APOE-ε4 count predicts age when prevalence of AD increases, then declines

The sample.

We studied the elderly (age ≥65) permanent resident population of Cache County, Utah, as of January 1, 1995. Being 91% members of the Church of Jesus Christ of Latter-day Saints, these individuals have a lifestyle that results in low rates of diseases associated with the use of alcohol and tobacco (several common cancers, hypertensive and atherosclerotic cardiovascular disease).9 Accordingly, they have low rates of mortality before age 85, notably for males whose median life expectancy at age 65 is the highest in the United States,10 and exceeds national norms by almost 10 years.11 The low burden of chronic disease also simplifies the differential diagnosis of dementia, especially among the oldest-old. Cache County has low rates of in- and out-migration, and the community is close-knit, with large families. These are advantages for sustained contact in longitudinal studies. The local culture is supportive of research, and its elderly had previously12 shown rates of participation exceeding 90%.

Estimating from the 1990 US Census, Utah state officials suggested there would be 5,650 eligible individuals, 746 aged 85 or older. We actually identified 5,956 individuals, representing 5,713 Medicare enrollees (list provided by the Health Care Financing Administration) and 243 others. When fieldwork began in April 1995, 98 individuals had moved away, and 181 had died before first contact. Of the 5,677 eligible individuals, 26 (0.5%) died before completion of enrollment, and a further 559 (9.8%) either refused or were not found. Thus, we enrolled 5,092 individuals (89.7% of those eligible). The various stages of investigation, depicted in figure 1, were as follows:

• Download figure
• Open in new tab
• Download powerpoint

Figure 1. Flow chart for screening and diagnostic procedures. 3MS = modified Mini-Mental State Examination; IQCODE = Jorm’s Informant Questionnaire on Cognitive Decline in the Elderly.

Diagnostic procedures.

Using procedures approved by our Institutional Review Boards, we attempted to identify and diagnose all prevalent dementia cases in the sample. We obtained informed consent at each stage from the subject or, where appropriate, from a responsible caregiver. Studies began with a baseline screening and risk factor interview in the participant’s home or in a nursing home (265 individuals). At the same visit we collected buccal scrapings from 4,932 participants (97%) for APOE genotyping.

Cognitive screening.

Cognitive screening relied mainly on an adaptation of the 100-point modified Mini-Mental State Examination (3MS)13 administered at the baseline interview. We used alternate versions with two new word list recall tasks of equivalent difficulty (validated in pilot work) as well as the original task. We adjusted 3MS scores for sensory deficits, discarding wrong or missing answers attributed to sensory difficulty and then calculating the percentage correct among the remaining items. To improve efficiency, we also normed results such that individuals with varying degrees of educational attainment obtained roughly equal mean scores.14

Proxy screening.

We administered proxy interviews for 386 individuals in any of four circumstances: 1) the individual could not complete the 3MS; 2) the individual scored below 15 of a possible 20 on a subset of orientation questions early in the interview; 3) the individual scored below 60 on the 3MS (and was therefore judged unable to provide reliable risk factor information); 4) the individual’s information was otherwise judged unreliable. Proxy interviews screened for cognitive difficulty using Jorm’s Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE),15 which assigns scores between 0 (no impairment) and 5 (extreme impairment). The proxy interview included the same questions about risk factors. Although the remaining interviews were conducted by proxy, we obtained full 3MS scores of <60 from 81 individuals.

Dementia questionnaire.

When the respondents’ 3MS scores were below 87 (the 25th percentile) or their IQCODE scores were above the recommended cutoff of 3.27, we investigated their cognitive functioning further by telephone interview of knowledgeable informants using the Dementia Questionnaire (DQ). The DQ is a 50-item semi-structured inventory of dementia symptoms and medical history designed to assist in the differential diagnosis of dementia.16 We also attempted a DQ interview for all individuals aged ≥90, regardless of their results on the 3MS or IQCODE. DQs were scored by a consensus of two or more clinicians as follows: 1) no impairment; 2) mild dysmnesia or other mild difficulty; 3) moderate cognitive difficulty probably not meeting criteria for dementia (e.g., only one cognitive domain impaired); 4) questionable dementia; 5) probable dementia.

Clinical assessment.

Individuals with DQ ratings of 4 or 5 were considered to have suspected dementia and were asked to undergo clinical assessment (CA). A research nurse and psychometrician conducted these assessments in the presence of a collateral informant at the individual’s residence (including nursing homes). The nurse recorded a narrative history of cognitive symptoms, a medical history, and current medications. The nurse then examined the individual using a brief protocol that included a standardized blood pressure measurement and a standardized neurologic examination.17 The psychometric technician interviewed the informant using the Neuropsychiatric Inventory18 and assessed cognitive and functional impairment using the Dementia Severity Rating Scale.19 The technician also obtained a family history and administered a 1-hour battery of neuropsychological tests17 to the individual. A 7-minute videotape segment recorded a brief standardized examination of mental status, insight, hand praxis, and gait.

Initial diagnostic conferences.

The results of all CAs were reviewed at initial diagnostic conferences. These conferences included a neuropsychologist (J.T.T.), a board-certified geriatric psychiatrist (J.C.S.B. or D.C.S.), and (except when no cognitive impairment was suspected) the examining nurse and technician. The psychiatrist dictated a chronology of the individual’s cognitive symptoms (if any) and the medical history (in standardized format, with date of onset for each condition when available). The psychiatrist’s note also recorded current medications, family history, results of the nurse’s physical and neurologic examinations and the neuropsychological battery, and a modified Hachinski ischemia score.20,21 Dementia was diagnosed using DSM-III-R criteria,22 except that we did not insist on a demonstrable deficit in both short-term and long-term memory. AD diagnoses followed NINCDS-ADRDA criteria,23 with the exception that a diagnosis of probable AD was deferred for neuroimaging results if these were forthcoming. Vascular dementia (VaD) diagnoses followed the NINDS-AIREN criteria24 as operationalized by Tatemichi et al.25 The conferees assigned a preliminary diagnostic category from seven broad categories as follows:17 probable AD,23 possible AD,23 dementia of unknown etiology (includes AD in the differential), demented-not AD, mild/ambiguous (denotes cognitive disorder not meeting criteria for dementia, with suspicion of prodromal AD), other (other cognitive syndrome not likely to represent prodromal AD, e.g., schizophrenia, stroke), and noncase (no notable cognitive compromise). These assignments reflected the clinical history and apparent functional impairment at least as much as neuropsychological test results. Conferees also recorded a rating on the Clinical Dementia Rating scale26 and estimated age at onset, defined as the year when the individual unambiguously met DSM-III-R criteria for dementia.

Follow-up laboratory testing and neuroimaging.

Follow-up laboratory testing and neuroimaging were requested for all individuals with the initial classification of probable AD or possible AD, dementia of undetermined etiology, or non-AD dementia. Tests completed on 189 individuals (62% of living individuals) included complete blood count, routine chemistries (CHEM-20), serum B₁₂, folate, thyroid function tests, and urinalysis, as well as standardized brain MRIs or (in four patients) CTs.

Physician examinations.

A board-certified geriatric psychiatrist (J.C.S.B. or D.C.S.) also examined demented individuals at their residence. These physicians were kept unaware of the results of the nurse’s neurologic examinations and the initial working diagnoses. They repeated the standardized neurologic examination and tested individuals with the Mini-Mental State Examination,27 also reviewing the clinical history (including interval history since assessment) with collateral informants and reviewing the Neuropsychiatric Inventory. Physicians visited 257 individuals (84% of those living when examinations were attempted) with working diagnoses of dementia. In 167 instances (65%) the physician’s diagnostic judgment agreed with the initial diagnosis. In other instances, discrepancies usually were minor (e.g., possible AD versus probable AD).

Final clinical diagnoses.

Final clinical diagnoses were assigned after review of all available information at consensus conferences that included J.C.S.B. and D.C.S., a board-certified neurologist, J.T.T., K.A.W.-B., and B.L.P. We now sought greater specificity, assigning diagnoses from a list of more than 30 categories (many indicated in figure 2). When making these assignments, we did not consider various dementing illnesses (notably AD and VaD) as exclusive if evidence suggested both were present. Apart from dementia diagnoses, we recorded a diagnosis of prodromal AD in individuals with VaD or other disorders who had shown periods of Alzheimer-like progressive cognitive decline (with no apparent responsible vascular insults). We also applied this term to individuals with progressive mild/ambiguous cognitive syndromes17 thought to probably represent early-stage AD. Similarly, we noted incidental strokes or minor vascular changes (either a history of TIAs or MRI evidence of multifocal white matter disease) in individuals with primary diagnoses of AD or other neurodegenerative conditions. Conference participants were required to note a rating (sometimes “no evidence of AD” or “no evidence of vascular pathology”) under headings of AD and VaD. When raters’ opinions varied, these were discussed until we reached a consensus.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2. Differential diagnostic assignment of 335 cases of dementia. The diagnoses of AD and of vascular dementia (VaD) or other specific disorders were not considered exclusive if individuals met criteria for both. Poss = possible; Prob = probable; Def = definite; PD = Parkinson’s disease with dementia; TBI = traumatic brain injury of sufficient extent to provoke dementia; Dem Unkwn Etiol = dementia of unknown etiology; ETOH Dem = alcoholic dementia; NPH = normal pressure hydrocephalus; Pick’s = Pick’s disease.

Postmortem studies.

We obtained names and dates of death for all deceased residents from vital statistics records and from daily review of newspaper obituaries (to capture deaths out of area). To date we have obtained 41 autopsy diagnoses made using CERAD criteria.28,29 Before their death, we had completed a CA and a variable degree of follow-up investigation on 34 of these individuals.

APOE genotyping.

Following the method of Richards et al.,30 APOE genotypes were determined using PCR amplification and a restriction isotyping method described by Saunders et al.31 APOE genotypes were not known to clinicians during the diagnostic process.

Fully examined subsample.

With intent to undertake a nested case-control study, we used iterative sampling without replacement to construct a stratified subsample of 960 enrollees. Regardless of their prior results on screening or DQ interview, each member of the subsample was asked to undergo CA and diagnostic review. The 838 subsample members who completed all phases of investigation (87.3%) included 2 nondemented individuals matched to each of the county’s prevalent AD cases by sex, 5-year age group, and number of ε4 alleles (4 individuals matched to cases aged <75 without APOE ε4/ε4).

Analyses.

We compared mortality and other attributes of responders versus nonresponders at successive stages of the protocol using t-tests for independent samples, chi-square tests, or Fisher’s exact test. To estimate the sensitivity of the 3MS/IQCODE screening procedure and the DQ, we considered the 193 demented members of the subsample who had screening scores. Of these, 190 had positive 3MS/IQCODE screening results (sensitivity 0.984). Two of these individuals lacked a DQ score; of the remaining 188, 162 had DQ ratings of 4 or 5 (conditional sensitivity estimate 0.862).

We estimated dementia prevalence after adjusting the observed number of cases for response rates and sensitivity of the screening measures. We first adjusted for response rate at the CA, 0.864. There was no further adjustment for the subsample who were all assigned for assessment, regardless of screening results. For others, we adjusted for completion rate at the DQ, 0.933. We also adjusted for sensitivity of the screening procedures (for individuals over age 90, we adjusted for the sensitivity of the DQ only; for others, we adjusted for the sensitivity of the 3MS/IQCODE and DQ in sequence, 0.848). Adjustment thus followed the following formula: where prevalence_i denotes prevalence in the i th stratum, c_i is the number of cases observed in the i th stratum of the subsample, r_i is the number of cases in the remainder, d is the response rate for the DQ, s_i is the estimated sensitivity for the i th stratum (see above), a is the completion rate at CA, and n_i is the number of individuals screened in the i th stratum. We estimated standard errors for prevalence in each stratum assuming a hypergeometric distribution for the number of cases in the stratum (from multi-stage sampling without replacement) and a binomial distribution for the estimated stratum prevalence (assuming Cache County is itself a sample from a much larger population.) We used Cochran’s adaptation of the Mantel-Haenszel procedure or Fisher’s exact test for estimation and evaluation of age-, sex-, and genotype-specific prevalence ratios and for bivariate comparison across various strata of the association between APOE and specific dementing disorders. Multiple logistic regression analyses employed the SAS statistical package. All statistical tests were two-tailed.

Final diagnoses and autopsy confirmation.

Figure 2 shows the distribution of dementia diagnoses, including autopsy diagnoses returned for 41 participants. Among the 34 individuals who had undergone CA before death, the positive predictive value of a clinical diagnosis of (probable or possible) AD, with or without comorbid diagnoses such as VaD or PD, was 85%. Four participants had a clinical diagnosis of AD that was not confirmed at autopsy. These included two with final diagnoses of diffuse Lewy body disease, one with autopsy findings “consistent with a history of acute hypoperfusion” (our clinical impression was that this event had triggered a typical Alzheimer process), and one with VaD (clinical diagnosis was also VaD but included a secondary diagnosis of AD that was not confirmed.) Five cases of definite AD had received other diagnoses before death. These included one individual with a clinical diagnosis of Parkinson’s dementia and one with VaD, both confirmed at autopsy but with an added secondary pathologic diagnosis of AD. One had been diagnosed clinically with frontal lobe dementia, and two had their clinical diagnosis deferred in the category “dementia, undetermined etiology.”

Prevalence estimates.

Table 2 shows the prevalence of AD and of all dementias by sex and 5-year age groups as estimated by equation (1). The prevalence figures are similar for men and women until age 85. In univariate comparisons the higher prevalence of dementia in women after this age was statistically significant (χ² = 6.20, df = 1, p = 0.02), but the individual comparisons at ages 85 to 89 and at ages 90+ were not. APOE genotype was not a strong predictor of prevalence at these late ages. There were 4 ε4/ε4 men and 10 ε4/ε4 women aged 85 and older. Eleven of these 14 individuals were examined, but only 4 were demented (3 with AD). The remainder, including one woman now aged 93, had little or no cognitive difficulty. There was a trend (odds ratio [OR] 2.39, confidence interval 0.83 to 5.39) toward association between genotype ε4/ε4 and VaD (either alone or mixed with AD).

Multiple logistic regression models.

These models were used to describe the probability π of a binary outcome (here, “caseness”) as a function of several terms. Each term in a given model denotes a variable attribute (e.g., age) multiplied by an estimated coefficient (e.g., β_age) that indicates the strength and direction of the attribute’s association with the outcome. We began using simple models with few terms. We then added terms in sequence, testing at each step for improved fit of the model using a likelihood ratio (LR) χ² statistic with df equal to the number of terms added. As is customary, we added additional terms as long as the LR χ² that compares the more complex model with its predecessor was statistically significant. We calculated ORs for the several attributes in each model as the natural logarithm of their estimated coefficients. Table 3 begins with our simplest model (model 1) that included only the effects of age, sex, and education. Better-fitting, more complex models added terms for presence of one or two ε4 alleles (model 2) and a term for age-squared (model 3). We next tested for effects of 2-way interactions (nonadditive effects) between age and presence of one or two ε4 alleles (model 4). Finally, we tested for interactions between sex and presence of one or two ε4 alleles (model 5). Throughout this exercise, we found that models that considered number of ε4 alleles fit much better than models with a binary (yes/no) rating for presence of ε4. Use of separate terms for presence of one or two ε4 alleles (model 2) yielded marginally better fit than a similar model with a single variable for ε4 count (values 0, 1, 2). However, separate terms produced better fit in the more complex models with interaction terms.

The improvement in logistic models with a term for age-squared was notable. This term, with its negative coefficient (OR <1), is responsible for the eventual decline in modeled prevalence (π) in all groups (figure 5). The positive coefficient for the linear age term β_age has a substantially larger absolute value than the negative coefficient for age-squared β_age². Modeled prevalence thus increases at first with age until the absolute value of the positive term (β_age × age) equals that of the negative term (β_age² × age²). The modeled prevalence declines thereafter.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5. Probability-by-age (π) that subjects will have prevalent AD predicted from a multiple logistic regression model (model 5 from table 4). Squares represent men; circles represent women. Black symbols indicate no ε4; gray symbols, one ε4; open symbols, two ε4’s. There was no difference in π between men and women with no ε4. Significant interaction terms for age-by-one ε4 and age-by-two ε4’s resulted in distributions around younger mean ages for these groups. Inclusion of interaction terms for sex-by-one ε4 and sex-by-two ε4’s suggested higher prevalence among women in these groups, especially ε4 homozygotes. In all groups, π reaches a maximum and then declines at later ages.

The significant interaction terms for age-by-one ε4 and age-by-two ε4’s reflect distinct distributions for ε4 heterozygotes and homozygotes, also depicted in figure 5. There was a trend toward further improvement of the model with addition of 2-way interaction terms for sex-by-one ε4 and sex-by-two ε4’s. With these terms, the main effect of sex (originally significant with OR 1.35, p = 0.048) vanished. This result suggests that the association between female sex and AD applied exclusively to those with ε4. Two further analyses corroborated this interpretation. A regression model based on the 3,339 individuals without ε4 showed no effect of sex (OR = 1.04, p = 0.88); by contrast, a model based on the 1,593 individuals with one or two ε4’s revealed significant association of AD with female sex (OR = 1.58, p = 0.02, table 4). Again, both of these models were improved by addition of a term for age-squared.

Figure 6 shows ORs plotted by age for the association of AD with one or two ε4’s (versus a reference group lacking ε4). In both groups, the association with AD varies strongly with age. At younger ages the OR with two ε4’s is extreme (there were no prevalent AD cases before age 70 in the reference group); the corresponding OR falls below unity at age 90. AD appears later in those with one ε4, so that the curves cross between ages 81 and 82. To investigate whether the crossing of the curves was merely an effect of extrapolation, we performed a truncated analysis in individuals 82 years of age and older with at least one ε4. Adding terms for the main effect of two ε4’s and interaction of this genotype with age did suggest lower AD prevalence in ε4 homozygotes, but the more complex model did not improve fit significantly (LR χ² = 1.80, df = 2, 0.25 < p < 0.5).

• Download figure
• Open in new tab
• Download powerpoint

Figure 6. Odds ratios for AD by age with one ε4 (open circles) and with two ε4’s (black squares) versus individuals with no ε4 as the reference group. The odds ratios were derived from multiple regression model 5, table 3. Age-related heterogeneity in the risk estimates is marked, especially for the group with two ε4’s. The odds ratio for ε4 heterozygotes remains above unity at all ages before 100. The figure suggests that the odds ratio for ε4 homozygotes dips below unity after age 89.

Population attributable risk for AD.

This is an estimate of the proportion of a population’s burden of disease that can be attributed to the presence of a risk factor. It depends on the strength of the factor’s association with disease (relative risk) and the population frequency of the factor. We used the adjusted ORs for one or two ε4’s from logistic model 5, along with the frequency of these genetic groups, to estimate the attributable risk for AD in the presence of one or two ε4 alleles. Using Levin’s formula,32,33 the risk attributable to genotypes including ε4 is estimated at 0.705, obtained by adding estimated risks of 0.526 and 0.179 for the disjoint groups bearing one ε4 and two ε4’s, respectively.