Changes in Predictive Performance of a Frailty Index with Availability of Clinical Domains

Abstract

Background:

The deficit-accumulation frailty index (FI) measures the proportion of deficits in a multi-domain assessment. Effects of missing data in frailty identification and risk prediction are unknown.

Methods:

We analyzed the 2011 data from 6206 community-dwelling older adults in the National Health in Aging Trends Study. A 41-variable FI was constructed with the following domains: comorbidities, activities of daily living (ADLs), instrumental activities of daily living, self-reported physical limitations, physical performance, and neuropsychiatric tests. We evaluated discrimination after removing single and multiple domains, comparing C-statistics for predicting 5-year risk of mortality and 1-year risks of disability and falls.

Results:

The full FI yielded a mean of 0.18 and C-statistics of 0.72 [95% confidence interval, 0.70–0.74] for mortality, 0.80 [0.77–0.82] for disability, and 0.66 [0.64–0.68] for falls. Removal of any single domain shifted the FI distribution, resulting in a mean FI ranging from 0.13 (removing comorbidities) to 0.20 (removing ADLs) and frailty prevalence (FI≥0.25) from 16.0% to 28.7%. Among robust participants models missing ADLs misclassified most often, (19% as pre-frail). Among pre-frail and frail participants missing comorbidities misclassified most often (69.2% from pre-frail to robust, 24% from frail to pre-frail, and 4.9% from frail to robust). Removal of any single domain minimally changed C-statistics: mortality, 0.71–0.73; disability, 0.79–0.80; and falls, 0.64–0.66. Removing neuropsychiatric testing and physical performance yielded comparable C-statistics of 0.70, 0.78, and 0.66 for mortality, ADLs, and falls respectively. However, removal of 3 or 4 domains based on likely availability decreased C-statistics for mortality (0.69, 0.66), disability (0.75, 0.70), and falls (0.64, 0.63), respectively.

Conclusion:

While FI discrimination is robust to missing information in any single domain, risk prediction is affected by absence of multiple domains. This work informs the application of FI as a clinical and research tool.

INTRODUCTION

The deficit-accumulation frailty index (FI) is a commonly used measure of frailty, which is calculated as the number of health deficits present divided by the total number of deficits considered. Because the FI approach focuses on quantifying the total amount of health deficits rather than measuring a specific set of variables, an FI can be constructed from any health-related variables that are biologically sensible and increase in prevalence with age.¹ Once a critical level of cumulative deficits is reached, these deficits collectively confer vulnerability to adverse health outcomes in older adults. The FI has been associated with disability, institutionalization, and mortality in general populations² as well as treatment outcomes in patient populations. ^3–5

A major strength of the FI is flexibility of implementation and adaptation to a variety of clinical settings. Although the original FI was based off comprehensive assessment of multiple health domains and physiologic systems,⁶ FIs have been adapted with different data sources, including clinical data,⁷ electronic health records,^8,9 and claims data.¹⁰ Previous work suggests that up to 50% of deficits included in an FI can be missing without compromising mortality prediction; the missing deficit variables are simply dropped from both the numerator and denominator.¹¹ Thus an FI can still be calculated, as long as a sufficient number of variables (i.e., ≥30)¹ are captured. However this is based on the assumption that deficit variables are missing completely at random.

In clinical practice, data are often unavailable or missing. For example, in studies comparing cognitive testing to medical chart reviewed physician diagnoses, dementia was often underdiagnosed by health care professionals, with estimates of missed diagnoses ranging from 25%–90.¹² Physical performance measurements, such as gait speed and grip strength, are also not routinely measured in primary care settings.¹³ Additionally, the Medicare Annual Wellness visit does not require all components that are part of a comprehensive geriatric assessment/frailty assessment. For example, gait speed, grip strength, and detailed cognitive testing are not mandatory. Since the presence and severity of frailty may guide clinical management establishing accuracy of frailty measurements is critical for real-world applications.

Here we investigate whether varying data availability affects the distribution and predictive accuracy of a deficit-accumulation FI. We assess prediction of mortality, worsening disability, and falls with a full 41-variable FI in a contemporary representative sample of Medicare beneficiaries. We then examine the impact of missing a single domain, and missing multiple clinical domains, on estimated prevalence of frailty and discrimination of the FI.

METHODS

Study Population

The National Health in Aging Trends Study (NHATS) is a nationally representative, prospective cohort study of Medicare beneficiaries aged 65 years and older.^14,15 NHATS is sponsored by the National Institute on Aging (grant number NIA U01AG032947) through a cooperative agreement with the Johns Hopkins Bloomberg School of Public Health. Trained research staff conducted annual in-home interviews of study participants, along with a proxy if available. Informed consent was obtained. The Hebrew SeniorLife Institutional Review Board approved this study.

We included all participants from the Round 1 (2011) NHATS assessment (n=8245). Participants were excluded if they were living in a nursing home (n=468), living in residential care (n=580), lost to follow-up within 1 year without survival data (n=26), or lacked at least 40 out of 41 variables to calculate FI measured (n=965). The final sample included 6206 participants.

Baseline Measurements

Age was reported categorically (65–69, 70–71, 75–79, 80–84, 85–89, ≥90 years). Sex, race, medical comorbidities, height and weight were self-reported. Depression and anxiety were measured using the 2-item Patient Health Questionnaire (PHQ-2) (score ≥3) and the 2-item Generalized Anxiety Disorder (GAD-2) (score ≥3), respectively.^14,16 Cognitive function was assessed by a dementia algorithm developed and validated by NHATS. The assessment includes orientation (day, date, month, year, President and Vice President), memory (immediate and delayed 10 word recall), and executive function (clock drawing). We considered abnormal cognitive performance as 1.5 standard deviations below the mean in two or more domains, based on the published NHATS standards.¹⁷ Gait speed (meters/second) and grip strength (kg) were measured as the average of two attempts. Time needed to complete 5 chair stands was based on a single attempt.

Frailty Index Construction

We used the deficit accumulation approach to calculate a full 41-variable FI, using items in the NHATS dataset typically collected as part of a comprehensive geriatric assessment (Supplemental Table 1). The variables, grouped by domain, are 12 medical history items (comorbidities), 6 activities of daily living (ADLs – feeding, dressing, grooming, bed mobility, bathing, toileting), 7 instrumental activities of daily living (IADLs – using the telephone, transportation, shopping, meal preparation, housework, medication management, managing finances), 7 self-reported physical tasks (pushing/pulling heavy objects, stooping or kneeling, lifting 10 lbs., reaching arms above shoulder, handing small objects, walking up a flight of stairs, walking half a mile, heavy housework), 3 performance measures (gait speed, grip strength, chair stands), 4 neuropsychiatric measures (self-reported diagnosis of dementia, abnormal cognitive testing, positive GAD-2 screen, positive PHQ-2 screen), and 2 nutritional measures (body mass index<21 kg/m², unintentional weight loss >10 lbs. in the past year).

Each item was scored as 1 if a deficit or condition was present, or 0 if absent (ie. if a participant had a disease they received a point, if they were unable to complete a task they received a point [Supplemental Table 1]). Partial points were given for gait speed, grip strength, chair stands, and cognitive tests. Inability to complete gait speed and chair stand for reasons other than space constraints was scored as a deficit.

We first calculated the full FI as the sum of deficits present, divided by the 41 items considered. We first removed individual domains and calculated the FI based on fewer variables. We then removed combinations of 2 or more domains based on the likelihood of having available data in clinical care settings, supported by literature¹⁸ and missing data frequency within the dataset itself (Supplemental Table 1). Performance tests were considered most likely to be missing, followed by in order, cognition, self-reported physical tasks, and IADLs. Of note, because FI by convention must contain at least 30 variables we avoided dropping more than that. This process yielded a total of 11 modified FIs in addition to the full FI.

Outcomes

The primary outcome was all-cause mortality at 5 years. Month and year of death were obtained from caregivers. Participants were followed until the earliest of last known follow-up interview, death, or 5 years from the baseline interview. As secondary outcomes, we assessed worsening ADL disability, defined as requiring assistance in more ADLs at 1 year compared to baseline, excluding those with maximal disability at baseline (n=28), and self-reported falls at 1 year among participants who were alive with complete first-year follow-up interview (n=5064).

Statistical Analysis

We describe the distribution of each FI according to its mean, median, interquartile range and total range. Calibration was assessed by comparing the mean of each FI to the full FI, and the proportion of the population that would be classified as frail was calculated based on a standard cutoff of ≥0.25).¹⁹ We repeated these analyses stratified by frailty status based on full FI (robust FI<0.15, pre-frail 0.15–0.24, frail ≥0.25), and assessed the proportion misclassified in each situation. For each index, we used Cox proportion hazards regression to model 5-year mortality and logistic regression to model 1-year ADL disability and falls. We used the somersd Stata package to calculate C-statistics with 95% confidence intervals (CIs) for each model.²⁰ This is a common rank order measure of model discrimination that represents the probability that a randomly selected participant will experience an outcome earlier than another patient that has a lower FI. A c-statistic ranges from 0.5 (random concordance) to 1 (perfect concordance).²¹ All analyses were performed using Stata survey procedures (version 15.1, StataCorp, Texas) to account for the complex sampling design of NHATS, and findings were weighted to reflect national estimates of the Medicare population.

RESULTS

Characteristics of study population

The study population included 3519 participants (42.2%) with age ≥75 years, 3512 (55.2%) women, and 4375 (82.3%) non-Hispanic white race (Table 1). The most prevalent self-reported comorbidity was hypertension (n=4143, 63.5%), and the least common was dementia (n=215, 2.5%); 582 (7.3%) reported ADL disability and 1724 (22.2%) reported IADL disability; cognitive impairment was present in 1211 (14.9%). Among those who completed performance measures, mean gait speed was 0.8 m/s, and mean time to complete 5 chair stands was 12.1 seconds. Those excluded tended to be frailer with more comorbidities (Supplemental Table 2). Among the entire sample performance tests had the most missing data (gait speed 14.5%, grip strength 15.7%, chair stands 25.6%) followed by cognitive testing (2.4%). The most complete data was that on comorbidities, with missing data ranging from 0–1.2% (Supplemental Table 2).

Table 1:

Characteristics of National Health in Aging Trends Study baseline 2011

Characteristic	n (%)
Age 75 years or older	3519 (42.2)
Female	3512 (55.2)
Race
Non-Hispanic White	4375 (82.3)
Black	1295 (7.8)
Other	536 (10.0)
Self-report comorbid conditions
Hypertension	4143 (63.5)
Vision impairment	3831 (61.5)
Arthritis	3350 (52.4)
Spine disease	2465 (40.1)
Diabetes	1536 (23.0)
Heart disease	1084 (16.6)
Cancer	1041 (16.0)
Lung disease	952 (15.4)
Heart attack	892 (13.2)
Hearing impairment	783 (11.3)
Stroke	633 (8.8)
Dementia	215 (2.5)
ADL disability	582 (7.3)
IADL disability	1724 (22.2)
Abnormal cognitive testing a	1211 (14.9)
PHQ-2 score ≥3	850 (12.7)
GAD-2 score ≥3	714 (10.8)
Unable to carry 10 pounds	1093 (13.4)
Unable to climb 10 steps	1143 (14.1)
Unable to walk 6 blocks	2179 (29.1)
Gait speed (mean ± SD in meters/sec)c	0.8 ± 0.5
Grip strength (mean ± SD in kg)	24.3 ± 12.5
Time to complete 5 chair stands (mean ± SD in sec)d	12.1 ± 4.3
assessment (n=6206)

Note. n indicates sample size; weighed percentages reflect national estimates. Abbreviations: ADL, Activities of daily living; FI, frailty index; GAD, General Anxiety Disorder 2-item; IADL, instrumental activities of daily living; PHQ-2 Patient Health Questionnaire-2

^aCognitive impairment defined by performance below 1.5 standard deviations from population mean in any cognitive domain (orientation, memory, executive function).

^bMobility impairment defined by inability to walk a block, or climb a flight of stairs.

^c500 participants unable to complete gait speed

^d1339 participants unable to complete chair stand

Removal of a single domain

The mean of the full FI was 0.18, and 21.5% of the population was considered frail, defined by an FI≥0.25 (Table 2). With removal of a single domain from the full FI, the number of the remaining variables ranged from 29 (removal of comorbidities) to 39 (removal of nutrition domain). The mean FI varied from 0.13 (when comorbidities were omitted) to 0.20 (when ADLs were omitted). The prevalence of frailty ranged from 16.0% (when comorbidities were omitted), to 28.7% (when ADLs were omitted). Stratified by frailty status, among robust participants the models missing ADLs misclassified more, (19% as pre-frail) (Figure 1, Supplemental Table 3). Among pre-frail and frail participants models missing comorbidities misclassified the most (69.2% from pre-frail to robust, 24% from frail to pre-frail, and 4.9% from frail to robust) (Supplemental Tables 4 and 5).

Table 2:

Characteristics of frailty index versions after removing various clinical domains (n=6206)

FI Version	Number of Variables	Percent with FI ≥0.25	Mean [95% Confidence Interval]	Median FI (Interquartile Range)
Full FI	41	21.5%	0.18 [0.17–0.18]	0.15 (0.09–0.23)
1 Domain Omitted
- Comorbidities	29	16.0%	0.13 [0.13–0.14]	0.08 (0.03–0.17)
- Performance	38	17.9%	0.16 [0.16–0.17]	0.13 (0.08–0.21)
- Cognition	37	23.7%	0.19 [0.18–0.19]	0.16 (0.10–0.24)
- Physical tasks	34	19.3%	0.17 [0.17–0.18]	0.15 (0.10–0.22)
- IADLs	34	28.2%	0.20 [0.19–0.20]	0.17 (0.10–0.26)
- ADLs	35	28.7%	0.20 [0.20–0.20]	0.17 (0.11–0.27)
- Nutrition	39	22.4%	0.18 [0.18–0.19]	0.15 (0.09–0.24)
2 Domains Omitted
- Performance and Cognition	34	19.5%	0.17 [0.17–0.17]	0.15 (0.09–0.21)
- Performance and Physical tasks	31	14.9%	0.15 [0.15–0.16]	0.13 (0.10–0.19)
3 Domains Omitted
- Performance, Cognition, and Physical tasks	27	17.7%	0.16 [0.16–0.17]	0.15 (0.07–0.22)
4 Domains Omitted
- Performance, Cognition, Physical tasks, and IADLs	20	33.7%	0.19 [0.19–0.20]	0.20 (0.10–0.25)

Note. Descriptive characteristics of frailty index versions after omitting single domains, and then multiple clinical domains of information based on likely patterns of data availability in clinical practice. Weighted percentages reflect national estimates. Abbreviations: ADL, activity of daily living; FI, frailty index; IADL, instrumental activity of daily living.

Figure 1. Misclassification of frailty with variable omission of clinical domains from a frailty index.

Abbreviations: ADLs – Activities of Daily Living, FI – Frailty Index, IADLs – Instrumental Activities of Daily Living. Misclassification of frailty status stratified by frailty level based on complete frailty index measurement. Omission of domains led to more misclassification among pre-frail and frail participants than robust participants.

The C-statistic for prediction of 5-year mortality with a full 41-variable FI was 0.72 [95% CI, 0.70–0.74] (Table 3). Omission of any single domain yielded C-statistics for 5-year mortality prediction ranging from 0.71 [0.69–0.72] with removal of performance domain, to 0.73 [0.71–0.75] with removal of comorbidities.

Table 3.

Discrimination for 5-year risk of mortality and 1-year risk of worsening disability and falls after removing clinical domains (n=6206)

FI Version	Number of Variables	C-statistic [95% CI] for 5-year Mortality (n=6206)	C-statistic [95% CI] for 1-year ADL Disability (n=5036)	C-statistic [95% CI] for 1-year Fall Incidence (n=5064)
Full FI	41	0.72 [0.70–0.74]	0.80 [0.77–0.82]	0.66 [0.64–0.68]
1 Domain Omitted
- Comorbidities	29	0.73 [0.71–0.75]	0.79 [0.76–0.82]	0.64 [0.62–0.66]
- Performance	38	0.71 [0.69–0.72]	0.79 [0.76–0.81]	0.66 [0.64–0.67]
- Cognition	37	0.72 [0.70–0.74]	0.79 [0.76–0.82]	0.66 [0.64–0.68]
- Physical tasks	34	0.72 [0.70–0.74]	0.79 [0.78–0.81]	0.65 [0.63–0.67]
- IADLs	34	0.72 [0.70–0.74]	0.79 [0.76–0.82]	0.66 [0.64–0.68]
- ADLs	35	0.72 [0.70–0.74]	0.80 [0.78–0.82]	0.66 [0.64–0.68]
- Nutrition	39	0.72 [0.70–0.73]	0.80 [0.77–0.83]	0.66 [0.64–0.68]
2 Domains Omitted
- Performance and Cognition	34	0.70 [0.68–0.72]	0.78 [0.75–0.81]	0.66 [0.64–0.67]
- Performance and Physical Tasks	31	0.69 [0.67–0.72]	0.76 [0.73–0.79]	0.64 [0.63–0.66]
3 Domains Omitted
- Performance, Cognition, and Physical Tasks	27	0.69 [0.67–0.71]	0.75 [0.72–0.78]	0.64 [0.62–0.66]
4 Domains Omitted
- Performance, Cognition, Physical Tasks, and IADLs	20	0.66 [0.64–0.68]	0.70 [0.67–0.73]	0.63 [0.61–0.65]

Note. Abbreviations: ADL, activity of daily living; FI, frailty index; CI, confidence interval; IADL, instrumental activity of daily living.

The C-statistics of 1-year worsening ADL disability and falls with a full 41-variable FI were 0.80 [0.77–0.85] and 0.66 [0.64–0.68], respectively. Removal of any single domain yielded minimal change in C-statistics ranging 0.79–0.80 for worsening disability and 0.64–0.66 for falls.

Removal of multiple clinical domains

Compared to 21.5% prevalence of frailty with a 41-variable FI, removal of performance and cognition domains resulted in a mean FI of 0.17 in the overall population, and classified 19.5% of the population as frail (Table 2). While only 1.1% of robust participants were misclassified, 30.9% and 13.3% of pre-frail and frail were misclassified, respectively (Figure 1). This model had a C-statistic of 0.70 [0.68–0.72] for 5-year mortality prediction, 0.78 [0.75–0.81] for 1-year worsening ADL disability, and 0.66 [0.64–0.67] for 1-year falls (Table 3). Omitting performance measures and self-reported physical tasks resulted in a mean FI of 0.15 and identified only 14.9% of the population as frail. 7.1% of robust participants were misclassified, compared to 35.6% of frail participants (Figure 1). This model had a C-statistic of 0.69 [0.67–0.72] for mortality, 0.76 [0.73–0.79] for ADL disability, and 0.64 [0.63–0.66] for falls.

After removal of 3 domains (performance, cognition, and physical tasks), mean FI was 0.16, and 17.7% of the population was identified as frail. Similarly less robust were misclassified than pre-frail or frail (5.9%, 50.9%, 31.3%, for robust, pre-frail and frail respectively). The C-statistic for 5-year mortality prediction after omitting 3 domains was 0.69 [0.67–0.71]. We observed a similar pattern for 1-year worsening ADL disability, where removal of 3 domains resulted in a C-statistic of 0.75 [0.72–0.78]. This pattern was less apparent for falls (0.64 [0.62–0.66]).

After removal of 4 domains (IADLs, performance, cognition, and physical tasks), mean FI was 0.19 and 33.7% of the population was frail. Misclassification of 49% of robust, 57.8% of pre-frail, and 19.4% of frail participants occurred. The C-statistic decreased further to 0.66 [0.64–0.68] for mortality, 0.70 [0.67–0.73] for worsening ADL disability, and 0.63 [0.61–0.65] for falls.

DISCUSSION

In this study, we simulated the effects of missing domains on reliability and discrimination of a FI, based on likely availability of clinical information in routine practice. We found that removal of a single clinical domain changes the prevalence of frailty and distribution of FI, thus leading to misclassification, most apparent in pre-frail and frail populations. These shifts however did not compromise discrimination for mortality, worsening ADL disability, or falls. Removal of the two domains resulted in relatively preserved discrimination; however removal of further domains progressively compromised discrimination of adverse outcomes with increasing misclassification of frailty status. This pattern was most apparent for mortality.

Our study provides insight into the performance ability of FI depending on how the measure is operationalized, and the components that may affect identification of frailty. We found that the prevalence of frailty estimated using the full FI (21.5%) is comparable to previously published estimates using FI in a community-dwelling population (22.7%).² However, depending on the definition of frailty used, frailty prevalence among community-dwelling older persons in various studies ranges from 4.0–59.1%.²² Even comparing different frailty measures within the same cohort, Theou and colleagues reported a prevalence ranging from 6.1–43.9%.²³ A previous NHATS study that used the frailty phenotype²⁴ estimated the prevalence to be 15.3% nationally. Interestingly, this estimate approximates the frailty prevalence using the FI that excluded comorbidities in our study (16.0%). Our study sheds light on how variation in the frailty prevalence may be a reflection of the clinical domains represented in differing frailty scales.

Although the original FI was initially developed from 70 variables,²⁵ subsequent work supported that this could be reduced to approximately 30–40 variables.² Further studies demonstrated that completely random omission up to 50% of variables from a 40-variable FI produced comparable discrimination of mortality to the full FI.^11,26 However in clinical practice, variables are unlikely to be missing completely at random. Complete frailty assessments can be time consuming,¹³ with assessment of physical performance frailty criteria, potentially taking 15–20 minutes alone, the majority of time for a typical provider visit.²⁷ Comprehensive cognitive tests can also be cumbersome, often taking 5–10 minutes,¹² and thus cognitive impairments may be under recognized. Consistent with this, a study in NHATS linked to Medicare claims found that among older adults with probable dementia, 39.5% were undiagnosed.²⁸ In our study, removal of cognition and physical performance tests did not result in significant compromise to FI discrimination. However, missing multiple domains may have ramifications on not only the distribution of FI values, but also the discrimination of an FI.

Removal of multiple domains was guided by clinical experience on which variables are least likely to be captured in routine care, based on requirements for annual Medicare wellness visits,¹⁸ missing data patterns in the NHATS dataset, published challenges in obtaining measurements, and input from colleagues at a national conference. Performance tests were considered most likely to be missing in routine clinical practice, followed by in order, cognitive testing, self-reported physical tasks, and IADLs. However, in electronic medical records or databases, cognitive impairments may be captured but questions regarding self-reported physical tasks, such as walking several blocks or lifting heavy objects, may not. Our results suggest that absence of performance variables may be compensated by self-reported physical task information, which captures physical fitness beyond ADLs and IADLs. Complete omission of performance tests and self-reported physical task information can alter predictive ability and identification of vulnerable older adults.

While discrimination was relatively preserved, removal of a single domain affected absolute FI value and frailty status, raising concerns for applying strict cutoffs in clinical practice. We observed the most misclassification with omission of comorbidities, IADLs, and ADLs. Interestingly the amount of misclassification and impact of domain omission varied by frailty status as well. In our study omission of comorbidities, performance measures and physical tasks tended to underestimate the FI, while omission of cognition, IADLs and ADLs tended to overestimate the FI. It may be that on the spectrum of deficits, IADL and ADL disability is accumulated relatively later, and are thus later markers of frailty compared to subtler changes in performance measures or comorbidities. Thus excluding functional disabilities highlights deficits from comorbidities among robust and pre-frail populations, while masking deficits among frail populations.

These findings have deep implications for application of the FI as a screening or diagnostic test. For example, an electronic health record^8,29 or insurance claims-based FI³⁰ may rely heavily on comorbidities while lacking data on functional or cognitive status, and thus may overestimate the FI. Because calibration and misclassification occurs with missingness, FI cutoffs should be implemented with careful attention to specific clinical contexts and potential ramifications. Ultimately these measures may need to be calibrated to an FI derived from comprehensive multi-domain assessment, or locally adjusted and contextualized to account for availability of domains.

Our study is not without limitations. Our study population was restricted to community-dwelling older Medicare beneficiaries, thus our findings cannot be generalized to those who are institutionalized or lack Medicare coverage. Those living in residential housing are likely to be frailer, with higher FI, with poorer model prediction. We restricted analysis to a population with at least 40 out of 41 FI variables measured at baseline; thus dropping domains from the FI would not allow for the inclusion of new participants. This ensures that variation in frailty prevalence and predictive performance of indices was due to the changes in FI variables, and not changes in the population analyzed. While it is true that fewer variables will inevitably lead to compromised discrimination to some degree, the effects observed are unlikely purely due to loss of information. In fact, omitting comorbidities resulted in the fewest number of variables, yet preserved discrimination. The outcomes of worsening ADL disability and falls were self-reported, and thus limited to those who were alive and completed annual follow-up; a population likely to be less frail. Thus these outcomes are likely underestimated. We recognize discrimination is only one aspect of model performance, and calibration was not formally tested. In fact the shifts in FI distributions imply that there are changes to calibration, and the potential for misclassification. Thus deficit composition may still affect overall model prediction. Given modest c-statistics of 0.72, the prognostic ability for even a full FI is likely inadequate in isolation. However this is comparable to many other prognostic models developed to predict mortality.^31,32 Furthermore it remains a useful measure to predict vulnerability to mortality and other adverse health outcomes such as disability and healthcare utilization.

In a nationally-representative sample of community-dwelling older adults in the United States, we demonstrate that FI performance is robust to the omission of a single clinical domain. While the absence of two domains most likely to be missing yielded reasonable discrimination, further omission of domains compromised FI performance. Furthermore, prevalence of frailty changes drastically depending the specific domain that is missing. Our findings help inform the development of FIs for translation into a variety of clinical and research settings.

Supplementary Material

Supplementary Material

Supplemental Table 1: Frailty index variable construction and details

Supplementary Table 2. Descriptive characteristics comparing those excluded due to <40 variables

Supplemental Table 3: Distribution of frailty index and frailty misclassification among robust participants

Supplemental Table 4: Distribution of frailty index and frailty misclassification among pre-frail participants

Supplemental Table 5: Distribution of frailty index and frailty misclassification among frail participants