Estimating Likelihood of Dementia in the Absence of Diagnostic Data: A Latent Dementia Index in 10 Genetically Informed Studies

Study samples

All samples used in the analyses are part of the IGEMS consortium [19, 20]. We included 10 studies from four countries (Sweden, Denmark, United States, and Australia), representing all IGEMS studies with both a functional ability measure and a cognitive assessment battery that includes memory. To the extent possible, we wished to know for each participant whether or not the person ever developed probable dementia. Therefore, we selected the most recent assessment wave for each participant with cognitive, memory, and/or functional ability data. No participant was included more than once. The individual studies are presented in Table 1 and described below. The combined sample size across all studies was 15,770. This study was approved by the University of Southern California Institutional Review Board.

Sweden

The four Swedish studies were drawn from the Swedish Twin Register, a population-based register of twins born in Sweden since 1886. Studies include the cross-sectional census: the Study of Dementia in Swedish Twins termed by the Swedish acronym HARMONY for Hälsa (health), ARv (genes), Miljö (environment), Och (and), NY (new) [21], and three longitudinal studies of specific subgroups: the Swedish Adoption/Twin Study of Aging (SATSA) [22], Ageing in Women and Men: A Longitudinal Study of Gender Differences in Health Behavior and Health among Elderly (GENDER) [23], and the Origins of Variance in the Oldest-Old (OCTO-Twin) [24]. Some individuals were part of both HARMONY and one of the three longitudinal studies. Those individuals were placed in the analytic sample that contained their most recent cognitive, memory, and/or functional ability data.

HARMONY entailed screening all living twins in the Swedish Twin Register aged 65+ years [25], then referring for clinical workup all twins who screened positive for cognitive impairment along with their co-twin and a control sample of healthy twin pairs. Individuals were evaluated between 1998 and 2004. The analytic sample included 1,381 individuals.

SATSA is a longitudinal study that began in 1984 and follows same-sex twins who indicated they had been separated before the age of 10 years and reared apart, and a matched sample of twins reared together. All SATSA twins who reached age 50 were invited to participate in in-person testing (IPT). We included only data from the fifth through ninth IPTs (1999–2012) when word list recall was collected. The analytic sample included 548 individuals.

OCTO-Twin includes same-sex twin pairs over the age of ∼80 (79–98 years) at baseline in 1991, with five follow-up waves obtained through 2002. The analytic sample included 561 individuals.

GENDER is a longitudinal study of opposite-sex twin pairs born between 1916 and 1925 who were aged 70–79 years at the first IPT and had three IPT waves between 1995 and 2005. The analytic sample included 479 individuals.

Denmark

The two Danish twin studies were drawn from the Danish Twin Registry and include the Longitudinal Study of Aging Danish Twins (LSADT) [26] and the Middle-Aged Danish Twins Study (MADT) [27]. LSADT is a cohort-sequential study of same sex twin pairs that began in 1995. Initially, LSADT included twins aged 75+ years, but as new cohorts were recruited the age minimum was lowered to 70, resulting in age ranges from 70 to 96 years. MADT is a longitudinal study of both same- and opposite-sex twin pairs born between 1931 and 1952 who were first assessed in 1998, resulting in age ranges from 45 to 68 years. The analytic sample included 4,626 in LSADT and 4,306 in MADT.

Australia

The Older Australian Twins Study (OATS) [28] started in 2006 with the recruitment of twins aged 65 and older, including twins from the Australian Twin Registry as well as new volunteers from the community. In-person evaluations included core components for diagnosing dementia, including medical and neuropsychological examinations, informant interview, blood collection, and brain imaging [28]. The analytic sample included 592 individuals.

United States

The three US studies include the Carolina African-American Twin Study of Aging (CAATSA) [29], the twin sample of the Midlife Development in the United States (MIDUS) [30], and the Vietnam Era Twin Study of Aging (VETSA) [31].

CAATSA is a cross-sectional sample of African American twins from North Carolina recruited by phone book/internet/credit service searches and voter registration records in 1999 through 2003 [29, 32]. The analytic sample included 675 individual twins ranging in age from 20 s through 80 s.

MIDUS is a nationally representative longitudinal study, which also collected a twin subsample of both same- and opposite-sex twin pairs ascertained using a random-digit dialing procedure and supplemented with additional twin pairs recruited through referrals given by non-twin participants [33]. Cognitive function and self-rated functional ability were assessed in two waves in 2004 to 2006 and in 2013 to 2014. The analytic sample included 1,341 individuals, ages ranging from 34 to 82 years.

VETSA is a longitudinal study of a national sample of male twins who served in the military during the Vietnam era (1965–1975), recruited through the Vietnam Era Twin Registry. All twins were 51 to 61 years of age at initial assessment wave and were 55 to 67 years of age at the second assessment wave that was used in the analyses. The analytic sample included 1,261 individuals.

Variables

The 10 IGEMS studies differ considerably in the number and type of memory, other cognitive, and functional ability measures available. Memory measures represent various episodic memory tasks. Other cognitive measures encompass domains of verbal ability, visuospatial ability, analytical reasoning, verbal fluency, processing speed, and working memory, with each study typically assessing only some domains. Functional ability measures encompass everyday activities that require cognition in order to function independently. In our analyses, we included measures of memory and other cognitive domains that were available in each participating study (as shown in Table 2 and described in detail in the Supplementary Material). Consistent with prior work [34], we converted raw memory and other cognitive scores into the proportion of maximum possible, a variation of percentage of maximum possible (POMP) values [35]. All scores were coded so that higher values indicated better performance. When there was a total correct score, the POMP score was the raw score divided by maximum possible (i.e., proportion correct). When raw score values were number of items generated during a timed task (e.g., verbal fluency), the POMP score was the number of items generated divided by the time allotted. When raw score values were times to complete a task with lower scores indicative of better performance (e.g., trails), we calculated the POMP score by subtracting time to complete from maximum time then dividing by maximum time. The proportions were then multiplied by 10 so that all scores ranged from 0 to 10; the transformation does not affect the covariance matrices analyzed [35].

Five studies (n = 3,490) from Sweden and Australia administered a complete diagnostic protocol and applied clinical diagnostic criteria to establish clinical diagnoses of dementia, including type of dementia, also whether the person met criteria for mild cognitive impairment (MCI) or cognitively impaired/not dementia (CIND). MCI and CIND were considered not-dementia in our dichotomous clinically diagnosed dementia variable, which was coded 1 = dementia (of any type) and 0 = not-dementia.

Swedish and Danish twin registries were both linked to population-based health registries that contained diagnostic information. There is a lag between the time a clinical diagnosis is made at a clinic and how soon the registry records the diagnostic code. Based on data from the three Swedish Twin Register longitudinal studies, it was estimated that a dementia ICD code appears in the National Patient Registry (NPR) on average 5.5 years after the diagnosis is first made [36, 37]. Thus, in our analyses of agreement between registry and LDI, we determined whether the participant had an ICD code for dementia in any registry within 5 years after the assessment wave used for scoring the LDI.

Data analysis

LDI scores

Raw data from each study sample were used in one-factor confirmatory factor analysis (CFA) models to estimate LDI in each of the 10 studies. CFA tests the hypothesis that the observed variables fit the latent structure of the model. Figure 1 illustrates the one-factor CFA model fit to each study sample. This model allows for existing variability in the number and type of memory, other cognitive, and functional ability measures administered across studies. For example, not every study included word list learning and recall among their memory measures. For studies that did, the number of words could vary from 10 to 16, the number of learning trials from 1 to 5, the delay before recall from 5 to 30 min.

Fig. 1

General Latent Dementia Index (LDI) Model (Variables Taken From HARMONY). LDI, latent dementia index; manifest variables include: COG₁ – COG_c, non-memory cognitive ability variables (e.g., block design, verbal fluency, digit span); MEM₁ – MEM_d, memory ability measures (e.g., immediate recall, delayed recall, and recognition); and functional ability, which is a composite measure of instrumental activities of daily living that require some level of cognitive functioning. Manifest variable intercepts and latent means are indicated by triangles with a value of 1 inserted in them. Residual variances (ɛ) are estimated for each manifest variable. * indicates that the parameter is estimated whereas values along paths (e.g., 0) indicate that the parameter is fixed to that value.

Consistent with prior approaches [9, 11, 13], cognitive ability was modeled as part of the LDI but with correlated residuals, which in effect splits the cognitive ability variance into a residual general cognitive ability component that represents baseline cognitive ability and a latent dementia indicator component that represents those cognitive abilities and functional areas affected by dementia. Both functional ability and memory were considered unique to the LDI. The separability of memory from general cognitive ability is already well-supported [38, 39] and matches the outcomes of prior δ studies [9, 13] that showed factor loadings of the memory measures on the general cognitive ability factor became non-significant after adding δ to the models.

The model in Fig. 1 saturates the correlations among the residuals of the non-memory cognitive ability variables, thereby statistically addressing differences among studies in the number of tests that assess the same domain. We specified one memory variable factor loading constrained to 1.0 to scale the LDI and allow the LDI variance to be estimated. The intercept of the same memory variable was constrained to 0 so that the mean of the LDI could be estimated. This scaling of means and variances has the effect of scaling the LDI scores by similar tasks across study so that studies can be coordinated meaningfully.

Analyses were conducted in Mplus 8.4 [40]. Maximum likelihood with robust standard errors was used as the estimator to account for violations of multivariate normality and missing data. Non-response was assumed to be missing at random or missing not at random. In both cases, full information maximum likelihood is the recommended approach for obtaining least biased parameter estimates and standard errors [41]. To obtain unbiased standard errors, we accounted for the dependence between twins in the same family. Factor scores were scaled such that lower scores indicate greater likelihood of dementia. Finally, age at testing was included in the model as a predictor of the cognitive, memory, and functional ability variables to center obtained scores at age 60 [42].

To pool all studies for coordinated analyses of the entire consortium data set, study sample LDI scores were re-centered to align with HARMONY LDI scores. The aligned individual scores were constructed by adding the difference between each individual’s sample mean and HARMONY LDI mean to each individual score (e.g., LDIAligned=LDIOriginal+(LDI-Original-LDI-Harmony) ). HARMONY is the study with the largest sample size and clinical diagnoses of participants.

LDI validity

Criterion validity was evaluated by estimating correlations of continuous LDI scores with dichotomously coded clinically diagnosed dementia. Subsequently, cut-off points were derived for the LDI (1 = likely to have dementia; 0 = unlikely to have dementia) using dementia prevalence within the study when available or otherwise based on prevalence in the population from which the study was drawn (e.g., national rate of dementia in Denmark was 9% in those 65+ years old [43]; in the US, 10% in those 65+ years-old [44, 45]. Using these cut-off points on the LDI, we calculated sensitivity (true positive rate), specificity (true negative rate), positive predictive value (proportion of people below cut-off who have the disease), false positive rate (those incorrectly identified as having the disease), negative predictive value (proportion of people above cut-off who do not have the disease), and accuracy (sum of true positives and true negatives, or those categorized correctly) compared with dichotomous clinical diagnosis.

Separately, we used the LDI as the reference standard and register diagnosis as the test to examine the specificity of the LDI. Given the Scandinavian national registries have shown approximately 50% sensitivity and nearly 100% positive predictive validity (i.e., there are minimal false positive cases [36, 37, 46]), we expected the LDI would show high specificity and capture nearly everyone who had a dementia code in a national registry.

Genetically informed analyses

We conducted univariate and bivariate quantitative genetic analyses with univariate twin models fit to the LDI scores (Fig. 2a). In classical twin (“ACE”) models, additive genetic (A) components are the cumulative additive effect of genotype and are estimated by virtue of the fact that MZ twins share 100% of their genes and DZ twins share 50% of their genes, on average. Shared environmental (C) components are the cumulative effect of any environment that makes twins reared in the same family more similar to one another and estimated under the assumption that shared environmental influences affect twins similarly regardless of genetic relatedness. Non-shared environmental (E) components are any environmental factor that makes twins different from one another, including measurement error. Broad heritability estimates of LDIs were estimated to infer whether additive genetic influence was similar across studies using their respective 95% confidence intervals.

Fig. 2

Univariate Twin Model (a) and Bivariate Twin Model (b) for One Twin Only. The univariate model in (a) gives the example of decomposing the observed latent dementia index (LDI) variance into random latent genetic (A), shared environmental (C), and non-shared environmental (E) variance components. In this model, LDI intercepts are estimated and residual variances of the LDI are constrained to be zero. Univariate ACE models were fit to dichotomized clinically diagnosed dementia and dichotomized LDI scores. In these models, thresholds replace intercepts. Effects of age at cognitive assessment (presented) and age² (not presented) were estimated. The subscripts 1 and 2 refer to twin 1 and twin 2, respectively. The bivariate model (b) estimates decomposes the LDI and clinically diagnosed dementia into random latent genetic (A), shared environmental (C), and non-shared environmental (E) variance components and estimates the correlations between the respective components (Cov(A), Cov(C), and Cov(E)). Effects of age at cognitive assessment (presented) and age² (not presented) were estimated.

Bivariate ACE models were fit to the continuous LDI scores and dichotomous clinical diagnoses (Fig. 2b, with observed and latent variables presented for one twin only for simplicity of presentation). Genetic and environmental covariances were estimated to construct the genetic and environmental correlations between LDI and clinically diagnosed dementia as product-moment correlations and reported with 95% confidence intervals. Intercepts and thresholds were constrained across twins and zygosity groups. Age and age-squared were included in all models to reduce bias in the shared environmental variance given that the twins are tested at nearly identical ages (r = 0.95). Proportions of variability in LDI and clinically diagnosed dementia attributed to genetic (A, or h²) and environmental (C, or c², and E, or e²) components were estimated by dividing these variance components by the total variance estimate.

Univariate ACE models were estimated with maximum likelihood, and bivariate twin models were estimated using weighted least squares because of including the dichotomous clinical diagnosis.

LDI scores

Table 2 shows the standardized factor loadings from the LDI model for the 10 studies (Supplementary Table 1 presents unstandardized results). Loadings were generally in the same range across all studies (0.20s to 0.80s, standardized), regardless of which cognitive tests were included. Loadings for the memory tests generally were more strongly associated with the LDI (0.30s to 0.90s, standardized) than the non-memory cognitive items. LDI factor means were lowest (i.e., greatest likelihood of dementia) in studies with highest proportion of individuals diagnosed with dementia (e.g., HARMONY, OCTO-Twin) and highest (i.e., least likelihood of dementia) in studies with the smallest proportion of individuals diagnosed with dementia (e.g., OATS, VETSA), with rates of dementia mirroring age distribution of the respective studies. As seen in Supplementary Table 1, variances for the LDI score were lower in US and Danish studies, suggesting relatively greater precision of the LDI mean estimate.

Figure 3 displays the LDI factor score histograms (Fig. 3a) and density functions (Fig. 3b) for all studies. The vertical black line in the plot represents the LDI cut-off point for categorizing individuals as likely to have dementia with values to the left of the black line categorized as “likely” and values to the right of the black line categorized as “unlikely”. This histogram provides support for a continuous LDI that distinguishes between individuals who have higher likelihood of dementia from those who have lower likelihood of dementia. The overlapping density curves depict the shape of each study’s LDI distribution, which is normally distributed. Although the distribution is symmetric and there is no noticeable skewness, the distribution is platykurtic with a value less than 3.00 (2.09), which suggests that there are fewer extreme values in our LDI distribution than a Gaussian distribution.

Fig. 3

Histogram (left panel) and Density Curves (right panel) of the 10 IGEMS studies (HARMONY units). Panel a is histogram of the probability densities for all 10 studies. Panel b is the Gaussian kernel density plots of all 10 studies. Units are converted to those estimated in the HARMONY. The solid black vertical line indicates the cut-off score (5.26) used to categorize twins with probable dementia (left of the line) and twins without probable dementia (right of the line).

LDI scores consistently correlated with clinically diagnosed dementia across studies (see lower portion of Table 2). All values are moderate to high, ranging from –0.41 in OATS to –0.75 in HARMONY. Correlations between LDI and mental status screening scores (e.g., Mini-Mental State Examination [MMSE] or Telephone Interview for Cognitive Status [TICS]) also ranged from a moderate value of 0.48 in CAATSA to a large value of 0.80 in OCTO-Twin.

Table 3 shows the criterion validity analyses for the studies that have clinical diagnoses available. Consistent with our selection of cut-off points that prioritize keeping false positives at a minimum, false positives were low (range: 1.7–7.0%) and specificity was high (90.5–98.2%) in all studies. Sensitivity varied across the five studies from HARMONY at 87% to SATSA at 51%. These sensitivity levels are highest in studies with higher disease prevalence. Separately, sensitivities for registry dementia codes within 5 years of neuropsychological testing were 88%, 93%, 93%, 89% for HARMONY, SATSA, GENDER, and OCTO-Twin, respectively.

Genetically informed analyses

Table 4 presents univariate ACE estimates and genetic and environmental correlations from bivariate ACE models for clinically diagnosed dementia where available, quantitative LDI, and dichotomized LDI. Pooled heritability estimates (h²) of the quantitative LDI scores (first set of columns in Table 4), dichotomous LDI scores (second set of columns in Table 4), and the dichotomously coded clinical diagnosis variable (third set of columns in Table 4) are comparable, showing h² = 53%, 59%, and 61%, respectively. Shared environmental variance (c²) components underlying LDI were negative across all studies and so were set to zero in the models. Non-shared environmental factors (e²) accounted for the remaining variance, with estimates of 47%, 41%, and 39% respectively for quantitative LDI, dichotomous LDI, and clinically diagnosed dementia. Results for individual studies were far more variable than the pooled results. These results suggest that LDI modestly underestimates h² and modestly overestimates e².

Moderate to strong genetic correlations were observed between the quantitative LDI scores and clinical diagnoses across all studies, with r_A = –0.66 for the combined sample (range: r_A = –0.81 to –0.48 for individual studies; see last set of columns in Table 4), suggesting genetic sources of variability in common to LDI and clinical diagnoses. Non-shared environmental correlations were small to moderate and significant, with r_E = –0.51 for the combined sample (range: r_E = –0.64 to –0.30 for individual studies). Twins who had higher LDI scores compared to their co-twins were, in fact, more likely to be diagnosed with dementia.

Conclusion

We extended prior research on latent variable approaches to estimating dementia likelihood in single studies to multiple studies. Developing a continuous measure of dementia likelihood that can be pooled across samples provides greater power than a dichotomous measure and increases the likely generalizability of findings. The consistency of our results across diverse studies (where dementia assessment may or may not have ever been intended) suggests this latent dementia indicator approach can be applied to any study of cognitive aging with everyday cognitive function measures. Most large surveys of aging include at least one measure of memory, one non-memory measure of cognition, and a measure of instrumental activities of daily living, which are the elements used for deriving the LDI scores. The implication is that many studies could adopt this method to coordinate estimation of likelihood of dementia diagnosis. Furthermore, the simplicity of the single factor LDI model specified in this study means that it can be applied easily across additional studies within research consortia.