Participants

A total of 90 healthy, right-handed, monolingual speakers of British English participated in this study. To control for variance in word-finding speed between monolinguals and bi-/multilinguals [32], only monolingual speakers of British English were included in this study. The sample size was based on a-priori power analysis using an online sample size calculator (https://clincalc.com/stats/samplesize.aspx; [33, 34]). The anticipated means, equating to approximately medium to large effect sizes, were based on previous literature on word-finding ability in healthy ageing [20, 21, 35–37]. The outcome variable was set as continuous (means). Power was set at 80% and the alpha level at .05 to obtain a large effect size (Cohen’s f = 0.40). This resulted in a sample size of 30 participants per age group. The participants were grouped into 30 younger (aged 18–30; 23 females), 30 middle-aged (aged 40–55; 22 females), and 30 older adults (aged 65–80; 17 females). The age bands were based on previous literature [e.g., 21, 25, 38, 39] and the 10-year age gap between the groups was introduced to increase the sensitivity to detect age-related effects.

Participants reported no history of or current neurological (e.g., stroke or epilepsy), psychiatric (e.g., schizophrenia or bipolar), speech, or language disorders. All participants had self-reported normal or corrected-to-normal hearing and vision. All participants had a score of 19 or less on the Beck Depression Inventory-II (BDI-II; [40]) and, hence, did not suffer from depression. In addition, all participants had a score of less than 3.6 on the Informant Questionnaire on Cognitive Decline Self-Report (IQCODE-SR; [41]), reflecting absence of cognitive difficulties. Years of education, occupational score, and frequency of leisure activity can be found in Table 1. The study was approved by the Research Ethics Committee of the Faculty of Science and Technology of Lancaster University. Participants gave digital consent to take part in the study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics of the measures to compute cognitive reserve for the periods preceding and coinciding with the COVID-19 pandemic separately.

https://doi.org/10.1371/journal.pone.0280566.t001

Materials

The BDI-II, the short IQCODE-SR, a demographics questionnaire, and the CR questionnaire were completed online via the Qualtrics XM Platform (Qualtrics, Provo, UT). The behavioural tasks for assessing word-finding ability (i.e., picture naming and verbal fluency) and general cognitive functions (i.e., cognitive processing speed, working memory, and inhibitory control) were completed online via the Gorilla Experiment Builder [42].

Cognitive reserve

CR was quantified through a General Activities Questionnaire comprising questions about social/leisure, intellectually stimulating, and productive activities. The questions were adopted from previously published studies [43, 44] and adapted where necessary by merging, adding, or modifying questionnaire items. The frequency of participation was rated on a 4-point scale from never (0) to every day (3). In addition, physical activity was measured through the Global Activity Questionnaire [45]. The CR questionnaire was split into two parts. The first part required participants to report the frequency of their activities for the period preceding the COVID-19 pandemic. That is, the participants were asked to report the frequency of their activities retrospectively. The second part of the CR questionnaire required participants to report the frequency of their activities for the period coinciding with the COVID-19 pandemic. Occupational attainment was obtained by computing the occupational scores using the UK Standard Occupational Classification system [46]. Internal consistency was calculated using Cronbach’s Alpha [47] and was α = .59 for all items used to compute CR. Years of education, occupational attainment, and scores for the General Activities Questionnaire and Global Activity Questionnaire were converted into z-scores within each age group. Following the method by Soldan and colleagues (2013), a fixed composite score for CR per individual was obtained by averaging over these standardised scores [48], for the periods preceding and coinciding with the COVID-19 pandemic separately. By averaging over the z-scores, education, occupational attainment, and general and physical activities were given equal weight when computing CR. This resulted in two CR composite scores per individual, one for the period preceding the COVID-19 pandemic and one for the period coinciding with the pandemic.

Picture naming

Pictures were obtained from the Center for Research in Language International Picture-Naming Project [49], and were carefully selected and balanced for word frequency (actions: M = 3.3, SD = 1.8; objects: M = 3.1, SD = 1.7), age of acquisition of the words (actions: M = 5.2, SD = 1.7; objects: M = 4.8, SD = 1.4), and had high picture-name agreement across individuals (actions: 91%, objects: 99%). Action and object words were not matched for word length. However, this should not affect our results as word length is unlikely to influence naming speed [50]. In addition, we did not directly compare naming speed between object and action naming. The final set contained 79 pictures of actions and 70 pictures of objects. For action naming, the items “vacuum”, “somersault”, “wait”, and “talk” were excluded as these items resulted in low accuracy or name agreement in this study. The item “cut” was excluded as this item was also used in a practice trial. For object naming, the items “football” and “skunk” were excluded as these items resulted in low name agreement across participants.

The picture naming tasks were presented in a blocked manner. That is, participants first completed the object naming task after which they completed the action naming task. Both the object and action picture-naming task started with four practice trials. The participant read the instructions on the computer screen and was instructed to name the pictures as quickly and accurately as possible. The participant first tested their microphone. Next, the participant completed four practice trials, which were recorded and played back to the participant. After the practice trials, the actual experiment started. All trials started with a fixation cross with a 500 ms duration, followed by a blank screen for 700 ms. After this, the picture was presented for 3000 ms. Differences in microphones used by the participants and the lags in audio recordings caused timing variability between participants and trials. Therefore, in order to indicate the start of the picture and measure timing variability, a trigger sound was added. The trigger sound enabled the experimenter to calculate the verbal reaction time more accurately. The stimuli were never repeated to avoid practice effects and were presented in a pseudo-random order, with the trials being randomised for frequency, age of acquisition, number of syllables in the picture names, and name agreement. Picture names starting with the same letter never followed each other. Both accuracy and reaction time were obtained using CheckVocal [51].

Verbal fluency

The verbal fluency task comprised tasks measuring semantic, action, and letter fluency. Participants completed the letter fluency task after the semantic and action fluency task so as to avoid the use of any cued strategies. For semantic fluency, the category prompts were: “animals”, “vehicles”, “fruits and vegetables”, “fluids”, and “writing utensils”. The action fluency task was prompted by “things people can do” and “things you can do to an egg”. Lastly, the letter fluency task included the letter prompts “S”, “M”, and “P”. The verbal fluency prompts were based on previous studies, which also have validated the use of these prompts [20, 21, 52, 53]. The participant had 60 seconds to produce aloud as many words as possible for each of the verbal fluency prompts. The responses were recorded and later transcribed by two transcribers in order to assess the interrater reliability. Composite scores for each of the three verbal fluency tasks were created to reduce the number of multiple comparisons. To obtain a composite score for the semantic, action, and letter fluency tasks separately, scores were standardised within age groups per verbal fluency prompt. The standardised scores were then averaged for each of the three verbal fluency tasks. This resulted in three different verbal fluency composite scores: one for semantic fluency, one for action fluency, and one for letter fluency.

Control tasks

In order to control for the influences of more general processing difficulties in ageing [54, 55], tasks for inhibitory control, working memory, and cognitive processing speed were assessed. The control tasks were counterbalanced and presented visually and, thus, did not require auditory processing. For all three tasks, scores were transformed into z-scores within age groups to create an average score per participant which reflects general cognitive processing ability. To investigate inhibitory control, a spatial Stroop task was used [56]. In this task, an arrow appeared on either the left or right side of the screen and the participant had to press the button corresponding to the direction the arrow was pointing. In incongruent trials (e.g., arrow pointing to the left but appearing on the right side of the screen), the participant had to inhibit the tendency to respond to the location of the arrow to select the correct response (i.e., the direction the arrow pointed to).

Working memory was assessed using a computerised visual adaptation of the Digit Ordering Test [56–58]. The participant was presented visually with a sequence of digits, ranging from 4 to 7 digits. Each digit in the sequence was presented for 700 ms with a 300 ms blank screen between the digits. After the sequence was presented, the participant had to type the sequence of digits in ascending order.

Finally, cognitive processing speed was assessed using the Deary-Liewald task [59], which included both Simple and Choice Reaction Time tasks. In the Simple Reaction Time task, the participant had to press the spacebar each time an X appeared in a square in the middle of the screen. In the Choice Reaction Time task, four empty squares were presented next to each other, and the X could appear in one of the four squares. The participant had to press the key corresponding to the square in which the stimulus appeared. A composite score for cognitive processing speed was created by averaging over the standardised scores of the Simple Reaction Time and Choice Reaction Time task [59]. Computerised versions of these three control tasks are widely used and even implemented in and validated for cognitive testing batteries such as the Cambridge Neuropsychological Test Automated Battery (CANTAB; [60]).

Procedure

The participant was first assessed for their eligibility via email, after which they logged onto the Gorilla Experiment Builder. After providing digital consent, the participant was directed to Qualtrics XM Platform to complete the demographics and CR questionnaires, which included the General Activities Questionnaire and the Global Activity Questionnaire. Next, the participant was redirected to the Gorilla Experiment builder to complete the behavioural tasks. The participant started with either the three control tasks or with the language production tasks, in a counterbalanced order. For the language production tasks, the participant started first with the verbal fluency task, after which they completed the picture-naming tasks. This order was fixed to avoid priming effects of picture naming on verbal fluency performance. At the end of the experiment, the participant was debriefed and thanked for their participation. Each participant completed the CR questionnaires and the behavioural tasks only once.

Statistical analysis

Data pre-processing and data analysis were done in R [61]. For data pre-processing, we used the package tidyverse [62]. For the analysis of the picture-naming data, linear mixed-effect models (LMMs) were employed, using the lme4 package [63]. The outcome variables were 1) reaction time of correct trials for object and action naming separately and 2) accuracy for object and action naming separately. As accuracy is a binomial variable, we used generalised linear mixed-models (GLMMs) as these are well-suited for binomial data [64]. Moreover, because the accuracy data in this study is binary (1 for correct or 0 for incorrect) with a highly skewed distribution (asymmetric distribution where there are more 1’s than 0’s), a binomial family with a complementary log-log (cloglog) link function was chosen [65] with a “Nelder Mead” optimiser [66]. Multiple linear regression was conducted on the verbal fluency data as there were no repeated measures. The outcome variables were 1) number of correctly produced words and 2) average frequency of the correctly produced words.

Following the preregistration, the data was trimmed using z-scores to reduce the probability of Type-II errors [67]. The dependent variables were transformed into z-scores within each age-group and outliers were identified as being above and below 2.5 SD. That is, picture-naming reaction times above or below 2.5 SD were considered outliers and subsequently removed (object naming: 4.4%, action naming: 7.9%). For verbal fluency, values for the number of correctly produced words and average frequency of the produced words above or below 2.5 SD were considered outliers. These were removed from the data before calculating the number of correctly produced words (semantic: 2.6%, letter: 2.9%, action: 1.7%) and average frequency of the produced words (semantic: 2.2%, letter: 2.6%, action: 2.2%). CR composite scores and scores on each of the three control measures were considered outliers if they had a value of 2.5 SD below or above the mean of each age group. Because there were no repeated measures for these scores and because they were important predictors in our analysis, the detected outliers were winsorised at +/- 2.5 SD. This allowed for the scores to remain on the extreme side of the distribution without loss of data. CR scores were missing for two participants and part of the CR scores was missing for three participants. For the control tasks, two participants had missing data for cognitive processing speed and three had missing data for working memory. Missing data was replaced using single imputation, where the mean per age group was imputed. This was done before creating any composite scores.

The predictor variables for all models were Age with three levels (younger adults aged 18–30; middle-aged adults aged 40–55; and older adults aged 65–80), the standardised CR composite score (continuous predictor), and an interaction between Age and CR composite score. The categorical variable Age was coded using Helmert contrasts such that the first contrast reflects the difference between middle-aged and younger adults, and the second contrast reflects the difference between older adults and the mean of the younger and middle-aged adult groups. The package emmeans was used to conduct post-hoc comparisons between age groups, including the Tukey Multiple Comparison test to obtain adjusted p-values [68]. Furthermore, we compared the models with the CR scores for the period preceding and coinciding with the COVID-19 pandemic. This was done to test whether CR scores for the period preceding or coinciding with the COVID-19 pandemic explained current word-finding ability better. Standardised scores for the general cognitive processing composite variable were added to the model as covariates.

For the picture-naming data, random effects for Item ID (i.e., variance caused by between-item variability) and Participant ID (i.e., variance caused by between-subject variability) were included [69]. To justify the inclusion of both random effects in our model, we compared the Akaike Information Criterion (AIC; [70]) values of the model with and without random effects. A lower AIC value indicates a better fit. For action naming reaction time, the AIC for inclusion of both random effects was substantially lower than the null model without random effects (32.9 and 2095.4 respectively). For object naming reaction time, the model which included both random effects resulted in a substantially lower AIC value compared to the null model without random effects (-796.3 and 1403.3 respectively). For action naming accuracy, the AIC for the model with both random effects was lower than the null model without random effects (2771.5 and 2994.4 respectively). For the model with both random effects for object naming accuracy, the AIC was also lower than the null model without random effects (1012.50 and 1070.89 respectively).

To assess goodness of fit, the assumptions of the different statistical models were investigated, and the AIC was compared against a null model. In addition, we reported the conditional and marginal R-squared (R²) for all models using the function “r2_nakagawa” of the package performance [71] for the LMMs and the function “r.squaredGLMM” of the package MuMIn for the GLMMs [72]. Moreover, goodness of fit of the GLMMs was assessed by testing for overdispersion, using “testDispersion” function of the DHARMa package [73]. Concordance and Somer’s D were calculated using the “somers2” function of the Hmisc package to assess the predictive power of the GLMMs [74].

Data availability

The study hypotheses, design, and statistical analyses were preregistered on aspredicted.org (https://aspredicted.org/b6we4.pdf). R code, Cognitive Reserve questionnaires, picture names, and additional online materials are openly available at the project’s GitHub repository (https://github.com/EliseJis/GBG-Online-Study). The dataset will be openly available via Lancaster University’s Pure repository. The dataset’s DOI will be made available after acceptance.

Picture naming reaction time data

The reaction times for action and object naming were analysed separately. Mean reaction times per age group were calculated for accurate items only. The mean reaction time for action naming was 1063.9 ms (SD = 348.3 ms) for younger adults, 1006.6 ms (SD = 312.4 ms) for middle-aged adults, and 1071.2 ms (SD = 326.8 ms) for older adults. The mean reaction time for object naming was 763.6 ms (SD = 183.2 ms), 792.9 ms (SD = 222.4 ms), and 881.0 ms (SD = 246.9 ms) for younger, middle-aged, and older adults, respectively.

The full model for both action and object naming with main predictors (Age and CR), the interaction term (Age*CR) and the covariates (the control tasks) converged without the need for optimisers. The assumptions for linearity and homoscedasticity were both met. To obtain normality of residuals, the outcome variable (reaction time) was log transformed.

Model formulas and results are reported in Tables 2 and 3 for action and object naming respectively. For action naming reaction time, none of the predictors were significant. For object naming reaction time, we found a main effect of age (β = 0.05, t(87.99) = 3.83), such that older adults were significantly slower than both younger (adjusted p = .003) and middle-aged adults (adjusted p = .001). The interaction between CR and Age did not reach significance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Results of the fixed effects for action naming reaction time.

https://doi.org/10.1371/journal.pone.0280566.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Results of the fixed effects for object naming reaction time.

https://doi.org/10.1371/journal.pone.0280566.t003

For action naming, approximately 16.2% and 16.8% of the variance not explained by our fixed effects was explained by the random effects for Participant ID and Item ID respectively. Marginal and conditional R² were calculated as a measure of model fit. Fixed effects explained 1.4% of the variance in the data (marginal R² = .014), whilst 34.0% of the variance was explained by the full model, including random effects (conditional R² = .340). For object naming, 24.3% and 9.2% of the variance not explained by our fixed effects was explained by the random effects for Participant ID and Item ID, respectively. Fixed effects explained 5.9% of the variance in the data (marginal R² = .059), whilst 37.4% of the variance was explained by the full model, including random effects (conditional R² = .374).

Picture naming accuracy data

Accuracy data for object and action naming was analysed separately. Mean accuracy for action naming was 94.7% for younger adults (SD = 22.4%), 95.4% for middle-aged adults (SD = 20.9%), and 93.5% for older adults (SD 24.7%). For object naming, the respective mean accuracies were 99% (SD = 10%), 98.7% (SD = 11.3%), and 97.3% (SD = 16.1%).

Model formulas and the results of both models are reported in Tables 4 and 5 for action and object naming respectively. For both models, there was a main effect for Age, in that older adults were less accurate in action and object naming than middle-aged adults (action naming: β = 0.20, z = 2.92, adjusted p = .010; object naming: β = -0.19, z = 2.33, adjusted p = .05) and less accurate in object naming than younger adults (β = 0.25, z = 2.53, adjusted p = .03). For action naming accuracy, the interaction between Age and CR reached statistical significance in middle-aged adults (see Fig 1). That is, CR had a positive effect on action naming accuracy (β = -0.11, z = -3.38, p < .001). Post-hoc comparisons revealed that the interaction was significant in middle-aged compared to younger adults (adjusted p = .002). For object naming accuracy, we found an interaction effect between Age and CR (β = -0.11, z = -2.37, p = .020). Post-hoc comparisons revealed that the interaction was significant in middle-aged compared to younger adults (adjusted p = .047; see Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The relationship between Cognitive Reserve (CR) and action naming accuracy per age group.

The Effect of CR is Significant in Middle-Aged Adults.

https://doi.org/10.1371/journal.pone.0280566.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The relationship between Cognitive Reserve (CR) and action naming accuracy per age group.

The Effect of CR is Significant in Middle-Aged Adults.

https://doi.org/10.1371/journal.pone.0280566.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Results of the fixed effects for action naming accuracy.

https://doi.org/10.1371/journal.pone.0280566.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Results of the fixed effects for object naming accuracy.

https://doi.org/10.1371/journal.pone.0280566.t005

Both the models for object and action naming showed a good fit for picture-naming accuracy and there was no overdispersion. For action naming, Concordance and Somer’s D were .82 and .65, respectively. For object naming, the respective values were .89 and .78. In addition, marginal and conditional R² for binomial distributions were calculated. Marginal R² was .008 for action naming and .012 for object naming. Conditional R² was .096 for action naming and .078 for object naming. However, logistic models often lead to low R² values [75] and low R² values are not an indication of poor model fit but could indicate a wider spread of data points instead [76, 77].

Verbal fluency data

Table 6 summarises the descriptive statistics of the verbal fluency tasks. For the models with the number of correctly produced words as the outcome variable, General Cognitive Processing was a significant predictor for action fluency only (β = 0.42, t(83) = 2.25, p = .03). None of the other predictors were significant. This was the case for all three verbal fluency variables (i.e., semantic, action, and letter fluency). For semantic fluency, Age was a significant predictor of average frequency of the produced words (β = -0.06, t(83) = 2.31), but the effect disappeared after correcting for multiple comparisons (p = .11). Moreover, none of the models predicted the data well, as was assessed through model fit diagnostics, including a non-significant overall model fit.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Descriptive statistics of the verbal fluency tasks (i.e., semantic, letter, and action fluency) per age group.

https://doi.org/10.1371/journal.pone.0280566.t006

The effect of COVID-19 on CR and subsequent task performance

It is possible that social isolation due to the COVID-19 pandemic led to changes in CR scores, which could impact current word-finding ability. Table 1 provides the descriptive statistics per age group for the CR scores preceding and coinciding the COVID-19 pandemic. To investigate whether changes in CR scores due to the COVID-19 pandemic affected behavioural performance, we performed an exploratory analysis whereby we ran all models again with the CR scores measured during the pandemic. Next, ANOVA for model comparisons and AIC were used to compare the models with CR scores preceding the COVID-19 pandemic to that with the CR scores coinciding with the pandemic for all picture-naming and verbal fluency models separately. The ANOVAs for model comparisons and AIC values revealed no differences between the models with CR scores preceding the COVID-19 pandemic and those with CR scores coinciding the pandemic.

Limitations

The online nature of the study meant that the participants were not in a controlled environment. We tried to control the study through initial screening, providing clear instructions with example pictures, giving the opportunity to test the microphone, and by asking the participants to complete the study in a quiet environment with as few distractions as possible. Future online studies would benefit from a researcher attending the session via a conference call whilst the participant is completing the study to increase participant-researcher interaction. Furthermore, the current study was restricted to a sample of monolingual speakers of British English to reduce the variance in word-finding speed, which can be affected by language background [32]. Hence, the results of this study cannot be generalised to multilingual speakers. Future studies should investigate this question using groups of speakers with varied language backgrounds to expand on the relationships between CR and word-finding abilities across the lifespan. Such future research will increase our understanding of the effects observed in the current study, which has a restricted sample.

In the current study, we transformed the picture-naming reaction times and the number of correctly produced words on the verbal fluency tasks into z-scores within each participant group. We then classified any score above or below 2.5 SD as an outlier, which was subsequently removed. It is theoretically possible that trimming data, including the above-mentioned method, can lead to distortions of the data set [84, 85]. However, recent research suggests that not excluding outliers leads to a greater negative bias [67]. That is, the probability of Type-II errors increases when the data is not trimmed. Using the z-score method mentioned in this study leads to the smallest bias. With increasing age, standard deviations in cognitive tasks increase, meaning more variance within older than younger samples. Hence, trimming the data, especially in older adults, may lead to masking of deviation effects. In the current study, we chose to trim the data within each of the age groups, reducing the chance of masking such deviations. In addition, it is likely that trimming the data using z-scores has led to lower bias than not trimming the data at all (see [83]).

We also note that the study was powered for a medium-large effect size, so any smaller effects sizes may have been missed. Future work could use the data collected here to inform data simulations or power calculations associated with specific statistical tests. For LMMs, Hox proposes to use the rule of thumb of 30 groups with 30 observations [86]. The current study has 30 participants per group and over 70 trials per participants, and, hence, would comply with Hox’s suggestion, so it is likely that the study is adequately powered.

The internal consistency of the CR questionnaire was poor (α = .59), indicating that the items in the questionnaire are multidimensional [87]. The low internal consistency score could be due to few questionnaire items or because they do not measure the same underlying construct. CR is a concept that is believed to be built up by lifestyle choices and lifetime experiences, such as physical activity and education [6]. Hence, CR in itself is a multidimensional construct and this multidimensionality might be reflected through the low Cronbach’s Alpha score. However, there is currently no consensus as to what lifestyle choices and lifetime experiences must be included to compute an individual’s CR level (for a discussion, please see [80]). To better capture CR, the current study included a range of CR proxies, including years of education, frequency of cognitively stimulating activities, and physical activities. Hence, the low Cronbach Alpha score may be a trade-off of using a more comprehensive CR measure.

Moreover, the use of different operating systems, browsers, and differences in microphone setup could have increased the variability in scores between participants. To reduce the timing variability in the audio recordings for the picture-naming tasks, the start of each trial was indicated with a trigger sound. To increase the reliability of the estimate of reaction time, we subtracted the trigger onset from voice onset. However, we could not control for the distance between participant and microphone, and this could have increased reaction time variability across participants and even across trials. We recommend that future online studies provide instructions on the distance between the participant and the computer (e.g., the participant should sit at arm’s length).

Lastly, the current study did not investigate the effect of vocabulary on picture-naming performance in older adults. Previous studies showed that vocabulary positively modulated word-finding ability in older adults [29] and that an increase in vocabulary masks subtle declines in word-finding ability [38, 88]. However, we found age-related declines in word-finding ability in older adults. This participant group was slower in object naming and achieved lower accuracy for action and object naming than younger and middle-aged adults. Because word-finding difficulties in older age were present in the current study, vocabulary could not have masked word-finding difficulties. However, future studies on CR need to clarify the effect of both vocabulary and CR on word finding across the lifespan to provide a clearer picture as to what resources are necessary in maintaining word-finding in older age.