Effects of Normal Cognitive Aging on Spoken Word Frequency: Older Adults Exhibit Higher Function Word Frequency and Lower Content Word Frequency than Young Adults

1.1. Cognitive Changes in Healthy Aging

The aging process is associated with normative brain alterations and well-characterized patterns of cognitive change [1-4], which include declines in fluid intelligence [5]. Specifically, cognitive aging in the absence of neurodegenerative conditions (such as Alzheimer’s disease or vascular dementia) is most typically associated with diminished fluid abilities such as executive functioning, episodic memory, and processing speed [6-8], and preservation of crystallized aspects of cognition in the context of aging [2, 5, 9-11]. These crystalized abilities include vocabulary knowledge, mathematical abilities, knowledge of general information, or ability to carry out specific skills, all of which reflect applied knowledge acquired over a lifetime. In many cases, the preservation of crystallized aspects of cognitive function may help to compensate for declines in fluid abilities [12, 13]. For example, having greater domain-specific knowledge can assist in the successful completion of day-to-day activities that would otherwise require problem solving or new processing [14]. In effect, although age-related patterns suggest older adults may process information more slowly, performance for many activities may remain intact through use of strategies learned from experience and practice [15].

1.2. Spontaneous Speech across Healthy Aging

Though these patterns of normative cognitive aging are well-documented [16, 17], much less is known about how they influence changes in spontaneous speech, the unscripted, unrehearsed, spoken utterances that are naturally-occurring during verbal communication [18]. Older adults (OAs) exhibit word-finding difficulty (i.e., less effective lexical retrieval) and dysfluencies related to declines in executive function and processing speed [19-22] in spoken discourse, which can lead to reliance on compensatory communicative strategies (e.g., circum- locutions) [23]. Such changes are reflected in spontaneous speech that is characterized by an increased number of pauses and repetitions (i.e., slower speech rate), production of fewer words overall, difficulty staying on topic, and even impoverished speech [2, 24-27]. However, in contrast to these declines, OAs also simultaneously exhibit preserved vocabulary [28-30], use of syntax [31], and lexical-semantic knowledge [32-34]. This selective preservation of speech function has been attributed to compensatory processes, as a combination of educational attainment, language-related experiences, and learned strategies over time [5, 35, 36] buffer the impact of cognitive decline, suggesting that crystalized abilities may compensate for reduced fluid abilities and obscure early signs of MCI or dementia [37].

1.3. Exploring Lexical-semantic Aspects of Spontaneous Speech: Word Frequency

Recent work has shown that lexical-semantic aspects of spontaneous speech can be quantified using automated detection tools to identify conditions like mild cognitive impairment (MCI) [38-42] and Alzheimer’s disease (AD) [43]. An especially important feature of such speech is word frequency (WF): the degree to which individuals use common, high-frequency words rather than less-common, lower frequency words [25, 44-47]. Referred to as the “Word Frequency Effect,” higher frequency words are generally easier to recognize, access, retrieve, and process than lower frequency words [44, 48-50]. WF is often used as an index in word recognition, naming, and lexical decision paradigms [46, 51-54], reflecting factors such as an individuals’ vocabulary size, language exposure, and mental lexicon [44]. However, these WF metrics are often dated [55] and have rarely been used to investigate speech changes associated with healthy cognitive aging at the word class level. This warrants further exploration of the intersectionality between lexical retrieval and cognitive functioning in aging samples across WF indices derived from regularly updated word corpora such as the Corpus of Contemporary American English (COCA).

Initial studies suggest that WF may be sensitive to age-related cognitive changes [56, 57]. However, the known declines in fluid cognitive abilities and relative stability of crystalized cognitive abilities would predict different trajectories for WF. On one hand, WF relies on crystalized aspects of language (e.g., vocabulary, general knowledge), which grows in early life and becomes relatively stable in later adulthood [30, 44, 58] suggesting that WF would remain stable. Consistent with this possibility, initial studies show that OAs use lower frequency words (i.e., less common, more specific words) in speech compared to younger adults, perhaps due to the accumulation of knowledge over a lifetime [32, 59-61]. In contrast, other studies show OAs may be more likely to use higher frequency words that would require less cognitive burden [56, 57, 62] associated with declines in fluid cognitive abilities like executive function. These findings raise the question whether WF would decline, remain stable, or even improve in the context of normal cognitive aging. Clarifying WF patterns among different word classes may further address this question regarding the impact of fluid and crystallized abilities.

1.4. Specific Word Classes

Although WF appears to be a promising speech feature that may be sensitive to cognitive changes, not all words have the same kinds of WF characteristics. While words may vary in their estimated WF, certain kinds of words are consistently high in frequency. For example, function words (FWs), also known as “closed class” words as there are only so many to choose from, vary in their frequency, but are generally very high frequency. On the other hand, content words (CWs), also known as “open class” words as new learning and exposure grows the lexicon, vary in their frequency, too, but much more widely, with some being very high frequency and others being exceptionally rare. In addition to determining the effect of healthy aging on WF generally, research is also needed to better understand potential changes in specific word subcategories related to normative cognitive aging [63].

FWs serve grammatical and structural purposes, such as linking sentences together (e.g., prepositions, articles, conjunctions, pronouns, auxiliary verbs), and are not based in conceptual semantic representations [64-66]. FWs are more automatically accessed and typically maintained across the course of aging [67], perhaps because they are overlearned and do not require extensive searching within the semantic network. In fact, OAs may rely more heavily on FWs to maintain fluency in conversation and compensate for difficulties retrieving semantically dense CWs, which could lead to the expression of less specific ideas or a shift in topic focus [65]. CWs carry specific information and semantic meaning (e.g., nouns, verbs, adjectives, adverbs) and tap into vocabulary complexity [64]. Past work has shown that OAs may use fewer content words than YAs, which could be partly attributable to cognitive changes that lead to difficulties with processing speed, memory, or lexical retrieval (i.e., naming) from the semantic network [21].

Disruption in speech-related function manifests differently in function vs. content words [68, 69], which may partly reflect their differential distribution patterns throughout the brain as alluded to in early work [66, 70]. For instance, FW use is relatively preserved early in MCI and AD, perhaps owing to its greater localization [66, 71]. In contrast, average CW frequency is distinguishable between healthy controls and persons with MCI and AD [72]. This is consistent with semantically-richer words being broadly distributed in brain areas including temporal regions disrupted early in the course of AD [73, 74]. This notion may also help to explain the differences in FW and CW frequency observed across neurological conditions with differential impact on subcortical vs. cortical structures (e.g., AD vs. Parkinson's disease) [75-77].

Despite its scientific and applied value, no study to date has examined the possible differential effects of normative cognitive aging on FW and CW in spontaneous speech. Such findings have the potential to provide new insight into the neurocognitive processes that mediate word choice during spontaneous speech and the extent to which changes in FW or CW frequency reflect normal vs. pathological cognitive aging.

1.5. Present Study

This study has two primary goals: 1) assess group differences between healthy younger adults (YAs) and older adults (OAs) on variables of spoken WF and 2) examine the extent to which spoken WF is associated with cognitive test performance. Based on established models that delineate more language experiences gathered over the lifespan refine the language system and mental lexicon [78, 79], it was hypothesized that OAs would produce words that are on average lower frequency (i.e., less common) compared to YAs. Given mixed findings in existing work regarding differential properties between FWs and CWs [66, 68, 80], we expected differences to emerge as a function of these WF values within age groups. If confirmed, these findings would suggest that differences in WF between healthy older and younger adults reflect changes in retrieval processes during spontaneous speech due to normative cognitive aging and any significant associations with cognitive test performance would identify possible mechanisms for these changes.

2.1. Participants

A total of 202 YAs (aged 18-23) and 116 community-dwelling OAs (aged 55-88) completed a remotely-administered cognitive test battery as part of a larger project. To be eligible, individuals had to meet the following criteria: age within specified ranges (YAs 18-25, OAs ≥ 55), be able to speak/read English, and have sufficient internet access to complete testing. Exclusion criteria included the self-reported presence of diagnosed memory disorder, stroke, or traumatic brain injury, current or past diagnosis of a severe psychiatric condition (e.g., schizophrenia, bipolar disorder), or reported gender as “other” to permit group comparisons and standardized score calculation as part of the larger project. To promote generalizability, individuals were not excluded on the basis of self-reported elevations in anxiety or depression scores, as there is no strong evidence from previous work that these conditions impact lexical processes and their prevalence is high in otherwise healthy samples [81, 82].

Potential participants for the current study were further excluded if the audio of spoken responses was undetectable due to poor video quality or insufficient data was collected due to experimenter error (e.g., the standardized administration protocol was not followed properly). Consistent with conventions in neuropsychological research, participants were also excluded if they exhibited questionable effort or engagement during testing, as this would lead to cognitive test scores not reflecting their true abilities. Poor cognitive effort was operationalized as falling below threshold on the Reliable Digit Span index (i.e., <7, calculated by summing the longest raw string forwards and backwards of which both trials were correct) [83, 84]. Exclusion due to technical issues, experimental error, and poor effort (total excluded: n=72 YAs; n=46 OAs) produced a final sample of 130 YAs (aged 18-22) and 70 OAs (aged 55-88) to be examined in primary analyses (N=200). Notably, excluded participants did not differ (all p > .05) from those retained in demographic variables (i.e., age, education, sex, race, ethnicity), suggesting these activities did not introduce selection bias.

2.2. Measures

2.3. Assessment of Speech

2.4. Neuropsychological Test Battery

A brief neuropsychological test battery was administered remotely through Microsoft Teams teleconferencing to assess cognitive function using commonly administered measures known to be valid through this modality. Raw, uncorrected test scores were used for the current analyses to facilitate comparison across age groups. These were gathered from the following measures within the following cognitive domains: a) Oral Trail Making Test A (OTMTA [93]) and Digit Span Forward (DSF [94]) (attention); b) Oral Trail Making Test B (OTMTB [93]), Cognitive Estimation Test (CET [95]), Digit Span Backward (DSB [94]), and Digit Span Sequencing (DSS [94]) (executive function); c) Craft Story-21 (CS-21 immediate and delayed verbatim recall [96] (episodic memory) and d) Animal/Vegetable and CFL Letter Fluency [7, 97] (language). MCI status was defined as having two test performances >1 SD below the mean within a single cognitive domain in accordance with past work [98].

2.5. Procedure

YAs self-selected to participate via an online research platform as part of their educational experience and OAs were recruited through a combination of an existing repository of persons who indicated interest in research participation, as well as new sign-ups following a talk given to a local senior living community. After providing consent, participants completed a recorded 60-minute videoconference session on Microsoft Teams consisting of neuropsychological testing, questionnaires, and speech tasks. Recordings of the speech tasks were extracted and reviewed for transcription and analysis. Participants were compensated for their time upon completion of study activities. YAs were provided academic research credits and OAs were compensated with a $25 gift card. All procedures were approved by the Kent State University Institutional Review Board (IRB#: 20-300).

2.6. Analytic Strategy

All analyses were performed in SPSS version 29.0.1.0 (171) and R version 2023.03.0 (Build 386). T tests compared YAs and OAs on WF indices and raw cognitive test performance to identify between-group differences. Pearson correlations determined whether WF indices were negatively associated with raw cognitive test performance among the whole sample and within each group. To mitigate risk of Type I error, p values were corrected for multiple comparisons using the widely used Benjamini-Hochberg method [99, 100]. This method was chosen as it is empirically validated, relatively conservative, and ideal for exploratory analyses with small associations [101], but still reduces risk of Type II error compared to some other corrections for multiple comparisons. For instance, Benjamini-Hochberg methods are often preferred and preserve more statistical power than other options such as Bonferroni methods [102]. Also, Benjamini-Hochberg methods are commonly used in past work examining lexical indices such as WF [42]. Regression analyses explored whether raw cognitive test scores predicted WF. Finally, a post-hoc ANCOVA was conducted to determine whether between-group differences remained significant after controlling for education.

3.1. Preliminary Analyses

Raw neuropsychological test scores were normally distributed with the exception of OTMTB (which was positively skewed and had skewness and kurtosis values greater than 2 and 7, respectively). However, this pattern is consistent with the distributions found in a normative dataset for this task and thus was not transformed [103]. Log transformed WF indices were normally distributed for CW, but not AW and FW. To avoid further modification of the already log-transformed COCA spoken WF indices, these were not transformed. This non-normal distribution is to be expected considering the higher frequency of both common words and FWs in spoken language and was considered during the interpretation of study findings.

3.2. Primary Analyses

3.3. Post-hoc Analyses

4.1. Summary of Study Findings

The current study examined indices of spoken WF and neuropsychological test performance in a sample of healthy younger and older adults. Analyses revealed higher AW and CW frequencies in young adults compared to older adults and these indices were associated with poorer performance on a test of executive function. Follow-up analyses revealed education did not account for all of these findings.

4.2. Differences in WF between YAs and OAs

Based on past work, it was hypothesized that healthy OAs would produce speech containing words that are less common and thus on average lower frequency than YAs [59-61] given the known protective effects of greater life experience, education, and exposure to words that maintain vocabulary, expand the mental lexicon with generally lower frequency words, and underlie WF [30, 104, 105]. This hypothesis was partially supported. Finding higher average WF for AWs and CWs in YAs is consistent with these hypotheses and previous WF research [32], suggesting that lifelong learning and verbal experiences may play a key role in sustaining the mental lexicon and shaping uncommon lexical-semantic decisions in speech.

However, OAs exhibited higher average FW frequency than YAs and the exact explanation for this inconsistent pattern across WF indices is unclear. One possibility is that the higher average FW frequency in OAs is due to increased reliance on more accessible, non-content words due to other speech difficulties that can arise with age (e.g., staying on topic, maintaining the flow of speech) [106]. This pattern could reflect the utilization of FWs as a compensatory mechanism when more specific content words are sought but not easily accessible or retrievable (e.g., FWs function as the “cement” that link CWs together [65]). However, OAs in the current sample used lower frequency CWs and performed better on most neuropsychological tests of language functioning compared to YAs (i.e., vegetable fluency, CFL), arguing against the need for frequent compensatory strategy use. An alternative explanation is that FWs leverage different neural processes than the more semantically-dense CWs. Broadly, FWs appear to be more left-hemisphere lateralized within the language network, whereas CW representations are bilaterally distributed across the neocortex [66], suggesting separability in the lexicon and differential lexical-semantic processing for words not rooted in semantic meaning. Indeed, processing conceptual-semantic information associated with CWs may actually involve the medial temporal lobe structures [107], which are also impacted in OAs. As there are relatively few “closed class” FWs to choose from, it is also possible that higher FW frequency in OAs reflects more developed and efficient grammatical processing abilities [108]. This limited set of higher frequency words may be more readily accessible in OAs with more language exposure, the utilization of these words becoming more efficient in spontaneous speech with aging [79]. This, coupled with their lower CW frequency, suggests OAs likely have richer linguistic systems in which they are more grammatically complete and know rarer words. Further, it is thought that OAs rely more heavily on automatic bottom-up processes during lexical access [109], which could explain why FW frequency is higher if they are often processed more readily than CWs [67]. Future studies are needed to clarify these and other possible explanations for the pattern of age effects found across WF indices.

4.3. Association between Word Frequency and the Cognitive Estimation Test

It was hypothesized that WF would be associated with cognitive performance as language production (e.g., word choice) requires the coordination of multiple cognitive abilities, including executive function and various memory processes [52, 56, 110]. We expected a differential pattern between specific word classes to emerge across age groups based on their distinctions observed in past work [66, 68, 79, 80].

Correlation analyses showed that poorer performance on a measure of reasoning and mental estimation (i.e., CET) was associated with higher average WF (AWs). Regression analyses also showed that CET was the strongest predictor of AW frequency and trended toward significance for CWs, consistent with the notion that the correlation with CWs disappeared after correcting for multiple comparisons. Finding a significant association between CET and WF – but not other measures of executive function – is somewhat unexpected. Past work suggests that CET may not exclusively measure executive functioning, but instead tap into multiple domains of cognition, including crystallized abilities closely tied to life experience [111, 112]. Recent cognitive aging research suggests cumulative experiences develop one’s knowledge structure [13, 113], so it is possible that less life experience may limit knowledge representations (contributing to worse CET performance) and restrict access to less common words in the YA group. No other neuropsychological task in the current study requires engagement with knowledge of this nature, which may explain why this association was not found in any other measure. However, the effects were small, suggesting that external factors (e.g., life experience) may drive this observed relationship with WF [114], rather than the specific mechanisms engaged during the CET. As such, future work must continue to account for individual differences across participants.

Other measures of executive function in the current study included commonly used instruments of working memory (DSB, DSS) and set-shifting (OTMTB) and these abilities have been associated with WF in past work [52]. Given these tasks recruit similar neuroanatomical regions and cognitive processes as the CET, it is unclear why performance on these measures were not also associated with WF. Past work shows that executive functions are highly related with many aspects of the language system that underlie WF [115], so our results may reflect a muddying of these processes. Further, our results may not detect the neural compensation that can allow for the maintenance of overall levels of cognitive function in OAs [116], or their association with lexical retrieval abilities. Future studies are needed to clarify these findings, as well as the better performance on the CET in OAs compared to YAs (which is counter to typical patterns of declining executive function with advancing age) [5, 117] that may relate to increased life experience.

4.4. Association between Word Frequency and Cognitive Performance within Age Groups

Findings for the association between WF and cognitive test performance within age groups were generally consistent with past work [16]. Within younger and older adult groups, higher AW frequency was significantly associated with better performance on a measure of attention (i.e., Digit Span Forward), aligning with the importance of attention systems in word retrieval [118]. In OAs, better performance on a measure of immediate episodic memory was also correlated with higher AW and CW frequency. It is possible that, similar to higher frequency words being more accessible, repeating a story that was just heard also taps into mechanisms involved in retrieving information from readily available storage [52]. Again, these effects were not large (potentially reflecting that WF is mediated by multiple cognitive abilities and no specific cognitive ability drives WF [115]) and will need to be explored in future work among specific age brackets.

4.5. Lack of Association between WF and Performance on Other Cognitive Tests

The lack of association between FWs and performance on any cognitive test across the full sample and within groups could reflect several possibilities. Some past work shows that FWs are minimally related to test performance as their use engages fewer cognitive resources [65, 69]. For instance, if the semantic representations of CWs cannot be accessed efficiently during speech, FWs may instead be sought as a compensatory mechanism because they are much more readily available and do not require additional cognitive resources associated with lexical search and retrieval (i.e., FWs are likely overlearned and more automatically processed) [65, 67, 116]. This automaticity of FW generation may explain why minimal correlations with other cognitive tests arose. However, as OAs still had better verbal fluency and lower AW and CW frequency than YAs, it is also possible that our findings may be a result of the range of FWs, perhaps driven by their inherently higher frequency, structural/grammatical role, and independence from vocabulary complexity, semantic meaning, experience, and language exposure [64, 104].

It is also possible that small effect sizes and the general lack of association between cognitive tests and WF indices in the current study reflect the notion that WF is the result of multiple, complex, and overlapping processes [119] and influenced by factors such as education and life experience [78]. Similarly, methodological choices such as the personally relevant nature of the expository speech task, the neuro- psychological measures used (e.g., remotely administered), and characteristics of the sample (e.g., minimal diversity, highly educated) could also have influenced the findings of the current study. Depending on the type or difficulty level of a speech task, different underlying cognitive abilities may be recruited and elicit various qualities of speech [120, 121]. For example, the speech prompt used in the current study (i.e., asking participants to describe an important person in their lives) may not have been sufficiently cognitively demanding to match the level of cognitive effort needed for performance on neuropsychological testing, thus reducing their observed correlations. It is also unclear to what extent remote administration of the speech task and cognitive measures may have had an impact on responses and performance compared to in-person administration, though existing evidence shows that teleneuropsychology is comparable to face-to-face administration [122].

4.6. The Role of Education

Findings from the current study also suggest that factors other than the cognitive abilities measured on testing are important contributors to WF and may influence word choice. It is largely accepted that older age and greater educational attainment are related to vocabulary size and lexical retrieval processes [104, 123-125]. Past work shows that individuals with less education take longer to make lexical decisions involving lower frequency words [126]. However, analyses from the current study revealed that between group differences in WF indices for AWs and CWs remained after controlling for education levels. This raises the possibility that lifestyle activities (e.g., reading for leisure, completing crossword puzzles, watching documentaries, socializing with other individuals) contribute to the maintenance of spoken language abilities with advancing age, which is consistent with past work showing increased engagement with cognitively stimulating activities is associated with better memory and verbal fluency performance [127]. Replication of these findings in other samples is needed.

4.7. General Discussion

Taken in combination, the current findings provide important new insight that informs work in this area. Overall, better cognitive test performance was associated with lower average AW frequency, consistent with the idea that the more precise utilization of lower frequency words is a marker of better cognitive function, language abilities, and/or greater knowledge which may promote access to those less common words [79, 128]. However, other findings are more difficult to interpret. For example, finding that OAs generated lower frequency AWs and CWs while also performing better on cognitive tests of episodic memory and attention than YAs is consistent with the known contribution of these mental abilities to WF [118, 129], but runs counter to the age-related cognitive decline typically seen in these cognitive abilities [4, 117]. The exact reason for this discrepancy is unclear, but may reflect the bolstering of the semantic network as a result of lifelong learning in OAs [78]. Also new to this literature, the pattern of results for FWs often diverged from that of CWs and AWs, supporting the possibility that production of different types of words may be mediated by different neural substrates and separable in the lexicon. The automaticity of FW generation may explain why minimal correlations with other cognitive tests arose. As effect sizes in the current study are small, additional work exploring differential properties between function and content WF in cognitive aging is warranted. Finally, results suggest that advancing age does not independently limit WF in the absence of pathological cognitive decline, highlighting the role of crystalized cognitive abilities that underlie WF during speech and the possible protective effect of life experience on language function.

When interpreting these results, it is crucial to note contextual factors can also influence the type and quality of spontaneous speech produced. For example, speech samples can be obtained from interviews or conversations (open or semi-structured), as well as traditional tests of verbal fluency, picture description and expository speech tasks [130-132], which may elicit varying WF values. Further, qualities and/or characteristics of speech may present differently in timed vs. flexible tasks; timed tasks (e.g., verbal fluency) rely more heavily on fluid processes (e.g., executive functions), whereas more flexible tasks (e.g., a semi-structured interview) may allow more opportunity to engage with compensatory strategies (e.g., word substitutions, increased pauses) to help navigate the fluid difficulties that arise in older age [133]. Picture description tasks that prompt individuals to talk about the elements of or events depicted in a drawing (e.g., Cookie Theft [134]) have often been used to assess lexical-semantic aspects of spoken language (e.g., word choice, word finding difficulties, empty speech, or repetitions) in healthy YAs and OAs [25, 135], as well as compare the syntactic speech characteristics (e.g., number of clauses, incomplete sentences) of OAs with and without neuro- degenerative pathology [130]. However, speech obtained from picture description tasks may produce less cognitive demand than other tasks (because the picture stimulus is ever-present) and inherently limit the variety of lexical choices in a response compared to other verbal tasks that are more open-ended and un-aided by a stimulus. Responding to open-ended questions or creating narratives in conversation may be more ecologically valid [136], as it better represents natural spontaneous speech without the aid of a visual stimulus. Overall, the association between WF and cognitive processes may be influenced by the nature of the speech elicitation task used to generate the speech sample and this important variable should be more fully investigated in future work.