This Article Abstract Full Text (PDF) Services Download to citation manager Citing Articles Citing Articles via HighWire PubMed PubMed Citation

Predicting 12-Month Mortality for Persons With Dementia

Departments of ¹ Social &
Behavioral Science, ² Medicine
⁴ Psychiatry, University of California, San Francisco.
³ Clay Software and Statistics, Ashland, Oregon.

Address correspondence to Robert Newcomer, Department of Social & Behavioral Sciences, University of California, 3333 California Street, Suite 455, San Francisco, CA 94118. E-mail: rjn{at}itsa.ucsf.edu

	Abstract

TOP Abstract Methods Results Discussion and Conclusions References

 
Objectives. We develop and test a model of 12-month mortality among persons () with organic dementia.

Methods. Data are from caregiver interviews and claims records for persons enrolled in the Medicare Alzheimer's Disease Demonstration Evaluation. Information covers the year prior to enrollment through 36 months. We used Proportional hazards models to predict time to death. We estimated two starting points: first, the date of randomization, and second, the date of maximum difficulty in performing two or more activities of daily living (ADLs).

Results. The final model includes age, gender, ADL difficulty, medical conditions, prior year hospitalizations, and whether a daughter was the primary caregiver. We combined hazard ratios to produce a cumulative mortality risk score. Model discrimination is reasonable for both models (c statistics of.72 and.69, respectively), and calibration tests were nonsignificant.

Discussion. The model's efficiency, as measured by the ratio of false positives (those predicted to die, but who lived) to true positives (those predicted to die and who did die) ranged from fewer than 1:1 to more than 4:1 as the model's sensitivity increased. This ratio was lower in the two or more ADL difficulty model. A validation test of the prediction model found comparable sensitivity and specificity (c statistic of.69) to the reference model.

Although patients with dementia have considerably higher mortality rates than age-matched controls, survival in patients with dementia is highly variable. A prognostic measure that could stratify patients with dementia into groups at differential risk for death would be useful to both clinical providers and policy makers. For example, such a measure would be helpful to providers who counsel patients and their families about advance planning and palliative care options such as hospice. In addition, a prognostic measure would be useful for risk adjustment, either to compare outcomes across different groups of providers or to more fairly compensate health systems that care for a greater proportion of higher risk patients.

Our model for the prediction of mortality in patients with dementia is guided by a conceptual framework, adapted from Iezzoni (1997). She views poor outcomes such as mortality as mediated by multiple domains of risk factors that include demographic, biomedical, and social factors. Risk factors signified by disease (or comorbidity), social factors, functional status, and health behaviors interact with each other and the effects of age and gender to increase the risk for mortality. The theoretical basis for Iezzoni's work is formed in part by a biopsychosocial model of health and illness. This perspective, as postulated by George Engel (1977), complements the prevailing biomedical formulations of illness severity (which focus primarily on disease diagnoses) by including the notion that the patient, the social context in which he lives, and the system devised by society to deal with illness should all be considered. When one views illness severity and prognosis through the eyes of the patient as in the biopsychosocial perspective, it becomes apparent that illness severity and prognosis are to some extent socially determined and must include dimensions important to patients, such as functional status (Covinsky & Landefeld, 1996). It is also apparent that social and environmental issues interact with biology, health practices, and individual definitions of health.

The importance of considering the multiple domains postulated by the biopsychosocial model (including those used by Iezzoni) when defining health status and predicting health outcomes in older people is well established by empirical work. For example, across all age ranges, men have higher rates of mortality than women (Fried et al., 1998). The number and severity of comorbid conditions are associated with mortality. Particular comorbid conditions such as congestive heart failure are often important components of comorbidity indices (Charlson, Pompei, Ales, & MacKenzie, 1987). Physical function, such as the ability to perform activities of daily living (ADLS) without assistance, has also been shown to predict mortality independently, and it may be at least as important a predictor as diagnoses. One example of this work is that ADL function has been shown to strongly predict mortality in hospitalized elders (Covinsky, Justice, Rosenthal, Palmer, & Landefeld, 1997; Walter et al., 2001). Within the current sample of persons with dementia, when comorbidity and other risk factors were held constant, the rate of 12-month mortality was 10% among those with no ADL limitations and more than 20% among those having two or more limitations.

The quality of the social networks available to an elder, another element in Iezzoni's framework, may be as important a predictor of health outcomes as traditional biomedical predictors. For example, Berkman reported that among older people hospitalized with myocardial infarction, lack of emotional support was associated with a threefold higher mortality over a 6-month period (Berkman, Leo-Summers, & Horwitz, 1992). The importance of social support is particularly important in patients with dementia, who are heavily reliant on the care of others. Socioeconomic status (SES), whether measured by education, income, occupation, or insurance status, has also been consistently associated with mortality and functional loss (Adler, Boyce, Chesney, Folkman, & Syme, 1993). Minority ethnicity has also been demonstrated to be a risk factor for higher mortality, although this is an effect only partially explained by SES (Guralnik, Land, Blazer, Fillenbaum, & Branch, 1993).

Although the discussion just given and most of our analysis focus on baseline risk factors, adverse outcomes in older people are probably the manifestation of both baseline risk factors and precipitating events. In older people, hospitalization is often used as a model of precipitating events (Gill, Williams, & Tinetti, 1999). For example, Morrison has demonstrated that in patients with end-stage dementia, mortality rates are very high following hospitalization for hip fracture or pneumonia (Morrison & Siu, 2000). Within the current sample, persons with dementia having a hospital stay for any cause had almost twice the mortality rate of those without a hospital stay. This rate increased fourfold to almost 50% among those discharged to nursing homes.

The development of any prognostic model is limited by the elements available in the data set. For example, we were not able to consider all of the risk domains (e.g., smoking) that have been formulated and tested by Iezzoni and others. However, the use of her framework informed our use of available elements and informs our thinking about how future studies may improve this model.

Mortality studies typically use an index event, such as a hospital stay or date of diagnosis, to provide the starting point or time zero for the time frame of interest. However, an index event such as one of these ignores information about the natural history and management of dementia that may have contributed to the hospital stay. An alternative event, the loss of independence in ADLs, has been suggested as a potentially more appropriate index event for persons with dementia (Gauthier, 1998). Our analysis examines whether this crucial outcome for dementia patients and their caregivers may be an effective start date in the prediction of mortality.

	METHODS

TOP Abstract Methods Results Discussion and Conclusions References

 
Data came from caregiver interviews and claims records compiled for persons enrolled in the Medicare Alzheimer's Disease Demonstration Evaluation (MADDE). This was an eight-site program funded by the Centers for Medicare & Medicaid Services (CMS). It operated from December of 1989 through November of 1994. MADDE eligibility included a physician-certified diagnosis of dementia (e.g., Alzheimer's disease, vascular dementia, dementia caused by degenerative diseases, infections, and trauma), eligibility for Parts A and B of the Medicare program, residence in a program catchment area, and not living in a nursing home at time of enrollment. Dementias resulting from tumors, toxins, drugs, or nutritional or psychiatric disorders were excluded. The intervention and control groups have been combined here, as the case management and home care service intervention has been shown to have no association with mortality risk or nursing home placement (Miller, Clay, Fox, & Newcomer, 1999), or with caregiver outcomes such as depression and burden that could be thought to influence the quality of informal care (Newcomer, Yordi, DuNah, Fox, & Wilkinson, 1999).

Excluded from the combined sample of 8,108 MADDE participants were (a) those who enrolled as managed care members () at program enrollment or during the year prior to enrollment; (b) those with no match to the Medicare eligibility claims file (); (c) those with less than 6 months of Medicare eligibility in the prior year (); and (d) those whose primary caregiver was either a group home operator or a paid provider and for whom no family member was defined as the primary caregiver (). The exclusion of managed care, those without a match to claims records, or less than 6 months of Medicare eligibility was necessary for the present analysis because we needed Medicare claims data to track hospital stays and diagnoses. The paid providers (which includes group home operators as well as live-in aides) were omitted because this group was too small for separate analysis.

We have divided the individuals included in the analysis into two cohorts: those who enrolled into MADDE between December of 1989 and March of 1991 () and those who enrolled between April and November of 1991 (). We interviewed the initial cohort at baseline (in their homes) and semiannually (by telephone) for 24 months (if in the community), and then at 36 months. We obtained assessment data for the second cohort only at baseline, and at 24 and 36 months. We used the initial sample to develop the mortality risk prediction model. We used the second cohort here in validation tests of the prediction model.

We obtained information on subject and primary caregiver attributes from assessment interviews conducted with the primary caregivers, and a cognitive screening examination conducted with the program applicants. Linked to these attributes were the subjects' Medicare claims, and, if applicable, date of permanent nursing entry and date of death. A primary caregiver was the one person who provided the "most assistance" to the person with dementia. No cases in the analysis were lost from claims or the observation of death, as we compiled Medicare eligibility and claims files for at least 36 months after study enrollment for all participants.

Measures and Prediction Model Specification
The dimensions of risk are shown in Tables 1–3. We have used two alternative measures to define the time zero or starting period for the prospective analysis of mortality risk: first, the date of randomization into the demonstration, and second, the date when the caregiver reported that the subject had maximum difficulty performing at least two of the following ADLs: bathing, dressing, eating, grooming, toileting, transferring, or walking (Katz, Ford, Moskowitz, Jackson, & Jaffee, 1963). We do not expressly consider date of institutionalization (defined by a permanent nursing home entry date), but we do retain cases in which subjects entered nursing homes after time zero in the analysis. This represents approximately 25% of the study sample, and just over 30% of those who died in the observation year.

We divided counts of subject behavior problems (Zarit, Reeves, & Bach-Peterson, 1980), such as asking repetitive questions, being suspicious or accusative, having trouble recognizing familiar people, and engaging in behavior potentially dangerous to self or others, into groups. This was to accommodate any curvilinear association between behavior problems and the risk of death. Another measure of cognitive status was the Mini-Mental State Examination (MMSE). The MMSE assesses orientation, recall, and ability to name objects. It was scored on a 30-point scale (Folstein, Folstein, & McHugh, 1975) and then grouped to address curvilinear relationships.

All measures for the randomization group were obtained at baseline. Among those having two or more maximum difficulty ADL limitations, 67% had this status at baseline.

Two other sets of measures are included in Table 1. These are the communities from which the sample was compiled and whether the case was in the MADDE treatment or control group. These measures were included in the initial Cox proportional hazards models as contextual controls, but they failed to retain statistical significance in the final model.

Caregiver Attributes
Caregiver measures are shown in Table 2. These include relationship to the subject, the caregiver's individual attributes, and items reflecting possible consequences of caregiving. These items were obtained at baseline for the randomization date group, and they were updated to reflect the information current at the time of the two maximum difficulty ADL classification. Caregivers' ADL limitations and IADLs (Lawton & Brody, 1969) used a slightly different set of activity items from the subject ADL and IADL scores, and asked whether "some" or "no" (but not "maximum") difficulty was experienced. The presence of some difficulty was used to count the number of limitations. Caregiver burden was measured by the 7-item Zarit Burden Scale (Zarit et al., 1980). Responses range from 0 = "never" to 4 = "almost always." The scale score was grouped into approximately equal aggregations as a way to adjust for any curvilinear relationships with mortality or nursing home placement (McCarty et al., 2000). Caregiver depression, measured by the Geriatric Depression scale, includes 15 items (Yesavage, Rink, Rose, & Aday, 1983).

The MADDE data have two sources of information on chronic health conditions. One of these is a listing of conditions included on the physician referral form required as part of the application for MADDE participation. This was completed by each subject's physician, but its emphasis was on delineating the nature of the dementia rather than other conditions. A more inclusive source of diagnoses is Medicare claims. Preliminary analyses found that claims-identified condition prevalence, except for diabetes, was more highly associated with mortality. The prevalence and mortality rates associated with a selected list of conditions are shown in Table 3. These conditions reflect the array of chronic conditions included in the Charlson Index (Charlson et al., 1987) and other recent analyses of mortality risk in the aged (e.g., Fillenbaum et al., 2000). A number of studies have tested the predictive validity of the Charlson Index and similar weighting processes and extended the application from medical records to claims records (e.g., Cleves, Sanchez, & Draheim, 1997; Deyo, Cherkin, & Ciol, 1992; D'Hoore, Bouckaert, & Tilquin, 1996; Kieszak, Flanders, Kosinski, Shipp, & Karp, 1999).

Hospital Use
An additional adjustment for condition classification was a measure counting the number of inpatient stays in the 12 months immediately preceding time zero. The hospital use distribution is shown in Table 3. The enrollment process into the MADDE program created a systematic truncation of hospital effects on mortality risk for those cases measured prospectively from date of randomization. Individuals could not enroll into MADDE if they were in a hospital. They had to be living in the community. Thus mortality occurring during or immediately following an inpatient stay is underrepresented in the MADDE enrollment. Hospital stays after date of enrollment were not constrained in this manner.

Statistical Methods
We modeled survival time in months by Cox proportional hazard regression, using PROC PHREG in SAS Version 8.2. We censored time at 12 months following the starting date, which was defined in two ways, as already discussed. We performed the initial model-reduction process by using backward stepwise procedures to address any collinearity among measures. We performed the model validation using the April–November MADDE enrollment cohort by applying the parameter estimates fitted on the main cohort. We assessed model calibration by comparing predicted deaths versus actual deaths within deciles of probability of death. This was tested with the Hosmer–Lemeshow chi-square test.

	RESULTS

TOP Abstract Methods Results Discussion and Conclusions References

 
The prevailing unadjusted likelihood of death among the mortality prediction sample was 14.3% when date of randomization was used as the starting point, and 20.9% when time zero shifted to the presence of at least two ADLs where maximum assistance was needed. Three sets of findings are presented. These attempt to broaden and refine these unadjusted estimates. The initial analyses used Cox proportional hazards methods to derive multivariate models of mortality. The linear parameter estimates obtained in these models were then combined into a cumulative risk score for each participant. The sensitivity and specificity of the derived risk scores and their relationship with mortality is tested in a second analysis. A third analysis assesses the predictive value of the risk scores by applying them to a second cohort of MADDE enrollees (i.e., those enrolling between April and December of 1991).

Proportional Hazard Models of Mortality
Table 4 shows the results for two models representing alternative approaches to measuring the effect of functional disability on the likelihood of death within 12 months. One uses the randomization date; the other uses the presence of at least two ADLs with maximum performance difficulty. We adjusted both models for the simultaneous risk associated with the subject, caregiver, and other attributes shown in Tables 1–3. Only those sets of measures having a significant association with the mortality are shown.

One caregiver attribute emerged as a risk factor. Daughters as caregivers were associated with a lower subject mortality risk, independent of subject age and gender, than were spouses. The final factor in the model is the history of inpatient utilization. This measure serves as a proxy measure for severity of illness. Three or more prior year inpatient stays were associated with an increased hazard of death. The absence of similar effects with recent hospital stays may be an artifact of the sample exclusion criteria noted earlier.

Predicting Risk of Mortality
The risk factors in the final models were combined to produce a cumulative risk score for each individual in the sample. We did this by summing the linear predictor (i.e., the logarithmic form of the hazard ratios) for each of the measures applicable to the case. We then tested these summed risk scores (here converted back to hazard ratios) relative to the likelihood of death. We used a series of cut points in the risk score to divide the population into high-risk cases predicted to die or low-risk cases predicted to live. This prediction has the following components:

• true positive cases (those predicted to die who do die) = A,
  
• false positive cases (those predicted to die who live) = B,
  
• false negative cases (those predicted to live who die) = C, and
  
• true negative cases (those predicted to live who live) = D.
  

Sensitivity and specificity are commonly used measures of how well a model discriminates or correctly distinguishes persons who live from those who die. In this case, sensitivity [A/()] is the percent of deaths correctly classified by our prediction rule (true +); specificity [D/()] is a measure of the percent of cases correctly classified as living (true -). A particular advantage of sensitivity and specificity measures is that their values are unaffected by the prevalence of death in the population (Ash & Shwartz, 1997).

The columns in Table 5 show the distribution of the sample cases among those predicted to die and their outcomes relative to death or remaining alive after 12 months. The rows show the distribution of risk scores divided into cut points based on approximately 10% increments in sensitivity. Evident from the table is a relatively constant trade-off between sensitivity and specificity, with the gain in sensitivity being achieved with a growth in the number of false positive cases (i.e., more persons are predicted to die than do die). In general, the randomization time-zero model performs marginally better than the any two maximum difficulty ADL model, with higher specificity and somewhat fewer false positive cases as sensitivity increases. The performance of these models is reflected in a c statistic, which is a comparison of the proportion of pairs in which the predicted probability of death is higher for the patient who died than for the patient who lived. Each tied pair is counted as one half. A c statistic also equals the area under a receiver operating characteristic curve. Both time zero models perform relatively similarly, although the randomization date model has a slightly higher c statistic (0.72 vs. 0.69). These values are consistent with c statistics reported from a number of risk-adjusted mortality models (e.g., Ash & Shwartz, 1997), and they reflect a 2 standard deviation discrimination improvement by both models over chance, which would be reflected by a c statistic of 0.5. There is approximately a 0.10 standard deviation difference between the two time-zero models.

The validation sample results (the third panel in Table 6) show somewhat lower discrimination () and calibration than those of the prediction model. Calibration is particularly problematic at deciles 4 and 6. The net effect is that over most of the risk deciles, the validation sample underestimates actual deaths by approximately 6%. These differences suggest some overfitting of the models and problems with transportability. However, the reduction in discrimination and the calibration problems are consistent with what is usually seen with predictive models. One possible explanation for the differences in model calibration examined was the geographic distribution of the sample cases, with location perhaps serving as a proxy for practice patterns and service supply. Expressed as a percentage of total subjects, the proportion of cases varied by less than 2% between the two samples among four of the sites, and by up to 6% among three sites.

	DISCUSSION AND CONCLUSIONS

TOP Abstract Methods Results Discussion and Conclusions References

 
This article has identified factors associated with mortality over 12 months. The analysis used 3 years of cohort data from a sample of persons known to have some form of organic dementia. Mortality over 12 months was 14.3% from the date of enrollment into the MADDE demonstration, and 20.9% among those having maximum difficulty performing two or more ADLs. The individual-level attributes of age, gender, the need for maximum assistance in ADLs, and selected medical conditions (i.e., congestive heart failure, chronic obstructive pulmonary disease, cancer, and diabetes) were each associated with an increase of 50% or more in the likelihood of mortality (i.e., a hazard ratio greater than 1.5). The number of prior year hospitalizations (an indicator of illness severity) was also associated with an increased likelihood of death. A daughter as the primary caregiver, rather than the spouse, was associated with reduced mortality.

These measures were combined to produce a cumulative risk score. This score was used to predict 12-month mortality. Its effectiveness, as measured by the ratio of false positives (those predicted to die, but who lived) to true positives (those predicted to die and who did die) ranged from fewer than 1:1 to more than 4:1 as the model's sensitivity increased. The ratio of false positives to true positives was generally lower when the prospective starting point for predicting mortality was defined as having maximum difficulty performing two or more ADLs, rather than when the start time of enrollment into the study was used, adjusting for the number of ADL limitations.

The validation test of the model was rigorous because it tested the model in subjects with some difference in their geographic distribution, and who enrolled at a different point of time than the subjects in the development sample. As a result, the validation tests not only the degree of model overfitting, but the degree of geographic and temporal transportability of the model (Justice, Covinsky, & Berlin, 1999). In spite of this rigorous validation test, there was only a moderate drop in the discrimination of the model and a moderate fall in the calibration of the model. The validation application underpredicted actual mortality by an average of 6%, or 15 cases out of 245 deaths. This suggests that our model may be useful in predicting mortality among patients with dementia in other geographic settings. Application among group home residents, those with paid care providers, and those after nursing home entry has not been tested.

These findings are encouraging. Moreover, the predictive efficiency of the models can likely be increased in future applications where limitations of the MADDE data set can be rectified. For example, approximately two thirds of the demonstration enrollees had at least two ADL limitations at the time of enrollment. It is possible that establishing a more precise measurement of the onset of maximum ADL dependency would improve predictive reliability. However, whether simply having more accurate information on the movement from one to two or two or higher limitations will be informative versus knowing the onset of particular limitations such as being unable to eat or transfer without maximum assistance has not been tested. What was observed is that mortality rates are higher among those with more ADL limitations.

A second area of limitation is the measurement of chronic conditions. A diagnosis establishes the types of care and resources required to treat the illness, and for this reason diagnoses have been a focal point for the development of risk-adjustment methods. The current analysis attempted to enumerate all known conditions, but claims sources likely underenumerated the conditions. Further, among the identified conditions, we were not able to distinguish primary disease process (other than dementia) from coexisting conditions; we also could not adjust for "severity." Studies having access to medical records would be able to more reliably determine the number of conditions, and possibly delineate the level of severity or stage, or the conditions directly associated with death. Such information could change the mortality risks for selected conditions over those reported here.

A third issue is more subtle, but it is related to the timing associated with diagnosis or functional limitation; it involves the selection of cases or patients. The MADDE sample had an inherent selection bias among program enrollees. All had to be living in the community (and presumably medically stable enough for the physician and families) to apply for program admission. This selection rule diminished the salience of the hospital stays in the weeks prior to admission, and it likely undercounted the proportion of the population with behavior and other challenges for which a nursing home placement was eminent. Although this is a possible limitation to the study findings, it is also important to consider that the selection of cases into any program or benefit will have to confront similar enrollment decision rules. Cases with eminent nursing home placements or with very high mortality risk (such as those who die during their hospital stay or shortly afterward) would likely be excluded from a palliative care or end-of-life risk-adjusted payment oriented to people in the community. Similarly, those without family caregivers might also be excluded, as they were in these analyses. In other words, replications and extensions of the present analyses have to recognize that the decision-making process permitting application or admission to a program could in various ways affect the salience of time-varying risk factors. In the current analyses, we likely underrepresented the prevalence of disease or condition severity in the dementia population (by the exclusion criteria). If this is true, then mortality risk may be higher in the presence of the identified risk factors than was represented by these analyses.

A fourth issue is that of caregiver attributes and their collinearity with each other. The models shown used caregiver relationship and found that the presence of a daughter as the primary caregiver lowered the risk of mortality compared with the presence of spousal caregivers. The underlying cause of this effect is not known. It may reflect a better ability to provide both instrumental assistance and disease management, or the attainment of better management through earlier nursing home placement than when there is a spousal caregiver. Among the expectations of a nursing home placement are close monitoring of medical conditions and drug management. Those in the MADDE sample entering nursing homes from the community (almost half of those entering nursing homes) had approximately a 40% lower 12-month mortality risk than the other half of the sample who entered nursing homes following a hospital stay. An effort was made in the analysis to unbundle the effects embedded within caregiver relationships. Those directly measured in the data set included education, health status, and age. There is overlap between relationship and caregiver age (highly correlated with spouse on the high end and with daughter on the low end). Age and education are also collinear, with older individuals generally having less education than their children. Caregiver education, age, and health status were separately considered in the original models, but they were not retained by the backward stepwise process that also included relationship. Attributes that may distinguish daughters from spouse caregivers, such as the comparative tolerance for the burdens of caregiving, were not directly measured. These are inferred from outcomes such as nursing home placements. Caregiver-perceived stress and burden was directly measured, but this measure was not associated with mortality risk, perhaps because of differential burden tolerance among caregivers. In any clinical application of the model, it would be prudent to consider caregiver physical and cognitive capabilities, educational level, and the ability to read and follow care management instructions. Further interacting with all of this are living arrangements and relationship. Spouses living with the subject likely have greater tolerance for the caregiver tasks than does a nonspouse, or caregivers not living with the subject. In short, the strengths and limitations of the caregiver must be appreciated and understood for their potential effects on patient outcomes, as well as for the nature and intensity of support that may be required.

A final issue is understanding the importance of age, gender, and race in the predictive models. We have presented these measures as individual-level attributes, but they potentially reflect dimensions that are subject to manipulations. For example, two dimensions are potentially adjusted for when age is controlled for. One of these is the physiological changes associated with aging. Increasing age increases the risk of death and the risk of complications from many procedures and treatments (see, e.g., Mangione et al., 1993). A second dimension is that of potential "ageism" in therapeutic choices for elderly patients. Separating ageism from changes in preferences among older patients further complicates the interpretation of treatment differences across age groups or by cognitive status.

Women and men differ anatomically, physiologically and hormonally. Whether the observed differences are a function of physiology or of patient and family preferences has not been resolved. Stratifying analyses by gender and documenting treatment selection decisions would be a possible means for addressing these concerns in future medical effectiveness studies. Several parallels arise when race and ethnicity are viewed as risk factors. In addition to well-documented differences in disease prevalence, there is the concern that ethnic minorities may have preferences that lead to differences in treatment selection. Such differences were not documented in the MADDE data set. The data used did separately measure income and education, but any cultural differences in health practices and behaviors (e.g., smoking, blood pressure, cholesterol level, body-mass index, and alcohol intake) were not measured. Analyses of the effect of race or ethnicity on outcomes of care can be better accomplished in results stratified by race and ethnicity, and where behaviors, practices, preferences, and treatments are documented.

Although a number of limitations and further measurement refinements have been posited for the mortality prediction model, the current model's potential salience should not be ignored. First, a prognostic measure that can stratify patients with dementia into groups at differential risk for death is useful to providers counseling patients and their families about advance planning and palliative care options, such as current treatment and future interventions; decisions about hospitalization; and referrals for nursing home care, hospice, or home health care. Our model identifies a combination of factors that contribute to an end-stage prognosis. A clinician using this model would have a starting prognosis and treatment plan that could be modified for individual patients on the basis of their comorbid condition severity. The main point is that the current work opens up an examination of the issue of end-of-life care for persons with dementia.

Second, the measures in this model are available in many administrative systems, as well as national and catchment area surveys. This lends itself to replication tests (such as those suggested previously) and application in clinical trials or outcomes studies designed to test the efficacy of interventions directed to those with dementia at the end of life, or studies of community-support programs directed at the caregivers of those at the end of life.

	Acknowledgments

 
This article was supported in part by Grant IIRG-01-2390 from the Alzheimer's Disease and Related Disorders Association, and Contract 500-89-0069 from the Centers for Medicare & Medicaid Services. The authors remain solely responsible for the conclusions presented.

	Footnotes

 
Decision Editor: Charles F. Longino, Jr., PhD

Received for publication July 31, 2002. Accepted for publication December 2, 2002.

	References

TOP Abstract Methods Results Discussion and Conclusions References

 

• Adler, N. E., Boyce, W., Chesney, M., Folkman, S., Syme, S. (1993). Socioeconomic inequalities in health. No easy solution. Journal of the American Medical Association, 269,3140-3145.[Abstract/Free Full Text]
• Ash, A., Shwartz, M. (1997). Evaluating the performance of risk-adjustment methods: Dichotomous outcomes. In L. I. Iezzoni (Ed.), Risk adjustment for measuring healthcare outcomes (pp. 427–469). Chicago: Health Administration Press.
• Bennett, C., Greenfield, S., Aronow, H., Ganz, P., Vogelzang, N., Elashoff, R. (1991). Patterns of care related to age of men with prostate cancer. Cancer, 67,2633-2641.[Medline]
• Berkman, L. F., Leo-Summers, L., Horowitz, R. (1992). Emotional support and survival after myocardial infarction. A prospective, population-based study of the elderly. Annals of Internal Medicine, 117,1003-1009.
• Charlson, M., Pompei, P., Ales, K., MacKenzie, R. (1987). A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. Journal of Chronic Disease, 40,373-383.[Medline]
• Cleves, M., Sanchez, N., Draheim, M. (1997). Evaluation of two competing methods for calculating Charlson's comorbidity index when analyzing short-term mortality from administrative data. Journal of Clinical Epidemiology, 50,903-908.[Medline]
• Covinsky, K. E., Justice, A. C., Rosenthal, G. E., Palmer, R. M., Landefeld, C. S. (1997). Measuring prognosis and case mix in hospitalized elders. The importance of functional status. Journal of Internal Medicine, 12,203-208.
• Covinsky, K. E, Landefeld, C. S. (1996). Using the biopsychosocial model in practice: Improving the care of patients with late life depression. Journal of General Internal Medicine, 11,249-250.[Medline]
• Deyo, R., Cherkin, D., Ciol, M. (1992). Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. Journal of Clinical Epidemiology, 45,613-619.[Medline]
• D'Hoore, W., Bouckaert, A., Tilquin, C. (1996). Practical considerations on the use of the Charlson comorbidity index with administrative data bases. Journal of Clinical Epidemiology, 49,1429-1433.[Medline]
• Engel, G. L. (1977). The need for a new medical model: A challenge for biomedicine. Science, 196,129-136.[Abstract/Free Full Text]
• Fillenbaum, G., Pieper, C., Cohen, H., Cornoni-Huntley, J., Guralnik, J. (2000). Comorbidity of five chronic health conditions in elderly community residents: Determinants and impact on mortality. Journal of Gerontology: Medical Sciences, 55A,M84-M89.[Abstract]
• Folstein, M., Folstein, S., McHugh, P. (1975). Mini-mental state: A practical method for grading the cognitive state of patients for the clinician. Journal of Psychiatric Research, 12,189-198.[Medline]
• Fried, L. P., Kronmal, R., Newman, A. B., Bild, D. E., Mittelmark, M. B., Polak, J. F., et al. (1998). Risk factors for 5-year mortality in older adults: The Cardiovascular Health Study. Journal of the American Medical Association, 279,585-592.[Abstract/Free Full Text]
• Gauthier, S. (1998). Update on diagnostic methods, natural history and outcome variables in Alzheimer's disease. Dementia and Geriatric Cognitive Disorders, 9,(Suppl. 3), 2-7.
• Gill, T. M., Williams, C. S., Tinetti, M. E. (1999). The combined effects of baseline vulnerability and acute hospital events on the development of functional dependence among community-living older persons. Journal of Gerontology: Medical Sciences, 54A,M377-M383.[Abstract]
• Guralnik, J. M., Land, K. C., Blazer, D., Fillenbaum, G. G., Branch, L. (1993). Educational status and active life expectancy among older blacks and whites. New England Journal of Medicine, 329,110-116.[Abstract/Free Full Text]
• Iezzoni, L. I. (1997). Dimensions in risk. In L. I. Iezzoni (Ed.), Risk adjustment for measuring healthcare outcomes (pp. 43–168). Chicago: Health Administration Press.
• Justice, A. C., Covinsky, K., Berlin, J. (1999). Assessing the generalizability of prognostic information. Annals of Internal Medicine, 130,515-524.[Abstract/Free Full Text]
• Katz, S., Ford, A., Moskowitz, R., Jackson, B., Jaffee, M. (1963). Studies of illness in the aged: The index of ADL, a standardized measure of biological and psychosocial functioning. Journal of the American Medical Association, 185,94-101.
• Kieszak S., Flanders, W. D., Kosinski, A., Shipp, C., Karp, H. (1999). A comparison of the Charlson comorbidity index derived from medical record data and administrative billing data. Journal of Clinical Epidemiology, 52,137-142.[Medline]
• Lawton, M., Brody, E. (1969). Assessment of older people: Self-sustaining and instrumental activities of daily living. The Gerontologist, 9,(3), 179-186.[Medline]
• Mangione, C., Marcantonio, E., Goldman, L., Cook, E., Donaldson, M., Sugarbaker, D., et al. (1993). Influence of age on measurement of health status in patients undergoing elective surgery. Journal of the American Geriatrics Society, 41,377-383.[Medline]
• McCarty, H., Roth, D., Goode, K., Owen, J., Harrell, L., Donovan, K., et al. (2000). Longitudinal course of behavioral problems during Alzheimer's Disease: Linear versus curvilinear patterns of decline. Journal of Gerontology: Medical Sciences, 55A,M200-M206.[Abstract/Free Full Text]
• Miller, R., Clay, T., Fox, P., Newcomer, R. (1999). Effects of the Medicare Alzheimer's disease demonstration on nursing home entry. Health Services Research, 34,691-713.[Medline]
• Morrison, R., Siu, A. L. (2000). Survival in end-stage dementia following acute illness. Journal of the American Medical Association, 284,1856-1860.[Free Full Text]
• Newcomer, R., Clay, T., Luxenberg, J., Miller, R. (1999). Misclassification and selection bias when identifying Alzheimer's disease solely from Medicare claims records. Journal of the American Geriatrics Society, 47,215-219.[Medline]
• Newcomer, R., Yordi, C., DuNah, R., Fox, P., Wilkinson, A. (1999). Effects of the Medicare Alzheimer's disease demonstration on caregiver burden and depression. Health Services Research, 34,669-689.[Medline]
• Walter, L. C., Brand, R. J., Counsell, S. R., Palmer, R. M., Landefeld, C. S., Fortinsky, R. H., et al. (2001). Development and validation of a prognostic index for 1-year mortality in older adults after hospitalization. Journal of the American Medical Association, 285,2987-2997.[Abstract/Free Full Text]
• Yesavage, J., Rink, T., Rose, T., Aday, M. (1983). Geriatric depression rating scale: Comparison with other self-report and psychiatric rating scales. In T. Crook, S. Ferris, & R. Bartusleds (Eds.), Assessment in geriatric psychopharmacology (pp. 153–167). New Canaan, CT: Mark Powley and Associates.
• Zarit, S., Reeves, K., Bach-Peterson, J. (1980). Relatives of the impaired elderly: Correlates of feelings of burden. The Gerontologist, 20,260-266.[Medline]

This article has been cited by other articles:

				M. J. McClendon, K. A. Smyth, and M. M. Neundorfer Long-term-care placement and survival of persons with Alzheimer's disease. J. Gerontol. B. Psychol. Sci. Soc. Sci., July 1, 2006; 61(4): P220 - P227. [Abstract] [Full Text] [PDF]

				D. C. Grabowski The Cost-Effectiveness of Noninstitutional Long-Term Care Services: Review and Synthesis of the Most Recent Evidence Med Care Res Rev, February 1, 2006; 63(1): 3 - 28. [Abstract] [PDF]

This Article Abstract Full Text (PDF) Services Download to citation manager Citing Articles Citing Articles via HighWire PubMed PubMed Citation