INTRODUCTION
No clear recommendations exist regarding the optimal normative reference samples that should be used to interpret cognitive performance in older adults undergoing assessment for dementia or mild cognitive impairment (MCI). Although it is common to standardize raw scores using published normative data, it is also possible to do this using the mean and standard deviation of a matched control group, or by establishing age-, sex-, and education-adjusted norms from regression coefficients derived from a healthy sample. In dementia investigations, the presence and severity of cognitive deficits vary considerably depending on the source of norms, with local norms generally resulting in higher estimates of amnestic MCI than published norms (Arsenault-Lapierre et al., Reference Arsenault-Lapierre, Whitehead, Belleville, Massoud, Bergman and Chertkow2011). Similar findings have been reported for IQ estimates: scores of Canadian subjects are consistently lower when interpreted using norms generated from a Canadian sample rather than an albeit larger American sample (Harrison, Armstrong, Harrison, Lange, & Iverson, Reference Harrison, Armstrong, Harrison, Lange and Iverson2014). Several other authors have additionally highlighted the importance of generating local normative data for use in non-North-American samples (e.g., Daugherty, Puente, Fasfous, Hidalgo-Ruzzante, & Pérez-Garcia, Reference Daugherty, Puente, Fasfous, Hidalgo-Ruzzante and Pérez-Garcia2017; Rivera & Arango-Lasprilla, Reference Rivera and Arango-Lasprilla2017).
Local norms (i.e., generated from the performance of a healthy matched control group) may result in more sensitive estimates of cognitive impairment because they account for biases known to affect test results, including cultural, educational, and linguistic factors (Anderson, Reference Anderson2001). In dementia research, this has only been investigated in ethnic or linguistic minorities for whom published norms are obviously inappropriate (e.g., non-English-speaking individuals). It is unclear if the source of norms affects disease estimates in individuals who closely resemble published normative samples (e.g., English-speaking, American, Caucasian, generally well educated). It is plausible that unobservable idiosyncrasies within communities may result in more accurate cognitive estimates using local norms, even when the published sample appears representative of the individual undergoing assessment in terms of ethnic or linguistic background. This question has important clinical and research implications for tracking disease prevalence rates, inclusion of appropriate subjects into clinical trials, patient management, and prognosis.
This study aimed to determine whether published norms or norms developed from a healthy locally recruited sample are most sensitive, specific, and accurate for detecting early dementia in the Uniform Data Set (UDS) (Weintraub et al., Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009), a widely used set of standardized clinical observations.
METHOD
Participants
Participants were 526 adults aged ≥50 from the first phase of the Alzheimer’s Disease Neuroimaging Initiative (ADNI; adni.loni.usc.edu) who were nondemented upon admission, had complete baseline data, and ≥2 visits. Data were collected between August 2005 and June 2011. There are three ADNI phases (ADNI1, ADNIGO, and ADNI2), but digit span and coding were dropped from the UDS after ADNI1, and because these scores were necessary to determine dementia risk in this study, only ADNI1 data were used. ADNI does not enroll participants with neurologic disease (other than Alzheimer’s disease), brain infection or lesions, mood disorders, schizophrenia, substance abuse, systemic illnesses, or unstable medical conditions; thus, the sample was free from these conditions. Those reporting “extenuating circumstances contributing to cognitive or functional change” were also excluded. Of the total sample, 205 participants were classified as cognitively normal at baseline based on Mini Mental State Exam ≥24, Clinical Dementia Rating Scale (CDR) = 0, no memory complaint, or objective memory loss [operationalized using education-adjusted performance on the Logical Memory II subscale of the Wechsler Memory Scale Revised (WMS-R)]. These participants contributed to generating “sample norms” (described later).
All participants were evaluated annually for dementia, which was diagnosed based on clinical judgment using the following criteria: (1) subjective cognitive complaint; (2) abnormal cognitive function (operationalized using education-adjusted performance on the Logical Memory II subscale of the WMS-R) representing a decline from previous and not better explained by another disorder; and (3) impacted functional independence (operationalized as CDR > 0). Procedures for establishing cognitive status in ADNI are described in detail at http://adni.loni.usc.edu/wp-content/uploads/2010/09/ADNI_GeneralProceduresManual.pdf. ADNI’s clinical diagnosis of dementia is considered valid per current definitions of a dementia syndrome (Chertkow, Feldman, Jacova, & Massoud, Reference Chertkow, Feldman, Jacova and Massoud2013) and is considered reliable across ethnic categories based on previous research indicating similar neuropathologic characteristics at autopsy between White and non-White patients with similar clinical dementia severity (Mehta et al., Reference Mehta, Yaffe, Pérez-Stable, Stewart, Barnes, Kurland and Miller2008).
ADNI centers obtained ethical approval from their institutional research boards, and subjects consented to participate.
Materials
The UDS (Weintraub et al., Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009) was administered at approximately yearly intervals, assessing episodic memory (Logical Memory story A immediate and delayed recall), language (animal fluency, naming), simple attention (digit span forward, Trails A, coding), and executive functions (digit span backward, Trails B).
Statistical Analyses
Participants who took longer than the maximum time allowed to complete Trails A (180 s) and Trails B (300 s) were assigned maximum values for these subsets, as per guidelines (Strauss, Sherman, & Spreen, Reference Strauss, Sherman and Spreen2006). Age and education were first centered (sample mean age and education were subtracted from each participant’s age and education, respectively), and both linear and quadratic relationships between the independent variables and test scores were explored. Age was found to have linear relationships with digit span backward, animal fluency, Trails A, and naming, and quadratic relationships with immediate and delayed story recall, digit span forward, Trails B, and coding. Education was found to have linear relationships with immediate and delayed story recall, digit span forward and backward and animal fluency, and quadratic relationships with Trails A and B, naming, and coding. The regression equations used to create sample norms, described below, account for these relationships.
Raw scores were standardized to Z-scores using published UDS norms, which were adjusted for age, sex, and education using a regression-based procedure in the original publication (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011). Raw scores were again standardized to Z-scores using regression-based norms derived from baseline data of ADNI control participants (n = 205), following the same procedure described by Shirk et al. (Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011). Specifically, multivariate regression models were created for individual tests, with age (in years, centered), sex (coded as 1 = male, 2 = female), and education (in years, centered) entered as predictors in all models. Where relationships between predictors and test scores were nonlinear, as described earlier, models included Age2 or Education2 instead of Age and Education. Predicted scores for each test were then estimated for individual subjects using the regression equation constant and weighted coefficients associated with each of the predictors in the model. The corresponding regression equation for each test can be written as: Predicted score = Constant + (β age*Age) + (β sex*Sex) + (β education*Education), substituting Age2 or Education2 instead of Age and Education in cases of quadratic relationships between variables. The three predictors were always included in all models, even if they did not contribute statistically significantly to the model. Although this may theoretically introduce error and impact the precision of regression estimates, comparisons of the models’ adjusted R 2 values indicated that the removal of nonsignificant predictors improved R 2 only by .004 at most (data not shown), suggesting that the error introduced by these variables is quite small. Furthermore, the original published sample also used nonsignificant predictors in all regression models (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011; Weintraub et al., Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009), resulting in arguably similar sources of error across both published and sample-based norms. There were no issues of multicollinearity between the predictors [variance of inflation factor (VIF) values 1.00–1.03]. Visual inspection of the model residuals found that they were normally distributed and homoscedastic. There were no multivariate outliers (all Mahalanobis distances < 13.81).
Z-scores were then estimated for individual subjects using the formula Z = (Raw score–Predicted score)/(Root of the mean square residual of the regression equation). Because the goal of this paper was to directly compare outcomes generated from published and sample-based norms, the creation of sample-based norms did not include race/ethnicity effects nor interactions between significant variables, to remain consistent with (and comparable to) the published normative dataset which did not include interactions or race/ethnicity effects (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011). Similarly, to ensure comparability of the datasets, all models included age, sex, and education even if they were not statistically significant predictors of cognitive outcomes, as this was the procedure used in the published normative dataset (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011).
Thus, two separate normative samples were used as benchmarks for “normal” performance: (1) a large (n = 3268) group of cognitively normal adults whose UDS results were published previously (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011; Weintraub et al., Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009; “published norms”) and (2) the local sample, cognitively normal controls within ADNI (n = 205; “sample norms”). Importantly, control participants in the ADNI sample met the same clinical criteria for “normal cognition” as those used in the published sample (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011; Weintraub et al., Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009), including (1) no cognitive impairment; (2) CDR = 0; (3) no functional impairment or dementia; and (4) available age, gender, and education information (Alzheimer’s Disease Neuroimaging Initiative, 2005).
Participants obtaining two scores < −1 SD within a cognitive domain were classified as “at risk for future dementia.” This was based on evidence that diagnostic accuracy is greatly increased by the use of at least two tests to establish impairment (Loewenstein et al., Reference Loewenstein, Acevedo, Potter, Schinka, Raj, Greig, Agron, Barker, Wu, Small, Schofield and Duara2009), as many cognitively normal adults commonly obtain one impaired score on a battery (Brooks, Iverson, Holdnack, & Feldman, Reference Brooks, Iverson, Holdnack and Feldman2008). A −1 SD cutoff was chosen to optimize the trade-off between sensitivity and specificity, consistent with prior work (Callahan, Ramirez, Berezuk, Duchesne, & Black, Reference Callahan, Ramirez, Berezuk, Duchesne and Black2015; Edmonds et al., Reference Edmonds, Delano-Wood, Jak, Galasko, Salmon and Bondi2016).
Risk was determined separately using published and sample norms. Both methods were compared using chi-square in terms of their sensitivity [true positives/(true positives + false negatives)], specificity [true negatives/(true negatives + false positives)], and accuracy [(true positives + true negatives)/total sample] to predict dementia at any follow-up visit. Positive predictive value (PPV) was calculated as (sensitivity × dementia base rate)/{[(sensitivity × base rate)] + [(1−specificity) × (1−base rate)]}, and negative predictive value (NPV) was calculated as [specificity × (1−base rate)]/[(specificity × (1−base rate)] + [(1−sensitivity) × base rate)]. PPV and NPV were calculated using dementia base rates from ADNI (27.4%), as this sample is enriched for dementia and is not representative of population base rates. In addition, classification parameters were examined at the test level following the approach described by Van Breukelen and Vlaeyen (Reference Van Breukelen and Vlaeyen2005), in which the difference between risk scores derived from published and sample norms is plotted against their average in a standard scatterplot, with each data point representing a single participant. This procedure allows to visually compare the overlap between both methods as participants’ individual test scores vary. Because differences in scores tended to be linked to sex (see Results), these results were plotted separately for males and females.
The threshold for statistical significance was p < .05. Effect sizes are reported as r and ω, where .1, .3, and .5 designate small, medium, and large effect sizes, respectively (Cohen, Reference Cohen1988).
RESULTS
Participants
All participants had complete UDS data for ≥2 visits (mean 2.9, range .5–4.5 years). Over the course of ADNI1, 144 participants (27.4%) developed dementia. There was no difference in follow-up length between participants who did and did not develop dementia (both 2.9 years; t = −.34, p = .733). Alzheimer’s disease was the primary presumed etiology in nearly all cases (96.5%).
Normative Sample
Participants had a mean baseline age of 75.1 years old, were predominantly White (90.9%), and generally well-educated (15.9 years education, on average), very similar to the characteristics of the published cohort described by Weintraub and colleagues (Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009).
Classification Results
Final multiple linear regression models for each neuropsychological score are reported in Table 1. Classifications using published and sample norms for all measures were highly correlated (r = .867–.997). Thus, most participants were classified similarly, regardless of norms used. Discrepancies constituted 9.1% of the sample, and published norms systematically tended to underestimate risk relative to sample norms (i.e., resulted in false negatives). Discrepant cases were more likely to be female (73%; χ 2 = 23.66, p < .001, ω = .21) than nondiscrepant cases (34%), but no other differences were apparent.
Final multiple linear regression models for each neuropsychological score, based on data from 205 healthy controls within ADNI
Notes. In all models, age and education were entered in years and centered. Sex was coded as 1 = male, 2 = female. Where relationships between predictors and test scores were nonlinear, models included Age2 or Education2.
In terms of classification parameters, published norms had 86.1% sensitivity (79.1−91.1), 62.0% specificity (56.9−66.9), and 68.6% accuracy (64.4−72.5). PPV was 46.1% (40.1−52.3), and NPV was 92.2% (88.1−95.1). In comparison, sample norms were more sensitive [91.0% (84.8−94.9), χ 2 = 4.03, p = .045, ω = .09], less specific [52.9% (47.7−58.0), χ 2 = 37.95, p < .001, ω = .27], and less accurate [63.3% (59.0−67.4), χ 2 = 16.86, p < .001, ω = .18]. PPV was slightly worse [42.1% (36.6−47.8), χ 2 = 1.30, p = .254, ω = .05], and NPV was slightly better [94.0% (89.6−96.6), χ 2 = .13, p = .715, ω = .02], but these differences were nonsignificant. Both published and sample norms identified impairments primarily in memory (86.6% and 90.3% of the sample, respectively), followed by attention (27.5% and 30.2%), language (27.5% and 24.0%), and executive function (12.3% and 13.6%). Published and sample norms identified multidomain impairments in 37.9% and 40.6% of cases, respectively.
Differences between published and sample norms, plotted by sex, are illustrated in Figure 1. These plots indicate, first, that differences between the two Z-score calculations (Y-axes) are more marked for some tests (e.g., story recall, naming) than for others (e.g., digit span, animal fluency). Second, differences were greater for female participants on story recall, backward digit span, animal fluency and coding, and slightly greater for men on forward digit span and naming. Lastly, differences were most significant as participants’ performance level decreased (X-axes) on tests of story recall and naming, but remained relatively constant across performance ranges for other tests.
Differences between scores generated using published and sample norms, plotted by sex.
Post hoc exploratory analyses were performed to determine whether racial minorities (White vs. non-White) and those with less than the sample’s median education (16 years) had different risk profiles based on their UDS scores relative to their White, better-educated counterparts. At the whole sample level (n = 526), there were no differences based on race or education. However, among those who eventually developed dementia (n = 144), there were significant differences in risk prediction using both published norms (t = −4.83, p < .001, r = .28) and sample norms (t = −3.78, p < .001, r = .22) between racial classes. Dementia risk calculated using both published and sample norms correctly identified 100% of the future dementia cases in non-Whites (n = 10), but only 85% and 90%, respectively, of future dementia cases in Whites (n = 134). There was also a significant difference in risk prediction based on education: sample norms scores correctly identified 98% of the future dementia cases in lower educated participants (n = 47), but only 88% of future dementia cases in highly educated participants (n = 97) (t = 2.58, p = .011, r = .20). There was no difference with regard to published norms (t = .27, p = .788, r = .02), which correctly identified 87% and 86% of low- and high-education incident dementia cases, respectively.
Other authors (e.g., Edmonds et al., Reference Edmonds, Delano-Wood, Jak, Galasko, Salmon and Bondi2016) have classified MCI participants based on impairment on only one test in multiple domains. To remain consistent with this literature, analyses were repeated after reclassifying participants based on any two scores < −1 SD, regardless of domain. These new classifications produced an additional “at risk” 61 individuals using published norms and 53 using sample norms (10%–11% of the sample). The sensitivity (95.1%) and NPV (95.7%) of sample norms remained slightly better than those of published norms (93.1% and 94.9%, respectively), and the specificity (48.7%), accuracy (60.8%), and PPV (40.6%) of published norms remained slightly better than those of sample norms (40.6%, 55.5% and 37.6%, respectively). Thus, results remained generally unchanged.
DISCUSSION
In the sample used in this study, published norms appeared slightly more accurate than sample-based norms to detect early dementia from neuropsychological test performance. However, consistent with prior research (Arsenault-Lapierre et al., Reference Arsenault-Lapierre, Whitehead, Belleville, Massoud, Bergman and Chertkow2011), sample-based norms generated higher estimates of impaired performance (particularly on tests of memory, coding, and naming, in which differences >1 SD between published and sample-based norms were not uncommon) and were thus more sensitive to early signs of dementia. This may reflect locally recruited healthy samples being the most representative benchmark for normal performance. Put differently, individuals recruited from similar contexts (e.g., same community, same research study) are arguably more similar to each other than to samples recruited from elsewhere, and deviations from normal within this context may be a stronger indicator of disease. In this study, sociodemographic differences were apparent between individuals identified as “at risk” by sample norms but not published norms, suggesting indeed that participants were more similar within than across samples. Published norms tended to underestimate dementia risk (i.e., result in false negatives), particularly in females, and this difference was most marked for story recall, backward digit span, animal fluency, and coding at the individual test level. This finding suggests that females’ cognitive performance within the published sample is more heterogeneous than that of locally recruited women, which may lead one to conclude that relatively lower performance is not cause for concern on the basis of published norms. Furthermore, post hoc analyses indicated that sample-based norms were best at identifying future dementia cases in participants with less than the sample’s median education (16 years). Accuracy of risk estimates for women and those with relatively lower education may thus be increased using locally generated norms. Certain community-specific characteristics (e.g., religious beliefs or lifestyle choices: Thygesen, Gimsing, Bautz, Hvidt, & Johansen, Reference Thygesen, Gimsing, Bautz, Hvidt and Johansen2017) have also been shown to impact cognitive health and may have played a role in driving our findings, though these data were not comprehensively collected in ADNI.
The major novel contribution of this work involves comparing results from normative samples that were highly similar to each other, in contrast to prior studies which compared results from different countries (e.g., Harrison et al., Reference Harrison, Armstrong, Harrison, Lange and Iverson2014) or different linguistic or cultural backgrounds (e.g., Arsenault-Lapierre et al., Reference Arsenault-Lapierre, Whitehead, Belleville, Massoud, Bergman and Chertkow2011). Participants used to create the sample norms in this study were highly similar to those used to generate published norms (Shirk et al., Reference Shirk, Mitchell, Shaughnessy, Sherman, Locascio, Weintraub and Atri2011; Weintraub et al., Reference Weintraub, Salmon, Mercaldo, Ferris, Graff-Radford, Chui, Cummings, DeCarli, Foster, Galasko, Peskind, Dietrich, Beekly, Kukull and Morris2009), suggesting that prediction can be impacted by norm selection even in the absence of overt demographic disparities. Nonetheless, it must be noted that the differences produced using different normative samples were relatively minor (reflected in small effect sizes) and may not be clinically significant in practice, particularly in smaller samples where the sensitivity/specificity trade-off may be more negligible and dementia prevalence rates are lower. To illustrate, applying a 7% population prevalence of dementia (Prince et al., Reference Prince, Bryce, Albanese, Wimo, Ribeiro and Ferri2013) to the ADNI classification parameters yields PPVs of 14.6% and 12.7% for published and sample norms, respectively, and NPVs of 98.3% and 98.7%. Thus, for most real-world applications, using norms generated from locally recruited healthy controls will likely produce essentially the same results as those obtained from demographically similar published normative data. Furthermore, the most accurate estimates of dementia risk are likely to result from normative data accounting more comprehensively for subject-specific factors, in particular race and ethnicity (Babulal et al., Reference Babulal, Quiroz, Albensi, Arenaza-Urquijo, Astell, Babiloni, Bahar-Fuchs, Bell, Bowman, Brickman, Chételat, Ciro, Cohen, Dilworth-Anderson, Dodge, Dreux, Edland, Esbensen, Evered, Ewers, Fargo, Fortea, Gonzalez, Gustafson, Head, Hendrix, Hofer, Johnson, Jutten, Kilborn, Lanctôt, Manly, Martins, Mielke, Morris, Murray, Oh, Parra, Rissman, Roe, Santos, Scarmeas, Schneider, Schupf, Sikkes, Snyder, Sohrabi, Stern, Strydom, Tang, Terrera, Teunissen, Melo van Lent, Weinborn, Wesselman, Wilcock, Zetterberg and O'Bryant2019), as well as within-person variability across tests (Holtzer, Verghese, Wang, Hall, & Lipton, Reference Holtzer, Verghese, Wang, Hall and Lipton2008) and longitudinal changes in performance (Heilbronner et al., Reference Heilbronner, Sweet, Attix, Krull, Henry and Hart2010).
Limitations to this study include the use of a dataset highly enriched for dementia; results may differ in settings with lower dementia base rates. Also, analyses only included cognitive tests within the UDS, and performance on other tests may be differently impacted by the source of norms. Lastly, one of the memory tests (delayed story recall) was used to diagnose dementia in ADNI, which may have contributed to some criterion contamination. Robustness of results should be replicated using completely independent measures. The requirement that both memory scores fall < −1 SD may also be viewed as a limitation, as impairment in delayed recall alone has been associated with dementia (e.g., Christensen, Hadzi-Pavlovic, & Jacomb, Reference Christensen, Hadzi-Pavlovic and Jacomb1991). However, between 6% and 15% of healthy older adults will obtain delayed recall scores < −1 SD (Palmer, Boone, Lesser, & Wohl, Reference Palmer, Boone, Lesser and Wohl1998), which, in this study, represents up to 79 potential false positives. Furthermore, ADNI classification criteria (not used here) use delayed recall as the sole measure to diagnose MCI, but 70% of these individuals do not progress to dementia (i.e., constitute false positives; Bondi et al., Reference Bondi, Edmonds, Jak, Clark, Delano-Wood, McDonald, Nation, Libon, Au, Galasko and Salmon2014). Thus, requiring two impaired memory scores was felt to be the most robust risk measure in this study.
ACKNOWLEDGMENTS
This work was supported by a Canada Research Chair to Dr. Callahan. Data collection and sharing for this project were funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. The author also gratefully acknowledges statistical assistance from Dr. Audrey Schnell and Dr. Jean-William P. Laliberté and insightful discussions and guidance from Dr. Larry Leach.
CONFLICTS OF INTEREST
The author has no conflicts to declare.
