Diffindo! Precise language comprehension in older adulthood revealed by event-related brain potential studies of domain knowledge

Abstract

Knowledge accumulates across the lifespan, yet most research into its effects on language comprehension has focused on young adult college students. This leaves a critical gap in understanding how knowledge impacts comprehension in other populations, including healthy older adults. Older adults bring greater language experience and knowledge to comprehension tasks but also must contend with changes in cognitive factors like processing speed and working memory. We still know little about how older adults use their knowledge in real time and thus how basic comprehension abilities, which seem relatively stable across the adulthood, may arise from the use of different types of representations or processing mechanisms. To address this, we assessed levels of domain-specific knowledge (in this case, of the fictional world of Harry Potter; HP) and used event-related potentials to measure how that knowledge influenced language processing dynamics as older adults (ages 50-81 years) read HP-related sentences. For sentences ending with expected information (true HP “facts”), greater domain knowledge was associated with larger effects of contextual support on the N400, an ERP component linked to semantic access. However, older adults did not show domain knowledge-based N400 modulations to infelicitous but contextually related words, suggesting that they tended to maintain a narrower scope of semantic activation. By contrast, an analysis of trial-level back-sorted according to accuracy on a subsequent sentence-completion task revealed that when older adults could produce the correct sentence completion, domain knowledge did increase the tendency to activate broadly, possibly suggesting that high-knowledge individuals had engaged more actively in anticipatory processing. We suggest that comprehension mechanisms among older adults may capitalize on maturing semantic networks, which become functionally organized for more selective but accurate processing.

Introduction

To make sense of language, individuals must rapidly construct meaning from ongoing input, a process that draws on both fluid cognitive abilities and extensive stores of knowledge in long-term memory (e.g., Kintsch, 1988). Successful language comprehension depends not only on processing speed and working memory, but also on quick and accurate access to accumulated knowledge. Compared to the college-aged samples that are most typically studied in research on language processing, older adults tend to show reductions in fluid abilities such as processing speed (Finkel, Reynolds, McArdle, & Pederson, 2007) and working and episodic memory (Lindenberger, 2014), albeit with considerable variability across individuals (Cabeza et al., 2018; Tucker-Drob, Brandmaier, & Lindenberger, 2019). Nonetheless, although some aspects of language production, such as word retrieval, may decline with age, many aspects of language comprehension appear relatively well maintained into healthy older adulthood (Burke & Shafto, 2008; Thornton & Light, 2006).

Critically, older adults also bring a wealth of domain knowledge and language experience to bear during comprehension, with measures of vocabulary and print exposure increasing across the lifespan and into older adulthood (Payne et al., 2016; Salthouse, 2019; Grolig, Tiffins-Richard, & Schroeder, 2020). These opposing trends—reduced fluid abilities but increased experience and knowledge—may mean that older adults comprehend language using different mechanisms from those that have been well-characterized in the young adult samples that dominate the literature. Indeed, mounting evidence suggests that successful language comprehension in older adulthood often arises via the use of different mechanisms compared to younger adults (e.g., Wlotko & Federmeier, 2012; reviewed in Wlotko et al., 2010, Payne & Silcox, 2019). An important component shaping how older adults comprehend may arise from how they use knowledge in real time, supporting processes such as integration and inference generation (Chin et al., 2015; Miller et al., 2004; Pitts et al., 2022).

Prior work has shown that comprehenders enjoy greater memory and comprehension for texts when they have pre-existing expertise in the content (i.e., domain knowledge; Hultsch & Dixon, 1983; Miller, 2001, 2009; Morrow, Leirer, Fitzsimmons, & Altieri, 1994). Moreover, self-paced reading studies suggest that older adults especially benefit from schematic knowledge, as when passages are preceded by titles rendering them more comprehensible (Miller & Stine-Morrow, 1998; Miller et al., 2006). Older adults also appear to take advantage of knowledge to augment integration processes and enhance inference in understanding text (Chin et al., 2015; Miller et al., 2004; cf. Stine-Morrow & McCall, 2022, for a review) and to understand and remember events (Pitts et al., 2022).

In light of this, understanding how knowledge is used in real-time during comprehension is essential, particularly in older adulthood when knowledge continues to grow. The present study aims to fill this gap, examining how both overall level of domain knowledge, as well as knowledge of specific information referenced by sentences, influences the moment-to-moment dynamics of language processing in older adults. To do this, we take advantage of event-related brain potentials (ERPs), which provide a useful tool for probing the mechanisms at work during real-time comprehension, as they are measures of instantaneous neuroelectric activity that can be measured non-invasively at exactly the moment when new information becomes available. In particular, N400 brain potentials have often been used as subtle probes of knowledge activation during real-time sentence processing. N400s are elicited by meaningful stimuli like words and are exquisitely sensitive to variables related to accessing meaning (reviewed in Kutas & Federmeier, 2000; 2011). They are larger when relatively more activation of lexico-semantic features is needed for comprehension, as in the case of accessing words presented in isolation and/or unsupported contexts compared to words presented in supportive contexts (Van Petten & Kutas, 1990; Van Petten, 2014), as well as for words with greater, compared to smaller, neighborhood densities (Holcomb, Grainger, & O’Rourke, 2002; Laszlo & Federmeier, 2011), and for low- vs. high-frequency words (although this effect can be mitigated by higher-level contextual effects; Van Petten & Kutas, 1990; Dambacher et al., 2006). They are smaller when information has previously been made available, as for words that are highly predictable compared to less expected in context (Kutas & Hillyard, 1984; Federmeier, Wlotko, De Ochoa-Dewald, & Kutas, 2007), as well as for words that are related vs. unrelated to the larger event being described or to particular (i.e., predictable) words that the context has made available to the comprehender (Federmeier & Kutas, 1999; Federmeier, Schul, & Kutas, 2010; Metusalem et al., 2012; Lau et al., 2013).

Studies examining N400 brain potentials reveal that younger and older adults often show subtle differences in the real-time processing of sentential context. Among young adults, N400s to content words in congruent sentences tend to decrease as word position in the sentence increases, reflecting the build-up of contextual information and its facilitation of incoming words; this pattern is qualitatively similar but quantitatively smaller among older adults (Van Petten & Kutas, 1990; Payne & Federmeier, 2018). Moreover, among younger adults, effects of word-level variables (e.g., orthographic neighborhood frequency and lexical frequency) on N400 amplitudes are typically reduced when contextual support is strong, such that lexical variables have a larger impact at the beginning vs. the end of sentences (Van Petten & Kutas, 1990; Van Petten, 1993; Payne, Lee, & Federmeier, 2015). By contrast, among older adults, lexical variables continue to have a relatively larger impact even as context accrues, suggesting that older adults do not rely on context in the same way as younger adults (Payne & Federmeier, 2018; Jongman & Federmeier, 2022). Given these findings, it is unsurprising that older adults seem to benefit somewhat less than younger adults from sentential context, such that N400 effects of sentential/contextual support are slightly weaker and/or delayed for older compared to young adults (Federmeier & Kutas, 2005; Wlotko & Federmeier, 2012). Thus, younger and older adults both tend to make use of contextual information, although there are subtle differences in the timing and extent to which they may use context to facilitate word processing.

Thus, how comprehenders rapidly make use of context is not identical in older and younger adults. This may be due in part to the real-time cognitive and neural dynamics of how older adults access long-term knowledge—a topic which has received limited study. As one way of probing what information is available to individuals as they process language in real time, researchers have recorded N400s to words that are “related anomalies” (linguistically implausible sentence continuations that have some relationship to the message and/or other contextually supported words) compared to (matched) anomalous words without that relationship (e.g., “They wanted to make the hotel look more like a tropical resort, so they planted a row of pines,” versus tulips, when one would expect palms). This approach has been fruitfully used to show that young adults demonstrate facilitated processing for information that is related to a contextually supported word by category/semantic association (e.g., Kutas & Hillyard, 1984; Federmeier & Kutas, 1999) or perceptual similarity (Amsel, DeLong, & Kutas, 2015). Young adults also show processing facilitation for thematically- and event-related knowledge that is linguistically unexpected but related to the discourse (Metusalem et al., 2012; Paczynski & Kuperberg, 2012). For example, after a paragraph discussing hiking, reading “…awed by the beautiful boots…” (where boots is linguistically unexpected but related to hiking events) elicits an N400 of reduced amplitude compared to an unexpected word (e.g., scissors) but of greater amplitude compared to an expected word (e.g., view). Moreover, it has been shown that the extent to which (categorically and/or thematically related) information can be activated in real-time by young adult comprehenders depends on their individual degree of relevant knowledge (Troyer & Kutas, 2020; Troyer, McRae, & Kutas, 2022). In sum, many studies of young adults suggest that they broadly activate related knowledge as they understand sentences, such that rich swaths of knowledge are available and have a rapid impact on word processing.

Overall, fewer studies investigating the time course of access to knowledge have been conducted among older adults. Federmeier and colleagues (2002; 2019) have found that, compared to younger adults, older adults show limited facilitation for category-related anomalies (e.g., the palm-pines-tulip example above), which may suggest that older adults are less likely to broadly activate semantic features of the best continuations relative to the young. Xu and colleagues (2020) also observed age-related differences in the extent to which younger and older adults were sensitive to world-knowledge violations (e.g., ‘Honey is bitter’ vs. ‘Honey is sweet’), with older adults showing reduced N400 effects. These patterns suggest that older adults may access their knowledge structures using different mechanisms compared to younger adults. However, a challenge for understanding when and how older adults activate their knowledge structures during language comprehension – and whether they do so differently from younger adults – is that it is generally difficult to obtain accurate measures of the knowledge that people have relevant to what they are reading. In the present study, we aimed to better understand how older adults use their existing knowledge to comprehend sentences as the text unfolds. To do so, we used a recent approach previously employed with young adults (Troyer & Kutas, 2020), which combined in-depth study of domain knowledge (the fictional world of Harry Potter) with the real-time precision of event-related brain potentials to track knowledge activation as older adults read sentences word by word.

In the present study, we measured older adults’ knowledge of a specific domain (the narrative world of Harry Potter; HP) and used this information to examine the extent to which they quickly used their knowledge, in the moment, as they understood language about the same domain. For the most part, we made use of naturally occurring variability across participants in this popular fictional domain. However, because these adults did not fall in the target demographic (i.e., children, adolescents, and young adults) of the book and movie series when they were initially released, we also conducted book clubs in which participants (re-)read the first three novels in the Harry Potter series. This was done to ensure that we would have at least some participants with recent experience with the book series. To do this, we partnered with a local lifelong learning institution to recruit participants.

In our study, we asked two primary questions. First, among older adults, does an individual’s degree of domain knowledge (about HP) have an influence on their ability to immediately access contextually supported word information? Prior work has found that this is true within young adults (Troyer & Kutas, 2018; Troyer, Urbach, & Kutas, 2019; Troyer, Kutas, Batterink, & McRae, 2024). Here, we examine this question for the first time within an older adult sample. Second, do older adult comprehenders cast a wide net of activation based on their domain knowledge, or, instead, activate knowledge in a manner that is more specified to the context? That is, in an older adult population, does domain knowledge-based activation spread to related (albeit contextually infelicitous) information (as previously observed among younger adults; Troyer & Kutas, 2020; Troyer et al., 2022)?

We quantified domain knowledge both at the level of the participant, comparing those with overall more and less knowledge of HP, and at the level of the trial. To assess trial level knowledge, we gave participants a post-ERP-experiment sentence-completion task, which allowed us back-sort the ERP data according to whether each participant actually knew and could produce from memory the correct ending for a particular sentence. Such trial-level knowledge has previously been shown to influence N400 responses to contextually supported endings among a university sample of young adults (Troyer & Kutas, 2020) but has never been examined among older adults.

To ask both questions, we focused on N400 brain potentials recorded to words completing sentences about HP. For example, readers knowledgeable about Harry Potter would expect a passage like “Professor McGonagall recruits Harry for the Gryffindor Quidditch team. She saw him save Neville’s…” to conclude with remembrall. They might also activate the concept of broomstick, which is involved in the event (Harry is flying on a broomstick when he saves the remembrall), but we would not expect them to activate dog, which is a concept not involved in the event. Our first question addressed whether the size of the N400 contextual support effect (comparing “correct” contextually supported words, like remembrall, to contextually unsupported and unrelated words, like dog) would be graded according to each individual’s overall degree of knowledge about the domain of HP and, at the trial level, whether that effect would depend on whether or not the participant knew (could retrieve) the correct ending. Such a finding would suggest that older adults can prepare even quite specific knowledge in advance so that they are better able to rapidly make sense of contextually supported words.

We also asked whether the size of an N400 related-anomaly effect—N400s to words that are anomalous in the sentence but contextually related, like broomstick in the above example, compared to matched words unrelated to the event being described—would be graded according to HP knowledge. As already discussed, there is evidence that young adults tend to activate information broadly, creating facilitation for the processing of even contextually anomalous words if these are related to predictable upcoming words or to the broader event picked out by the sentence – and those with more knowledge do so to a greater extent (Troyer & Kutas, 2020; Troyer, McRae, & Kutas, 2022). This kind of wide-ranging knowledge activation can allow for flexibility in processing if the discourse does not unfold as expected. However, maintaining information that is not relevant for immediate comprehension goals may also be taxing on cognitive resources, raising the question of whether this kind of processing strategy may be less apt for older adults. Given that we measured knowledge at both the individual participant level (i.e., looking at variability in domain knowledge across participants) as well as at the trial level (allowing us to also examine knowledge differences in specific knowledge at the sentence/fact level within participants of relatively stronger or weaker knowledge), our design is well-suited to pinpoint quite specific effects of knowledge on older adults’ tendencies to activate and maintain broadly related knowledge during sentence comprehension.

In addition to N400s, we were interested in effects on posteriorly distributed post-N400 positivities. These have been found to be modulated by domain knowledge in younger adults (Troyer & Kutas, 2020; Troyer, Kutas, & Federmeier, 2022), but (to our knowledge) late parietal positivities have not frequently been examined among older adults. The exact functional significance of these late positivities is not well-defined. In memory and other non-language studies, late positive potentials (often called late positive components, or LPCs) have been compared to the domain-general P3b ERP component (reviewed in Leckey & Federmeier, 2020; see Contier, Höger, & Rabovsky, 2024) and have been viewed as indices of explicit processes including evaluation of stimuli (reviewed in Polich, 2007) and memory retrieval (reviewed in Rugg & Curran, 2007). In studies of language comprehension, they have been theorized to reflect processes involved in updating working memory (Van Petten & Luka, 2012), semantic integration (Brouwer, Fitz, & Hoeks, 2012), and/or attempts to (re-)engage with difficult-to-integrate words (Brothers, Wlotko, Warnke, & Federmeier, 2020). Thus, late positivities following N400s may involve episodic re-evaluation and possibly memory-related processes, both of which may differ as a function of each individual’s knowledge of the sentence content. Given that episodic memory processes are known to be impacted as a function of cognitive aging (Lindenberger, 2014), we were particularly interested in whether domain knowledge (and/or trial-specific knowledge) might influence processes reflected by these potentials.

Method

Participants

A total of 52 participants from the Urbana-Champaign community aged 50+ (mean age = 65 years, range = [50, 81]) took part in the study for cash. All participants provided informed consent reviewed by the Institutional Review Board at the University of Illinois at Urbana-Champaign. Of these, we excluded one participant from data analysis due to high rate of artifact (> 50%) in the EEG data. Data reported therefore reflect 51 participants (40 self-identified as female; 11 as male). Due to experimenter error, the first block of the experiment (control sentences) was not recorded for two participants, leaving an N of 49 for the control block only. Table 1 summarizes several characteristics of the full sample as well as subgroups of the sample according to their HP knowledge (high knowledge vs. low knowledge based on a median split of the HP Quiz, where the median was 5 out of 10), including measures related to HP knowledge, demographic variables such as age and education, and individual differences measures (described in more detail below). Correlations are presented in Table 2.

Table 1.

Mean, standard deviation, and range for measures of individual differences.

	All participants			High HP group			Low HP group
	Mean	SD	Range	Mean	SD	Range	Mean	SD	Range
HP measures
HP Quiz ***	5.16	(2.16)	[1-10]	7.04	(1.12)	[6-10]	2.95	(1.05)	[1-4]
HP Self Score ***	21.84	(19.00)	[0-62]	33.92	(17.01)	[0-61]	8.00	(13.79)	[0-62]
Average # of times each HP book was read ***	1.0	(1.0)	[0-4+]	1.5	(1.1)	[0-4+]	0.3	(0.5)	[0-1.7]
Average # of times each HP movie was watched§ ***	1.5	(1.4)	[0-4+]	2.4	(1.4)	[0-4+]	0.5	(0.9)	[0-4]
Estimated age at first exposure to HP books/movies	46.7	(11.5)	[26-81]	42.5	(8.7)	[26-62]	54.8	(12.76)	[32-81]
% of HP sentences estimated to be true	48.5	(17.3)	[8-91]	45.9	(16.1)	[8-85]	51.4	(21.0)	[10-91]
% of HP sentences estimated known ***	47.3	(34.7)	[0-95]	65.5	27.5	[0-95]	19.4	(28.2)	[0-91]
Non-HP measures
ART	0.63	(0.23)	[0.09-0.92]	0.63	(0.24)	[0.09-0.92]	0.60	(0.24)	[0.09-0.92]
MRT	0.41	(0.10)	[0.20-0.75]	0.41	(0.10)	[0.20-0.63]	0.40	(0.10)	[0.28-0.75]
Print Exposure	0.04	(0.81)	[−1.87-2.31]	0.03	(0.83)	[−1.89-1.65]	−0.06	(0.83)	[−1.43-2.31]
GKQ *	6.88	(1.81)	[2-9]	7.46	(1.47)	[4-9]	6.20	(2.04)	[2-9]
MoCA	28.06	(1.67)	[24-30]	28.00	(1.59)	[25-30]	27.90	(1.74)	[24-30]
VF	15.88	(4.81)	[8-29]	16.17	(4.96)	[9-29]	15.80	(4.79)	[9-28]
Age **	65.10	(8.23)	[50-81]	63.25	(7.01)	[51-76]	69.10	(7.97)	[50-81]
Education, n (%)	N	%		N	%		N	%
No response	1	(2%)		0	(0%)		1	(5%)
High school graduated	10	(20%)		5	(21%)		4	(20%)
College graduated	11	(22%)		7	(29%)		3	(15%)
Post-graduate degree	29	(57%)		12	(50%)		12	(60%)

Note: Differences between high- and low-HP knowledge groups (defined by the HP quiz score) are indicated as: *** p < .0001, ** p < .01; * p < .05. ART=Author Recognition Test; MRT=Magazine Recognition Test; GKQ=General Knowledge Quiz; MoCA=Montreal Cognitive Assessment; VF=Verbal Fluency.

Table 2.

Intercorrelations (Pearson’s r) among behavioral measures of individual differences.

		2	3	4	5	6	7	8	9
1	HP Quiz	0.68	0.10	0.03	0.09	0.33	0.03	0.09	-0.30
2	HP Self Score		0.08	−0.07	0.04	0.32	−0.01	−0.09	-0.33
3	ART			0.41	0.95	0.68	0.14	0.25	0.14
4	MRT				0.67	0.30	−0.13	0.00	−0.29
5	Print Exposure					0.65	0.07	0.20	0.02
6	GKQ						0.17	0.33	−0.16
7	MoCA							0.42	−0.08
8	VF								0.06
9	Age

Note: r values above an absolute value of 0.28 are significant at α = .05 (highlighted in lighter gray); r values above an absolute value of .36 are significant at α = .01 (highlighted in darker gray). ART=Author Recognition Test; MRT=Magazine Recognition Test; GKQ=General Knowledge Quiz; MoCA=Montreal Cognitive Assessment; VF=Verbal Fluency.

Materials

Sentence materials

Forty (general-topic) control and 156 HP two-sentence passages in which the last word of the second sentence was the critical word were used in the EEG reading experiment (Table 3). All materials were taken from Troyer and Kutas (2020), where a more detailed description of how materials were constructed and normed can be found. As in Troyer and Kutas (2020), we first presented all control materials in a single block, and then the HP sentences were presented across 5 blocks. Prior to the control block, participants were given six practice trials of control sentences; prior to the first HP block, participants were given an additional four practice trials of HP sentences.

Table 3.

Sample experimental stimuli.

Control sentences
Sentence frame		Supported	Unsupported
We had been watching the blue jay for days. The bird laid her eggs in the		nest	yard
The grade-schoolers stood on the corner and waited. They rode to school on the		bus	train
The vampire moved in. He bit his victim on the		neck	shoulder
For his second birthday, Leonard got a stuffed sheep. His mother baked him a big chocolate		cake	pie
The detective arrived at the office. Within minutes he spilled his thermos full of		coffee	tea
Alicia’s first client was a failure. But her second was a		success	triumph
HP sentences
Sentence frame	Supported	Related	Unrelated
There is one main bank in the wizarding world. It is run by	goblins	werewolves	Alohomora
Sybill Trelawney is a Hogwarts professor. She teaches	Divination	Transfiguration	basilisk
There is a branch of magic focused on changing the form of objects. It is called	Transfiguration	Divination	Alley
Ginny opens the Chamber of Secrets and unleashes an ancient monster. It turns out to be a	basilisk	diary	Shunpike
Professor McGonagall recruits Harry for the Gryffindor Quidditch team. She saw him save Neville’s	remembrall	broomstick	dog
Before Harry’s second year, he is rescued by Ron, Fred, and George Weasley. They pick him up in a flying	car	Dursleys	Draco

Control sentences ended either in a supported or unsupported critical word, where supported words were defined as the best completion (mean cloze of best completion = 94%; range = 89-100%). Unsupported words were plausible alternatives that were semantically related to the best completions but never produced during cloze norming.

HP sentences ended in a critical word that came from one of three conditions: supported (i.e., the correct word according to the HP stories), unsupported but related, or unsupported and unrelated. Related endings were words that related closely to the HP-specific events/activities described by the sentence pair and/or that were words from the same semantic category as the supported/correct word, within the HP universe (e.g., “Snitch” and “Quaffle” are both balls used in Quidditch).

Note that in Troyer and Kutas (2020) items were counterbalanced across three separate lists, such that across the experiment each sentence occurred with each of the three possible ending types. Here, instead, the majority (49/51) of participants were run on the same experimental list.¹ Given that (1) we were primarily interested in the relationship between an individual-level variable (HP knowledge) and the brain’s response to each ending type, and (2) there can be substantial between-subjects variability in ERP data among older adults, we reasoned that using the exact same materials across participants would allow us to better understand the individual variability (HP knowledge) in which we were interested. Importantly, prior patterns observed in young adults were not different across lists, being driven instead by domain knowledge (which could not be controlled across list condition).

Measures of individual differences

Harry Potter knowledge quiz.

Our primary estimate of participants’ knowledge about Harry Potter was taken from a trivia-style quiz containing 10 multiple-choice questions (4 choices per question; α = .68). HP quiz score (henceforth, “HP knowledge”) was the correct number of answers out of 10 (z-transformed for statistics in all of our regression analyses). For visualization purposes only, we plot high- versus low-HP-knowledge groups according to a median split based on this Harry Potter knowledge score.

Harry Potter self-report.

Participants also completed a questionnaire about their experience with the Harry Potter books, movies, audiobooks, and other Harry Potter-related activities (details in Troyer & Kutas, 2018). We report scores for the number of times they had read each book and seen each movie, where responses to these items was capped at 4+.

General knowledge quiz.

Participants’ knowledge about general topics including popular culture, science, geography, politics, religion, literature, and art was estimated using a 10-question, multiple-choice general knowledge quiz, shortened from an original 30-question version (GKQ; Troyer & Kutas, 2018). Scores reflect total correct out of 10.

Measures of print exposure.

Print exposure was measured using adaptations of the Author (ART) and Magazine (MRT) recognition tests (Stanovich & West, 1989). Scores reflect the proportion of correct items selected minus incorrect items (foils) selected. We also computed an aggregate print exposure score reflecting an average of ART and MRT scores.

MoCA.

An experimenter administered the Montreal Cognitive Assessment (MoCA; Nasreddine et al., 2005) to each participant. The MoCA is often used as a clinical screening tool for mild cognitive impairment (MCI). MoCA scores can range from 0-30, and scores <=20 can be conservatively interpreted as indicating increased risk of MCI (Waldron-Perrine & Axelrod, 2012). All participants in our sample scored at or above 24 on this measure. We also report the verbal fluency measure from this test, which reflects the total number of words beginning with the letter “b” that participants could report in one minute.

Procedures

Book clubs

Prior to participating in the study, a subset of the participants (N=8) took part in a book club led by the first author. During these book club sessions, which took place over the course of 8 weeks (meeting either weekly or biweekly), participants read and then met to discuss each of the first three books from the Harry Potter series by J.K. Rowling. This ensured that at least some of the participants recruited would have recent experience with the books.

Laboratory tasks

After providing informed consent, participants completed a set of questionnaires including a health history, a handedness questionnaire, and the ART/MRT. An experimenter then administered the MoCA. Participants then underwent set-up for the EEG experiment, which they completed next. After the EEG experiment, participants completed the computerized sentence completion task. They were also asked to estimate (1) approximately what percentage of the HP sentences presented during the ERP experiment had actually been true and (2) approximately what percentage of the HP sentences presented during the ERP experiment they had actually known. They next completed a set of individual difference measures in the following order: Harry Potter self-report, Harry Potter knowledge quiz, general knowledge quiz. Finally, they completed a questionnaire that asked about their thoughts on the experiment, and an experimenter debriefed them on the purpose of the experiment. Participants were given plenty of opportunities for breaks in between tasks. The entire experimental testing session typically took between three to four hours.

EEG experiment

Participants were instructed to remain relaxed and still during the EEG experiment to minimize muscle artifact. They were told they would be reading two-sentence stories across six blocks (first block about general topics, and the remaining five blocks about the world of Harry Potter) for meaning. They were told they would be asked questions about what they read after the EEG recording session and were instructed to focus on understanding the meaning of each sentence pair. They were given practice trials before the first block (the start of the general topic / control trials) and also before the second block (the start of the Harry Potter trials).

During the EEG experiment, participants sat approximately 100 cm in front of a cathode-ray tube monitor. The background of the screen was black and words were presented in white type. Each trial began with a blank screen for two seconds. Then, the first sentence of each pair appeared on the screen until the participant pressed a button to advance to the next sentence. After their button press, a crosshair appeared in the center of the screen for a duration that varied randomly between 900 and 1100 ms. Participants were instructed to focus on the crosshair and not to move their eyes or blink while it was on the screen. The second sentence was then presented one word at a time right above the crosshair. Each word was presented for 200 ms with an interstimulus interval of 300 ms (SOA of 500 ms). After the sentence-final word disappeared, the crosshair stayed on the screen for a duration that randomly varied between 900 and 1100 ms.

Sentence completion task

Following the EEG session, participants completed a modified cloze production / sentence completion task. Participants were asked to provide the best completion (according to the stories of Harry Potter) for the same pairs of sentences about the fictional world of Harry Potter, the final word of which was missing. They were reminded that some of the sentences that they had read had not in fact ended in the correct word; however, they were encouraged to provide the word they actually believed to be correct, and they were told that they could guess if they were not sure. They were also told that it was permissible to leave sentences blank if they felt they did not know the answer. For scoring, participants’ typed responses were manually corrected for spelling.

EEG recording

The electroencephalogram (EEG) was recorded from 26 equally-spaced electrodes attached to an elastic electrode cap (see Ganis, Kutas, & Sereno, 1996, Figure 9, for visualization and comparison with the International 10-20 system) . Six additional electrodes were attached, including two placed lateral to the outer canthus of each eye (for use in monitoring eye movements), two placed under each eye (for use in monitoring blinks), and two placed behind the ears on each mastoid (for use as reference). Prior to recording, electrode impedances were lowered to 5 kΩ. The signal was amplified using a BrainVision BrainAmp DC amplifier set at a bandpass of .02-250 Hz and a sampling rate of 1 kHz. During recording, the online reference was the left mastoid; after recording, all cap electrodes were re-referenced offline to an average of the left and right mastoid.

Data analysis

EEG

Filtering, artifact correction, and epoching.

All electrodes were re-referenced offline to an averaged mastoid reference. In addition, the left horizontal eye channel was re-referenced offline to the right horizontal eye channel to create a bipolar channel to monitor for horizontal eye movements. Data were digitally lowpass filtered at 30 Hz. We extracted single-trial epochs of EEG data from the continuous recordings from 200 ms prior to the onset of a critical word until 900 ms post-critical word. For each electrode, a 200 ms pre-stimulus baseline was subtracted from the waveform. Trials contaminated by eye movements, blinks, muscle activity, blocking, or other artifact were removed from subsequent analysis. This resulted in an exclusion of 9.8% of trials total: 11.8% Control-Supported; 10.6% Control-Unsupported; 9.0% HP-Supported; 9.7% HP-Related, 9.7% HP-Unrelated. Thus, the average number of trials contributing to each condition was 17.6 items (SD = 2.8) for Control-Supported; 17.9 items for Control-Unsupported (SD = 3.1); 47.3 items (SD = 6.0) for HP-Supported; 46.9 items (SD = 6.1) for HP-Related; and 46.9 items (SD = 5.8) for HP-Unrelated. Condition-based ERPs for each participant were created by averaging from 200 ms prior to onset of a critical word until 900 ms post-critical word.

Extraction of time windows.

For statistical analyses, we measured ERPs across two time windows: (1) 300 to 500 ms after critical word onset, corresponding to the typical peak of the N400 and (2) 600 to 900 ms after word onset, during which late positive complex (LPC) activity is often observed. For N400 potentials to critical words in both sentence types, we planned to focus on a centro-parietal region of interest (ROI) where N400 and LPC effects are typically most prominent, averaging across 8 electrodes placed at midline central, midline parietal, left and right mediocentral, left and right mediolateral central, and left and right medial occipital locations (see Federmeier & Kutas, 1999, for a visualization of these electrodes, therein labeled MiCe, LMCe, RMCe, MiPa, LDPa, RDPa, LMOc, and RMOc). Based on results from a study employing the same materials in young adults (Troyer & Kutas 2020), we planned to analyze LPC effects (in the same centro-parietal region) in the late positivity time window only for the HP sentences.

Statistical analysis

Our statistical analyses focused on HP sentences, with a separate analysis conducted for control sentences in the N400 time period only. We examined the effect of contextual support (for both control and HP sentences) and related anomaly (for HP sentences only) on ERP amplitudes in both time periods of interest (N400, LPC).

Our main research questions centered around whether HP knowledge would modulate the neural response to words that were contextually supported and/or related to the HP sentences. In testing this, it was important to also rule out the possibility that such a relationship was due to other individual differences co-occurring in the group. To do this, we also examined (1) whether HP knowledge influenced the neural response to effects of contextual support in the control sentences and (2) whether other measured individual differences better accounted for effects of contextual support and/or related anomaly in the HP sentences. For the latter, we tested models incorporating age, MoCA score, verbal fluency (VF), general knowledge test (GKT) performance, and an aggregate measure of print exposure (combining ART/MRT).

For all ERP statistical analyses, we used linear mixed-effects models (Baayen, Davidson, & Bates, 2008) fit to ERPs from each time window (N400, LPC) over the 8-channel centro-parietal ROI (MiCe, LMCe, RMCe, MiPa, LDPa, RDPa, LMOc, and RMOc). For the majority of analyses (all but the analyses including trial-level accuracy), mixed-effects models were fit to participant-averaged data and included by-participant random intercept terms. All used sum coding for categorical variables and z-transformed values for continuous variables. For statistical inferences about predictors of interest, we conducted model comparison using Chi-square tests on nested models, comparing a full model with a reduced model that omitted the term being tested, using the anova() function in R version 4.0.2. This approach is standard practice in psycholinguistic studies with mixed-effect designs (see Schad et al., 2020). For example, to test interaction effects, we compared models including an interaction (e.g., ending type X HP knowledge) against models with only additive fixed effects (ending type + HP knowledge). To test effects of single fixed effects (e.g., ending type), we compared models using the effect vs. models with only an intercept term (and no fixed effects). Where relevant, we report the full model in tables in the supplementary materials (SM Tables 1-28) with the Satterthwaite option for denominator degrees of freedom for F statistics. We additionally report the models most important for our main points in the main text.

Finally, we also examined the ERPs to HP sentences (N400s and LPCs) as a function of participants’ trial-level responses in the post-ERP-experiment sentence completion task. For these analyses, we fit mixed-effects models incorporating by-participant and by-item random intercept terms to trial-level ERPs.

Results

Behavioral Data

Measures of individual differences

Descriptive statistics of all individual measures are provided in Table 1. Our primary measure of interest was HP knowledge obtained from the 10-question HP trivia quiz. The median score for this measure was 5 (Mean = 5.16, SD = 2.16, Range = [1, 10]). This distribution of scores was similar to that of the younger sample in Troyer and Kutas (2020), who showed a median of 6 (Mean = 6.06, SD = 2.50, Range = [1, 10]).

Table 2 provides correlations between measures of individual differences. Of particular note, our primary measure of interest, HP knowledge (as measured by a trivia quiz), showed a moderate positive correlation with performance on the general knowledge trivia quiz (r = .33, p < .05). This is not surprising given that both measures tap into individuals’ ability to store and recognize factual knowledge. HP knowledge was also moderately negatively correlated with age (r = −.30, p < .05). The negative correlation with age was also observed for the self-reported measure of HP experience (r = −.33, p < .05), but not significant for the general knowledge trivia quiz (r = −.16, n.s.). This pattern suggests that relatively younger older adults were more familiar with and knowledgeable about the world of HP than the oldest older adults in our sample.

Post-ERP study sentence completion task

Across all participants, accuracy (defined as the total proportion of trials for which each participant provided the correct, i.e., supported, word) on the post-ERP study sentence completion task was 37% (range: 0-90%). Accuracy on this task was strongly correlated with participants’ HP quiz scores, r = .82, p < .0001; see Figure 1, which shows a comparison with the results obtained among younger adults in Troyer & Kutas (2020).

Figure 1.

Accuracy on the post-experiment sentence completion task is plotted against HP knowledge (Z-scored) by condition for young adult data from Troyer and Kutas (2020) (left) and the older adult data reported here (right).

Because participants had actually viewed the correct words during the ERP task only for the Supported condition, we expected that accuracy might be boosted specifically for this condition. Figure 1 demonstrates exactly this pattern. A repeated measures ANOVA indicated that ending type had a significant effect on accuracy, F(2, 100) = 114.53, p < .0001, with higher mean accuracy for sentence completions in the Supported condition (Mean = 47%, CI = [39%, 55%]) compared to the Related (Mean = 33%, CI = [26%, 41%]), t(50) = 11.30, p < .0001, or Unrelated (Mean = 29%, CI = [22%, 37%]), t(50) = 11.99, p < .0001, conditions. There was also a significant difference between the Related and Unrelated conditions, t(50) = 4.32, p < .0001. Nested model comparison using linear mixed-effects modeling indicated that there was no interaction between Ending Type and HP Quiz Score, χ²(3) = 1.78, p = 0.41 (SM Table 1), suggesting that having read the correct word or (to a lesser degree) a related word boosted comprehension across all participants compared to having read unrelated words.

ERP Data

Subject-averaged ERP data

Figure 2 shows the grand average ERPs to critical words in HP sentences for all participants across 26 scalp electrodes from 200 ms prior to onset of the critical word to 900 ms post-critical words, which manifest the characteristic pattern of ERPs to visually presented words. Across most electrodes, we observed a negative-going peak around 100 ms (N1), a positive-going peak around 200 ms (P2), and a negative-going potential in the N400 time period (300-500 ms) which was visibly modulated by type of critical word across the back of the head. In general, contextually supported words elicited the most positive-going activity in the N400 time period, and contextually unsupported words elicited negative-going activity in this time period.

Figure 2.

Grand average ERPs across all participants for critical words of each type for HP sentences.

Because we were specifically interested in how domain knowledge would influence processing of words in HP sentence contexts, we also provide whole-head plots for a high-HP-knowledge subgroup (N = 24) and a low-HP-knowledge subgroup (N = 20) (Figure 3).

Figure 3.

Grand average ERPs for High (top) and Low (bottom) HP knowledge subgroups (based on a median split) to critical words in the HP sentences.

N400 time period (300-500 ms).

Figure 4 presents plots of ERPs to critical words in the ROI for both Control and HP sentences in panels a and b, respectively. Ending type had a significant effect on N400 amplitudes (χ²(2) = 38.93, p < .0001; SM Table 2). As expected, we observed significant N400 effects of both contextual support (Supported vs. Unrelated endings; χ²(1) = 25.08, p < .0001; SM Table 3) and related anomaly (Related vs. Unrelated endings; χ²(1) = 3.86, p < .05; SM Table 4). We also observed a significant difference between Supported and Related endings (χ²(1) = 21.38, p < .0001; SM Table 5). In sum, there was a large effect of contextual support (Supported < Unrelated endings) and a small but statistically reliable effect of related anomaly (Related < Unrelated endings).

Figure 4.

ERPs are plotted to critical words from control (a) and HP (b) sentences for an ROI based on an average of 8 centro-parietal scalp electrodes (c). (d) N400 amplitude is plotted against HP knowledge (z-scored) by condition.

We next examined the impact of HP knowledge on these N400 effects of contextual support and related anomaly in HP sentences. To do so, we first tested a model in which HP knowledge interacted with ending type to predict N400 amplitudes against one in which there were only additive effects. The more complex model was supported, suggesting that HP knowledge did modulate overall differences in ending type, χ²(2) = 10.29, p < .01 (SM Table 6). Planned follow-up analyses revealed that HP knowledge modulated the effects of contextual support (Supported vs. Unrelated), χ²(1) = 7.13, p < .01 (SM Table 7), but not of related anomaly (Related vs. Unrelated), χ²(1) = 0.77, p = .379 (SM Table 8; see Figure 4). In other words, the effects of contextual support (i.e., the difference between contextually supported and unrelated items; first vs third panels in Figure 4d) increased with knowledge, but the related anomaly effect did not (second vs third panels in Figure 4d). There was a significant positive correlation between HP knowledge and N400 amplitude for Supported endings, r = .32, p < .05, but not for Related, r = .16, p = .27, or Unrelated, r = .07, p = .60, endings.

To address the possibility that other individual differences were driving the interaction between HP knowledge and ending type for the effect of contextual support on N400 amplitudes, we also tested a model including fixed effects of ending type (supported, unrelated endings only) and our measured individual differences (print exposure, general knowledge quiz, verbal fluency, MoCA, and age) as well as interactions between ending type and each of these measures against a model with all fixed effects and a single interaction term between HP knowledge and ending type. We found that that the simpler model was preferred, χ²(5) = 2.47, p = 0.785 (the complex model is reported in Table 4 for reference). This supports the interpretation that other existing individual differences did not better account for the individual variability in the N400 effect of contextual support compared to HP knowledge alone (at least given the present measures and analyses).

Table 4.

ROI analysis for the N400 time period effect of contextual support; complex model incorporating interaction terms with individual differences of print exposure (ART/MRT), general knowledge test (GKT), age, verbal fluency (VF), and MoCA scores.

	Estimate	SE	df	t-value	Pr(> ∣t∣)
Intercept	2.388	0.309	59.804	7.721	0.000
Ending Type	1.421	0.245	44	5.802	0.000
HP Knowledge	−0.17	0.347	59.804	−0.49	0.626
ART/MRT	−0.37	0.485	59.804	−0.763	0.449
GKT	0.701	0.431	59.804	1.627	0.109
Age	−0.351	0.332	59.804	−1.056	0.295
VF	0.036	0.363	59.804	0.099	0.921
MoCA	0.044	0.349	59.804	0.128	0.899
Ending Type: HP Knowledge	0.719	0.275	44	2.619	0.012
Ending Type: ART/MRT	−0.042	0.384	44	−0.109	0.914
Ending Type:GKT	−0.221	0.341	44	−0.647	0.521
Ending Type:Age	0.109	0.276	44	0.395	0.766
Ending Type:VF	0.28	0.288	44	0.976	0.335
Ending Type:MoCA	0.079	0.263	44	0.299	0.695

Finally, to confirm that the effect of HP knowledge on the N400 effects were specific to domain-specific comprehension, we examined the relationship between HP knowledge and N400 effects of contextual support in our control sentences. Although we observed a numerical difference between Supported and Unsupported endings in the expected direction in the N400 time period, we did not observe a main effect of ending type on N400 amplitudes in the control sentences (χ²(1) = 0.378, p = 0.539) (SM Table 9).² Importantly, however, responses to the control sentences did not interact with HP knowledge (χ²(1) = 0.514, p = 0.473; SM Table 10), suggesting no generalized difference between the groups in their processing of non-HP sentences.

LPC time period (600-900 ms).

LPC responses to the HP sentences also showed a main effect of ending type, χ²(2) = 13.68, p < .01 (SM Table 11). This was driven by more positive-going LPCs for the Unrelated condition compared both to the Supported, χ²(1) = 10.50, p < .01 (SM Table 12), and Related, χ²(1) = 5.07, p < .05 (SM Table 13), conditions. The Related endings elicited a numerically more positive-going LPC compared to the Supported endings, but the difference did not reach significance, χ²(1) = 3.02, p = .08 (SM Table 14).

There was no significant interaction between HP knowledge and ending type in the LPC time period, χ²(2) = 2.169, p = 0.340 (SM Table 15). However, HP knowledge did have a main effect overall, χ²(1) = 3.946, p < .05 (SM Table 16), with more positive-going LPCs for individuals with relatively higher compared to lower HP knowledge (compare ERPs for high vs. low-knowledge subgroups in Figure 4b).

Single-trial ERP data

We sorted ERP trials based on participants’ responses during the post-ERP sentence completion task according to whether they correctly produced the appropriate (i.e., supported) ending to each sentence, or not. We expected that accuracy on this task would relate to N400 amplitudes for supported sentences, with a reduction for correct compared to incorrect trials. Based on findings from Troyer & Kutas (2020), we further expected no relationship between accuracy and N400 responses to the other two types of sentences and no effect of accuracy on LPC amplitudes. However, if older adults show additional benefits from trial-level, specific knowledge, it is possible that we might observe a different pattern of results, possibly facilitating N400s for contextually related (though anomalous) words and/or influencing the size of LPCs (which may reflect more explicit/episodic memory processes) for supported or related words.

N400 time period (300-500 ms).

We first asked whether there was any effect of accuracy on N400 amplitudes and whether this was modulated by ending type. In the centro-parietal ROI examined, we found an overall effect of accuracy on N400 amplitudes, χ²(2) = 215.67, p < .0001 (full model in SM Table 17), indicating that N400s were more positive for correct responses than for incorrect responses. This main effect was moderated by ending type in the two-way interaction, χ²(2) = 202.65, p < .0001 (SM Table 18). Planned follow-up tests revealed that this interaction was driven by the significant effect of accuracy on N400s to supported words, χ²(1) = 118.44, p < .0001 (SM Table 19), with reduced N400s to correct compared to incorrect trials, but not on N400s to the related or unrelated words.

We next asked whether this pattern was additionally modulated by domain knowledge (Figure 5). To test this, we fit a model with a three-way interaction of sentence-completion accuracy, sentence ending type, and HP knowledge (as well as all fixed effects of each and all two-way interactions).³ When tested against a nested model that dropped the three-way interaction, the more complex model was preferred, suggesting that HP knowledge did have an influence on the overall pattern of results, χ²(2) = 90.051, p < .0001 (full model reported in SM Table 20). Follow-up comparisons indicated that for both supported, χ²(1) = 13.225, p < .0001 (SM Table 21), and related, χ²(1) = 17.112, p < .0001 (SM Table 22), words (but not unrelated words), HP knowledge led to larger effects of accuracy on N400 amplitudes.

Figure 5.

ERPs averaged over single trials in the centroparietal ROI are plotted by condition and by accuracy on the post-experiment sentence completion task and by HP knowledge subgroup (based on a median split). Because these ERPs are computed over single trials, and because there are substantial differences in the number of trials included in the “correct” vs. “incorrect” bins across participants, we plot error bars that reflect a 95% confidence interval computed over single trials.

LPC time period (600-900 ms).

We also asked whether accuracy on the sentence-completion task influenced late positive components. We found that there was an overall effect of accuracy on LPCs, with more positive-going potentials for correct compared to incorrect words, χ²(2) = 14.718, p < .001 (SM Table 23). In addition, accuracy interacted with condition, χ²(2) = 28.957, p < .0001 (SM Table 24). Follow-up comparisons showed that the overall effects of accuracy on LPCs were driven by the Supported condition, χ²(1) = 12.679, p < .001 (SM Table 25), and were not significant for Related or Unrelated words.

Moreover, this pattern was modulated by HP knowledge, χ²(2) = 16.876, p < .001 (SM Table 26). Follow-up comparisons indicated that this followed a similar pattern as in the N400 window, with greater domain knowledge increasing the influence of accuracy on LPCs for both Supported, χ²(1) = 19.712, p < .0001 (SM Table 27), and Related, χ²(1) = 7.652, p < .001 (SM Table 28), but not Unrelated words.⁴ Thus, among our older adult sample, domain knowledge did seem to lead to broader activation that facilitated related but anomalous words—but only when the older adults knew and could later readily produce the correct/appropriate sentence completion.

Post-hoc trial-level analysis of linguistic plausibility

Although domain knowledge did not facilitate the processing of words constituting the related anomalies, which were typically linguistically inappropriate but thematically related to the sentence context and/or semantically related to the correct sentence completions, there was, however, a small but significant related anomaly effect, with N400 amplitudes to the related words being slightly reduced compared to those of the unrelated words. Given that this effect was not influenced by domain knowledge, we wondered whether it might be driven by more general properties of word plausibility based on statistics of the English language, and not the fact that these words were related via events and categories from the Harry Potter novels, per se. For example, consider this item, completed with a contextually supported word: “Sybill Trelawney is a Hogwarts professor. She teaches Divination.” Participants might instead read the word “Transfiguration” as the related anomaly vs. “Basilisk” as the unrelated anomaly. In this case, the supported and related words (“Divination” and “Transfiguration”) are much more similar linguistically compared to the unrelated continuation (“Basilisk”). Although it was not our intention for related vs. unrelated words to differ in this way, we wondered if differences in this level of linguistic “plausibility” might have contributed to ease of semantic access to the critical words, possibly leading to the observed related-anomaly effects.

To examine this possibility, we conducted a behavioral plausibility ratings study in a separate group of older adult participants. We asked this group to rate, based on their knowledge of the English language alone (and not of the fictional world of HP), the linguistic plausibility of each completion for the HP sentences (more details provided in Supplementary Materials, Effects of plausibility on N400 brain potentials). As would be expected based on the design of our materials, ratings of linguistic plausibility were highest for correct words (i.e., those in the Supported condition). Critically, ratings of linguistic plausibility were also higher for words used in the Related condition than for those used in the Unrelated condition. This suggests that the observed effects of related anomaly on N400 amplitudes may have been driven by subtle differences in the linguistic appropriateness of the words in the related and unrelated conditions, respectively, and not by HP domain knowledge.

Discussion

In the present study, we investigated when and to what extent domain knowledge influences access to meaning during language comprehension among a group of older adults. We examined the amplitude of brain potentials associated with the retrieval of lexico-semantic information (N400s) and subsequent cognitive processes related to stimulus evaluation and memory mechanisms (post-N400 LPCs) as a function of sentence-final words’ contextual support and individual-level domain knowledge relating to the content of the sentences.

Accessing contextually supported words

We found that greater domain knowledge was associated with greater reduction in the amplitude of N400 responses to words that were contextually supported (versus contextually unsupported words). As such, our study is the first to show that individual differences in domain knowledge can have immediate consequences on the initial stages of semantic access to word meaning within a sample of older adults. This finding is important for several reasons.

Perhaps most importantly, our findings suggest that older adults are able to rapidly make sense of sentential contexts (describing a world of fiction) through the lens of their domain knowledge. This pattern of results are consistent with previous work directly comparing older and younger adults finding that both groups use sentential context relatively quickly to facilitate the processing of contextually supported words, albeit perhaps less rapidly and/or to a somewhat smaller extent than younger adults do (reviewed in Payne & Silcox, 2019; Wlotko et al., 2010).

Our findings additionally show that the knowledge immediately available to older adults includes that of a domain with which they cannot have had a lifetime of experience. Indeed, of the 44 (out of 52) older adults in our sample who reported prior experience with the Harry Potter books (released between 1997-2007) and/or movies (released between 2001-2011), the median estimated age of first exposure was 45 years. Moreover, the domain in question was a fictional world, one which is likely not encountered in day-to-day life. This suggests that even fictional information gleaned in mid-to-late adulthood can be fully integrated into semantic memory and made available immediately as a function of individuals’ degree of knowledge for use during sentence comprehension.

In addition to examining levels of domain knowledge at the participant level, we used a post-ERP sentence completion task to examine the impact of knowledge of specific facts at the trial level. Unsurprisingly, domain knowledge increased the number of items that people could correctly complete. The ability to complete a sentence with the appropriate/correct completion suggests that the knowledge is available in memory with high fidelity. Although we cannot know for certain that this information was available with exactly the same fidelity (and the requisite speed) in the ERP study itself, the objective measure of sentence-completion accuracy is a good approximation of which fictional “facts” individuals knew well enough to recall, and thus which sentence completions were most readily available in the moment (and perhaps even prior to their receipt) during the ERP study for each of the sentence ending types. Indeed, our analysis using accuracy on the sentence completion task to back-sort ERP trials into bins of known vs. unknown trials revealed that, across participants, N400 amplitudes to contextually supported words were reduced when participants could provide the correct word on the completion task, compared to cases in which they provided an incorrect or no response.

This pattern of results was similar to that previously observed among younger adults in Troyer & Kutas (2020). However, in that dataset, domain knowledge did not additionally influence N400 patterns once trial-level effects of the ability to produce the correct sentence completion were accounted for. In contrast, among older adults, individuals with more domain knowledge were not only able to produce correct responses to more items, but also showed enhanced N400 facilitation for those correct items compared to items correctly produced by those with lower overall HP knowledge. One possible explanation for this discrepancy is that older adults may encounter retrieval difficulty, possibly leading to tip-of-the-tongue states in which retrieving specific words is impaired (Burke & Light, 1981; Burke, MacKay, Worthley, & Wade, 1991). As such, even when older adults actually know the correct word, they might not be able to retrieve and produce it. Thus, high-knowledge older adults may possess more knowledge than might be indicated behaviorally, but which may be detectable using the more subtle measure of ERPs. Whatever the precise mechanism, these data support the interpretation that older adults with relatively greater domain knowledge can retrieve more facts and activate richer information from items correctly retrieved than could those with weaker domain knowledge.

Accessing contextually related (linguistically inappropriate) words

Of particular interest in our study was whether older adults would show evidence of routine, broad activation of information related to the context as a function of their domain knowledge. We probed for this using related anomalies – linguistically implausible completions that were semantically associated with the context. Overall, older adults did show a (small) related anomaly effect on N400 potentials, with the brain’s response to related anomalies being slightly more positive-going than that elicited by the unrelated anomalies. However, this effect was not significantly influenced by domain knowledge, and a post-hoc examination revealed that it was likely due to subtle differences between the Related and Unrelated endings in overall linguistic plausibility (i.e., without respect to the world of HP). In contrast to this pattern, many studies, including prior work looking at effects of HP domain knowledge, have noted that younger adults seem to show wide-ranging semantic activation during sentence comprehension, bringing online rich swaths of information thematically and/or semantically related to the sentence context even when that information is linguistically implausible and not the correct continuation of the sentence. The present findings are consistent with prior work probing categorical knowledge activation among older adults, who on average do not show related anomaly effects (Federmeier et al., 2002; Federmeier & Kutas, 2019). Similarly, we find that older adults’ domain knowledge-based activations seem more narrowly focused on information that best fits the unfolding language context.

Although we did not observe an overall relationship between domain knowledge and related anomaly effects at the subject level, a somewhat different pattern appeared when we examined trial-level ERPs, back-sorted by participants’ item-level accuracy on the post-ERP-experiment sentence completion task. Older adults with relatively greater domain knowledge showed a reliably larger reduction in N400s to related anomalies for sentences that they later completed accurately compared to those which they could not accurately complete. Thus, once we were able to pinpoint item-level knowledge in a more specific way, our findings suggest that domain knowledge did affect how older adults processed related anomalies. For those with high domain knowledge overall, semantic activation was broader, extending to thematically- and conceptually-related content that facilitated processing of the related anomalies, for particularly strong and high-fidelity knowledge (that better afforded accurate retrieval).

Moreover, accuracy on the post-ERP-experiment sentence completion task was slightly greater for sentences that had been continued with Related compared to Unrelated endings. This suggests that older adults may have benefited from having read a related (compared to unrelated) word when it came to accurately retrieving (and producing) the correct word from memory. In other words, having read a related word during the ERP session may have primed and facilitated later retrieval of the correct word.

One interpretation of these patterns is that older adults adopt different processing strategies when they are more vs. less likely to have strong knowledge of the information in the sentence contents. One possibility, consistent with theories of language prediction as involving mechanisms used in language production (Federmeier, 2007; Pickering & Garrod, 2013; Dell & Chang, 2014), is that the ability to (rapidly) produce a sentence continuation (as in our sentence-completion task) allows the same individuals to potentially predict that information during a comprehension task and thus more deeply engage with semantically related information, such as that probed by the related anomalies in the present study. Although older adults have generally been shown to engage in active predictive processes to a lesser degree than younger adults (reviewed in Wlotko & Federmeier, 2012; Payne & Silcox, 2019; cf. Dave, et al., 2018; Cheimariou & Morett, 2024), recent behavioral work suggests that prediction in older adults may fluctuate on a trial-by-trial basis (Jongman et al., 2023). Our findings suggest that rich domain knowledge may permit some information to be retrieved rapidly and with high fidelity, contributing to this use of predictive processes.

A related possibility, consistent with theories of prediction as a learning mechanism (Dell & Chang, 2014; Fitz & Chang, 2019), is that older adults may engage in predictive mechanisms opportunistically in moments where they stand to learn new information (see discussion in Ryskin & Nieuwland, 2023). Because older adults already have a lifetime of experience with language, most language comprehension may not seem likely to afford such opportunities; by contrast, our sentence materials describing the fictional world of Harry Potter might have yielded just the right environment for knowledgeable individuals to verify extant (but relatively less stable) knowledge of a fictional world (an opportunity that may not be afforded for language materials describing everyday events), leading them to engage more actively in moments where they wished to check their existing knowledge.

Effects on late positive components

In our study, across all individuals (regardless of their domain knowledge), LPCs were largest for unrelated (compared to contextually supported and/or related) words completing fictional facts about Harry Potter. These findings are consistent with those of many sentence comprehension studies testing younger adults in which similar, posteriorly distributed LPCs have often (but not always) been observed to unexpected words that are anomalous in sentence contexts (Thornhill & Van Petten, 2012; DeLong, Quante, & Kutas, 2014; reviewed in Van Petten & Luka, 2012). The present study extends the above research, conducted in younger adults, to show that older adults also show enhanced posterior positivities to anomalous (compared to contextually supported or related) words, even when they are all presented in fictional contexts.

We also found that older adults with relatively greater (compared to less) domain knowledge showed larger LPCs to words completing sentences about that domain—regardless of the ending type. In other words, domain knowledge was associated not only with an increase in LPCs to the anomalous (unsupported/unrelated) words, but also to words that were contextually supported and/or related. We tentatively speculate that this may occur in cases where individuals possess the relevant knowledge to (potentially successfully) comprehend an incoming word, but wherein that information is not immediately available to them and requires a further search through memory. Since older adults may experience difficulty in immediately accessing all relevant information (Ramscar, Hendrix, Shaoul, Milin, & Baayen, 2014), they may selectively invest in deeper consideration of information—leading to an enhancement of memory search processes reflected by LPCs—only when relevant knowledge is highly available (cf. Miller et al, 2004; Chin et al., 2015).

Age-related differences in the structure of (and access to) knowledge

Our findings show that, on the whole, older adults can readily use their domain knowledge (when it is available) to process appropriate incoming words, but that this knowledge activation does not typically extend to linguistically anomalous thematically and semantically related words. Given that processing by younger adults has typically shown patterns of broad activation, this suggests differences in either the structure of older adults’ semantic networks or the dynamics by which that information is accessed.

Recent work investigating age-related changes in semantic networks (typically constructed from verbal fluency tasks and/or semantic relatedness tasks) has argued that older adults have more modular networks, which are less flexible compared to those of younger adults, leading to longer path lengths on average between words/concepts (Cosgrove et al., 2021; Cosgrove, Beaty, Diaz, & Kenett, 2023; Wulff, Hills, & Mata, 2022). These network analyses suggest that, across the lifespan, connectivity becomes less uniform, reflecting strengthening of certain connections and weakening of others. Such an increase in modularity could have benefits given age-related declines in fluid abilities (Park et al., 2002; Park & Gutchess, 2006), including the structural integrity of fronto-temporal regions involved in the rapid semantic access required for language processing (Kemmotsu et al., 2012). With more limited fluid resources available, older adults might adopt strategies that do not involve a wide scope of activation of semantically related content. It seems likely to us that such a maturational pattern of changes in the organization of semantic networks to be more modular in older adulthood might not reflect any deficiency; rather, it might be an adaptive response enabling greater functioning given the dual challenges of having larger amounts of knowledge to organize and relatively more limited fluid abilities.

Our results are thus in line with the conclusion that older adults, who possess greater knowledge and weaker fluid abilities compared to younger adults, may balance their strengths and weaknesses to activate semantic content judiciously (cf. compensatory processing; Baltes, 1997; Park & Reuter-Lorenz, 2009). In other words, they may limit parallel activation of multiple types of conceptually related knowledge in favor of prioritizing the most relevant or appropriate semantic content. The benefit of such an approach is a reduced need to deploy costly working memory resources to keep different types of information activated simultaneously. The downside is reduced flexibility in terms of processing less expected input. Indeed, a great deal of evidence suggests that older adults (especially those with relatively weaker verbal fluency abilities) may experience difficulty quickly adjusting to less typical linguistic content (e.g., Federmeier, Schul, & Kutas, 2010; Wlotko, Federmeier, & Kutas, 2012).

At a broad level, our findings are consistent with developmental theories of neural plasticity highlighting that in early exploratory phases, neural systems are less stable and processing is more variable, with more stable neural ensembles and processes as the system matures (e.g., Lindenberger & Lövdén, 2019). Extrapolating to knowledge, younger adults’ semantic networks may be more malleable and open to neural changes due to plasticity whereas older adults’ semantic networks are at a later maturational stage and are therefore more stable. Applied to language comprehension, this might translate well into a theory where young adults engage in wide-range knowledge activation, including information that might not be immediately relevant for the sentential content but that could be useful to have active for understanding downstream language in the discourse (see discussion in Metusalem et al., 2012) and/or for learning about the world through language. Older adults, by contrast, may benefit less from activating such information, either due to their extensive experience and world knowledge, limitations on fluid abilities, or both.

In sum, it seems that as older adults understand incoming words in a sentence, they tend to narrowly and precisely activate their semantic networks. Our findings suggest that this activation is most typically restricted to information that is linguistically consistent with the immediately preceding sentence context, given its interpretation through the lens of their knowledge. This strategy might allow older adults to focus their resources more fully on semantic content in the here-and-now, although at the expense of pre-activating contextually related information that could prove useful for processing additional linguistic content in the future.

Limitations and future directions

A major limitation of the present study is that we did not have good control over when the individuals first acquired their domain knowledge of Harry Potter. Although we attempted to recruit individuals who would be brand-new to the Harry Potter books using small book clubs, and therefore be able to compare these individuals with those who had read the material long ago, only two participants from this group of participants had never before read the Harry Potter books. In general, those with lower estimated domain knowledge also tended to have been older when they first acquired the knowledge (see Table 2). This limited our ability to draw any strong conclusions about how knowledge organization might change during adulthood, over time. It is possible that recency of acquisition might influence the breadth of knowledge activation during language comprehension, but this is something we did not control for and thus could not examine in the present study.

In addition, we could not control for all individual differences that were likely to (and in some cases, did) co-occur with individual variability in domain knowledge. Notably, individuals with greater knowledge of Harry Potter were on average younger than those with less knowledge. However, we did not observe a relationship between domain knowledge and related anomaly effects even among the (relatively) younger half of the older adults tested, who (on average) had slightly greater domain knowledge compared to the older half of the participants. Therefore, the fact that age was correlated with HP knowledge seems unlikely to impact the conclusions we draw from our results. In addition, performance on a general knowledge quiz was also moderately correlated with performance on our measure of Harry Potter knowledge, perhaps suggesting that the general ability to quickly retrieve information from long-term memory positively impacted participants’ HP knowledge scores to an extent. However, ERP analyses which included individual differences beyond HP knowledge did not better explain our findings. This strongly suggests that it was indeed the HP knowledge and not other existing individual differences that drove the HP knowledge-based ERP effects described here.

In addition, another limitation of the present approach is that, given the individual-differences/domain knowledge approach we employ here, it remains difficult to directly compare findings among younger and older adults. The younger adults who took part in previous ERP language studies using the HP domain were exposed to the fictional world of Harry Potter from childhood; by contrast, the older adults in the present study were (on average) first exposed to this fictional world in middle age. Given this, it is difficult to directly compare previous results from the younger adult study with the present findings.⁵ This makes it impossible to make strong causal claims about age-related differences between the two groups. This difficulty would remain present even if we were to concurrently test individuals across a range of ages using the present HP domain paradigm. Moreover, even for knowledge domains that do not have this “age of acquisition” property (e.g., a domain that both younger and older adults would have learned at a similar point in development), the domain-knowledge approach is limited by the fact that this approach requires us to look at a domain-knowledge variable that is only defined across subjects.

To combat these limitations, in future work, we aim to examine how knowledge is used during language comprehension from a broad, lifespan perspective by examining within-participant differences in domain knowledge (e.g., by investigating multiple domains in which individuals have relatively stronger vs. weaker knowledge), in samples that span adulthood. We also aim to focus on domains wherein individuals are likely to have begun to gain the knowledge at a similar age (e.g., music, cooking, etc.). This will allow us to better disentangle the contributions of biological/maturational factors associated with aging from within-individual differences resulting from knowledge and expertise.

Conclusions

Our study investigated the timing and extent to which older adults use their rich knowledge in the moment during sentence reading. We found that variation in the degree of older adults’ relevant domain knowledge had an influence on the earliest stages of semantic access to contextually supported words, inferred from N400 effects. However, older adults in the present study seemed to focus on those words/concepts that fit the currently unfolding language, such that domain knowledge did not typically facilitate the activation and processing of related but anomalous words. However, in cases where knowledge was most strongly available (inferred from a sentence-completion production task), having more domain knowledge did seem to facilitate the processing of linguistically inappropriate but conceptually related knowledge.

These findings highlight the variation in language comprehension processes that emerge as a function of differences in relevant knowledge among older adults. Our results are consistent with the notion that it is a fallacy to view development as complete by early adulthood (Baltes, Staudinger, & Lindenberger, 1999) and emphasize the importance of characterizing the mechanisms and representations underlying language and cognition across the adult lifespan in future research. Moreover, we suggest that differences in processing patterns among age groups should not necessarily be assumed to reflect “deficits” due to aging. Instead, they may reflect functional changes in cognitive systems that adapt to meet the changing experience and goals of individuals as they age.