Abstract
Objective: To determine age- and sex-specific AD prevalence rates for whites using each of four severity-based definitions of a case: a person diagnosed as having at least questionable AD, one diagnosed as having at least mild AD, one diagnosed as having moderate or severe AD, and one diagnosed as having severe AD.
Methods: Data from 21 studies of Europeans and North Americans were pooled.
Results: Empirically, one study, from East Boston, is an outlier. Applied to the US population of 1996, the obtained estimates yield a prevalence of 1.7 to 1.9 million cases, depending on whether the outlier study is excluded or included.
Conclusions: With disease severity taken into account, estimates of AD prevalence from white populations are more consistent than is usually acknowledged. By including disease severity in the case definition, variations in reported rates are much reduced. The outlier status of the East Boston study appears to result from the use of a definition of AD that differs from that used in the other 20 studies. Alternative explanations of the discrepancy between these estimates and the common estimate of 4 million cases are discussed.
Accurate prevalence rate estimates of AD by age and sex for different levels of severity would aid both etiologic research and health care planning. It has been noted that published estimates of age- and sex-specific prevalence fluctuate widely1; for example, between 7% for men older than 85 years in the Framingham Study2 and 54% for those in Southern California.3 Moreover, few studies present such specific rates for varying levels of severity.
Researchers at Johns Hopkins University1 conducted a meta-analysis of 15 prevalence studies from several continents using a logistic regression model to include the effects of predictors. They found that the extrabinomial variance (the amount of variance beyond that accounted for by the model) in the prevalence estimates was reduced by 76% when they added eight methodologic variables to a model that initially included only age as a predictor. These methodologic variables are inclusion of mild cases of dementia, inclusion of institutionalized subjects, use of CT for diagnosis, use of laboratory studies for diagnosis, use of Hachinski Ischemic Scale score for diagnosis, use of sampling, adjustment for false negatives, and the type of community ascertained. Estimating the independent contributions of some of these variables, however, led to surprising or paradoxical results. Gender did not make a difference, studies that included the institutionalized variable did not tend to have greater AD prevalence, studies that used laboratory studies found more AD, and studies that adjusted for false negatives tended to yield lower prevalence rates. The Johns Hopkins study1 was further able to reduce the amount of extrabinomial variation by eliminating two outliers with very high prevalence rates—the studies from Southern California3 and East Boston.4
The Southern California study,3 conducted from 1980 to 1982, evaluated 817 adults age 65 to 99 years in a white, predominantly middle-class retirement community, and estimated the prevalence of AD, including questionable cases. The East Boston study,4 begun in 1982, evaluated 3,623 adults older than 64 years of age in a white, urban, working-class community, and estimated the prevalence of AD. Both studies excluded institutionalized residents; employed two-stage designs, correcting for false negatives at stage one; and have been interpreted as using the equivalent of the criteria of the National Institute of Neurologic and Communicative Disorders and Stroke–Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) for the diagnosis of AD, although the East Boston study4 did not include the explicit assessment of social/occupational functioning.
We seek a parsimonious model of prevalence from which to make estimates—one including common biological predictors, such as age, sex, and race, and one that emphasizes disease definition. Our interest in definition centers on how much variation remains when disease severity is taken into account. Severity has been rated in different ways. For example, as measured by the Clinical Dementia Rating,5 disease severity has five levels: no dementia, questionable dementia, mild dementia, moderate dementia, and severe dementia. It is clear that the latter three levels represent cases of dementia, but not all of those rated as questionable progress to diagnosable dementia, and most prevalence studies define a patient with AD as a person with at least mild dementia.
The severity of dementia is an important variable in determining the prevalence rate of AD. Inspection of the prevalence rates from published studies suggests that prevalence is in fact relatively low, as logic would dictate, when only moderate and severe cases are included,2 and is relatively high when even cases of questionable AD are included.3 How large a role severity plays in accounting for variation in AD prevalence has not been studied systematically.
Besides the inclusion of severity, we also tried to reduce the variation among estimates in two other ways: by including sex as a predictor and by only including studies of primarily white populations. Gender, although not statistically significant in the Johns Hopkins analysis,1 has frequently been a predictor of prevalence, with women usually having the higher prevalence rates.2,6-9 Although the reasons are not well understood, studies of Asians tend to yield lower prevalence estimates than do studies of whites.10 Evidence indicating that blacks may have higher rates than European Americans11 also supports the possibility of racial differences.
Methods.
Selection of studies.
The selection criteria for studies of AD prevalence were similar to those of the Johns Hopkins study,1 with the addition of limiting the studies to whites. The criteria included the reporting of population-based, age- and sex-specific AD prevalence rates (or data from which these can be computed), diagnosed by NINCDS-ADRDA or equivalent criteria, with sample sizes (denominators), for white populations. We found 21 studies that satisfied these criteria2-4,6-9,12-25 (table 1). For some of these, data from the article describing the study were supplemented by data from another source or replaced by such data when the latter were more appropriate for our analysis or were thought to be more accurate (see table 1).26-28
AD prevalence study characteristics
Data.
We employed four different disease definitions: questionable or more severe dementia (questionable+), mild or more severe dementia (mild+), moderate or more severe dementia (moderate+), and severe dementia. We coded severity as given by the original authors. For 19 of the 21 studies, only one set of age- and sex-specific prevalence rates was presented, and therefore from each of these studies there was available only a single set of rates contributing to the analysis. One of these studies defined a case as questionable+, 15 of them defined it as mild+, and three of them defined AD as moderate+ (see table 1). One of the other two studies3 presented the data needed to obtain rates using all four definitions of a case, thereby contributing four sets of rates for analysis. The remaining study15 presented the data needed to compute rates using three definitions of a case of AD—mild+, moderate+, and severe—thereby contributing three sets of rates.
Whenever rates for pure AD cases were given, these were used in the analysis, even if rates for mixed AD cases were also given. If the only AD rates presented were for combined pure and mixed cases, then these were used. If the age- and sex-specific rates were corrected for false negatives, these were used; otherwise, the uncorrected rates were used.
Data analysis.
Severity-specific prevalence was assumed to be a logistic function of age and sex, and the multiple logistic regression capability of the SAS software package (SAS Institute, Cary, NC) was used to estimate parameters and to estimate prevalence rates. The individual data from all studies able to contribute to a given definition of a case were pooled. Thus, there were two studies contributing to the analysis based on the definition of a case as questionable+, 17 studies contributing to the analysis based on mild+, five studies contributing data to the analysis based on moderate+, and two studies contributing to the analysis of severe. To help identify possible outliers, models including each study as a predictor (dummy) variable were also analyzed.
In addition to pooling data, for comparisons of special interest the data from individual studies were also fit to a logistic model, including age and sex as predictors, so that the estimates from such studies could be compared with the overall estimates from pooled data reflecting a specific definition of a case. Microsoft Excel (97 SR-2, Redmond, WA) was used to generate plots of estimated prevalence as a function of age for both pooled and individual study data.
Age was coded as a continuous variable, using the midpoint of the age interval presented in the original study. All intervals with a lower limit of at least 60 years were included. Open-ended age intervals of 80 years or 85 years and older were assumed arbitrarily to extend to 95 years, whereas those listed as 90 years or 95 years and older were assumed to extend to 99 years.
We judged goodness of fit by the overdispersion parameter used in the Johns Hopkins study.1 This statistic, like the familiar chi-square statistic, is expected to equal one when all the variation in the data are accounted for by the specific model being tested, and is expected to be more than one when the data deviate from the model. In the former case, all of the variation is the result of chance and, for logistic models, is referred to as “binomial.” In the latter case, some of the variation represents the error in the model, and is often referred to as “extrabinomial.” Alternative logistic models including different sets of explanatory variables can be compared in terms of their overdispersion parameters. The amount of previously unexplained variation in the data now explained by a new model with an additional variable or set of variables is measured by the percent reduction in the extrabinomial variation between the new and old models.
Results.
Table 2 presents the estimated age- and sex-specific AD prevalence rates based on the model for each of the four definitions of a case of AD. As expected, prevalence increases with age and decreases with the level of severity used to define a case. The tendency toward a greater AD prevalence for women, observed when a case is defined as mild+, as moderate+, or as severe, is reversed when a case is defined as questionable+. Moreover, there is a clear dose–response curve relating the male-to-female prevalence ratio to case definition (severity). The ratio, which varies little over age for a given definition, declines systematically from 1.14 for questionable+ to 0.12 for severe.
Prevalence of AD by severity, age, and sex per 1000 individuals
Inspection of the data for mild+ reveals that the East Boston study4 may be an outlier. This was confirmed by a series of analyses including each study as a predictor. When East Boston was the reference study, all other studies were significant sources of variation. When any of the other studies served as the reference, East Boston was always significant, and each of the others only occasionally so, depending on the reference. When the study with rates closest to the estimated rates8 was used as a reference, only East Boston was a significant source of variation. In addition, East Boston consistently had the highest odds ratio (OR), which was always more than 2.0, regardless of which other study was used as a reference. Without East Boston, the same series of analyses was repeated, in which no study was consistently significant, consistently had the highest OR, or consistently had an OR more than 2.0. Estimated prevalence rates for mild+ excluding East Boston are also included in table 2.
When a case of AD is defined in the most usual way, as mild+, the overdispersion parameter was 5.87 (based on 17 studies). In contrast, the overdispersion parameters for questionable+ (two studies), moderate+ (five studies), and severe (two studies) were lower at 3.41, 1.35, and 2.39 respectively. In the Johns Hopkins study,1 this parameter diminished by 64% (from 3.6 to 1.3, with all nine covariates included) when data from East Boston4 and Southern California3 were eliminated. Without the East Boston data and retaining the Southern California data, in our study the overdispersion parameter was 2.01. Thus, East Boston accounts for about two-thirds of the extrabinomial variation in the mild+ prevalence rates analyzed here ([5.87 − 2.01]/5.87 = 0.66).
The figure compares logistic model estimates for women from selected individual studies with pooled data estimates for selected definitions of a case of AD. The large discrepancy between the Southern California curve3 at the top and the Framingham curve2 in the middle represents the large discrepancy among prevalence estimates commonly reported in the literature.
Figure. Estimated prevalence of AD for women according to severity. ......S. Calif., questionable+; ——East Boston, mild+; –·–·–S. Calif., mild+; –⋄–Pooled, mild+; – – – – Framingham, moderate+; –×–Pooled, moderate+.
The data from the Southern California study3 have been cited as being in agreement with the East Boston data.4 However, this similarity is between East Boston’s rates using the mild+ definition of a case and Southern California’s rates using the questionable+ definition. Once the definition of a case is taken into account, a different picture emerges. The East Boston curve is far above the Southern California mild+ curve, whereas the latter curve is very close to the pooled mild+ curve. In addition, it can be seen that the Framingham data2 do not represent especially low prevalence rates, just ones reflecting moderate+ AD. The Framingham curve is relatively close to the pooled moderate+ curve, as would be expected.
When the age- and sex-specific prevalence rates obtained here for mild+, excluding the data from the outlier East Boston study,4 are applied to the US population of 1996, the total prevalence of AD is estimated to be 1.7 million. When we include all the East Boston data as representing mild+ dementia, our model yields an estimate of 1.9 million cases of AD.
Discussion.
With disease severity taken into account, estimates of AD prevalence from white populations, with the exception of the East Boston study,4 are more consistent than is usually acknowledged. Moreover, we have shown, empirically, that the East Boston study4 is an outlier. The pooled data led to prevalence estimates with and without East Boston of 1.9 million and 1.7 million cases of AD respectively.
Controlling for severity, in addition to age, race, and sex, shrinks the variation not explained by the model and helps to interpret discrepancies noted in the literature. The pooled prevalence estimates for given case definitions are more useful than those generated when severity is ignored. The separate estimates for different levels of severity can be used in different ways. For example, the prevalence of moderate and severe cases may be important in health services planning, whereas a more liberal definition may be most useful in demographic modeling.
By taking severity into account in the definition of a case of AD, not only is the total amount of unexplained variation reduced, but specific differences in reported rates become less puzzling. For example, the wide gap between the Framingham2 and Southern California3 prevalence rates noted earlier can be recognized as what would be expected when comparing AD prevalence for cases defined as moderate+ with ones defined as questionable+.
The data from the East Boston study,4 using the mild+ definition of a case, are similar to those from the Southern California study,3 using the questionable+ definition. An interpretation of this may be derived from the observation that the East Boston study4 defined AD purely psychometrically, on the basis of cognitive tests, and did not include a functional test.26,29 Most of the other studies did require such a test, often based on an interview with an informant (see table 1). The specific tests used in the East Boston study are described in several of the articles on the study.4,30,31 That study’s definition of AD, in terms of the cutoffs employed, was relatively liberal. It required as evidence of memory impairment failure on any one of several tests a score of ≤2 of 6 points on the test of delayed memory, a score of ≤7 of 17 points on the test of delayed recognition memory span, or a decrement of ≥2 points between immediate and delayed memory. Patients also had to be impaired in at least one of three other cognitive abilities: a score of ≤10 of 15 points on the confrontation naming test, a score of ≤5 of 11 points on the copying test of spatial ability, or a score of ≤6 points on the test of visual abstraction.
Furthermore, in the East Boston study,4 the level of disease severity was assigned on the basis of how many cognitive tests were failed. This stands in contrast to most other approaches, which used some kind of scale based on functional assessment. In the Southern California study,3 for example, although cognitive and other kinds of clinical testing contributed substantially to case ascertainment, severity (including the designation of questionable cases) was determined by a social function questionnaire administered to alternate informants.3
These characteristics of the East Boston study4 may have resulted in a definition of a case closer to questionable+ than to mild+, as has also been noted by others.29 Because the NINCDS-ADRDA criteria for AD require impaired activities of daily living, it is likely on logical grounds that a group of people diagnosed without any testing for this requirement will include “cases” that would not ordinarily (i.e., with functional testing) qualify as mild+, but that might be questionable. The tendency of the East Boston researchers to find AD, as opposed to other dementing diseases, in a relatively high percentage of cases further increases their AD prevalence rates beyond the contribution of their liberal case definition. In the East Boston study,4 84% of the patients with dementia were diagnosed as having pure AD, whereas for the other studies2,3,6-9,12-25 the median pure AD rate was 57.5% (range, 42 to 81%; see table 1).
Our overall prevalence estimate for the US based on all the data, with or without the East Boston data, is considerably smaller than the estimate of approximately 4 million cases, based on the East Boston study alone.32 Our estimate obtained when the East Boston data are included as representing mild+ dementia is relatively close to that of another study33 using very different methods, but also including the data from East Boston. Most studies estimating the prevalence of mild+ dementia are in relatively close agreement, and their pooled data lead to the lower estimate. Relying mainly on the East Boston study4 may be misleading because it appears to be an outlier that may have used an unusually liberal definition of a dementia case and of AD relative to other dementias.
It is possible that the East Boston rates are the correct ones and that all or most of the other studies reviewed used tests for mild dementia that were too insensitive to detect many of the true cases in their samples. In this circumstance, the oft-cited estimate of 4 million cases, based on the East Boston study,4 would be correct, and the 20 other prevalence studies cited here would have to be considered to have collectively generated highly flawed estimates. Future research on the comparative validity of various clinical approaches to diagnosis may be needed to decide whether the East Boston approach is more valid than that used in the other studies.
Many of the questionable cases will develop clinical AD, whereas other questionable cases retain the same level of functioning or recover. For example, it has been shown recently that subjects with only memory impairment and little or no functional impairment develop AD at a high rate.34 It is possible that a significant proportion of questionable cases already had the neuropathologic substrate at the time of their diagnosis, and that most of the data analyzed here do not include these important cases. Thus, by combining these cases with the mild+ cases already counted, the “true” prevalence would be greater than the estimates presented here. The reality is that population-based epidemiologic surveys of AD, including most of those included in this and other reviews, rely on clinical definitions of probable/possible AD and by their very nature cannot economically include the very mild cases of AD.
If the NINCDS-ADRDA clinical criteria, requiring functional tests, were to be discarded on the grounds that most people with certain minimal cognitive deficits, regardless of functional status, really do have AD, then the East Boston data might be the most appropriate for obtaining overall prevalence estimates, as would other data, such as any questionable+ data, also requiring only evidence of slight cognitive impairment. Although the standard criteria are generally accepted for epidemiologic work, the development of more sensitive diagnostic techniques34,35 may be making it possible to diagnose AD accurately on the basis of more limited impairment than the NINCDS-ADRDA criteria require. Being able to detect these early cases is of potentially great usefulness in studies of prevention and other clinical research.
It is of interest that our model, including only two biological predictors but also controlling for race and severity, indicates less than twice as much unexplained variance as does the Johns Hopkins study,1 which used nine predictors (5.9 versus 3.6). In view of the paradoxical results they obtained in interpreting some of their methodologic variables, it would seem that the model based on age, sex, and severity, with race held constant, provides a more parsimonious and biologically satisfying account of how AD prevalence is determined.
The prevalence estimates obtained, based on the white population alone, should be improved when rates from other groups become available. The change might be small because the US population is still primarily white.
The data show an intriguing sex difference. Men are more likely to have questionable AD, but the male-to-female prevalence ratio declines systematically as severity increases. One hypothesis is that more women than men advance to the more severe levels of AD because women with the disease live longer than men with the disease. Data on time spent at a given level of severity would be helpful to clarify this picture.
Differences between the sexes are also sometimes observed in studies of AD incidence.36,37 Such studies are the best way to assess the rates of disease occurrence, and they can be used, combined with estimates of mortality, to derive prevalence estimates.33 Severity-specific prevalence estimates could be similarly derived from incidence and mortality data, included estimates of how long patients remain at each level of severity.
The following recommendations, some of which have been mentioned previously by other reviewers, should enhance the ability to integrate AD prevalence data in the future:
• Reporting operational diagnostic methods for determining level of severity
• Reporting rates at all levels of severity
• Reporting by race
• Reporting by sex
• Reporting age groups clearly, including the upper end of the oldest group
• Reporting rates of pure and mixed AD separately
• Correcting rates for false negatives in two-stage designs
An obvious limitation of this study is its reliance on severity ratings as given by the original researchers. Although it might be possible with a sufficient knowledge of each study’s diagnostic methods to transform their data to a common scale of severity, the needed information is, by and large, not available. The severity ratings that are available, however, explain much of the variation in the data in spite of the ratings not being rigorously equated across studies.
A related limitation concerns differences in the diagnostic methods used in the studies. As stated when describing our selection of studies, all of those reviewed used the NINCDS-ADRDA or equivalent diagnostic criteria. In most cases, adherence to these criteria was declared explicitly by the researchers involved, and in the other cases, we accepted the judgment of other reviewers.1 Although the criteria for AD are similar across studies, their implementation varies in that different diagnostic tests were used for similar purposes in different studies. For example, some used CT to help diagnose the kind of dementia of a particular patient, others used MRI, and others did not use any imaging modality. This is a problem endemic to information synthesis. One can, as in the Johns Hopkins study,1 code for particular methods, but this has not always led to easily interpretable results. For example, in the Johns Hopkins study,1 the use of laboratory studies for diagnosis was associated with a higher AD prevalence rate, but it is not clear why. Even when seemingly relevant variables, including many of the methodologic variables investigated in that study, are ignored, as was done in this investigation, the control for severity, age, race, sex, and one outlier significantly limits the unexplained variation among the estimates.
The main limitation of the current study stems from the pooling method used to combine the data. This method, unlike some other meta-analytic approaches, weights the data according to the sample size of the contributing studies. To the extent that the rates from the different studies are heterogeneous, the pooled estimates will be biased toward the rates of the larger studies. It is for this reason that we recommend excluding the data from the apparent outlier, East Boston, in estimating the mild+ rates. When this is done, the rates are reasonably homogeneous, the pooling approach is valid, and the results are relatively orderly.
Footnotes
See also page 166
The authors wrote this article independently of their affiliation with the US General Accounting Office. The views expressed herein are those of the authors and do not necessarily reflect those of the US General Accounting Office.
- Received March 8, 1999.
- Accepted in final form March 13, 2000.
References
Letters: Rapid online correspondence
REQUIREMENTS
You must ensure that your Disclosures have been updated within the previous six months. Please go to our Submission Site to add or update your Disclosure information.
Your co-authors must send a completed Publishing Agreement Form to Neurology Staff (not necessary for the lead/corresponding author as the form below will suffice) before you upload your comment.
If you are responding to a comment that was written about an article you originally authored:
You (and co-authors) do not need to fill out forms or check disclosures as author forms are still valid
and apply to letter.
Submission specifications:
- Submissions must be < 200 words with < 5 references. Reference 1 must be the article on which you are commenting.
- Submissions should not have more than 5 authors. (Exception: original author replies can include all original authors of the article)
- Submit only on articles published within 6 months of issue date.
- Do not be redundant. Read any comments already posted on the article prior to submission.
- Submitted comments are subject to editing and editor review prior to posting.
You May Also be Interested in
Cutaneous α-Synuclein Signatures in Patients With Multiple System Atrophy and Parkinson Disease
Dr. Rizwan S. Akhtar and Dr. Sarah Brooker