Highlights
-
• Bilingualism’s cognitive benefits remain debated despite rigorous research.
-
• Existing work on bilingual effects assumes that task performance is stable.
-
• We investigate bilingual effects on time-sensitive performance measures.
-
• Different measures reveal distinct links between language experience and cognition.
-
• Conventional methods may contribute to inconsistent findings in bilingual research.
1. Introduction
Growing interest and heated debate surround the assertion that bilingualism, the use of more than one language, confers non-linguistic benefits on cognitive function (Antoniou, Reference Antoniou2019; Bialystok, Reference Bialystok2024; Paap, Reference Paap2019). While non-linguistic bilingual “advantages,” referred to more generally as bilingual effects (Barac & Bialystok, Reference Barac and Bialystok2012; Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023), have been documented since at least the 1960s (Peal & Lambert, Reference Peal and Lambert1962), it was not until the early 1990s that attention shifted toward cognitive control (Bialystok, Reference Bialystok and Harris1992). Cognitive control, sometimes referred to as executive function, is a set of goal directed attentional functions including updating, shifting, and inhibitory control, which are essential in controlling thought and behavior (Friedman & Miyake, Reference Friedman and Miyake2017; Miyake et al., Reference Miyake, Friedman, Emerson, Witzki, Howerter and Wager2000). Proponents of the view that bilingualism impacts cognitive control argue that the experience of managing two languages, which are thought to be simultaneously active (Abutalebi & Green, Reference Abutalebi and Green2007; Costa et al., Reference Costa, Miozzo and Caramazza1999; Kroll et al., Reference Kroll, Dussias, Bogulski and Kroff2012; Marian & Spivey, Reference Marian and Spivey2003), places higher demands on this system, resulting in more efficient cognitive control with higher levels of bilingual experience (Bialystok & Craik, Reference Bialystok and Craik2022).
Earlier theoretical accounts framed bilingual effects on cognitive control as a form of transfer. Specifically, because the cognitive demands associated with successful communication in bilinguals, such as the need to inhibit an active but unneeded language, rely on domain general inhibitory control (Green, Reference Green1998), bilingual language experience was thought to transfer to the nonverbal domain. As if follows, bilinguals were expected to demonstrate improved inhibitory control relative to monolinguals who do not experience these cognitive demands (e.g., Bialystok et al., Reference Bialystok, Craik, Klein and Viswanathan2004). Arguments that bilingual language experience transfers to domain-general inhibitory control soon became contentious as results across studies supported that bilinguals outperformed monolinguals on a range of measures that did not rely exclusively on inhibitory control (Hilchey & Klein, Reference Hilchey and Klein2011), a finding initially attributed to the overlap across dimensions of cognitive control (Friedman & Miyake, Reference Friedman and Miyake2017). Reported bilingual effects in preverbal infants raised in bilingual environments (e.g., D’Souza & D’Souza, Reference D’Souza and D’Souza2021; D’Souza et al., Reference D’Souza, Brady, Haensel and D’Souza2020) further called into question the validity of the transfer perspective as enhanced cognitive control was observed in the absence of the very language control experience thought to be essential for its development. Recently, in light of conflicting findings but also strong evidence supporting bilingual effects on cognitive control (Grundy, Reference Grundy2020), Ellen Bialystok proposed a new theoretical framework arguing that bilingual experience modifies cognitive control not through transfer, but through adaptation of the underlying attention system (Bialystok, Reference Bialystok2024). A central feature of this adaptationist account is that bilingual experience modifies the efficiency of cognitive control in a graded manner, with effects observed in both linguistic and nonlinguistic domains.
Investigations of bilingual effects on cognitive control generally administer behavioral tasks including the Simon task (Simon & Rudell, Reference Simon and Rudell1967), Flanker task (Eriksen & Eriksen, Reference Eriksen and Eriksen1974), and Attention Network Test (ANT; Fan et al., Reference Fan, McCandliss, Sommer, Raz and Posner2002), relying on between-groups designs which compare task performance between monolingual and bilingual samples. Performance on these tasks is typically quantified by averaging reaction times (RT) or accuracy rates to reflect performance overall and under different task conditions (e.g., congruent trials), with scores for each participant and group analyzed using analysis of variance or similar statistical tests (Privitera & Weekes, Reference Privitera and Weekes2023). Systematic reviews and meta-analyses of these previous studies have reported mixed findings, calling into question the veracity of bilingual effects on cognitive control (Donnelly et al., Reference Donnelly, Brooks and Homer2019; Giovannoli et al., Reference Giovannoli, Martella, Federico, Pirchio and Casagrande2020; Grundy, Reference Grundy2020; Van den Noort et al., Reference Van den Noort, Struys, Bosch, Jaswetz, Perriard, Yeo, Barisch, Vermeire, Lee and Lim2019; Ware et al., Reference Ware, Kirkovski and Lum2020).
Commonplace methodological practices across previous studies may, in part, explain mixed findings reported across evidence syntheses. An almost exclusive reliance on monolingual-bilingual comparisons may create conditions under which authentic bilingual effects on cognitive control are difficult to identify. The shift toward using exclusively bilingual between-group (e.g., Grundy, Reference Grundy2020; Xie & Pisano, Reference Xie and Pisano2019) and within-group designs (e.g., Privitera et al., Reference Privitera, Momenian and Weekes2022; Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023; Xie et al., Reference Xie, Privitera and Huang2024) has partially addressed these issues, aligning with recent critiques that monolinguals are not an appropriate control against which to compare the performance of bilinguals (Rothman et al., Reference Rothman, Bayram, DeLuca, Di Pisa, Dunabeitia, Gharibi, Hao, Kolb, Kubota and Kupisch2023). Additionally, the practice of assigning categorical language status labels to participants (Luk & Bialystok, Reference Luk and Bialystok2013), reducing monolingualism and bilingualism to uninformative, unidimensional constructs, ignores the heterogeneity inherent in language experience (De Bruin, Reference De Bruin2019; Gullifer et al., Reference Gullifer, Kousaie, Gilbert, Grant, Giroud, Coulter, Klein, Baum, Phillips and Titone2021). Operationalizing bilingualism as a multidimensional, continuous variable and more comprehensively assessing and modeling differences can support identification of meaningful associations between separable dimensions of language experience and cognitive control.
Despite the progress made in addressing pervasive methodological concerns, one significant issue persists across nearly all previous studies: the assumption that task performance is stable through time. Put another way, almost everything we know about bilingual effects on cognitive control is based on evidence that presupposes that performance on a given task does not change. The statistical impact of this assumption is evidenced in the commonplace use of aggregated data during analysis, reducing the influence of differences in performance across trials within participants and of individual differences between participants (Speelman & McGann, Reference Speelman and McGann2013). While some studies ask participants to complete multiple blocks of trials to model changes over time (e.g., Abutalebi et al., Reference Abutalebi, Della Rosa, Green, Hernandez, Scifo, Keim, Cappa and Costa2012; Costa et al., Reference Costa, Hernández, Costa-Faidella and Sebastián-Gallés2009), data within each block are aggregated during analysis, preventing the identification of changes occurring within a single block. To address the limits associated with these methods, there is a need to identify additional measures that best reflect nontrivial differences in performance on cognitive control tasks.
The adoption of alternative measures of task performance in the investigation of bilingual effects on cognitive control is not without precedent, although their use remains extremely uncommon. Using data collected from a Flanker task, previous studies have reported significant bilingual effects on non-decision time extracted through diffusion modeling (Ong et al., Reference Ong, Sewell, Weekes, McKague and Abutalebi2017), incongruent trial τ component and congruent trial μ component after ex-Gaussian analysis (Abutalebi et al., Reference Abutalebi, Guidi, Borsa, Canini, Della Rosa, Parris and Weekes2015), and sequential congruency effects (Grundy et al., Reference Grundy, Chung-Fat-Yim, Friesen, Mak and Bialystok2017). While these approaches utilized theoretically sound indices capturing underexplored dimensions of task performance, none account for changes in performance occurring over time. Accordingly, there is a need to identify measures of performance that address the limitations of aggregated data and challenge the false assumption that performance on cognitive control tasks remains stable over time.
A potential solution can be found in a recent paper from Cochrane and Green (Reference Cochrane and Green2021) investigating the relationship between fluid intelligence and time-sensitive measures derived from a cognitive control (i.e., working memory/updating) task. Unlike conventional measures of average performance, these time-sensitive performance trajectory measures included separate indices reflecting initial and final (asymptotic) task performance and the rate of change (time taken to change) during the task (Kattner et al., Reference Kattner, Cochrane and Green2017). The authors identified novel links between fluid intelligence and performance trajectory measures with higher fluid intelligence associated with better final task performance and faster rate of change. Additionally, while conventional performance measures were almost perfectly correlated with final task performance, no such relationship with rate of change was identified, supporting an independent link with fluid intelligence. Crucially, findings from this study support that performance on cognitive control tasks is not stable over time, and that time-sensitive measures of performance trajectory capture the dynamic nature of task performance in manner that is not confounded by aggregation of performance over time.
Extending the work of Cochrane and Green (Reference Cochrane and Green2021), the present study explored the temporal dynamics of bilingual effects on cognitive control. Through the use of a widely-adopted cognitive control task (i.e., Simon task), we investigated the relationship between separable dimensions of language experience and both conventional measures of aggregated performance and time-sensitive measures of performance trajectory. If bilingual effects manifest as graded differences in the efficiency and deployment of cognitive control (Bialystok, Reference Bialystok2024; Bialystok & Craik, Reference Bialystok and Craik2022), then higher levels of bilingual experience should be associated with faster initial task performance or rate of change. However, as the impact of improved cognitive control efficiency may be diluted over successive trials, these associations may not emerge on measures of overall or final performance. Due to limited previous work investigating the impact of separable dimensions of language experience on cognitive control (e.g., Anthony & Blumenfeld, Reference Anthony and Blumenfeld2019; Novitskiy et al., Reference Novitskiy, Shtyrov and Myachykov2019; Privitera et al., Reference Privitera, Momenian and Weekes2022; Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023; Tao et al., Reference Tao, Marzecová, Taft, Asanowicz and Wodniecka2011; Xie & Pisano, Reference Xie and Pisano2019; Yow & Li, Reference Yow and Li2015) including the absence of any study utilizing the performance trajectory measures used in the present study, we make no specific predictions about the nature of association between language experience and performance measures. However, previous work does suggest that separate dimensions of language experience may demonstrate opposing relationships with a given measure of cognitive control (Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023), resulting in a negative association between bilingual experience and cognitive control. Although previous work has provided some guidance to inform predictions, considering that the present study is the first investigation of bilingual effects on cognitive control utilizing time-sensitive performance trajectory measures, the nature of our analyses should be considered exploratory.
2. Materials and methods
2.1. Participants
Fifty-seven Mandarin-English speaking bilingual young adults (39 females; M age = 19.74 years, SDage = 0.99 years) were recruited from a public university in a central province of Mainland China. All participants were native Mandarin (L1) speakers who had generally began learning English (L2) in primary school in accordance with the national education language policy. Written informed consent was collected from all participants before data collection. Approval for this study was granted by Huazhong University of Science and Technology. Participants received ¥20 (~$3 USD) for completing the study.
2.2. Language experience and background measures
Language experience data for Mandarin and English were collected using a bilingual version of the Language History Questionnaire (LHQ-3; Li et al., Reference Li, Zhang, Yu and Zhao2020). This tool consisted of a series of self-report questions designed to measure proficiency, immersion, and dominance in all languages a participant uses. Proficiency was calculated based on participants’ self-reported ability to use a language, rated on a scale from 1 (“very poor”) to 7 (“excellent”). Immersion was assessed using both age of acquisition (AoA) and years of use for each language. Dominance reflected the number of hours participants reported engaging in specific activities in a given language. Reported data were used to generate aggregate scores ranging from 0 to 1 for proficiency, immersion, and dominance in both Mandarin and English. Equal weight was given to the dimensions of listening, speaking, reading, and writing in the calculation of each aggregate score, and dominance scores were further weighted by reported proficiency. Aggregate scores for Mandarin and English dominance were also used to calculate an L2/L1 dominance ratio which, depending on a participant’s pattern of language usage, may exceed 1. Objective English proficiency was assessed using a 25-item multiple-choice grammar and vocabulary test taken from Transparent Language (https://www.transparent.com/) with scores calculated as the proportion of correct responses (ranging from 0 to 1). The decision to use this test was based on prior validation work and its established use in assessing objective proficiency in heterogeneous samples of bilingual participants (e.g., Branzi et al., Reference Branzi, Della Rosa, Canini, Costa and Abutalebi2016; Del Maschio et al., Reference Del Maschio, Del Mauro, Bellini, Abutalebi and Sulpizio2022; Privitera, Reference Privitera2024b; Privitera, Li, et al., Reference Privitera, Li, Zhou and Wang2023). In total, this resulted in three measures of L1 experience: subjective L1 proficiency, L1 dominance, and L1 immersion, and six measures of L2 experience: L2 AoA, subjective L2 proficiency, objective L2 proficiency, L2 dominance, L2 immersion, and L2/L1 dominance ratio. Finally, participants reported on basic demographic details, language switching frequency, other languages used beyond Mandarin and English, and parental education level, computed as the mean of the mother’s and father’s education levels (ranging from 1 = elementary school to 6 = doctorate) as a proxy for socioeconomic status (SES; Wermelinger et al., Reference Wermelinger, Gampe and Daum2017).
2.3. Nonverbal intelligence measure
Differences in nonverbal intelligence were assessed using an abbreviated form of Raven’s Advanced Progressive Matrices (Raven, 1962). A total of 36 items from Set I were administered online using the Chinese survey website SoJump, and presented in a manner such that the difficulty progressively increased with each subsequent item. Prior to the start of the test, participants were told that they had 10 minutes to complete as many items as possible. In the event a particular item was too difficult, participants were told to make a guess and skip to the next item. The total number of items out of 36 answered correctly in 10 minutes was taken as a participant’s measure of nonverbal intelligence. While these administrative procedures deviate considerably from those generally followed, abbreviated versions of Raven’s Advanced Progressive Matrices have reported similar psychometric properties to the full task (e.g., Arthur et al., Reference Arthur, Tubre, Paul and Sanchez-Ku1999; Hamel & Schmittmann, Reference Hamel and Schmittmann2006).
2.4. Simon task
Participants completed a two-color Simon task administered online (Privitera et al., Reference Privitera, Momenian and Weekes2022). Prior to the start of the task, participants were instructed to place their left index finger on the “Q” key and their right index finger on the “P” key of their computer’s keyboard. Each trial began with the presentation of a black fixation cross in the center of a white background for 300 ms. Based on the trial condition, a blue or brown square target stimulus (2.54 × 2.54 cm) was presented to the left, right, or directly over the previous location of the fixation cross. Participants were instructed to press either the “Q” or “P” button based only on the color of the target stimulus (button and color mapping counterbalanced). The combination of target stimulus color and location resulted in three trial conditions: congruent, where a target stimulus appeared on the same side as the correct response key, incongruent, where the target stimulus appeared on the opposite side, and neutral, where the target stimulus was presented in the center of the screen. Target stimuli remained on screen until a response was given after which a blank screen was presented for 500 ms. In total, participants completed six practice trials with feedback and 150 experimental trials without feedback. Experimental trials contained equal numbers of each possible condition and were presented in the same pseudorandomized order to ensure that participants’ task experience was comparable and that any observed changes over time could not be attributed to differences in trial order.
2.5. General administration procedures
All aspects of the experiment were carried out using the Gorilla online experiment builder (Anwyl-Irvine et al., Reference Anwyl-Irvine, Massonnié, Flitton, Kirkham and Evershed2020) with all written content including directions and consent forms provided simultaneously in both Chinese and English. Interested participants were sent an access link through either email or WeChat along with brief instructions asking them to find a quiet place free of distractions in which to complete the experiment on a desktop computer or laptop. The experimental portal was configured in such a way that accessing the study link using either a tablet or smartphone would result in automatic rejection of the participant. After successfully entering the experimental portal, participants were asked to maximize the size of their browser screen and were again reminded not to use their phone or engage in any other distracting behavior. After providing consent, participants completed the Simon task, followed by the LHQ-3, the objective English test, and then the nonverbal intelligence assessment (link to SoJump embedded in Gorilla experiment) before debriefing. Completion of all aspects of the experiment took around 40 minutes for each participant and breaks were available after each task.
2.6. Statistical analysis
Prior to analysis, incorrect trials, and trials with RTs shorter than 150 ms and longer than 2,000 ms were trimmed in alignment with previous recommendations when investigating bilingual effects on cognition (Zhou & Krott, Reference Zhou and Krott2016), including cognitive control specifically (Privitera, Reference Privitera2024a). Initially, average overall RT and separate average RTs for each task condition were computed for participants to generate conventional measures of task performance. Performance on the Simon task was then modeled in R (Version 4.3.1; R Core Team, Reference Raven2023) using the TEbrm function from the TEfits package (Cochrane, Reference Cochrane2020), which used nonlinear Bayesian modeling in Stan using the brms package (Bürkner, Reference Bürkner2017). The TEbrm function allows for the modeling of task performance as a continuous function of time using a maximum-likelihood model that nonlinearly relates time (e.g., trial number) to the outcome variable (e.g., RT). Participant-level task performance is summarized with three mean parameter estimates: initial RT, the estimated value at trial one; final (asymptotic) RT, the estimated value at a theoretical infinite amount of elapsed time; and rate of change, the base-2 log of the time taken to change halfway from initial to asymptotic values, with higher rates corresponding to longer periods of time needed to reach asymptotic performance, reflecting differences in task learning (A. Cochrane, personal communication, February 4, 2025). These three parameter estimates, which we collectively refer to as measures of performance trajectory (Cochrane & Green, Reference Cochrane and Green2021), were extracted from models for further analysis using the coef function in brms. More details about the TEbrm function TEfits package can be found on the package creator’s website (https://github.com/akcochrane/TEfits). Two separate models were run to generate participant-level performance trajectory parameters for overall task performance and performance for the three different task conditions. Both models were estimated following an ex-Gaussian distribution with priors set to default, and were run with two chains for 7,000 iterations for each chain with 2,000 iterations per chain discarded as warmup, resulting in a total of 10,000 post-warmup draws. R-hat values for both models were all equal to 1, indicating that models had converged appropriately across all chains.
All additional analyses were performed in JASP (JASP Team, 2024). Pearson or Spearman two-tailed correlations were run between conventional and performance trajectory measures, language experience, and background measures. To obtain more robust estimates, 95% confidence intervals for correlation coefficients were calculated using 1,000 bootstrap replicates. Finally, the variance in task performance accounted for by language experience was tested using generalized linear models for each performance trajectory or conventional task performance measure that was significantly or marginally significantly correlated with a language experience measure. All continuous predictor variables were standardized (z-score) prior to modeling. Models each contained a single performance trajectory or conventional task performance measure as the outcome and correlated measure(s) of language experience and control background variables as predictors. Collinearity for each model was assessed using variance inflation factor (VIF), and variables with values above five were removed (Craney & Surles, Reference Craney and Surles2002). The appropriateness of final models was confirmed via inspection of residual plots and fit indices including Akaike information criterion (AIC; Stoica & Selen, Reference Stoica and Selen2004). Robust estimates of regression coefficients and their associated 95% confidence intervals were calculated using 5,000 bootstrap replicates.
3. Results
Participant language experience and other background variables are summarized in Table 1. As expected, considerable heterogeneity in language experience was observed across participants. Of note is the observation that some participants reported learning English either much earlier or later than expected based on national educational language policy. Taking all assessed dimensions of language experience into consideration, our sample can be most accurately described as bilingual, but Mandarin dominant. Due to a technical issue in which data from completed assessments were not recorded by SoJump for unknown reasons, nonverbal intelligence data were not available for 10 participants. To minimize the impact, analyses utilized the full dataset where possible, while analyses incorporating nonverbal intelligence as a variable were conducted on the reduced dataset. Full data and analyses can be accessed on Open Science Framework (https://doi.org/10.17605/OSF.IO/H5XVG).
Table 1. Demographic and language history data of final sample (n = 56)
3.1. Simon task results
Prior to analyses, all data (150 trials) were removed for one participant who had an overall accuracy of less than 5% on the task. Data trimming resulted in the removal of 420 incorrect trials and 40 trials with RTs shorter than 150 ms and longer than 2,000 ms. Final analyses included a total of 7,940 trials from 56 participants (38 females; M age = 19.73 years, SDage = 1.00 year). Simon task conventional performance and performance trajectory measure results are summarized in Table 2. No analysis of accuracy data was performed due to very low error rates. Replication of the classic Simon effect (i.e., slower RTs on incongruent relative to congruent trials) was confirmed via one-tailed paired samples t-tests for conventional performance and initial and final performance trajectory measures (ps < .001). Finally, expected performance improvements (i.e., faster RTs on final relative to initial performance trajectory measures) were tested using one-tailed paired samples t-tests, with significant improvements observed for overall, neutral, and incongruent trial performance (ps ≤ .008), but not congruent trial performance (p = .211).
Table 2. Summary of Simon task performance by item congruency condition
Note: Rate of change corresponds to the [base-2] log of the time taken to change halfway from the initial to the final (asymptotic) RT. A higher rate of change corresponds to a longer period of time needed to reach asymptotic performance.
3.1.1. Correlation results
Conventional and performance trajectory measures. Correlation analysis results are presented in Table 3. All conventional performance measures, representing average RT, were significantly positively correlated with their respective initial and final RT performance trajectory measures. Final RT measures were all significantly positively correlated with their respective rate of change measures, indicating that slower learners tended to have worse asymptotic performance. Significant positive correlations were also identified between average incongruent and neutral trial RT and their respective rate of change measures, indicating that slower learners also tended to have worse overall performance on incongruent and neutral trials. Finally, no initial or final RT performance trajectory measures were significantly correlated with each other, supporting that these measures are independent of one another.
Table 3. Correlations between conventional and performance trajectory measures
Note. Values in square brackets represent 95% confidence intervals for each correlation based on 1,000 bootstrapped replicates. All correlations were calculated with df = 54. OA: Overall, AVG: Average, CON: Congruent, RoC: Rate of change, INC: Incongruent, NEU: Neutral. Bold values indicate statistically significant results: * = p ≤ .05, ** = p ≤ .01, *** = p ≤ .001.
a Symbol identifies variables that were non-normally distributed, with results generated using Spearman correlations.
Language experience and background measures. Correlation analysis results are presented in Table 4. Age was significantly positively correlated with L2/L1 dominance ratio and marginally significantly correlated with L2 subjective proficiency and L2 dominance, supporting that older participants tended to use more English relative to Mandarin and demonstrate higher proficiency. This relationship was further evidenced in significant positive correlations between L2/L1 dominance ratio and L2 subjective and objective proficiency and L2 immersion, as well as L2 dominance and L2 subjective and objective proficiency. L2 AoA was significantly negatively correlated with nonverbal intelligence and marginally significantly negatively correlated with SES, indicating that those with began learning English earlier tended to be of higher status with higher nonverbal intelligence. There was a significant positive correlation between the number of languages used and frequency of language switching, supporting that those using a higher number of languages were more likely to engage in switching. Language switching was also significantly positively correlated with all measures of L2 experience and negatively correlated with L2 AoA, supporting that participants with higher levels of bilingualism tended to engage in more switching. While a significant positive correlation was identified between L2 subjective and objective proficiency, L1 subjective proficiency was only significantly positively correlated with L2 subjective proficiency. Finally, no significant correlation was identified between SES and nonverbal intelligence.
Table 4. Correlations between language experience and background measures
Note: Values in square brackets represent 95% confidence intervals for each correlation based on 1,000 bootstrapped replicates. All correlations were calculated with df = 54 except correlations with the variable NV-IQ which were calculated with df = 45 due to missing data. SES: socioeconomic status, NV-IQ: nonverbal intelligence, NUM Lang: number of languages used, L2: second language, AoA: age of acquisition, PROS: subjective proficiency, L1: native language, IMM: immersion, DOM: Dominance, PROO: objective proficiency. Bold values indicate statistically significant results: ^ = p ≤ .10 * = p ≤ .05, ** = p ≤ .01, *** = p ≤ .001.
a Identifies variables that were non-normally distributed, with results generated using Spearman correlations.
Performance measures and language experience and background measures. Full correlation analysis results are presented in Table 5. L2 AoA was significantly negatively associated with average overall and congruent trial RT, congruent, incongruent, and neutral trial initial RT, and marginally significantly negatively correlated with overall initial and average neutral trial RT. A nearly identical pattern of results, albeit with positive correlations, was observed for L2 immersion, an aggregate score that gives equal weight to both AoA and years of use. Specifically, L2 immersion was significantly positively correlated with average overall and congruent trial RT, overall, congruent, incongruent, and neutral initial trial RT, and marginally significantly positively correlated with average incongruent and neutral trial RT. Across all rate of change measurements only one significant correlation was observed: a negative correlation between overall rate of change and SES, supporting that those with higher status tended to demonstrate faster learning on the Simon task. Overall rate of change was also marginally significantly negatively correlated with L2/L1 dominance ratio, indicating that participants who used more English relative to Mandarin tended to demonstrate faster learning. L1 subjective proficiency was significantly positively correlated with congruent trial final RT, and marginally significantly positively correlated with incongruent and neutral trial final RT, indicating that those with higher Mandarin proficiency tended to reach lower levels of performance across the three Simon task trial conditions. Finally, the number of languages used was marginally significantly negatively correlated with overall, incongruent, and neutral trial average RT, but the correlation with average congruent trial RT was not significant (p = .143).
Table 5. Correlations between performance, language experience, and background measures
Note: Values in square brackets represent 95% confidence intervals for each correlation based on 1,000 bootstrapped replicates. All correlations were calculated with df= 54 except correlations with the variable NV-IQ which were calculated with df=45 due to missing data. SES: Socioeconomic status, NV-IQ: Nonverbal intelligence, NUM Lang: Number of languages used, L2: Second language, AoA: Age of acquisition, PROS: Subjective proficiency, L1: Native language, IMM: Immersion, DOM: Dominance, PROO: Objective proficiency, OA: Overall, AVG: Average, CON: Congruent, RoC: Rate of change, INC: Incongruent, NEU: Neutral. Bold values indicate statistically significant results: ^ = p ≤ .10 * = p ≤ .05, ** = p ≤ .01, *** = p ≤ .001.
a Identifies variables that were non-normally distributed, with results generated using Spearman correlations.
3.1.2. Regression results
Informed by the results of the correlation analysis, separate generalized linear models were constructed for each conventional and performance trajectory measure that was either significantly or marginally significantly correlated with a measure of language experience. As summarized in Table 5, L2 AoA, L2 immersion, and L1 proficiency were all significantly correlated with Simon task performance while number of languages used and L2/L1 dominance ratio were marginally significantly correlated. Given the strong significant correlation between L2 AoA and L2 immersion (Table 4), the observation that all performance measures correlated with L2 AoA were also correlated with L2 immersion, and the more informative nature of the L2 immersion measure which gives equal weight to both AoA and years of use, the decision was made to exclude L2 AoA from models in favor of the L2 immersion measure. Finally, due to the observed significant correlation between SES and Simon task performance (Table 5), and the importance of accounting for differences in SES when investigating the impact of bilingualism on cognitive control (Morton & Harper, Reference Morton and Harper2007; Naeem et al., Reference Naeem, Filippi, Periche-Tomas, Papageorgiou and Bright2018; Xie & Pisano, Reference Xie and Pisano2019), SES was included as a control variable on the first step of all models. Accordingly, models containing SES and language experience variables will be compared against null models containing only SES, not intercept only models. Nonverbal intelligence was not included in any model due to the absence of any significant or marginally significant correlations with conventional or performance trajectory measures (Table 5).
Conventional performance measures. Full analysis results are presented in Table 6. The model for average overall RT including SES alone was not significant, R 2 = .01, F(1, 54) = 0.279, p = .600, explaining 1% of the variance in performance. The addition of L2 immersion and number of languages used on the second step resulted in a significantly improved model, ΔF(2, 52) = 3.995, p = .024, explaining an additional 13% of variance in performance, although the overall model was marginally significant, R 2 = .14, F(3, 52) = 2.767, p = .051. The model for average congruent trial RT with SES alone was not significant, R 2 = .00, F(1, 54) = 0.184, p = .670, and explained less than 1% of the variance in performance. The addition of L2 immersion on the second step resulted in a significantly improved model, ΔF(1, 53) = 5.914, p = .018, explaining an additional 10% of variance in performance, but the overall model was marginally significant, R 2 = .10, F(2, 53) = 3.058, p = .055. The model for average incongruent trial RT with SES alone was not significant, R 2 = .00, F(1, 54) = 0.132, p = .718, and the addition of L2 immersion and number of languages used on the second step marginally improved the model, ΔF(2, 52) = 3.057, p = .056, with the final model accounting for 11% of variance in performance, but the overall model was not significant, R 2 = .11, F(3, 52) = 2.086, p = .113. Finally, the model for average neutral trial RT with SES alone was not significant, R 2 = .01, F(1, 54) = 0.611, p = .438, and explained 1% of the variance in performance. The addition of L2 immersion and number of languages used on the second step resulted in a significantly improved model, ΔF(2, 52) = 4.029, p = .024, explaining an additional 13% of variance in performance, R 2 = .14, F(3, 52) = 2.913, p = .043. To summarize, although only the overall model for average neutral trial RT was significant, the presence of significant coefficients supports that higher levels of L2 immersion were generally associated with slower average performance across all trial types (marginally significant coefficient for incongruent trials) while higher numbers of languages used were associated with faster average performance on all but congruent trials.
Table 6. Regression analysis results for conventional performance measures
Note: P-values reported next to each model represent results of overall model ANOVA. Coefficient estimates are based on the median of bootstrap distributions after 5,000 replicates. All models met the assumptions for linear regression. SE: Standard error, CI: Confidence interval, SES: Socioeconomics status, NUM Lang: Number of languages used, L2 IMM: Second language immersion. Bold values indicate statistically significant results.
Performance trajectory: Initial RT. Full analysis results are presented in Table 7. The model for initial overall RT including SES alone was not significant, R 2 = .00, F(1, 54) = 0.025, p = .874, explaining less than 1% of the variance in performance. The addition of L2 immersion on the second step resulted in a significantly improved model, ΔF(1, 53) = 8.644, p = .005, explaining an additional 14% of variance in performance, R 2 = .14, F(2, 53) = 4.336, p = .018. The model for initial congruent RT including SES alone was not significant, R 2 = .01, F(1, 54) = 0.291, p = .592, explaining 1% of the variance in performance. The addition of L2 immersion on the second step resulted in a significantly improved model, ΔF(1, 53) = 11.221, p = .001, explaining an additional 17% of variance in performance, R 2 = .18, F(2, 53) = 5.783, p = .005. The model for initial incongruent RT including SES alone was not significant, R 2 = .00, F(1, 54) = 0.151, p = .699, explaining less than 1% of the variance in performance. The addition of L2 immersion on the second step resulted in a significantly improved model, ΔF(1, 53) = 11.164, p = .002, explaining an additional 17% of variance in performance, R 2 = .18, F(2, 53) = 5.672, p = .006. Finally, the model for initial neutral RT including SES alone was not significant, R 2 = .01, F(1, 54) = 0.396, p = .532, explaining 1% of the variance in performance. The addition of L2 immersion on the second step resulted in a significantly improved model, ΔF(1, 53) = 10.375, p = .002, explaining an additional 16% of variance in performance, R 2 = .17, F(2, 53) = 5.420, p = .007. While controlling for SES, higher levels of L2 immersion were associated with slower performance on all measures of initial performance.
Table 7. Regression analysis results for performance trajectory measures
Note: P-values reported next to each model represent results of overall model ANOVA. Coefficient estimates are based on the median of bootstrap distributions after 5,000 replicates. All models met the assumptions for linear regression. SE: Standard error, CI: Confidence interval, SES: Socioeconomics status, NUM Lang: Number of languages used, L2 IMM: Second language immersion. Bold values indicate statistically significant results.
Performance trajectory: Final RT. Full analysis results are presented in Table 7. The model for final congruent trial RT including SES alone was not significant, R 2 = .00, F(1, 54) = 0.102, p = .750, explaining less than 1% of the variance in performance. The addition of L1 proficiency on the second step resulted in a significantly improved model, ΔF(1, 53) = 4.155, p = .047, explaining an additional 7% of variance in performance, but the overall model was not significant, R 2 = .07, F(2, 53) = 2.132, p = .129. The model for final incongruent trial RT including SES alone was not significant, R 2 = .00, F(1, 54) = 0.045, p = .833, explaining less than 1% of the variance in performance. The addition of L1 proficiency and age on the second step resulted in a marginally significantly improved model, ΔF(2, 52) = 2.822, p = .069, explaining an additional 10% of variance in performance, but the overall model was not significant, R 2 = .10, F(3, 52) = 1.897, p = .141.Finally, the model for final neutral trial RT including SES alone was not significant, R 2 = .00, F(1, 54) = 0.143, p = .707, explaining less than 1% of the variance in performance. The addition of L1 proficiency on the second step resulted in a marginally improved model, ΔF(1, 53) = 3.931, p = .053, explaining an additional 7% of variance in performance, R 2 = .07, F(2, 53) = 2.041, p = .140. Together, the lack of significant overall model tests and coefficients for predictors within each model suggests that these predictors do not have a meaningful relationship with final task performance.
Performance trajectory: Rate of change. Full analysis results are presented in Table 7. Pre-modeling data checking identified the presence of outliers in our dataset, which were successfully addressed through log transformation. The model for overall rate of change including SES alone was significant, R 2 = .15, F(1, 54) = 9.615, p = .003, explaining 15% of the variance in performance. The addition of L2/L1 dominance ratio on the second step resulted in a significantly improved model, ΔF(1, 53) = 5.470, p = .023, explaining an additional 8% of variance in performance, R 2 = .23, F(2, 53) = 7.940, p < .001. While controlling for SES, higher usage of English relative to Mandarin was associated with faster overall learning.
4. Discussion
In a significant departure from all previous work, the present study investigated the relationship between language experience and Simon task performance modeled as a continuous function of time. The novelty of this study lies in the use of behavioral measures that do not assume stable task performance, therefore treating the Simon task as a learning task. We identified distinct patterns of results across both conventional performance and performance trajectory measures while controlling for SES, with each supporting a different relationship between bilingualism and cognitive control. While differences in some aspects of language experience, most reliably L2 immersion, were generally predictive of average and initial task performance, no significant associations with final task performance were identified. Additionally, L2/L1 dominance ratio was the only variable predictive of overall rate of change on the Simon task. Taken together, our findings provide additional support for the idea that different aspects of bilingual experience are predictive of different dimensions of cognitive control (Yurtsever et al., Reference Yurtsever, Sotebeer and Grundy2024), extending these findings into the domain of performance trajectory measures (Cochrane & Green, Reference Cochrane and Green2021).
4.1. Trajectory measures capture unique aspects of performance
Correlations between conventional performance measures and their respective measures of initial and final performance suggest that these separate indices reflect similar aspects of performance, but the absence of significant correlations between measures of initial and final performance support that these measures are independent. We interpret this pattern of results as support for initial and final performance measures reflecting distinct and independent aspects of performance at different stages, with conventional performance measures (i.e., average) reflecting an integrative measure of these independent contributions. Significant positive correlations between rate of change and final but not initial performance further supports the independence of these measures, suggesting that heterogeneity in learning is more strongly exhibited in final performance. Significant positive correlations for rate of change and conventional performance measures were only observed for incongruent and neutral trials, suggesting that overall performance on congruent trials is less sensitive to differences in learning. This interpretation is further supported by the absence of a significant reduction in RT between initial and final performance measures, suggesting that congruent trial performance may begin at near ceiling levels, potentially due to their faciliatory influence (Simon & Rudell, Reference Simon and Rudell1967).
4.2. Distinct relationships of L2 immersion and multilingualism with task performance
Although models for overall, congruent, and incongruent trial RT did not significantly explain variance in task performance, L2 immersion and number of languages used exhibited meaningful relationships with all measures after accounting for differences in SES. Specifically, higher levels of L2 immersion, reflecting both earlier AoA and more years of L2 use (Li et al., Reference Li, Zhang, Yu and Zhao2020), were generally associated with slower average performance across all trial types while higher numbers of languages used were associated with faster average performance on all but congruent trials. Similarly, the same association between L2 immersion and performance was observed for initial performance measures, although all models were significant and no meaningful relationship was identified for number of languages used. This pattern of results highlights that the relationship between bilingual experience and cognitive control can differ significantly depending on the specific dimension of language experience under investigation (Yurtsever et al., Reference Yurtsever, Sotebeer and Grundy2024).
Although negative associations between cognitive control and bilingualism in general (reviewed in Lehtonen et al., Reference Lehtonen, Soveri, Laine, Järvenpää, De Bruin and Antfolk2018), and L2 immersion specifically (Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023), have been reported, previous work has generally identified either positive (e.g., Luk et al., Reference Luk, De Sa and Bialystok2011; Yow & Li, Reference Yow and Li2015) or null associations (e.g., Pelham & Abrams, Reference Pelham and Abrams2014) with the highly related measure of AoA. One explanation for our observed findings relates to the characteristics of our sample: Mandarin-English bilinguals who were Mandarin dominant, had lower English proficiency, and lived in a Mandarin-immersive environment. Positive bilingual effects on cognitive control have been reported in late bilingual young adults (Bak et al., Reference Bak, Vega-Mendoza and Sorace2014; Luk et al., Reference Luk, De Sa and Bialystok2011), but studies of this kind generally recruit samples who are highly proficient in their L2 and living in L2 dominant environments (e.g., international students studying abroad), conditions under which significant language control demands are exerted (Green & Abutalebi, Reference Green and Abutalebi2013). Taking the characteristics of our sample into consideration, along with previous reports of a similar relationship in a linguistically comparable but slightly younger sample (Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023), we interpret the negative association between L2 immersion and cognitive control as reflecting the atypical environment our study was conducted in which did not necessitate proficient, habitual use of English. This interpretation aligns with prior empirical findings that emphasize the role of linguistic context in accounting for conflicting results regarding the relationship between bilingual experience and cognitive control (e.g., Woumans et al., Reference Woumans, Ceuleers, Van der Linden, Szmalec and Duyck2015), as well as broader theoretical frameworks emphasizing the influence of interactional context (Green & Abutalebi, Reference Green and Abutalebi2013). Further work is needed to more clearly identify factors that modulate the relationship between L2 immersion and cognitive control.
Relative to managing two languages, the use of three or more can be considered a more intensive form of language experience, which would be expected to exert a more pronounced influence on cognitive control (Bialystok, Reference Bialystok2024). Support for this prediction remains limited, in part because studies often group individuals who use more than one language into a single ‘bilingual’ category, obscuring distinctions between bilingual and multilingual experiences (e.g., Bialystok et al., Reference Bialystok, Poarch, Luo and Craik2014). Previous empirical studies have reported mixed results, with participants using three or more languages demonstrating superior performance on only some dimensions of cognitive control compared to bilinguals (e.g., Durand López, Reference Durand López2021; Madrazo & Bernardo, Reference Madrazo and Bernardo2018; Poarch & Van Hell, Reference Poarch and Van Hell2012; reviewed in Schroeder & Marian, Reference Schroeder and Marian2017). In one previous study conducted using a linguistically similar sample, Privitera, Momenian, and Weekes (Reference Privitera, Momenian and Weekes2023) observed a main effect of number of languages used on the Attention Network Test (ANT; Fan et al., Reference Fan, McCandliss, Sommer, Raz and Posner2002), with higher numbers of languages associated with better cognitive control, manifesting as faster RTs on all trial types. While our pattern of results would support a similar conclusion, the absence of significant overall models for all but average neutral trial performance force us to interpret this finding with caution. The strength in our conclusions is further tempered due to the use of a measure of multilingualism that is limited in scope, only asking participants to indicate what languages they used beyond Mandarin and English. Although lacking granularity, the measure of multilingualism used in the present study is similar to those used in previous investigations of the impact of speaking more than two languages on cognitive control (e.g., Durand López, Reference Durand López2021; Guðmundsdóttir & Lesk, Reference Guðmundsdóttir and Lesk2019; Poarch & Van Hell, Reference Poarch and Van Hell2012). Future investigations should consider including more detailed assessments of multilingualism to better understand the nature of this result.
4.3. No bilingual effects associated with final performance measures
In stark contrast with conventional and initial performance measures, no language experience variables exhibited a meaningful relationship with final performance measures after accounting for differences in SES. Differences between the pattern of results observed for initial and final performance measures suggests that bilingual effects on cognitive control may emerge more readily toward the beginning of the task or, conversely, when using tasks with fewer experimental trials. Aligning with this interpretation, comparisons between bilingual and monolingual samples support a negative association between the number of experimental trials and the magnitude of bilingual effects on cognitive control in adults (reviewed in Hilchey et al., Reference Hilchey, Saint-Aubin and Klein2015; Hilchey & Klein, Reference Hilchey and Klein2011), although conflicting results have also been reported (Abutalebi et al., Reference Abutalebi, Della Rosa, Green, Hernandez, Scifo, Keim, Cappa and Costa2012). As posited by Costa et al. (Reference Costa, Hernández and Sebastián-Gallés2008), experience with a task is negatively associated with its cognitive demands, with well-practiced tasks being unlikely to identify bilingual effects. Empirical evidence supporting this claim was later identified (Costa et al., Reference Costa, Hernández, Costa-Faidella and Sebastián-Gallés2009), serving as a key component of current theoretical accounts of bilingual effects on cognitive control (Bialystok, Reference Bialystok2024; Bialystok & Craik, Reference Bialystok and Craik2022). The near-ceiling accuracy rates observed on the Simon task suggest it was not particularly challenging, which may have obscured any effects of bilingual language experience on final performance measures, especially considering the number of trials on the task (i.e., 150). Our finding, consistent with previous evidence syntheses (Hilchey & Klein, Reference Hilchey and Klein2011; Hilchey et al., Reference Hilchey, Saint-Aubin and Klein2015), underscores the importance of identifying methodological factors that may prevent the identification of authentic bilingual effects on cognitive control, and further highlights the utility of alternative measures that capture different aspects of performance.
4.4. Higher L2 relative to L1 usage associated with faster learning
A novel contribution of the present study is the identification of a significant bilingual effect on rate of change, a performance trajectory measure capturing differences in learning. Higher usage of English relative to Mandarin was associated with faster overall learning on the Simon task. Investigations of the relationship between bilingualism and learning suggest a positive association (Adesope et al., Reference Adesope, Lavin, Thompson and Ungerleider2010), although studies on specific kinds of learning, such as statistical learning, have reported mixed findings, potentially due to unaccounted heterogeneity in bilingual experience (Bulgarelli et al., Reference Bulgarelli, Lebkuecher and Weiss2018). Improved learning associated with higher levels of bilingual experience likely reflects the broader impact of bilingualism on cognitive control given its crucial role in learning and memory (Diamond, Reference Diamond2013; Duff et al., Reference Duff, Schoenberg, Scott and Adams2005).
Higher use of English relative to Mandarin has been previously associated with both positive and negative bilingual effects on cognitive control. Mandarin-English bilingual adolescents in China demonstrated a negative bilingual effect on a Flanker task, with higher L2/L1 dominance ratio associated with slower global RTs, but no influence on Simon task performance (Privitera et al., Reference Privitera, Momenian and Weekes2022). An identical finding was also observed on no-cue trials from the ANT, but emerged alongside a positive bilingual effect on orienting network function, with higher levels of English use relative to Mandarin associated with faster responses on spatial cue trials (Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023). Considering that the present study only identified a positive bilingual effect of L2/L1 dominance ratio on overall rate of change and not any conventional, initial, or final performance measure, and that no explicit orienting cues were presented on the Simon task, our finding may reflect the unmeasured influence of improved orienting network function on learning in general (e.g., Leclercq & Seitz, Reference Leclercq and Seitz2012; Pederson & Guion-Anderson, Reference Pederson and Guion-Anderson2010). Support for this interpretation can be found in the same study by Privitera, Momenian, and Weekes (Reference Privitera, Momenian and Weekes2023) in which a positive bilingual effect of L2/L1 dominance ratio on orienting, together with a positive correlation between L2/L1 dominance ratio and L2 proficiency, was interpreted as increased sensitivity to environmental cues supporting the identification of opportunities to use and, subsequently, improve (i.e., learn) English. The same correlation between L2/L1 dominance ratio and L2 proficiency was identified in the present study, supporting a similar interpretation. Alternatively, the cognitive demands placed on bilinguals who use their L2 more relative to their L1 in an L1-dominant environment may support the emergence of bilingual effects that are more readily observed on measures of learning. Considering that our participants were recruited from Mainland China, a Mandarin-dominant environment, those with higher L2/L1 dominance ratios may be more adept at learning when, where, and with whom they can use English. However, this interpretation should be considered speculative in the absence of strong evidence.
To summarize, the present study identified distinct associations between separable dimensions of bilingual language experience and both conventional and performance trajectory measures from a Simon task. While earlier accounts predicted that bilingual effects would manifest exclusively on incongruent trials or interference scores, reflecting improved inhibitory control (Bialystok, Reference Bialystok2001), more recent accounts suggest these effects are likely more general (Bialystok, Reference Bialystok2024; Bialystok & Craik, Reference Bialystok and Craik2022; Hilchey et al., Reference Hilchey, Saint-Aubin and Klein2015; Hilchey & Klein, Reference Hilchey and Klein2011). Our pattern of results aligns with these more recent accounts, although this included both positive and negative bilingual effects. Additionally, historic methodological reliance on categorical language status labels (Luk & Bialystok, Reference Luk and Bialystok2013) and ecologically flawed monolingual-bilingual comparisons (Rothman et al., Reference Rothman, Bayram, DeLuca, Di Pisa, Dunabeitia, Gharibi, Hao, Kolb, Kubota and Kupisch2023) complicated the interpretation of our findings regarding the relationship between a given dimension of bilingual experience and cognitive control. Negative associations between dimensions of bilingual language experience and cognitive control may consistently be present, but their identification can be obscured by these common methodological approaches and by the masking effects of positive and null relationships with other dimensions. Only recently have investigations of bilingual effects on cognitive control considered separable dimensions of language experience, with growing evidence supporting that these relationships can differ across dimensions of both language experience and cognitive control (e.g., Privitera, Momenian, & Weekes, Reference Privitera, Momenian and Weekes2023; Xie et al., Reference Xie, Privitera and Huang2024; Xie & Pisano, Reference Xie and Pisano2019; Yow & Li, Reference Yow and Li2015; Yurtsever et al., Reference Yurtsever, Sotebeer and Grundy2024). For these reasons, interpretations described above may change considerably as future investigations better characterize the impact of separable dimensions of bilingual experience and cognitive control.
4.5. Limitations
The present study represents the first investigation of how bilingual language experience impacts time-sensitive measures of performance trajectory derived from Simon task performance. Accordingly, these results should be interpreted with caution as they require further validation through replication across diverse bilingual samples and a range of cognitive control tasks. Additionally, we did not observe robust bilingual effects on cognitive control, and the effect sizes associated with significant findings were generally modest. This may have resulted due to our reliance on a single behavioral task (Ware et al., Reference Ware, Kirkovski and Lum2020) or the use of a large number of experimental trials (Hilchey & Klein, Reference Hilchey and Klein2011). These methodological decisions were made based on the desire to reduce the time burden on participants while ensuring a sufficient number of trials for generation of performance trajectory measures (Cochrane & Green, Reference Cochrane and Green2021). While we employed bootstrapping in our analyses to improve the accuracy of estimates and help address sample size issues, these limitations nonetheless underscore the need for further research with larger samples and more sensitive measures. Finally, as described in our discussion, the unique linguistic profile of our sample may reduce generalizability of our findings to other bilingual populations.
5. Conclusions
There is a need to reconsider how bilingualism influences cognitive control and which analytical approaches most effectively reveal these effects. Findings from the present study suggest that the nearly universal reliance on overall task performance measures may be partially responsible for mixed results across previous reports. Furthermore, their use perpetuates the assumption that performance on tasks measuring cognitive control is stable, an assumption our data do not support. We suggest that complementing conventional performance measures with those that reflect different aspects of task performance, including learning, should be considered best practice. Finally, our observed pattern of results underscores the importance of investigating the contribution of separable dimensions of language experience in order to understand the complex relationship between bilingualism and cognitive control.
Acknowledgements
Thank you to Dr Xia Lihua at Huazhong University of Science and Technology for supporting the ethical approval and data collection for this study and Dr Aaron Cochrane at Brown University for his extraordinary patience in answering my seemingly endless stream of questions about data analysis.
Data availability statement
Full data and analyses can be accessed on Open Science Framework (https://doi.org/10.17605/OSF.IO/H5XVG).
Competing interests
The authors have no competing interests to declare.
Open access
