Development and External Validation of Models to Predict Need for Nursing Home Level of Care in Community-Dwelling Older Adults With Dementia

W James Deardorff; Sun Y Jeon; Deborah E Barnes; W John Boscardin; Kenneth M Langa; Kenneth E Covinsky; Susan L Mitchell; Sei J Lee; Alexander K Smith

doi:10.1001/jamainternmed.2023.6548

. 2023 Dec 4;184(1):81–91. doi: 10.1001/jamainternmed.2023.6548

Development and External Validation of Models to Predict Need for Nursing Home Level of Care in Community-Dwelling Older Adults With Dementia

W James Deardorff ^1,^2,^✉, Sun Y Jeon ^1,², Deborah E Barnes ^3,⁴, W John Boscardin ^1,^2,³, Kenneth M Langa ^5,^6,^7,⁸, Kenneth E Covinsky ^1,⁹, Susan L Mitchell ^10,¹¹, Sei J Lee ^1,², Alexander K Smith ^1,²

¹Division of Geriatrics, Department of Medicine, University of California, San Francisco

²Geriatrics, Palliative and Extended Care Service Line, San Francisco Veterans Affairs Health Care System, San Francisco, California

³Department of Epidemiology and Biostatistics, University of California, San Francisco

⁴Department of Psychiatry and Behavioral Sciences, University of California, San Francisco

⁵Department of Internal Medicine, School of Medicine, University of Michigan, Ann Arbor

⁶Institute for Healthcare Policy and Innovation, University of Michigan, Ann Arbor

⁷Veterans Affairs Ann Arbor Center for Clinical Management Research, Ann Arbor, Michigan

⁸Institute for Social Research, University of Michigan, Ann Arbor

⁹Associate Editor, JAMA Internal Medicine

¹⁰Hinda and Arthur Marcus Institute for Aging Research, Hebrew SeniorLife, Boston, Massachusetts

¹¹Department of Medicine, Beth Israel Deaconess Medical Center, Boston, Massachusetts

Accepted for Publication: October 9, 2023.

Published Online: December 4, 2023. doi:10.1001/jamainternmed.2023.6548

^✉

Corresponding Author: W. James Deardorff, MD, Division of Geriatrics, Department of Medicine, University of California, San Francisco, 490 Illinois St, Floor 8, San Francisco, CA 94158 (william.deardorff@ucsf.edu).

Author Contributions: Dr Deardorff had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Drs Lee and Smith contributed equally as co–senior authors.

Concept and design: Deardorff, Barnes, Covinsky, Mitchell, Lee, Smith.

Acquisition, analysis, or interpretation of data: Deardorff, Jeon, Barnes, Boscardin, Langa, Mitchell, Lee, Smith.

Drafting of the manuscript: Deardorff.

Critical review of the manuscript for important intellectual content: All authors.

Statistical analysis: Deardorff, Jeon, Boscardin.

Obtained funding: Lee, Smith.

Administrative, technical, or material support: Covinsky, Lee.

Supervision: Boscardin, Lee, Smith.

Conflict of Interest Disclosures: Dr Barnes is a co-founder, chief science adviser, and stockholder in Together Senior Health, which offers evidence-based, online programs for people living with dementia and care partners. Dr Langa reported grants from the National Institutes of Health/National Institute on Aging and the Alzheimer’s Association outside the submitted work. No other disclosures were reported.

Funding/Support: This work was supported by grants from the National Institute on Aging (T32-AG000212 and R03AG082859 to Dr Deardorff; R01AG057751 to Drs Smith, Lee, and Barnes; K24AG068312 to Dr Smith; K24AG066998 to Dr Lee; P01AG066605 to Drs Covinsky, Smith, Lee, Jeon, and Boscardin; and P30AG044281 to Drs Covinsky, Deardorff, Smith, Lee, Jeon, and Boscardin). Dr Deardorff was also supported by the National Center for Advancing Translational Sciences of the National Institutes of Health under award number KL2TR001870. The Health and Retirement Study is funded by National Institute on Aging grant U01AG009740 and performed at the Institute for Social Research at the University of Michigan. The National Health and Aging Trends Study was produced and distributed by www.nhats.org, with funding from the National Institute on Aging (grant number U01AG32947).

Role of the Funder/Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Disclaimer: Dr Covinsky is an Associate Editor of JAMA Internal Medicine but was not involved in any of the decisions regarding review of the manuscript or its acceptance. The content of this article is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Data Sharing Statement: See Supplement 2.

^✉

Corresponding author.

PMCID: PMC10696518 PMID: 38048097

This prognostic study leverages 2 nationally representative prospective cohorts, the Health and Retirement Study and the National Health and Aging Trends Study, to develop and validate models to predict need for nursing home level of care among older adults with probable dementia.

Key Points

Question

Can need for nursing home level of care (NHLOC) in community-dwelling older adults with dementia be accurately predicted?

Findings

This prognostic study used 2 nationally representative cohorts of community-dwelling older adults with probable dementia (from 1998-2016 and 2011-2019) to develop and perform an external validation of 2 models to predict need for NHLOC based on self-report and proxy responses. The final models included readily available clinical predictors (eg, demographics, health factors, functional measures) to predict need for NHLOC with moderate discrimination and excellent calibration.

Meaning

Predictions using these models for community-dwelling older adults with dementia may help inform conversations between clinicians, patients, and families related to future planning and care management.

Abstract

Importance

Most older adults living with dementia ultimately need nursing home level of care (NHLOC).

Objective

To develop models to predict need for NHLOC among older adults with probable dementia using self-report and proxy reports to aid patients and family with planning and care management.

Design, Setting, and Participants

This prognostic study included data from 1998 to 2016 from the Health and Retirement Study (development cohort) and from 2011 to 2019 from the National Health and Aging Trends Study (validation cohort). Participants were community-dwelling adults 65 years and older with probable dementia. Data analysis was conducted between January 2022 and October 2023.

Exposures

Candidate predictors included demographics, behavioral/health factors, functional measures, and chronic conditions.

Main Outcomes and Measures

The primary outcome was need for NHLOC defined as (1) 3 or more activities of daily living (ADL) dependencies, (2) 2 or more ADL dependencies and presence of wandering/need for supervision, or (3) needing help with eating. A Weibull survival model incorporating interval censoring and competing risk of death was used. Imputation-stable variable selection was used to develop 2 models: one using proxy responses and another using self-responses. Model performance was assessed by discrimination (integrated area under the receiver operating characteristic curve [iAUC]) and calibration (calibration plots).

Results

Of 3327 participants with probable dementia in the Health and Retirement Study, the mean (SD) age was 82.4 (7.4) years and 2301 (survey-weighted 70%) were female. At the end of follow-up, 2107 participants (63.3%) were classified as needing NHLOC. Predictors for both final models included age, baseline ADL and instrumental ADL dependencies, and driving status. The proxy model added body mass index and falls history. The self-respondent model added female sex, incontinence, and date recall. Optimism-corrected iAUC after bootstrap internal validation was 0.72 (95% CI, 0.70-0.75) in the proxy model and 0.64 (95% CI, 0.62-0.66) in the self-respondent model. On external validation in the National Health and Aging Trends Study (n = 1712), iAUC in the proxy and self-respondent models was 0.66 (95% CI, 0.61-0.70) and 0.64 (95% CI, 0.62-0.67), respectively. There was excellent calibration across the range of predicted risk.

Conclusions and Relevance

This prognostic study showed that relatively simple models using self-report or proxy responses can predict need for NHLOC in community-dwelling older adults with probable dementia with moderate discrimination and excellent calibration. These estimates may help guide discussions with patients and families in future care planning.

Introduction

Dementia is one of the leading indications for nursing home (NH) placement, with an estimated 50% of NH residents having a dementia diagnosis.¹ Among individuals with dementia diagnosed at age 70 years, NH admission would be expected for approximately 75% of surviving individuals within 10 years compared with only 4% of the general population.² High NH admission rates are often related to functional impairments, medical comorbidities, and behavioral symptoms (eg, wandering, hallucinations), which make home care challenging.^3,4,5

Deciding when to enter an NH is complicated and often influenced by sociocultural and systems/environmental factors. Individuals with similar levels of functional impairment may have dramatically different NH admission times based on cultural beliefs, financial considerations, presence of family and caregiver support, occurrence of crisis events, and systems factors (eg, differential access to in-home services and federal/state benefits).^{6,7,8,9,10,11,12,13} Given the complexity surrounding this decision, providing an estimate of when an individual with dementia might be expected to need NH level of care (NHLOC) based on functional impairments and/or behavioral issues is important to guide future planning.

Several prediction models among older adults with and without dementia have been developed to predict either time until NH admission or onset of functional impairment, such as dependence in activities of daily living (ADL).^{14,15,16,17,18,19,20,21,22,23,24,25,26} Previous models are often limited by small and outdated samples, not incorporating competing risk of death, or not focused on individuals with dementia. These models are often at high risk of bias due to poor analytic methods or lack of reporting (eg, not providing information about calibration). When reported, model performance measures were typically moderate, with C statistics in the range of 0.6 to 0.7 and adequate calibration. Previous models that predict NH admission may be problematic, as this outcome differs greatly across socioeconomic and cultural groups of older adults with similar levels of disability. Many individuals with dementia remain at home despite needing NHLOC. Providing estimates of time to needing NHLOC rather than time to NH admission is more relevant across these groups as an indication of high-level care needs, whether provided in a NH or at home. Therefore, our objective was to leverage 2 nationally representative prospective cohorts, the Health and Retirement Study (HRS) and the National Health and Aging Trends Study (NHATS), to develop and validate models to predict need for NHLOC among older adults with probable dementia.

Methods

Study Design and Data Sources

For the development cohort, we used data between 1998 and 2016 from HRS, a nationally representative survey of US adults in which participants are interviewed every 2 years.²⁷ We included community-dwelling adults 65 years and older with probable dementia ascertained by a validated algorithm (eMethods 1 in Supplement 1).²⁸ We created 2 models based on whether interview responses were from the individual (self-respondent model) or a proxy, typically a spouse or family member (proxy model). Proxy interviews were generally conducted for participants who could not participate in the interview (eg, physical/cognitive reasons) (eMethods 1 in Supplement 1). Certain predictors and their strengths of association with need for NHLOC are likely to vary based on dementia severity for which proxy reporting is a surrogate. While proxy reports can differ from what patients may report if able, our approach to create 2 models mirrors the clinical practice of asking proxies for information to corroborate patient self-reports or when patients are unable to respond. Individuals were assigned to the self-respondent or proxy model based on how the interview was conducted at the interview wave when the individual was first classified with dementia.

For external validation, we used data between 2011 and 2019 from NHATS, a nationally representative study of Medicare beneficiaries aged 65 years and older.²⁹ We included participants with probable dementia based on validated cutoffs and classified them into the proxy and self-respondent models in the same manner (eMethods 1 in Supplement 1).³⁰

The study was reviewed and approved with a waiver for informed consent (due to publicly available and deidentified data) by the University of California, San Francisco Committee on Human Research. The study design and reporting were guided by the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guidelines (eTable 1 in Supplement 1).

Outcomes

We chose need for NHLOC rather than NH admission as the primary outcome because we believed it represented a more relevant outcome for individuals from different cultural and socioeconomic backgrounds. There are several ways to define need for NHLOC (eMethods 2 in Supplement 1). Based on previous recommendations,^31,32 we chose a definition that combined functional impairments and behavioral issues. Using functional impairment alone would sometimes leave out individuals with little functional impairment but behavioral problems that posed substantial safety risks to themselves or others.³³ The outcome was a composite of either: (1) 3 or more ADL dependencies (eg, bathing/showering, getting in/out of bed, dressing, toileting, walking across the room), (2) 2 or more ADL dependencies and proxy report that the individual wanders or cannot be left alone, or (3) eating dependency (eg, needing help cutting up food). In general, dependency with an ADL was defined by needing help with the task. Participants classified as needing NHLOC at baseline were excluded. The primary outcome was ascertained from subsequent interviews. For participants who died, postdeath interviews with next of kin were used to ascertain NHLOC need in the months prior to death (eMethods 2 in Supplement 1).

Candidate Predictors

Candidate predictors were identified by reviewing published prediction models and meta-analyses.^13,14,15,18 Candidate predictors were collected at the time of the interview when individuals were first classified with dementia (detailed definitions are provided in eTables 2-4 in Supplement 1). Briefly, predictors included demographics (eg, age, sex), health/behavioral factors (eg, body mass index [BMI], smoking status), comorbidities, and functional measures (eg, baseline ADL and instrumental ADL [IADL] difficulties/dependencies). Baseline ADL/IADL dependency scores were obtained by summing the number of dependencies in ADL (eg, bathing/showering, getting in/out of bed, dressing, toileting) and IADL (eg, using a telephone, preparing a meal, managing medications, managing money, shopping for groceries). Candidate predictors unique to the self-respondent model included calendar date recall (day of week, month, year) and questions about mood.

Model Development

We used a parametric survival model with Weibull distribution to model time to NHLOC need. We incorporated competing risks and interval censoring (eMethods 3, eFigure 1, and eTable 5 in Supplement 1).³⁴ We developed 2 separate models: one incorporating proxy responses (proxy model) and another using self-report (self-respondent model). To create parsimonious models, we performed variable selection using backward selection combined with multiple imputation.^35,36,37,38 We first created 20 imputed data sets under the missing at random assumption. We performed backward selection with P < .05 for elimination in each of the imputed data sets. Variables selected in more than 60% of the 20 imputed data sets (≥12 times) were considered. We explored including certain predictors such as age, BMI, and baseline ADL/IADL difficulties/dependencies as categorical or continuous (modeled linearly or with splines). We incorporated HRS survey weights and strata at each stage. Final coefficients were obtained by pooling results across imputations.

Model Evaluation

Model performance was assessed through discrimination and calibration (eMethods 4 in Supplement 1). For discrimination, we calculated the integrated area under the receiver operating characteristic curve (iAUC), which averages all available AUC values over time. We calculated time-specific AUC values at 2, 5, and 10 years. Calibration, which refers to the agreement between observed and predicted probability of need for NHLOC, was assessed graphically at fixed time points. We assessed internal validity via bootstrapping to quantify model performance optimism. We repeated the entire modeling process by performing 20 imputations followed by drawing 30 bootstrapped samples in each imputed data set. Individual iAUC values following bootstrapping were averaged to obtain optimism-corrected iAUC values. We conducted an external validation using NHATS. Because we only had follow-up in NHATS from 2011 to 2019, we did not examine model performance at 10 years. As sensitivity analyses, we explored different modeling strategies (eg, interval-censored Cox model, Fine-Gray competing risk model, alternative approaches to handling competing event cases, inclusion of social determinants of health variables) and examined model performance across subgroups by self- or proxy-reported race and ethnicity (including Hispanic, non-Hispanic Black, non-Hispanic White, and other, which encompassed American Indian, Alaska Native, Asian, Pacific Islander, and other race and was grouped together owing to small sample sizes) (eMethods 3 in Supplement 1).

Statistical Analysis

Final models will be made available on ePrognosis (https://eprognosis.ucsf.edu/nhloc.php) for users to obtain need for NHLOC estimates.³⁹ Statistical analyses were performed using SAS, version 9.4 (SAS Institute Inc); Stata, version 17.0 (StataCorp); and R, version 4.3.0 (R Project for Statistical Computing). A 2-sided P < .05 was considered statistically significant.

Results

Cohort Characteristics

Of the 43 398 participants 65 years and older in the HRS between 1998 and 2016, 3327 participants with probable dementia not needing NHLOC at baseline were included (2312 in the self-respondent model and 1015 in the proxy model) (eFigure 2 in Supplement 1). Mean (SD) age was 82.4 (7.4) years (Table 1). The proxy cohort was younger, with more baseline ADL/IADL difficulties/dependencies. The NHATS cohort included 1712 participants with probable dementia (eFigure 3 in Supplement 1). Compared with HRS, participants in NHATS were younger and generally reported fewer comorbidities (Table 1).

Table 1. Baseline Characteristics of Participants.

Characteristic	No. (survey-weighted %)^a
	Health and Retirement Study, 1998-2016 (development cohort)			National Health and Aging Trends Study, 2011-2019 (validation cohort)
	Total cohort (n = 3327)	Self-respondent model (n = 2312)	Proxy model (n = 1015)	Total cohort (n = 1712)	Self-respondent model (n = 1327)	Proxy model (n = 385)
Classified as NHLOC during follow-up	2107 (63.3)	1424 (61.6)	683 (67.3)	806 (47.1)	592 (44.6)	214 (55.6)
Age at baseline, mean (SD), y	82.4 (7.4)	83.3 (7.0)	80.2 (7.8)	NR^b	NR^b	NR^b
Age category, y
65-69	207 (5.8)	104 (3.8)	103 (10.3)	86 (9.4)	66 (8.9)	20 (11.1)
70-74	302 (8.7)	149 (5.9)	153 (15.1)	178 (16.0)	138 (15.6)	40 (17.7)
75-79	598 (17.5)	383 (15.8)	215 (21.3)	321 (23.0)	262 (24.2)	59 (18.8)
80-84	856 (26.1)	618 (27.5)	238 (22.9)	434 (22.8)	354 (24.1)	80 (18.0)
85-89	803 (24.5)	614 (27.3)	189 (18.2)	388 (18.0)	298 (17.7)	90 (19.1)
≥90	561 (17.5)	444 (19.7)	117 (12.2)	305 (10.8)	209 (9.5)	96 (15.4)
Sex
Female	2301 (69.6)	1611 (70.5)	690 (67.8)	985 (54.1)	764 (54.8)	221 (51.5)
Male	1026 (30.4)	701 (29.5)	325 (32.2)	727 (45.9)	563 (45.2)	164 (48.5)
Race and ethnicity^c
Hispanic	394 (9.7)	278 (9.9)	116 (9.4)	64 (5.2)	37 (3.8)	27 (10.9)
Non-Hispanic Black	600 (11.9)	448 (12.7)	152 (10.1)	503 (12.6)	386 (12.5)	117 (13.1)
Non-Hispanic White	2256 (76)	1549 (75.7)	707 (76.7)	932 (68.4)	734 (69.8)	198 (62.4)
Other^d	77 (2.4)	37 (1.8)	40 (3.8)	175 (12.5)	139 (12.3)	36 (13.3)
Missing	0	0	0	38 (1.3)	31 (1.5)	7 (0.3)
BMI
<18.5	209 (6.5)	133 (5.9)	76 (7.8)	85 (4.3)	52 (3.2)	33 (8.4)
18.5 to <25	1579 (48.6)	1097 (48.6)	482 (48.7)	654 (37.0)	502 (35.6)	152 (42.4)
25 to <30	993 (29.2)	716 (30.3)	277 (26.6)	533 (31.8)	422 (33.4)	111 (25.5)
≥30	448 (13.0)	295 (12.2)	153 (14.7)	311 (20.3)	252 (21.1)	59 (17.2)
Missing	98 (2.8)	71 (3.0)	27 (2.2)	129 (6.6)	99 (6.6)	30 (6.4)
≥1 ADL dependency at baseline	637 (18.6)	369 (15.3)	268 (26.2)	347 (17.8)	215 (14.5)	132 (30.1)
Missing	25 (0.7)	16 (0.7)	9 (0.7)	2 (0.1)	1 (0)	1 (0.2)
≥1 IADL dependency at baseline	1598 (48.6)	889 (38.9)	709 (70.8)	1008 (52.8)	702 (47.8)	306 (71.6)
Missing	27 (0.8)	17 (0.7)	10 (0.8)	1 (0)	1 (0)	0
Difficulty walking several blocks	2031 (60.7)	1320 (56.6)	711 (70.2)	1024 (54.1)	733 (50.6)	291 (67.1)
Missing	18 (0.6)	13 (0.6)	5 (0.4)	12 (0.5)	11 (0.6)	1 (0.2)
Incontinence	905 (28.2)	586 (26.7)	319 (31.5)	526 (27.6)	383 (25.9)	143 (33.7)
Missing	41 (1.2)	12 (0.7)	29 (2.2)	11 (0.5)	8 (0.6)	3 (0.3)
Falls in the past 2 y	1459 (44.6)	971 (42.8)	488 (48.8)	658 (40.5)	483 (38.3)	175 (48.7)
Missing	15 (0.4)	9 (0.4)	6 (0.5)	7 (0.4)	6 (0.4)	1 (0.4)
Driving status
Never drove	191 (4.9)	132 (4.9)	59 (4.9)	247 (13.0)	165 (11.0)	82 (20.4)
Not driving	1714 (50.0)	1077 (45.6)	637 (60.1)	792 (43.0)	582 (40.6)	210 (52.0)
Still driving	1191 (37.1)	914 (40.4)	277 (29.5)	672 (43.9)	579 (48.3)	93 (27.6)
Missing	231 (8.0)	189 (9.2)	42 (5.5)	1 (0)	1 (0)	0
Comorbidities
Heart disease	1235 (37.9)	823 (36.3)	412 (41.7)	509 (30.1)	387 (29.8)	122 (31.1)
Missing	5 (0.2)	2 (0.1)	3 (0.4)	7 (0.3)	4 (0.2)	3 (1.0)
Lung disease	369 (11.6)	229 (10.3)	140 (14.6)	330 (21.2)	274 (22.6)	56 (16.0)
Missing	8 (0.3)	3 (0.1)	5 (0.7)	2 (0.2)	2 (0.2)	0
Stroke	629 (19.6)	398 (17.5)	231 (24.4)	227 (13.0)	160 (12.1)	67 (16.2)
Missing	7 (0.2)	4 (0.2)	3 (0.3)	7 (0.4)	5 (0.4)	2 (0.5)
Diabetes	739 (21.7)	496 (20.6)	243 (24.3)	506 (31.7)	400 (32.5)	106 (28.6)
Missing	10 (0.3)	6 (0.2)	4 (0.3)	1 (0)	1 (0)	0
Cancer	552 (17.1)	370 (16.6)	182 (18.3)	242 (13.7)	170 (12.4)	72 (18.5)
Missing	7 (0.2)	6 (0.3)	1 (0.1)	2 (0.1)	2 (0.1)	0

Open in a new tab

Abbreviations: ADL, activities of daily living; BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); IADL, instrumental ADL; NHLOC, nursing home level of care; NR, not reported.

^{^a}

Numbers within each column represent the raw unweighted number of participants within each cohort. The weighted percentages within each column represent the column percentage based on the weighted sample size after using national survey weights from the cohort studies.

^{^b}

Mean age for individuals in the National Health and Aging Trends Study is not reported because continuous age is not available.

^{^c}

Race and ethnicity were assessed to understand the demographic profile of individuals in the Health and Retirement Study and the National Health and Aging Trends Study. Race and ethnicity were self-reported by participants or proxies.

^{^d}

The other category includes American Indian, Alaska Native, Asian, Pacific Islander, and other race and was grouped together owing to small samples sizes.

In HRS, the median (IQR) follow-up time was 2.62 (1.11-5.26) years in the proxy cohort and 2.91 (1.31-5.05) years in the self-respondent cohort. Over the follow-up period, 2107 participants (63.3%) were classified as needing NHLOC (Table 1), including 1424 (61.6%) in the self-respondent model and 683 (67.3%) in the proxy model (eTable 6 in Supplement 1). In NHATS, 806 participants (47.1%) were classified as needing NHLOC, including 592 (44.6%) in the self-respondent model and 214 (55.6%) in proxy model (Table 1 and eTable 6 in Supplement 1). Cumulative incidence curves are shown in eFigures 4 and 5 in Supplement 1. Need for NHLOC was predominantly evaluated through proxy responses, either in a core or exit interview (eTable 7 in Supplement 1). For participants who died during follow-up prior to NHLOC classification (approximately 40%-50%), approximately 50% to 60% were classified as needing NHLOC based on exit interviews (eTable 8 in Supplement 1).

Model Development

eTables 9 and 10 in Supplement 1 summarize unadjusted hazard ratios. Table 2 includes the final variables selected for the proxy and self-respondent models along with regression coefficients. Both models included age, baseline ADL/IADL dependencies, and driving status; the proxy model additionally included BMI and falls history, while the self-respondent model additionally included female sex, incontinence, and number of incorrect orientation items (day of week, month, year).

Table 2. Multivariable-Adjusted Hazard Ratios and Regression Coefficients for the Final Variables Included in the Proxy and Self-Respondent Models Predicting Need for Nursing Home Level of Care^a.

Variable	Proxy model			Self-respondent model
Variable	Adjusted hazard ratio (95% CI)	β Coefficient (95% CI)	P value	Adjusted hazard ratio (95% CI)	β Coefficient (95% CI)	P value
Age category, y
65-69	1 [Reference]	0 [Reference]	NA	1 [Reference]	0 [Reference]	NA
70-74	1.31 (0.94 to 1.82)	0.27 (−0.07 to 0.60)	.12	1.53 (1.04 to 2.25)	0.42 (0.04 to 0.81)	.03
75-79	1.51 (1.11 to 2.06)	0.41 (0.10 to 0.72)	.01	1.56 (1.10 to 2.22)	0.45 (0.09 to 0.80)	.01
80-84	1.30 (0.96 to 1.77)	0.26 (−0.05 to 0.57)	.09	1.73 (1.23 to 2.43)	0.55 (0.20 to 0.89)	.002
85-89	1.54 (1.12 to 2.12)	0.43 (0.12 to 0.75)	.01	1.94 (1.38 to 2.73)	0.66 (0.32 to 1.01)	<.001
≥90	1.78 (1.26 to 2.52)	0.58 (0.23 to 0.93)	.001	1.65 (1.15 to 2.37)	0.50 (0.14 to 0.86)	.01
Female sex	NA	NA	NA	1.19 (1.03 to 1.38)	0.18 (0.03 to 0.32)	.02
BMI category
<18.5	1.38 (1.03 to 1.87)	0.33 (0.03 to 0.62)	.03	NA	NA	NA
18.5 to <25	1 [Reference]	0 [Reference]	NA	NA	NA	NA
25 to <30	1.15 (0.96 to 1.39)	0.14 (−0.04 to 0.33)	.13	NA	NA	NA
≥30	1.18 (0.93 to 1.50)	0.16 (−0.08 to 0.40)	.18	NA	NA	NA
ADL/IADL dependency count^b	1.23 (1.18 to 1.28)	0.20 (0.16 to 0.24)	<.001	1.08 (1.03 to 1.13)	0.07 (0.03 to 0.12)	.002
Driving status
Still driving	1 [Reference]	0 [Reference]	NA	1 [Reference]	0 [Reference]	NA
Not driving	1.34 (1.10 to 1.65)	0.29 (0.09 to 0.50)	.005	1.15 (1.00 to 1.33)	0.14 (−0.00 to 0.28)	.05
Never drove	0.77 (0.53 to 1.13)	−0.26 (−0.64 to 0.12)	.18	0.85 (0.64 to 1.12)	−0.17 (−0.45 to 0.11)	.24
Falls in the past 2 y	1.28 (1.09 to 1.50)	0.25 (0.09 to 0.41)	.002	NA	NA	NA
Incontinence	NA	NA	NA	1.17 (1.01 to 1.35)	0.15 (0.01 to 0.30)	.04
No. of incorrect date recall items^c
0	NA	NA	NA	1 [Reference]	0 [Reference]	NA
1	NA	NA	NA	1.18 (1.02 to 1.38)	0.17 (0.02 to 0.32)	.03
2	NA	NA	NA	1.73 (1.43 to 2.09)	0.55 (0.36 to 0.74)	<.001
3	NA	NA	NA	1.89 (1.56 to 2.29)	0.64 (0.45 to 0.83)	<.001

Open in a new tab

Abbreviations: ADL, activities of daily living; BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); IADL, instrumental ADL; NA, not applicable.

^{^a}

For the self-respondent model, the Weibull intercept β₀ = −7.8945 and the Weibull ancillary parameter P = .8235. For the proxy model, the Weibull intercept β₀ = −7.1725 and the Weibull ancillary parameter P = .7154. See the eAppendix in Supplement 1 for an example calculation using the final model.

^{^b}

The corresponding hazard ratio indicates the increase in hazard of nursing home level of care for each additional ADL/IADL dependency at baseline. ADL include needing help with bathing/showering, getting in or out of bed, dressing, and using the toilet. Eating was not included because an individual needing help with eating at baseline was excluded because they met the definition of nursing home level of care. IADL include using a telephone, preparing a meal, managing medications, managing money such as paying bills, and shopping for groceries. ADL/IADL dependency indicates needing help with an ADL or IADL (range, 0-7, as individuals with ≥3 ADL dependencies were excluded because they were considered to need nursing home level of care at baseline).

^{^c}

A simple date recall score was created based on the number of items incorrectly recalled when individuals are asked what the day of the week, current month, and year is (range, 0-3, where a score of 0 indicates that all items were correctly recalled).

Model Performance

After bootstrap internal validation, optimism-corrected iAUC was 0.72 (95% CI, 0.70-0.75) in the proxy model and 0.64 (95% CI, 0.62-0.66) in the self-respondent model (Table 3). AUC values in the proxy model at 2, 5, and 10 years were 0.74 (95% CI, 0.73-0.75), 0.75 (95% CI, 0.74-0.76), and 0.69 (95% CI, 0.68-0.71), respectively (Table 3 and eFigure 6 in Supplement 1). AUC values in the self-respondent model at 2, 5, and 10 years were 0.65 (95% CI, 0.62-0.67), 0.63 (95% CI, 0.61-0.65), and 0.61 (95% CI, 0.59-0.62), respectively. Calibration plots at 2, 5, and 10 years indicated excellent calibration across the range of predicted risk (Figure 1).

Table 3. Discrimination of the Nursing Home Level of Care Models.

Time	HRS (development cohort)			NHATS (validation cohort)
Time	No. remaining at risk (n = 1015)^a	Apparent AUC (95% CI)	Optimism-corrected AUC (95% CI)	No. remaining at risk (n = 385)^a	AUC (95% CI)
Proxy model
iAUC	NR	0.72 (0.70-0.75)	0.72 (0.70-0.75)	NR	0.66 (0.61-0.70)
2 y	577	0.75 (0.71-0.78)	0.74 (0.73-0.75)	225	0.67 (0.61-0.72)
5 y	274	0.76 (0.73-0.79)	0.75 (0.74-0.76)	93	0.66 (0.60-0.72)
10 y	78	0.70 (0.66-0.73)	0.69 (0.68-0.71)	NR	NR^b
Time	No. remaining at risk (n = 2312)^a	Apparent AUC (95% CI)	Optimism-corrected AUC (95% CI)	No. remaining at risk (n = 1327)^a	AUC (95% CI)
Self-respondent model
iAUC	NR	0.64 (0.62-0.66)	0.64 (0.62-0.66)	NR	0.64 (0.62-0.67)
2 y	1448	0.66 (0.63-0.68)	0.65 (0.62-0.67)	890	0.64 (0.60-0.68)
5 y	589	0.64 (0.62-0.66)	0.63 (0.61-0.65)	348	0.64 (0.60-0.67)
10 y	107	0.61 (0.58-0.64)	0.61 (0.59-0.62)	NR	NR^b

Open in a new tab

Abbreviations: AUC, area under the receiver operating characteristic curve; HRS, Health and Retirement Study; iAUC, integrated AUC; NHATS, National Health and Aging Trends Study; NR, not reported.

^{^a}

The number remaining at risk indicates the number of individuals remaining in the study cohorts at specific time points as individuals either are classified as needing nursing home level of care (outcome event), have a competing event (die without being considered as needing nursing home level of care), or are censored. See eFigures 4 and 5 in Supplement 1 for the unadjusted cumulative incidence curves.

^{^b}

The AUC at 10 years is not reported for the NHATS external validation cohort because the period of observation was not long enough to obtain estimates at 10 years.

Figure 1. — Calibration plots at the 2-, 5-, and 10-year time points indicate the agreement between the predicted nursing home level of care risk and observed nursing home level of care risk. Data from the Health and Retirement Study (HRS; development cohort) are presented as blue circles, and data from the National Health and Aging Trends Study (NHATS; validation cohort) are presented as orange circles. The samples were divided into fifths by splitting at quintiles of predicted risk. The blue circles and orange circles represent the mean risk estimates of the outcome from each fifth of predicted risk. Perfect predictions would have the circles at the 45° dotted line. A density plot displays the distribution of predicted risk in HRS (blue shaded) and NHATS (orange shaded). Calibration in NHATS was not assessed at 10 years because the duration of follow-up was not long enough.

On external validation in NHATS, the calibration slope (SE) for the proxy model was 0.77 (0.13) and 0.92 (0.10) for the self-respondent model. eFigure 7 in Supplement 1 shows the histogram of the prognostic index. iAUC in the proxy and self-respondent models were 0.66 (95% CI, 0.61-0.70) and 0.64 (95% CI, 0.62-0.67), respectively (Table 3). Compared with HRS, time-specific AUC values in NHATS were comparable in the self-respondent model and lower in the proxy model. For example, 2-year AUC in the proxy model was 0.67 (95% CI, 0.61-0.72) in NHATS and 0.74 (95% CI, 0.73-0.75) in HRS (Table 3 and eFigure 8 in Supplement 1). Calibration in NHATS was similar to HRS (Figure 1). Sensitivity analyses demonstrated similar performance across various modeling strategies and racial and ethnic subgroups in HRS and NHATS (eMethods 3, eTables 11-18, and eFigures 9-12 in Supplement 1). For illustration, Figure 2 displays baseline characteristics and median predicted time to needing NHLOC for 10 randomly selected participants within fifths of predicted risk. An example calculation using the model is provided in the eAppendix in Supplement 1.

Figure 2. — A and B, Baseline characteristics for randomly selected participants in the Health and Retirement Study from the self-respondent cohort (5 participants) and proxy cohort (5 participants) within each fifth of predicted risk. The color and shading of the individual boxes indicate the direction and magnitude of the specific baseline characteristic’s prognostic effect on time to needing NHLOC, respectively. For example, going from light pink to dark pink shading indicates increasing hazard of needing NHLOC. Green shading indicates that the characteristic is associated with lower hazard of needing NHLOC. C and D, The median predicted time to needing NHLOC in years is displayed for 10 participants. The ends of the whiskers indicate the 25th and 75th percentiles of predicted time to needing NHLOC. ADL indicates activities of daily living; BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); IADL, instrumental ADL.

Discussion

Using 2 nationally representative samples of US community-dwelling older adults with probable dementia, we developed and performed an external validation of models to predict need for NHLOC based on self-responses and proxy responses. We included a minimal number of predictors commonly collected in clinical practice, including demographics, baseline functional status, health/behavioral factors, and date recall.

These models may be used to enhance conversations with patients and families about future planning by giving estimates of when an individual with dementia may need high-level care, whether provided in a NH or at home. Predicting need for NHLOC rather than NH admission provides a standardized metric applicable across cultural and socioeconomic backgrounds. The present models may be used even for families who would avoid having their loved one enter a NH by allowing them to estimate when they may need to hire caregivers or become the individual’s full-time home caregiver. At population levels, these models may identify individuals that most benefit from in-home and community-based services.

In clinical practice, those with early-stage dementia will start out as the primary source of medical information. Eventually, persons with dementia will progress to where a surrogate becomes the primary source of information. Use of these models should mirror this clinical reality, with the self-respondent model used in early-stage dementia and proxy model used in later stages. In cases where both self-respondent and proxy information is available, it should be noted that HRS and NHATS turned to the proxy only when the respondent was unable/unwilling to provide information. While the final predictors in these models are slightly different, they have considerable overlap, with age and baseline ADL/IADL dependencies being prominent prognostic factors.

A few studies are worth noting in comparison. Among individuals with dementia, Yaffe et al¹⁸ developed a model predicting time to NH placement using the Medicare Alzheimer Disease Demonstration and Evaluation study (1989-1994). The final model included patient variables (ethnicity, living alone, ≥1 ADL dependency, Mini-Mental State Examination, difficult behaviors) and caregiver variables (age, relationship, Zarit Burden Scale). The C statistic was 0.66 and 0.63 in the development and validation cohort, respectively. The most closely related model to the present study is Stern et al¹⁹ and related studies,^20,40,41 which predicted time to NH level care (defined by an interviewer’s impression that the individual’s current needs were equivalent to facility-level care or a score on the Dependence Scale of 4-5).⁴² The original model included age at dementia onset, duration of dementia, modified Mini-Mental State Examination, extrapyramidal signs, and psychotic symptoms.¹⁹ An updated 2021 study²⁰ validated a similar Cox model and applied a longitudinal grade of membership model using additional predictors. The primary disadvantage includes unclear generalizability, as it was developed using approximately 250 individuals in the predictors 1 (1989-2001) and predictors 2 (1997-2007) cohorts from 3 memory clinics at academic research centers. However, it has subsequently been validated in a community-based cohort.⁴³ Among individuals without dementia, Wolff et al¹⁵ developed a model predicting NH admission using the National Long-Term Care Survey and NHATS that performed similarly to the present models (C statistic of 0.65 after internal validation).

We found higher discrimination in the proxy model compared with the self-respondent model (eg, iAUC in HRS of 0.72 vs 0.64, respectively). This may be because participants in the self-respondent cohort provided inaccurate functional assessments compared with proxies due to cognitive impairment (ie, lack of insight) or desire to downplay deficits. This could contribute to greater bias in the outcome, which was frequently assessed later by a proxy. Greater measurement error in baseline functional status might make it more difficult for the model to discriminate.

We found lower discrimination in the proxy model on external validation in NHATS (iAUC, 0.66) compared with HRS (iAUC, 0.72), which could be explained by several factors. First, differences in the case mix of the validation sample may result in decreased model performance.^44,45,46 We used NHATS as the external validation cohort to provide estimates of predictive performance for the intended target setting, namely community-dwelling older adults with dementia not necessarily seen in specialized memory clinics.⁴⁷ While HRS and NHATS both aim to create nationally representative samples, participants from NHATS were older and different with respect to the proportion who needed NHLOC (67.3% and 55.6% in HRS and NHATS proxy cohorts, respectively). This may be due to differences in how participants were classified as having dementia or selected into the proxy interview. Second, heterogeneity of predictor or outcome measurements across settings has been shown to affect model performance.^48,49,50 This may be particularly relevant for functional status, as a previous study showed that the proportions of individuals receiving help with various ADL differed between NHATS and HRS, partially due to variation in the wording of questions and skip patterns.⁵¹

The present models did not perform as well on measures of discrimination when compared with other models among community-dwelling older adults for outcomes such as mortality or hospitalization.⁵² This is likely because NHLOC is harder to predict and more subjective. However, these models may still provide added value over current practice, whereby clinician estimates are prone to substantial underestimation or overestimation.⁵³ Discrimination of the models was on par with other widely used calculators,^54,55 and calibration was excellent.

Strengths and Limitations

This study has several strengths. We used 2 large, nationally representative cohorts to develop and externally validate the models. Predictors were readily obtainable and could be inputted into an online calculator. Compared with prior studies, the present model predicted need for NHLOC using a definition more applicable to individuals with dementia.

This study also has limitations. First, participants were classified with dementia using an algorithm that may be subject to misclassification and does not necessarily represent the way dementia diagnoses are made in clinic. While this algorithm has high accuracy in validation studies against criterion-standard clinical diagnoses, accuracy is reduced in certain subgroups (eg, racial and ethnic minority groups, less-educated individuals).^28,56,57 Second, information on dementia cause (eg, Alzheimer disease, vascular disease) or severity was not available. Third, the outcome may be subject to misclassification, which may be more relevant to the self-respondent model. Concordance between functional ability based on self-report, proxy report, and objective measures is complex. While some studies show that persons with dementia may underestimate their functional impairment, this can vary based on individual-level factors.^58,59,60 Finally, the models were developed among older US adults, which may not be applicable to international contexts. While our definition of NHLOC can still apply across geographic locations as a marker of high-level care needs, futures studies should validate the model in other countries. Additionally, future studies should evaluate model performance in newer cohorts.

Conclusions

This prognostic study showed that relatively simple models using self-report or proxy responses can predict need for NHLOC among older adults with probable dementia with moderate discrimination and excellent calibration. Given that most individuals with dementia ultimately need NHLOC, model estimates may help frame conversations between patients and families/caregivers regarding care planning.

Supplement 1.

eMethods 1. Additional details on the Health and Retirement Study and National Health and Aging Trends Study

eMethods 2. Additional details on defining the primary outcome of nursing home level of care

eMethods 3. Additional details on model development, including sensitivity analyses for an interval-censored model incorporating competing risks and survey weights

eMethods 4. Additional details on assessing model performance

eAppendix. Example calculation using the final model

eFigure 1. Schematic of possible scenarios that could occur during follow-up when assessing nursing home level of care

eFigure 2. Flow chart of participants from the Health and Retirement Study 1998-2016

eFigure 3. Flow chart of participants from the National Health and Aging Trends Study 2011-2019

eFigure 4. Unadjusted cumulative incidence for the outcome of nursing home level of care among proxy and self-respondents in the Health and Retirement Study (development cohort)

eFigure 5. Unadjusted cumulative incidence for the outcome of nursing home level of care among proxy and self-respondents in the National Health and Aging Trends Study (validation cohort)

eFigure 6. Graph of the area under the receiver operating characteristic curve over time with 95% confidence limits for the final proxy and self-respondent models following bootstrap internal validation in the Health and Retirement Study

eFigure 7. Histogram of the prognostic index in the Health and Retirement Study (development cohort) and National Health and Aging Trends Study (validation cohort) for the (A) proxy and (B) self-respondent models

eFigure 8. Graph of the area under the receiver operating characteristic curve over time with 95% confidence limits for the final proxy and self-respondent models following bootstrap internal validation in the National Health and Aging Trends Study

eFigure 9. Comparison of model calibration in the proxy and self-respondent models for the primary analysis and two sensitivity analyses comparing alternative modeling approaches for individuals who died during follow-up in the Health and Retirement Study

eFigure 10. Comparison of model calibration in the proxy and self-respondent models for the primary analysis and two sensitivity analyses comparing alternative modeling approaches for individuals who died during follow-up in the National Health and Aging Trends Study

eFigure 11. Comparison of model predicted probabilities for time to needing nursing home level of care in the proxy model for the primary analysis and sensitivity analysis 1 (left) and sensitivity analysis 2 (right) comparing alternative modeling approaches for individuals who died during follow-up

eFigure 12. Comparison of model predicted probabilities for time to needing nursing home level of care in the self-respondent model for the primary analysis and sensitivity analysis 1 (left) and sensitivity analysis 2 (right) comparing alternative modeling approaches for individuals who died during follow-up

eTable 1. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) checklist for prediction model development and validation

eTable 2. Full description of candidate predictors for the proxy model in the Health and Retirement Study (HRS), including categorization scheme

eTable 3. Additional candidate predictors that were included in the predictor pool for the self-respondent model in the Health and Retirement Study (HRS), including categorization scheme

eTable 4. Full description of predictors in the National Health and Aging Trends Study (NHATS) that were included in external validation

eTable 5. Four scenarios that individuals could fall into when determining time of nursing home level of care accounting for interval censoring and competing risks

eTable 6. Number of individuals who were classified as nursing home level of care during the follow-up period by individual components of the criteria in the Health and Retirement Study and National Health and Aging Trends Study cohorts

eTable 7. Number of individuals assessed as becoming nursing home level of care based on who provided the information (proxy vs. self-respondent)

eTable 8. Handling of deaths among individuals not previously classified as nursing home level of care during follow-up and how their status was ascertained

eTable 9. Unadjusted hazard ratios of all the candidate predictors and adjusted hazard ratios of the selected variables in the final proxy model

eTable 10. Unadjusted hazard ratios of all the candidate predictors and adjusted hazard ratios of the selected variables in the final self-respondent model

eTable 11. Model discrimination across racial/ethnic subgroups in the proxy and self-respondent models for the Health and Retirement Study (development cohort) and National Health and Aging Trends Study (validation cohort)

eTable 12. Comparison of interval-censored parametric regression models using Weibull, exponential, and Gompertz distributions

eTable 13. Comparison of model hazard ratios and discrimination for the proxy and self-respondent models in the primary interval-censored Weibull model and sensitivity analysis using interval-censored Cox model both without multiple imputation or survey features

eTable 14. Sensitivity analyses calculating model discrimination in the Health and Retirement Study when modifying the value of time low in the interval-censored Weibull model and modifying the definition of survival time in the discrimination calculation

eTable 15. Comparison of model hazard ratios and discrimination for interval-censored parametric Weibull model incorporating competing risks and Fine-Gray competing risk model

eTable 16. Comparison of model hazard ratios in the proxy-respondent model for two sensitivity analyses comparing alternative modeling approaches for individuals who died during follow-up

eTable 17. Comparison of model hazard ratios in the self-respondent model for two sensitivity analyses comparing alternative modeling approaches for individuals who died during follow-up

eTable 18. Comparison of model discrimination in the proxy and self-respondent models for two sensitivity analyses comparing alternative modeling approaches for individuals who died during follow-up

eReferences

Click here for additional data file.^{(1.2MB, pdf)}

Supplement 2.

Data Sharing Statement

Click here for additional data file.^{(15.8KB, pdf)}

References

1.2023 Alzheimer’s disease facts and figures. Alzheimers Dement. 2023;19(4):1598-1695. doi: 10.1002/alz.13016 [DOI] [PubMed] [Google Scholar]
2.Arrighi HM, Neumann PJ, Lieberburg IM, Townsend RJ. Lethality of Alzheimer disease and its impact on nursing home placement. Alzheimer Dis Assoc Disord. 2010;24(1):90-95. doi: 10.1097/WAD.0b013e31819fe7d1 [DOI] [PubMed] [Google Scholar]
3.Gaugler JE, Yu F, Krichbaum K, Wyman JF. Predictors of nursing home admission for persons with dementia. Med Care. 2009;47(2):191-198. doi: 10.1097/MLR.0b013e31818457ce [DOI] [PubMed] [Google Scholar]
4.Toot S, Swinson T, Devine M, Challis D, Orrell M. Causes of nursing home placement for older people with dementia: a systematic review and meta-analysis. Int Psychogeriatr. 2017;29(2):195-208. doi: 10.1017/S1041610216001654 [DOI] [PubMed] [Google Scholar]
5.Mjørud M, Selbæk G, Bjertness E, et al. Time from dementia diagnosis to nursing-home admission and death among persons with dementia: a multistate survival analysis. PLoS One. 2020;15(12):e0243513. doi: 10.1371/journal.pone.0243513 [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Duan-Porter W, Ullman K, Rosebush C, et al. Interventions to prevent or delay long-term nursing home placement for adults with impairments—a systematic review of reviews. J Gen Intern Med. 2020;35(7):2118-2129. doi: 10.1007/s11606-019-05568-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Shepherd-Banigan M, James HJ, Smith VA, et al. Drivers of long-term care considerations by persons with cognitive impairment. J Appl Gerontol. 2021;40(6):648-660. doi: 10.1177/0733464820903908 [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Thomeer MB, Mudrazija S, Angel JL. How do race and Hispanic ethnicity affect nursing home admission? evidence from the Health and Retirement Study. J Gerontol B Psychol Sci Soc Sci. 2015;70(4):628-638. doi: 10.1093/geronb/gbu114 [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Yan D, Wang S, Temkin-Greener H, Cai S. The use patterns of Medicaid home and community based services among Medicare/Medicaid beneficiaries with dementia. Front Public Health. 2021;9:708402. doi: 10.3389/fpubh.2021.708402 [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Cai X, Temkin-Greener H. Nursing home admissions among Medicaid HCBS enrollees: evidence of racial/ethnic disparities or differences? Med Care. 2015;53(7):566-573. doi: 10.1097/MLR.0000000000000379 [DOI] [PubMed] [Google Scholar]
11.Yan D, Wang S, Temkin-Greener H, Cai S. HCBS service spending and nursing home placement for patients with Alzheimer’s disease and related dementias: does race matter? J Appl Gerontol. 2022;41(3):638-649. doi: 10.1177/07334648211048187 [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Gaugler JE, Duval S, Anderson KA, Kane RL. Predicting nursing home admission in the U.S: a meta-analysis. BMC Geriatr. 2007;7(1):13. doi: 10.1186/1471-2318-7-13 [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Cepoiu-Martin M, Tam-Tham H, Patten S, Maxwell CJ, Hogan DB. Predictors of long-term care placement in persons with dementia: a systematic review and meta-analysis. Int J Geriatr Psychiatry. 2016;31(11):1151-1171. doi: 10.1002/gps.4449 [DOI] [PubMed] [Google Scholar]
14.Van Grootven B, van Achterberg T. Prediction models for functional status in community dwelling older adults: a systematic review. BMC Geriatr. 2022;22(1):465. doi: 10.1186/s12877-022-03156-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Wolff JL, Mulcahy J, Roth DL, et al. Long-term nursing home entry: a prognostic model for older adults with a family or unpaid caregiver. J Am Geriatr Soc. 2018;66(10):1887-1894. doi: 10.1111/jgs.15447 [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Greiner MA, Qualls LG, Iwata I, et al. Predicting nursing home placement among home- and community-based services program participants. Am J Manag Care. 2014;20(12):e535-e536. [PubMed] [Google Scholar]
17.Kim DH, Newman AB, Lipsitz LA. Prediction of severe, persistent activity-of-daily-living disability in older adults. Am J Epidemiol. 2013;178(7):1085-1093. doi: 10.1093/aje/kwt097 [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Yaffe K, Fox P, Newcomer R, et al. Patient and caregiver characteristics and nursing home placement in patients with dementia. JAMA. 2002;287(16):2090-2097. doi: 10.1001/jama.287.16.2090 [DOI] [PubMed] [Google Scholar]
19.Stern Y, Tang MX, Albert MS, et al. Predicting time to nursing home care and death in individuals with Alzheimer disease. JAMA. 1997;277(10):806-812. doi: 10.1001/jama.1997.03540340040030 [DOI] [PubMed] [Google Scholar]
20.Stern Y, Stallard E, Kinosian B, et al. Validation and demonstration of a new comprehensive model of Alzheimer’s disease progression. Alzheimers Dement. 2021;17(10):1698-1708. doi: 10.1002/alz.12336 [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Kurichi JE, Bogner HR, Streim JE, et al. Predicting 3-year mortality and admission to acute-care hospitals, skilled nursing facilities, and long-term care facilities in Medicare beneficiaries. Arch Gerontol Geriatr. 2017;73:248-256. doi: 10.1016/j.archger.2017.08.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Covinsky KE, Hilton J, Lindquist K, Dudley RA. Development and validation of an index to predict activity of daily living dependence in community-dwelling elders. Med Care. 2006;44(2):149-157. doi: 10.1097/01.mlr.0000196955.99704.64 [DOI] [PubMed] [Google Scholar]
23.Taş U, Steyerberg EW, Bierma-Zeinstra SM, Hofman A, Koes BW, Verhagen AP. Age, gender and disability predict future disability in older people: the Rotterdam Study. BMC Geriatr. 2011;11(1):22. doi: 10.1186/1471-2318-11-22 [DOI] [PMC free article] [PubMed] [Google Scholar]
24.den Ouden MEM, Schuurmans MJ, Mueller-Schotte S, van der Schouw YT. Identification of high-risk individuals for the development of disability in activities of daily living: a ten-year follow-up study. Exp Gerontol. 2013;48(4):437-443. doi: 10.1016/j.exger.2013.02.002 [DOI] [PubMed] [Google Scholar]
25.Nüesch E, Pablo P, Dale CE, et al. Incident disability in older adults: prediction models based on two British prospective cohort studies. Age Ageing. 2015;44(2):275-282. doi: 10.1093/ageing/afu159 [DOI] [PubMed] [Google Scholar]
26.Jonkman NH, Colpo M, Klenk J, et al. Development of a clinical prediction model for the onset of functional decline in people aged 65-75 years: pooled analysis of four European cohort studies. BMC Geriatr. 2019;19(1):179. doi: 10.1186/s12877-019-1192-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Fisher GG, Ryan LH. Overview of the Health and Retirement Study and introduction to the special issue. Work Aging Retire. 2018;4(1):1-9. doi: 10.1093/workar/wax032 [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Wu Q, Tchetgen EJ, Osypuk TL, White K, Mujahid M, Maria Glymour M. Combining direct and proxy assessments to reduce attrition bias in a longitudinal study. Alzheimer Dis Assoc Disord. 2013;27(3):207-212. doi: 10.1097/WAD.0b013e31826cfe90 [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Freedman VA, Kasper JD. Cohort profile: the National Health and Aging Trends Study (NHATS). Int J Epidemiol. 2019;48(4):1044-1045g. doi: 10.1093/ije/dyz109 [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Kasper JD, Freedman VA, Spillman B. Classification of persons by dementia status in the National Health and Aging Trends Study. National Health and Aging Trends Study . July 2013. Accessed February 8, 2022. https://www.nhats.org/sites/default/files/inline-files/DementiaTechnicalPaperJuly_2_4_2013_10_23_15.pdf
31.O’Keeffe J. People with dementia: can they meet Medicaid level-of-care criteria for admission to nursing homes and home and community-based waiver programs? AARP Public Policy Institute . August 1999. Accessed June 12, 2022. https://assets.aarp.org/rgcenter/health/9912_dementia.pdf
32.O’Keeffe J, Tilly J, Lucas C. Medicaid eligibility criteria for long term care services: access for people with Alzheimer’s disease and other dementias. Alzheimer’s Association . May 2006. Accessed January 6, 2023. https://www.alz.org/documents/national/Medicaideligibilityissues.pdf
33.Fox P, Maslow K, Zhang X. Long-term care eligibility criteria for people with Alzheimer’s disease. Health Care Financ Rev. 1999;20(4):67-85. [PMC free article] [PubMed] [Google Scholar]
34.Wolbers M, Koller MT, Witteman JCM, Steyerberg EW. Prognostic models with competing risks: methods and application to coronary risk prediction. Epidemiology. 2009;20(4):555-561. doi: 10.1097/EDE.0b013e3181a39056 [DOI] [PubMed] [Google Scholar]
35.Heymans MW, van Buuren S, Knol DL, van Mechelen W, de Vet HC. Variable selection under multiple imputation using the bootstrap in a prognostic study. BMC Med Res Methodol. 2007;7(1):33. doi: 10.1186/1471-2288-7-33 [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Vergouw D, Heymans MW, Peat GM, et al. The search for stable prognostic models in multiple imputed data sets. BMC Med Res Methodol. 2010;10(1):81. doi: 10.1186/1471-2288-10-81 [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Austin PC, Lee DS, Ko DT, White IR. Effect of variable selection strategy on the performance of prognostic models when using multiple imputation. Circ Cardiovasc Qual Outcomes. 2019;12(11):e005927. doi: 10.1161/CIRCOUTCOMES.119.005927 [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Schomaker M, Heumann C. Bootstrap inference when using multiple imputation. Stat Med. 2018;37(14):2252-2266. doi: 10.1002/sim.7654 [DOI] [PMC free article] [PubMed] [Google Scholar]
39.ePrognosis. Accessed September 18, 2023. https://eprognosis.ucsf.edu/
40.Razlighi QR, Stallard E, Brandt J, et al. A new algorithm for predicting time to disease endpoints in Alzheimer’s disease patients. J Alzheimers Dis. 2014;38(3):661-668. doi: 10.3233/JAD-131142 [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Stallard E, Kinosian B, Stern Y. Personalized predictive modeling for patients with Alzheimer’s disease using an extension of Sullivan’s life table model. Alzheimers Res Ther. 2017;9(1):75. doi: 10.1186/s13195-017-0302-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Stern Y, Albert SM, Sano M, et al. Assessing patient dependence in Alzheimer’s disease. J Gerontol. 1994;49(5):M216-M222. doi: 10.1093/geronj/49.5.M216 [DOI] [PubMed] [Google Scholar]
43.Stallard E, Kociolek A, Jin Z, et al. Validation of a multivariate prediction model of the clinical progression of Alzheimer’s disease in a community-dwelling multiethnic cohort. J Alzheimers Dis. 2023;95(1):93-117. doi: 10.3233/JAD-220811 [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Vergouwe Y, Moons KGM, Steyerberg EW. External validity of risk models: use of benchmark values to disentangle a case-mix effect from incorrect coefficients. Am J Epidemiol. 2010;172(8):971-980. doi: 10.1093/aje/kwq223 [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Riley RD, Ensor J, Snell KIE, et al. External validation of clinical prediction models using big datasets from e-health records or IPD meta-analysis: opportunities and challenges. BMJ. 2016;353:i3140. doi: 10.1136/bmj.i3140 [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Nieboer D, van der Ploeg T, Steyerberg EW. Assessing discriminative performance at external validation of clinical prediction models. PLoS One. 2016;11(2):e0148820. doi: 10.1371/journal.pone.0148820 [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Sperrin M, Riley RD, Collins GS, Martin GP. Targeted validation: validating clinical prediction models in their intended population and setting. Diagn Progn Res. 2022;6(1):24. doi: 10.1186/s41512-022-00136-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Luijken K, Wynants L, van Smeden M, Van Calster B, Steyerberg EW, Groenwold RHH; Collaborators . Changing predictor measurement procedures affected the performance of prediction models in clinical examples. J Clin Epidemiol. 2020;119:7-18. doi: 10.1016/j.jclinepi.2019.11.001 [DOI] [PubMed] [Google Scholar]
49.Luijken K, Groenwold RHH, Van Calster B, Steyerberg EW, van Smeden M. Impact of predictor measurement heterogeneity across settings on the performance of prediction models: a measurement error perspective. Stat Med. 2019;38(18):3444-3459. doi: 10.1002/sim.8183 [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Pajouheshnia R, van Smeden M, Peelen LM, Groenwold RHH. How variation in predictor measurement affects the discriminative ability and transportability of a prediction model. J Clin Epidemiol. 2019;105:136-141. doi: 10.1016/j.jclinepi.2018.09.001 [DOI] [PubMed] [Google Scholar]
51.Ankuda CK, Covinsky K, Freedman VA, et al. The devil’s in the details: variation in estimates of late-life activity limitations across national cohort studies. J Am Geriatr Soc. 2023;71(3):858-868. doi: 10.1111/jgs.18158 [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Yourman LC, Lee SJ, Schonberg MA, Widera EW, Smith AK. Prognostic indices for older adults: a systematic review. JAMA. 2012;307(2):182-192. doi: 10.1001/jama.2011.1966 [DOI] [PMC free article] [PubMed] [Google Scholar]
53.Christakis NA, Lamont EB. Extent and determinants of error in doctors’ prognoses in terminally ill patients: prospective cohort study. BMJ. 2000;320(7233):469-472. doi: 10.1136/bmj.320.7233.469 [DOI] [PMC free article] [PubMed] [Google Scholar]
54.Rockhill B, Spiegelman D, Byrne C, Hunter DJ, Colditz GA. Validation of the Gail et al. model of breast cancer risk prediction and implications for chemoprevention. J Natl Cancer Inst. 2001;93(5):358-366. doi: 10.1093/jnci/93.5.358 [DOI] [PubMed] [Google Scholar]
55.Lip GYH, Nieuwlaat R, Pisters R, Lane DA, Crijns HJGM. Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: the Euro Heart Survey on atrial fibrillation. Chest. 2010;137(2):263-272. doi: 10.1378/chest.09-1584 [DOI] [PubMed] [Google Scholar]
56.Gianattasio KZ, Wu Q, Glymour MM, Power MC. Comparison of methods for algorithmic classification of dementia status in the health and retirement study. Epidemiology. 2019;30(2):291-302. doi: 10.1097/EDE.0000000000000945 [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Gianattasio KZ, Prather C, Glymour MM, Ciarleglio A, Power MC. Racial disparities and temporal trends in dementia misdiagnosis risk in the United States. Alzheimers Dement (N Y). 2019;5:891-898. doi: 10.1016/j.trci.2019.11.008 [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Farias ST, Mungas D, Jagust W. Degree of discrepancy between self and other-reported everyday functioning by cognitive status: dementia, mild cognitive impairment, and healthy elders. Int J Geriatr Psychiatry. 2005;20(9):827-834. doi: 10.1002/gps.1367 [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Martyr A, Gamble LD, Nelis SM, et al. ; IDEAL Study Team . Predictors of awareness of functional ability in people with dementia: the contribution of personality, cognition, and neuropsychiatric symptoms—findings from the IDEAL program. Dement Geriatr Cogn Disord. 2022;51(3):221-232. doi: 10.1159/000524607 [DOI] [PubMed] [Google Scholar]
60.Martyr A, Clare L. Awareness of functional ability in people with early-stage dementia. Int J Geriatr Psychiatry. 2018;33(1):31-38. doi: 10.1002/gps.4664 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials