| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
a Palmetto Richland Memorial Hospital, Division of Geriatrics Services, and Professor, Division of Geriatric Medicine, University of South Carolina School of Medicine, Columbia, SC
b Center for Demographic Studies, Duke University, Durham, NC
c Medfocus, Inc., Des Plains, IL
d University of South Carolina Schools of Public Health, Columbia, SC
e University of South Carolina Schools of Medicine, Columbia, SC
f Palmetto Senior Care, Palmetto Richland Memorial Hospital, Columbia, SC
Correspondence: Darryl Wieland, PhD, MPH, Research Director, Palmetto Richland Memorial Hospital, Division of Geriatrics Services, and Professor, Division of Geriatric Medicine, University of South Carolina School of Medicine, 9 Medical Park, #630, Columbia, SC 29203. E-mail: isbjorn{at}aol.com.
Decision Editor: Vernon L. Greene, PhD
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionReferences |
|---|
). Using grade-of-membership analysis, we classified participants on the basis of their specific diseases, impairments, and disabilities. The classification was reviewed by a physician panel to produce clinical profiles, which were then validated against participants' PACE tenure, demographics, supports, and health. Cognitive impairment, incontinence, and activities of daily living disabilities were influential in producing eight types, which correspond predictably to responses in tenure (the more disabled, ill types likely to be in PACE longer), demographics, health, and informal support.
Key Words: Community long-term care • Frail elderly people • Case-mix classification • Grade-of-membership analysis
Until late 1997, the Program of All-Inclusive Care for the Elderly (PACE) was a Health Care Financing Administration-waivered demonstration extending the care and financing model developed by On Lok in San Francisco (Eng, Pedulla, Eleazer, McCann, and Fox 1997
). PACE is targeted to frail, disabled, noninstitutionalized elderly persons whom states certify as requiring nursing home care. For each enrollee, PACE operates under full capitation, receiving payments at predetermined rates from Medicare, Medicaid, and/or private-pay sources and assuming responsibility for any needed medical or supportive services. Because of low voluntary disenrollment (4.7% of those served in 1998), PACE provides lifetime care for most participants. The program uses community-based interdisciplinary team assessment and management with integrated medical care in homes and day centers, aiming to bolster participants' quality of life and to avoid inappropriate institutional care. By March 1997 (the time of our study), 12 fully capitated sites operated in nine states under demonstration status. PACE's legal establishment as a Medicare provider under the 1997 Balanced Budget Act is leading to its steady expansion. As of July 1999, 25 capitated and 8 developmental sites in 18 states served over 6,000 participants.
The objective of the present study is to make use of the full range of disease, impairment, and disability data available for the PACE demonstration's population in order to identify salient clinical profiles and their relative prevalence. We describe the methods and results of this classification, with emphasis on the interpretation of the clinical types and their validation.
Understanding the heterogeneous PACE population in terms of coherent clinical groupings is critical to optimizing the program as it expands. It will be difficult to address differences in outcomes, service patterns, and costs across sites, or between established and expansion sites, without an effective, independent, clinically based adjustment for participants' clinical status. Given sufficient data, similar adjustments can be used to compare outcomes and costs of PACE with those of other provider types by profiling larger populations of those at risk of, eligible for, or receiving long-term care. Such a clinically based classification scheme can be applied to improvement of quality, development of equitable reimbursement, and appropriate targeting and assurance of access for those needing special or heavy care.
Our work closely parallels Manton's (Manton, Cornelius and Woodbury 1995) classification of nursing-home residents. Specifically, an advanced multivariate technique—grade-of-membership (GoM) analysis—develops the PACE participant types. Information on PACE participants' assessed diseases, impairment, and disabilities (i.e., internal variables) is used to estimate the number of participant types and the likelihood of each type's response on each of the disease, impairment, and disability factors. We then describe the process by which the detailed quantitative solution is reviewed both to develop narrative interpretations of the solution (clinical profiles) and to validate their content (i.e., ensure that there are no major data or analytic problems). Predictive validation is provided by conditionally estimating the responses for each of the types on a variety of demographic, health, and informal support (external) variables.
We chose GoM over alternative case-mix classification techniques for several reasons. First, most discriminant procedures do not perform well in describing the clinical status of elderly patients for whom there are large numbers of variables indicative of multiple comorbid diseases, impairments, and disabilities. A problem of "crisp-set" methods is that classification within a diverse elderly population of a complex clinical state, as reflected by multiple measures of disease, impairments and disabilities, is unlikely to produce a small number of classical symptom patterns that exactly describe the individuals' attributes. In comparison, GoM suffers from fewer constraints regarding the form and nature of data and the classification of sampling units. For example, GoM has been shown to increase in classificatory precision given large sets of discretely coded variables. Further, participants may be reflected in more than one clinical profile, but to differing degrees, efficiently explained by using a set of weights (Manton, Woodbury, and Tolley 1994). GoM has been applied to a wide range of data describing elderly and disabled populations, wherein its solutions were substantively and theoretically meaningful and robust to missing data problems and measurement deficiencies (Manton and Woodbury 1991).
|
Methods |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
. Most participants had either an admission or a quarterly comprehensive follow-up assessment during this period. For others, we recovered health assessment information from December 1996, to form a complete prevalence sample representing open cases in late 1996/early 1997.
All variables were derived from the PACE minimum data set (On Lok, Inc 1994). This consisted of information submitted by sites through the demonstration's data entry and submission software, DataPACE. DataPACE management was provided under HCFA contract by On Lok, which provided program statistics and evaluation data. DataPACE contains participant-level demographic and program information, periodic health and functional assessments, monthly service utilization statistics, inpatient episodes, and annual informal support assessments. On Lok provided sites with training, retraining, and technical assistance on site and by telephone. Nurses responsible for data entry at each site participated in semiannual reliability tests based on videotaped participant assessments. On Lok furnished the software, provided initial and ongoing technical assistance, reviewed data, and provided feedback about problems to correct errors.
The GoM Model
A multivariate classification technique based on fuzzy-set mathematics, the GoM model is a general multivariate procedure for analyzing high-dimensional discrete response data (Manton, Woodbury, and Tolley 1994). Using maximum likelihood principles, GoM estimates two types of parameters: (a) the probability (
kl) of a particular response on a given variable (e.g., 
) for one of K analytically defined types where l is the given variable and (b) an individual's degree of membership (gik ) in each of the K types where i is the individual subject. In other words, GoM analysis simultaneously generates nosological types while quantifying the degree of an individual's membership in each. The procedure does not assume that most individuals are classifiable as classical cases into discrete groups. Rather, participants will have a grade of membership, or partial membership, in more than one profile, reflecting the logic of fuzzy partitions. Given this, the reported prevalence of each profile in a given sample consists of the sum of its grades of membership. The GoM software used in this project is an MS-DOS product under commercial development by the Duke University Center for Demographic Studies and is available through its webstite (http://cds.duke.edu).
Selection of Internal (Classification Generating) and External (Profile Validating) Variables
We used 85 DataPACE elements, measures of specific diseases, impairments, and disabilities, as internal variables (i.e., factors generating the GoM classification and indicating dimensions of participants' clinical status). Fifty-eight diseases were recorded from the medical problem lists in participants' charts and registered as present or absent. The Short Portable Mental Status Questionnaire (SPMSQ; Pfeiffer 1975) is performed at admission and quarterly. We stratified participants into SPMSQ error groups (0, 1–3, 4–6, and 7–10), each of which captured roughly 20–30%, in order to reduce the number of response probabilities (kl ). We retained DataPACE coding for the remaining impairment and disability variables. PACE teams assess participants for specific communication (expressive, receptive) and sensory (vision, hearing) impairments. These are categorized as totally impaired, partially impaired, or not impaired. Similarly, participants were assessed for the frequency of behavioral problems (wandering, verbal disruption, aggressive behavior, regressive behavior, and hallucinations): often (once/week or more), occasional (less than once/week), or never. Bladder and bowel incontinence were assessed as absent, infrequent, or frequent. Activities of daily living (ADL) of bathing, dressing, grooming, toileting, transferring, and feeding were measured as independent, requiring some help or supervision, or dependent. Walking was similarly measured. Use of mechanical assistive devices in these ADL was also registered. Instrumental activities of daily living (IADL) of meal preparation, shopping, housework, laundry, heavy chores, managing money, taking medications, and transportation were scored as independent, requiring some help, or dependent.
External variables are not used in estimating the K types but are conditionally estimated in association with them and can be applied to predictive validation. For this analysis, these variables included demographic data on participant age, gender, marital and ethnic/minority status, living arrangement and household composition, and education. We also estimated PACE tenure and self-rated health. Others included the number of household and nonhousehold informal caregivers, primary caregiver's age and sex, and type of informal support given (personal care, meal preparation, housework, money management, assistance with medications, transportation).
Profile Interpretation and Validation
The GoM quantitative solution was presented to a panel of experienced PACE clinicians in order to develop narrative profile descriptions. Specifically, we identified among internal variables those having greatest impact (denoted by relatively high values of the statistic H). We also constructed figures depicting response probability profiles on these factors. Using this information, the panel made predictions concerning response probabilities on unreviewed but related internal variables. For example, given a high impact score for SPMSQ errors, the panel might suggest that profiles characterized by many errors might be likely to have high probabilities of dementia. Working through the internal variables and their marginal and type-specific response probabilities, such clinical expectations were confirmed and other predictions made in branching fashion until extensive profile descriptions were obtained. Finally, the panel reviewed the fully described profiles for their clinical coherence (content validity).
We assessed predictive validity by constructing a series of figures depicting the conditionally estimated response probabilities (kl) on the external tenure, self-rated health, demographic, and informal support variables for the fully developed types. These figures are identical in format to those used in the clinical review procedure. Here, too, the PACE clinician panel was consulted to assess the coherence of the observed response probabilities.
|
Results |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
2 (
in the Wilson-Hilferty t-test approximation). The addition of a ninth type did not significantly improve the solution. Physician panel review for the solution's clinical coherence (see below) was consistent with content validity.
Table 1 displays selected information for the GoM solution. The eight profiles comprise the table's columns and are ranked by weighted prevalence. The row headings list internal variables together with the sample frequency/prevalence or marginal probability. Also linked to these internal variables are their solution impact factors (H). In the table body are the profile-specific response probabilities (kl) for selected internal disease, impairment, and disability variables. (A complete table is available from us.)
|
had any membership in Type 5, but 369 of these (12.7% of all participants) had Type 5 membership in excess of 0.5. The greater sample prevalence of Type 5 relative to Types 6–8 was accounted for by few participants having had equally high grades of membership in the latter. The clinician panel's knowledge of the nature, importance, and relationships among the represented disease, impairment, and disability variables is key to interpreting the GoM solution. First, we illustrate the process of generating clinical interpretations for these data. Then we summarize the results of panel review and associate these summaries with the profiles' conditionally estimated health, demographic, and social support characteristics.
Cognitive impairment, coded at levels of SPMSQ errors, had the largest impact factor 
; Table 1 ) and showed differentially distributed response probabilities (kl; Fig. 1). The probability was 100% that Types 3 and 4 have three or fewer SPMSQ errors, whereas the population prevalence of this few errors was 48.3% (Table 1 ). In contrast, for Types 5 and 7, the probability of more than six SPMSQ errors is 100%. The clinical review panel's first expectation was that cognitively impaired types would also be characterized by specific diagnoses and disabilities. These SPMSQ error likelihoods were in fact congruent with those for the prevalent dementia diagnosis (47.8% of PACE participants; Fig. 2), which also had high impact 
. For dementia, a probability of 100% was found for Types 1, 2, 5, and 7 and 0% for Types 3, 4, and 6. (The dementia question-relevance factor for Type 8 is 0, meaning that the diagnosis is not germane.)
|
|
The Dementia Series of Profiles
Type 1 was the most prevalent and was characterized by moderate cognitive impairment (Fig. 1 and Fig. 2). The profile was likely to require human help or supervision with grooming, bathing, and dressing but to be independent in toileting, transferring, walking, and feeding. These patients were dependent in IADL, including the cognitive tasks of money management and taking medications (Table 1 ). This profile had several salient associations with the conditionally estimated external variables (Table 2 ). First, compared with the population or marginal frequency, Type 1 had a high likelihood of being enrolled in PACE for less than 1 year (68.3% vs. 35.3%; see also Fig. 3). For self-reported health, Type 1 was less likely than Types 5 and 7 (more impaired dementia profiles) to have missing data and more likely than all responding PACE participants to report good or better health (81.7% vs. 56%). Relative to all participants, Type 1 was less likely to live at home alone and more likely to live at home with others or in a group home. The type was likely to receive informal help with medications.
|
|
Type 5 was highly cognitively impaired and likely to be fully dependent in IADL and ADL, except that Type 5 participants were partly dependent in feeding. Further, this profile was more likely to require feeding assistive devices (e.g., gastrostomy tubes). Impaired receptive and expressive communication was likely, as was frequent double incontinence. Type 5 was associated with long program tenure (68.6% for more than 2 years vs. 44.1% of all participants; Table 2 ). Regarding self-reported health, both Types 5 and 7 were likely not to respond. Type 5 was associated with living in a nursing home and was unlikely to live at home alone. Like Type 2, this profile had a higher than marginal likelihood of non-White ethnicity.
Type 7 had the severe cognitive impairment common with Type 5 but was also likely to exhibit a range of behavior problems and to carry diagnoses of psychosis and depression/anxiety (Table 1 ). This profile was associated with requiring feeding help and was unlikely to use ADL assistive devices. Type 7 was hearing impaired, impaired in expressive and receptive communication, and frequently bladder incontinent. Type 7 was likely to be older than the PACE population (64.0% older than 85 vs. 39.0% overall; Table 2 ). The type was highly associated with being White and was especially likely to live in nursing or group homes (81.6% vs. 7.1% for all participants). Consequently, Type 7 was less likely to have household caregivers or to receive various types of informal care (Table 2 ).
The Physical Disability Profiles
Type 3 describes cognitively intact participants, independent in ADL (Table 1 ). They were dependent in heavy chores and required help/supervision in transportation, meal preparation, shopping, and laundry but were more likely to be independent or requiring help only in taking medications and managing money. Among conditionally estimated demographic characteristics, Type 3 was younger, with a 49.7% likelihood of being younger than 75 vs. 26.3% for participants overall (Table 2 ). The Type 3 participant was very likely to live at home, especially alone. Type 3 was also somewhat less likely to be married and more likely divorced or never married. Correspondingly, Type 3 was more likely to have no identified household caregiver and was also unlikely to receive various forms of informal care, especially help with medications and money management, although these were needs identified for the profile. The small likelihood of receiving informal personal care was consistent with this type's relative ADL independence.
Type 4—like Type 3—participants were cognitively intact but had increased likelihood of chronic, disabling diseases other than dementia, for example, diabetes and other endocrine, nutritional, and metabolic diseases; depression/anxiety; hypertension and cerebrovascular disease; and arthritis and other musculoskeletal conditions, as well as other unspecified diseases (Table 1 ). These participants were independent feeders but required help with grooming, bathing, and dressing. Like Type 2, they were more likely to use a variety of ADL assistive devices. They required help with meal preparation, money management, and medications but were fully dependent in other IADL. This type was associated with longer PACE tenure (64.8% likelihood of more than 2 years vs. 44.1% marginally; Table 2 and Fig. 3). Type 4 was also associated with fair or worse self-reported health (69.8% likelihood vs. 44% for all participants). This was the youngest profile (65.8% likelihood younger than 75 vs. 26.3% for all participants). Type 4 was somewhat more likely to be Hispanic and less likely to be White. Type 4 was similar to Type 3 in its greater likelihood of living at home, especially alone, but was not as likely as Type 3 of being deprived of household caregivers. Despite their IADL needs and average caregiver availability, Type 4 participants were less likely to receive informal help with medications, transportation, and money management.
The "Chronic Illness" Profiles
Type 6 participants were likely to have more than a few SPMSQ errors, but—unlike Types 1 and 2—were unlikely to be diagnosed with dementia (Table 1 , Fig. 2). Like Type 4, this profile was likely to have hypertension, arthritis, and depression/anxiety. Other chronic conditions were also likely (e.g., cancer, anemia, ear diseases, esophageal reflux, chronic constipation, osteoporosis, fractures, and medication allergies). Type 6 participants were visually and hearing impaired, were independent in feeding, and were likely to use ADL assistive devices. Type 6 required help with medications and was dependent in other IADL. Type 6 participants, like Type 8, had particularly salient health, demographic, and support characteristics (Table 2 ). This was the longest tenured type, with enrollment for less than 1 year very unlikely (kl = 0). Fair to poor self-rated health was more likely for Type 6 than for the general PACE population (62.7% vs. 44%). This was the oldest profile (81.1% older than 85 vs. 37.7 overall), comprising widowed White women, likely to live in group homes or alone. Although household caregivers were unlikely, Type 6 was more likely to have multiple non-household informal caregivers visiting once or twice weekly. Nevertheless, Type 6 was not more likely than participants overall to receive any type of informal care and was much less likely to receive informal help with meals, medications, or personal care.
Type 8 presented a mixed chronic disease picture, with high probabilities for a range of diagnoses, including diabetes, depression, heart diseases, hypertension, cerebrovascular disease, chronic obstructive pulmonary disease, renal failure, and prostate disease (Table 1 ). Dementia (as well as basic ADL and some additional internal variables) was not germane in generating this profile. However, like Type 6, a moderate number of SPMSQ errors were likely (Fig. 1). These patients required help in feeding and were dependent in most IADL. They were more likely than others to use a urine catheter. Like Type 6, this long-tenured profile was likely to claim fair or poor health (Table 2 , Fig. 3). Demographically, however, Type 8 was quite distinctive, consisting of younger-old, less educated, non-White men, tending to be married, but very likely living at home with others. Correspondingly, multiple household caregivers were very likely, as was informal personal care and IADL assistance (except, notably, money management).
|
Discussion |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
The conditionally estimated demographic, health, and informal support probabilities support the profiles' predictive validity. First, the more disabled profiles in each series were associated with long PACE tenure and older age, reflecting aging in place of the population. Both the pattern of missing data and the probabilities for different levels of self-rated health were also congruent with the profiles' differing cognitive, illness, and disability properties. Although the relationship between the differing levels of need captured in the profiles on the one hand, and the demographic and informal support probabilities on the other, was more complex, it also supports the classification's predictive validity. For example, as would be expected, living at home alone was unlikely for the dementia profiles (relative to the overall population and especially to the physical disability/chronic illness types), with the probability of living in group and/or nursing homes increasing for the more regressed types. In contrast, living at home alone and/or with others was more likely for the nondemented profiles.
The likelihood of receiving the various types of informal support was related both to profiled needs, the profiles' domestic situation characteristics, and concurrent provision of PACE services. For instance, informal help with medications—a cognitive IADL dependency—was more likely for Type 1, the early- to midstage dementia type, living at home with others. Such help was much less likely for Type 7, a late-stage type highly likely to live in group or nursing homes, and most of the nondemented profiles, which were both unlikely to have the need for this help and unlikely to have household informal caregivers.
An underpinning of the model is the grading of individual participants' membership into more than one profile. A given individual might have partial membership in Type 6 due to certain characteristics (e.g., being depressed, having osteoporosis, hip fracture, urinary tract infection, etc.) best fitting a probability sequence of that profile, but having other characteristics (e.g., impairments and disabilities) better fitting to one or more others. Consider Type 3, which in its complete expression describes individuals with fewer care needs than the nursing-home level required for PACE enrollees. Even among the small minority with high membership in this profile, partial memberships in other types may be more than sufficient to qualify them for PACE in their state. Manton and colleagues 1995 study of nursing home demonstration residents identified a similar, less ill and disabled profile (stable), which had a higher weighted prevalence than our Type 3 (23.3% vs. 15.8%) and was the most prevalent of their 11 resident types (Lamb and Willert 1999).
This article presents a first step in our analysis of the PACE population. We clinically profiled persons participating in PACE in early 1997. The profiles were based solely on indicators of diseases, impairments, and disabilities (a selection of which are presented Table 1 ). Using the external variables, we examined the profiles' predictive validity. In Table 2 , we noted the high probability of certain types belonging to specific ethnic minority groups. This reflects the differing ethnicity of dually eligible elderly persons in different PACE markets (e.g., Asian Americans in downtown San Francisco, Hispanics in El Paso and the Bronx), and the variation in the weighted prevalence of profiles across sites. We are now extending our work to consider the probability of PACE service responses to the profiles (e.g., hours per month of in-home personal care). In this work, we analyze discrepant response probabilities (a profile having high probabilities of both no hours and very many hours of in-home care) in terms of different approaches to care taken at those PACE sites at which the profiles in question are especially prevalent. In identifying participant subgroups having such variable care inputs, we can pose questions of cost effectiveness by linking to available cost and outcome information, contributing to development of evidence-based standards of PACE care as the program expands nationwide.
|
Acknowledgments |
|---|
Received for publication February 17, 1999. Accepted for publication December 7, 1999.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  S. M. Friedman, D. M. Steinwachs, P. J. Rathouz, L. C. Burton, and D. B. Mukamel Characteristics Predicting Nursing Home Admission in the Program of All-Inclusive Care for Elderly People Gerontologist, April 1, 2005; 45(2): 157 - 166. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
