Analysis of Time Trends in Alzheimer’s Disease and Related Dementias Using Partitioning Approach

Abstract

Background:

Understanding the dynamics of epidemiologic trends in Alzheimer’s disease (AD) and related dementias (ADRD) and their epidemiologic causes is vital to providing important insights into reducing the burden associated with these conditions.

Objective:

To model the time trends in age-adjusted AD/ADRD prevalence and incidence-based mortality (IBM), and identify the main causes of the changes in these measures over time in terms of interpretable epidemiologic quantities.

Methods:

Trend decomposition was applied to a 5%sample of Medicare beneficiaries between 1991 and 2017.

Results:

Prevalence of AD was increasing between 1992 and 2011 and declining thereafter, while IBM increased over the study period with a significant slowdown in its rate of growth from 2011 onwards. For ADRD, prevalence and IBM increased through 2014 prior to taking a downwards turn. The primary determinant responsible for declines in prevalence and IBM was the deceleration in the increase and eventual decrease in incidence rates though changes in relative survival began to affect the overall trends in prevalence/IBM in a noticeable manner after 2008. Other components showed only minor effects.

Conclusion:

The prevalence and IBM of ADRD is expected to continue to decrease. The directions of these trends for AD are not clear because AD incidence, the main contributing component, is decreasing but at a decreasing rate suggesting a possible reversal. Furthermore, emerging treatments may contribute through their effects on survival. Improving ascertainment of AD played an important role in trends of AD/ADRD over the 1991-2009/10 period but this effect has exhausted itself by 2017.

Keywords

Alzheimer’s disease decomposition dementia incidence Medicare partitioning relative survival time trends

INTRODUCTION

The number of people with Alzheimer’s disease (AD) and related dementia (ADRD) in the U.S. is expected to double by the year 2060 [1, 2] with associated healthcare costs comparable to the financial burden of heart disease and cancer [2, 3]. AD accounts for 60–80%of dementia cases diagnosed [4], and in clinical practice, its diagnosis often intertwines with other dementias: about 50%of individuals with AD also had dementia due to cerebrovascular disease or Lewy body disease [3] and about 30%of individuals who have been diagnosed with AD had other types of dementia on autopsy [3]. Therefore, understanding the dynamics of current and historical epidemiologic trends in AD/ADRD as well as their epidemiologic causes is vital to providing important insights into reducing the health and human burden associated with these conditions.

Two fundamental epidemiologic measures indicative of the expected level of disease-related burden are the prevalence proportion and incidence-based mortality (IBM). These measures are easily estimated from survey, medical record, and/or administrative claim data and their time trends reflect important information on periods of success and failure in disease care and prevention. At the same time, inferences based on these measures can be misleading. Observed changes in disease prevalence and IBM result from two simultaneously occurring processes: changes over time in disease incidence, and patient survival. These processes are the targets of specific health interventions aimed at reducing disease incidence, increasing survival, or both. However, even if such health interventions are highly effective and have no adverse trade-off effects, the resulting improvements in population health may not be clearly observable in the data.

For prevalence, one source of confounding is the way improvements in incidence and survival jointly affect the observable prevalence proportion. In an ideal situation, successful actions of public health policy and the healthcare system act conjointly to decrease the incidence of a disease through prevention and increase survival after disease onset through treatment. However, these improvements in population health will not, necessarily, lead to a decrease in the levels of population prevalence. This is because lower incidence decreases disease prevalence while better survival increases it. The resulting change in the prevalence proportion can then increase, decrease, or remain constant, depending on the relative magnitude of the two effects. In contrast, improvements in incidence and survival act concordantly to reduce IBM: improved incidence reduces the total number of people with the disease while improved survival further reduces the number of associated deaths. However, the resulting picture is still not clearly interpretable due to the confounding effect of simultaneous changes in general mortality rates not specific to the disease in question.

A further complication in the study of prevalence/mortality trends are the uncertainties introduced by the source of data being utilized for analysis. Indeed, the problem of correctly ascertaining a definitive diagnosis of AD as distinct and separate from other types of dementia outside of relatively small cohort studies is an important barrier to obtaining accurate, nationally representative estimates of AD/ADRD prevalence. This source of uncertainty is exacerbated by the fact that AD can and does co-exist with dementias of other etiology. In the case of AD/ADRD mortality the primary source of national-level data is information obtained from death certificates [4]. However, literature on other chronic diseases has identified significant differences between mortality trends based on death certificates and IBM. For example, the median proportion of diabetes mellitus reported in death certificates among deceased individuals known to have the disease was only 43%[5] with significant variation associated with the origin of the death certificate [6]. The extent to which death certificate data accurately reflects AD/ADRD mortality has not been widely studied; however, recent research showed that the percentage of deaths attributable to dementia can be underestimated if the calculation is based on death certificates [7]. Furthermore, death certificate-based mortality trends of AD, ADRD, and AD/ADRD combined (Fig. 1) demonstrate a relationship between these causes that requires additional clarification. Specifically, there is a partial decline in AD-related mortality observed in 2012–2013 which is wholly overpowered by a rapid increase in mortality from ADRD. The combined trend shows a relatively smooth pattern of mortality, suggesting some trade-off in the recorded mortality from these two causes.

Fig. 1

Multiple Cause of Death-based Age-Adjusted (65+) Mortality Rates. Mortality with Alzheimer’s disease (AD; blue dots), Alzheimer’s disease related dementia (ADRD; red dots), and AD/ADRD combined (AD+ADRD; black dots) using underlying cause of death and multiple causes of death information.

Additional information on time trends of AD/ADRD mortality can be obtained from health trajectories reconstructed from administrative claim records such as those from the Medicare system. Previous studies using such data for the study of other chronic conditions such as diabetes mellitus were both adequately sensitive and highly specific, when compared to studies utilizing death certificate data alone [5]; while those focused on the relationships between dementia diagnoses made using the ADAMS protocol [8] and Medicare claims found that Medicare claims and ADAMS agreed on 85%of subjects with a kappa statistic of 0.70 with the sensitivity/specificity of Medicare claims being 0.85/0.89 for dementia and 0.64/0.95 for AD [9, 10]. Furthermore, such data provide an important resource for obtaining nationally representative estimates of AD/ADRD incidence and prevalence. Although it is true that such estimates are limited in that they are based on the presence of an International Classification of Disease (ICD) code on a billing record and cannot be linked to a medical chart or autopsy report for diagnosis validation, the scope of the data allows us to estimate reasonably accurate bounds for the real trend in AD/ADRD incidence/prevalence by varying the range of ICD codes and the strictness of the parameters used in the algorithm used to identify AD/ADRD onset from Medicare data.

In this study, we use Medicare data together with an innovative method of partitioning analysis [11 –14] to identify the time trends in age-adjusted AD/ADRD prevalence, IBM, incidence and survival; develop and estimate general models to predict these measures; demonstrate their high accuracy in comparison to empiric estimates; and identify the main causes of the changes in prevalence and IBM over time in terms of interpretable epidemiologic measures. The partitioning analysis utilized has been designed to explicitly quantify how prevalence and IBM trends are affected by the relative contributions of changes in incidence, relative survival, and mortality in the general population as well as to account for changes in disease prevalence at the bounds restricting the age-time period where data are available. Combined with a Medicare-based dataset of sufficient statistical power to obtain nationally representative estimates even for diseases that are relatively rare on the populational level, this approach allows us to mitigate the effect of the uncertainties involved in traditional prevalence and mortality studies.

METHODS

Data sources

Administrative claim records drawn from a nationally representative 5%sample of the total U.S. Medicare population (5%-Medicare) spanning the 1991–2017 period were utilized. The 5%-Medicare database provided data on the diagnoses made during episodes of care paid for by either Medicare Part A (facility-based services), or Medicare Part B (professional services), dates of death, as well as basic demographic characteristics. The age-time area in which data are available for this study is shown in Supplementary Figure 1. Death certificate-based mortality trends for comparison were generated using the CDC-Wonder online tool [15].

Empiric analysis

Using Medicare enrollment files, we identified the first and last month/year during which an individual was enrolled in a traditional Medicare fee-for-service plan with both Parts A and B coverage. These two points in time served as the bounds over which an individual was followed. Then, we calculated the proportion of time (in months) within these bounds that the individual spent in a Medicare Advantage plan that did not share claims information with the Centers for Medicare and Medicaid Services. If the total proportion of time spent in Medicare Advantage was greater than 20%, then the individual was dropped from the analysis. Disease ascertainment algorithms described in detail in other publications [16] were used to identify the presence of AD (ICD-9:331.0; ICD-10: G30.x) and ADRD (ICD-9:290.x or 294.2x; ICD-10: F01.x, F03.x) as well as the individuals’ age at entry, death/censoring, and any time within the available time-series. Briefly, we required the presence of two distinct claims with a diagnosis of AD/ADRD within 90 days of each other with the earliest date in the pair designated as the date of onset. If death occurred within 90 days of the first diagnosis of AD/ADRD, this was treated as a confirmatory record. Once the date of onset was identified, the individual was considered diagnosed during the entire follow-up period. Sensitivity testing found that our results were not sensitive to variation in the Medicare Advantage cutoff point or to variation of the 90-day verification period.

The total number of individuals within the age-time area of the study was 4,632,534. Of these, 348,932 were identified as AD cases, and 748,737 as ADRD cases. Disease prevalence proportions, incidence rates, incidence-based mortality, and all-cause mortality were then calculated in two-dimensional age-time-specific bins spanning the age-time plane for which data was available. We used 29 age groups (25 single-year groups for ages 65–90, four aggregated age-groups: 90–91, 92–94, 95–99, and 100–110) and single-year calendar time groups for 1992–2017 period. A single individual trajectory contributed to multiple age-time bins over the course of its follow-up period. Age-specific prevalence, incidence, incidence-based mortality, and all-cause mortality in each bin were calculated as the ratio of the actual number of cases detected in each bin (or the number of person-years with the disease for the calculation of prevalence) over the total number of person-years in that bin (Supplementary Figure 2). Age-adjusted measures were then calculated for each calendar year using population weights from standard population counts for the year 2000. Standard errors and confidence intervals for all age-adjusted measures were calculated based on total number of cases and person-years [17].

Partitioning analysis

This study uses a recently developed partitioning approach for disease prevalence and mortality based on an explicit representation of prevalence and mortality with no simplifying assumptions [11]. In brief, the method 1) predicts trends in prevalence and mortality, 2) decomposes (or partitions) them into their constituent components, and 3) calculates the relative impact each component has on the overall trend as well as intertemporal changes in the strength and direction of these impacts. The specific outcome measures used in this study are age-specific and age-adjusted AD/ADRD prevalence and incidence-based mortality; the constituent components are AD/ADRD incidence, relative survival, morbidity at the bounds defined by the availability of data (65 years of age and the year 1992), and mortality for the general population (for incidence-based mortality only).

Formally, age-adjusted prevalence, P (y) = P₀ (y) + P₀₀ (y) + P_is (y), is represented as the sum of three positive contributions from individuals with pre-existing prevalence at the age-65-boundary (P₀ (y)), pre-existing prevalence at the time-1992-boundary (P₀₀ (y)), and, disease onset after 1992 and age 65 (P_is (y)). Similarly, age-adjusted incidence-based mortality, M (y) = M₀ (y)+ M₀₀ (y) + M_is (y) + M_Pμ (y), is represented as the sum of the contributions of mortality among individuals with pre-existing prevalence at the age-65-boundary (M₀ (y)), pre-existing prevalence at the time-1992-boundary (M₀₀ (y)), as well as two other contributions associated with individuals with disease onset after 1992 and age 65: mortality because of the presence of the disease (M_is (y)) and mortality interpretable as non-disease-specific (M_Pμ (y)).

Partitioning of age-adjusted prevalence and incidence-based mortality is formally obtained by differentiation over calendar time (year). Thus, partitioning for prevalence, P′ (y)/P (y) = T₀ (y) + T₀₀ (y) + T_inc (y) + T_sur (y) is determined by four components. Two major contributions reflect the effects of change in disease incidence (T_inc (y)) and relative survival (T_sur (y)) over time. The two remaining contributions reflect the effects from age (T₀ (y)) and time (T₀₀ (y)) boundaries. Similarly, partitioning for mortality is given by five components: $M^{'} (y) / M (y) = T_{μ} (y) + {\hat{T}}_{0} (y) + {\hat{T}}_{00} (y) + {\hat{T}}_{inc} (y) + {\hat{T}}_{sur} (y)$ , with specific components reflecting the change over time in disease incidence ( ${\hat{T}}_{inc} (y)$ ), relative survival ( ${\hat{T}}_{sur} (y)$ ), morbidity at 65 ( ${\hat{T}}_{0} (y)$ ) and 1992 ( ${\hat{T}}_{00} (y)$ ), and mortality in the general population (T_μ (y)). Change over time in the latter component influences the trend in the incidence-based mortality through its effect on other-cause mortality.

The analysis involves the design and estimation of separate models for 1) the incidence rate, 2) relative survival after AD/ADRD diagnosis as well as of individuals prevalent at 65 and/or 1992, 3) prevalence at the age boundary (age 65), 4) prevalence at the time boundary (year 1992), and 5) mortality in the general population. The partitioning components are expressed in terms of derivatives of these functions with respect to time. The model for any function contains parameters to describe the function for each year of diagnosis, and the parameters of B-splines (with equidistant knots including 4 inner knots and boundary knots on the year boundaries, the parameters of the boundary knots being equal to those of the closest inner knot) that are used to model the relationships between year-specific model parameters and evaluate the y-dependences of the function. B-splines allow explicit calculation of derivatives without requiring additional simplifying assumptions. The distribution of age (and time after onset for relative survival) is modeled using 1) the generalized 4-parameter Armitage-Doll model [18, 19] with additional individual predisposition parameterized by gamma or inverse Gaussian distributions [20] (for incidence), 2) the Weibull model [21, 22] for time after disease onset with the quadratic function of age for the shape and scale parameters (for relative survival), and 3) the Gompertz model (for mortality in the general population). These models are either standard for these outcomes or demonstrated better goodness-of-fit characteristics in respect to the alternatives (specifically, we tested survival models from refs. [23 –28] as well as all models defined in Section 7.3 of ref. [29]). The model parameters are estimated using non-linear least squares for age-specific incidence, mortality, and prevalence at boundaries, and the likelihood-based approach for relative survival [30].

RESULTS

Figures 2 3 compare the fitted incidence and relative survival rates predicted by our models with those empirically calculated from the data. Apart from 2006, the predicted and empirical rates showed a high level of agreement which allows us to conclude that the incidence and relative survival models that are used in the partitioning analysis were appropriate. Incidence of AD shows a local maximum in 2004 after which it smoothed out over the 2005–2013 period resuming its growth 2014 + . In contrast, ADRD reaches a maximum in 2012-2013 after which it declines.

Fig. 2

Incidence. Medicare-based empirical (dots) and modeled (lines) estimates of age-adjusted incidence, per 1,000, of Alzheimer’s disease (AD; red dots/lines), and Alzheimer’s disease related dementia (ADRD; blue dots/lines) for American older adults aged 65 + .

Fig. 3

Relative Survival. Medicare-based empirical (dots) and modeled (lines) estimates of age-adjusted relative survival, of Alzheimer’s disease (AD; red dots/lines), and Alzheimer’s disease related dementia (ADRD; blue dots/lines) for American older adults aged 65 + .

Figure 4 compares the predicted and empirical estimates of our models for prevalence and incidence-based mortality and presents trends in the components of these models. The modeled and empirical estimates were in agreement with each other. For AD, prevalence increased monotonically until 2011, flattening out thereafter. For ADRD, prevalence demonstrated a monotonic increase until 2015 followed by a gradual decline. Mortality for both conditions increases continuously over the entire study period. As expected, the main contributor to the prevalence of both conditions comes from the prevalence of individuals diagnosed during the study age-time area (P_is (y)). Among the two main contributions to incidence-based mortality, mortality caused by the presence of the disease, M_is (y), had a higher contribution compared to non-disease-specific mortality, M_Pμ (y). As expected, the contributions of the 1992 bound, P₀₀ (y) and M₀₀ (y), rapidly declined with time. Of interest is the gradual increase of the contribution of incidence at the age-boundary (P₀ (y)) suggesting an increase in the number of diagnoses before age 65.

Fig. 4

Prevalence and Incidence-based Mortality. Prevalence (P (y) = P₀ (y) + P₀₀ (y) + P_is (y)) and incidence-based mortality (M (y) = M_Pμ (y) + M₀ (y) + M₀₀ (y) + M_is (y)) models and their components: empirical prevalence/mortality (black dots), modeled prevalence/mortality (P/M, black lines), contribution of individuals with onset after age 65 and year 1992 to prevalence/mortality (P_is/M_is, green lines), contribution of the time-boundary cohorts to prevalence/mortality (P₀₀/M₀₀, pink lines), contribution of pre-existing prevalence at the age-boundary to prevalence/mortality (P₀/M₀, blue lines), contribution of relative mortality in the general population to mortality (M_Pμ, brown lines).

The results of the partitioning analysis for prevalence and incidence-based mortality are presented in Fig. 5 (the two most influential components, incidence, and relative survival, are shown) and Supplementary Figures 3 and 4 (all components are shown). Note that unlike Figs. 2 , and 4, these graphs show changes in rates rather than the rates themselves, therefore the difference between two points indicates the speed of the change, while positive/negative values indicate the direction of the change.

Fig. 5

Decomposition of Prevalence and Incidence-based Mortality. Change in prevalence/mortality (black dots), change in the contribution of incidence (green squares), change in the contribution of survival (red open dots).

Partitioning the rate of change in AD/ADRD prevalence showed that increases in prevalence between 1994 and their respective peaks (2011 for AD; 2015 for ADRD) were primarily due to increased incidence (T_inc (y)) and, to a much lesser extent, variations in relative survival (T_sur (y)). The relative contribution of incidence reached its maximum strength in 2001 for AD and 1998 for ADRD after which it steadily declined, although ADRD demonstrated a rebound to a lesser local maximum in 2009.

Increased incidence accounted for almost 100%of the prevalence increase in 1998–2000. Notable contributions from improved relative survival appeared in 2008 accounting for about 46%of the total change for AD (compared to 44%from incidence), and 25%for ADRD (compared to 65%from incidence). In all cases, the derivative of the incidence and relative survival functions was zero at the point where prevalence reaches its peak; after which AD/ADRD incidence and survival continue to decrease, but the incidence function of AD reverses its declining trend and begins to increase at a relatively high rate of change. The effect of pre-existing prevalence at the time-boundary (T₀₀ (y)) decreases over the study-time period becoming negligible in 2002, while the effect of pre-existing prevalence at the age-boundary (T₀ (y)) is negligible throughout the entire study time-period.

Incidence-based mortality for AD and ADRD increased over the entire study period at a rapidly decreasing rate, reaching their local minima in 2008 for AD and in 2004 for ADRD (the latter is followed by a rebound to a local maximum in 2011 with a subsequent decline). The effect of incidence ( ${\hat{T}}_{inc} (y)$ ) was dominant with respect to relative survival ( ${\hat{T}}_{sur} (y)$ ) though to a much lesser extent than observed for prevalence. The effect of pre-existing prevalence at the time-boundary ( ${\hat{T}}_{0} (y)$ ) decreased over the study-time period becoming negligible in 2002 while the effects of pre-existing prevalence at the age-boundary ( ${\hat{T}}_{0} (y)$ ) and mortality in the general population (T_μ (y)) were negligible over the entire study period.

The results of sex-specific decomposition of prevalence and mortality for AD are shown in Supplementary Figure 3. In females, incidence has a relatively stronger contribution to the overall prevalence trend than in males while survival demonstrates a relatively weaker contribution. Of note is that the effect of the 1992 cohort is almost entirely associated with the female subgroup. In contrast, the decomposition of mortality shows comparable trends between the two sexes. Sex-specific decomposition of ADRD (Supplementary Figure 4) shows similar relationships between male and female trends with two exceptions: 1) the contribution of the 1992 cohort is now notable in both males and females (although its effect in males is still weaker) and 2) the contribution of pre-existing morbidity at age 65 is consistently higher in males (although this effect is still small when compared to that of incidence and survival).

To test the stability of our results to variations in the way the study population was selected from the Medicare data and how the onset of AD/ADRD was defined we conducted several sensitivity studies. Supplementary Figures 5 and 6 show that our results were not sensitive to these variations.

DISCUSSION

In this study, we decomposed the historical trends in AD/ADRD prevalence and IBM into their epidemiologic causal components to identify the sources of the changes in the epidemiology of these diseases. The prevalence of AD was increasing between 1992 and 2011 after which it started to decline. IBM of AD increased over the entire study period, although with a significant slowdown in its rate of growth from 2011 onwards. For ADRD, both prevalence and IBM increased through 2014 prior to taking a downwards turn. Change in incidence was the primary factor determining the beneficial trends in AD/ADRD prevalence and IBM. The rate of growth in AD/ADRD incidence was slowing from about 2000 onwards, eventually switching directions and starting to decline on or about 2010. This trend was especially pronounced for ADRD. In terms of magnitude, the effect of incidence overpowered the effects (positive and negative) of changes in relative survival which exhibited a distinct wave-like pattern until about 2008, when incidence and relative survival began to have comparable effects on the overall trends in AD/ADRD prevalence and IBM. Other components such as changes in prevalence at 65 and mortality in the general population showed only minor effects.

Existing literature suggests that the prevalence [31 –36] of AD/ADRD in the U.S. may have declined in the past 25 years [31], though a consensus has not been reached [37]. One of the likely sources of these discordant findings, is differences in methods of AD/ADRD ascertainment across disparate data sources. Studies reporting an expected increase in AD/ADRD prevalence often used clinically diagnosed AD/ADRD, administrative data such as Medicare [38], and/or estimates derived from overall trends in U.S. population structure [4, 39] coupled with increases in life expectancy of people diagnosed with AD/ADRD [40 –44]. In contrast, many studies reporting declines in AD/ADRD prevalence were based on levels of cognitive impairment (CI), obtained through cognitive function evaluations, as indicators of AD/ADRD presence. Indeed, levels of CI in the U.S. have been decreasing, likely due to increasing levels of education and improved control of cardiovascular risk factors [4 , 45–47]. However, CI and AD/ADRD, although closely related are not always interchangeable. This is underlined by a recent study which used Medicare administrative data linked to the results of cognitive evaluations, to show that while the prevalence of AD/ADRD has been going up, the level of CI at the time of each new AD/ADRD diagnosis has been going down [48].

The majority of nationally representative estimates of mortality from AD/ADRD are based on death certificate data [4]; however, recent work has started to question the validity of such estimates [7 , 50]. In this study we cannot validate death-certificate-based rates directly, but we can use individual-level Medicare records to verify whether the trend patterns between the two sources are in agreement. The closest proxy for a death-certificate-based mortality rate obtainable with our data is the component of the total IBM model representing disease-specific mortality (M_is (y)).

Figure 6 shows mortality trends for AD, ADRD, and AD+ADRD combined based on estimates of M_is (y) from our model and the multiple cause of death database using both underlying and secondary causes. Overall, M_is (y) generally reproduces the patterns provided by death certificate data, especially for patterns of mortality from AD or ADRD combined (AD+ADRD). The difference in magnitude is explained by the inherent differences between an incidence-based and a cause-specific rate as well as known underreporting of dementia deaths in death certificates [34, 35]. Two other important distinctions exist: 1) when AD and ADRD are combined in the IBM model, the magnitude of the estimates does not change while the death-certificate-based rates of AD, ADRD, and AD+ADRD are more additive in nature; 2) the IBM model of AD does not show the dip in death-certificate-based AD rates in 2012-2013, but matches the corresponding uptake in ADRD diagnoses over the same time period. The cause of these differences is straightforward. Individuals diagnosed with AD in Medicare data almost always had a diagnosis of ADRD at some point of their Medicare history. In contrast, in death certificates, the presence of both AD and ADRD in the multiple-cause-of-death fields were relatively rare. This causes the addition of AD cases to the existing ADRD subgroup to have a much stronger effect in cause-of-death data than in Medicare data where most of the individuals contributing to AD IBM were already contributing to the ADRD IBM rate. A further, observation that can be made from Fig. 6 is that IBM from ADRD+AD can be lower than mortality from ADRD alone. In Medicare data this can occur only if a noticeable fraction of individuals with both an AD and ADRD diagnosis has AD diagnosed before ADRD. For these individuals, survival from AD+ADRD is better compared to survival from ADRD alone even though the date of death/censoring is unchanged. Combined with the higher levels of ADRD IBM mortality versus AD+ADRD in the 2010–2015 period, this provides additional evidence for the hypothesis of earlier ascertainment of AD [48]. Since the excess of mortality from AD+ADRD versus ADRD is detected in 2010–2016, this effect partly explains the dip in death-certificate-based AD rates in 2012-2013: a large fraction of diagnoses of AD was diagnosed earlier and the uptake in ADRD mortality is then just a compensation of the deficit of AD cases. However, this cannot give the complete explanation of this phenomenon, and it is likely that administrative factors such as cause-of-death coding habits and/or changing standards for placing an AD diagnosis may have also played a role.

Fig. 6

Differences in Age-Adjusted Mortality Rates. Mortality for Alzheimer’s disease (multiple cause of death-based: blue squares; Medicare-based: blue solid line), Alzheimer’s disease related dementia (multiple cause of death-based: red open dots; Medicare-based: red dashed line), and AD/ADRD combined (multiple cause of death-based: black dots; Medicare-based: black solid line).

Our study shows that the most influential epidemiological determinant of the trends of AD/ADRD prevalence and IBM over the 1991–2017 period was AD/ADRD incidence. We found that the contribution of AD/ADRD incidence has been increasing but at a decreasing rate (Fig. 5; green lines) until finally reaching a tipping point in 2010 for AD and 2013 for ADRD and then starting to decline. These results support recent findings on trends of claim-based AD/ADRD incidence obtained using Medicare data: increased incidence in 1994–2005 [51] with a maximum in 2009 followed by a decline in 2007–2014 [52], but do not directly support recent findings of decreased [53, 54] or unchanging/mixed [35 , 55] AD/ADRD incidence obtained in studies centered on non-nationally-representative cohorts. In our study, the incidence trend had two effects, the direct effect of relative decreases in incidence slowing and eventually reversing prevalence growth, and the indirect effect of the relative contribution of incidence falling to levels that allowed changes in AD/ADRD survival (Fig. 5; red lines) to play a visible role in determining the overall trend. Note that although change in incidence is the strongest factor in determining both prevalence and IBM trends, its effect on IBM is relatively weaker and occurs after a time lag; this allows disease specific survival to play an important role much earlier as the contribution of incidence declines.

Changes in survival after an AD/ADRD diagnoses was the second most powerful factor influencing prevalence/IBM although in absolute terms the effect of this component was relatively weak until 2008. Similar to the patterns of survival in Fig. 3, the time pattern of the contribution of survival (Fig. 5; both panels; red lines) has a wave form that is inverted for prevalence and IBM, representing the opposite effects that a change in survival has on prevalence/IBM trends. One explanation for the existence of such patterns was proposed in [14], where the authors detected similar patterns in their study of bladder cancer and concluded that they were generated by the effect of accumulated survivors who have successfully survived bladder cancer during previous years reaching the end of their natural lifespan. Generally, increased survival has two effects on IBM that, over time, push this rate in opposite directions: 1) the direct decline of mortality from the disease; and 2) increasing prevalence of survivors and associated delayed increase of mortality from other reasons. Which tendency dominates depends on the magnitudes of M_is (y) and M_Pμ (y). In the case of bladder cancer M_is (y) < M_Pμ (y) and therefore the effect of accumulated survivors was observed. In the case of lung cancer, this relationship was dominated by M_is (y) and no accumulation of survivors occurs [13]. In the case of AD/ADRD, M_is (y) > M_Pμ (y), though this difference is relatively small, so although the effect of accumulated survivors is expected, its magnitude is fairly minor.

Finally, the epidemiologic picture presented by the patterns identified in this study is characterized by increased incidence/prevalence of AD/ADRD until a maximum around 2008–2010 with simultaneous improvements in survival with a maximum in 2008–2010 after which survival started decline (see Figs. 2 –4 for age-adjusted rates and Supplementary Figure 7 for age-specific rates). The combination of increasing incidence combined with improved survival suggests that, ceteris paribus, diagnoses of AD/ADRD are occurring earlier in the disease’s course and therefore at earlier stages of severity. This picture is consistent with and provides additional support for the inferences on the prevailing trends in stage ascertainment and disease severity in AD/ADRD health outcomes made in [48]. Specifically, analyses of cognitive scores linked to Medicare records for the same individuals showed that until 2008, higher incidence rates of AD, ADRD, and other neurocognitive disorders were accompanied by a higher fraction of cases with better cognitive status which led to both higher incidence rates and improved survival. For AD this effect started declining after 2008 (Fig. 4 of [48]). The additional years of Medicare data available in this study suggest that deceleration of the effects associated with improved AD ascertainment identified in [48] have continued over the 2011–2017 period and that by 2017 the associated epidemiological effects have likely been exhausted.

We acknowledge several study limitations. The calculations were based on empiric parametric models for age dependence of incidence and fractions of stage at diagnosis as well as age and time after diagnosis dependence for survival. Although we carefully constructed such models based on their empiric properties or widely accepted modeling approaches, statistical uncertainties still remain. The base decomposition method used in this study cannot account for simultaneous changes in other behavioral, socioeconomic and health-related risk factors which also affect the observed epidemiological characteristics of AD/ADRD. An expansion of the decomposition method used in this study to account for this limitation is currently underway. The estimates of the model for relative survival is estimated based on a 60-month follow-up after the date of diagnosis; however, the real follow-up and potential follow-up time decreases as the time series approaches the cutoff date (12/31/2017). Medicare data is administrative billing data not designed for research and therefore has a number of known limitations. Most relevant is that this data is not linked to medical charts, autopsy reports, or other ways of independent validation of a diagnosis present in the claim. Combined with the difficulty of placing an AD diagnosis, this introduces a certain level of uncertainty in any estimate based on such data. To address this limitation, we included both AD and ADRD in our analysis with the understanding that the misclassification of some portion of individuals with AD through misdiagnosis is possible. Study of both the narrow (AD) and wide (ADRD) category of individuals with probable AD and comparing/contrasting the resulting trends gives us a better picture of the real-world situation. We have also conducted a number of sensitivity studies, varying the conditions under which disease onset is identified and the cut-off point at which an individual switching in and out of traditional Medicare (e.g., fee-for-service churn) was excluded from further analysis (Supplementary Figures 5 and 6) and found no major differences between the results presented in this study. Finally, available Medicare data does not capture individuals enrolled in Medicare Advantage—a private managed care alternative to traditional Medicare. Enrollees in Medicare Advantage are likely to have lower proportions of AD/ADRD since hierarchical condition categories involving dementia were not included in the models determining their compensation until 2020. However, the magnitude of the potential bias associated with this exclusion varies with the size of the MA population which decreased from about 18%of the total in 1999 to a low of 13%in 2005 and then demonstrated a persistent increasing trend doubling from 16%in 2006 to 33%in 2017 [56]. These statistics show that for the majority of the 2004 + period, individuals in MA were in the minority. Furthermore, we saw no shocks in prevalence, IBM, or any of their components over the 2004–2007 period. So, if any associated change occurred, it was gradual and did not stand out from the general trends over that time period. Therefore, even though it is still possible that deviation between mean health statuses of Medicare Advantage individuals and individuals under the traditional fee-for-service plans, could bias our estimates, we do not believe that the magnitude and direction of such potential bias would affect the conclusions of our study.

In sum, the primary determinant responsible for the beneficial trends in prevalence and IBM was the deceleration and eventual decrease in incidence rates of AD/ADRD, though the changes in relative survival began to affect the overall trends in prevalence/IBM in a noticeable manner after 2008. Based on the results of this study we expect the prevalence of ADRD to continue to decrease, however, the same cannot be said for AD: although AD incidence continues to decrease in the most recent years of data available, it is doing so at a decreasing rate suggesting a possible reversal in the near future. IBM for ADRD is expected to continue its decline in the future, while the direction of IBM for AD is not clear: based on our decomposition results alone it appears that AD IBM will increase in the near future; however, newly approved treatment may reverse this emerging trend. The results on time trends of epidemiologic measures identified in this study confirm the important role of changes in AD ascertainment of the epidemiological trends of this condition over the 1991–2008/10 period but show that the changes associated with this effect have been exhausted in the period 2011–2017. Finally, we conclude that mortality trends based on death certificate data are a reliable source of information, at least in application to AD/ADRD and in comparison with the trends derived from Medicare records.

Footnotes

ACKNOWLEDGMENTS

This study was supported by the National Institute on Aging (R01-AG066133, R01-AG046860, R01-AG057801) and the Department of Defense (W81XWH-20-1-0253). The sponsors had no role in the design and conduct of this study.

Authors’ disclosures available online ().

The supplementary material is available in the electronic version of this article: .

References

Matthews

, Xu

, Gaglioti

, Holt

, Croft

, Mack

, McGuire

(2019) Racial and ethnic estimates of Alzheimer’s disease and related dementias in the United States (2015–2060) in adults aged≥65 years. Alzheimers Dement 15, 17–24.

Wortmann

(2012) Dementia: A global health priority - highlights from an ADI and World Health Organization report. Alzheimers Res Ther 4, 40.

Hurd

, Martorell

, Delavande

, Mullen

, Langa

(2013) Monetary costs of dementia in the United States. N Engl J Med 368, 1326–1334.

Alzheimer’s Association (2020) 2020 Alzheimer’s disease facts and figures. Alzheimers Dement 16, 391–460.

Saydah

, Geiss

, Tierney

, Benjamin

, Engelgau

, Brancati

(2004) Review of the performance of methods to identify diabetes cases among vital statistics, administrative, and survey data. Ann Epidemiol 14, 507–516.

McEwen

, Pomeroy

, Onyemere

, Herman

(2008) Are primary care physicians more likely to record diabetes on death certificates? Diabetes Care 31, 508–510.

Stokes

, Weiss

, Lundberg

, Xie

, Kim

, Preston

, Crimmins

(2020) Estimates of the association of dementia with US mortality levels using linked survey and mortality records. JAMA Neurol 77, 1543–1550.

Plassman

, Langa

, Fisher

, Heeringa

, Weir

, Ofstedal

, Burke

, Hurd

, Potter

, Rodgers

(2007) Prevalence of dementia in the United States: The aging, demographics, and memory study. Neuroepidemiology 29, 125–132.

Taylor

Jr , Østbye

, Langa

, Weir

, Plassman

(2009) The accuracy of Medicare claims as an epidemiological tool: The case of dementia revisited. J Alzheimers Dis 17, 807–815.

10.

Goodman

, Lochner

, Thambisetty

, Wingo

, Posner

, Ling

(2017) Prevalence of dementia subtypes in United States Medicare fee-for-service beneficiaries, 2011–2013. Alzheimers Dement 13, 28–37.

11.

Akushevich

, Yashkin

, Kravchenko

, Fang

, Arbeev

, Sloan

, Yashin

(2017) Theory of partitioning of disease prevalence and mortality in observational data. Theor Popul Biol 114, 117–127.

12.

Akushevich

, Yashkin

, Kravchenko

, Fang

, Arbeev

, Sloan

, Yashin

(2018) Identifying the causes of the changes in the prevalence patterns of diabetes in older US adults: A new trend partitioning approach. J Diabetes Complications 32, 362–367.

13.

Akushevich

, Kravchenko

, Yashkin

, Fang

, Yashin

(2019) Partitioning of time trends in prevalence and mortality of lung cancer. Stat Med 38, 3184–3203.

14.

Akushevich

, Yashkin

, Inman

, Sloan

(2020) Partitioning of time trends in prevalence and mortality of bladder cancer in the US. Ann Epidemiol 47, 25–29.

15.

(2018) Underlying Cause of Death 1999-2017 on CDC WONDER Online Database, released December, 2018. Data are from the Multiple Cause of Death Files, 1999-2017, as compiled from data provided by 818 the 57 vital statistics jurisdictions through the Vital Statistics Cooperative Program. Centers for Disease Control and Prevention 821 (CDC), National Center for Health Statistics (NCHS). http://wonder.cdc.gov/ucd-icd10.html. Accessed on May 6, 2019.

16.

Akushevich

, Kravchenko

, Ukraintseva

, Arbeev

, Yashin

(2012) Age patterns of incidence of geriatric disease in the US elderly population: Medicare-based analysis. J Am Geriatr Soc 60, 323–327.

17.

Keyfitz

(1966) Sampling variance of standardized mortality rates. Hum Biol 38, 309–317.

18.

Licher

, van der Willik

, Vinke

, Yilmaz

, Fani

, Schagen

, Ikram

(2019) Alzheimer’s disease as a multistage process: An analysis from a population-based cohort study. Aging (Albany NY) 11, 1163.

19.

Webster

(2019) Multi-stage models for the failure of complex systems, cascading disasters, and the onset of disease. PloS One 14, e0216422.

20.

Kravchenko

, Akushevich

, Abernethy

, Lyerly

(2012) Evaluating the number of stages in development of squamous cell and adenocarcinomas across cancer sites using human population-based cancer modeling. PloS One 7, e37430.

21.

Carroll

(2003) On the use and utility of the Weibull model in the analysis of survival data. Control Clin Trials 24, 682–701.

22.

Brookmeyer

, Corrada

, Curriero

, Kawas

(2002) Survival following a diagnosis of Alzheimer disease. Arch Neurol 59, 1764–1767.

23.

Shayan

, Ayatollahi

, Zare

(2011) A parametric method for cumulative incidence modeling with a new four-parameter log-logistic distribution. Theor Biol Med Model 8, 43.

24.

Lemonte

(2014) The beta log-logistic distribution. Braz J Probab Stat 28, 313–332.

25.

Mudholkar

, Srivastava

(1993) Exponentiated Weibull family for analyzing bathtub failure-rate data. IEEE Trans Reliab 42, 299–302.

26.

De Santana

TVF

, Ortega

, Cordeiro

, Silva

(2012) The Kumaraswamy-log-logistic distribution. J Stat Theory Appl 11, 265–291.

27.

Ramos

MWA

, Cordeiro

, Marinho

PRD

, Dias

CRB

, Hamedani

(2013) The Zografos-Balakrishnan log-logistic distribution: Properties and applications. J Stat Theory Appl 12, 225–244.

28.

Hashimoto

, Ortega

, Cordeiro

, Hamedani

(2017) The log-gamma-logistic regression model: Estimation, sensibility and residual analysis. J Stat Theory Appl 16, 547–564.

29.

Manton

, Akushevich

, Kravchenko

(2008) Cancer mortality and morbidity patterns in the US population: An interdisciplinary approach, Springer Science & Business Media.

30.

Dickman

, Sloggett

, Hills

, Hakulinen

(2004) Regression models for relative survival. Stat Med 23, 51–64.

31.

Langa

, Larson

, Crimmins

, Faul

, Levine

, Kabeto

, Weir

(2017) A comparison of the prevalence of dementia in the United States in 2000 and 2012. JAMA Intern Med 177, 51–58.

32.

Sheffield

, Peek

(2011) Changes in the prevalence of cognitive impairment among older Americans, 1993–2004: Overall trends and differences by race/ethnicity. Am J Epidemiol 174, 274–283.

33.

Hudomiet

, Hurd

, Rohwedder

(2018) Dementia prevalence in the United States in 2000 and 2012: Estimates based on a nationally representative study. J Gerontol B Psychol Sci Soc Sci 73, S10–S19.

34.

Freedman

, Kasper

, Spillman

, Plassman

(2018) Short-term changes in the prevalence of probable dementia: An analysis of the 2011–2015 National Health and Aging Trends Study. J Gerontol B Psychol Sci Soc Sci 73, S48–S56.

35.

Rocca

, Petersen

, Knopman

, Hebert

, Evans

, Hall

, Gao

, Unverzagt

, Langa

, Larson

(2011) Trends in the incidence and prevalence of Alzheimer’s disease, dementia, and cognitive impairment in the United States. Alzheimers Dement 7, 80–93.

36.

Hall

, Gao

, Baiyewu

, Lane

, Gureje

, Shen

, Ogunniyi

, Murrell

, Unverzagt

, Dickens

(2009) Prevalence rates for dementia and Alzheimer’s disease in African Americans: 1992 versus 2001. Alzheimers Dement 5, 227–233.

37.

Rajan

, Weuve

, Barnes

, Wilson

, Evans

(2019) Prevalence and incidence of clinically diagnosed Alzheimer’s disease dementia from 1994 to 2012 in a population study. Alzheimers Dement 15, 1–7.

38.

Akushevich

, Kravchenko

, Yashkin

, Yashin

(2018) Time trends in the prevalence of cancer and non-cancer diseases among older US adults: Medicare-based analysis. Exp Gerontol 110, 267–276.

39.

Hebert

, Beckett

, Scherr

, Evans

(2001) Annual incidence of Alzheimer disease in the United States projected to the years 2000 through 2050. Alzheimer Dis Assoc Disord 15, 169–173.

40.

Tom

, Hubbard

, Crane

, Haneuse

, Bowen

, McCormick

, McCurry

, Larson

(2015) Characterization of dementia and Alzheimer’s disease in an older population: Updated incidence and life expectancy with and without dementia. Am J Public Health 105, 408–413.

41.

Helzner

, Scarmeas

, Cosentino

, Tang

, Schupf

, Stern

(2008) Survival in Alzheimer disease: A multiethnic, population-based study of incident cases. Neurology 71, 1489–1495.

42.

Xie

, Brayne

, Matthews

(2008) Survival times in people with dementia: Analysis from population based cohort study with 14 year follow-up. BMJ 336, 258–262.

43.

Brodaty

, Seeher

, Gibson

(2012) Dementia time to death: A systematic literature review on survival time and years of life lost in people with dementia. Int Psychogeriatr 24, 1034.

44.

Todd

, Barr

, Roberts

, Passmore

(2013) Survival in dementia and predictors of mortality: A review. Int J Geriatr Psychiatry 28, 1109–1124.

45.

Crimmins

, Saito

, Kim

, Zhang

, Sasson

, Hayward

(2018) Educational differences in the prevalence of dementia and life expectancy with dementia: Changes from 2000 to 2010. J Gerontol B Psychol Sci Soc Sci 73, S20–S28.

46.

Choi

, Schoeni

, Martin

, Langa

(2018) Trends in the prevalence and disparity in cognitive limitations of Americans 55–69 years old. J Gerontol B Psychol Sci Soc Sci 73, S29–S37.

47.

Stallard

, Yashin

(2016) LTC Morbidity Improvement Study: Estimates for the Non-Insured US Elderly Population Based on the National Long Term Care Survey 1984–2004. Society of Actuaries, Schaumburg, IL.

48.

Akushevich

, Yashkin

, Kravchenko

, Ukraintseva

, Stallard

, Yashin

(2018) Time trends in the prevalence of neurocognitive disorders and cognitive impairment in the United States: The effects of disease severity and improved ascertainment. J Alzheimers Dis 64, 137–148.

49.

Romero

, Benito-León

, Louis

, Bermejo-Pareja

(2014) Under reporting of dementia deaths on death certificates: A systematic review of population-based cohort studies. J Alzheimers Dis 41, 213–221.

50.

Romero

, Benito-León

, Mitchell

, Trincado

, Bermejo-Pareja

(2014) Under reporting of dementia deaths on death certificates using data from a population-based study (NEDICES). J Alzheimers Dis 39, 741–748.

51.

Akushevich

, Kravchenko

, Ukraintseva

, Arbeev

, Yashin

(2013) Time trends of incidence of age-associated diseases in the US elderly population: Medicare-based analysis. Age Ageing 42, 494–500.

52.

Kirson

, Meadows

, Desai

, Smith

, Cheung

, Zuckerman

, Matthews

(2020) Temporal and geographic variation in the incidence of Alzheimer’s disease diagnosis in the US between 2007 and 2014. J Am Geriatr Soc 68, 346–353.

53.

Satizabal

, Beiser

, Chouraki

, Chêne

, Dufouil

, Seshadri

(2016) Incidence of dementia over three decades in the Framingham Heart Study. N Engl J Med 374, 523–532.

54.

Sullivan

, Dodge

, Hughes

, Chang

C-CH

, Zhu

, Liu

, Ganguli

(2019) Declining incident dementia rates across four population-based birth cohorts. J Gerontol A Biol Sci Med Sci 74, 1439–1445.

55.

Hebert

, Bienias

, Aggarwal

, Wilson

, Bennett

, Shah

, Evans

(2010) Change in risk of Alzheimer disease over time. Neurology 75, 786–791.

56.

Jacobson

, Damico

, Neuman

, Gold

(2017) Medicare Advantage 2017 spotlight: Enrollment market update. KFF. https://www.kff.org/medicare/issue-brief/medicare-advantage-2017-spotlight-enrollment-market-update/ Accessed January 14, 2020.