Recognition memory shielded from semantic but not perceptual interference in normal aging
Introduction
Long-term declarative memory function declines in old age. Deficits in older adults have been reported for several forms of long-term declarative memory including episodic, semantic, and spatial memory (Bäckman and Nilsson, 1996, Newman and Kaszniak, 2000, Rönnlund et al., 2005, Schaie, 2005). In particular, age-related impairments are frequently found in tasks that require the retrieval of information via associations, such as in contextual, associative, and source memory tasks (Chalfonte and Johnson, 1996, Duarte et al., 2008, Naveh-Benjamin, 2000, Schacter et al., 1991, Schacter et al., 1994, Silver et al., 2012, Spencer and Raz, 1995; but see Campbell et al., 2010). Regarding the neuroanatomical changes associated with age-related memory decline, research has revealed disproportionate deterioration of medial temporal lobe (MTL) structures (Devitt and Schacter, 2016, Raz et al., 1998, Raz et al., 2004, Wang et al., 2002). Given the cognitive effects of aging, this finding fits with the well-documented role of the MTL in long-term declarative memory (Scoville and Milner, 1957, Squire and Wixted, 2011, Squire and Zola-morgan, 1991) and, more specifically, in associative or relational memory (e.g., Cohen et al., 1999; Eichenbaum et al., 1994; Giovanello et al., 2004).
One theory of the role of MTL structures in cognition – the Representational-Hierarchical (R-H) account (Bussey and Saksida, 2002, Cowell et al., 2006, Sadil and Cowell, 2017) – has begun to amass empirical support for its explanation of mnemonic and perceptual deficits following brain damage (Barense et al., 2005, Barense et al., 2007, Barense et al., 2012a, Bartko et al., 2010, Bartko et al., 2007b, Bartko et al., 2007a, Lee et al., 2005a, Lee et al., 2005b, Lee et al., 2007, McTighe et al., 2010). However, to our knowledge this theory has not been applied to the effects of aging on memory (but see Scheerer and Marrone, 2014). And yet it makes clear predictions for memory performance that distinguish it from other theories of age-related memory loss (e.g., Brainerd et al., 2009; Naveh-Benjamin, 2000; Yassa and Stark, 2011). In this article, we outline those predictions and report an empirical study that tested them.
Briefly, the R-H account assumes that the ventral visual pathway contains a hierarchically organized system of representations that culminates in the MTL. At the start of the pathway, in visual cortex, the simple features of visual stimuli (e.g., color, orientation) are represented individually, and those features are brought together into cohesive conjunctions of increasing complexity in progressively more anterior regions. MTL structures contain conjunctive representations of whole stimuli such as objects, scenes and episodic events. These representations determine the role of MTL structures in cognition: whenever a cognitive task requires discrimination between stimuli or events containing shared features, conjunctive MTL representations are needed to resolve the feature ambiguity (Bussey and Saksida, 2002, Bussey et al., 2002). This prediction has been confirmed using tasks that are traditional tests of memory (e.g., Bartko et al., 2010, Bartko et al., 2007b; McTighe et al., 2010; Yeung et al., 2013) as well as tasks that tap perception (e.g., Barense et al., 2005; Barense et al., 2012a; Barense et al., 2012b; Bartko et al., 2007a, Bartko et al., 2007b; Lee, 2006; Lee et al., 2005b; Lee et al., 2006; Lee et al., 2008), lending support to the proposal that the role of MTL structures in cognition is determined by representational content, rather than by specialization for a particular cognitive process.
If normal aging compromises MTL structures and the representations contained therein, then the R-H account makes predictions for age-related changes in memory. Although cognitive aging is a complex phenomenon that likely stems from widespread changes in brain structure and function (Cabeza et al., 2016, Fjell and Walhovd, 2010), to the extent that MTL structures are disproportionately affected by aging, the R-H account predicts that age-related impairments in recognition memory should share some properties with the impairments seen after MTL damage. In this study we test whether the R-H account can explain age-related memory changes by focusing on one of its key predictions for memory function, regarding the resolution of different types of mnemonic interference. Specifically, the R-H account makes clear predictions for “feature-level interference” – a kind of perceptual interference in which features possessed by the to-be-discriminated test stimuli appear repeatedly during a task. According to the R-H account, such interference should impair the discrimination of old and new items when there is damage to conjunctive representations in MTL, because these representations are necessary for uniquely specifying an object (or event) and distinguishing it from similar items with shared features (Cowell et al., 2006). In support of this, studies have shown that feature-level interference causes impairments in recognition memory in individuals with MTL damage (Bartko et al., 2010, Bartko et al., 2007b, McTighe et al., 2010, Yeung et al., 2013). To the extent that aging damages the representations and processes normally carried out in MTL, the R-H account predicts the same interference effect in older relative to younger adults. However, since this prediction derives from the claim that conjunctive representations in MTL are compromised, a second, more counterintuitive prediction for interference also follows from the R-H account. Consider a scenario in which the repeatedly appearing (i.e., interfering) attributes of the to-be-discriminated stimuli reside at the level of complex conjunctions, or high-level associations – effectively creating “conjunction-level” interference. In this case, older adults should be paradoxically shielded from interference, because they are less able to encode that interference than young adults.
We operationalized these predictions for healthy older adults using a version of the Deese-Roediger-McDermott (DRM) recognition memory paradigm (Deese, 1959, Roediger and McDermott, 1995). In this paradigm, participants are presented with categories of related items at study, then tested with both studied items and unstudied lures, with some of the lures related in theme to the studied items. Participants of all ages tend to endorse related lures as studied with a higher probability than unrelated lures, a memory impairment that could be explained in terms of interference between similar items (see Gallo, 2006 for a review). Our goal was to use this paradigm to create both feature-level and conjunction-level interference.
Word stimuli constitute the largest and most easily manipulated source of items for use with the DRM paradigm. Given that we had previously applied the R-H account only to visual stimuli, in order to use words, we needed to specify the format of word representations within the assumed representational hierarchy. Popular models of visual word recognition and models of language production assume that humans possess two separate representational hierarchies for words – one stream for phonemic/orthographic features and the other for semantic content (Coltheart et al., 2001, Dell, 1986). This assumption is justified intuitively by considering that two very orthographically similar words such as TIDE and TILE have extremely dissimilar and separable semantic content. We therefore took our inspiration from these prominent models of word representation and blended their assumption with those of the R-H account, yielding a model with two separate streams for word representations that each contain a hierarchical continuum of increasingly complex representations (Fig. 1). We propose that the phonemic/orthographic hierarchy culminates in a conjunctive representation of the combination of phonemic/orthographic features that uniquely defines a whole word, assuming the word has no homonyms (Fig. 1, left column). In contrast, the semantic hierarchy culminates at the level of semantic themes or “gist”, which is represented via the semantic associations between individual words, rather than being contained within individual word representations themselves (Fig. 1, right column). We further suggest that older adults have compromised representations at the top level of both hierarchies: aging renders whole-word level representations of the conjunctions of phonemic/orthographic features less precise, and renders representations of the semantic associations between words less robust (Fig. 1, bottom row).
The notion that the MTL and nearby temporal cortex are involved in representing individual words phonemically/orthographically, particularly when those words are similar, and that aging can affect these representations, was inspired by a number of empirically-based claims: that perirhinal cortex is situated at the top of the visual word processing hierarchy (Liuzzi et al., 2015); that perirhinal cortex is engaged specifically when abstract visual stimuli share a high degree of feature-overlap (Barense et al., 2012a, Barense et al., 2010); and that older adults suffer increased interference, or competition, between phonological neighbors (Abrams et al., 2007, Gordon and Kurczek, 2013).1 The notion that semantic properties and semantic associations depend upon MTL structures is supported by a considerable body of research (Barense et al., 2010, Bruffaerts et al., 2013, Chadwick et al., 2016, Clarke and Tyler, 2014, Manns et al., 2003, Warrington and Shallice, 1984).
By using the DRM paradigm to create interference that affects these two putative processing streams at different levels, we created the conditions of feature-level and conjunction-level interference necessary for testing our predictions. Specifically, we created categories of items possessing shared phonemic/orthographic features but few semantic associations (e.g., BEST, LEST, NEST, PEST, REST, TEST, VEST, WELT, WENT, WEPT), and categories of items that shared a semantic theme but few phonemic-orthographic features (e.g., ASSAILANT, BANDIT, BOOTY, BURGLARY, HOLDUP, MUGGING, STEALING, STICKUP, THEFT, WALLET). In the phonemic/orthographic categories, the shared information across items that provides the source of interference resides at the sub-word level: the letter- and phoneme- components appear repeatedly. This causes the representations of simple perceptual features to appear familiar, even for novel words. In young adults (Fig. 1, top row), intact phonemic/orthographic representations at the whole-word level can resolve this interference because the unique conjunction corresponding to an individual stimulus is not repeated and does not become familiar for novel items. But in older adults the conjunctive word-level representations of phonemes/orthography are compromised (Fig. 1, bottom row), forcing reliance upon feature-level representations, thereby increasing the effects of interference. Thus, the phonemic/orthographic categories provide a “feature-level interference” condition in which older adults should be less able to resolve the interference than young adults. In semantic categories, the shared information across items resides at the supra-word level: inter-item semantic associations. When a semantic theme occurs repeatedly, word-level semantic representations for words that were unstudied but related to the theme begin to accrue familiarity owing to inter-item semantic associative activation. That is, the effects of semantic interference on novel test items are mediated by the higher-level semantic associations. Because these associations are assumed to be weaker in older adults than in young adults, this interference is less potent for older adults. Thus, the semantic categories provide a “conjunction-level interference” condition in which older adults should be paradoxically shielded from interference, relative to young adults.2
Despite a number of prior studies investigating age-related changes in semantic and perceptual memory using the DRM paradigm, these predictions have not yet been tested. Difficulty in drawing relevant inferences from the literature stems from two sources. First, some prior DRM studies of memory in older adults are subject to limitations in task design or data analysis that allow for non-mnemonic explanations of the data, either in terms of differences in response bias between groups and conditions, or in terms of item confounds. Second, perhaps because of the variety of task designs and analyses employed, findings from prior DRM studies comparing perceptual and semantic interference in older adults are conflicting: some have reported age-related deficits for perceptual but not semantic interference (Ly et al., 2013), others have observed greater deficits for semantic than perceptual (Koutstaal et al., 2003), and others still have found deficits for both (Budson et al., 2003, Pidgeon and Morcom, 2014).
Thus, the goal of the present study was to test whether the R-H account of MTL amnesia can explain how the effects of interference on memory change in normal aging, with an empirical study that avoids some of the confounds that complicate the interpretation of prior, related research. Specifically, we included three important features in our experimental design and analysis, as follows.
First, we used signal detection theory (SDT) to extract a measure of discriminability, d′. We use d′ because in a standard memory task like the DRM paradigm the R-H account makes predictions for recognition memory performance, rather than for false memory (i.e., false alarm (FA) rates). In standard recognition memory tasks with a yes/no test format, response bias plays a role, therefore assumptions about decision criteria are required to make a priori predictions for hit and FA rates. The R-H account makes no such assumptions and so, for the DRM paradigm, it makes predictions only for memory sensitivity, as indexed by a dependent variable such as d′. (But see McTighe et al. (2010) and Yeung et al. (2013) for incidental recognition memory tasks in rats and humans, respectively. In these tasks, decision criteria had little or no influence on memory performance – which was indexed by exploration time or eye fixations – and hence the R-H account was able to make predictions specifically concerning false memory for these studies). Further, the d′ measure confers two key advantages over hit and FA rates. The first advantage is that d′, given the assumptions of SDT, lies on an interval scale. The inferential test that we planned to use – analysis of variance (ANOVA) – assumes that the dependent variable is normally distributed with homogeneous variance; these properties are frequently violated by proportion scores, such as hit and FA rates, but much less often violated by an interval scale measurement such as d′. Furthermore, the use of an interval scale was critical because our key prediction was an interaction between the experimental manipulations (small/large category size, perceptual/semantic interference type) and participant group (older/young age). While main effects are testable regardless of the measurement scale, a non-crossover interaction observed with a dependent variable that does not lie on an interval scale (such as hit or FA rate) may be "removable" and hence uninterpretable (Loftus, 1978, Wagenmakers et al., 2012). The second advantage of d′ is that, unlike FA rate, it is unconfounded by possible differences in response bias between groups or conditions. In particular, we were concerned that participants may be able to maintain different criteria for different categories, even when items from all categories are interleaved at test. For example, the presence of systematically differing word properties across semantic versus phonemic/orthographic categories (e.g., word frequency, discussed further below) might differentially affect response criteria for the two conditions.3 Moreover, the propensity to use distinct criteria for different categories might differ between young and older participants, such that any group differences in the hit or FA rate reflect differences in response bias rather than memory. Given these important advantages of d′, we report hit and FA rates for completeness only. To draw conclusions about memory, we focus exclusively on d′, which tests the specific predictions of our theoretical account, conforms to the assumptions of our statistical tests, and provides an index of memory performance that is independent from response bias.
Second, similar to many prior DRM studies, we interleaved the presentation of items from all categories in both study and test phases. This served two main purposes: (1) it decreased the salience of the shared themes within each category, increasing the chance that older adults with putatively compromised semantic/associative representations would fail to apprehend those associations; and (2) it helped avoid any confounding effects of fatigue on performance for words drawn from different category sizes (category size was a critical manipulation in our experimental design, as discussed below). That is, if words from small and large sized categories are tested in a blocked format, poorer performance on larger-sized categories could be induced either by additional interference from the category or by fatigue from being tested on a longer test list. In addition, interleaved presentation may have reduced participants’ tendencies to use different response criteria across different category types at test (Shiffrin et al., 1995), but we are doubtful that it can eliminate this tendency entirely, as noted above.
Third, and most importantly, we manipulated category size, using categories of 2 and 8 items at study, in order to provide a measure of mnemonic interference for each condition that is not subject to item confounds. In a task design comparing perceptual with semantic interference, the experimenter's goal would typically be to observe an effect of interference type, allowing the conclusion that participants are differentially susceptible to the two types of interference (i.e., the build-up of interfering information in memory during the study phase). However, if only one category size is employed (e.g., Ly et al., 2013), any effect of interference type might be equally well explained by differences in inherent properties of words in the two interference types, such as frequency, pre-experimental familiarity, word length, and so on. That is, because words cannot be counterbalanced across the two conditions, their inherent properties are not controlled for when only a single category size is used. Instead, if each condition includes two category sizes (across which inherent word properties remain constant, if the assignment of particular words to category sizes is counterbalanced across participants) and we examine the difference between d′ for 2-item and 8-item categories, we can measure the influence of study-related interference on recognition performance for each condition, in the absence of item confounds. Thus, our key dependent variable was an interference measure – the difference between d′ for 2-item and d′ for 8-item categories. (The importance of using a dependent variable that lies on an interval scale is further illustrated in this context: in order for the proposed difference score to be interpretable, we must be confident that the psychological distance between values of 1 and 2 on the dependent variable is the same as the psychological distance between 2 and 3. This is theoretically true for d′, but not true for probability scores such as hit or FA rates. This issue is the same as the point originally made by Loftus (1978; and expanded on by Wagenmakers et al., 2012) when he argued that a statistical interaction – a difference of differences – obtained using response probabilities as a dependent variable is often uninteresting. Only a subset of such interactions are interpretable, because any non-crossover interaction may be removed or augmented depending on exactly how the response probability non-linearly transforms the psychological variable of interest. Hits and FAs non-linearly transform the underlying “memory strength” variable of interest, whereas d′ provides a linear measure.)
Our predictions can be summarized as follows. The R-H account assumes that stimulus representations become more complex with progression through the visual and auditory processing pathways toward anterior and MTL, giving rise to representations of whole words and the semantic relations between them. It also assumes that the more anterior, more complex representations (both word-level and semantic-level) are compromised in older adults, but intact in the young. The R-H account thus predicts an interaction for recognition memory performance between age (i.e., compromised versus intact anterior representations) and interference type (i.e., perceptual versus semantic). In the case of perceptual interference, low-level features of the words appear repeatedly. When anterior representations are compromised (in older but not young participants), participants must rely on posterior feature-based representations; this reduces the discriminability of old from new items, because even new items contain familiar features. This effect is greater in large than in small perceptual categories, because large categories entail more feature repetition. In the case of semantic interference, the compromised anterior representations (in older but not young) have the opposite effect. The build-up of interference across semantic categories occurs at a higher level of representation – the semantic relations between different words. Now, older participants are paradoxically shielded, because the high-level associations mediating the semantic interference are less robust than in young adults. Therefore, using the d′2-d′8 difference measure, our critical prediction was an interaction between age group and interference type: older adults, relative to young adults, should suffer greater interference for phonemic/orthographic than for semantic categories.
Section snippets
Participants
A total of 120 participants were recruited from the University of Massachusetts-Amherst and the local community, including 40 older adults and 80 young adults. One older adult was excluded before analysis because MATLAB quit unexpectedly during the study phase. A further seven (three older and four younger) subjects were excluded during analysis (see Section 2.4.1. Signal detection model). Of the remaining participants, older participants were between the ages of 60 and 92 years old (M = 71.4,
Neuropsychological test performance
Results of the neuropsychological battery are shown in Table 1. Older adults demonstrated intact group performance on all cognitive tasks, with average performance within the normal range relative to established norms or within established passing cutoff scores (Crum et al., 1993, Folstein et al., 1975, Tombaugh, 2004, Warrington and James, 1991, Wechsler, 2008, Wechsler, 2009). Individually, all older participants included in the analysis passed the study's inclusion criterion of a score
Discussion
The key result was a differential effect of perceptual and semantic interference on recognition memory performance in older compared with younger adults. Specifically, recognition memory in older adults was impaired relatively more by perceptual interference and less by semantic interference than was the case in young adults. This finding was predicted by the R-H account of cognition, applied here for the first time to memory performance in aging.
Several theories have previously been put
Acknowledgements
This work was supported by the National Science Foundation, United States, Award BCS-1554871. The sponsors had no involvement in the execution or reporting of this work. We thank David Huber and Jeff Starns for helpful discussions.
References (109)
- et al.
The human medial temporal lobe processes online representations of complex objects
Neuropsychologia
(2007) - et al.
Intact memory for irrelevant information impairs perception in amnesia
Neuron
(2012) - et al.
Heightened susceptibility to interference in an animal model of amnesia: impairment in encoding, storage, retrieval – or all three?
Neuropsychologia
(2010) - et al.
Semantic versus phonological false recognition in aging and Alzheimer's disease
Brain Cogn.
(2003) - et al.
Specialization within the ventral stream: the case for the visual word form area
NeuroImage
(2004) - et al.
Effects of aging on true and false memory formation: an fMRI study
Neuropsychologia
(2007) - et al.
The influence of perceptual similarity and individual differences on false memories in aging
Neurobiol. Aging
(2018) - et al.
False memories with age: neural and cognitive underpinnings
Neuropsychologia
(2016) - et al.
Mini-mental state: a practical method for grading the cognitive state of patients for the clinician
J. Psychiatr. Res.
(1975) WAIS-III IQs, Horn's theory, and generational changes from young adulthood to old age
Intelligence
(2001)
Gist-based false recognition of pictures in older and younger adults
J. Mem. Lang.
Perirhinal cortex activity during visual object discrimination: an event-related fMRI study
NeuroImage
Perceptual deficits in amnesia: challenging the medial temporal lobe “mnemonic” view
Neuropsychologia
Differing profiles of face and scene discrimination deficits in semantic dementia and Alzheimer's disease
Neuropsychologia
Left perirhinal cortex codes for similarity in meaning between written words: comparison with auditory word input
Neuropsychologia
Semantic memory and the human hippocampus
Neuron
False memories and aging
Trends Cogn. Sci.
Impairment in associative memory in healthy aging is distinct from that in other types of episodic memory
Psychiatry Res.
Trail making test A and B: normative data stratified by age and education
Arch. Clin. Neuropsychol.
Why a superman cannot help a tsunami: activation of grammatical class influences resolution of young and older adults' tip-of-the-tongue states
Psychol. Aging
Semantic memory functioning across the adult life span
Eur. Psychol.
Pattern separation in the human hippocampal CA3 and dentate gyrus
Science
Functional specialization in the human medial temporal lobe
J. Neurosci.
Interactions of memory and perception in amnesia: the figure-ground perspective
Cereb. Cortex
Influence of conceptual knowledge on visual object discrimination: insights from semantic dementia and MTL amnesia
Cereb. Cortex
Perceptual functions of perirhinal cortex in rats: zero-delay object recognition and simultaneous oddity discriminations
J. Neurosci.
Perirhinal cortex resolves feature ambiguity in configural object recognition and perceptual oddity tasks
Learn. Mem.
The Wechsler-Bellevue performance of white males past age 50
J. Gerontol.
Trichotomous processes in early memory development, aging, and neurocognitive impairment: a unified theory
Psychol. Rev.
Similarity of fMRI activity patterns in left perirhinal cortex reflects semantic similarity between words
J. Neurosci.
The organization of visual object representations: a connectionist model of effects of lesions in perirhinal cortex
Eur. J. Neurosci.
Perirhinal cortex resolves feature ambiguity in complex visual discriminations
Eur. J. Neurosci.
Hyper-binding
Psychol. Sci.
Semantic representations in the temporal pole predict false memories
Proc. Natl. Acad. Sci. USA
Feature memory and binding in young and older adults
Mem. Cogn.
Object-specific semantic coding in human perirhinal cortex
J. Neurosci.
Hippocampal system and declarative (relational) memory: summarizing the data from functional neuroimaging studies
Hippocampus
DRC: a dual route cascaded model of visual word recognition and reading aloud
Psychol. Rev.
Computational models of perirhinal cortex function
Hippocampus
Why does brain damage impair memory? A connectionist model of object recognition memory in perirhinal cortex
J. Neurosci.
Population-based norms for the Mini-Mental State Examination by age and educational level
JAMA
On the prediction of occurrence of particular verbal intrusions in immediate recall
J. Exp. Psychol.
A spreading-activation theory of retrieval in sentence production
Psychol. Rev.
Age-related differences in brain activity during true and false memory retrieval
J. Cogn. Neurosci.
Separate effects of word frequency and age of acquisition in recogntion and recall
J. Exp. Psychol.
Remembering” emotional words is based on response bias, not recollection
Psychon. Bull. Rev.
The effects of aging on the neural correlates of subjective and objective recollection
Cereb. Cortex
Two functional components of the hippocampal memory system
Behav. Brain Sci.
Structural brain changes in aging: courses, causes and cognitive consequences
Rev. Neurosci.
Cited by (8)
Are two words recalled or recognised as one? How age-of-acquisition affects memory for compound words
2023, Journal of Memory and Language