INTRODUCTION
The notion that there are memory systems that are able to operate independently from the hippocampus first became apparent in the case study H.M. H.M. acquired declarative memory impairment due to the surgical removal of parts of his medial temporal lobes including his hippocampi (Milner, Corkin, & Teuber, Reference Milner, Corkin and Teuber1968) after which he was no longer able to learn new facts and events. In contrast, on procedural learning tasks such as the Mirror Tracing task, his performance improved with repetition, showing an ability to learn (Corkin, Reference Corkin1968). Squire (Reference Squire2004) defined declarative memory as the conscious recollection about facts and events. As can be inferred from H.M., this ability heavily relies on the medial temporal lobe (Eichenbaum & Lipton, Reference Eichenbaum and Lipton2008; van Strien, Cappaert, & Witter, Reference van Strien, Cappaert and Witter2009), among other brain areas. In contrast, non-declarative memory was defined as a collection of abilities that are expressed through the benefit from previous experiences rather than recollection.
Early researchers distinguished between the non-declarative” (or “implicit”) and “declarative” (or “explicit”) memory systems based on whether or not these types of memory were reliant on the medial temporal lobe (Henke, Reference Henke2010; Squire, Reference Squire2004). Other researchers distinguished between these types of memory based on the presence of conscious awareness: they found that conscious awareness of the learned information was required for declarative memory while it was not required for non-declarative memory (Squire & Dede, Reference Squire and Dede2015). Both binary ways of distinguishing between declarative and non-declarative memory are now thought to be too simplistic, but the notion that there are multiple memory systems remains well accepted (Henke, Reference Henke2010). Two of the main subtypes of non-declarative memory are repetition priming and procedural learning. Repetition priming is defined as the facilitation of retrieval that results from previous exposure, which can be achieved outside of conscious awareness (Schacter et al., Reference Schacter, Eich, Gabel, Glisky, Graf, Jacoby and Shimamura1987). Procedural learning encompasses skill learning, pattern learning, and habit learning (Squire & Dede, Reference Squire and Dede2015). Skill learning tasks typically involve observable skills that improve with practice. Pattern learning tasks are tasks in which a pattern is repeated multiple times, with improved performance resulting from this repetition.
Cases like H.M. (Milner et al., Reference Milner, Corkin and Teuber1968) inspired researchers to try to find double-dissociations between these different types of learning and memory. Impaired procedural learning with intact declarative learning was found in individuals with basal ganglia and cerebellum-related disorders such as individuals with early Parkinson’s and Huntington’s disease, and individuals with cerebellar atrophy (Knowlton, Squire, Paulsen, Swerdlow, & Swenson, Reference Knowlton, Squire, Paulsen, Swerdlow and Swenson1996; Sanes, Dimitrov, & Hallett, Reference Sanes, Dimitrov and Hallett1990; Seger, Reference Seger2006). A similar double-dissociation was found for perceptual repetition priming but not for conceptual repetition priming (Gabrieli, Fleischman, Keane, Reminger, & Morrell, Reference Gabrieli, Fleischman, Keane, Reminger and Morrell1995; Keane, Gabrieli, Mapstone, Johnson, & Corkin, Reference Keane, Gabrieli, Mapstone, Johnson and Corkin1995). Other studies demonstrated a double-dissociation between priming and skill learning (Butters, Heindel, & Salmon, Reference Butters, Heindel and Salmon1990; Deweer et al., Reference Deweer, Ergis, Fossati, Pillon, Boller, Agid and Dubois1994; Yeates & Enrile, Reference Yeates and Enrile2005), suggesting that these different subtypes of non-declarative learning are also mediated by different brain areas.
Alzheimer’s disease (AD) is the most common cause of dementia, characterized by progressive cognitive decline that hampers independent functioning (Alzheimer’s Association, 2020). AD-related neurofibrillary lesions are typically found first in the medial perirhinal and entorhinal cortices and then spread to the neocortex and the other parts of the parahippocampal gyrus as the disease progresses (Braak & Braak, Reference Braak and Braak1991; Braak, Tredeci, Schultz, & Braak, Reference Braak, Tredeci, Schultz and Braak2000; Krumm et al., Reference Krumm, Kivisaari, Probst, Monsch, Reinhardt, Ulmer and Taylor2016). Other brain areas such as the primary motor and sensory cortices, the brain stem, and the cerebellum remain intact until the disease progresses (Fox et al., Reference Fox, Crum, Scahill, Stevens, Janssen and Rossor2001; Whitwell et al., Reference Whitwell, Shiung, Przybelski, Weigand, Knopman, Boeve and Jack2008).
In line with the typical spread in AD pathology, declarative learning and memory decline in the early stages of AD as atrophy tends to occur first in the hippocampus and entorhinal cortex (Stoub, Rogalski, Leurgans, Bennett, & deToledo-Morrell, Reference Stoub, Rogalski, Leurgans, Bennett and deToledo-Morrell2010). Brain structures involved in priming tasks are dependent on task characteristics (Schacter, Dobbins, & Schnyer, Reference Schacter, Dobbins and Schnyer2004), however, conceptual repetition priming is thought to rely on the inferior frontal and temporal regions including the fusiform gyrus and parahippocampal gyrus (Buckner, Koutstaal, Schacter, & Rosen, Reference Buckner, Koutstaal, Schacter and Rosen2000; De Wit et al., Reference De Wit, Marsiske, Kessels, Johnson, Thangwaritorn, O’Shea and Smith2021; Schacter, Wig, & Stevens, Reference Schacter, Wig and Stevens2007), which are likely impacted once the AD pathology spreads throughout the temporal lobe. Procedural learning is thought to rely on the subcortical brain areas that are thought to remain intact until the severe stages of AD (Knowlton et al., Reference Knowlton, Squire, Paulsen, Swerdlow and Swenson1996; Sanes et al., Reference Sanes, Dimitrov and Hallett1990; Seger, Reference Seger2006). As a result, both repetition priming and procedural learning are thought to remain intact in the very early stages of AD, with procedural learning remaining intact until the disease progresses to the severe stages (De Wit et al., Reference De Wit, Marsiske, Kessels, Johnson, Thangwaritorn, O’Shea and Smith2021, Reference De Wit, Marsiske, O’Shea, Kessels, Kurasz, DeFeis and Smith2020).
The pre-dementia stage during which cognitive impairment is present but does not significantly interfere with daily functioning is typically referred to as the mild cognitive impairment (MCI) phase of AD (Albert et al., Reference Albert, Dekosky, Dickson, Dubois, Feldman, Fox and Phelps2011). When declarative memory impairment is present in MCI regardless of the etiology, the syndrome is referred to as amnestic MCI (aMCI; Petersen, Reference Petersen2004). Because pharmacological interventions to date fail to halt or reverse cognitive decline in AD, behavioral interventions offer a promising approach to help preserve cognitive health and function in MCI due to AD (Chandler et al., Reference Chandler, Locke, Crook, Fields, Ball, Phatak and Smith2019; National Academies of Sciences, Reference Leshner, Landis, Stroud and Downey2017; Ngandu et al., Reference Ngandu, Lehtisalo, Solomon, Levälahti, Ahtiluoto, Antikainen and Kivipelto2015; Olazarán et al., Reference Olazarán, Reisberg, Clare, Cruz, Peña-Casanova, del Ser and Muñiz2010).
The existence of non-hippocampus-reliant memory systems, such as the procedural learning system, has inspired researchers to develop behavioral interventions for individuals with early AD that utilize procedural learning to help compensate for the declarative learning and memory deficit (Fleischman, Reference Fleischman2007; Greenaway, Hanna, Lepore, & Smith, Reference Greenaway, Hanna, Lepore and Smith2008). However, the literature regarding the degree of sparing in procedural learning and repetition priming across the AD severity spectrum remains inconclusive, which may prevent the incorporation of techniques based on these memory systems into clinical practice. Two meta-analyses that assessed procedural learning and repetition priming comparing individuals with aMCI or AD dementia to controls highlighted that only three studies assessed procedural learning and only four studies thus far assessed repetition priming in MCI compared to controls (De Wit et al., Reference De Wit, Marsiske, Kessels, Johnson, Thangwaritorn, O’Shea and Smith2021, Reference De Wit, Marsiske, O’Shea, Kessels, Kurasz, DeFeis and Smith2020). Procedural learning was shown to be similar in individuals with aMCI/AD dementia and controls, with no statistical or meaningful differences, suggesting sparing in aMCI/AD dementia (De Wit et al., Reference De Wit, Marsiske, O’Shea, Kessels, Kurasz, DeFeis and Smith2020). However, the literature on repetition priming in AD dementia is less conclusive (De Wit et al., Reference De Wit, Marsiske, Kessels, Johnson, Thangwaritorn, O’Shea and Smith2021) and, therefore, more studies on repetition priming in the MCI phase of AD are needed.
Another concern related to the literature on sparing of procedural learning and repetition priming is that many studies attempting to assess whether these types of learning remain spared in AD assessed only procedural learning or only repetition priming in their sample and typically do not report effect sizes. Drawing conclusions regarding memory sparing based on the absence of a significant difference between groups is problematic because of the role that statistical power plays in such comparisons. In addition, drawing conclusions about memory sparing based on the absence of a significant difference between groups results in a binary distinction of spared versus impaired functioning. This binary distinction fails to consider the possibility of a continuous degree of sparing/impairment across the AD severity spectrum in procedural learning and repetition priming. Such a continuum might be expected based on the patterns of localization and spread of AD-related pathology. The use of Bayesian statistics, which allows researcher to draw more continuous conclusions about the presence of an effect versus the null hypothesis along with the strength of the evidence (Lakens, McLatchie, Isager, Scheel, & Dienes, Reference Lakens, McLatchie, Isager, Scheel and Dienes2020) has not been applied yet in the context of memory sparing in AD or aMCI.
The present study extends the literature by assessing procedural memory, conceptual repetition priming, and declarative learning in individuals with aMCI and cognitively healthy older adults in a single study. We aim to provide an overview across the different types of learning by reporting between-group effect sizes as well as Bayes Factors for declarative learning, conceptual repetition priming, and procedural learning in individuals with aMCI and cognitively unimpaired older adults, measured within the same group of individuals. Given the neuropathological staging discussed above, when comparing aMCI and cognitively unimpaired older adults, we hypothesize to see a large between-group effect size and a Bayes Factor in favor of a between-group difference on declarative learning as declarative impairments define the diagnosis of aMCI. We hypothesize to see a small-to-moderate between-group effect size and a Bayes Factor in favor of a between-group difference on conceptual repetition priming. Last, we hypothesize to observe the smallest between-group effect size and a Bayes Factor in favor of the null hypothesis when comparing both groups on procedural learning.
METHODS
Design
The current, transnational study builds on the “Physical Exercise And Cognitive Engagement Outcomes For Mild Neurocognitive Disorder” (PEACEOFMND) study and the “Memory-Dementia & Mild Cognitive Impairment” (M-DeMi) study. The PEACEOFMND study is a multi-site, group-randomized trial that took place at the University of Florida, Mayo Jacksonville, and Tallahassee Memorial HealthCare, Florida, USA. The present study utilizes data collected during the baseline visit prior to any intervention (De Wit et al., Reference De Wit, O’Shea, Chandler, Bhaskar, Tanner, Vemuri and Smith2018). The M-DeMi study is a cross-sectional imaging study in cognitively unimpaired older adults and individuals with aMCI at Radboud University in Nijmegen, the Netherlands. The M-DeMi study aims to investigate different factors related to learning and memory performances. Both studies recruited individuals with aMCI and cognitively unimpaired older adults. The study samples and the eligibility screening processes are discussed in more detail in the Participants and the Procedures section below.
Measures
Screening measures
The Telephone Interview for Cognitive Status for Memory
The Modified Telephone Interview for Cognitive Status for Memory (TICS-M; Welsh, Breitner, & Magruder-Habib, Reference Welsh, Breitner and Magruder-Habib1993) was administered to all participants in the PEACEOFMND study. The TICS-M is a cognitive phone screener that is effective in identifying individuals with aMCI (Cook, Marsiske, & McCoy, Reference Cook, Marsiske and McCoy2009; Graff-Radford et al., Reference Graff-Radford, Ferman, Lucas, Johnson, Parfitt, Heckman and Crook2006). The 50-item version of the TICS-M was used for the present study.
The Clinical Dementia Rating scale
The Clinical Dementia Rating (CDR) was administered as part of both the PEACEOFMND and M-DeMi studies (Morris, Reference Morris1997). The CDR is a tool that assesses the influence of cognitive impairment on ADLs. The CDR is widely accepted as a reliable and valid measure for cognitive severity staging in individuals with cognitive impairment (Morris, Reference Morris1993). The current study only included participants with a CDR of 0 or 0.5.
Cognitive measures
The Dementia Rating Scale-2
The Dementia Rating Scale-2 (DRS-2) was used to assess global cognitive functioning in the individuals with aMCI that were part of the PEACEOFMND study. The DRS-2 is an instrument that assesses several cognitive domains including conceptualization, attention, initiation/perseveration, construction, and declarative memory (Jurica, Leitten, & Mattis, Reference Jurica, Leitten and Mattis2001).
The Montreal Cognitive Assessment
The Montreal Cognitive Assessment (MoCA) was used to assess global cognitive functioning in all participants of the M-DeMi study. The MoCA is a cognitive screener that assesses multiple cognitive domains including visuospatial/executive functioning, construction, naming, declarative memory, attention, and abstraction (Nasreddine et al., Reference Nasreddine, Phillips, Bedirian, Charbonneau, Whitehead, Collin and Chertkow2005).
The Rey Auditory Verbal Learning Test
The Rey Auditory Verbal Learning Test (AVLT; Rey, Reference Rey1964) was administered to assess declarative learning and memory. The AVLT consists of five consecutive 15-word list trials that were read to the participants. Total Recall was measured by the total number of words recalled after five trials (Trials 1–5). The Learning Score was calculated by multiplying the number of words on the first trial by five and subtracting this from the Total Recall score (Ivnik et al., Reference Ivnik, Malec, Tangalos, Petersen, Kokmen and Kurland1990). Finally, delayed recall is determined by the number of words recalled after a 30-minute delay. This task was only administered to the individuals in the PEACEOFMND study.
The Mirror Tracing Task
The Mirror Tracing Task (MTT; Corkin, Reference Corkin1968) was used to assess visuomotor procedural learning. Participants were asked to trace the outline of a five-point star without touching the edges while viewing their hand and the stimulus in a mirror, during which the direct vision of hand and the stimulus was blocked. Completion time and errors were calculated for each trial. Two blocks of five trials, separated by an interval of 15 minutes, were administered. To mitigate frustration, a 300 second time limit per trial was used. The main outcome measures were the individual’s linear change slopes across the 10 blocks for both errors (maximum of 100 errors) and time (in seconds) per trial. The generation of these slopes is described in more detail in the Slope Generation paragraph of the Data Analyses section.
The Serial Reaction Time Task
The Serial Reaction Time Task (SRTT) is a procedural learning task that involves pattern learning. The task in the present study was developed based on the paradigm of a previous study (Gabriel et al., Reference Gabriel, Maillart, Stefaniak, Lejeune, Desmottes and Meulemans2013). It was administered on an Android-run tablet (Samsung SM-T580NZWAXAR 10.1 Galaxy Tab A T580) using OpenSesame software (Mathôt, Schreij, & Theeuwes, Reference Mathôt, Schreij and Theeuwes2012). Participants were required to touch the cue as quickly possible. Each trial ended when the participants touched the cue or after 5000 milliseconds. A cue could appear in each quadrant of the screen, and the locations where a cue might appear remained visible at all times. The first six blocks were learning blocks during which the cue appears in a 12-element sequence that is repeated eight times per block, resulting in a total of 48 repetitions of the 12-element sequence. The seventh block was a transfer block during which a different 12-element sequence is repeated eight times.
Several procedural memory outcome measures were generated from this task all based on the median reaction time per block of block 1 through 7. First, a “pattern learning slope” was generated on the first six blocks. Additionally, a simple “pattern-transfer score” was generated based on the median reaction time of block 6 to 7. The generation of these slopes using is described in more detail in the Slope Generation paragraph of the Data Analyses section.
The Word-Stem Completion Task
The Word-Stem Completion Task (WSCT; Warrington & Weiskrantz, Reference Warrington and Weiskrantz1974) consisted of a set of 40 concrete nouns of 5-7 letters (e.g. “LEMON”) with a mean occurrence frequency of 40 per million (Rajaram & Roediger, Reference Rajaram and Roediger1993). The test occurred in two phases. First, participants were shown 13 words on a rating scale. Of these 13 words, one was added to the start and two words were added at the end of the rating scale to reduce primacy and recency effects, respectively; these extra three words were not part of the 40-word set. Participants were asked to rate how much they liked each word on a 5-point Likert scale. They were not told that they would be asked about the words again at a later stage. Immediately after, twenty visual word-stems on cards (e.g. LEM) were shown, one at a time, and participants were asked to complete the word-stems with the first word that came to mind. Fifty percent of the stems could be completed using the words from the rating scale they were previously exposed to, and the other fifty percent of the stems could not be completed with words they were previously exposed to in order to determine baseline guessing rates. During the second phase, which occurred right after the completion of the first phase, the full procedure was repeated using a new 13-word list and 20 word-stems. The main outcome variable was the ratio of the number of stems that were completed with the 20 words from the set of 20 words to which they were previously exposed divided by the number of stems that were completed with any word from the 40-wordset in total. This task was only administered to the individuals in the PEACEOFMND study.
Participants
The Study Characteristics of the samples can be found in Tables 1 and 2. Both studies recruited individuals with clinical diagnoses of aMCI (Albert et al., Reference Albert, Dekosky, Dickson, Dubois, Feldman, Fox and Phelps2011) and cognitively unimpaired older adults. Clinical diagnoses were made based on available medical records, often including neuro-imaging and neuropsychological evaluations. Additional criteria for participation are listed in Tables A1 and A2. The study eligibility screening process is discussed in more detail as part of the Procedures section below. The sample demographics for the participants are listed in Table 1. Overall, 85 cognitively unimpaired controls and 81 participants with aMCI participated in the study. The inclusion of individuals with aMCI was based on clinical diagnoses rather than cut-off scores on cognitive screeners. The lowest score on the MoCA for individuals as part of the M-DeMi study was 22. The lowest score on the TICS-M for individuals as part of the PEACEOFMND study was 21.
Sample demographics for the PEACEOFMND subsample
Sample demographics for the M-DeMi subsample
* Dutch educational system categorized into levels from 1 = less than 6 years of primary education to 7 = university degree (Duits & Kessels, Reference Duits and Kessels2014).
Study Procedures
Informed consent
All procedures were completed in accordance with the Helsinki Declaration. Only participants who provided both verbal and written informed consent were able to participate as approved by the Institutional Review Board of Medical Ethics Committee at each study site.
Study eligibility
Eligibility was determined by medical record review and a phone screening. Questions were asked about medications, language, physical abilities and basic demographic information such as age, gender, and contraindications to MRI (due to parent study aims). The criteria and procedures to ensure that the individuals with aMCI were in the MCI phase and that controls did not have cognitive impairment differed between the studies: as part of the PEACEOFMND study, the TICS-M (Welsh et al., Reference Welsh, Breitner and Magruder-Habib1993) was administered to both the individual with aMCI as well as the study partners, and the CDR was also administered. As part of the M-DeMi study, the MoCA (Nasreddine et al., Reference Nasreddine, Phillips, Bedirian, Charbonneau, Whitehead, Collin and Chertkow2005) was administered to all participants.
Test visit
During the test visit, several cognitive measures were administered to both the individuals with aMCI and the cognitively unimpaired controls. The PEACEOFMND study used the DRS-2 and the M-DeMi study used the MoCA. The MTT and SRTT were administered as part of both studies. As part of the PEACEOFMND study, the AVLT and the WSCT were also administered.
Data Analysis
General analysis information and testing statistical assumptions
Data analyses were conducted using IBM® SPSS (Armonk, 2007) and R (Team, Reference Team2015). Prior to data analysis, all dependent variables were converted to Z-scores checked for their distribution, incorrect values, and outliers. Throughout this study, Z-scores of > 3 above or below the full sample mean were considered outliers. While the distributions were close to normal, the dependent variables that remained statistically non-normally distributed after excluding outliers were normalized by a Blom transformation (Blom, Reference Blom1954). Age and level of education were explored as covariates in all analyses.
Procedural Learning Slope Generation
The MTT time and error slopes were generated for all participants that completed more than six trials and the SRTT pattern learning slope was generated for those who completed all six learning blocks. For the MTT, there were 27 participants that discontinued the MTT before or during the 6th trial, which was 15.9% of the participants who completed the MTT. For the SRTT, there were three participants (all individuals with aMCI from the PEACEOFMND study) that were excluded as their Z-score of the standard deviation of the median reaction time for the first 6 blocks was higher than three. Eight participants (two controls and six individuals with aMCI; one of the individuals with aMCI was from the M-DeMi otherwise all PEACEOFMND) were excluded because they made 30 or more errors during the six learning blocks.
For the MTT, both error and time slopes were generated based on the errors per trial and time to completion (in seconds) per trial, respectively. The trial variable was centered and served as the time IV for the MTT slopes. For the SRTT, a ‘pattern learning’ slope was generated based on the median reaction time per block of the six learning blocks. The block variable was centered and served as the time IV for the SRTT pattern learning slope. Three slopes were generated by outputting the coefficients when running individual linear Ordinary Least Squares regressions with the time variable as the IV in SPSS. Additionally, a “pattern-transfer” simple slope score was generated based on the median reaction time of block 6 to block 7 using the same method. The SRTT pattern learning slope, SRTT pattern-transfer score, MTT error slope, and MTT speed slope slopes were outputted and served as the procedural learning DVs in the main analyses.
Main analyses
One-way analysis of covariance (ANCOVA) models were run; separate models were run for each of the dependent variables (i.e. the AVLT learning score, SRTT pattern learning slope, SRTT pattern-transfer score, MTT error slope, and MTT speed slope). Group (aMCI vs. healthy controls) served as the independent variable. The study (PEACEOFMND vs. M-DeMi), level of education, and age were included as covariates. The F-values outputted by the ANCOVA models were used to calculate effect sizes.
Effect size calculations
Between-group Hedges’ g effect sizes comparing performance in aMCI to cognitively unimpaired older adults were calculated for each of the dependent variables in four sets of analyses. Cohen’s ds and Hedges’ gs are used to describe the standardized mean difference of an effect between groups and can be used to compare effects across measures or studies (Lakens, Reference Lakens2013).
In the first, third, and fourth sets of analyses, Cohen’s ds were calculated without controlling for age and education in the full sample, and the PEACEOFMND and M-DeMi subsamples, respectively. In these three sets, Cohen’s ds were calculated using the samples ns, means, and SDs. The Practical Meta-Analysis Effect Size Calculator by Wilson was used to derive separate between-group effect sizes (Cohen’s ds) and effect size variances (retrieved in January 2021; Wilson, n.d.). Because Cohen’s d can be inflated in small sample sizes, we converted all Cohen’s d values to Hedges’ g values. The procedural learning slopes were aggregated using the Robust Variance Estimation (RVE) method function for linear mixed effect models (Fisher & Tipton, Reference Fisher and Tipton2015) using R studio (Team, Reference Team2015).
Effect size assumptions
Throughout this manuscript, a negative Hedges’ g indicates that there was more learning in individuals with aMCI than in cognitively unimpaired controls. Further, a priori assumptions about the bounds of Hedges’ g were made regarding what would be considered a meaningful effect. Per Cohen’s criteria, a Hedges’ g of 0.200 is considered ‘small,’ 0.500 is considered ʼmedium,’ and 0.800 is considered ʼlarge’ (Cohen, Reference Cohen1977). Because a Hedges’ g of 0.200 is typically considered small yet meaningful, a Hedges’ g smaller than the absolute value of 0.200 is considered a trivial or non-meaningful difference for the present study.
Bayesian statistics
We repeated the analyses using Bayesian ANCOVAs in JASP (JASP Team, 2020) to calculate Bayes Factors (BF). BFs were used to determine how likely to be true the results obtained with each group comparison were, given the data. The main advantage of using Bayesian statistics is that it provides a quantification of the evidence in support of the null hypothesis, rather than only against it (Lakens et al., Reference Lakens, McLatchie, Isager, Scheel and Dienes2020). An uninformative prior was used across Bayesian analyses to allow closer alignment with null hypothesis testing. Further, using an uninformative prior allowed us to keep the prior consistent across all three types of memory, which facilitates more accurate comparisons across tasks. To facilitate our interpretation of the BFs, the benchmarks of Jeffreys (Reference Jeffreys1961) were used. Per these benchmarks, a BF10 of ≥ 100 provides decisive evidence for H 1 , a BF10 of 10–30 provides very strong evidence for H 1 , a BF of 3–10 provides substantial evidence for H 1 , a BF of 1–3 provides anecdotal evidence for H 1 , a BF of 1 provides no evidence. Further, a BF10 between 1/3 and 1 provides anecdotal evidence for H 0, and a BF10 between 1/10 and 1/3 provides substantial evidence for H 0.
RESULTS
Results of Assumption Testing
After the slope generation, several outliers in scores (z-score > ±3) were identified and removed (n = 1 for AVLT learning score; n = 2 for MTT error slope; n = 0 for MTT time slope; n = 1 for SRTT pattern learning slope; n = 1 for SRTT pattern-transfer simple slope, and n = 2 for WSCT ratio score).
Main Results Comparing Individuals with Amnestic MCI to Controls
The Hedges’ gs for all four sets of analyses are listed in Table 3. The results for these same analyses in the subsamples are listed in Table A3 to allow the comparison of these results separately for the PEACEOFMND and M-DeMi studies. The output of the ANCOVAs used to calculate Hedges’ g controlled for age and education can be found in more detail in Table A4, and the information used to calculate Hedges’ g for the full sample and the subsamples without controlling for the level of age and education can be found in Table A5.
Hedges’ gs comparing individuals with aMCI to cognitively unimpaired older adults
Note. g indicates Hedges’ g; gv indicates the variance of Hedges’ g; BF indicates Bayes Factors * indicates a significant difference between the groups at p ≤ .05; The signs (negative vs. positive) of the Hedges’ gs were reversed when required such that, for all scores, negative Hedges’ gs indicate more learning in individuals with aMCI than in cognitively unimpaired controls.
Our results showed a significant learning difference (p < .001) on the AVLT, with a medium-to-large effect size (g = 0.658) and a Bayes Factor indicative of decisive evidence for the presence of a between-group difference (BF10 = 859.046) between cognitively unimpaired older adults and individuals with aMCI.
For the WSCT, no significant difference (p = .067) was found between cognitively unimpaired older adults and individuals with aMCI; although a small-to-medium effect size (g = .333) was obtained and the Bayes Factor provided anecdotal evidence for the presence of a between-group difference (BF10 = 2.267).
For procedural learning, when comparing cognitively unimpaired older adults to individuals with aMCI and while controlling for age and education, we did not find any statically significant differences and found that Hedges’ g was smaller than the a priori set boundary of a meaningful difference for both the SRTT learning slope (g = 0.175, p = 0.317) and the SRTT pattern-transfer score (g = 0.099, p = 0.570). When looking at the BF, the SRTT learning slope provided substantial evidence for the null hypothesis (BF10 = 0.283) and the SRTT pattern-transfer score provided anecdotal evidence for the null hypothesis (BF10 = 0.437).
The pattern of findings for the MTT was more complex: we found no significant difference between cognitively unimpaired older adults and individuals with aMCI for the MTT error slope (p = 0.682) or time slope (p = 0.204). The Hedges’ g was larger than the a priori set boundary of a meaningful difference for both the MTT error slope (g = –0.230) and the MTT time slope (g = 0.243). However, these two effect sizes were in opposite directions, with the cognitively unimpaired older adults showing more learning on the time slope and the individuals with aMCI showing more learning on the error slope. When looking at the Bayes Factors, both the MTT error and time slopes provided anecdotal evidence for the null hypothesis (MTT error slope: BF10 = 0.656; MTT time slope: BF10 = 0.464).
Aggregating all four Hedges’ gs comparing cognitively unimpaired older adults to individuals with aMCI yielded a meta-Hedges’ g of 0.075 (SE = 0.067, 95% CI [–0.765, 0.914]. This aggregated Hedges’ g is smaller than the a priori set boundary of a meaningful difference. However, the confidence intervals expand beyond the a priori set boundary of a meaningful difference. Further, these aggregates are exploratory and should be interpreted with caution.
DISCUSSION
The present study assessed declarative learning, conceptual repetition priming, and procedural learning in individuals with aMCI in order to provide a direct comparison and nuanced overview of different aspects of learning, by examining between-group effect sizes across memory types in individuals with aMCI compared to cognitively unimpaired older adults. The current results suggest that in aMCI, declarative memory is impacted the most, followed by conceptual repetition priming, with procedural memory remaining largely spared.
Consistent with our hypotheses, when comparing individuals with aMCI and cognitively unimpaired older adults, our results showed a large between-group effect size and a Bayes Factor in favor of the presence of a between-group difference on declarative learning. This finding is expected because it is the presence of declarative memory deficits that generally leads to the diagnosis of aMCI. What is of greater interest is the pattern of effect sizes for procedural learning and conceptual repetition priming. For procedural learning, overall trivial but variable differences were seen, with Bayes Factors that were less conclusive but were somewhat in favor of the null hypothesis. The results regarding conceptual repetition priming fell in between the declarative and procedural learning, with a small-to-moderate between-group effect size and a Bayes Factor that was not conclusive but was slightly in favor of the presence of a between-group difference. These results suggest that a small-to-medium group difference between the individuals with aMCI and controls may be present in our sample but may not be detectable with our sample size due to statistical power.
Our priming results are consistent with the results from a recent meta-analysis on repetition priming that showed that the between-group effect sizes comparing individuals with aMCI to controls varied based on task characteristics, with the differences between individuals with aMCI and controls being the largest for tasks similar to the WSCT that require (1) conceptual processing and (2) production rather than the identification of primes in the test phase (De Wit et al., Reference De Wit, Marsiske, Kessels, Johnson, Thangwaritorn, O’Shea and Smith2021). Our procedural learning results, while variable, were overall consistent with the results from a recent meta-analysis regarding procedural learning that demonstrated trivial between-group effect sizes (Hedges’ g of <0.200) for individuals with aMCI and controls on procedural learning tasks (De Wit et al., Reference De Wit, Marsiske, O’Shea, Kessels, Kurasz, DeFeis and Smith2020). These meta-analyses were only able to include three procedural learning and four priming studies that assessed individuals with aMCI in comparison to cognitively unimpaired older adults, thereby limiting ability to draw conclusions regarding sparing of these types of memory in the MCI phase of AD. The present study contributes to the literature by providing a larger sample of individuals with aMCI compared to cognitively unimpaired older adults on measures of procedural learning and priming and is of additional value as it provides a concurrent view of impairment and sparing across declarative memory, conceptual repetition priming, and procedural learning by (1) assessing these three types of learning in the same groups of individuals and (2) providing effect sizes as well as BFs. Given that our results regarding procedural learning in individuals with aMCI remain inconclusive, more research on this topic is warranted. While future research using similar types of procedural learning tasks may be helpful, the use of novel procedural learning paradigms that more closely resemble the procedural learning component of the use of a memory compensation tool may be of even more clinical benefit.
While not all aMCI subjects have an underlying pathology of AD, prior research suggests that AD is the most common etiology of aMCI (Petersen, Reference Petersenm2016). Thus, our findings are consistent with the model that AD pathology impacts brain areas that are important for declarative memory, including the entorhinal cortex and hippocampus, followed by brain areas important for conceptual priming, including other parts of the parahippocampal gyrus and the fusiform gyrus; our results contribute to the evidence-base that suggests that subcortical areas important for procedural memory are not impacted until the more severe stages of the disease.
Limitations and Generalizability
A limitation of the present study is its cross-sectional nature. A longitudinal design could more accurately depict this change, and would also address the limitation that we do not know whether our aMCI subjects are indeed destined for AD dementia. However, longitudinal testing for skill learning represents a unique challenge due to practice effects: skills are only new during the first exposure. Similarly, procedural learning is often thought of as an unconscious process (Squire, Reference Squire2004), and repeated testing can result in an increased awareness of the goal of procedural learning tasks, which is particularly true for the SRTT, and this increased awareness can interfere with researchers’ ability to limit the influence of declarative learning on test performance. Thus, procedural learning does not lend itself well for repeated testing or longitudinal assessments. While it may be possible to asses procedural memory over time as evidenced by sustained performance gains on procedural memory tasks, this reflects a somewhat different process than procedural learning (much like the difference between declarative learning, i.e., the encoding process, vs. declarative memory, i.e. the information that is retained over time). Similarly, priming is often thought of as an unconscious process (Squire, Reference Squire2004), and repeated testing can result in an increased awareness of the goal of the priming task in participants, which could interfere with researchers’ ability to limit the influence of declarative learning on test performance. Another limitation of the present study is that the declarative learning (AVLT) and conceptual repetition priming (WSCT) tasks were only administered as part of the PEACEOFMND study and not as part of the M-DeMi study. Further, while the cross-national nature of the current study adds to external validity, another limitation of our study is that most of our participants were highly educated and from a predominantly Caucasian population with the ethnicity of the ethnic majority in each country. Another limitation of the current study is that slightly different inclusion criteria were used for both parent studies, although we do not believe that this impacted our results. Last, the current study did not include biomarkers of AD pathology. Future studies should recruit a more diverse sample and include AD biomarkers.
Implications and Conclusion
It is well known that declarative memory is impaired in both aMCI due to AD and AD dementia, but the extent to which non-declarative learning is spared remains unclear. The results of the present study suggest that conceptual priming may not remain intact in aMCI while procedural learning does appear to remain intact. Based on our findings, we recommend for studies that aim to build on spared functions in aMCI not to rely on conceptual repetition priming techniques. However, procedural memory-based cognitive interventions have potential utility in individuals with aMCI. Given that most procedural learning tasks used visuomotor learning, using a behavioral compensation technique that most heavily relies on a visuomotor component is currently the most supported approach. As an example, focusing the behavioral compensation intervention on the act of physically locating, opening, and using the compensatory tool consistently throughout the day may be most helpful, in line with the interventions used by several recent studies (Greenaway et al., Reference Greenaway, Hanna, Lepore and Smith2008; Vasquez, Lloyd-Kuzik, & Moscovitch, Reference Vasquez, Lloyd-Kuzik and Moscovitch2021). As part of this approach, individuals with aMCI should be trained to use their compensatory tool consistently when they either (1) have forgotten information or (2) are hoping to remember specific information. Given that most compensatory tools also require the understanding of how to use the tool and considering that the complexity of such tools can vary widely, we recommend direct testing of any procedural memory-based cognitive interventions in an aMCI population before providing such interventions clinically.
ACKNOWLEDGEMENTS
We thank Dr. Inti Brazil for his guidance regarding Bayesian statistics.
FINANCIAL SUPPORT
Financial support was provided by the Ed and Ethel Moore Alzheimer’s Disease Research Program of The Florida Department of Health [7AZA1], the Gravitation Grant from the Netherlands Organization for Scientific Research to the Language in Interaction Consortium (024.001.006), and Alzheimer Nederland (# WE.2017-05). ClinicalTrials.gov. Identifier: NCT03095170.
CONFLICT OF INTEREST
The authors have nothing to disclose.
ETHICAL STANDARDS
All procedures were completed in accordance with the Helsinki Declaration. Both parent studies were approved by the Institutional Review Board or the Medical Ethics Committee at each study site.
APPENDIX:
Inclusion and exclusion criteria for participation in the PEACEOFMND study
Note: aMCI: amnestic Mild Cognitive Impairment. MRI: Magnetic Resonance Imaging; CDR: Clinical Dementia Rating Scale; TICS-M: Telephone Interview for Cognitive Status for Memory.
Inclusion and exclusion criteria for participation in the M-DeMi study
Note: aMCI: amnestic Mild Cognitive Impairment; MRI: Magnetic Resonance Imaging; CDR: Clinical Dementia Rating Scale.
Hedges’ gs comparing individuals with aMCI to cognitively unimpaired older adults for the PEACEOFMND and M-DeMi subsamples without covariates
Note. g indicates Hedges’ g; gv indicates the variance of Hedges’ g; * indicates a significant difference between the groups at p ≤.05; The signs (negative and positive) of the scores were reversed when required such that, for all scores, negative Hedges’ gs indicate more learning in individuals with aMCI than in cognitively unimpaired controls.
One-way analysis of covariance models in aMCI versus cognitively unimpaired older adults
Note. MSS: Sum of Squares. F: F-values p: p-values. MS: Mean Square values. F-values were used to calculate Hedges’ gs controlling for age and education. * indicates significance at p ≤ .05; ** indicates significance at p ≤ .01.
Means (M), Standard Deviations (DVs), and sample sizes (ns) for the control group and the MCI groups that were used to calculate Hedges’ gs
Note. M: Group mean, n: number of participants per group. SD: Standard Deviation of the group mean. ** indicates significance at p ≤ .01; *** indicates significance at p ≤ .001.
Open access
