Abstract
Objectives: To estimate annual transition probabilities (i.e., the likelihood that a patient will move from one disease stage to another in a given time period) for AD progression. Transition probabilities are estimated by disease stages (mild, moderate, severe) and settings of care (community, nursing home), accounting for differences in age, gender, and behavioral symptoms as well as the length of time a patient has been in a disease stage.
Methods: Using data from the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD), the authors employed a modified survival analysis to estimate stage-to-stage and stage-to-nursing home transition probabilities. To account for individual variability, a Cox proportional hazards model was fit to the CERAD data to estimate hazard ratios for gender, age (50 to 64, 65 to 74, and more than 75 years), and level of behavioral symptoms (low/high, according to responses to the Behavioral Rating Scale for Dementia) for each of the key stage-to-stage and stage-to-nursing home transitions.
Results: The transition probabilities underscore the rapid progression of patients into more severe disease stages and into nursing homes and the differences among population subgroups. In general, male gender, age under 65, and high level of behavioral symptoms were associated with higher transition probabilities to more severe disease stages. Disease progression is roughly constant as a function of the time a patient has spent in a particular stage.
Conclusions: Transition probabilities provide a useful means of characterizing AD progression. Economic models of interventions for AD should consider the varied course of progression for different population subgroups, particularly those defined by high levels of behavioral symptoms.
As new drugs for AD are approved for use, interest in their cost-effectiveness has arisen. In recent years, a number of economic evaluations of these treatments have been published in the medical literature.1-5⇓⇓⇓⇓ But the previous economic evaluations suffer from three potentially important limitations. First, they have modeled the natural history of disease for generalized patient cohorts without stratifying by individual characteristics such as age and gender. Second, they have defined AD and measured its progression solely on the basis of cognitive functioning, typically measured with the Mini-Mental State Examination6 or Clinical Dementia Rating (CDR),7 without accounting for other distinguishing features such as behavioral symptoms. Third, at least some of the analyses have used Markov models, which assume that disease progression is independent of the length of time a patient has been in a particular disease stage.1,3,4⇓⇓
See also pages 943, 964, and 972
Both demographic and behavioral factors may be important features in the prognosis and management of the disease and thus integral to the development of more accurate economic models. Several studies8-10⇓⇓ have reported that younger age at disease onset is associated with faster rates of decline, for example (though others have reported the opposite effect or minimal impacts of age).11-14⇓⇓⇓ Researchers have also reported that male gender is a positive predictor of nursing home admission and death in patients with AD.15,16⇓
It is widely recognized that patients with AD experience a variety of coincident behavioral or psychiatric symptoms, which may be the presenting complaint or may emerge in the course of disease.17-20⇓⇓⇓ Manifestations include depression, apathy, agitation, aggression, and sleep disruption17-19⇓⇓ as well as psychotic symptoms such as delusions and hallucinations.19
A number of researchers have concluded that various behavioral changes (notably psychosis and aggression) are associated with more rapid rates of cognitive and functional disease progression.10,13,18,21⇓⇓⇓ Numerous studies have also found that certain behavioral disturbances and psychiatric symptoms10,22-25⇓⇓⇓⇓ as well as sleep disruption26,27⇓ are important contributors in decisions to institutionalize patients.
In this article, we present estimates of stage-to-stage and stage-to-nursing home transition probabilities for AD. Transition probabilities reflect the likelihood that a patient will move from one disease stage to another in a given time period. Over the years, they have been used widely in economic models of health and medical treatments.3 In addition to adjusting for age, gender, and behavioral symptoms, we also examine how transition probabilities vary with the length of time a patient has spent in a particular disease stage.
Data and methods.
Data sources.
To estimate transition probabilities, we used data from the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD), a longitudinal database of 1145 dementia patients who were examined annually by clinicians in 22 major medical centers in the United States between 1986 and 1995.28 All subjects in CERAD received a standardized assessment battery at entry and annual follow-up examinations, which included a medical and neurologic examination, a battery of neuropsychological measures, and stage of disease as assessed by the CDR. Demographic information as well as date of nursing home admission and date of death were also obtained. Details about the population and assessments have been described elsewhere.28
Because of its large and diverse population and because patients were assessed annually over several years and followed into nursing homes, CERAD provides an excellent source for the estimation of transition probabilities, adjusted for age and gender. Roughly 40% of the CERAD population was male. Approximately 17% were under 65 years of age, 38% were between 65 and 74, and 45% were 75 or older.
CERAD also permits an estimation of transition probabilities for groups defined by behavioral characteristics. During the course of the CERAD data collection, 345 patients were administered the Behavior Rating Scale for Dementia (BRSD), an instrument designed to measure the presence and frequency of occurrence of a wide range of psychopathology, including mood changes, vegetative changes, apathy, anxiety, agitation, aggression, hallucinations, and delusions.29 It also includes pathologic behaviors such as wandering and repetitive questions, though these are less commonly associated with specific psychiatric disorders. Patients in CERAD received up to four administrations of the BRSD over time.29 In total, CERAD contains 551 BRSD observations (299 with the original BRSD and 252 with a slightly revised version) for the 345 patients administered the questionnaire. Patients receiving the BRSD were mostly in the mild (CDR = 0.5 or 1; 50.2%) or moderate (CDR = 2; 37.3%) disease stages. Roughly half were women, and over 90% lived at home.
The large number of patients undergoing annual assessments also permits an estimation of time-varying annual transition probabilities, defined as the probability of transitioning from one stage to another, conditional on the amount of time spent in a given stage. For example, the probability of transitioning from the mild to moderate stage for patients who had been in the mild stage for 1 year may differ from the transition probability for patients who had been mild for, say, 2, 3, or 4 years.
One challenge in estimating transition probabilities using CERAD involves missing data. Although 1,145 patients in CERAD had assessments on entry, the numbers dropped off with each visit, as patients died or were lost to follow-up. For example, only 774 patients were assessed for their first visit, 626 for the second, 534 their third, 367 their fourth, 260 their fifth, 180 their sixth, 85 their seventh, and 16 their eighth.
Missing data can also be an issue because even some patients who completed an annual assessment sometimes had data missing in the record. As one example, of the 3,987 visits captured in the database, annual CDR assessments were missing for 24 visits (0.6%).
Despite these problems, CERAD remains the largest database of its kind and is likely the best available source for calculating transition probabilities. In the estimates that follow, patients lost to follow-up were treated as censored data. Where “between-visit” CDR assessments were missing (e.g., present on visit 1 and 3 but not on visit 2), we assumed that the CDR stage at the missing visit was identical to the preceding assessment. In sensitivity analyses, we repeated our analyses under alternative assumptions (e.g., that the missing value stage had increased by a 1-unit increment relative to the preceding visit), with no change in the overall results. In our estimates of time-varying probabilities, we provide in a footnote to table 4 the number of patients who were present for different lengths of time. Finally, 55 patients were missing data on nursing home status and were excluded from the analyses of the probability of institutionalization, conditional on disease stage.
Time-varying annual transition probabilities
Estimating transition probabilities.
Using the CERAD data, we estimated annual stage-to-stage and stage-to-nursing home transition probabilities for population subgroups defined by gender, age (50 to 64 years, 65 to 74 years, and 75 years or over), and level of behavioral symptoms. Disease stage was classified as mild, moderate, or severe, based on the CDR classification. We also estimated stage-to-dead transition probabilities. Following the practice used in CERAD, we assumed that once entering a nursing home, a patient remained institutionalized.
We classified patients as having a “high” or “low” level of behavioral symptoms, based on the presence of items listed for each patient on the BSRD. The BRSD consists of 48 specific psychopathologic signs and symptoms.29 The mean (SD) number of items for the 345 patients was 14.6 (7.0) (median = 14.0). Symptoms experienced by over half of the respondents included purposeless behavior, agitation, verbal repetitiveness, irritability, loss of interest, loss of enjoyment, tiredness, physical signs of anxiety, altered sleep pattern, restlessness, and diurnal pattern of confusion.
We defined patients as having a high degree of behavioral impairment if they had more than 12 items present in the previous month at the time of initial assessment. This threshold was chosen based on logistic regression analyses to predict nursing home entry, using different cutoff values for the number of behavioral symptoms as the key independent variable, after controlling for disease stage (based on CDR), age, and gender. For example, when the high behavioral impairment group was defined as having more than 12 behavioral items, the odds ratio relative to the low behavioral group was 2.24 (p = 0.01), which was greater and more highly significant than that of other potential thresholds. Note that our analysis relies on the first administration of the BRSD. Although some patients had multiple BRSD assessments, the number of follow-up visits was too small for the purposes of this analysis.
Transition probabilities were estimated for each of the stage-to-stage transitions, using a modified survival analysis that adjusts for the discrete nature of the data as well as for differential follow-up times in CERAD.3 For example, to estimate the probability of transitioning from mild to moderate AD, we divided the number of annual mild to moderate transitions or “events” in CERAD by the total number of annual transitions by individuals who began a year in the mild stage.
The probability of transitioning from a community setting to a nursing home was estimated separately for each disease stage in similar fashion. That is, conditional on their disease stage, community-based patients were assigned a probability of being in one of two settings, community or nursing home, in the following period. We assumed that disease stage-to-stage transitions occurred independently of setting. For example, the probability of moving from the mild to the moderate AD stage did not depend on whether a patient resided in the community or a nursing home.
Finally, we fit a Cox proportional hazards model to the CERAD data to estimate hazard ratios for gender and age associated with each of the key stage-to-stage and stage-to-nursing home transitions.30 The commonly used Cox model provides a means of conducting regression analysis that incorporates time to event (in this case, the time spent in a particular disease stage before progressing to another stage) as a key explanatory variable. Thus, we specify the hazard function as a function of time and the covariates under consideration. Similarly, we fit a Cox proportional hazards model to the 345 patients in CERAD for whom BRSD data were collected and estimated hazard ratios for level of behavioral symptoms, adjusting for gender and age, for each of the key transitions. Using similar methodology, we estimate time-varying annual transition probabilities, that is, stage-to-stage and stage-to-nursing home transition probabilities, conditional on the time patients spent in a given stage.
Results.
Stage-to-stage and stage-to-nursing home transitions.
Table 1 presents the annual stage-to-stage and stage-to-nursing home transition probabilities estimated from the aggregate CERAD sample. The table also presents the combined stage and nursing home transition matrix, obtained by multiplying stage-to-stage and stage-to-nursing home transitions for each possible combination of stage and setting. For example, patients who begin in the mild/community stage have a 59.1% chance (0.614 × [1 − 0.038]) of remaining in that stage in the following year.
Estimated annual transition probabilities from CERAD
The figure shows the disease progression of a hypothetical cohort of community-based patients with mild AD over time. The numbers are obtained by recursively multiplying the 7 × 7 matrix at the bottom of table 1. After 1 year, 59.1% of patients remain in the mild/community stage; by the fourth year, this percentage has declined to 14.4%. On the other hand, the percentage of patients in the severe/nursing home stage increases over time, from 0.2% in year 1 to 14.8% in year 4.
Figure. Natural history of AD based on transition probabilities from the Consortium to Establish a Registry for Alzheimer’s Disease (assuming 100% at cohort is in the mild/community stage at year 0); comm = community; nh = nursing home; purple = mild; magenta = moderate; yellow = severe; blue = dead.
Adjustments for age and gender.
The hazard ratios associated with gender and age are presented in table 2. As reflected in the ratios for the stage-to-stage transitions, disease progression is somewhat more rapid in men than women. For example, compared with women, men are 1.16 times more likely to transition from the mild to the moderate stage, controlling for age (p = 0.14). Men have a higher hazard of dying for each of the three stages and in the case of the moderate-to-dead and severe-to-dead transitions, differences are significant (p > 0.05). Men have somewhat lower rates of transitioning from the mild or moderate stage to the nursing home, but differences are not significant.
Hazard ratios associated with gender and age
In general, older age groups have lower hazard ratios for stage-to-stage transitions and higher ratios for mortality and stage-to-nursing home admission, though few of the ratios are statistically significant. The exception is the hazard of death for the over-75 group, which is substantially and statistically significantly higher than for the 50- to 65-year-old group.
Adjusting for behavioral symptoms.
Patients with high levels of behavioral symptoms tend to have greater hazards of transitioning than those with low levels of symptoms (table 3). For example, the high behavioral group has a hazard ratio of 1.33 (p = 0.09) of transitioning from mild to moderate compared with the low behavioral group. The high behavioral symptom group has higher hazard ratio of transitioning to nursing home than the mild and moderate groups, though differences are not significant.
Hazard ratios associated with high and low levels of behavioral symptoms
Adjusting for time in stage.
Finally, table 4 shows that transition probabilities are generally constant with the number of years spent in a particular disease stage. For example, the stage-to-stage transitions show that 32.7% of patients who have been mild for only 1 year transition to the moderate stage in the following year compared with 34.3% for patients who have been mild for 2 years and 30.8% for patients who have been mild for 3 years. A roughly similar pattern holds for the mild-to-severe and moderate-to-severe transitions and for the stage-to-dead and stage-to-nursing home transitions.
Discussion.
Over the years, researchers have measured AD progression in various ways, including monitoring change in scores in levels of cognitive functioning,10,11,13,14⇓⇓⇓ employing survival analysis to measure time to institutionalization and death,10,15,16,31⇓⇓⇓ and tracking clinical markers of disease advancement such as loss of instrumental activities of daily living and failure to recall words on the Mini-Mental State Examination.32
Measuring disease progression in terms of transition probabilities can provide a useful alternative means of characterizing the course of chronic illnesses and of projecting the path of patient populations over time. The transition probabilities for AD estimated here from the CERAD database could potentially be helpful to clinicians, health insurers, and policy makers in understanding and predicting the trajectory of AD cohorts over time. The data should also prove useful to researchers developing models of disease progression and estimating the cost-effectiveness of interventions for AD in different population subgroups. As one example, one might apply the risk ratios for transitioning from moderate to severe AD reported for patients on selegiline or α-tocopherol, compared with placebo,33 and apply them to our “natural history” transition probabilities to model the consequences of these treatments over time.
Our results highlight several important points. First, they underscore the rapid and progressive nature of the disease. The overall transition probabilities imply, for example, that the majority of a cohort of mild, community-based patients with AD progresses to the severe/nursing home stage (25.3%) or dies (25.5%) in 5 years. In general, our results are consistent with other reports that degree of dementia stage is an important predictor of time to institutionalization and death.10,16,31⇓⇓
The data also underscore the varied course of progression for different AD subgroups. In general, male gender, age under 65 years, and high level of behavioral symptoms are associated with higher transition probabilities to more severe disease stages. These results suggest that richer economic models, which account for individual variability and behavioral features, are needed. Such models, which incorporate multidimensional features of the disease, are currently under development.34
Behavioral symptoms may be a particularly important independent factor in disease progression and nursing home placement, and future models and cost-effectiveness analyses would do well to consider this dimension. Our findings are relevant in light of ongoing research to determine the impact of new drugs for AD on behavioral changes and the fact that new antidepressants and antipsychotics are emerging as first-line treatment of psychiatric problems in AD and other dementias.21,35-37⇓⇓⇓
Our results also suggest that transition probabilities do not vary with the length of time a patient has spent in a particular stage. That is, patients who have remained in the mild or moderate stage for several years are as likely to progress as those who have been in a stage for only 1 or 2 years, for example.
In some ways, this result is counterintuitive, because one might anticipate progression to be more rapid in patients who have remained in a stage for several years. The data may reflect the fact that some patients are “rapid progressors,” whereas others achieve a plateau and thereafter progress more slowly. It should be emphasized, however, that the sample size for our estimates of many individual transitions is small and the results must be viewed with caution. The small numbers make adjusting the time-dependent transition probabilities for age or gender difficult, for example. Moreover, the data do not account for how long patients may have had symptoms before their initial classification in CERAD.
Still, the results provide some support for the use of Markov models in economic analyses of drug treatments for AD. Although Markov models are often useful for studying chronic and progressive diseases, for which short-term data must be used to forecast long-term outcomes,38,39⇓ they impose a “memory-less” assumption: That is, an individual’s probability of transitioning to a different stage does not depend on the history of disease progression. Our results suggest that this may be a reasonable assumption in studies of cohorts with AD, though more study into this area is needed.
It is important to note several limitations in our analysis. First, the sample sizes for many of the transitions examined were small. In particular, the small number of patients making transitions from the moderate and severe stages limited our ability to provide more precise estimates.
Another issue involves the generalizability of estimates from CERAD, which drew its subjects from academic medical centers. One potential problem, for example, is that our definition of mild AD includes individuals with CDR stages 0.5 and 1.0. Potentially, this could result in variability because of the inclusion of “possible AD” cases that do not, in fact, have AD but another variant of dementia. However, only a small percentage (3%) of all cases in CERAD had a CDR score of 0.5 at baseline, and exclusion of these cases in sensitivity analyses did not measurably change results.
In addition, a small percentage of the CERAD population (4.3%) undergoes a “backward transition” from moderate to mild AD, despite the progressive nature of AD, which should preclude such reversals. This may reflect variations in clinical assessments, the presence of depression or psychosis, or adjustments of medications, which may mask the actual disease stage at any particular time.
A related concern involves the representativeness of the population administered the BRSD, most of whom had mild or moderate AD. Despite these potential drawbacks, CERAD remains the largest AD registry available, and its population was fairly diverse in terms of the items considered in this analysis.
Third, we adjusted our transition probabilities solely on the basis of age, gender, behavior symptoms, and time in stage. Other factors may influence disease progression and nursing home placement, including a patient’s duration of disease at initial visit, socioeconomic status, marital status, severity of dementia as measured by activities of daily living, and the availability, capability, resources, health, and burden of caregivers.10,16,31,40,41⇓⇓⇓⇓ We also assumed that disease progression was independent of residence, though caregivers’ capabilities and other factors may be covariates. It will be important to investigate these relationships in the future.
Fourth, our measure of behavioral disturbances is based on a one-time assessment with the BRSD. Researchers have noted that the presence of behavioral symptoms at any given time is not necessarily indicative of the temporal pattern of behaviors.17-20,42,43⇓⇓⇓⇓⇓ A related issue is that we measured only behavioral symptoms occurring in the prior month, which may be too infrequent a period to detect some clinically important behavioral features.29 Furthermore, our composite measure of behavioral disturbances simply reflects the occurrence of problems, without weighting the importance of any one over another. It is possible that certain types of behaviors (e.g., agitation or delusions) are more important than others in predicting disease progression and nursing home placement or that certain combinations of behaviors are better predictors. Researchers have reported, for example, that some behavioral disturbances (e.g., agitation, irritability) tend to occur in tandem.18 Also, it is important to note that other instruments measure behavioral symptoms differently than the BRSD does29 and that the field has often lacked consensus in defining and measuring behaviors.18 Thus, it is important to view this work as a first step with an ongoing need to refine estimates.
It will also be important to examine the costs and health-related quality of life in AD associated with age, gender, and levels of behavioral impairment, after controlling for degree of cognitive impairment. Most of the work in the area to date has related costs and health-related quality of life to levels of cognition but has not incorporated these other factors.44-47⇓⇓⇓ The strength of future cost-effectiveness analyses of treatments for AD will depend on such factors.
Acknowledgments
Support provided by Janssen Pharmaceutica Products L.P., Titusville, NJ.
Acknowledgment
The authors are grateful to Gerda Fillenbaum for help in obtaining data from CERAD and for helpful comments on an earlier draft of this manuscript.
- Received August 28, 2000.
- Accepted May 12, 2001.
References
Disputes & Debates: Rapid online correspondence
NOTE: All authors' disclosures must be entered and current in our database before comments can be posted. Enter and update disclosures at http://submit.neurology.org. Exception: replies to comments concerning an article you originally authored do not require updated disclosures.
- Stay timely. Submit only on articles published within the last 8 weeks.
- Do not be redundant. Read any comments already posted on the article prior to submission.
- 200 words maximum.
- 5 references maximum. Reference 1 must be the article on which you are commenting.
- 5 authors maximum. Exception: replies can include all original authors of the article.
- Submitted comments are subject to editing and editor review prior to posting.