Missing‐Data in Epidemiological Studies of Age‐Associated Cognitive Decline

Dementia is becoming a major public health problem as the life expectancy of the population continues to increase. In 1997, 2.32 million Americans were affected by Alzheimer's disease, which accounts for almost 50% of all the cases of dementia. The prevalence of persons living with Alzheimer's disease is 4.3%, 8.5%, 16.0%, and 28.5% among 75-,80-, 85-, and 90-year-olds, respectively. Projections from large epidemiological studies indicate that, in the next 50 years, the annual number of new cases of Alzheimer's disease is expected to increase from 360,000 in 1997 to 1.14 million in 2047, a threefold increase.1 The trends for non-Alzheimer's dementia will likely show similar or only slightly lower figures. The longitudinal study by Bryane et al., published in this issue,2 suggests that these epidemiological data, although dramatic, probably still underestimate the true burden of dementia in older populations. The authors examined the change in cognitive performance of an initial cohort of 2130 unselected 75-year-old persons in Cambridge, UK, who were screened using Folstein's Mini-Mental State Examination (MMSE). Those who were judged not demented were followed up 2.4, 6, and 9 years after the baseline assessment. From the available data, only a modest change in MMSE score, from 24.3 to 22.9, was observed after 9 years. However, 35%, 47%, and 54% of the participants who were alive at each follow-up had cumulatively dropped-out at the 2.4-, 6-, and 9-year follow-up visits, respectively. Therefore, only a fraction of those potentially eligible were actually re-examined. To correct for attrition, the authors imputed the mean MMSE score in the participants who had dropped-out at each wave under two different assumptions: that drop-out is a random phenomenon (“missing-at-random”) or, conversely, that it is associated with identifiable characteristics (“informative drop-out”). In the first scenario, the hypothesized rate of change in MMSE score was similar to the one observed, although the absolute values were lower, and the mean MMSE score at year 9 was 21.7. An alternative - and probably more realistic - scenario was depicted by taking into account that the risk of attrition was inversely related to baseline MMSE. This implies that the reported data were likely biased and, therefore, underestimated the true change in MMSE score because persistence in the study was itself associated with a more preserved cognitive performance. Indeed, under the assumption of an “informative drop-out,” the rate of decline in MMSE score was much more rapid, and the mean scores at year 9 ranged from 15.8 to 19.2, according to different hypothesized rates of change. As the authors suggest, this analysis implies that age-associated cognitive decline might be more severe, and/or might affect a larger proportion of the aging population, than previously thought. The consequences for both the clinician and the public health administrator are evident. These data, together with other recent observations, also have relevant methodological implications for aging research. In epidemiology, nonresponse and attrition are rarely, if ever, random phenomena. This seems to be particularly true in studies of cognitive decline in older adults. In a cross-sectional survey on the prevalence of dementia in a Dutch population, Boersma et al. observed that nonresponse was associated with age, gender, marital status, housing condition, and the general practitioner's judgment about the participant's cognitive status.3 Moreover, the patterns of these associations differed among three subgroups of subjects, identified on the basis of the reason for nonparticipation such as refusal, illness or death, or inability to be traced. Estimated prevalence of dementia substantially differed in the three subgroups as well. Therefore, non-responders should be considered and, possibly, characterized to minimize potential bias. Proxy informants, such as next-of-kin or participant's general practitioner, should be contacted whenever possible. The study by Bryane et al.2 demonstrates that estimates of longitudinal changes in cognitive performance could be dramatically affected by assumptions made on randomness of loss to follow-up. Missing-at-random can only dilute the strength of an association, but selective drop-out and non-response may potentially distort the result of an observational study. The effect of selective attrition can be even more dramatic in clinical trials. Previous studies from the Systolic Hypertension in the Elderly Program (SHEP), a randomized clinical trail of diuretic versus placebo for the treatment of isolated systolic hypertension, have shown no long-term difference in cognitive status between active treatment and placebo.4,5 However, we have recently analyzed the data from SHEP to evaluate the presence and the impact of missing data on cognitive outcomes.6 Participants who were alive and eligible for the annual cognitive evaluations during the follow-up, but missed the assessments, were significantly older, less educated, and predominantly non-white race compared with those who were assessed. More importantly, assignment to the placebo group and occurrence of non-fatal cardiovascular events were independent predictors of missed cognitive assessments during the follow-up after adjusting for confounders. Thus, cognitive assessments in SHEP were likely biased toward the null effect, and this might have obscured the appraisal of a protective effect of antihypertensive treatment on cognitive outcomes. To confirm this hypothesis, sensitivity analyses were performed, assuming increasing proportions of subjects with poor cognitive function among those who were eligible but who missed the assessment. In these scenarios, when 20 to 30% of the dropouts were considered cognitively impaired and were added to those who actually screened positive at each annual dementia evaluation, assignment to active treatment reduced the risk of mental deterioration. This suggests that the apparent lack of any beneficial effect of diuretics on the prevention of dementia in the hypertensive participants of SHEP should be seriously questioned. Taken together, the results of these three studies strongly indicate the need for a more accurate conduction of epidemiological aging research. Indeed, nonresponse and dropout are much more likely in older than in younger participants as a result of illness, disability, cognitive impairment, and institutionalization. These conditions are frequent outcomes of either observational or intervention studies, the results of which can, therefore, be altered completely by selective attrition of subjects who had the event of interest. Hence, when epidemiological studies of older adults are being designed, adequate human and financial resources should be allocated up-front to minimize the proportion of nonresponders and dropouts. For example, home and nursing home visits should be considered for those unable to reach the study clinic. As Dorn stated in 1950, “The only method of handling persons lost to follow-up is not to have any.”7 Participation of older, frail individuals in epidemiological studies is increasing, parallel to the expanding proportion of the very old population in our societies. The methodology of aging research must comply with this new challenge.