Skip to main content
NCBI home page
The Journals of Gerontology Series B: Psychological Sciences and Social Sciences logoLink to The Journals of Gerontology Series B: Psychological Sciences and Social Sciences
. 2016 Mar 11;72(6):947–955. doi: 10.1093/geronb/gbw004

Exportation and Validation of Latent Constructs for Dementia Case Finding in a Mexican American Population–based Cohort

Donald R Royall 1,2,3,1, Raymond F Palmer 3, Kyriakos S Markides 4
PMCID: PMC5927021  PMID: 26968639

Abstract

Background

The latent variable “δ” has been validated as a dementia phenotype. δ can be extracted from Spearman’s general intelligence factor “g” in any data set that contains measures of cognition and instrumental activities of daily living (IADL). We used δ composites (“d-scores”) to estimate the prevalence of dementia in the Hispanic Established Population for Epidemiological Studies in the Elderly (H-EPESE).

Method

δ was constructed from Mini-Mental State Examination, a clock-drawing task (CLOX), and IADL. δ’s H-EPESE factor weights were validated in the well-characterized Texas Alzheimer’s Research and Care Consortium (TARCC). Optimal thresholds for the discrimination between “Alzheimer’s disease” (AD) versus normal controls (NCs) were determined by receiver operating characteristic curve. Those thresholds were used to estimate the prevalence of dementia in H-EPESE.

Results

Each δ homolog fits its source’s data well. d-scores were strongly associated with Clinical Dementia Rating scale Sum of Boxes (r = .74–.85, all p < .001], and accurately distinguished AD cases from NCs, in both Mexican Americans (MAs) and non-Hispanic Whites (NHWs) [c = 0.94–0.96]. The TARCC MA threshold estimated the prevalence of dementia at 21.4% in H-EPESE. The NHW threshold estimated the prevalence of dementia at 21.0%.

Conclusions

It is possible to export δ composites from populations to well-characterized cohorts for validation.

Keywords: Aging, Cognition, Dementia, Functional status, g

The prevalence of dementia in minority populations is difficult to ascertain. Population-based cohorts are often constrained with regard to the psychometrics they can obtain, and those are further limited by measurement error, including educational, cultural, and linguistic bias. Better-characterized cohorts are usually convenience samples of uncertain generalizability.

The valid diagnosis of dementia and dementia severity in population-based cohorts would represent a considerable advance because it would unlock valuable information on dementia’s risk factors and longitudinal progression. Many epidemiological samples are carefully selected to preserve their generalizability to entire populations. They often have longitudinal information extending, in some cases, back to midlife characteristics, and may include information on social and lifestyle characteristics, dietary habits, health care access, and so on, which are often neglected in “better-characterized” convenience samples.

The prevalence of dementia in Mexican Americans (MAs) is particularly pressing, as this is a rapidly growing demographic in the United States. Not only are MAs likely to dominate the elderly populations of several U.S. states in coming years, but MAs as a whole exhibit unexpected features with regard to their presentations. They may be at higher risk for dementia than non-Hispanic Whites (NHWs) and present at more advanced stages, or with distinct neuropsychiatric profiles (Alzheimer’s Association, 2010; Espino et al., 2002; Salazar, Palmer, & Royall, 2015). Furthermore, even the biomarkers of dementia may differ between MAs and NHWs (Farrer et al., 1997; O’Bryant et al., 2013a, 2013b; Royall & Palmer, 2015).

We have reported a novel approach to dementia case finding that could help establish the prevalence of dementia in population-based samples, even across cultural and linguistic barriers (Royall et al., 2016a). Our approach employs confirmatory factor analysis in a structural equation modeling (SEM) framework to construct a novel bifactor model which results in a latent proxy for dementia severity, that is, “δ” (for “dementia”) (Royall, Palmer, & O’Bryant, 2012).

The latent variable δ represents the “cognitive correlates of functional status.” Its unique bifactor construction explicitly parses Spearman’s general intelligence factor “g” (Spearman & Wynn Jones, 1951) into two orthogonal fractions, that is, one that is related to a target indicator and one that is not. The latter latent variable is labeled “g′” to distinguish it from g itself. If the target indicator is a measure of instrumental activities of daily living (IADL), then the former latent variable is δ.

As a lucky happenstance, δ’s derivation from g also ensures that it can be extracted from almost any psychometric battery, provided that it also includes a measure of IADL. We have explored this claim down to item-level data (Royall et al., 2016b). Thus, it appears that δ might be constructed post hoc from almost any existing data set, or prospectively from variables selected for other agendas (e.g., brevity, cost, low administrative burden, or availability in translation). Because there are so many possible batteries from which to construct δ, we refer to each specific embodiment as a δ “homolog.” Latent variables extracted from the same bifactor model, but directed at other targets (e.g., depression), are referred to as δ “orthologs.”

We have validated δ homologs in multiple data sets, including well-characterized participants participating in the Texas Alzheimer’s Research and Care Consortium (TARCC) study, which includes MA participants (Royall & Palmer, 2013; Royall, Palmer, & O’Bryant, 2012). δ (i) is uniquely related to dementia severity as measured by the Clinical Dementia Rating scale Sum of Boxes (CDR-SOB), (ii) accurately distinguishes cases with Alzheimer’s disease (AD), and mild cognitive impairment (MCI), from each other and from controls (Royall et al., 2012), (iii) is specifically associated with the default mode network by voxel-based morphometry (Royall et al., 2012, 2013), (iv) accurately diagnoses dementia regardless of etiology (Gavett et al., 2015), and (v) is superior to conventional psychometrics with regard to the prediction of prospective MCI to dementia conversion, prediction of hippocampal atrophy, prediction of apolipoprotein E e4 allele burden, and prediction of an AD-specific cerebrospinal fluid biomarker profile (Koppara et al., 2016).

As a latent variable, δ is inherently resistant to measurement error. Moreover, because δ is related to g, it appears to share g’s robust indifference to the cognitive measures used to construct it (i.e., its indicators). Thus, δ homologs can be constructed from almost any ad hoc combination of cognitive and functional status measures. We have successfully engineered δ homologs from large batteries of formal measures, brief batteries of “bedside” screeners, and even the item sets of individual scales (Koppara et al., 2016; Royall, Palmer, & O’Bryant, 2012, Royall et al., 2015, 2016b).

We recently validated an MA-specific δ homolog (dMA) in TARCC (Royall & Palmer, 2013). dMA was indicated by two simple screening tests, the Mini-Mental State Examination (MMSE) (Folstein, Folstein & McHugh, 1975) and the CLOX (an executive clock-drawing task [CDT]) (Royall, Cordes, & Polk, 1998). These measures were chosen because they are also available in the Hispanic Established Population for Epidemiological Studies in the Elderly (H-EPESE). H-EPESE is the largest population-based cohort of MA adults ever assembled, and yet its psychometrics are limited to those two measures, administered by lay interviewers (Royall, Espino, Polk, Palmer, & Markides, 2004). In this analysis, we reconstruct dMA homologs in both TARCC and H-EPESE. TARCC is a well-characterized clinic-based convenience sample. H-EPESE is a less well-characterized population-based cohort.

We propose to export dMA homolog factor weights from H-EPESE to TARCC and to validate them separately among MA and NHW TARCC participants, and to compare their diagnostic accuracies to dMA homologs constructed locally in those TARCC subsets. From those analyses, it should be possible to define optimal thresholds for the discrimination between AD and normal controls (NCs) in both ethnicities, and then use those to diagnose the presence or absence of dementia in the H-EPESE participants. If concordance is achieved, we will have demonstrated δ’s potential to be exported from a population into a well-characterized sample for validation and to use those thresholds to estimate the prevalence of dementia in the original population sample.

Method

Participants

Texas Alzheimer’s Research Consortium

These represent Visit 1 data from the TARCC cohort (circa 2008–2014). The cohort had N = 2,016 TARCC participants (920 cases of AD, 277 cases of MCI, and 819 controls). The methodology of the TARCC project has been described in detail elsewhere (Waring et al., 2008). Each participant underwent a standardized annual examination at their respective evaluation site that includes a medical evaluation, neuropsychological testing, and clinical interview. Diagnosis of AD status was based on National Institute for Neurological Communicative Disorders and Stroke—Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria (McKhann et al., 1984). Controls performed within normal limits on their psychometric assessments. Institutional Review Board approval was obtained at each site, and written informed consent was obtained for all participants.

Hispanic Established Population for Epidemiological Studies in the Elderly

The H-EPESE is a longitudinal study of a representative sample of MAs, aged 65 years, residing in the Southwestern United States. The sample was drawn using area probability sampling procedures to represent MA older adults in five Southwestern states, Texas, New Mexico, Colorado, Arizona, and California. As of the 2010 census, more than 85% of all elderly MA resided in these states (Ennis, Ríos-Vargas, & Albert, 2011).

Participants were interviewed in their own homes. The response rate for eligible respondents was 83%. The total number of participants surveyed at baseline (Fall, 1993–Spring 1994) was 3,050. Of these, 1,713 (56.2%) were still available at Wave 3 (Fall–Spring 1998–1999), when the CLOX was introduced. Wave 3 data were used in this analysis. A valid CLOX (either CLOX1 or CLOX2) was obtained in 1,202 participants (70.2%). Of these, 1,165 had a valid CLOX1 and 1,202 had a valid CLOX2.

TARCC Clinical Variables

Depressive symptoms were assessed using the 30-item Geriatric Depression Scale (GDS) (Maixner et al., 1995; Yesavage et al., 1982). GDS scores range from 0 to 30. Higher scores represent worse conditions. A cutpoint of 9–10 best discriminates clinically depressed from nondepressed elderly adults.

The CDR-SB

The CDR (Hughes, Berg, Danziger, Coben, & Martin, 1982) was used to evaluate dementia severity. This rating assesses the patient’s cognitive ability to function in six domains—memory, orientation, judgment and problem solving, community affairs, home and hobbies, and personal care. Information is collected during an interview with the patient’s caregiver.

δ’s Cognitive Indicators (both data sets)

The MMSE (Folstein, Folstein, & McHugh, 1975) is a well-known and widely used test for screening cognitive impairment. Scores range from 0 to 30. Scores less than 24 reflect cognitive impairment. The MMSE has significant educational and cultural biases. A cutpoint of 21/30 is recommended in elderly MA (Raji, Al Snih Ray, Patel, & Markides, 2004).

CLOX: An Executive CDT

The CLOX (Royall, Cordes, & Polk, 1998) is a brief executive control function (ECF) measure based on a CDT and is divided into two parts. CLOX1 is an unprompted task that is sensitive to executive control. CLOX2 is a copied version that is less dependent on executive skills. CLOX1 is more “executive” than other comparable CDTs (Royall et al., 1999). Each CLOX subtest is scored on a 15-point scale. Lower CLOX scores are impaired. Cutpoints of 10/15 (CLOX1) and 12/15 (CLOX2) represent the 5th percentiles for young adult controls.

δ’s Target Indicator (both data sets)

Each δ homolog used a measure of caregiver-rated IADL as its target indicator. Lawton and Brody’s (1969) assessment was used in both H-EPESE and TARCC. The ability to use the telephone, shopping, food preparation, housekeeping, laundry, use of transportation and ability to handle finances were rated on Likert scales ranging from 0 (no impairment) to 3 (specific incapacity). Total IADL scores were summed across items.

Statistical Analyses

The sequence of analyses is presented in Figure 1. A dMA homolog was constructed in a SEM framework, in a combined sample, using Analysis of Moment Structures software (Arbuckle, 2006). The latent dMA variable was indicated by MMSE, CLOX1, CLOX2, and IADL and was adjusted for age, gender, and education. The model’s factor invariance across the two cohorts was tested by comparing the χ2 difference between a restricted model (separately holding each δ factor weight to be equal across cohorts) and an unrestricted model (allowing the loadings to be freely estimated between cohorts). A nonsignificant χ2 test indicates invariance (Bollen & Long, 1993).

Figure 1.

Figure 1.

Open in a new tab

Analysis design. H-EPESE = Hispanic Established Population for Epidemiological Studies in the Elderly; MA = Mexican American; NHW = non-Hispanic White; TARCC = Texas Alzheimer’s Research and Care Consortium. The numbers refer to analysis paths in Tables 2 and 3.

Identical cohort invariant dMA bifactor structures (Figure 2) were then applied to the H-EPESE and TARCC data separately, and their fit tested in the respective samples. In TARCC, the model was stratified on ethnicity (MA n = 1,082; NHW n = 1,098). Thus, we obtained three unique dMA homologs with identical indicators but with freely estimated sample-specific factor loadings (i.e., for H-EPESE, TARCC MA, and TARCC NHW). The homolog parameter estimates are provided in Supplementary Material.

Figure 2.

Figure 2.

Open in a new tab

Mexican American—specific δ homolog (dMA) or generic model used to construct dMA in all three samples. To ease comparisons across samples, dMA is scaled to MMSE, and all psychometric indicators are standardized to the distribution of the sample within which the homolog is constructed. CLOX1 = unprompted clock drawing from CLOX: an executive clock-drawing task; CLOX2 = copied clock drawing; IADL = instrumental activities of daily living; MMSE = Mini-Mental State Examination.

In the well-characterized TARCC participants, the two ethnicity-specific dMA homologs were validated by testing their receiver operating curve characteristic curve (ROC curve)or area under the ROC curve (AUC) for the discrimination between AD versus NCs (Paths 1 and 2 in Figure 1) and by their correlations with CDR-SOB. The association with CDR-SOB was tested both for each raw latent construct and its extracted composite.

Having confirmed the validity of each ethnicity-specific δ homologs in their own well-characterized samples, each homolog’s factor weights were “exported” to observed data in the other ethnic subgroup (Paths 3 and 4 in Figure 1). The validity of the exported homologs was again tested by ROC curve or AUC and by their associations with CDR-SOB.

Next, the H-EPESE dMA homolog was similarly exported to TARCC MA and NHW subsamples. The validity of its factor weights in TARCC data was tested by ROC curve or AUC and by association with CDR-SOB (Paths 5 and 6 in Figure 1).

Next, optimal dMA thresholds for the discrimination between AD and NCs were determined separately by ROC curve in TARCC, in each ethnicity-specific subset. These thresholds were then applied to H-EPESE data (Paths 7 and 8, Figure 1). This resulted in two independent estimates of AD’s prevalence in the H-EPESE cohort: H-EPESE to H-EPESE via TARCC MA and H-EPESE to H-EPESE via TARCC NHW.

Missing data

Only the ROC curves were limited to complete cases. Elsewhere, we used full-information maximum likelihood methods to address missing data (Graham, 2009; Schafer & Graham, 2002).

Fit indices

The validity of structural models was assessed using two common test statistics. A nonsignificant χ2 signifies that the data are consistent with the model (Bollen & Long, 1993). The comparative fit index (CFI), with values ranging between 0 and 1, compares the specified model with a model of no change (Bentler, 1990). CFI values below 0.95 suggest model misspecification. Values of 0.95 or greater indicate adequate-to-excellent fit. A root mean square error of approximation (RMSEA) of 0.05 or less indicates a close fit to the data, with models below 0.05 considered “good” fit, and up to 0.08 as “acceptable” (Browne & Cudeck 1993). All three fit statistics should be simultaneously considered to assess the adequacy of the models to the data.

ROC curves

The diagnostic performance or accuracy of a test to discriminate diseased from normal cases can be evaluated using ROC curve analysis (Metz, 1978; Zweig & Campbell, 1993). Briefly, the true-positive rate (sensitivity) is plotted as a function of the false-positive rate (1.00-specificity) for different cutoff points of a parameter. Each point on the ROC curve represents a sensitivity/specificity pairing corresponding to a particular decision threshold. The AUC is a measure of how well a parameter can distinguish between two diagnostic groups (diseased or normal).

Statistical differences in AUCs were tested by DeLong’s method (DeLong, DeLong, & Clarke-Pearson, 1988) using MedCalc Statistical Software version 14.12 (MedCalc Software bvba, Ostend Belgium www.medcalc.org; 2014).

Results

Descriptive statistics are presented in Table 1. The H-EPESE sample was significantly older and less well educated than both the TARCC MA and NHW subsets. 80% of H-EPESE participants were interviewed in Spanish compared with 20% of TARCC’s.

Table 1.

Descriptive Statistics of Three Data Sets

N Minimum Maximum Mean SD
H-EPESE
 Gender (% female) 3,050 — — 42% —
 Education 3,002 0 20 4.85 3.90
 Age 1,980 70 104 77.40 6.07
 MMSE 1,876 0 30 21.05 7.29
 IADL 1,970 0 10 2.61 3.46
 CLOX1 1,108 0 15 8.37 3.25
 CLOX2 1,192 1 15 12.26 2.40
 Valid N (listwise) 1,064
TARCC MA
 Gender (% female) 1,082 — — 68% —
 Education 1,082 0 23 10.01 4.72
 Age 1,082 50 94 66.41 9.24
 MMSE 1,082 5 30 26.23 4.08
 IADL 1,068 3 31 9.15 3.86
 CLOX1 1,012 0 15 11.18 2.66
 CLOX2 1,010 0 15 13.27 1.72
 Valid N (listwise) 1,010
TARCC NHW
 Gender (% female) 1,898 — — 58% —
 Education 1,898 3 23 14.96 2.799
 Age 1,898 50 102 73.50 8.991
 MMSE 1,897 2.00 30.00 24.88 5.221
 IADL 1,396 1.00 5.00 1.31 0.710
 CLOX1 878 0.00 15.00 11.69 2.715
 CLOX2 871 2.00 15.00 13.34 1.819
 Valid N (listwise) 869
Open in a new tab

Notes. H-EPESE = Hispanic Established Population for Epidemiological Studies in the Elderly; IADL = instrumental activities of daily living; MA = Mexican American; MMSE = Mini-Mental State Examination; NHW = non-Hispanic White; TARCC = Texas Alzheimer’s Research and Care Consortium.

The bifactor dMA model had excellent fit regardless of the sample in which it was constructed (H-EPESE: χ2 = 16.0 (4), p < .01, CFI = 0.98, RMSEA = 0.03; TARCC MA: χ2 = 2.1 (4), p = .55, CFI = 0.99, RMSEA = 0.01; NHW: χ2 = 1.4 (4), p = .85, CFI = 0.99, RMSEA = 0.01). Model parameter estimates are provided in Supplementary Tables. Factorial invariance was achieved across samples. The unrestricted versus the restricted model for; CLOX1 135.7 (DF = 14) vs 139.1 (DF = 15); CLOX2 139.1 (DF = 15) vs 141.3 (DF = 16); MMSE 141.3 (df = 16) vs 143.3 (DF = 17); IADL 141.3 (DF = 16) vs 143.5 (DF = 17). Each χ2 difference was less than 3.84 and therefore not significantly different with 1 DF.

In the well-characterized samples, each dMA homolog was strongly associated with CDR-SOB, although their composites were significantly less strongly associated with CDR-SOB than were the corresponding latent variables. Both dMA homologs also achieved high AUCs for AD versus NCs (TARCC MA: AUC = 0.94; TARCC NHW: AUC = .95; Table 2). Each ethnicity-specific composite performed well in the other’s ethnic subgroup (Table 2).

Table 2.

Receiver Operating Curve Results

Path (Figure 1) Source of δ weights (sample) AUCa r for composite d with CDR-SOB* r for latent δ with CDR-SOB*
1 TARCC MA (TARCC MA) .942 −.83 −.88
2 TARCC NHW (TARCC NHW) .945 −.85 −.98
3 TARCC NHW (TARCC MA) .960 −.85
4 TARCC MA (TARCC NHW) .963 −.82
5 H-EPESE (TARCC MA) .937 −.74
6 H-EPESE (TARCC NHW) .941 −.77
Open in a new tab

Notes. AUC = area under the receiver operating characteristic curve; CDR-SOB = Clinical Dementia Rating scale Sum of Boxes; H-EPESE = Hispanic Established Population for Epidemiological Studies in the Elderly; MA = Mexican American; NHW = non-Hispanic White; TARCC = Texas Alzheimer’s Research and Care Consortium.

aNo significant crossgroup differences by DeLong’s method.

*p < .001.

The H-EPESE dMA homologs’ factor weights also correlated strongly with CDR-SOB and achieved high AUC regardless of which well-characterized convenience sample they were exported to (Table 2). Regardless of the data set, there were no significant differences between the AUCs achieved by dMA within a cohort vs. those achieved by either imported homolog.

In each well-characterized cohort, optimal dMA thresholds were chosen by ROC for AD’s discrimination from NCs, both for the local dMA homolog and for the imported H-EPESE homolog. These thresholds and their respective sensitivities and specificities are presented in Table 3.

Table 3.

dMA Thresholds and Estimated Prevalence of Dementia

Path (Figure 1) dMA homolog source (sample) Thresholda Sensitivitya Specificitya Prevalence of ADb
7 H-EPESE (TARCC NHW) −0.3752 0.90 0.84 21.4
8 H-EPESE (TARCC MA) −0.3849 0.88 0.80 21.0
Open in a new tab

Notes. AD = Alzheimer’s disease; dMA = MA-specific δ homolog; H-EPESE = Hispanic Established Population for Epidemiological Studies in the Elderly; MA = Mexican American; NHW = non-Hispanic White; TARCC = Texas Alzheimer’s Research and Care Consortium.

aIn TARCC.

bWhen applied to H-EPESE.

Finally, we exported the TARCC dMA homologs’ factor weights to H-EPESE. Using the thresholds in Table 3, we estimated the prevalence of all-cause dementia in H-EPESE. That prevalence was estimated at 21.4% by H-EPESE’s homolog using the threshold validated in TARCC MA, and at 21.0% by H-EPESE’s homolog using the threshold validated in TARCC NHW.

Discussion

This analysis confirms the potential of latent δ homologs to achieve accurate dementia diagnoses from very brief cognitive batteries. We specifically confirm dMA’s strong AUC for the discrimination of AD versus controls on the basis of MMSE and CLOX scores (Royall & Palmer, 2013). This analysis goes beyond that report to demonstrate that (i) a δ homolog’s factor weights can be applied to other samples and ethnicities and that (ii) it may be possible to export δ composites specifically from populations into well-characterized cohorts for validation.

The latter finding is of singular importance. The MMSE and CLOX are simple bedside measures that can be easily obtained by trained psychometricians in the field. Both are available in multiple translations, including Spanish. It may be feasible then to diagnose dementia by such measures in H-EPESE, or similar settings and populations that do not lend themselves to comprehensive psychometric assessment and expert consensus adjudication of clinical diagnoses. Furthermore, although we benefitted from MMSE and CLOX’s overlap in TARCC’s and H-EPESE’s data sets, δ’s indifference to its indicators suggests that other δ homologs could be developed from the cognitive assessments in many other population-based cohorts, even from the item sets of a single measure (Royall et al., 2016b).

Furthermore, we appear to have succeeded in estimating the prevalence of all-cause dementia in a MA population–based sample on the basis of psychometrics alone. This is possible for three reasons: (i) IADL, and therefore dementia as we have previously shown, have no association with observed cognitive performance except through its association with δ, and ultimately with g. This implies that domain-specific variance in cognitive performance (as distinct from g) is irrelevant to dementia case finding. (ii) Both g and δ are indifferent to their indicators. As a consequence, dementia might be diagnosed from any cognitive battery by this method, and the need for the so-called “comprehensive” batteries is undermined. (iii) δ is a latent variable, and therefore relatively free of measurement error. This likely includes cross-sample differences in age, education, comorbidities, medication usage, sample frames, language of assessment, and so on, which might otherwise pose substantial obstacles to the equation of cases across samples on the basis of any observed cognitive measure.

Both our estimates concur with a 21% prevalence of all-cause dementia in MAs. For comparison, Haan and colleagues (2003) estimated a prevalence of 4.8% among community-dwelling MAs older than 60 years. However, the prevalence was strongly influenced by age. In those aged 85 and older, the prevalence was 31%. H-EPESE’s mean age is 77.4 years, and so our intermediate prevalence estimate appears consistent with that of Haan and colleagues’.

Like H-EPESE, Haan and colleagues’ participants were interviewed in their homes, in a 2-hour interview administered by nonclinicians. In contrast to our approach, the adjudication of dementia status required clinical examination and formal psychometric testing. Referral for psychometric testing was based on performance below the 20th percentile on simple bedside screens, augmented by a 20% random selection. Because of the high prevalence of low educational attainment and lack of English proficiency among the participants, these thresholds were adjusted statistically. “Impairment” of formal psychometrics was set at the 10th percentile relative to a normative sample (of community-dwelling MA?) and followed by neurological examination. In comparison with this arduous process, our estimate derives from bedside psychometrics alone, obtained in the participants’ homes by trained lay persons.

However, there are also several caveats to our finding. First, CLOX only became available in H-EPESE’s Wave 3. H-EPESE’s generalizability to the MA population may have been eroded by nonrandom attrition at prior waves. Attrition may select for a more demented sample, as more intact individuals move out of state, or as the more demented persons, or their caregivers, become motivated to keep their longitudinal contacts with the investigators. Had CLOX been available in H-EPESE’s first wave, we might have arrived at a different and more accurate estimate of all-cause dementia.

Second, there are limitations to δ’s reification as a composite score. It is inevitable that some measurement error is reintroduced at this step, and composite scores seldom work as well as the latent variables that generate them. For example, we reported a correlation of r = .71 (p < .001) between an item-level d-score composite and CDR scores (Royall et al., 2016b). The latent δ homolog that generated it correlated (r = .78) with CDR (p < .001).

A greater potential challenge is the problem of “factor score indeterminacy” (Grice, 2001). An essential limitation of the common factor model is that an infinite number of unique factor score composites can be derived from any factor. Although all might be consistent with the factor’s loadings, some composites may be orthogonal to others, or even inversely related, potentially resulting in wildly discrepant participant rankings, depending on the composite selected.

However, these can be divided into “determinant” and “indeterminant” fractions (Guttman, 1955). Fortunately, many common factor score estimates are highly intercorrelated and yield identical reproduced covariance matrices (Beauducel, 2007). Several statistical methods are available to test a factor’s determinacy. We have tested δ in TARCC (Royall, Palmer, & O’Bryant, 2012) by Grice’s “Refined Factor Score Evaluation Program (Equation 5)” (Grice, 2001) and found its determinacy to be adequate (i.e., having a total item squared multiple correlation = .84). This method maximizes composite validity and is recommended when the factor composite scores are to be used as “observed” variables in subsequent analyses (e.g., as predictors). However, factor determinacy may need to be tested for each individual δ homolog or ortholog before it can be validated as a clinical phenotype.

Finally, δ appears to be agnostic to dementia’s etiology. Gavett and colleagues (2015) demonstrated δ’s ROC curve or AUC of .96 for all-cause dementia in the National Alzheimer’s Coordinating Center’s Unified Data Set which contained approximately 14,000 dementia cases of multiple etiologies. δ has the potential to reveal the disabling, and therefore arguably “dementing” aspects of a number of conditions, not currently considered to be dementias, e.g., by clinicians who ignore the essential disabling character of their associated cognitive impairments. This list might include, but not be limited to, certain medical conditions (e.g., diabetes mellitus, HIV, and other infectious or viral diseases), postoperative cognitive decline, postchemotherapy cognitive decline (popularly known as “Chemobrain”), traumatic brain injury, certain neuropsychiatric illnesses (e.g., major depression and schizophrenia [previously known as “dementia praecox”]), and the dementia of normal aging (also referred to as “senility”). Thus, it remains to be seen whether the cases identified with “dementia” by this method suffer from AD, from other conditions, or from combinations of potentially dementing pathologies.

Our approach is amenable to further refinements. For example, dMA’s indicators (i.e., CLOX and MMSE) can be embedded in larger batteries, with more comprehensive psychometric assessments (as in Royall & Palmer, 2013; Royall et al., 2016a). While comprehensive assessment may not be necessary for accurate dementia case finding, δ’s diagnostic accuracy does appear to benefit from larger batteries and from formal psychometrics (Koppara et al., 2016). δ ROC curves or AUCs as high as .999 have been reported for AD versus NCs, compared with the current weaker discriminations (i.e., 0.92–0.97). Thus, dMA’s diagnostic accuracy in H-EPESE might be further improved if its factor weights were informed by multivariate associations with a broad range of formal measures (e.g., by exporting a restricted composite from a more comprehensive TARCC-based dMA homolog). We have successfully exported such a restricted composite from TARCC to Japan for validation and achieved an ROC of .97 in Japanese data despite the fact that it was informed only by MMSE and CLOX1/2, but not by IADL (Royall et al., 2016a).

It remains to be seen whether a δ homolog can be constructed de novo in a well-characterized convenience sample and then exported into a population-based cohort. Differences in the observed distributions of their indicators and the effects on measurement error on their factor weights may prevent this. We intend to address these issues in future investigations.

In summary, we have successfully validated a latent dementia proxy from a population-based cohort in two well-characterized ethnically diverse samples. This could represent a significant advance in dementia epidemiology. This method opens the door to valid dementia case finding in large population-based cohorts. Moreover, our approach is modular and not limited to dementia case finding. It can easily be modified to assess other clinical syndromes or phenotypes.

Supplementary Material

Supplementary data is available at The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences online.

Funding

This study was made possible by the Texas Alzheimer’s Research and Care Consortium (TARCC) funded by the state of Texas through the Texas Council on Alzheimer’s Disease and Related Disorders and National Institute on Aging (2R01AG010939-21).

Supplementary Material

Supplemental_Table
Click here for additional data file. (22.3KB, docx)

Acknowledgments

D. R. Royall and R. F. Palmer have disclosed the results of these analyses to the University of Texas Health Science Center at San Antonio (UTHSCSA), which has filed patent application 2012.039.US1.HSCS and provisional patents 61/603,226 and 61/671,858 relating to the latent variable δ’s construction and biomarkers. D. R. Royall and R. F. Palmer have been engaged as Consultants by Actavis + Allergan (formerly Forest Laboratories) regarding δ.

Investigators from the Texas Alzheimer’s Research and Care Consortium

Rachelle Doody, MD, PhD, Mimi M. Dang, MD, Valory Pavlik, PhD, Wen Chan, PhD, Paul Massman, PhD, Eveleen Darby, Monica Rodriguear, RN, Aisha Khaleeq (Baylor College of Medicine, Houston, TX)

Chuang-Kuo Wu, MD, PhD, Matthew Lambert, PhD, Victoria Perez, Michelle Hernandez (Texas Tech University Health Sciences Center, Lubbock, TX)

Thomas Fairchild, PhD, Janice Knebl, DO, Sid E. O’Bryant, PhD, James R. Hall, PhD, Leigh Johnson, PhD, Robert C. Barber, PhD, Douglas Mains, Lisa Alvarez, Rosemary McCallum (University of North Texas Health Science Center, Fort Worth, TX)

Perrie Adams, PhD, Munro Cullum, PhD, Roger Rosenberg, MD, Benjamin Williams, MD, PhD, Mary Quiceno, MD, Joan Reisch, PhD, Ryan Huebinger, PhD, Natalie Martinez, Janet Smith (University of Texas Southwestern Medical Center, Dallas, TX)

Donald Royall, MD, Raymond Palmer, PhD, Marsha Polk (University of Texas Health Science Center, San Antonio, TX)

Farida Sohrabji, PhD, Steve Balsis, PhD, Rajesh Miranda, PhD (Texas A&M University Health Science Center, Round Rock, TX)

Stephen C. Waring, DVM, PhD (Essentia Institute of Rural Health, Duluth, MN)

Kirk C. Wilhelmsen, MD, PhD, Jeffrey L. Tilson, PhD, Scott Chasse, PhD (University of North Carolina, Chapel Hill, NC)

References

  1. Alzheimer’s Association. (2010). Alzheimer’s disease facts and figures. Special report: Race, ethnicity and Alzheimer’s disease. Retrieved from http://www.alz.org/documents_custom/report_alzfactsfigures2010.pdf (Accessed September 24, 2015). [Google Scholar]
  2. Arbuckle J. L. (2006). Analysis of Moment Structures-AMOS (Version 7.0) [Computer Program]. SPSS, Chicago, IL. [Google Scholar]
  3. Beauducel A. (2007). In spite of indeterminacy many common factor score estimates yield an identical reproduced covariance matrix. Psychometrika, 72, 437–441. [Google Scholar]
  4. Bentler P. M. (1990). Comparative fit indexes in structural models. Psychological Bulletin, 107, 238–246. [DOI] [PubMed] [Google Scholar]
  5. Bollen K. A., & Long J. S (1993). Testing structural equation models. Thousand Oaks, CA: Sage. [Google Scholar]
  6. Browne M., & Cudeck R (1993). Alternative ways of assessing model fit. In Bollen K. A., Long J. S. (Eds.), Testing structural equation models (pp. 136–162). Thousand Oaks, CA: Sage. [Google Scholar]
  7. DeLong E. R. DeLong D. M., & Clarke-Pearson D. L (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics, 44, 837–845. [PubMed] [Google Scholar]
  8. Ennis, S.R., Ríos-Vargas, M., & Albert, N.G. (2011). The Hispanic Population: 2010: 2010 Census Briefs. Retrieved from http://www.census.gov/prod/cen2010/briefs/c2010br-04.pdf
  9. Espino D. V. DelAguila D. Mouton C. P. Alford C. Parker R. W. Miles T. P., & Olivares O (2002). Characteristics of Mexican-American elders with dementia presenting to a community-based memory evaluation program. Ethnicity & Disease, 12, 517–521. [PubMed] [Google Scholar]
  10. Farrer L. A. Cupples A. Haines J. L. Hyman B. Kukull W. A. Mayeux R., … van Duijn C. M (1997). Effects of age, sex, and ethnicity on the association between apolipoprotein E genotype and Alzheimer disease: A meta-analysis. Journal of the American Medical Association, 278, 1349–1356. [PubMed] [Google Scholar]
  11. Folstein M. F. Folstein S. E., & McHugh P. R (1975). “Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician. Journal of Psychiatric Research, 12, 189–198. [DOI] [PubMed] [Google Scholar]
  12. Gavett B. E. Vudy V. Jeffrey M. John S. E. Gurnani A., & Adams J (2015). The δ latent dementia phenotype in the NACC UDS: Cross-validation and extension. Neuropsychology, 29, 344–352. doi:10.1037/neu0000128 [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Graham J. W. (2009). Missing data analysis: Making it work in the real world. Annual Review of Psychology, 6, 549–576. doi:10.1146/annurev.psych.58.110405.085530 [DOI] [PubMed] [Google Scholar]
  14. Grice J. W. (2001). Computing and evaluating factor scores. Psychological Methods, 6, 430–450. doi:10.3389/fpsyg.2015. 01007 [PubMed] [Google Scholar]
  15. Guttman L. (1955). The determinancy of factor score matrices with applications for five other problems of common factor theory. British Journal of Statistical Psychology, 8, 65–82. [Google Scholar]
  16. Haan M. N., Mungas D. M., Gonzalez H. M., Ortiz T. A., Acharya A., Jagust W. J. (2003). Prevalence of dementia in older Latinos: The influence of type 2 diabetes mellitus, stroke and genetic factors. Journal of the American Geriatrics Society, 51, 169–177. [DOI] [PubMed] [Google Scholar]
  17. Hughes C. P. Berg L. Danziger W. L. Coben L. A., & Martin R. L (1982). A new clinical scale for the staging of dementia. British Journal of Psychiatry, 140, 566–572. [DOI] [PubMed] [Google Scholar]
  18. Koppara A. Wolfsgruber S. Kleineidam L. Schmidtke K. Frölich L. Kurz A., … Wagner M (2016). The latent dementia phenotype δ is associated with CSF biomarkers of Alzheimer Disease and predicts conversion to AD dementia in subjects with MCI. Journal of Alzheimer’s Disease, 49, 547–560. doi:10.3233/JAD-150257 [DOI] [PubMed] [Google Scholar]
  19. Lawton M. P., & Brody E. M (1969). Assessment of older people: Self-maintaining and instrumental activities of daily living. The Gerontologist, 9, 179–186. [PubMed] [Google Scholar]
  20. Maixner S. M. Burke W. J. Roccaforte W. H. Wengel S. P., & Potter J. F (1995). A comparison of two depression scales in a geriatric assessment clinic. American Journal of Geriatric Psychiatry, 3, 60–67. [DOI] [PubMed] [Google Scholar]
  21. McKhann D. Drockman D. Folstein M. Katzman R. Price D., & Stadlan E. M (1984). Clinical diagnosis of Alzheimer’s disease: Report of the NINCDS-ADRDA Work Group. Neurology, 34, 939–944. [DOI] [PubMed] [Google Scholar]
  22. Metz C. E. (1978). Basic principles of ROC analysis. Seminars in Nuclear Medicine, 8, 283–298. [DOI] [PubMed] [Google Scholar]
  23. O’Bryant S. E. Johnson L. Edwards M. Soares H. Devous M. D. Ross S., … Hall J; Texas Alzheimer’s Research & Care Consortium (2013. a). The link between C-reactive protein and Alzheimer’s disease among Mexican Americans. Journal of Alzheimer’s Disease, 34, 701–706. doi:10.3233/JAD-122071 [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. O’Bryant S. E. Xiao G. Edwards M. Devous M. Gupta V. B. Martins R., … Barber R (2013. b). Biomarkers of Alzheimer’s disease among Mexican Americans. Journal of Alzheimer’s Disease, 34, 841–849. doi:10.3233/JAD-122074 [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Raji M. A. Al Snih S. Ray L. A. Patel K. V., & Markides K. S (2004). Cognitive status and incident disability in older Mexican Americans: Findings from the Hispanic established population for the epidemiological study of the elderly. Ethnicity & Disease, 14, 26–31. [PubMed] [Google Scholar]
  26. Royall D. R., & Palmer R. F (2013). Validation of a latent construct for dementia case-finding in Mexican-Americans. Journal of Alzheimer’s Disease, 37, 89–97. doi:10.3233/JAD-130353 [DOI] [PubMed] [Google Scholar]
  27. Royall D. R., & Palmer R. F (2015). Ethnicity moderates dementia’s biomarkers. Journal of Alzheimer’s Disease, 43, 275–287. doi:10.3233/JAD-140264 [DOI] [PubMed] [Google Scholar]
  28. Royall D. R. Cordes J. A., & Polk M (1998). CLOX: An executive clock drawing task. Journal of Neurology, Neurosurgery, and Psychiatry, 64, 588–594. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Royall, D. R., Espino, D. V., Polk, M. J., Palmer, R., & Markides, K. S. (2004). Prevalence and patterns of executive impairment in the Hispanic EPES Study. International Journal of Geriatric Psychiatry, 19, 926–934. [DOI] [PubMed] [Google Scholar]
  30. Royall D. R. Palmer R. F., & O’Bryant S. E (2012). Validation of a latent variable representing the dementing process. Journal of Alzheimer’s Disease, 30, 639–649. doi:10.3233/JAD-2012-120055 [DOI] [PubMed] [Google Scholar]
  31. Royall, D. R., Matsuoka, T., Palmer, R. F., Kato, Y., Taniguchi, S., Ogawa, M., … Narumoto, J. (2015). Greater than the sum of its parts: δ Improves upon a battery’s diagnostic performance. Neuropsychology, 29, 683–692. doi:10.3233/JAD-150261 [DOI] [PubMed]
  32. Royall D. R. Mulroy A. Chiodo L. K., & Polk M. J (1999). Clock drawing is sensitive to executive control: A comparison of six methods. Journal of Gerontology: Psychological Science, 54B, P328–P333. [DOI] [PubMed] [Google Scholar]
  33. Royall D. R. Palmer R. F. Vidoni E. D., & Honea R. A (2012). The default mode network may be the key substrate of depression related cognitive changes. Journal of Alzheimer’s Disease, 34, 547–560. doi:10.3233/JAD-121639 [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Royall D. R. Palmer R. F. Vidoni E. D. Honea R. A., & Burns J. M (2012). The default mode network and related right hemisphere structures may be the key substrates of dementia. Journal of Alzheimer’s Disease, 32, 467–478. doi:10.3233/JAD-2012-120424 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Royall D. R. Palmer R. F. Matsuoka T. Kato Y. Taniguchi S. Ogawa M., … Narumoto J (2016. a). δ Scores are exportable across cultural and linguistic boundaries. Journal of Alzheimer’s Disease, 49, 561–570. [DOI] [PubMed] [Google Scholar]
  36. Royall D. R. Palmer R. F. Matsuoka T. Kato Y. Taniguchi S. Ogawa M., … Narumoto J (2016. b). Greater than the sum of its parts: δ can be constructed from item-level data. Journal of Alzheimer’s Disease, 49, 571–579. doi:10.3233/JAD-150250 [DOI] [PubMed] [Google Scholar]
  37. Salazar R. Palmer R. F., & Royall D. R (2015). Neuropsychiatric symptoms in community-dwelling Mexican-Americans: Results from the Hispanic EPESE Study. International Journal of Geriatric Psychiatry, 30, 300–307. doi:10.1002/gps.4141 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Schafer J. L., Graham J. W. (2002). Missing data: Our view of the state of the art. Psychological Methods, 7, 147–177. [PubMed] [Google Scholar]
  39. Spearman C., & Wynn Jones L. L (1951). Human ability. London, UK: Macmillan. [Google Scholar]
  40. Waring S. O’Bryant S. E. Reisch J. S. Diaz-Arrastia R. Knebl J., & Doody R; for the Texas Alzheimer’s Research Consortium (2008). The Texas Alzheimer’s Research Consortium longitudinal research cohort: Study design and baseline characteristics. Texas Public Health Journal, 60, 9–13. [Google Scholar]
  41. Yesavage J. A. Brink T. L. Rose T. L. Lum O. Huang V. Adey M., & Leirer V. O (1982). Development and validation of a geriatric depression screening scale: A preliminary report. Journal of the American Geriatrics Society, 29, 164–171. [DOI] [PubMed] [Google Scholar]
  42. Zweig M. H., Campbell G. (1993). Receiver-operating characteristic (ROC) plots: A fundamental evaluation tool in clinical medicine. Clinical Chemistry, 39, 561–577. [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental_Table
Click here for additional data file. (22.3KB, docx)

Articles from The Journals of Gerontology Series B: Psychological Sciences and Social Sciences are provided here courtesy of Oxford University Press

RESOURCES