)[d]	0.18 (0.14)[d]
Long Term Care Survey Frailty Index	0.14 (0.11)	0.13 (0.10)	0.14 (0.11)[d]	0.11 (0.09)	0.17 (0.11)[d]	0.13 (0.11)	0.14 (0.10)[d]	0.15 (0.11)[d]
WHOAFC and self-reported health	0.17 (0.20)	0.14 (0.19)	0.19 (0.21)[d]	0.13 (0.19)	0.21 (0.22)[d]	0.15 (0.19)	0.18 (0.21)[d]	0.20 (0.22)[d]
HRCA Vulnerability Index	0.16 (0.18)	0.14 (0.17)	0.18 (0.19)[d]	0.13 (0.16)	0.20 (0.19)[d]	0.14 (0.17)	0.16 (0.18)[d]	0.20 (0.19)[d]
CSHA Clinical Frailty Scale	0.33 (0.22)	0.31 (0.21)	0.34 (0.22)[d]	0.29 (0.19)	0.37 (0.24)[d]	0.31 (0.21)	0.33 (0.22)[d]	0.36 (0.23)[d]
Vulnerable Elders Survey	0.19 (0.20)	0.16 (0.18)	0.22 (0.21)[d]	0.12 (0.16)	0.27 (0.21)[d]	0.18 (0.20)	0.20 (0.20)[d]	0.21 (0.20)[d]

Abbreviations: BLSA, Beijing Longitudinal Study of Ageing; BMI, body mass index; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSHA, Canadian Study of Health and Aging; FiND, Frail Non-Disabled; HRCA, Hebrew Rehabilitation Center for Aged; SHARE, Survey of Health, Ageing and Retirement in Europe; WHOAFC, World Health Organization Assessment of Functional Capacity; ZED, Zutphen Elderly Study.

a Standard deviation was calculated according to the rules of Rubin.

b Least-squares regression was used for continuous dependent variables.

c Reference category.

dP < 0.05 for women compared with men, older persons (>70 years) compared with younger persons (≤70 years), ex-smokers compared with never smokers, and current smokers compared with never smokers.

e Inverted scale.

Table 5 displays the median κ values. It also shows the median prediction interval widths and absolute error of under-/overestimation in analyses based on model A and model B.

Table 5.

Cohen's Kappa Coefficients for Agreement Between Frailty Scores and Results of Bland-Altman Model Analysis for Participants in Wave 2 of the English Longitudinal Study of Ageing, 2004–2005

Frailty Model and Measure	Median κ	Width of Prediction Interval		Absolute Error
		M35FS[a]	PFS[b]	M35FS[a]	PFS[b]
“Phenotype of frailty” model
Beaver Dam Eye Study Index	0.318	0.747	0.795	0.123	0.140
Physical Frailty Index	0.298	0.732	0.801	0.113	0.135
FiND Questionnaire	0.508	0.415	0.593	0.025	0.058
Frail Scale	0.391	0.421	0.598	0.099	0.090
Phenotype of Frailty	0.402	0.501	0.663	0.048	0.065
Modified Phenotype of Frailty	0.451	0.427	0.624	0.075	0.096
SOF Index	0.254	0.597	0.736	0.089	0.097
Short Physical Performance Battery	0.396	0.499	0.672	0.102	0.113
ZED1 (Physical Activity and Low Energy)	0.363	0.688	0.759	0.067	0.063
ZED2 (Physical Activity and Weight Loss)	0.191	0.760	0.818	0.057	0.061
ZED3 (Physical Activity and Low BMI)	0.195	0.730	0.907	0.050	0.063
Multidimensional model
Brief Frailty Index	0.316	0.629	0.746	0.073	0.085
Modified Frailty Score	0.293	0.454	0.640	0.076	0.096
CGAST	0.419	0.347	0.552	0.057	0.048
Conselice Study of Brain Aging Score	0.387	0.430	0.600	0.099	0.111
Edmonton Frail Scale	0.454	0.242	0.454	0.092	0.075
Frailty Staging System	0.447	0.500	0.621	0.054	0.068
G-8 geriatric screening tool	0.352	0.355	0.531	0.013	0.064
Groningen Frailty Indicator	0.513	0.280	0.492	0.014	0.065
Health Status Form	0.430	0.485	0.663	0.082	0.072
Static/Dynamic Frailty Index	0.389	0.429	0.612	0.104	0.120
Screening Instrument	0.344	0.479	0.662	0.061	0.053
Sherbrooke Postal Questionnaire	0.305	0.515	0.580	0.015	0.065
Inter-Frail Questionnaire	0.445	0.385	0.699	0.015	0.064
Tilburg Frailty Indicator	0.472	0.339	0.569	0.091	0.111
“Accumulation of deficits” model
Comprehensive Geriatric Assessment	0.493	0.212	0.424	0.042	0.059
Evaluative Frailty Index for Physical Activity	0.536	0.219	0.462	0.024	0.061
40-item Frailty Index	0.535	0.201	0.450	0.035	0.060
70-item Frailty Index (SHARE)	0.518	0.231	0.468	0.021	0.063
Frailty Index (BLSA)	0.500	0.234	0.449	0.060	0.060
Long Term Care Survey Frailty Index	0.435	0.236	0.440	0.080	0.066
Disability model
WHOAFC and self-reported health	0.463	0.436	0.616	0.071	0.065
HRCA Vulnerability Index	0.444	0.391	0.589	0.076	0.066
CSHA Clinical Frailty Scale	0.380	0.561	0.721	0.086	0.093
Vulnerable Elders Survey	0.437	0.392	0.602	0.050	0.050

Abbreviations: BLSA, Beijing Longitudinal Study of Ageing; BMI, body mass index; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSHA, Canadian Study of Health and Aging; FiND, Frail Non-Disabled; HRCA, Hebrew Rehabilitation Center for Aged; M35FS, mean of 35 frailty scores; PFS, pair of frailty scores; SHARE, Survey of Health, Ageing and Retirement in Europe; SOF, Study of Osteoporotic Fractures; WHOAFC, World Health Organization Assessment of Functional Capacity; ZED, Zutphen Elderly Study.

a Bland-Altman analysis with x-axis equal to the M35FS and y-axis equal to the difference between each score and M35FS; absolute error was calculated at the median value of the M35FS.

b Bland-Altman analysis with x-axis equal to the mean of PFS and y-axis equal to the difference between each score and the mean of PFS; absolute error was calculated at the median value of the mean of PFS.

Some FS show over- or underestimation, which can be seen when the regression line deviates from 0 at the median point of frailty (0.18) (Web Figure 1). The degree of over- or underestimation can vary according to frailty level. Some scores show wider prediction intervals than others. On the right side of each Bland-Altman plot, a density plot displays the distribution of the error.

The FS that showed the narrowest prediction interval widths were the 40-item Frailty Index with model A and the Comprehensive Geriatric Assessment with model B. Both FS belong to the “accumulation of deficits” model category.

Figure 1 shows a heat map of κ values for all 595 pairs of scores. The scores are grouped by frailty model category and then sorted by each score's median κ within each category. The highest κ value was observed for the Evaluative Frailty Index for Physical Activity (Table 5 and Figure 1). κ values ranged from 0.10 to 0.83 and were ≥0.80 for 0.8% of pairs, ≥0.60 and <0.80 for 10.4% of pairs, ≥0.40 and <0.60 for 35.3% of pairs, ≥0.20 and <0.40 for 45.9% of pairs, and <0.20 for 7.6% of pairs (details of estimates and 95% confidence intervals are shown in Web Table 3). For the 29 FS that had a published cutoff point, additional results with κ calculated using these cutoffs are shown in Web Table 4.

Figure 1.

Agreement (calculated with Cohen's κ) between pairs of frailty scores (595 combined pairs of scores) among participants in wave 2 of the English Longitudinal Study of Ageing, 2004–2005. The plot is sorted by frailty model and then from highest (red) to lowest (blue) median value of Cohen's κ coefficient. BDE, Beaver Dam Eye Study Index; BFI, Brief Frailty Index; BMI, body mass index; MFS, Modified Frailty Score; CGA, Comprehensive Geriatric Assessment; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSBA, Conselice Study of Brain Aging Score; EFIP, Evaluative Frailty Index for Physical Activity; EFS, Edmonton Frail Scale; FI40, 40-item Frailty Index; FI70, 70-item Frailty Index; FIBLSA, Frailty Index Beijing Longitudinal Study of Ageing; PFI, Physical Frailty Index; FiND, Frail Non-Disabled (FiND) Questionnaire; FS, Frail Scale; FSS, Frailty Staging System; G8, G-8 geriatric screening tool; GFI, Groningen Frailty Indicator; HRCA, Hebrew Rehabilitation Center for Aged Vulnerability Index; HSF, Health Status Form; NLTCS, National Long Term Care Survey Frailty Index; PHF, Phenotype of Frailty; MPHF, Modified Phenotype of Frailty; SDFI, Static/Dynamic Frailty Index; SHCFS, Canadian Study of Health and Aging Clinical Frailty Scale; SI, Screening Instrument; SOF, Study of Osteoporotic Fractures; SPPB, Short Physical Performance Battery; SPQ, Sherbrooke Postal Questionnaire; IFQ, Inter-Frail Questionnaire; TFI, Tilburg Frailty Indicator; VES13, Vulnerable Elders Survey; WHOAFC, World Health Organization Assessment of Functional Capacity; WHRH, WHOAFC and self-reported health; ZED1, Zutphen Elderly Study (Physical Activity and Low Energy); ZED2, Zutphen Elderly Study (Physical Activity and Weight Loss); ZED3, Zutphen Elderly Study (Physical Activity and Low BMI).

Prediction interval widths obtained with model B are plotted as a heat map in Figure 2, grouped by frailty model category. The narrowest median prediction interval was found for the 40-item Frailty Index with model A (Table 5) and the Comprehensive Geriatric Assessment with model B (Table 5 and Figure 2). Both FS belong to the “accumulation of deficits” model.

Figure 2.

Prediction interval widths obtained with Bland-Altman models for all 595 combined pairs of frailty scores, English Longitudinal Study of Ageing, wave 2 (2004–2005). The narrowest prediction interval widths are shown in red, and the widest are shown in blue. The plot is sorted by frailty model and then by the narrowest prediction interval. BDE, Beaver Dam Eye Study Index; BFI, Brief Frailty Index; BMI, body mass index; MFS, Modified Frailty Score; CGA, Comprehensive Geriatric Assessment; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSBA, Conselice Study of Brain Aging Score; EFIP, Evaluative Frailty Index for Physical Activity; EFS, Edmonton Frail Scale; FI40, 40-item Frailty Index; FI70, 70-item Frailty Index; FIBLSA, Frailty Index Beijing Longitudinal Study of Ageing; PFI, Physical Frailty Index; FiND, Frail Non-Disabled (FiND) Questionnaire; FS, Frail Scale; FSS, Frailty Staging System; G8, G-8 geriatric screening tool; GFI, Groningen Frailty Indicator; HRCA, Hebrew Rehabilitation Center for Aged Vulnerability Index; HSF, Health Status Form; NLTCS, National Long Term Care Survey Frailty Index; PHF, Phenotype of Frailty; MPHF, Modified Phenotype of Frailty; SDFI, Static/Dynamic Frailty Index; SHCFS, Canadian Study of Health and Aging Clinical Frailty Scale; SI, Screening Instrument; SOF, Study of Osteoporotic Fractures; SPPB, Short Physical Performance Battery; SPQ, Sherbrooke Postal Questionnaire; IFQ, Inter-Frail Questionnaire; TFI, Tilburg Frailty Indicator; VES13, Vulnerable Elders Survey; WHOAFC, World Health Organization Assessment of Functional Capacity; WHRH, WHOAFC and self-reported health; ZED1, Zutphen Elderly Study (Physical Activity and Low Energy); ZED2, Zutphen Elderly Study (Physical Activity and Weight Loss); ZED3, Zutphen Elderly Study (Physical Activity and Low BMI).

Figure 3 (grouped by model) shows a heat map of the absolute error calculated with model B. The lowest absolute error with model B was found for the Comprehensive Geriatric Assessment Screening Tests and, with model A, the G-8 geriatric screening tool (Table 5). Both FS belonged to the “multidimensional” model (Table 5 and Figure 3). Web Figures 2–4 show results from the same analysis as Figures 1–3, grouped according to the stated target population. Web Figures 5–11 are heat maps of κ values stratified by sex, age, and smoking status. Plots of model B are shown in the Web Appendix.

Figure 3.

Absolute error (over-/underestimation) of frailty in the median frailty value from the modified Bland-Altman model obtained with all 595 combined pairs of frailty scores, English Longitudinal Study of Ageing, wave 2 (2004–2005). The over-/underestimation is the absolute value of the intercept plus the product of the slope and the median. The intercept and slope are obtained from the Bland-Altman model. The median is calculated as the median value of the mean of 2 frailty scores for each pair. The lowest absolute errors are shown in red, and the highest are shown in blue. The plot is sorted by frailty model and then by the lowest absolute error. BDE, Beaver Dam Eye Study Index; BFI, Brief Frailty Index; BMI, body mass index; MFS, Modified Frailty Score; CGA, Comprehensive Geriatric Assessment; CGAST, Comprehensive Geriatric Assessment Screening Tests; CSBA, Conselice Study of Brain Aging Score; EFIP, Evaluative Frailty Index for Physical Activity; EFS, Edmonton Frail Scale; FI40, 40-item Frailty Index; FI70, 70-item Frailty Index; FIBLSA, Frailty Index Beijing Longitudinal Study of Ageing; PFI, Physical Frailty Index; FiND, Frail Non-Disabled (FiND) Questionnaire; FS, Frail Scale; FSS, Frailty Staging System; G8, G-8 geriatric screening tool; GFI, Groningen Frailty Indicator; HRCA, Hebrew Rehabilitation Center for Aged Vulnerability Index; HSF, Health Status Form; NLTCS, National Long Term Care Survey Frailty Index; PHF, Phenotype of Frailty; MPHF, Modified Phenotype of Frailty; SDFI, Static/Dynamic Frailty Index; SHCFS, Canadian Study of Health and Aging Clinical Frailty Scale; SI, Screening Instrument; SOF, Study of Osteoporotic Fractures; SPPB, Short Physical Performance Battery; SPQ, Sherbrooke Postal Questionnaire; IFQ, Inter-Frail Questionnaire; TFI, Tilburg Frailty Indicator; VES13, Vulnerable Elders Survey; WHOAFC, World Health Organization Assessment of Functional Capacity; WHRH, WHOAFC and self-reported health; ZED1, Zutphen Elderly Study (Physical Activity and Low Energy); ZED2, Zutphen Elderly Study (Physical Activity and Weight Loss); ZED3, Zutphen Elderly Study (Physical Activity and Low BMI).

DISCUSSION

We studied the cross-sectional agreement between 35 FS in an elderly population and found a wide range of agreement. Out of 595 pairs of scores, almost 90% had a κ value under 0.6. Our results, based on both traditional and modified Bland-Altman models, indicated that FS belonging to the “accumulation of deficits” model with many variables had higher median agreement (Figure 1) and narrower prediction intervals (Figure 2) and that FS belonging to the “multidimensional” model had lower absolute errors (Figure 3). Our results support our initial hypothesis that different FS classify different subsets of the population as frail.

Using the published cutoff values for each frailty score, we found very wide variation in the prevalence of frailty, as previously reported by others (12, 54–56). Scores that define solely frail and nonfrail categories generally yielded a higher frailty prevalence than scores that also define an intermediate “prefrail” state. Even though some variation is to be expected due to the fact that scores have been developed according to different underlying concepts of frailty, our finding of a 70-fold difference between the highest prevalences (Short Physical Performance Battery: 65.0% in men and 72.4% in women) and the lowest prevalences (Zutphen Elderly Study (Physical Activity and Low BMI): 0.8% in men and 1.0% in women) indicates that published estimates of frailty prevalence, and consequently our insight into the magnitude of the frailty problem, is dependent to an overwhelming degree on the chosen instrument and cutoff level. Comparisons with prevalence estimates from other populations, such as those published in 2012 in a systematic review (55), therefore need to be undertaken with caution and preferably only between studies using the same instrument.

Our findings also highlight that the general recommendation that scores and their cutoff levels be recalibrated (by modification of the weights attached to each item and/or revision of the optimal cutoff level) before being applied outside their original population is highly applicable and important in the field of frailty. When we regarded FS on a continuous scale from 0 to 1, the between-score variation was still large but less pronounced (a 2.7-fold difference in mean frailty score between the highest (0.35) and the lowest (0.13) scores). This indicates that the problem of the wide divergence in prevalence estimates is due firstly to lack of generalizability of cutoff values across different populations and secondly to different characteristics of the scores themselves. The lack of a uniform understanding of what constitutes frailty is ultimately what underlies the large number of different scores used to measure it and the resulting issues that arise when attempting to compare results.

Given the outlined issues with the use of published cutoff levels, we focused our study of agreement on identification of the 20% of individuals who are the most frail. We found that in some cases agreement was as low as 0.1 (10%), which, with a prevalence of 0.2, means that approximately 30% of individuals would be classified differently. The highest agreement (0.83 or 83%) translates to about 6% of individuals being classified differently, at the predefined prevalence of 0.2. Only 11.3% of pairs of scores had a κ value of 0.6 or higher, indicating that only a small minority of score pairs would identify the same individuals as being frail with an acceptable level of consistency. In clinical practice, these low levels of agreement would lead to the selection of largely different people for further examination or treatment, depending on which tool was implemented.

As a summary measure of agreement, κ has the disadvantage of valuing correct classification of the presence or absence of the condition in equal measure. Judgement of whether or not this is appropriate will depend on the context in which a score is used. If it is used as part of a sequence of screening steps, sensitivity is likely to be more important than specificity, while if the score is used to guide treatment initiation, specificity will be equally important. Additionally, in a research context, this measure depends on the prevalence of the condition (with a very low prevalence, κ will be very low, even with high agreement between raters) (57).

We examined agreement across the entire spectrum of frailty based on both traditional and modified Bland-Altman analyses. Traditional pairwise Bland-Altman models regard the mean of each pair of measures as an indicator of the “true” value. In our modified Bland-Altman models, we calculated the M35FS to generate a global indicator of the “true” level of frailty. Although use of the M35FS as a proxy for the “true” level of frailty involves making a number of assumptions, such as assigning equal importance to each of the studied scores, we feel that this approach best captures the agreement between each score and the global level of frailty in the absence of an accepted gold standard. The complementary pairwise analyses based on traditional Bland-Altman models largely confirmed the finding of better agreement for FS with numerous variables and lower error for FS from the multidimensional model category (Table 5).

Several scores tended to progressively under- or overestimate frailty at higher levels of “true” frailty, indicating that they would require recalibration not only of the distribution or cutoff level but also of the relative weight attached to each underlying variable to avoid producing biased frailty estimates in the ELSA population. Several scores showed remarkably wide prediction intervals, indicating a poor capacity to accurately assess the “true” level of frailty.

The narrowest prediction intervals were observed for the FS from the “accumulation of deficits” model category, which were also the FS with the best agreement across all score pairs based on κ values. In general, scores based on a larger number of variables tended to have narrower prediction intervals and higher overall agreement, but with a certain degree of underestimation in the higher frailty ranges. Scores that were multidimensional tended to have less error at the median point of frailty.

While features such as accuracy, overall agreement, and bias are important considerations guiding a choice of score for research or clinical practice, practical feasibility is likely to be as important. Although we observed the highest overall agreement between scores derived from numerous variables, these scores may be difficult to implement in practice due to the high demands on time, expertise, and equipment required to obtain a valid and complete set of the necessary data. When we categorized our results by the intended setting (clinical or community-based) in supplementary analyses, we observed similar variability in agreement both within and between these 2 kinds of settings (Web Figures 2 and 3). Which score strikes the optimal balance between feasibility and performance is likely to be different in each situation. Our results may help guide these decisions.

The main strengths of the present study are that we analyzed agreement between the most comprehensive list of FS examined to date in a large sample representative of an elderly general population, based on data including self-reported and objective measures of determinants and characteristics of frailty. We applied 3 different approaches to the study of agreement, finding broadly consistent results. In addition, we applied multiple imputation, using a state-of-the art method.

The main limitation of our study is that in the absence of an external gold standard, our analyses of agreement between the continuous scores depended on an internal proxy for “true” frailty, defined either specifically for each pair of scores or globally as the M35FS. Our adaptation of some scores to the data available in ELSA may have led to some degree of distortion in comparison with the original score definition. However, this affected only a minority of scores and is unlikely to have determined our main findings. Finally, due to the cross-sectional design of the present analysis, we cannot draw conclusions regarding dynamic features of the scores (such as longitudinal stability) or about other aspects (such as external validity) with regard to frailty outcomes.

Our comparative study of different features of agreement in a wide set of published FS showed marked heterogeneity in the degree to which various FS over-/underestimate frailty and agree on the identification of particular individuals as frail. Different scores are based on different concepts of frailty, and most pairs cannot be assumed to be interchangeable. Research results based on different scores cannot be compared, pooled, or summarized directly. Our results support a multidimensional concept of frailty that includes many variables.

Click here for additional data file.