The Memory Binding Test in a Longitudinal Study of Cognitive Aging and Preclinical Disease

Abstract

Objective:

The Memory Binding Test (MBT) shows promise in detecting early cognitive changes associated with Alzheimer’s Disease (AD). This study assesses the psychometric properties (i.e., construct and criterion validity, test-retest reliability) of the MBT and its sensitivity to incipient disease and incident cognitive impairment.

Method:

149 cognitively unimpaired adults ages 45–85 completed the MBT and neuropsychological tests at baseline; 132 returned for 2-year follow-up. Based on neuroradiological ratings of amyloid PET and MRI markers at baseline, they were categorized as healthy (n = 94) or having preclinical disease (n = 55, either on the AD continuum or having non-AD pathologic change). Construct validity was assessed by the associations between MBT scores, demographics, and neuropsychological scores within the healthy group. Criterion validity was assessed by testing how MBT scores correlate with AD biomarkers, differ and discriminate between groups at baseline, and predict incident cognitive impairment.

Results:

MBT scores decreased with age and were strongly associated with memory and global cognition. MBT scores were largely not associated with amyloid, hippocampal volume, or AD signature cortical volume, but related to white matter lesion volume in the preclinical disease group. The preclinical groups performed worse on MBT immediate free recall at baseline, but no scores predicted incident cognitive impairment at follow-up. Most scales demonstrated modest test-retest reliability.

Conclusions:

This study demonstrates that the MBT has adequate construct validity in cognitively unimpaired adults, moderate sensitivity to preclinical disease cross-sectionally, and limited prognostic utility. Careful consideration of demographic influences on score interpretation remains necessary.

Associative memory, or the ability to encode and store information based on the relationships between elements, declines with age (Gutchess & Park, 2009; Naveh-Benjamin, 2000). Older adults exhibit deficits in binding—the integration of multiple pieces of information into one unit—during encoding (Naveh-Benjamin, 2000; Saverino et al., 2016) and demonstrate suboptimal retrieval strategies (Cohn et al., 2008; Hertzog et al., 2013). In addition, associative memory appears to be particularly susceptible to neurodegenerative disease (Bastin et al., 2014; Polcher et al., 2017). Deficits in associative memory are among the earliest cognitive changes in the preclinical stages of Alzheimer’s disease (AD; Rentz et al., 2013), where biomarkers indicate the presence of pathology in otherwise asymptomatic individuals (Dubois et al., 2016). Thus, there is great interest in leveraging associative memory measures as low-cost, low-burden tools for identifying individuals at risk of developing neurodegenerative conditions, monitoring disease progression, and serving as sensitive endpoints for clinical trials.

One such measure is the Memory Binding Test (MBT), previously the Memory Capacity Test, which measures associative memory by testing the ability to bind related words based on semantic cues (Buschke, 2014). The MBT was developed as an evolution of the Free and Cued Selective Reminding Test (FCSRT; Buschke, 1984; Grober & Buschke, 1987), an episodic memory measure that was widely used in earlier efforts to detect AD-related cognitive changes. The FCSRT was designed to improve upon the limited sensitivity of traditional memory measures to subtle memory deficits in early disease stages (Rentz et al., 2013) by incorporating 1) controlled learning (i.e., searching for information based on category cues) that ensures the specificity of encoding and 2) cued retrieval (Grober & Buschke, 1987). These test features are intended to reduce variability due to individualized learning/compensatory strategies and to minimize the confounding effects of deficits in attention or processing speed due to aging or other conditions, providing a more direct assessment of medial temporal lobe-mediated memory function. Indeed, the FCSRT demonstrates sensitivity to memory impairment and AD pathology in the prodromal (i.e., mild cognitive impairment; MCI) and dementia disease stages (Di Stefano et al., 2015; Grober et al., 2018; Grober et al., 2019; Grober et al., 2022a; Grober et al., 2022b; Mura et al., 2014; Wagner et al., 2012). Building upon these core features of the FCSRT, the MBT is posited to further enhance sensitivity to very early memory dysfunction by exploiting specific vulnerabilities that appear in the preclinical stage—namely susceptibility to proactive and retroactive semantic interference and deficits in associative binding (Buschke, 2014; Loewenstein et al., 2018). Unlike the FCSRT, the MBT presents two word lists serially with each word pair (i.e., one from each list) tied to common semantic cues, introducing interference from competing semantically similar targets and allowing for the direct evaluation of associative memory binding.

One approach to evaluating the utility of the MBT as a screening tool for early disease-related cognitive changes is to test its association with relevant biomarkers. The most recent AD research framework (Jack et al., 2018) describes three biomarker profiles that are biologically defined by the presence of β-amyloid (A+), pathologic tau (T+), or neurodegeneration (N+): 1) the AD continuum, defined by the presence of β-amyloid with or without other biomarkers; 2) non-AD pathologic change, defined by the presence of tau or neurodegeneration without β-amyloid; and 3) normal AD biomarkers. This framework identifies individuals as having preclinical disease if they are cognitively unimpaired in the presence of AT(N) markers. One prior study of the MBT in biomarker-confirmed preclinical AD found that compared to A-/N- participants, A+/N- participants performed worse on delayed free recall and A+/N+ participants performed worse on both the free and paired (i.e., cued) delayed recall (Papp et al., 2015). Prior prospective studies found that among cognitively unimpaired adults, poorer MBT paired (i.e., cued) recall at baseline was associated with an increased risk for developing amnestic MCI (aMCI) over 4–7-years (Mowrey et al., 2016) and dementia over 13 years (Mowrey et al., 2018), supporting its prognostic utility. Although this preliminary evidence suggests that the MBT is a promising tool, more work is needed to ascertain its psychometric properties. Previous studies show moderate convergent validity of the MBT with other measures of associative memory (i.e., the FCSRT; Gramunt et al., 2016), but the extent to which MBT scores correlate with pertinent demographic factors and conventional neuropsychological tests has not been reported, and its ability to discriminate between biomarker groups has yet to be replicated or extended. For an early-detection measure like the MBT to reach wide clinical or research utility, we need to understand how it functions in individuals with and without known biological risk of disease at the critical juncture between normal and abnormal cognition.

Thus, this study sought to assess the construct validity of the MBT in healthy adults and the criterion validity of the MBT in identifying preclinical disease and predicting clinical conversion to MCI. We used an adapted biomarker framework that includes white matter hyperintensities (WMH) as a vascular biomarker, in recognition that cerebrovascular factors contribute to AD pathogenesis (Brickman et al., 2018). This is in line with the prevailing research framework which states, “The AT(N) biomarker scheme is expandable to incorporate new biomarkers… For example, a vascular biomarker group could be added, that is, ATV(N)” (Jack et al., 2018, p. 544). A tau biomarker was not collected in this dataset; thus, the preclinical disease group was defined via ratings of β-amyloid on PET (A+), medial temporal lobe atrophy (N+), and WMH (V+) using validated clinician rating scales. We further parsed this preclinical group into AD continuum and non-AD pathologic change groups, in accordance with the research framework.

This study employed a two-step approach to provide a systematic evaluation of the psychometric properties of the MBT in older adults with and without preclinical disease. First, we assessed construct validity in the healthy group by examining associations between MBT scores and 1) demographic variables and 2) composite cognitive scores using standard neuropsychological tests. We expect that, like most cognitive tests, MBT scores will correlate with demographic variables such as age and education. Having a clear understanding of how the MBT functions in healthy individuals will directly inform how it may be used to detect associative memory deficits in the earliest disease stages. For example, by quantifying how MBT scores correlate with demographic factors, we can better distinguish patterns of scores representing disease-related impairment vs. expected performance based on age, educational attainment, or other factors. Next, we assessed criterion validity by testing whether MBT scores differ between healthy and preclinical groups and evaluating how MBT scores relate to continuous A(VN) biomarkers cross-sectionally. We build substantially upon existing evidence of the MBT’s discriminability by testing whether MBT scores add unique variance in discriminating healthy and preclinical groups (AD continuum and non-AD pathologic change) over and above demographic and composite cognitive scores from standard neuropsychological tests. We then used data collected at 2-year follow-up to conduct a novel evaluation of whether MBT scores can predict incident MCI in individuals with and without biomarker-confirmed preclinical disease. We hypothesize that the MBT will exhibit superior sensitivity and predictive utility compared to standard neuropsychological tests. That is, since the MBT is designed to specifically measure associative memory, which is known to exhibit declines prior to more domain-general impairments, we expect MBT scores to more sensitively discriminate between healthy and preclinical groups at baseline and better predict conversion to MCI compared to standard cognitive scores.

Methods

Transparency and Openness

In the following sections, we describe how we determined our sample, all data exclusions, and all measures from the study used in the preparation of this manuscript. This study was approved by the institutional review board. This study was not preregistered.

Participants

Data were drawn from an ongoing observational cohort study of community-dwelling adults (N = 165), which aims to evaluate changes in brain white matter integrity in aging and preclinical AD. For the present study, we used data from visits that had completed data collection at the time of analysis, which were the baseline and 2-year follow-up visits that were conducted between 2013 and 2022. Figure 1 presents the flow of procedures and participants involved in the current study. All participants provided written informed consent at enrollment and upon returning for the follow-up visit. Participants completed amyloid PET at baseline only and brain MRI, clinical assessments, and neuropsychological testing at both baseline and follow-up. To be eligible, participants had no MRI/PET contraindications, no history of severe/unstable conditions that affect cognition (e.g., stroke, seizures, serious mental illness, current substance abuse), spoke English as their first/primary language, and had no self-reported cognitive difficulties exceeding age-related complaints. To increase experimental rigor beyond subjective self-reports of cognition, we retroactively excluded any participants from this analysis if they exhibited objective evidence of cognitive impairment on the Montreal Cognitive Assessment (MoCA; Nasreddine et al., 2005), indicated by an age- and education-adjusted score of z ≤ −1.0 (Rossetti et al., 2011; n = 3). The absence of self-reported and objective cognitive impairments is in line with the field’s standard definitions of the cognitively unimpaired stage of preclinical AD (Dubois et al., 2016; Jack et al., 2018). Participants were also excluded from analyses if they had incidental MRI findings (n = 2) or incomplete neuroimaging or neuropsychological test data (n = 11).

Figure 1.

Flow of study procedures and participants.

The resulting baseline sample of N=149 were adults ages 45–85 (M_age = 67.15 years, SD = 9.53) who were mostly female (69.8%; 104 female sex), White (91.9%; racial identity: 137 White, 11 Black/African American, 1 Asian), non-Hispanic ethnicity (2 indicated White/Hispanic), and had a college education on average (M_edu = 16.09 years, SD = 2.56, range: 8 – 25 years). This study collected data on self-reported biological sex only, and did not collect information about gender identity, sexual orientation, or nativity/immigration history. Participants were categorized into preclinical and healthy groups, as described below, and there were minimal group differences in medical history (cf. Supplementary Table S1).

All study procedures besides amyloid PET were repeated at the 2-year follow-up visit, with the addition of the administration of the Clinical Dementia Rating Scale (CDR; Morris, 1993) to ascertain and stage incident cognitive impairment. Of the 149 baseline participants included in this study, 132 had the complete neuropsychological testing and CDR ratings needed for the planned longitudinal analyses (see Figure 1). The CDR was administered and scored by study personnel who had undergone training to be certified raters (i.e., clinician investigator, clinician post-doctoral fellow, or study coordinator under the supervision of a clinician). The rater conducted this semi-structured interview with a knowledgeable informant and the participant that queries the participant’s cognitive and functional status. Using the standardized scoring criteria, the rater then assigned scores for the six domains of function: memory, orientation, judgment/problem solving, community affairs, home and hobbies, and personal care. These 6 scores were entered into the open-source scoring algorithm to generate a CDR Global Score, with 0 indicating normal cognition/function and 0.5 indicating incident MCI.

Neuroimaging Data Acquisition and Analysis

Acquisition: MRI and PET

Detailed information about MRI and PET scan parameters and analysis were reported previously in greater detail (Benitez et al., 2022). MRI scans were conducted on a 3T Prisma^fit MRI system (Siemens Healthineers, Erlangen, Germany). Images utilized in this study were acquired with the following sequences: (1) T1-weighted MPRAGE with TR/TI/TE = 2300/900/2.26 ms, FOV = 256 × 256 mm², parallel imaging factor = 2, 1mm³ isotropic voxels; (2) T2-weighted FLAIR with TR/TI/TE = 9000/2500/92.0 ms, FOV = 220 × 220 mm², parallel imaging factor = 2, voxel size 0.9 × 0.9 × 2 mm³.

PET scans were obtained as a 3D acquisition on a Siemens Biograph mCT Flow PET/CT scanner approximately 50 min following the injection of 370 MBq (10 mCi) of ¹⁸F-florbetapir (Amyvid™). Images were acquired with a 128 × 128 matrix (zoom × 2) and were reconstructed using iterative or row action maximization likelihood algorithms. No participants reported adverse reactions to florbetapir.

Analysis: Discrete Measures to identify Preclinical Disease

Participants were categorized as having preclinical disease (n = 55) based on: 1) amyloid positivity (A+; Doraiswamy et al., 2012) on florbetapir PET scan, determined by a board-certified radiologist and trained Amyvid reader, 2) evidence of neurodegenerative changes (N+) based on mean bi-hemispheric medial temporal lobe atrophy rating (Scheltens et al., 1992) of ≥ 1.5 (Pereira et al., 2014; N+), and/or 3) evidence of significant cerebrovascular disease (V+; Prins et al., 2004) based on Fazekas rating (Fazekas et al., 1987) ≥ 2 in periventricular or deep WM, both rated by a board-certified neuroradiologist. The preclinical group was further categorized into AD continuum (A+/V±/N±) and non-AD pathologic change (A-/V+/N-, A-/V-/N+, or A-/V+/N+; Figure 2).

Figure 2.

AD Biomarkers within the Preclinical Group. Values within the Venn diagram indicate the number and percent of the 55 total preclinical participants with each pathology. The preclinical group is further divided into those who are amyloid PET positive, regardless of other pathology (AD continuum, in purple) and those with MTA or WMH but are PET negative (non-AD pathologic change, in blue). MTA: medial temporal lobe atrophy; WMH: white matter hyperintensities.

Analysis: Continuous Measures to Assess Criterion Validity

The three neuroimaging biomarkers were also used in continuous form for correlational and group difference analyses, calculated in the same manner as we recently reported (Benitez et al., 2022) and are briefly summarized here.

Amyloid PET.

We quantified brain cortical amyloid burden by replicating published approaches that calculate the ratio of cortical to cerebellar amyloid, cited here. First, standard uptake values (SUV) were calculated for six cortical regions used widely in the literature (Clark et al., 2011; Clark et al., 2012; Joshi et al., 2015) due to their anatomical relevance in AD and sensitivity to florbetapir uptake: anterior cingulate, posterior cingulate, parietal lobe, medial orbito-frontal lobe, middle temporal lobe, and precuneus. Cortical SUV values were then normalized to the mean SUV of the cerebellum (reference region) to derive the SUV ratio (SUVr; Clark et al., 2011). The continuous metric of amyloidosis was the non-weighted mean SUVr (mSUVr) across all six cortical regions.

Neurodegeneration.

We selected hippocampal atrophy and the AD signature, an indicator of cortical thinning in AD-relevant regions (Wang et al., 2015; 2016), as measures of neurodegeneration as both have demonstrated strong diagnostic accuracy across the AD continuum (Allison et al., 2019). To facilitate interpretation and analysis, we computed these volumes as follows. Normative volumetric z-scores were generated from FreeSurfer v6.0-segmented T1-weighted images that were submitted to the NOMIS software (Potvin et al., 2021) which uses a normative database of ~7,000 cognitively intact individuals and adjusts for age, sex, intracranial volume, and image quality. From these we calculated 1) hippocampal atrophy: mean of bilateral hippocampal volume segment z-scores and 2) AD signature volume: mean of bilateral cortical thickness z-scores for entorhinal, inferior temporal, middle temporal, inferior parietal, fusiform, and precuneus (Wang et al., 2015).

WMH lesion volume.

The lesion growth algorithm (Schmidt et al., 2012) within the LST toolbox 3.0 was used to identify WMH and calculate lesion volume from co-registered T1-weighted and T2-FLAIR images. Total WMH lesion volume was quantified in ml.

Neuropsychological Measures

Memory Binding Test.

The MBT measures associative memory by testing participants’ ability to bind words to category cues and resist semantic proactive interference (Buschke, 2014). It consists of two visually-presented 16-item word lists that correspond to 16 category cues (e.g., cue: Color; list 1 word: Brown, list 2 word: Yellow). Administration involves 1) a learning phase followed immediately by immediate paired recall and free recall (~10–15 minutes), 2) a 30-minute delay, and 3) delayed free recall and delayed cued paired recall (~5 minutes). Learning condition: First, the MBT employs a controlled learning phase in which participants were presented with the 16 words from list 1, shown printed on paper 4 at a time, and asked to identify which of the words match a verbally presented category cue (e.g., cue: Country; response options: Brown, Spain, Paul, Captain). Encoding was then tested by prompting participants to recall each of the 16 words in response to the verbally presented category cues (e.g., cue: Country; correct response: Spain). The learning and encoding phases were then repeated with new words from List 2 (e.g., cue: Country; new response options: Yellow, Canada, Harry, General). Immediate paired recall condition: To assess semantic binding, participants were then asked to recall the two items from both lists when verbally presented with each category cue (e.g., cue: Country; correct responses: Spain and Canada;). Immediate free recall condition: Next, without delay, they were asked to recall as many words as they could from both lists without cueing. 30-min delay: During the delay, participants completed non-verbal tests of processing speed and working memory to minimize task interference. Delayed free recall condition: After the delay, participants were first asked to recall all the words without cueing. Delayed paired recall condition: Last, they were asked to recall the two words belonging to each category when verbally presented with each cue (e.g., cue: Color, correct responses: Brown and Yellow).

We included nine MBT scores that are most consistently evaluated in previous studies of the MBT: eight arithmetical scores (i.e., sum of correct words or word pairs from a trial) and one derived score of semantic proactive interference (SPI; Buschke et al., 2017), which we included given that vulnerability to this type of interference relates to AD biomarkers (Crocco et al., 2018; Loewenstein et al., 2016). These scores were defined in accordance with previously published work (Gramunt et al., 2016); descriptions, calculations, and ranges are detailed in Table 1.

Table 1.

The Memory Binding Test Procedure and Score Descriptions.

MBT Variables	Description/Calculation	Range
Learning Condition	Participants learn List 1 and then are asked to recall them in response to cue (Cued Recall List 1). They then learn List 2 and are asked to recall them in response to cue (Cued Recall List 2).
Semantic Proactive Interference (SPI)	(Cued Recall List 2/Cued Recall List 1)*100	%a
Immediate Paired Recall Condition	Participants asked to recall words from both lists in response to cue.
Paired Recall Pairs (Immed. PRP)	Total number of correct semantic word pairs	0–16
Total Paired Recall (Immed. TPR)	Total number of correct words from either list	0–32
Immediate Free Recall Condition	Participants asked to recall words from both lists with no cue.
Free Recall Pairs (Immed. FRP)	Total number of correct semantic word pairs	0–16
Total Free Recall (Immed. TFR)	Total number of correct words from either list	0–32
Delayed Free Recall Condition	After 30-min. delay, participants asked to recall words from both lists with no cue.
Free Recall Pairs (Del. FRP)	Total number of correct semantic word pairs	0–16
Total Free Recall (Del. TFR)	Total number of correct words from either list	0–32
Delayed Paired Recall Condition	Participants asked to recall words from both lists in response to cue.
Paired Recall Pairs (Del. PRP)	Total number of correct semantic word pairs	0–16
Total Paired Recall (Del. TPR)	Total number of correct words from either list	0–32

Note.

^aLower percentage (<100%) is indicative of greater proactive interference from List 1 on List 2 performance

Standard Neuropsychological Battery.

Participants were administered version 3 of the Alzheimer Disease Centers’ Neuropsychological Test Battery in the Uniform Data Set (UDS 3.0; see Besser et al., 2018 and Weintraub, 2018 for detailed description). For the purposes of data reduction and to control for demographic influences on these criterion variables, raw scores were converted to demographically-adjusted (i.e., age, sex, years of education) factor z-scores based on published normative equations (Kiselica et al., 2020). The factor scores were calculated from: 1) memory (Craft story immediate and delayed recalls, verbatim scores), 2) visual (Benson figure copy and delayed recall, total scores), 3) language (animal fluency, vegetable fluency, multilingual naming test, total correct), 4) attention (number span forward and backward total scores), 5) speed/executive (time to completion on Trail Making Tests A and B), 6) and a general factor score calculated from these 5 factor scores.

Statistical Analysis

All analyses were conducted in SPSS v28 and RStudio v2023.03.0+386. All variables were assessed for normality and nonparametric tests were used in analyses involving non-normally distributed variables. Group differences in demographics, biomarkers, and neuropsychological factor scores were tested using t-tests or Mann-Whitney U tests for continuous variables and chi square tests for categorical variables.

We assessed construct validity using data from the healthy group. Spearman rank correlations were conducted between MBT scores, age, and education. Mann-Whitney U tests were conducted to assess whether MBT performance differed by sex and race, for which we dichotomized participants into White (n = 137) and non-White (n = 12) since one participant identified as neither White nor Black. Spearman rank correlations assessed the associations between MBT performance and neuropsychological factor scores. Differences in the resulting correlations between each pair of factor scores (dependent groups) were tested using the ‘cocor’ R package, which computes a Fisher’s Z statistic and p-value.

We assessed criterion validity using the full dataset. Age was covaried in these analyses given that the preclinical group was significantly older than the healthy group and age was associated with higher amyloid (rho = 0.31, p < .001), WMH (rho = 0.63, p < .001), and non-significant but directionally lower hippocampal (rho = −0.10, p = .213) and AD signature (rho = −0.11, p = .197) volumes. Group differences in MBT performance and neuropsychological scores covarying age were tested using non-parametric ANCOVAs (ranked Quade’s Test) with Bonferroni correction for post hoc pairwise comparisons of significant effects. Partial Spearman correlations between MBT scores and each of the continuous AD biomarkers, covarying age, were conducted in each group. Logistic regressions evaluated prediction of group using age, neuropsychological factor scores, and MBT performance as predictors. Models were compared using likelihood ratio tests (X²).

For longitudinal analyses, we first tested whether the subset of participants who did and did not have complete follow-up data differed in age or years of education (Mann-Whitney U tests) and sex or race (chi square tests). A chi square test was conducted to examine whether the proportion of participants with follow-up data differed between the healthy group and full preclinical group. In the subset of participants who returned for 2-year follow-up, we used a Mann-Whitney U test to assess differences in follow-up interval between the healthy group and full preclinical group. Test-retest reliability of each MBT score was assessed by computing the intraclass correlation coefficient (ICC). ICCs were calculated using the ‘icc’ function from the ‘IRR’ R package specifying the measurement of absolute agreement with two-way random effects models and single ratings (per observation) in which both subjects and raters are considered as random effects (Koo & Li, 2016). This is appropriate given that due to the longitudinal nature of this study, the personnel administering the MBT varied across participants and over visits. ICC values range from 0–1 with higher scores indicating higher test-retest reliability. We then tested for group differences in MBT score changes over time by conducting repeated measures ANCOVAs with time as a within-subjects factor, group as a between-subjects factor, and age and follow-up interval as covariates, with Bonferroni correction for post hoc pairwise comparisons of significant effects.

A chi square test was used to evaluate differences in the proportion of participants who converted to MCI between the healthy and full preclinical groups. Last, logistic regressions evaluated prediction of conversion to MCI based on baseline biomarkers, MBT scores, and standard neuropsychological test scores, with age at baseline and follow-up interval included as covariates. Variance inflation factor (VIF) was used to identify problematic levels of multicollinearity (VIF ≥ 5), in which cases separate regression models were used for each predictor of interest. A single logistic regression model was used to test all four AD biomarkers (amyloid, hippocampal volume, WMH lesion volume, and AD signature volume) as predictors since these variables were not highly collinear (VIF’s < 1.04). Separate models were used for each of the 9 MBT scores (VIF’s > 10.17) and each of the 5 standard neuropsychological factor scores as these scores exhibited high multicollinearity.

The data used in this study are not publicly available as participants were not asked to consent to data sharing. Analysis code for this study may be requested from the corresponding author.

Results

Table 2 summarizes demographics, characteristics, and tests of group differences. The full preclinical group, and both preclinical subgroups individually, were older than the healthy group, but the two preclinical subgroups did not differ significantly in age. There were no differences across groups in years of education or sex. The overall sample was predominantly White (8%), and all non-White participants (n = 12) were in the healthy group. As expected, compared to the healthy group, the preclinical group had significantly different continuous biomarker variables (i.e., higher amyloid, lower hippocampal volume, and greater WMH) except AD signature volume. The two preclinical subgroups also differed such that amyloid was higher in the AD continuum than both the non-AD pathologic change and healthy groups (which did not differ from each other) and WMH burden was higher in the non-AD pathologic change than AD continuum group, both of which had higher levels than the healthy group. As is expected when sampling cognitive performance from the general population (Binder et al., 2009), we observed variability in scores on the standard neuropsychological tests. The “normative” group from which the factor scores were derived generally performed better than our community-based sample that had average z-scores below 0. Nonetheless, the pattern of performance was consistent with group allocation such that the median performance for the healthy group was within 1 SD of the normative mean in all domains, whereas the median performance for the full preclinical disease group (and both preclinical subgroups separately) was < −1 SD in memory, speed/executive, and the general composite score.

Table 2.

Sample demographics and characteristics.

	Healthy (n = 94)		pAD (n = 55)		Group Differences (Healthy vs. Full Preclinical)		AD Continuum (n = 25)		Non-AD Pathologic Change (n = 30)		Group Differences (Healthy vs. AD Continuum vs. Non-AD)
	Mean/no.	SD/%	Mean/no.	SD/%	Stat	p-value	Mean/no.	SD/%	Mean/no.	SD/%	F(2,146) a	p-value
Age (years)	64.2	8.7	72.3	8.7	U = 3913.5	< .001 ***	70.9	9.9	73.4	7.5	15.54	< .001 ***
Education (years)	16.2	2.6	15.9	2.5	U = 2421	.508	15.9	2.5	15.8	2.5	0.37	.689
Sex	70	74.5%	34	61.8%	X2 = 2.07	.150	16	64.0%	18	60.0%	X2 = 2.74	.254
Race					X2 = 7.64	.022 *					X2 = 7.64	.022 *
White	82	87.2%	55	100.0%			25	100.0%	30	100.0%
Black or African American	11	11.7%	0	0.0%			0	0.0%	0	0.0%
Asian	1	1.1%	0	0.0%			0	0.0%	0	0.0%
Biomarkers: Clinician Ratings	no.	%	no.	%			no.	%	no.	%
Amyloid PET positive	-	-	25	45.5%			25	100.0%	-	-
Medial Temporal Lobe Atrophy
0	67	71.3%	16	29.1%			11	44.0%	5	16.7%
0.5	8	85.0%	9	16.4%			6	24.0%	3	10.0%
1	19	20.2%	17	30.9%			6	24.0%	11	36.7%
1.5	-	-	7	12.7%			0	0.0%	7	23.0%
2	-	-	5	9.1%			2	8.0%	3	3.0%
3	-	-	1	18.0%			0	0.0%	1	10.0%
White Matter Hyperintensities
Periventricular WM 0	28	29.8%	5	9.1%			5	20.0%	0	0.0%
1	66	70.2%	24	43.6%			13	52.0%	11	36.7%
2	-	-	20	36.4%			6	24.0%	14	46.7%
3	-	-	6	10.9%			1	4.0%	5	16.7%
Deep WM 0	31	33.0%	4	7.3%			4	16.0%	0	0.0%
1	63	67.0%	30	54.5%			16	64.0%	14	46.7%
2	-	-	17	30.9%			4	16.0%	13	43.3%
3	-	-	4	7.3%			1	4.0%	3	10.0%
Biomarkers: Continuous	Mean	SD	Mean	SD	Stat	p-value	Mean	SD	Mean	SD	Stat	p-value
Amyloidosis (mSUVr)	1.17	0.1	1.34	0.3	U = 3377	.002 **	1.57	0.3	1.15	0.1	71.69	< .001 ***
Hippocampal volume (z-score)	0.29	0.8	−0.06	0.9	t(147) = −2.43	.016 *	−0.06	0.9	−0.06	0.9	2.94	0.056
AD signature volume (z-score)	0.05	0.7	−0.14	0.8	U = 2222	.154	−0.27	0.8	−0.04	0.8	1.82	0.166
WMH volume (ml)	1.91	2.0	8.32	9.1	t(147) = 6.55	< .001 ***	5.72	6.0	10.48	10.7	27.72	< .001 ***
Memory Binding Test Raw Scores	Median	IQR	Median	IQR	F(1,146) a	p-value	Median	IQR	Median	IQR	F(2,146) a	p-value
Learning Condition
SPI	94	13.3	93	18.8	0.55	.459	94	15.6	88	19.1	0.41	0.665
Word Pairs (max 16)
Immed. PRP	14	3.0	13	3.5	0.01	.919	13	3.5	13	4.0	0.30	0.744
Delayed PRP	14	3.0	13	3.5	0.14	.704	7	5.0	7	4.0	0.23	0.792
Immed. FRP	9	3.0	7	4.0	3.73	.055	7	4.0	7	4.3	2.06	0.131
Delayed FRP	9	3.8	7	4.0	2.62	.107	13	4.0	13	4.0	1.90	0.153
Total Recalls (max 32)
Immed. TPR	30	3.0	29	4.0	0.01	.905	29	3.5	29	4.0	0.39	0.676
Delayed TPR	30	3.0	29	4.5	0.07	.793	16	7.0	16	8.0	0.42	0.658
Immed. TFR	19	5.0	16	7.0	6.54	.011 *	16	7.5	15	8.0	3.37	0.037 *
Delayed TFR	19	6.0	15	7.5	2.22	.138	29	4.5	29	4.5	1.50	0.227
Neuropsychological Factor Z-Scores	Median	IQR	Median	IQR	F(1,146) a	p-value	Median	IQR	Median	IQR	F(2,146) a	p-value
Memory	−0.92	1.9	−1.17	1.6	0.04	.839	−1.17	1.5	−1.15	2.1	0.34	0.712
Visual	−0.55	1.6	−0.57	2.1	0.15	.697	−0.62	1.9	−0.53	2.4	0.17	0.845
Language	−0.90	2.1	−0.72	1.8	0.20	.656	−0.75	1.5	−0.66	2.0	0.46	0.631
Attention	−0.43	1.3	−0.18	1.5	0.36	.549	−0.69	1.6	−0.06	1.6	0.85	0.430
Speed/Executive	−0.36	1.9	−1.05	1.9	4.32	.039 *	−1.00	1.2	−1.12	2.3	2.31	0.103
General	−0.98	2.0	−1.04	1.6	0.15	.695	−1.04	1.7	−1.14	1.7	0.08	0.926

Note.

^aNon-parametric ANCOVA covarying for age.

^*p < .05

^**p < .01

^***p < .001.

Abbreviations: SPI = semantic proactive interference, PRP = paired recall pairs, FRP = free recall pairs, TPR = total paired recall, TFR = total free recall.

Construct Validity of the MBT in Healthy Adults

Cross-Sectional Associations between MBT and Demographic Variables.

We first examined how performance on the MBT varied by age, education, sex, and race (Tables 2 and 3). Older age was associated with poorer performance on all conditions. Scatterplots are shown for the four total recall scores in Figure 3. The strongest magnitude correlations were for delayed free recall scores (delayed FRP rho = −0.38 and delayed TFR rho = −0.40). Higher education was associated with better performance only on the paired recall condition (immediate PRP, immediate TPR, and delayed TPR) and with more modest effects (maximum rho = 0.24). MBT performance did not differ by sex (p’s > .119) or race (i.e., p’s > .052); note that the number of non-White participants (n = 12) was small which limits our ability to statistically test for differences by race.

Table 3.

Relationships between MBT scores, demographic variables, and factor scores derived from standard neuropsychological measures in the Healthy group (n = 94).

MBT Variables	Age	Edu	Memory	Visual	Language	Attention	Speed/ Executive	General
Learning Condition
SPI	−0.27*	0.18	0.18	0.22 *	0.14	0.04	−0.01	0.12
Word Pairs
Immed. PRP	−0.27**	0.24 *	0.38 ***	0.18	0.30 **	0.18	0.26 *	0.37 ***
Delayed PRP	−0.33**	0.18	0.42 ***	0.21 *	0.31 **	0.14	0.28 **	0.40 ***
Immed. FRP	−0.28**	0.20	0.43 ***	0.20	0.37 ***	0.19	0.28 **	0.44 ***
Delayed FRP	−0.38***	0.14	0.36 ***	0.10	0.27 **	0.11	0.26 *	0.35 ***
Total Recalls
Immed. TPR	−0.26**	0.24 *	0.37 ***	0.22 *	0.31 **	0.16	0.25 *	0.38 ***
Delayed TPR	−0.33**	0.21 *	0.38 ***	0.22 *	0.30 **	0.10	0.25 *	0.37 ***
Immed. TFR	−0.32**	0.17	0.48 ***	0.24 *	0.36 ***	0.15	0.26 *	0.44 ***
Delayed TFR	−0.40***	0.12	0.36 ***	0.14	0.23 *	0.10	0.26 *	0.33 **

Note. Reported values are Spearman rank correlation coefficients.

^*p < .05

^**p < .01

^***p < .001.

Abbreviations: SPI = semantic proactive interference, PRP = paired recall pairs, FRP = free recall pairs, TPR = total paired recall, TFR = total free recall.

Figure 3.

Associations between four primary MBT scores (y-axes) and age (x-axes) in healthy adults. Scatterplots depict the significant negative relationships between performance on immediate and delayed, paired and free recall conditions of the MBT. Each variable has a maximum score of 32 words recalled, indicated by the horizontal gray line. Points represent individual participants’ scores.

Cross-Sectional Associations between MBT and Standard Neuropsychological Tests.

We next evaluated how well performance on the MBT corresponded with composite scores derived from standard neuropsychological measures (Table 3). Descriptively, MBT performance was most consistently and strongly related to memory (8/9 MBT scores, average rho = .37) and the general composite factor scores (8/9, average rho = .36). The majority (8/9) of MBT scores were also associated with the language and speed/executive factor scores, but with weaker correlations (language: average rho = .29, speed/executive: average rho = .23). Fewer MBT scores were associated with visual factor scores (5/9, average rho = 0.19) and none with attention factor scores. Although age was not covaried given that the factor scores are age-adjusted, this pattern of results remained when it was covaried. We next tested for differences in the magnitude of the correlations between MBT scores and memory versus the other factor scores. MBT performance was more strongly related to memory than attention factor scores for delayed PRP (Z = 2.16, p = .031), immediate TFR (Z = 2.61, p = .009), delayed TFR (Z = 1.96, p = .050), and delayed TPR (Z = 2.12, p = .034). There was also a stronger association with memory than visual factor scores for immediate FRP (Z = 2.02, p = .043), delayed FRP (Z = 2.21, p = .027), and immediate TFR (Z = 2.17, p = .030). The association between MBT scores and memory did not differ significantly from those for language (p’s > .260), speed/executive (p’s > .072), or the general composite factor scores (p’s > .652).

Criterion Validity of the MBT

Cross-sectional Analyses

Group Differences in MBT and Standard Neuropsychological Tests.

Tests of group differences are reported in Table 2. The preclinical group performed worse than the healthy group on MBT immediate TFR (Figure 4A), and marginally worse on immediate and delayed FRP, when controlling for age. Of note, when age was not covaried, the groups differed on 6 of the 9 MBT scores (immediate PRP, immediate and delayed FRP, delayed TPR, and immediate and delayed TFR), highlighting the susceptibility of the MBT to age-related effects. When testing for differences in MBT performance among the healthy and two preclinical subgroups while covarying age, the only significant effect was for immediate TFR (F(2,146) = 3.37, p = .037). The AD continuum group performed worse than the healthy group (p_adj = .042), the non-AD pathologic change group performed marginally worse than the healthy group (p_adj = .111), and there was no difference between the AD continuum and non-AD pathologic change groups (Figure 4B). For the standard neuropsychological measures, the only significant difference was lower speed/executive scores in the preclinical group than the healthy group, and there were no significant differences between the preclinical subgroups.

Figure 4.

Group differences in MBT performance. A) Boxplots indicate median scores (y-axis) and are superimposed on violin plots showing the distribution of scores on each MBT variable (x-axis) for the healthy group (green, left of pair) and preclinical disease group (purple, right of pair). The first 4 MBT variables (word pairs) have a maximum score of 16, and the last four variables (total recalls) have a maximum score of 32. B) Boxplots show the main effect of group (x-axis), and results of pairwise comparisons, on immediate total free recall (y-axis) for the healthy (green, left) and two specific biomarker preclinical subgroups: AD continuum (middle, purple) and non-AD pathologic change (right, blue). Points represent individual participants’ scores. This variable has a maximum score of 32 words. ^†p ≤ .11, *p ≤ .05. **p ≤ .01, ***p ≤ .001.

Cross-Sectional Associations between MBT and Biomarkers.

We assessed the associations between MBT performance and biomarkers, covarying age, in the healthy group, full preclinical group, and preclinical subgroups separately (Table 4). MBT performance was not associated with mSUVr in any group. MBT scores were not related to hippocampal volume in the healthy or full preclinical group, but lower hippocampal volume was moderately associated with worse delayed TFR in the AD continuum subgroup (rho = 0.42). Lower AD signature volume was associated with poorer MBT performance on all scores except SPI in the healthy group only, with small-moderate effects (rho’s = 0.21 – 0.31). Greater WMH lesion volume was associated with lower scores on immediate and delayed TFR and delayed FRP in the full preclinical group and both subgroups, with moderate to strong effects (rho’s = −0.39 – −0.52).

Table 4.

Spearman Rank Partial Correlations between MBT scores and AD biomarkers, covarying age, in the healthy group (n = 94), full preclinical disease group (n = 55), and the preclinical biomarker subgroups: AD-Continuum (n = 25) and non-AD pathologic change (n = 30).

	Amyloid Burden				Hippocampal Volume				AD Signature Volume				WMH Volume
	Healthy	Preclinical			Healthy	Preclinical			Healthy	Preclinical			Healthy	Preclinical
		Full	AD	Non-AD		Full	AD	Non-AD		Full	AD	Non-AD		Full	AD	Non-AD
Learning Condition
SPI	0.09	−0.07	0.05	−0.25	−0.04	−0.21	−0.35	−0.07	0.09	−0.12	−0.07	−0.09	0.14	−0.18	−0.14	−0.10
Word Pairs
Immed. PRP	0.00	−0.20	−0.15	−0.17	0.04	−0.12	−0.09	−0.12	0.31 **	−0.12	−0.02	−0.2	0.04	−0.12	0.05	−0.24
Delayed PRP	0.05	−0.13	−0.11	−0.05	0.02	−0.17	−0.18	−0.17	0.23 *	−0.15	−0.07	−0.2	0.05	−0.20	0.01	−0.34
Immed. FRP	0.06	−0.16	−0.13	−0.06	0.01	0.03	−0.06	0.15	0.24 *	−0.01	0.06	−0.08	−0.06	−0.39**	−0.33	−0.49**
Delayed FRP	0.03	−0.09	0.10	0.02	0.02	0.18	0.32	0.10	0.26 *	−0.15	0.01	−0.27	−0.12	−0.41**	−0.50*	−0.41*
Total Recalls
Immed. TPR	0.01	−0.20	−0.13	−0.16	0.03	−0.15	−0.13	−0.14	0.30 **	−0.17	−0.05	−0.28	0.03	−0.11	0.05	−0.24
Delayed TPR	0.07	−0.16	−0.11	−0.04	0.04	−0.18	−0.17	−0.20	0.24 *	−0.15	−0.05	−0.23	0.05	−0.15	0.03	−0.30
Immed. TFR	0.08	−0.12	−0.13	−0.07	0.02	0.15	0.13	0.15	0.21 *	−0.04	0.03	−0.11	−0.09	−0.47***	−0.52**	−0.48**
Delayed TFR	0.02	−0.09	0.13	−0.05	0.04	0.19	0.42 *	0.03	0.25 *	−0.08	0.09	−0.23	−0.17	−0.39**	−0.45*	−0.42*

Note. Correlations are presented for the preclinical group overall (Full) and for the AD-continuum subgroup (AD) and non-AD pathologic change subgroup (Non-AD). Age is covaried in all correlations. Amyloid burden refers to mSUVr, Hippocampal and AD Signature volumes are normed z-scores, WMH volume refers to lesion volume in ml.

^*p ≤ .05.

^**p ≤ .01,

^***p ≤ .001.

Abbreviations: SPI = semantic proactive interference, PRP = paired recall pairs, FRP = free recall pairs, TPR = total paired recall, TFR = total free recall.

Discriminability of the MBT.

We examined whether the MBT could discriminate between groups better than age and the standard neuropsychological measures (Table 5). Age was a significant predictor of being in the healthy vs. full preclinical group, explaining 23% of the variance and correctly classifying 70.5% of cases (OR = 1.12). Next, we included MBT immediate TFR, the only score remaining different between groups when age was covaried. Both higher age (OR = 1.10) and poorer MBT performance were associated with a higher likelihood of preclinical status (OR = 0.90), with improved model fit over age alone (X² = 6.79, p = .009). This model explained 29% of the variance and correctly classified 71.8% of cases. The addition of the neuropsychological factor scores did not improve model fit (X² = 7.59, p = .267). None of the factor scores were significant predictors of preclinical status, while higher age (OR = 1.09) and lower MBT performance (OR = 0.89) remained significant.

Table 5.

Results of Logistic Regression Models Predicting Group Status.

		B	SE	Wald X2	p-value	Odds Ratio	95% CI	AUC
	Outcome: Healthy vs. Preclinical
Model 1.	Age	0.110	0.024	21.46	< .001 ***	1.12	1.07 – 1.17	0.757
Model 2.	Age	0.097	0.025	15.57	< .001 ***	1.10	1.05 – 1.16	0.757
	MBT immediate TFR	−0.105	0.042	6.42	.011 *	0.90	0.83 – 0.98	0.701
Model 3.	Age	0.083	0.033	6.40	.011 *	1.09	1.02 – 1.16	0.757
	MBT total immediate free recall	−0.120	0.050	5.85	.016 *	0.89	0.80 – 0.98	0.701
	Memory factor score	0.041	3.173	0.00	.990	1.04	0.00 – 522.77	0.524
	Visual factor score	−0.064	1.486	0.00	.966	0.94	0.05 – 17.26	0.526
	Language factor score	0.118	4.510	0.00	.979	1.13	0.00 – 7.77 e3	0.464
	Attention factor score	0.124	4.004	0.00	.975	1.13	0.00 – 2.90 e3	0.442
	Speed/Executive factor score	−0.471	4.375	0.01	.914	0.62	0.00 – 3.30 e3	0.653
	General factor score	0.357	11.391	0.00	.975	1.43	0.00 – 7.10 e9	0.537
	Outcome: Healthy vs. AD Continuum
Model 1.	Age	0.089	0.029	9.42	.002 **	1.09	1.03 – 1.16	0.710
Model 2.	Age	0.072	0.030	5.67	.017 *	1.08	1.01 – 1.14	0.710
	MBT immediate TFR	−0.112	0.052	4.61	.032 *	0.89	0.81 – 0.99	0.707
Model 3.	Age	0.065	0.040	2.70	.100	1.07	0.99 – 1.15	0.710
	MBT immediate TFR	−0.157	0.065	5.77	.016 *	0.86	0.75 – 0.97	0.707
	Memory factor score	0.090	3.964	0.00	.982	1.09	0.00 – 2.59 e3	0.554
	Visual factor score	−0.029	1.858	0.00	.988	0.97	0.03 – 37.04	0.506
	Language factor score	0.492	5.653	0.01	.931	1.64	0.00 – 1.06 e5	0.428
	Attention factor score	0.068	5.031	0.00	.989	1.07	0.00 – 2.05 e4	0.498
	Speed/Executive factor score	−0.083	5.466	0.00	.988	0.92	0.00 – 4.13 e4	0.625
	General factor score	−0.187	14.272	0.00	.990	0.83	0.00 – 1.17 e12	0.530
	Outcome: Healthy vs. Non-AD Pathologic Change
Model 1.	Age	0.145	0.034	18.09	< .001 ***	1.16	1.08 – 1.24	0.796
Model 2.	Age	0.133	0.035	14.71	< .001 ***	1.14	1.07 – 1.22	0.796
	MBT immediate TFR	−0.079	0.050	2.57	.109	0.92	0.84 – 1.02	0.696

Note.

^*p ≤ .05.

^**p ≤ .01

^***p ≤ .001.

Belonging to the non-AD pathologic change vs. healthy groups was predicted by age only (OR = 1.14), with a trend-level effect of MBT total immediate free recall (OR = 0.92). In contrast, AD continuum status was predicted by both age (OR = 1.08) and MBT total immediate free recall (OR = 0.89), with improved fit over age alone (X² = 4.85, p = .028). Again, the addition of the neuropsychological factor scores did not improve fit (X² = 5.20, p = .517), and the only variable remaining significant predictor of AD continuum status was MBT total immediate free recall (OR = 0.86).

Longitudinal Analyses

The participants who did (n = 132) and did not have complete follow-up data (n = 17) did not differ significantly in age (U = 1070.5, p = .761), years of education (U = 1106.0, p = .924), sex (X² = 0.84, p = .359), or race (X² = 1.31, p = .519). The proportion of participants with follow-up data did not differ between the healthy group (83/94 = 88.3%) and full preclinical group (49/55 = 89.1%; X² = 0.22, p = .883). The median follow-up interval was 2.09 [1.91, 2.28] years and this did not differ significantly between the healthy group (median = 2.10 [1.89, 2.28]) and preclinical group (median = 2.04 [1.94, 2.25]; U = 2115.5, p = .701).

Change in MBT Performance Over Time.

We tested for group differences in change in each of the 9 MBT scores from baseline to follow-up controlling for baseline age and follow-up interval (Table 6). When comparing change in MBT scores over time between the healthy group and full preclinical group, there were no significant Group X Time interactions. When comparing the healthy group and two preclinical subgroups, there was one Group X Time interaction for SPI (F(2,127) = 3.25, p = .042). Pairwise comparisons indicated that the AD-continuum group exhibited a significant within-groups decline in SPI (F(1,127) = 5.20, p = .024) whereas the healthy and non-AD pathologic change groups did not (p’s > .228); however, there were no significant group differences in the magnitude of change in SPI (p’s > .358).

Table 6.

Tests of Within- and Between-Group Differences and Interaction Effects in Participants with Complete Follow-Up Data (N = 132).

Panel A. Comparison of Healthy and Full Preclinical Group
	Healthy (n = 83)		pAD (n = 49)				A. Between-Group Differences at Baseline		B. Between-Group Differences at Follow-Up		C. Within-Group Differences over Time				D. Group x Time Interactions
MBT (Raw Scores)	Baseline	Follow-Up	Baseline	Follow-Up							Healthy		pAD
	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)			F	p	F	p	F	p	F	p	F	p
Learning Condition
SPI	93.75 (13.3)	93.75 (13.4)	92.86 (18.8)	93.33 (18.8)			0.9	.344	1.11	.294	1.53	.219	2.15	.145	3.42	.067
Word Pairs
Immed. PRP	14 (3)	14 (3)	13 (3.5)	13 (4)			0.01	.932	1.66	.199	1.33	.251	0.63	.428	1.65	.202
Delayed PRP	14 (3)	15 (4)	13 (3.5)	13 (4)			0.23	.635	1.49	.225	0.59	.443	0.44	.506	0.92	.338
Immed. FRP	9 (3)	9 (4)	7 (4)	8 (4)			5.57	.020 *	3.06	.083	0.55	.461	0.03	.869	0.09	.763
Delayed FPR	9 (3.8)	10 (4)	7 (4)	8 (3)			5.08	.026 *	2.72	.101	0.02	.901	0.63	.427	0.29	.589
Total Recalls
Immed. TPR	30 (3)	30 (3.5)	29 (4)	29 (4)			0.17	.680	1.49	.225	1.59	.210	1.73	.191	3.05	.083
Delayed TPR	30 (3)	31 (4)	29 (4.5)	29 (4)			0.1	.752	1.15	.285	0.24	.624	0.64	.426	0.81	.369
Immed. TFR	19 (5)	20 (7)	16 (7)	16 (8)			10.21	.002 **	5.77	.018 *	1.06	.306	0.18	.672	0.07	.788
Delayed TFR	19 (6)	20 (7)	15 (7.5)	17 (8)			4.91	.029 *	2.99	.086	0.04	.852	0.12	.733	0.14	.710
Panel B. Comparison of Healthy and Preclinical Subgroups							A. Between-Group Differences at Baseline		B. Between-Group Differences at Follow-Up		C. Within-Group Differences over Time				D. Group x Time Interactions
	Healthy (n = 83)		AD Continuum (n = 21)		Non-AD Pathologic Change (n = 28)
MBT (Raw Scores)	Baseline	Follow-Up	Baseline	Follow-Up	Baseline	Follow-Up					AD Continuum		Non-AD Pathologic Change
	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	F	p	F	p	F	p	F	p	F	p
Learning Condition
SPI	93.75 (13.3)	93.75 (13.4)	93.75 (12.5)	86.67 (25)	87.5 (18.8)	93.75 (18.8)	0.68	.509	1.25	.289	5.20	.024 *	0.00	.979	3.25	.042 *
Word Pairs
Immed. PRP	14 (3)	14 (3)	13 (3)	12 (4)	13 (3.5)	14 (3.3)	0.16	.851	2.39	.096	2.4	.137	0.06	.804	1.67	.193
Delayed PRP	14 (3)	15 (4)	7 (5)	8 (5)	7 (3.8)	7.5 (4.3)	0.19	.825	1.45	.239	1.4	.239	0.02	.882	0.96	.387
Immed. FRP	9 (3)	9 (4)	7 (3)	7 (3)	7 (4)	8 (3.3)	3.37	.038 *	1.83	.164	0.00	.996	0.05	.828	0.06	.946
Delayed FPR	9 (3.8)	10 (4)	13 (3)	12 (4)	13 (3.8)	14 (3)	3.53	.032 *	2.21	.114	0.32	.575	0.33	.567	0.15	.864
Total Recalls
Immed. TPR	30 (3)	30 (3.5)	29 (3)	28 (4)	29 (4)	30 (4)	0.21	.812	2.37	.097	3.5	.064	0.02	.902	2.42	.093
Delayed TPR	30 (3)	31 (4)	16 (7)	17 (10)	16 (7.3)	16 (6.5)	0.09	.915	1.34	.266	1.86	.175	0.02	.897	1.03	.360
Immed. TFR	19 (5)	20 (7)	16 (6)	17 (8)	15 (7.5)	16.5 (6.3)	5.42	.005 **	3.10	.048 *	0.08	.776	0.10	.751	0.04	.964
Delayed TFR	19 (6)	20 (7)	29 (4)	28 (4)	29 (4)	30 (4)	3.24	.042 *	2.18	.118	0.06	.815	0.06	.802	0.07	.934

Note. Results presented in Columns A through D are from repeated measures ANCOVAs controlling for age and follow-up interval (F(1,60)).

^*p ≤ .05

^**p ≤ .01

^***p ≤ .001.

Abbreviations: SPI = semantic proactive interference, PRP = paired recall pairs, FRP = free recall pairs, TPR = total paired recall, TFR = total free recall.

Predictive Validity of the MBT and Standard Neuropsychological Tests.

We examined whether incident MCI, based on a CDR rating of = 0.5 at follow-up, was predicted by baseline biomarker status, MBT scores, and neuropsychological test scores, controlling for age and follow-up interval. Of the 132 participants with complete follow-up data, there was a significantly larger proportion of the preclinical group who developed incident cognitive impairment (9/49; 18.4%) compared to the healthy group (5/83; 6.0%) (X² = 4.95, p = .026). In a model including all biomarkers (amyloid, hippocampal volume, WMH lesion volume, and AD signature volume), only elevated amyloid at baseline was associated with increased risk of conversion (OR = 18.49). There was also a marginal effect of smaller hippocampal volume at baseline being associated with an increased risk of conversion (OR = 2.10). In separate models, none of the baseline MBT scores or neuropsychological factor scores successfully predicted conversion.

Test-retest Reliability of the MBT.

We assessed the test-retest reliability of the 9 MBT scores over the approximately 2-year follow-up interval in each group separately. Intraclass correlation coefficients (ICC) for each score are provided in Table 7, along with information needed for reliable change index calculation (i.e., means, SDs, and the correlation between scores at each time point). For the healthy group, all MBT scores had moderate reliabilities (ICCs between 0.5 and 0.75), with the lowest being immediate TFR and SPI at the bottom of that range. In general, reliabilities tended to be lower in the preclinical disease group. Most still fell in the moderate range except delayed TFR, which was somewhat lower (ICC = 0.42) and SPI, which was notably poor (ICC = 0.28).

Table 7.

Descriptive statistics and test-retest reliability over 2 years for each MBT variable.

	Subsample with Complete Follow-Up Data (n = 132)				Healthy (n = 83)				pAD (n = 49)
	Baseline Mean (SD)	Follow-Up Mean (SD)	r p	ICC	Baseline Mean (SD)	Follow-Up Mean (SD)	r p	ICC	Baseline Mean (SD)	Follow-Up Mean (SD)	r p	ICC
Learning Condition
SPI	90.2 (12.12)	90.29 (12.52)	0.39	0.40	90.16 (11.28)	91.11 (12.2)	0.49	0.49	90.27 (13.55)	88.89 (13.06)	0.27	0.28
Word Pairs
Immed. PRP	13.23 (2.44)	13.32 (2.62)	0.55	0.55	13.41 (2.56)	13.70 (2.56)	0.57	0.57	12.92 (2.21)	12.67 (2.63)	0.49	0.49
Delayed PRP	13.15 (2.64)	13.20 (3.10)	0.60	0.60	13.45 (2.69)	13.61 (2.98)	0.62	0.62	12.65 (2.5)	12.49 (3.20)	0.55	0.54
Immed. FRP	8.2 (2.85)	8.03 (3.22)	0.62	0.61	8.88 (2.80)	8.65 (3.30)	0.56	0.56	7.04 (2.58)	6.98 (2.82)	0.64	0.64
Delayed FPR	8.17 (3.10)	8.32 (3.20)	0.60	0.60	8.88 (2.89)	8.94 (3.02)	0.61	0.61	6.96 (3.09)	7.27 (3.25)	0.49	0.50
Total Recalls
Immed. TPR	29.03 (3.16)	29.08 (3.00)	0.59	0.59	29.16 (3.54)	29.48 (2.92)	0.64	0.62	28.82 (2.41)	28.39 (3.04)	0.50	0.49
Delayed TPR	28.92 (3.23)	28.89 (3.65)	0.59	0.59	29.23 (3.27)	29.33 (3.57)	0.63	0.63	28.41 (3.14)	28.16 (3.71)	0.45	0.46
Immed. TFR	18.33 (4.82)	17.86 (5.53)	0.56	0.56	19.69 (4.58)	19.11 (5.53)	0.51	0.50	16.02 (4.34)	15.73 (4.89)	0.52	0.52
Delayed TFR	17.96 (5.56)	17.99 (5.89)	0.58	0.58	19.23 (5.02)	19.17 (5.32)	0.67	0.66	15.82 (5.82)	16.00 (6.32)	0.42	0.42

Note. The values reported for r_p are the Pearson correlation coefficients between scores at baseline and follow-up. ICC = intraclass correlation.

Discussion

This study assessed the construct validity of the MBT in healthy adults and its sensitivity to preclinical disease. As is typical of neuropsychological tests, MBT scores were highly confounded by age but nonetheless demonstrated expected associations with standard neuropsychological measures in healthy adults, even after covarying for age. The MBT demonstrated modest sensitivity to preclinical disease. Cross-sectionally, the healthy and preclinical disease groups differed only in performance on the free recall, rather than the paired recall condition, and MBT scores were largely unrelated to AD biomarkers in the preclinical groups. Although free recall scores discriminated between healthy and preclinical groups at baseline over and above important factors such as age and performance on other neuropsychological measures, there were minimal within- or between-group differences in MBT scores over 2-year follow-up and none of the baseline MBT scores predicted incident cognitive impairment. Thus, our findings provide some support for the utility of this test in detecting early AD-related cognitive deficits cross-sectionally but reveal several limitations of this measure.

Construct Validity of the MBT

In healthy adults, the MBT exhibited evidence of convergent and discriminant validity with other standard neuropsychological tests. MBT scores showed strongest associations with memory (medium to large effect sizes) and general cognitive composite factor scores (medium effect sizes) but weaker associations with language and speed/executive and little association with visual and attention scores. Although correlations between MBT scores and memory were numerically larger than those for the other factor scores, most of these differences were not statistically significant. This is unsurprising given the factor structure of the UDS v3.0 neuropsychological battery, in which the five lower order factors covary (moderately to very strongly) amongst each other (Kiselica et al., 2020). Similarly, memory factor scores in our dataset had sizable associations with other domains (rho’s = 0.11 – 0.51), likely reflecting method covariance. Nonetheless, the general pattern of findings suggests that MBT scores are more related to episodic memory than non-memory domain, expanding upon previous evidence of convergent validity of the MBT with another test of associative memory (the FCSRT; Gramunt et al., 2016) by providing novel evidence of convergent and discriminant validity.

Of the demographic factors, age was most strongly related to MBT performance in the healthy group (medium effect sizes). Although unsurprising given the well-established decline of associative memory with age (Naveh-Benjamin, 2000), this poses a challenge to using the MBT as an early indicator of AD risk given that age is also a strong predictor of amyloid accumulation (Jansen et al., 2015) and development of MCI and AD (Campbell et al., 2013; Villemagne et al., 2011). Previous work exploring age-stratified MBT cut-scores found no significant difference in discriminability of healthy vs. aMCI (Buschke et al., 2017), but this age range was limited compared to our study (72–90 vs. 45–85 years in our study). Higher education was associated (to a more modest degree) with better paired recall performance. Educational attainment is a proxy for cognitive reserve (Stern, 2012), and individuals with more education may have utilized more effective strategies (e.g., mnemonics) for creating semantic associations (Frankenmolen et al., 2018). MBT performance did not differ by race or sex, in contrast to well-documented sex differences on other episodic and associative memory tests (Asperholm et al., 2019; Rentz et al., 2017). However, this study was underpowered to detect race and sex effects given the sample composition (72% Female, 94% White), and this lack of representativeness greatly limits generalizability. Thus, critical next steps are using larger, more diverse samples to establish whether the MBT demonstrates metric invariance across groups, validate normative data and possible demographic adjustments, and establish robust estimates of sensitivity and specificity across stages of the AD continuum (Parra, 2022).

Criterion Validity of the MBT

When assessing associations with biomarkers, we found that MBT performance was not associated with amyloid burden in any group but poorer performance on certain MBT conditions was associated with lower hippocampal volume in the AD continuum group, greater WMH lesion volume in both preclinical groups, and lower AD signature volume in the healthy group. The lack of significant associations between MBT performance and cerebral amyloid is inconsistent with the existing literature in cognitively unimpaired adults (Rentz et al., 2011; Vannini et al., 2012), including studies using the MBT (Gagliardi et al., 2019; Rentz et al., 2010). Of note, these findings remained when age was not covaried and do not appear to be due to a lack of accumulated pathology given the observed ranges of amyloidosis (mSUVr range: 0.90–1.23). However, it is possible that variations in biomarker parameterization, such as PET radiotracer and cortical regions sampled, between our study and others’ contributed to this discrepancy.

The AD signature marker comprises cortical regions that exhibit neurodeneration early in the AD course (Allison et al., 2019), and is a closer proxy for tau accumulation than β-amyloid (Wang et al., 2015). Therefore, the association between MBT scores and AD signature volume in the healthy group could conceivably indicate correspondence between associative memory and tau accumulation (a direct measure of which was not available in this dataset). However, a more likely alternative is that this reflects a link between associative memory and global brain volume, which may be indirectly represented by the multi-region AD signature volumes. This aligns with existing evidence that, even in cognitively unimpaired adults, global brain atrophy is moderately associated with individual differences in most cognitive domains (Fletcher et al., 2018).

Associations with hippocampal volume were isolated to MBT delayed free recall and observed in the AD continuum group only, which is unsurprising given the classic triad of medial temporal lobe atrophy, amyloidosis, and amnestic presentation in AD. This is consistent with previous findings in cognitively unimpaired adults (de Flores et al., 2015; Shing et al., 2011; Thomann et al., 2013) and adults with subjective cognitive decline and aMCI using the MBT (Markova et al., 2022). Within both preclinical groups, poorer performance on MBT free recall conditions was associated with greater WMH lesion volume, with medium to large effect sizes. Importantly, the magnitude of these effects remained even after removing participants with just WMH (n = 19), confirming that this was not attributable to using an expanded A(VN) framework. There is increasing recognition that cerebrovascular disease is a core feature of AD (Lee et al., 2016). WMH correlate with amyloid and tau accumulation (Kandel et al., 2016; Kester et al., 2014; Thanprasertsuk et al., 2014), predict AD onset both in conjunction with and independent of other AD biomarkers (Brickman et al., 2012, 2015; Mortamais et al., 2014), and precede AD-related cognitive decline (Brickman et al., 2008; Lo & Jagust, 2012; Mortamais et al., 2014). Further, WMH may result from AD-related degenerative processes, rather than solely reflecting ischemic disease (McAleese et al., 2017). Thus, these findings suggest that the MBT is sensitive to cognitive deficits related to both AD and/or vascular processes.

In support of the utility of the MBT for detecting subtle cognitive deficits, we found that the preclinical group performed worse than the healthy group on MBT immediate free recall, but not most other neuropsychological measures. Moreover, MBT total immediate free recall scores discriminated between group status at baseline in models including age and other neuropsychological scores. Notably, the MBT was more successful at discriminating the AD continuum, rather than non-AD pathologic change, subgroup from the healthy group. The observed AUCs (>0.70) are approaching the previously reported value for discriminating healthy controls from aMCI using MBT immediate paired recall (AUC = 0.79; Buschke et al., 2017). Thus, in addition to this prior evidence that the MBT can distinguish between those with AD, aMCI, and cognitively unimpaired controls, our findings suggest that this sensitivity may extend into the preclinical stage.

However, this finding is tempered by longitudinal results that are less promising. Specifically, over the 2-year follow-up period, we found no group differences in the magnitude of change in performance on the MBT, and none of the baseline scores successfully predicted incident cognitive impairment. These findings suggest that although the MBT can distinguish preclinical groups cross-sectionally at baseline, it may have limited sensitivity to disease progression and/or prognostic utility for conversion to MCI. This lack of effects may be understood with consideration of the temporal pattern of neuropathological and memory changes in AD. A model proposed by Didic et al. (2011) posits stages of memory dysfunction that follow neuropathologic progression, beginning with deficits in “context-free” memory with initial accumulation in transentorhinal (i.e., sub-hippocampal) cortex, followed by deficits in “context-rich” memory due to hippocampal involvement. Prior studies have substantiated this model, finding a temporal gradient of sensitivity in which context-free memory measures appear more sensitive to early/preclinical disease stages whereas context-rich memory measures perform better in more advanced disease stages (Forno et al., 2022; Parra et al., 2022). In this model, the MBT is considered context-rich as it relies on contextual information to support encoding and retrieval. It therefore may be missing the very earliest changes (e.g., in context-free memory and conjunctive rather than relational binding) that may be present in our preclinical groups (Bastin & Delhaye, 2023). However, it is also important to consider that we evaluated change over a relatively short 2-year interval, as compared to 4–13 years in prior prospective studies that found that the MBT could predict incident MCI and dementia (Mowrey et al., 2016, 2018). Relatedly, only 14 participants converted to MCI over the follow-up period, limiting our statistical power to test these effects. Therefore, future work with larger samples and additional follow-up points is needed to more definitively test the MBT’s predictive utility.

A surprising finding was that the effects we observed were largely localized to the free recall condition. The lack of findings for paired recall (i.e., cued) conditions, which would be expected to be most sensitive to associative binding, may be due to it being less challenging. Although previous work reports fewer ceiling effects on the MBT than similar measures (Gramunt et al., 2016), we still observed ceiling effects on the paired recall condition (57% of the healthy group and 40% of the preclinical group achieved scores >30/32 on total paired recall conditions at baseline). Notably, ceiling effects appear to be more pronounced in our study, with total paired recall scores being 4–6 words greater on average than previously studies of cognitively intact adults (Buschke et al., 2017; Gramunt et al., 2016). The reason for this is unclear but nonetheless suggests that the paired recall component (in which participants are cued) may not be challenging enough for cognitively unimpaired individuals, likely resulting in greater range restriction and reduced sensitivity to subtle group differences. This hypothesis is supported by a similar dissociation reported previously, with poorer MBT free recall in earlier disease stages (A+/N-) and poorer cued (i.e., paired) recall in more advanced stages (A+/N+; Papp et al., 2015). We observed strongest effects for the immediate conditions, likely reflecting the greater influence of age on delayed recall performance, which showed significant effects only when age was not covaried.

When considering the use of a measure that is designed to detect subtle impairment, it is important to consider its clinical utility. The time demands of the MBT (~45 minutes including a 30-minute delay) are slightly longer than those required for single-list verbal episodic memory measures (e.g., Rey Auditory Verbal Memory Test), but comparable to measures that similarly incorporate cueing and assessment of proactive interference (e.g., the California Verbal Learning Test). However, the MBT offers a more specific measure of associative binding as it utilizes controlled learning to ensure encoding and involves two wordlists with common semantic cues (Loewenstein et al., 2018). As there are other measures that show promising sensitivity to cognitive changes in preclinical AD, such as measures of cross-modal associative binding, change detection/discrimination, retention of complex objects, and dual tasking, among others (see Rentz et al., 2013 for review), future work may examine the relative merits of the MBT compared to these. Norms for the MBT do not currently exist, and our results suggest that creating age-adjusted norms may be an important next step in advancing the MBT. A small number of healthy individuals produced low scores, which likely represents normative variation or low effort/engagement, warranting consideration in certain evaluation contexts. Estimates of test-retest reliability appear to be adequate in both our study (ICCs = 0.40 – 0.61) and previous work in healthy adults and aMCI (ICCs = 0.58 – 0.85; Buschke et al., 2017); however, alternative forms are likely needed for serial assessment as practice effects are substantial (10–25% score increase over 6 weeks; Gramunt et al., 2016).

Constraints on Generality

Our study has several strengths that support its contribution to this literature as well as inherent limitations to generality that should be considered. First, we provide novel information about the convergent/divergent validity of the MBT and its incremental utility over neuropsychological measures (from the UDS v3.0) used in clinical and research settings. However, given that this battery includes only one short verbal episodic memory test (i.e., Craft Story), future work is needed to replicate these findings in comparison to more comprehensive memory tests. We chose to include the 9 MBT scores that are commonly evaluated in prior studies to facilitate interpretation of our findings within the context of the existing literature. This is further justified by the fact that these 9 scores map most directly onto the central constructs that this measure was designed to assess and that are most impacted early in the disease course of AD. Of note, correlational results were not corrected for multiple comparisons given that the goal of this study was to evaluate the pattern of associations across the multiple variables. We hope that subsequent studies will benefit from this comprehensive view and generate more directed, specific hypotheses with appropriate statistical inferential corrections. Although there are additional metrics that could be derived from the MBT, this was beyond the scope of this paper but may be a worthy pursuit for future studies.

By utilizing an expanded definition of preclinical disease, our sample is more representative of the heterogenous mix of amyloidosis, neurodegeneration, and vascular pathology observed in our target population of adults at risk of developing AD (Jack et al., 2018). A related limitation is that a measure of tau, which is highly implicated in AD risk, was not available in this study. However, we included an AD cortical signature metric that is a close proxy of tau distribution (Wang et al., 2015), strengthening our confidence that these results are applicable to the ATV(N) biomarker framework. Another feature of this study that supports its generalizability is that we took a two-pronged approach to assessing the MBT’s criterion validity by 1) testing for differences in the healthy vs. preclinical groups, which mirrors the categorical determinations made in clinical practice, and 2) assessing associations with continuous biomarker metrics to elucidate relationships that may be missed by dichotomization and to avoid circularity/criterion contamination. Future work with sufficient power could further illuminate whether these effects localize to specific biomarkers.

To further aid in generalizability, we were not overly stringent in our definition of cognitive impairment for use as an exclusionary criterion since many neurologically health adults will produce one or more “impaired” scores on standard neuropsychological testing (Palmer et al., 1998) due to a variety of situational and/or sociodemographic factors. By utilizing age- and education-adjusted MoCA scores, our intention was to be more inclusive of the full range of subtle cognitive deficits present among adults with and without biomarker evidence of preclinical disease. However, one potential downside of this approach to weigh against the improved representativeness is that it may limit the specificity of our findings. Despite our use of a broader definitions of preclinical disease and cognitive impairment, we acknowledge that the ability to generalize our findings to all adults at risk of developing AD is significantly limited by the composition of this study’s sample, which was majority White, non-Hispanic, college educated, and female sex. The underrepresentation of racial-, ethnic-, socioeconomic-, and sex/gender-diverse individuals in this study was likely driven by a recruitment approach that relied heavily on convenience sampling and the use of inclusion/exclusion criteria that may have systematically excluded individuals with conditions that are disproportionately experienced by certain minority groups. Future work should utilize culturally sensitive study design and recruitment methods that address systemic barriers that may impede members of marginalized communities from participating in research, with the goal of achieving better representation of the socio-demographic, racial, and ethnic composition of the target region (Parra, 2022). Critical next steps for studies of the MBT are to use larger, more diverse samples to replicate these findings, establish whether the MBT demonstrates metric invariance across groups, validate normative data and possible demographic adjustments, and establish robust estimates of sensitivity and specificity across stages of the AD continuum.

Conclusions

This study demonstrated that the MBT, a measure of associative memory, has adequate construct validity in healthy adults and demonstrates modest criterion validity with respect to evidence of preclinical AD in cognitively unimpaired adults. Our findings suggest that the MBT may be able to detect early AD-related cognitive deficits cross-sectionally, but this will require careful consideration of age as a confounding variable. Although the MBT exhibited adequate reliability over a 2-year interval, there were minimal changes in how the groups performed over that time span. Further, baseline performance on the MBT was not predictive of conversion to MCI, suggesting that it may have limited prognostic utility.

Key Points.

• Question: What are the psychometric properties of the Memory Binding Test (MBT) in healthy adults and those with biomarker evidence of preclinical neurodegenerative disease? How well can performance on the MBT predict progression to mild cognitive impairment (MCI)?

• Findings: The MBT is sensitive to early disease-related deficits in associative memory cross-sectionally despite being highly confounded by demographic factors, particularly age. Performance on the MBT at baseline did not predict incident MCI over 2 years.

• Importance: Although the MBT may be useful for detecting subtle cognitive deficits in individuals with biomarker evidence of preclinical disease, the MBT may be less useful for predicting who will progress to MCI in the near future.

• Next steps: Future work should test the generality of these findings using larger and more diverse samples, including additional biomarkers (e.g., tau), and comparing the relative merits of the MBT to other measures that are sensitive to cognitive change in preclinical disease. Additionally, the development of age-adjusted norms for the MBT may be warranted.

Data Availability:

The data used in this study are not publicly available as participants were not asked to consent to data sharing. Analysis code for this study may be requested from the corresponding author.