The Relationships Between Age, Sex, and the Incidence of Dementia and Alzheimer Disease: Title and subTitle BreakA Meta-analysis FREE

AS THE AGE distribution of the United States and world population shifts, the dementing disorders, especially Alzheimer disease (AD), are emerging as a major health problem. The assessment of prevalence and incidence of the disease provides crucial input for public health professionals in determining and allocating health care resources as well as identifying critical risk factors that may be amenable to preventive intervention. Many prevalence studies on dementia and AD have been conducted in various populations.^{1 - 24} Reported prevalence rates from those studies vary considerably and the variations, as indicated by Corrada et al,²⁵ are largely due to methodological differences, namely, clinical diagnostic criteria, sampling strategies, and statistical analysis procedures. Despite the varying magnitude of the reported prevalence rates, all studies show a positive association between age and prevalence rates.²⁶ However, the trend of the association with age has been subject to considerable discussion. One view is that dementia and AD are age dependent, ie, the inevitable consequence of the aging process. This view predicts that if we live long enough, we would all be demented. The other view is that dementia and AD are age related, as in cancer and aging, where the relationship to age is simply an expression of other biological risk factors. This view implies that the disease can be separated from aging and eliminated with the removal of the risk factors. The latter hypothesis is supported by the evidence that prevalence rates seem to level off or even decline in the very-old age groups.²⁷

The relationship between sex and AD has been inconsistent across studies although in many studies women are reported to have higher rates of AD than men even after adjusting for differential survival. Significant differences between sex usually occur, however, in the oldest-age categories where there are few men and even fewer with AD, making estimates unreliable.^{6 ,12 ,14 ,16 ,24 ,26 ,28} The association between sex and AD assumes a greater significance as there is now increasing evidence that estrogen replacement therapy in postmenopausal women improves cognitive function and reduces the risk for both cognitive impairment and AD.^{29 - 31}

Prevalence studies, although valuable, have limitations. The changes in observed prevalence estimates with age cannot really answer the question about age dependency or age relatedness because prevalence is influenced by both survival and disease incidence.³² Incidence, which is the rate of new cases in a population, is considered a better measure of disease risk. Since epidemiological studies on incidence of dementia and AD can be expensive and time consuming, only a limited number of incidence studies have been published. All incidence studies also report a positive association between age and incidence rates of dementia and AD. However, the trend of association with age and the effect of sex are not clear from individual studies because each study is limited in sample size, especially in the very-old age groups. We conducted a meta-analysis to investigate the relationships between age, sex, and the incidence of dementia and AD using published data.

A systematic search through MEDLINE (from 1966 to mid 1997) on "incidence dementia" and "incidence Alzheimer's disease" was conducted. Reference lists from identified articles were also reviewed to locate published studies. To minimize methodological variations of the studies in the meta-analysis, studies were selected using the following inclusion criteria: (1) The study is population based using personal interviews and examinations of the study subjects. (2) The study uses standardized clinical diagnotic criteria; specifically, DSM-III or DSM-III-R³³ for the diagnosis of dementia, National Institutes of Neurological and Communicative Disorders and Stroke–Alzheimer's Disease and Related Disorders Association³⁴ for the diagnosis of AD. (3) Age-specific incidence rates are reported. (4) The study is published in the English language.

Sixty-eight articles were identified as the result of the MEDLINE search, 59 of which were written in English. Twelve studies met the inclusion criteria for the meta-analysis on the incidence of dementia.^{28 ,35 - 45} Ten of these studies also reported sex- and age-specific incidence rates of dementia. Eight studies met the criteria for the meta-analysis on the incidence of AD, with 7 of them also reporting sex-specific incidence rates of AD.^{35 ,37 ,40 - 42 ,46 - 47} Most of excluded studies did not report age-specific incidence rates and some were excluded for not using standard diagnostic criteria.

Annual incidence rates were extracted from all included studies. In cases where only sex-specific incidence rates were reported,^{40 - 41} incidence rates combining men and women were calculated using information and methods specified in the articles.

Two well-known incidence studies are excluded for the following reasons: the Rochester, Minn, study⁴⁸ which used medical records rather than personal interviews and the Liverpool, England,¹³ study which used AGECAT for diagnosis of dementia rather than DSM-III.

Since all studies reported incidence rates by age groups, we recorded incidence rates using the midpoint of each age category. In an open-ended group such as those 85 years and older, we added the distance between the lower age limit to the midpoint for the other age groups in the same study. The studies with age- and sex-specific incidence rates were first analyzed to test for sex effect. Analyses were then conducted using all studies reporting age-specific incidence rates. Analyses were carried out for the incidence of dementia and for the incidence of AD separately.

It is likely that the relationship between incidence of dementia and/or AD and age is not constant over all ages. To test a level-off effect in incidence rates and a differential increase of incidence rates with age, we included a second order of age as a covariate in the analyses. A significant age squared implies that the increase in incidence rates varies with age. It would also indicate the age of a rate decline, if such decline exists. A meaningful level-off effect, we believe, should occur at an age where there is a considerable proportion of the population still at risk.

A mixed-effect model approach was adopted for the meta-analysis. The model was an extension to the random effect models of DerSimonian and Laird,⁴⁹ which allows the studies included in the meta-analysis to be heterogeneous. Details on the mixed-effect models used for the analyses are available from us on request. The modeling was completed in 2 stages. At the first stage, we assumed that the logits of reported incidence rate, ie, log (rate / [1 − rate]), were normally distributed with the logits of true incidence rate as means and variances. At the second stage, we assumed that means from the normal distributions were linearly associated with age, age squared, and sex, with a random study effect. For the analyses on dementia incidence, we assumed that the random study effect increases linearly with age. The underlying assumption for the mixed-effect model is that the effects in the studies included in the meta-analysis can be seen as a random sample of the effects observed in all possible studies meeting the inclusion criteria. The advantages of the mixed-effect model are 2-fold. First, it allows us to model the incidence rates from heterogeneous studies. Second, random measurement errors of the reported incidence rates are taken into consideration in the mixed-effect model.

Likelihood ratio tests were used to test homogeneity of the studies in a mixed-effect model⁵⁰ with age and age² as continuous variables, and sex as a binary variable if applicable. The difference in the −2 log likelihood functions with a random study effect and the −2 log likelihood function without the random effect was known to follow a 50:50 mixture of 2 χ² distributions.^{50 - 51} Tests for the fixed effects (age, age squared, and sex) were performed using F tests.⁵² Variance estimates for the random effect parameters were derived using the restricted maximum likelihood method.⁵³ All analyses were performed using the PROC MIXED procedure found in SAS.^{6 ,12 ,54} Estimated overall annual incidence rates were derived using the approximation to the logistic function for marginal means in a mixed-effect logistic model proposed by Zeger et al⁵⁵ at each medium-age point for various age groups. Confidence intervals for the incidence estimates were obtained using this approximation and the variance-covariance matrices of the fixed-effect parameters from the mixed-effect models.

To demonstrate the effect of age on the estimated incidence rates, we also calculated the incidence rate ratios, defined as the incidence rate of an age group divided by the incidence rate of the age group 5 years younger.

Summary information on the studies included in the meta-analysis is given in Table 1. In addition to reporting the sample sizes and age ranges covered by each study, we also included the following information: (1) whether sampling was used for incidence clinical diagnosis; (2) whether the study included institutionalized subjects; (3) length of follow-up; (4) number of follow-up examinations; and (5) methods of excluding putative demented subjects from baseline, eg, using clinical diagnoses, or below a cut-off point of a screening test. Some of these characteristics are highly correlated. For example, the studies with small sample size tend to clinically diagnose everyone while studies with a large sample size used 2-phase sampling.

Because of the inclusion criteria for using standardized diagnostic criteria, all studies included in the meta-analysis were published after 1991. Reported age-specific incidence rates of dementia from the 12 studies are seen in Figure 1. (The annual incidence rates of dementia stratified by age and sex along with sample sizes are available from us on request.) The youngest age group studied was the 50- to 59-year-olds and several studies reported incidence rates for the 90 and older groups. Reported age-specific incidence rates of AD from the 8 studies are shown in Figure 2. (Annual incidence rates of AD stratified by age and sex along with sample sizes are also available from upon request.)

The first set of analyses was performed using the 10 studies reporting both age- and sex-specific incidence rates of dementia. The test for homogeneity indicates the studies are heterogeneous (c²_10,0=17.14,⁵¹ P =.03). Both age and age squared are significantly related to the prediction of incident dementia (F₉=43.49, P<.001 and F₆₃=26.48, P<.001, respectively). However, female sex is not a significant predictor of incident dementia (F₆₃=2.32, P =.133).

All 12 studies with age-specific incidence rates were used to investigate the relationship between age and the incidence of dementia. Both age and age squared are again significant predictors of incidence of dementia (F₃₄=40.62, P<.001 and F₃₄=25.48, P<.001, respectively). The significance of a quadratic age effect suggests that the increase in incidence rates of dementia varies with the increase in age. The slowing down of the increase in incidence is illustrated in Figure 3, where calculated incidence rate ratio is plotted against age. The incidence rate ratio indicates that for every 5-year increase in age, incidence rates of dementia triple before age 63 and double between ages 64 and 75, and the ratio declines to 1.5 at around age 84. However, for the entire age range covered in the meta-analysis, the rate ratio is always above 1, showing no sign of a rate decline. The phenomenon of a differential increase in incidence rates can also be observed from Table 2, which gives the estimated annual incidence rates and 95% confidence intervals.

The maximum point of predicted incidence rates from this model is at age 98. Theoretically, the model predicts a rate decline after age 98. However, since age 98 is at the upper bound of our data, any prediction beyond this point is not reasonable.

The analysis on the incidence of AD parallels those on the incidence of dementia. The first set of analyses was performed using the 7 studies with age- and sex-specific incidence rates of AD. Age and age squared are shown to be significant predictors of incident AD (F₄₂=15.6, P=.0003 and F₄₂=10.68, P =.002, respectively). Female gender is also significantly related to the prediction of incidence of AD (F₄₂=9.37, P =.004). Specifically, parameter estimates indicate that women are at greater risk of developing incidence of AD with an estimated odds ratio of 1.56 (95% confidence interval, 1.16-2.10).

When all 8 studies with age-specific incidence rate for AD are analyzed, both age and age squared are again significant predictors of incidence of AD (F₂₂=10.07, P=.004 and F₂₂=6.20, P=.021, respectively). The significance of the age squared term again indicates a differential increase in incidence rates of AD with increasing age. It is interesting to observe from Figure 3 that the patterns on the incidence rate changes are almost identical for dementia and AD. Estimated annual incidence rates of AD are presented along with 95% confidence intervals in Table 3.

Results from the meta-analysis indicate that the increase in incidence rates of both dementia and AD slows down with increasing age, although the incidence rates themselves do not decline. For every 5-year increase in age, both dementia and AD incidence rates triple before age 64, double before age 75, and drop down to an increase of 1.5 times around age 85. This slowing down of age-related increase in incidence rates lends support to the hypothesis that both dementia and AD are age related rather than age dependent, with the hopeful corollary that it is possible that preventable risk factors can be identified. Most authorities would already agree that at least one major cause of dementia, cerebrovascular disease, is amenable to preventive strategies.⁵⁶ However, while incidence studies are less influenced by differential mortality than prevalence studies, they do not entirely eliminate the possibility that the oldest subjects with cognitive impairment are more likely to die of comorbid causes before they reach the criteria threshold for dementia, thus influencing rates in the oldest-old category.

Although we do not rule out completely the possibility that the slowing down in the increase of incidence rates is because of small sample sizes for the older age groups, we believe that this is unlikely. If we restricted our analyses only to the large studies (ie, with >1000 subjects), the slowing down effect is still apparent.

Whether or not there is a leveling off of actual rates of incidence with age, as has been suggested by Ritchie and Kildea,²⁷ is not so clear, however. Although the increase in incidence rates slows down with increasing age, there is no evidence of a rate decline at least to age 98 which is the upper limit of our data.

It is important to note that previous studies on dementia and AD have mostly assumed logistic models linear in age when modeling the relationship between age and the disease. Our results demonstrate that a more complicated relationship between age and dementia and/or AD exists.

In contrast to the age effect, the sex effect seemed to be confined to AD with women being at a significantly higher risk for developing an incidence of AD. The differences in the sex effect between dementia and AD are most likely explained by the fact that men are at a higher risk than women for vascular dementia. The reasons for the apparent higher risk for women to develop AD are still uncertain. It has been proposed that there is a sex apolipoprotein genotype interaction, with women who are ϵ3 ϵ4 heterozygotes being at greater risk for AD than men who are ϵ3 ϵ4 heterozygotes.^{57 - 59} The results of the meta-analysis by the APOE and Alzheimer Disease MetaAnalysis Consortium⁶⁰ suggest, however, that women have a higher degree of susceptibility than men regardless of apolipoprotein genotype. It has also been suggested that lack of estrogen or other hormonal changes in postmenopausal women either by themselves or in association with other factors account for the increased risk. Our results are consistent with this hypothesis.^{29 - 31}

The significant female effect on incidence of AD may be due to selection bias in that the studies not reporting sex-specific rates may not have found a significant difference between men and women. However, this is unlikely because the meta-analyses on dementia studies with reported sex-specific rates did not find women at significantly higher risk for incidence of dementia. Furthermore, the difference in sex effect on dementia and AD remains even after we restrict the analysis to only the 7 studies that reported sex-specific rates for both dementia and AD.

It is possible that both our findings, the slowing down of increase in incidence with age and the higher risk of AD for women, are due to survival effects, ie, older subjects are biologically different (harder) than younger subjects. However, this hypothesis can only be tested by longitudinal studies where baseline characteristics may be compared between survivors and the subjects who died.

Meta-analysis was developed as a systematic approach to identify, abstract, and integrate the results of different studies. If performed correctly, a meta-analysis can answer questions that may not be answered from each individual study because of the small sample sizes. However, there are methodological limitations to all meta-analyses. Ideally, a weighted analysis with each reported rate weighted by its precision is preferred.⁶¹ However, since some studies did not report SEs or confidence intervals, the results reported here are based on unweighted analyses. Nevertheless, when we used approximate SEs derived for these studies to perform a weighted analysis, the results changed very little and our conclusions remained the same.

Results from meta-analysis using summarized data can be different from those obtained from using individual subject data.⁶² However, in the studies of dementia and AD, synthesis would be difficult even if data on individual subjects were available because different studies used different sampling and estimation methods. For example, the Cambridge study^{28 ,46} and the East Boston study⁴⁷ both used complex sampling designs at baseline prevalence and at incidence waves. The Cambridge study used weighting from stratified sampling to estimate incidence rates, while the East Boston study derived incidence rates using logistic models. The Framingham study³⁷ had unequal numbers of follow-up times for the study subjects and survival analysis techniques were used for the estimation of incidence rates. Therefore, synthesis of individual subject data would be difficult, if not impossible.

In summary, in this meta-analysis of published incidence studies of both dementia and AD, it would appear that the age-related increase in incidence rates slows down with increasing age, although the incidence rates themselves do not show a decline. Women are at significantly greater risk for developing AD.

Reprints: Hugh C. Hendrie, MB ChB, Professor and Chairman, Department of Psychiatry, Indiana University School of Medicine, 541 Clinical Dr, Room 298, Indianapolis, IN 46202-5111.

Accepted for publication June 11, 1998.

Supported by grants PHS RO1 AGO9956-06 and P30 AG10133-06 from the National Institute on Aging, Rockville, Md.

We would like to acknowledge Taketo Yoshitake, MD, DMSc, Medical Corporation, Department of Japan International Cooperation Agency (JICA) for providing us with the Hisayama Study of Dementia incidence data.