Characteristics of the cohorts
All six studies chose their source populations according to geographical area and age. The Hisayama study, CFAS and the Cache County study selected for a population aged over 65, while HAAS selected for over 72, and CC75C for over 75. The Vantaa 85+ study included all individuals aged 85 and over thus representing the very elderly. Interpretation of findings must take these design strategies into account. For example, in a study of people surviving to 85+ years differential survival rates may affect genetic findings such as true homozygous for Apoe E4 individuals because they are at increased risk for premature death from cardiovascular disease and similar causes, and therefore likely to be underrepresented in populations of older individuals [
90]. Studying a very elderly population may introduce new sources of bias but also provide new insights.
By definition all six studies sought community-dwelling and institutionalised respondents. This requirement to include institutionalised persons in a representative community sample in dementia studies has been challenged. Based on a meta-analysis covering prevalence studies from 1945 to 1985, Jorm and colleagues found no effect on the overall prevalence rates of dementia [
91]. They argued that institutionalised individuals represent only a small proportion of the elderly population and thus effects are marginal. However, there are secular changes in institutional rates and variations across countries and cultures [
92]. In a German study of people over 75 (LEILA75+) it was found that institutionalised individuals were seven times more likely to be demented than community-dwelling individuals, but they were also older, more often single and less often married [
93]. In the CFAS population, 62% of those living in institutions were demented and represented 34% of those with dementia in the whole study cohort [
94].
Each study population has characteristics which influence how generalisable the findings are, but which may also relate to the rationale behind selecting the study area. The Cache County study sample was chosen because it was not typical of the present USA population as a whole. A great majority of the respondents did not smoke or drink alcohol and 91% were members of the Church of Jesus Christ of Latter-Day Saints – factors which may contribute to the average longevity among men in the study that exceeds the USA national average by almost 10 years. HAAS is also restrictive in the sense that it only includes men, who do not represent the majority of dementia sufferers in the population. It was created out of several cohorts with a different initial purpose, which was the investigation of the effect of migration on vascular risk. It could be argued that these two studies are more akin to the selected cohort studies than to the true population-based studies in that there may be limited generalisability from their findings. CFAS is the only multicentre study mixing populations from both rural and urban settings but has little ethnic variation.
To date these studies have not presented their findings with any attempt to weight back to the population. For now, study size, initial response rate, and attrition are important factors that determine whether a study using an autopsy endpoint is truly population-based. Large study populations are needed for adequate statistical power to detect relationships between variables, such as the effect of specific gene polymorphisms. Initial response rates and follow-up rates need to be high, and maintained, to retain the representativeness of the group. All six studies showed high initial response rates. These range from 79.8% (HAAS, sampled from the 65% follow-up cohort studied from 1965–1971), to 92% (Vantaa 85+ and CC75C) and a remarkable 99.7% for the Hisayama study in the 1992-screening wave. Attrition varied across the studies. Due to the greater age of the baseline Vantaa 85+ respondents had a higher death rate between each follow-up compared to the other studies. Approximately one-half of the subjects had died by the 3-year follow-up but 97.3% of survivors were re-examined [
95]. In CC75C, the two-year incidence wave achieved 67% follow-up of all respondents, and an 81% follow-up of all survivors [
23]. Similar reporting of attrition is made in CFAS, HAAS and the Cache County study publications.
The problem of non-response can be especially important for the population study of cognitive status in the elderly because cognitive compromise is a predictor of non-response. Non-response in the Cache County study was related to lower levels of cognitive functioning, as measured by the MMSE, than initial responders [
74]. They also reported that the sample under-represented females, younger individuals and people from other ethnic or religious backgrounds than members of the Church of Jesus Christ of Latter-Day Saints. The CC75C study also reported a lower baseline MMSE score (-1.17) in those who did not participate at follow-up, although 15% of these non-responders were initially interviewed but did not complete the assessment [
36]. It is likely that these MMSE scores were disproportionately low at the interview and so the effect of MMSE on not being interviewed at all will not be as high as -1.17 [
96]. In CFAS, looking at longitudinal attrition, predictors for drop out due to death were being older, male, having impaired activities of daily living, poor self-perceived health, poor cognitive ability and smoking. Similarly individuals who refused were more likely to have poor cognitive ability but had less years in full-time education and were more often living in their own home though less likely to be living alone [
97]. Further investigation into the characteristics of those lost to follow-up after initial enrolment was not reported by the other selected studies. This also includes investigating those lost to follow-up because of death. The Vantaa 85+ study and CFAS report that death certificates were checked to identify new dementia cases among those who were not demented in the last interview. This strategy is undermined because the recording of dementia on death certificates is notoriously unreliable [
98‐
100]. Other longitudinal studies of ageing and dementia have also suggested that mental status scores predict attrition [
101] so that the influence of non-responders on the findings should not be minimised. In spite of this, none of the six studies could be excluded according to this criterion. These are all large population studies dealing with the older population where in any case high attrition is expected and where ethical approval usually does not allow approach of those who decline future participation.
How representative were the autopsy samples compared to all deaths in the initial cohorts?
In volunteer cohorts such as the Nun Study, very high rates of autopsy amongst the deaths are achieved. How do population-based studies compare? The six studies vary in the proportion of brains collected in relation to their sample populations. The Hisayama study and Vantaa 85+ have autopsied between 70% and 82% (depending on years) and 52% of respondents respectively. These are remarkable autopsy rates in the context of a global decline in autopsies [
102]. HAAS' autopsies were discussed with all examined participants at the 1991–93 and 1994–96 examinations and the study has collected 559 brains so far (equal to 15% of the number of participants at the 1991 baseline). CFAS has performed 470 autopsies and another 250 are pledged. This is equal to around 40% of those approached to take part in the donation programme. Factors such as whole body donation, non-notification of death, coordination difficulties or changes of decisions at or before death, all result in erosion of the 'population representativeness' of the autopsy cohort. The possibility that systematic biases relate to the willingness of respondents to countenance brain donation has not often been studied. In one study of healthy elderly people, agreement to brain donation was related only to age and the Cornell Depression Scale score. Older respondents (≥ 85 years of age) were more likely to consent to donation than younger ones [
103]. Another study investigating the differences in pre-morbid clinical diagnoses between autopsied and non-autopsied dementia patients, the autopsied individuals differed in age, race, and interval between the last clinical contact and death compared to those who did not [
104]. Those patients with early onset of disease (<65) were more likely to consent to autopsy than those whose illness began somewhat later in life. A key motivating factor for agreeing to post-mortem examination is personal knowledge of an AD patient [
105]. These findings are reassuring and suggest that even though autopsies are voluntary, they do not seem to be skewed to one segment of the demented population.
HAAS, Vantaa 85+, CFAS and CC75C have all addressed the representativeness of their autopsy group. In HAAS the autopsy sample was similar to the total cohort in terms of sociodemographic terms, cognitive status and late life cholesterol levels. The autopsy rate was approximately 50% among the participants who had been recognised as demented with the autopsy rates being similar for the four clinically diagnosed subtypes of dementia [
45,
61]. Vantaa 85+ reported slight differences in the age and sex between the autopsied and the non-autopsied subpopulations [
27]. CFAS and CC75C used sampling strategies that enriched the autopsy sample to include older and more impaired respondents but other than these characteristics there was no clear indication that the donor sample differed from the general population in basic sociodemographic indicators [
20,
34].
Even though brain donations may not be biased regarding dementia status, other factors could introduce error into the autopsy process such as lack of notification of the study team because of the place of death: hospital, institution or home. The rapid accessibility to the body and the idea that it is likely that cognitive decline has a difference course in patients who are institutionalised compared to those who remain at home until the end of their life both play a part in the generalisability of the findings. One must assess the impact of only including a small proportion of individuals who die at home in an autopsy study [
106]. HAAS clearly reported the place of death of its cohort members whom have undergone autopsy, this being 71% in the hospital. 13.1% at home, 11.2% in a nursing home and 4.6% in a hospice or other place [
61] although none of the studies have reported on 'missed' autopsies.
Case identification, examination and in life diagnosis
Even if the initial cohort is population-based, studies may lose their population representation because of their case identification strategy. This arises if they rely on studying cases that have already come in contact with the health care system. For example, one of the excluded studies, the University of Washington Alzheimer's Disease Patient Registry (ADPR), uses a base population comprising members of a Group Health Cooperative with an attrition rate less than 1% per year. However, cases are found through the primary care physician network, referral from neurologists and mental health services or by monitoring events leading to the entry of demented patients into the health-care system [
107]. Dementia cases remaining in the community are therefore under-represented. There is no organised population-based or primary-care effort directed at the early detection of dementia in the USA and in the UK primary care physicians have reported that fewer than 50% have adequate basic and post-qualifying training for identifying dementia [
108]. One can fairly assume that many dementia cases go unnoticed in the health care system and that such a design results in a sample with limited population representation.
All the studies reported here have used direct interviewing of the populations identified with standardised assessments (Table
2). Clinical assignment of a dementia diagnosis was undertaken in three of the studies by consensus between expert clinicians. The Cache County study set up diagnostic conferences including geropsychiatric, neurology, neuropsychology and cognitive neuroscience expertise and the final diagnosis was achieved using information from all phases of the study including the autopsy results. HAAS and Vantaa 85+ achieved clinical diagnosis through smaller clinician panels and CC75C used a single psychiatrist with inter-rater reliability tests. CFAS based its findings on a previously validated clinically based computer-assisted taxonomy (AGECAT) administered by trained lay research interviewers [
109].
Neuropathological case confirmation
The comprehensiveness and comparability of the neuropathological investigation varies more in population settings. All more or less closely followed the Consortium to Establish a Registry for Alzheimer's Disease (CERAD) neuropathology protocol [
13]. This is in contrast to the broader variation in methods for clinical diagnosis (Table
2). Details of neuropathological investigations are yet not available in the published literature for the Cache County Study. In Vantaa 85+, besides completing the CERAD protocol, the NIA-RIA criteria was also adopted which emphasises topographic staging of neurofibrillary changes in addition to neuritic plaques [
110,
111], and the investigators routinely sampled additional areas of the brain supplementing the usual techniques with immunocytochemical methods to enhance detection of amyloid and Lewy bodies. Regarding how many neuropathologists carried out the examination: in CFAS, the Hisayama study and HAAS, postmortem investigation was performed by several neuropathologists, while in the Vantaa 85+ study, for one study looking at brain infarctions, the examination was done by one pathologist. No information is given for the Cache County Study. Comparability of assessment has been checked within CFAS and HAAS.
The CERAD protocol gives guidelines for tissue fixation, tissue processing, sectioning and tissue staining. It provides a simple semiquantitative systematic gathering of information on clinical, neuropsychological and neuropathologic aspects reducing subjective interpretation [
13,
112]. Nonetheless, the assignment of probability of a diagnosis of AD by the CERAD approach is based on concepts of clinicopathological correlation derived from cohort studies of selected cases and 'controls' so that it may represent a biased view of what burden of lesions are associated with a positive diagnosis for dementia. For instance, CERAD uses abundant neuritic plaques as a reference for AD case definition, reinforcing the concept of AD as a disorder of amyloidosis. The relationship between lesion burden and cognitive status, and the interactions between lesions, differs in a population sample from those in selected secondary referral cohorts of demented people [
34]. It is increasingly acknowledged that the proportion of normal elderly individuals with a substantial neuropathological burden is not negligible [
34,
106,
113]. The logical aim of neuropathological studies of dementia, if no neuropathological 'gold standard' exists, would be to collect information in an unbiased way to test which diagnostic criteria are most valid and reliable.
The CERAD protocol was developed in the 1980s and is based on a paradigm which heavily relied on the significance of lesions such as neurofibrillary tangles and amyloid plaques as the pathological substrate of cognitive decline. The accumulation of data from many studies over the last 15 years underline the limitations of this approach. The CERAD protocol ignores a number of pathological and biochemical advances which have identified additional targets, or refined the basis for quantifying pathologies, that are likely to have a closer link with cognitive decline or are additional pathologies that are not taken into account. These newer approaches include: biochemical estimations of soluble amyloid peptide load and aberrant tau accumulation in brain regions, synaptic density measurements, glial responses such as astrogliosis and microgliosis, severity of white matter attenuation. Finally CERAD provides no basis for assessing the significance of microvascular pathologies in the causation of cognitive decline.
A significant limitation is imposed through the necessity to adopt a validated pathology protocol with documented inter-laboratory comparability. In CERAD, 83% of raters showed consistency for plaque determination, whereas only 66% showed consistency for tangle determination. Despite this, apart from bias resulting from knowledge of dementia status before death, most of the other types of variation should produce non-systematic bias and therefore underestimate any relationship [
11]. Mirra has remarked that 'whether this variation is produced by differences in the phase of the moon or other indeterminate factors remains unclear, but, in my view, standardisation of methodology will be difficult to achieve' [
114]. The six studies report few measures taken to assure or measure quality in the neuropathological investigation. Blinding of the neuropathologist to clinical and risk factor data is reported for Vantaa 85+, CC75C and CFAS. Neuropathological inter-rater reliability was only described by CFAS.
Timing of data collection
Each study up-dated their data in life around every 2 years. In population studies it has been accepted that it is not feasible to have regular clinical follow-up until death for each member of a large study cohort despite the desirability of minimising the time interval between the last examination and death [
106]. This interval is critical in determining whether pathological features at autopsy are a reflection of the cognitive state at last interview. Health checks were done at least once a year in the Hisayama study. For Vantaa 85+ the average length of time from the last clinical examination to death was 0.99 y (range 0.05 – 2.35 y), for HAAS, the average was 2.4 y (SD 0.9 y) and for CFAS 1.2y (range of 3 d – 4.2 y; 81% within 2 y). In CC75C the study design and timing of the interview waves suggests the interval between last clinical examination and death is usually less than 3 years unless follow up interview was refused. Details have not been published yet for the Cache County study. Using various informant interviews may help in gaining a truer cognitive picture of the participant if delays between follow-up and death are considered too long. Studies which report the use of informant interviews can be found in (Table
2).
The timing of autopsy and methods of handling after death are also important for molecular studies. Formalin fixing plays a role in any future neuropathological examination and affects the sensitivity of the immunocytochemistry results. This variable depends on the work force available to complete the neuropathological examinations at the time of death and is thus dependent on the level of funding for the study and the relative importance given to such studies in their research and cultural contexts. None of the studies have reported time between death and autopsy although this is usually available in mortuary records.
Comparing and combining studies
Research based on community samples has the advantage that it allows the calculation of attributable risk, population attributable risk and population excess risk, which can all estimate how important a given factor is in a particular population. Each study has used similar neuropathological standards but this does not necessarily imply that the results are directly comparable. This is illustrated by the example of the classification of 'possible' and 'probable' cases of dementia. HAAS has described mutually exclusive subcategories of dementias: AD as primary cause, Vascular Dementia (VaD) as primary cause, mixed AD/VaD and other types of dementias [
63] rather than risking misclassification of mixed cases; whereas in the Vantaa 85+ study, no 'mixed dementia' category is mentioned [
27]. Combining of clinical information should be possible if one goes back to specific measures within each study. Unfortunately there is little overlap in the areas investigated amongst the six existing population-based neuropathological studies so that a systematic review of their results is not possible.
Redefining the field
Despite all the effort put into setting up the studies and gathering evidence, these datasets remain as yet under-tapped resources. So far most population-based studies have used post-mortem information to test or validate clinical diagnoses. Few have taken the opportunity to challenge and re-characterise existing criteria. Until relatively recently clinicopathological studies have concentrated on clinically clear-cut cases of AD but these population-based studies have confirmed that there is considerable overlap in pathologies found in the demented and non-demented [
20,
34,
45]. For example in HAAS, 33% of demented subjects' condition could not be attributed to any of four primary pathogenic processes or to a combination of them (vascular lesions, AD lesion patterns, hippocampal sclerosis and cortical Lewy bodies). The same heterogeneity of lesions was found in CC75C. In CFAS' first 209 autopsies, both cerebrovascular (78%) and Alzheimer type (70%) pathology were common. Dementia was present in 48% of whom 64% had features indicating probable or definite AD. 33% of the non-demented people, however, had equivalent densities of neocortical neuritic plaques. Some degree of neocortical neurofibrillary pathology was found in 61% of demented and 34% of non-demented individuals. Vascular lesions were equally common in both groups, although the proportion with multiple vascular pathology was higher in the demented group (46% vs. 33%). In the same way, Vantaa 85+ findings show that the prevalence of neuropathologically defined AD was 33% whereas the prevalence of clinically diagnosed AD was 16%. Although there was a correlation between clinical AD and neuropathological AD, 55% of individuals with neuropathologic AD were either in the non-demented group or in the Vascular dementia (VaD) or other clinical non-AD groups. On the other hand, 35% of those with clinical AD did not fulfil the neuropathologic criteria for AD used in the study. The Cache County study reported 21% of mixed dementia diagnosis using a consensus between clinical and autopsy findings.