The Influence of Multilingual Experience on Executive Function and Structure Learning: Effects in Young Adults in the UK and Singapore

Abstract

Most studies regarding the relationship between multilingualism and cognitive control reduce linguistic diversity to a dichotomous comparison, viz., monolinguals vs. bilinguals, failing to capture the multifactorial nature of multilingualism. Language research is largely restricted to the Global North, albeit most of the world’s population resides in the Global South, limiting the interpretability of the existing literature. Cognitive performance is assessed using very few tasks, yielding unreliable measurements. In this study, we identify the manner in which multilingual experiences influence cognitive performance in diverse sociolinguistic contexts. Young adults from the UK (n = 51, mean age = 24.0, SD = 3.18) and Singapore (n = 36, mean age = 21.3, SD = 2.15) were tested using an extensive battery of cognitive tasks, including cognitive flexibility (CF), working memory (WM), inhibition, and structure learning (SL). Information on language proficiency, use, age of acquisition, and frequency of switching was collected. The effects of various linguistic factors on the cognitive performance of each group were assessed using multiple linear regression models. The UK and Singapore samples exhibited significantly different linguistic profiles, which in turn dissimilarly influenced their cognitive performance. Our study underscores the necessity for more research in the Global South, challenging the prevailing Northern-centric focus on the multilingualism–cognition relationship.

1. Introduction

The effects of multiple-language control on non-linguistic cognitive control are highly debated. Behavioural studies have reported better performance for bilinguals across a range of cognitive tests, including inhibition tasks (Bialystok et al. 2004; Hernández et al. 2010); cognitive flexibility (CF) tasks, for example, task-switching (Prior and Macwhinney 2010; Stasenko et al. 2017); and on working memory (WM), both visuospatial (Bialystok 2009; Kerrigan et al. 2017) and verbal (Antón et al. 2019; Blom et al. 2014). Advantages in statistical learning have also been reported (Bonifacci et al. 2011; Verhagen and de Bree 2021). However, an important segment of the literature reports null or even negative results (Antón et al. 2014; Carlson and Meltzoff 2008; Duñabeitia et al. 2014; Paap et al. 2014; Paap and Greenberg 2013).

The inconsistency in the literature can be attributed to multiple reasons. Most studies simply perform cross-sectional comparisons of monolingual vs. bilingual groups. Such designs reflect a binary understanding of bilingualism. They fail to capture the diversity and multifactorial nature of the phenomenon and cannot account for differences within bi-/multilingual groups (Luk and Bialystok 2013). Factors such as proficiency, language use, age of acquisition, or frequency of language switching are not considered, yet they may be a source of meaningful differences.

Additionally, these between-group comparisons may be confounded by variables such as socio-economic status (SES), education, intelligence, or immigration status, which often correlate with general cognitive performance (Fuller-Thomson and Kuh 2014; Goldberg et al. 2008; Kirk et al. 2014; Yang et al. 2011). In other words, differences between monolingual and bilingual groups in regard to/on non-linguistic variables may be driving spurious effects or masking genuine impacts (Bak 2015), so it is very important that these variables are controlled for.

The limited scope of cognitive assessments, which usually encompass one task per cognitive component, neglect the intricate architecture of executive functions (EFs) (Jylkkä et al. 2018). While EFs encompass inhibition, WM, and CF (Miyake et al. 2000), the interrelatedness and potential subcomponents of these functions remain underexplored.

Finally, most existing literature is focused on the Global North, even though the majority of the world’s population resides in the Global South (Makoni et al. 2022). This unavoidably produces a distorted image of the relationship between bi-/multilingualism and cognitive processing (Bak and Alladi 2016). Different sociolinguistic contexts are highly likely to influence this relationship, so research within different environments is crucial. For example, while countries in the Global North frequently have a single official national language, several countries in the Global South have multiple official languages. And while multilingualism may be present in certain areas of the Global North as well, societal perspectives tend to be different. Monoglossic ideologies are more dominant in the Global North (especially in Europe), whereas numerous contexts in the Global South embrace multilingualism as standard practice (Bunk and Wiese 2024).

This study examines the impact of managing multiple languages on domain-general cognitive control, addressing the aforementioned weaknesses. Instead of comparing distinct groups of monolinguals and bilinguals, we are operationalising “lingualism” as a continuum (Wigdorowitz et al. 2022), breaking it down into specific continuous linguistic variables (Gallo et al. 2021) and evaluating their effect on cognitive performance, while controlling for background variables, such as SES, intelligence, and age. The term “bilingual” is used to refer to people using two or more languages (or dialects) in their daily lives (Grosjean 2021). This view of bilingualism places more emphasis on language use rather than fluency. When the term “multilingual” is used, it is employed to emphasise that speakers use more than two languages.

Following recent studies, we have included various continuous measures of linguistic experience. These include age of onset of second-language acquisition (L2 AoA) (Luk et al. 2011; Soveri et al. 2011), balanced proficiency, and language use (Yow and Li 2015), as well as the relatively new measure of language entropy, which reflects the social diversity of language use (Gullifer and Titone 2020). High entropy indicates the use of multiple languages in an integrated fashion, and low entropy indicates compartmentalised use in distinct contexts or the use of a single language. We also evaluate the frequency and context of language switching (Jylkkä et al. 2017; Verreyt et al. 2016; Woumans et al. 2019), in line with the adaptive control hypothesis (ACH) (Green and Abutalebi 2013). The ACH proposes that there are distinct contexts of bi-/multilingual language use. In the single-language context, bilinguals use each of their languages in distinct environments. In the dual-language context, they use their languages within the same environment but with different interlocutors. In dense code-switching, they use their languages with the same interlocutor, making frequent switches between languages, even within a sentence. Each context is assumed to pose unique linguistic and cognitive demands.

To measure the influences of linguistic experience on EF and statistical learning, we employ an extensive battery of cognitive tests, which includes tasks on CF, WM, inhibition, as well as a statistical learning paradigm, here a specific instantiation of learning under uncertainty, termed structure learning (SL) (Wang et al. 2017a, 2017b). We thus hope to detect more nuanced effects and pinpoint specific aspects of cognition susceptible to multiple-language experience.

Finally, we investigate the effects of linguistic experience on cognition in two groups, one from the UK and one from Singapore. We thus broaden the scope of the existing literature by including a country from the Global South1 (World Population Review 2023). While research in Singapore is not as scarce as it is in other countries of the Global South (e.g., Hartanto and Yang 2016; Yow and Li 2015), the Singaporean sociolinguistic context remains underexplored. Ooi et al. (2018) compared bilinguals from Singapore with monolinguals and bilinguals in Edinburgh, who differed in switching tendencies and L2 AoA. While the authors focused on specific linguistic experiences to distinguish the groups and did not employ a binary categorisation, they still used a between-group design. In any case, assessing bilingual (and multilingual) populations from different countries and sociolinguistic environments is very informative because such investigations help to determine how the relationship between multilingualism and cognitive processing can vary as a function of linguistic context.

Singapore is an extremely interesting case, as it is a highly multilingual environment, with four official languages (English, Mandarin, Malay, Tamil), along with several other dialects (Ooi et al. 2018). Not all Singaporeans speak or understand all the official languages. In fact, recent research indicates that over the past decades, there has been a shift to “English Plus”; that is, Singaporeans mostly speak English, along with one other language, depending on their ethnic group (Cavallaro and Chin 2014). This is due to language policies: English has been defined as the main medium of instruction in the educational system, and the mother tongues are taught as L2s (Dixon 2005). A recent study with young adults (university students) indicates that most of the participants exhibited bilingual or trilingual identities (Siemund et al. 2014).

While many Singaporeans may not actively use more than two languages, they are still exposed to significant linguistic diversity, as speakers of the various ethnic and linguistic groups live side by side and interact on a daily basis (Leimgruber 2013). Singlish, a creolised form of English arising due to the language contact between British colonisers and indigenous Chinese, Malay, and Indian populations, is also frequently used and is part of the Singaporean identity (Siemund et al. 2014). The use of Singlish further increases the linguistic variety to which Singaporeans are exposed.

The United Kingdom (UK), on the other hand, is predominantly monolingual, with English as its official language, reported as the main language by 91.1% (52.6 million) of usual residents, according to Census 2021 data. The most common main languages other than English (English or Welsh in Wales) are: Polish (1.1%, 612,000), Romanian (0.8%, 472,000), Panjabi (0.5%, 291,000), and Urdu (0.5%, 270,000) (Office for National Statistics 2022). Research suggests that there is significantly less linguistic diversity in the UK than in more multilingual countries (e.g., South Africa) (Wigdorowitz et al. 2022), with the majority of the population speaking only English (Gough 2023). However, there may be variation in the linguistic diversity in the UK (and Europe, more generally) in different areas (e.g., rural vs. urban). When compared to more rural areas or smaller cities, larger cities such as London, Brussels, or Berlin may exhibit more linguistic diversity as a result of frequent interaction between groups, even though societal attitudes still privilege the major languages differently than is the case in Singapore (Bunk and Wiese 2024; Siemund et al. 2014).

In summary, this study sheds light on the relationship between multilingual experience and cognitive ability, using a variety of tasks and examining two diverse sociolinguistic contexts. To this effect, the study has two main aims: (i) to assess the differences in the linguistic profiles of young adults in the UK and Singapore; (ii) to identify whether different multilingualism experience factors influence cognitive performance and determine whether the relationship between linguistic experience and cognitive performance is different in these two contexts (UK and Singapore).

2. Materials and Methods

2.1. Participants

All participants were university students in either the UK (University of Cambridge, Cambridge, UK) or Singapore (Nanyang Technological University—NTU, Singapore). The UK sample included 51 participants (33 females; aged 19–30; M = 24.0; SD = 3.18), and the Singapore sample comprised 36 participants (27 females; aged 18–27; M = 21.3; SD = 2.15). The participants received monetary compensation for the participation (GBP 8 per hour of participation for participants in the UK and SGD 20 for the participants in Singapore).

All participants exhibited normal or corrected-to-normal vision and had no history of neurological impairment or brain injury.

All participants rated themselves as highly proficient in English2. We did not focus on a specific language combination, hoping to increase the ecological validity of the study and make our samples more representative in terms of the language realities of each of the contexts. We encouraged participants to report all the languages they have used in their lifespan, but they were asked to indicate proficiency and daily use for each language at the time of the study. In the UK, the languages other than English reported by the participants were: Afrikaans, Arabic, Bulgarian, Danish, French, German, Greek, Hokkien, Indonesian, Italian, Korean, Mandarin, Norwegian, Portuguese, Russian, Serbian, Sinhalese, Spanish, and isiZulu. In Singapore, the languages, other than English, reported were: Cantonese, French, German, Hindi, Hokkien, Indonesian, Japanese, Korean, Malay, Mandarin, Tamil, Telugu, and Teochew. More information on self-reported language proficiency and use in the different groups is presented in Section 3.1.

2.2. Materials and Procedure

2.2.1. Procedure

Participants were tested individually following a remote guided protocol for supervised web-based testing that was developed during the pandemic [for a detailed description of the protocol, see Leong et al. (2022)]. To ensure high-quality data, we required participants to have access to a personal computer with Windows 7 or 10 or MAC OS, minimum 8GB RAM, a webcam and microphone, administrator rights to download and install software, and a reliable internet connection. We ensured that all the requirements were met by employing a screening questionnaire administered before accepting individuals for participation. Furthermore, by using the remote guided testing (RGT) protocol, we could make sure that participants were concentrating when performing the tasks, as they were connected via a video call with the experimenter, sharing their screens throughout testing.

The participants also signed a consent form and completed a demographics and language questionnaire before beginning the testing session. During the testing session, participants were assessed on a battery of cognitive tasks measuring EF and SL. They also completed a test of non-verbal intelligence, in which they were required to solve matrix puzzles of increasing difficulty. A test of verbal intelligence (the Vocabulary subtest from the Wechsler Abbreviated Scale of Intelligence, 2nd edition (WASI-II); Wechsler (2011)) was also administered. For the remaining cognitive tests, five pseudo-randomised orders of tasks were created, and each participant was assigned to one of these. The testing session lasted for about 2.5 to 3 h in total. Frequent breaks between tasks were planned by the experimenter, but we also checked with the participants throughout the process to see if they needed additional breaks.

The study was approved by the University of Cambridge and NTU Singapore review boards.

2.2.2. Background Measures

Demographics

Information on demographics, including gender, age, and SES, was collected using an extensive demographics questionnaire (Appendix A). For SES, we used a composite score ranging from 0 to 1, which comprised measures of parental and maternal education, as well as annual income. Higher scores indicate higher SES.

Intelligence3

• Non-Verbal Intelligence (NVIQ)

To measure NVIQ, we used a seven-item progressive matrices test, inspired by the Raven’s Standard Progressive Matrices (Raven et al. 2000). The task was hosted on the iABC website, a Cambridge University in-house platform (I-ABC: A Personalised Learning Study—Adaptive Brain Lab n.d.). Participants were presented with a 3 × 3 grid of images, with the final image left blank, and were asked to select which of six options best completed the grid.

The duration of the task is about 7 min.

• Verbal Intelligence (VIQ)

We used the Wechsler Abbreviated Scale of Intelligence, 2nd edition (WASI-II) (Wechsler 2011) Vocabulary subtest to measure the participants’ VIQ. The Vocabulary subtest requires participants to define up to 30 words. It assesses participants’ word understanding and reflects language development, expressive language skills, and retrieval of information from long-term memory. Raw scores range from 3 to 59 (for this age group), but they were transformed based on the age-related standards. The adjusted scores ranged from 20 to 80.

The duration of the task is about 6 min.

The background demographics for each group are summarised in

Table 1. Means and standard deviations (SD) for Age, SES, intelligence.

Background Measure	UK		Singapore
	Mean	SD	Mean	SD	t	p
Age	24.0	3.18	21.3	2.15	−4.36	<0.0001
SES a	0.59	0.17	0.45	0.17	−3.76	0.0003
NVIQ b	0.65	0.24	0.48	0.24	−3.31	0.001
VIQ c	63.5	10.5	51.7	7.89	−5.68	<0.0001

^a SES values could range from 0 to 1. ^b NVIQ values could range from 0 to 1. ^c VIQ values could range 20 to 80.

and

Table 2. Information on gender, dominance in English, and videogame usage.

Background Measure	UK	Singapore
	N	N	X2	p
Female	33 (of 51)	27 (of 36)	1.05	0.307
English-dominant (use)	42 (of 51)	31 (of 36)	0.22	0.638
Videogame usage	18 (of 51)	20 (of 36)	3.52	0.061

. The groups differed significantly in regard to age, SES, NVIQ, and VIQ.

For this reason, as discussed in the Section 2.3, we did not perform direct comparisons between the two contexts; that is, we did not add Country as a predictor in our models. Instead, we employed separate models for each context. Given the great differences in these demographic measures, as well as other factors outside of our experimental control (such as the educational system), a difference between countries would not necessarily be meaningful or easily interpreted. It is the language-related effects present in each country that are of interest to us, rather than the effect of the country per se.

2.2.3. Language Measures

In our attempt to quantify multilingualism as a multifactorial continuum, we used an extensive questionnaire, collecting information on several language background variables. Specifically, we required participants to provide the following information: languages they use (or have used), age of acquisition of each language, and hours (and percentage) of daily use. They also reported the context of acquisition and the context of use, but these data are not analysed in the present study, as we only focused on continuous variables.

Following the adaptive control hypothesis (ACH) (Green and Abutalebi 2013), we asked participants to rate how frequently they use their languages in a single-language, a dual-language, and in a dense code-switching context. The values ranged from 1 (never) to 5 (always). The full language questionnaire, with the exact phrasing used in the questions, can be found in Appendix B.

Finally, the questionnaire included a section in which participants had to rate their proficiency in listening, reading, spoken interaction, spoken production, and writing in each of their languages. To prevent participants from providing mechanical responses, we did not ask them to simply select a value, but provided them with different descriptions of proficiency, based on the Common European Framework of Reference for Languages (CEFR) (Council of Europe 2020). These were then converted to values ranging from 1 to 6.4

Based on this information and following the methods in recent literature regarding the effects of multilingualism on cognitive control (Li et al. 2021; Yow and Li 2015), we calculated the following indices: Balanced Proficiency, Balanced Usage, and Entropy. Balanced Proficiency was calculated as the difference between Proficiency₁ (the highest proficiency score of the two most proficient languages) and Proficiency₂ (the lowest proficiency score of the two languages).5 Balanced Usage was calculated as the difference between Usage₁ (the highest usage score of the two most used languages) and Usage₂ (the lowest usage score of the two languages).6 Higher scores in Balanced Proficiency and Balanced Usage indicated greater differences between the languages, and, therefore, greater “imbalance”, whereas a value of 0 indicated perfect balance. We focused on the first two languages to make our results comparable to those from existing studies, but also because all of our participants mainly used (and were most proficient in) the first two languages they reported. Nevertheless, as we possessed information on all the languages that each participant used, we also calculated a richer metric of linguistic diversity, that is, Language Entropy (Gullifer et al. 2018; Gullifer and Titone 2020). This is particularly relevant for the Singapore context, which is highly multilingual, with four official languages (English, Mandarin, Malay, and Tamil) (Siemund et al. 2014). We used the Language Entropy R package (Gullifer and Titone 2018) for our evaluation. Unlike the Balanced Usage metric, Language Entropy allowed us to take into account the usage of all of a participant’s languages. The participants were asked to list all of their languages and the percentage of use for each one. The sum of the percentages added up to 100%. The percentages were converted to proportions (e.g., 47% to 0.47), and the following equation was used to calculate Language Entropy:

$$H=-\sum _{i=1}^n{P}_i{log}_2{(P}_i)$$

Here, n represents the possible languages, and P_i is the proportion of use of language_i. Language Entropy scores range from 0 to log₂(n). Lower entropy scores indicate lower diversity and the use of predominantly one language in a compartmentalised fashion. Higher entropy scores indicate higher language diversity and a more balanced use of languages, as well as language switching. Participants who only used one language had a Language Entropy score of 0.

2.2.4. EF and SL Measures

Cognitive Flexibility

A total of five tasks were administered to measure CF.

(1)

Task-Set switching (TSS)

In Task-set switching (TSS), adapted from Kehagia et al. (2017), participants had to categorise the stimuli on their screens based on certain rules. For each trial, a digit (1 to 9) and a letter (A, E, I, U, C, F, T, or X) were presented in the centre of the screen. The participants were required to switch between two tasks, one concerning the number and one the letter. The target stimuli in each trial were accompanied by a cue indicating whether the participants had to focus on the number or the letter. A circular shape (a vertical or horizontal ellipse) was used as a cue for the letter task. An angular shape (square or rhomb) was used as a cue for the number task. The stimuli (letter and number) were presented one above the other. In the letter task, when the circular-shape cue appeared, participants had to decide whether the presented letter was a vowel or a consonant. For vowels, participants had to press the left-arrow key on their keyboard, and for consonants, they were to press the right-arrow key. In the number task, the task with the angular-shaped cue, participants had to decide whether the number was odd or even. They had to press the left-arrow key for odd numbers and the right-arrow key for even numbers.

There were four blocks in total, each comprising 49 trials. The first trial of each block was excluded from the analysis, as it could not be regarded as either a switch or repeat. A switch trial is defined as a trial in which the participant has to perform a task different from the one they performed in the previous trial, i.e., switching from performing the number task to performing the letter task, and the switch was indicated by a change in cue (from angular to circular). In repeat trials, participants would perform the same task as the one performed in the previous trial, (e.g., continue performing the number task), so the cue would not change (remaining angular). The ratio of switch and repeat trials in each block was 1:2. In each trial, participants initially saw the cue for 300 ms, and then the letter and number pair appeared on the screen, inside the cue. The cue and stimuli remained on screen until the participants provided a response or for a maximum 2000 ms. After that, a fixation cross appeared in the centre of the screen for a variable inter-trial interval of 1700, 1825 or 1950 ms. No feedback was offered after selection. Before the experimental part of the task began, a very detailed set of instructions was provided to the participants, and they practiced on five sample trials, during which they were given feedback for their responses. The task was administered through Gorilla (Gorilla 2022). The duration of the task is about 25 min (see Figure 1).

The mean accuracy for switch and non-switch trials was calculated for each participant. The mean reaction time (RT) for switch and non-switch trials was also calculated, after error trials were removed, along with trials faster than 300 ms, and trials > 2.5 SD from the mean. The TSS switch cost was then calculated as follows:

• TSS switch Accuracy cost = mean Accuracy in switch trials − mean Accuracy in non-switch trials
• TSS switch RT cost = mean RT in switch trials − mean RT in non-switch trials

(2)

Trail Making Test (TMT)

The Trail Making Test (TMT) (Armitage 1946) is a test of visual attention and task switching. We administered a computerised version of the task hosted on Millisecond-Inquisit (Inquisit 6 2021). The participants were asked to use the mouse to connect a series of nodes as quickly and accurately as possible, without interrupting the course of the line. When the participants made an error, the software automatically interrupted the line and required them to continue from the last correct node they had reached.

The task consists of two trails, Trail A and Trail B. In Trail A, participants are required to connect nodes numbered from 1 to 25, in ascending order. In Trail B, numbers alternate with letters, so participants must switch between the two, going from 1 to A to 2 to B to 3 to C and so on, up to the number 13. Participants completed two short training trials, one before Trail A and one before Trail B. The duration of the task is about 3 min (see Figure 2).

The outcome measures yielded from TMT are latency (response time) and number of errors for Trail A and Trail B. We used latency scores to calculate the Trail B:A ratio, which was then entered in our analyses.

(3)

Wisconsin Card Sorting Test (WCST)

The Wisconsin Card Sorting Test (WCST) evaluates the ability to adapt to unannounced rule changes (Grant and Berg 1948). We implemented a computerised version of the task, using the iABC platform.

The participants are required to sort cards into four different categories. No instructions are given regarding the categorisation rules, but participants receive trial-level feedback after each selection, that is, whether they sorted the card correctly. The four categories appear as four different cards, each containing spaceships of different shapes and colours. These are: one blue spaceship of shape 1, two orange spaceships of shape 2, three yellow spaceships of shape 3, and four green spaceships of shape 4. The cards that the participants are asked to sort have similar designs and vary in colour (four options), shape (four options), and number (four options). In this version of the task, the participants must sort 60 trials in total. Every 10 trials, the rule changes. Each rule (colour, shape, number) is tested two times. Participants have a maximum of 5000 ms in which to give a response. The duration of the task is about 3 min (see Figure 3).

The number of perseverative errors, that is, erroneous responses showing that the participant has failed to switch to the new relevant rule, was used in analyses.

(4)

Intra-Extra Dimensional Set Shift (IED) task

The Intra-Extra Dimensional Set Shift (IED) task (Robbins et al. 1994) is a test of rule acquisition and reversal. It is considered an analogue of the WCST, yet with fewer dimensions (colour and shape). The IED task involves visual discrimination, set formation, shifting, and flexibility of attention. The task was administered through the CANTAB^® cognitive assessment software (Cambridge Cognition 2019).

The participants are presented with two cards in each trial and are asked to choose one of the two. They receive feedback after each selection (the words “correct” or “incorrect” appear on the screen). They must figure out the rule through trial and error. After they reach the criterion of six correct responses, the stimuli and/or rule changes. In the first stages, the task only involves simple stimuli consisting of one dimension, for example, two white lines differing in shape. As the task proceeds, the stimuli become more complex, for example, white lines and pink shapes; thus, the number of potentially relevant dimensions increases to two, i.e., shape and colour. The rule shifts are initially intra-dimensional; thus, for example, the white lines remain the only relevant dimension (colour). Later in the task, an extra-dimensional shift in the rule occurs, so the pink shapes become the pertinent dimension (shape). There are nine stages in total, and the extra-dimensional shift occurs in Stage 8. If the participants fail to reach the criterion of six correct responses after 50 trials, the task terminates. Explicit instructions are offered through voice-over technology in the first few trials of the task until the participant is familiarised with the process. The duration of the task is about 7 min (see Figure 4).

The outcome measures included the total number of errors (which included an adjustment of 25 errors per stage, for participants who did not complete all stages) and the number of errors in Stage 8, which was the extra-dimensional shift stage. Again, for participants who did not reach Stage 8, 25 errors were assumed. The former is an index of the participants’ overall efficiency in attempting the test, whereas the latter is a measure of their ability to shift attentional set.

(5)

Probabilistic Reversal (PR) Learning task

In the Probabilistic Reversal (PR) Learning task (Cools et al. 2002), participants were presented with two stimuli (two different spaceships), each with a specific probability of being correct (80% vs. 20%). The participants were asked to select one of the two spaceships, and they would receive feedback after each selection. The rule defining the most probably correct stimulus changed at some point without warning, and the probabilities were then reversed. The participants had to learn and adapt to the new rule. There were 80 trials in total, 40 of each probabilistic structure. The task was administered through the iABC platform. The duration of the task is about 4 min (see Figure 5).

The outcome measure used was the number of perseverative errors participants made after the probabilities had been switched.

Inhibition

(1)

Stroop task

The Stroop task (Stroop 1935) measures the interference of word meaning with the naming of the colour in which the words are written, as indicated by RT and Accuracy differences between colour-meaning incongruent and congruent conditions. It is a task of inhibitory control. We implemented a computerised version of the task requiring keyboard responding, hosted on the Millisecond-Inquisit platform (Inquisit 6 2021).

Participants see names of colours, written in colour, and are asked to indicate the colour of the word (and not its meaning) through key press, as quickly and accurately as they can. Specifically, they were required to press “D” for red, “F” for green, “J” for blue, and “K” for yellow. The colour of the presented word may or may not match the meaning (e.g., “RED” printed in red or green). In congruent trials, the meaning of the word and its colour are the same, whereas in incongruent trials, they are not. The stimuli stay on the screen until a response is provided, and latencies are measured from the onset of the stimuli. If the participants give an incorrect response, an X is shown on the screen for 400 ms, and then a blank screen appears for 200 ms. If their response is correct, no feedback is given. There were 180 trials in total, 75% of which were congruent and 25% incongruent. The different keys and the colours they corresponded to were present on the screen throughout the task. The duration of the task is about 8 min (see Figure 6).

The mean accuracy for the congruent and incongruent trials was calculated for each participant. The mean RT for the congruent and incongruent trials was also calculated, after error trials were removed, along with trials faster than 300 ms, and trials > 2.5 SD, from the mean. The Stroop cost was then calculated as follows:

• Stroop Accuracy cost = mean Accuracy in incongruent trials − mean Accuracy in congruent trials
• Stroop RT cost = mean RT in incongruent trials − mean RT in congruent trials

Working Memory

Two tasks were used to measure WM, one for visuospatial and one for verbal WM.

(1)

Spatial Working Memory (SWM) task

We used the Spatial Working Memory (SWM) task from the CANTAB^® (Cambridge, UK) cognitive assessment battery (Cambridge Cognition 2019; Robbins et al. 1994). This is a self-ordered EF task requiring the retention and manipulation of visuospatial information.

The participants view some coloured boxes on the screen. They are asked to find all the yellow tokens and are told there is one in each box, so once they have found it, they need not revisit that box again. In the first stage, there are 4 boxes, and this number gradually increases to 6, 8, and 12. The colour and position of the boxes change as the participants progress through the stages. Explicit instructions are offered through voice-over technology in the first few trials of the task until the participant is familiarised with the process. The duration of the task is about 8 min (see Figure 7).

Outcome measures include the number of errors in each stage (selecting the same box twice in a run, even though it has been found to be empty, and revisiting boxes in which a token has already been found), as well as strategy. Strategy reflects the number of times a participant begins a new search pattern at the same box where they started their previous search. Starting one’s search consistently from the same box demonstrates planned strategy rather than random selection. Therefore, a lower score indicates better strategy, with a score of 1 indicating that the participant always began their searches at the same box.

(2)

Backward Digit Span (BDS) task

We measured executive verbal Working Memory with the Backward Digit Span (BDS) task (Blackburn and Benton 1957). The stimuli (random numbers from 1 to 9) were presented auditorily, and participants were asked to repeat them in reverse order. There were eight stages in total, each with two trials. The participants started at Stage 1 with two digits and moved to the next stage if they answered at least one trial correctly. In Stage 8, the final stage, participants had to recall nine numbers backwards. The administration stopped once a participant answered both trials in a stage incorrectly. A practice trial using two numbers was offered in the beginning. The duration of the task is about 8 min.

The outcome measure was the number of correct responses, ranging from 0 to 16.

Statistical Learning

Statistical learning was measured through a specific instantiation of learning under uncertainty, termed the structure learning (SL) task.

(1)

Structure Learning (SL) task

To measure statistical learning ability, we used a novel methodological framework, coined Structure Learning (SL) (Karlaftis et al. 2019; Wang et al. 2017a, 2017b). A detailed description of the SL task can be found in (Wang et al. 2017a, 2017b). Briefly, it is a visual statistical learning paradigm, in which participants are required to learn complex context-based probabilistic contingencies and not just to memorise simple frequency statistics. Participants are exposed to sequences of visual stimuli and are periodically asked to predict the upcoming symbol, without receiving trial-level feedback (see Figure 8).

In the version of the task employed in this study, four highly discriminable visual symbols were used. The sequences of symbols were generated by a 1st-level Markov model, which means that each new symbol depends on the symbol that appeared immediately before it. However, at each time point, the generated symbol is determined probabilistically. For each given symbol, only one of two others are allowed to follow, one with high (80%) and one with low (20%) probability. This makes the sequence stochastic and not deterministic. In other words, symbol A is most likely (80%) followed by symbol B, but 20% of the time, it will be followed by symbol C. The model is summarised below:

Level 1		Target
		A	B	C	D
Context	A		0.8	0.2
	B			0.8	0.2
	C	0.2			0.8
	D	0.8	0.2

Participants were familiarised with the process over five practice trials. The duration of the task is about 55 min.

The outcome measures used for the analyses were a performance index (PI) and a strategy index [strategy integral curve difference (ICD)]. The former indicates how closely the participants’ responses reflected the presented probabilistic structure, with higher scores denoting better performance. The latter indicates the strategy they used. Strategy ICD values close to 0 reflected a matching strategy, and larger positive values reflected a maximising strategy. Negative values indicated random performance. Participants employ a matching strategy when they select symbols by trying to follow the presented structure as closely as possible, while with the maximisation strategy, participants consistently select the most probable symbol.

The performance index (PI) was calculated as the difference in performance between the final two blocks (Blocks 6 and 7) and the initial two blocks (Blocks 1 and 2). This estimate is preferrable to an average PI score, as it more accurately depicts improvement in performance and hence, learning. Strategy ICD was computed across the seven blocks. Details on the computation of PI and strategy ICD can be found in the original papers by Wang et al. (2017a, 2017b).

2.3. Analysis

In this study, we aimed to assess the influence of multiple multilingual experience factors, attempting to identify those that best capture the effects of multilingualism on cognitive performance. We also tried to evaluate cognitive performance in a thorough and extensive manner, using a variety of tasks of CF, inhibition, WM, and SL.

Multiple linear regression models were used separately for each of the cognitive measures (described in Section 2.2.4), with the language background variables entered as predictors (described in Section 2.2.3), following the methods of recent papers employing a similar approach (Li et al. 2021; Soveri et al. 2011; Yow and Li 2015). We employed separate models for each context; that is, we did not add Country as a predictor in the models but conducted distinct analyses for the UK and Singapore. This is because: (1) the two samples differed significantly regarding most background variables (Age, SES, NVIQ, VIQ); (2) the UK and Singapore are distinct environments with different education systems, which may also have dissimilar impacts on EF development. This means that even if a significant Country effect were to arise, it would not necessarily be meaningful, and it could have been driven by factors outside our control and experimental design. It is the language-related effects present in each country that are of interest to us, rather than the effect of country per se.

Before our main analysis, we performed a data-cleaning procedure on the tasks that included RT indices. Specifically, in TSS, the first trial of each block was removed of the analysis, as it could not be evaluated as either a switch or repeat. When calculating RT costs, we had to remove the error and post-error trials because the type of trial could not be determined after an error had occurred. We then removed responses faster than 300 ms, as well as those that were 2.5 SD slower than the mean. For the Stroop task, we excluded trials with responses faster than 300 ms, as well as those that were 2.5 SD above the mean in the incongruent and congruent condition, respectively.

In the UK group, one participant was excluded from the TSS analysis due to too many excessively slow or fast responses, which together resulted in more than 35% data loss for that participant. One participant was not included in the TMT analysis because due to technical issues, their responses were not recorded. We decided to not exclude the participants completely (i.e., we kept their scores for the tasks they had completed without issues), as removing them would reduce our sample size and because generalised linear models (GLM), which were used for our analysis, can handle missing values.

To simplify the interpretation of the analyses and graphs, we switched the signs in the Accuracy cost measures. Specifically, the Accuracy cost in TSS was calculated as the difference in accuracy between the switch and non-switch trials (or repeat and pure trials, for the mixing cost). As participants generally perform worse on switch trials, this difference (or Accuracy cost) was usually negative, with values closer to (or above) zero, signifying better performance or more flexibility, and more negative values signifying worse performance. However, for RT costs, the opposite pattern emerges: higher RTs (slower responses) are associated with switch trials, so the difference between switch and non-switch trials is usually positive, with values closer to (or below) zero indicating better performance, and more positive values indicating worse performance. Therefore, to avoid confusion, we switched the signs for Accuracy cost so that lower values indicate better performance for this measure as well. The same logic applies for the Stroop task, this time regarding the difference between the incongruent and congruent trials.

The multiple regression models for our main analysis were run in R version 4.2.2 (R Core Team 2021), using the glm() (generalised linear models) function. The best combination of factors was determined using the stepAIC() function of the MASS package in R (Venables and Ripley 2002). The direction of the search was set to “both” (default), so a forward–backward search was performed, which, at each step, decides whether to include or exclude a predictor. A predictor that was included/excluded previously can later be included/excluded, until the lowest Akaike information criterion (AIC) is reached. After identifying the best structure including the language-related predictors of interest, we then performed warranted inclusion and gradually added standardised VIQ and NVIQ scores, age, and SES to the models to determine whether their inclusion would improve the model fit (χ², AIC) (Cunnings 2012). Importantly, there were no issues with collinearity between the background variables and language variables in the models reported. Correlation matrices, with all background and language measures, are included in Appendix C. The plots were generated using the ggplot2 package (Wickham 2016).

3. Results

3.1. Language Profiles in the UK and Singapore

3.1.1. Differences between Groups

The mean scores and standard deviations (SDs) of each group regarding the language background measures of interest are presented in

Table 3. Means and standard deviations (SD) for language background measures.

	UK		Singapore
	Mean	SD	Mean	SD	MW U	p
Age of L2 Acquisition (L2 AoA) (in years)	8.47	4.24	2.98	1.95	178	<0.0001
Usage1	0.87	0.15	0.81	0.16
Usage2	0.13	0.15	0.19	0.16
Proficiency1	5.98	0.07	5.97	0.14
Proficiency2	4.24	2.11	4.39	1.11
Balanced Usage (Usage1 − Usage2) a	0.75	0.30	0.62	0.32	688	0.046
Balanced Proficiency (Proficiency1 − Proficiency2) b	1.74	2.12	1.57	1.11	1038.5	0.295
Language Entropy c	0.50	0.44	0.64	0.39	1068	0.196
Single-Language d	3.86	1.22	3.89	1.01	883.5	0.756
Dual-Language d	2.67	1.28	3.36	0.90	1205.5	0.011
Dense Code-Switching d	2.31	1.14	3.19	0.86	1327.5	0.0003

^a Difference between Usage₁ and Usage₂. Usage₁ indicates use of most frequent language (irrespective of whether it was L1 or L2 or two L1s, in certain cases). Usage₂ indicates use of second most frequent language (irrespective of whether it was L1 or L2). Usage of each language was rated on a scale from 0 to 1. Higher values in Balanced Usage indicate a greater difference between languages, thus, more unbalanced profiles. ^b Difference between Proficiency₁ and Proficiency₂. Proficency₁ indicates the most proficient language. Proficency₂ indicates the second most proficient language. Values of Proficiency in each language ranged from 0 (no proficiency) to 6. Higher values in Balanced Proficiency indicate a greater difference between languages, thus more unbalanced profiles. ^c Higher Entropy scores indicate higher language diversity and more balanced use of languages. Entropy scores are based on all of a participants’ languages. ^d Based on the ACH, participants were asked to rate how frequently they use their languages in a Single-Language, a Dual-Language, and a Dense Code-Switching context. Values range from 1 (never) to 5 (always).

. Pairwise comparisons revealed that the groups differed significantly in: (a) L2 AoA (Mann–Whitney U = 178, p < 0.0001), with the Singapore group acquiring their L2 significantly earlier; (b) Balanced Usage (Mann–Whitney U = 688, p = 0.046), with the Singapore group being more balanced in the use of their two languages; (c) Dual-Language context use (Mann–Whitney U = 1205.5, p = 0.011), with the Singapore group using their languages in a dual-language context more frequently; (d) Code-Switching (Mann–Whitney U = 1327.5, p = 0.0003), with the Singapore group using their languages in a dense code-switching context more frequently.

3.1.2. Differences between Groups

The different language variables of interest showed different correlations in the two groups. These can be found in

Table 4. Correlation matrix for the language background measures in the UK sample.

	1	2	3	4	5	6	7
Balanced Usage	–
Language Entropy	−0.85 ***	–
Balanced Proficiency	0.54 ***	−0.68 ***	–
L2 AoA	0.11	−0.20	0.45 *	–
Single-Language	0.03	0.02	0.02	−0.02	–
Dual-Language	−0.34	0.34	−0.32	−0.35	0.05	–
Code-Switching	−0.33	0.28	−0.27	−0.35	−0.10	0.80 ***	–

Asterisks indicate significance after Holm–Bonferroni correction (Holm 1979). * p < 0.05. *** p < 0.001.

for the UK sample and

Table 5. Correlation matrix for the language background measures in the SG sample.

	1	2	3	4	5	6	7
Balanced Usage	–
Language Entropy	−0.84 ***	–
Balanced Proficiency	0.13	−0.17	–
L2 AoA	−0.08	−0.03	0.18	–
Single-Language	0.03	0.14	−0.12	−0.07	–
Dual-Language	−0.36	0.35	−0.16	−0.05	−0.24	–
Code-Switching	−0.41	0.26	−0.17	0.02	−0.14	0.69 ***	–

Asterisks indicate significance (p < 0.05) after Holm–Bonferroni correction. ** p < 0.01. *** p < 0.001.

for the Singapore sample.

• UK Sample

Balanced Usage significantly correlated with Language Entropy (r = −0.85, p < 0.001), even after applying a Holm–Bonferroni correction for multiple comparisons. This is not unexpected, as both indices reflect balance of use, but the Language Entropy index is a slightly “richer” metric, as it encompasses all languages an individual uses and not just two. Both Balanced Usage and Language Entropy significantly correlated with Balanced Proficiency (r = 0.54 and r = −0.68 respectively, ps < 0.001). Balanced Proficiency also correlated with L2 AoA (r = −0.85, p = 0.032), with more unbalanced participants indicating a later L2 AoA. Dual-Language context use significantly correlated with Dense Code-Switching (r = 0.80, p < 0.001). Finally, Dual-Language context use and Code-Switching showed high correlations with Balanced Usage, Entropy, and Balanced Proficiency, but these were not significant after applying the Holm–Bonferroni correction (p > 0.05). All correlations for the UK sample can be seen in Table 4.

• Singapore Sample

Similar to the results for the UK sample, Balanced Usage significantly correlated with Language Entropy (r = −0.83, p < 0.001) in the Singapore sample as well, even after applying a Holm–Bonferroni correction for multiple comparisons. Dual-Language context use significantly correlated with Dense Code-Switching (r = 0.69, p < 0.001). Finally, Dual-Language context use and Code-Switching showed high correlations with Balanced Usage and Entropy and Balanced Proficiency, but these were not significant after applying the Holm–Bonferroni correction (p > 0.05). No other correlations were found to be significant in the Singapore sample. All correlations for the Singapore sample can be seen in Table 5.

3.2. Language Effects on Cognitive Performance

As some of our language background variables correlated significantly with each other, potentially leading to issues of multicollinearity (Dormann et al. 2013), we performed a generalised linear regression, using only specific combinations of factors, to produce feasible and interpretable models (Kuhn and Johnson 2013; Yow and Li 2015).

The combinations for the UK sample were the following:

• Language Entropy and L2 AoA;
• Balanced Proficiency, Single-Language context use, Code-Switching;
• Balanced Proficiency, Single-Language context use, Dual-Language context use.

Language Entropy significantly correlated with Balanced Usage, and as Entropy is a richer metric, we decided to use it in our models. In addition, Entropy significantly correlated with Balanced Proficiency, so they were not entered in the same models. Correlations with Dual-Language and Code-Switching did not reach significance after performing the Holm–Bonferroni correction, but they were still high, so we chose not to include these indices in the same model. Models including combinations ii and iii did not differ significantly, so we will only report models with Code-Switching as a predictor, where relevant.

The combinations for the Singapore sample were the following:

• Language Entropy, L2 AoA, Balanced Proficiency, Single-Language context use, Code-Switching;
• Language Entropy, L2 AoA, Balanced Proficiency, Single-Language context use, Dual-Language context use.

Again, as Language Entropy significantly correlated with Balanced Usage, we only used Entropy in our models. The correlations present in the UK sample did not apply in the Singapore sample, so we could add all the predictors in one model, with the exception of Code-Switching and Dual-Language context, which again correlated significantly. We therefore ran separate models with Code-Switching and Dual-Language Switching, but again, these yielded virtually identical results, so we will only report models with Code-Switching as a predictor. As suggested by Ooi et al. (2018), dual-language and dense-code switching contexts are difficult to distinguish in the linguistic environment of Singapore, as it is very likely that individuals engage in both contexts with similar frequency.

As described in Section 2.3, the best combination of predictors was identified using stepwise model selection with AIC values. Importantly, after entering the language-related predictors of interest, we performed warranted inclusion of age, SES, NVIQ, and VIQ to assess whether any language effects remained after controlling for potential confounding variables. We report the best models for each cognitive measure (lowest AIC). Only models that were significant and improved the base model (intercept only) are reported. Full model results can be found in Appendix D.

3.2.1. UK Sample

1.

Cognitive Flexibility (CF)

• Task-Set Switching (TSS)

• Accuracy cost

For the TSS task, the best model for Accuracy cost included VIQ as a predictor [AIC = −104.46, F(1,48) = 7.84, p = 0.007, R² = 0.14, R²_adjusted = 0.12], and the VIQ coefficient was significant (β = 0.003, t = 2870; p = 0.007).

2.

Working Memory (WM)

• Spatial Working Memory (SWM)

• Strategy score

The best model for SWM Strategy included L2 AoA, Age, SES, NVIQ, and VIQ [AIC = 251.78, F(5,39) = 4.03, p = 0.005, R² = 0.34, R²_adjusted = 0.26]. The effect of SES (β = −9.861, t = −3.081; p = 0.004) and VIQ (β = −0.113, t = −2.105; p = 0.042) were significant, with higher SES and VIQ leading to better Strategy. None of the language background measures had a significant effect.

However, as there was a correlation between Age and VIQ (which did not remain significant after correcting for multiple comparisons), we also ran separate models. The model including VIQ was better than the model including Age [AIC = 252.97, F(4,40) = 4.13, p = 0.007, R² = 0.29, R²_adjusted = 0.22]. The remaining outputs were virtually the same as those for the previously reported model. Both can be found in Appendix D.

3.

Structure Learning (SL)

• SL Performance Index (PI)

The best model for SL PI improvement included Balanced Proficiency, Code-Switching, and VIQ [AIC = −68.21, F(5,39) = 6.26, p = 0.001, R² = 0.29, R²_adjusted = 0.24]. The effects of Balanced Proficiency and VIQ were significant (β = −0.022, t = 2.635, p = 0.011; β = 0.006, t = 3.486, p = 0.001). Higher Balanced Proficiency scores indicate less balanced profiles, so participants with less balanced proficiency showed reduced learning compared to the results for the more balanced profiles. Higher VIQ also correlated with increased performance.

However, as there were correlations between Balanced Proficiency and VIQ, and Code-Switching and VIQ (which did not remain significant after correcting for multiple comparisons), we also ran separate models. The model including VIQ only was the best [AIC = −64.52, F(1,49) = 9.98, p = 0.003, R² = 0.17, R²_adjusted = 0.15]. The effect of VIQ was significant (β = 0.005, t = 3.158, p = 0.003). The outputs for both models can be found in Appendix D.

The significant models for the UK sample are summarised in

Table 6. Summary of the significant multiple regression models for the UK sample: background variables as predictors of performance in EF and SL tasks.

Variable	TSS Accuracy Cost	SWM Strategy	SL PI
	β	β	β
(Intercept)	−0.082	26.193 ***	−0.184
L2 AoA	–	−0.095	–
Balanced Proficiency	–	–	−0.022 *
Code-Switching	–	–	−0.021
VIQ	0.003 **	−0.081	0.006 **
NVIQ	–	−3.867	–
SES	–	−9.915 **	–
R2	0.14	0.29	0.29
F	7.84 **	4.13 **	6.26 **

Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

.

3.2.2. Singapore Sample

1.

Cognitive Flexibility (CF)

• Trail Making Test (TMT) B:A Ratio

The best model for TMT B:A Ratio included L2 AoA and Balanced Proficiency [AIC = 27.80, F(2,33) = 5.90, p = 0.006, R² = 0.26, R²_adjusted = 0.22]. The effect of Balanced Proficiency was significant (β = 0.123, t = 2.377, p = 0.002), and the effect of L2 AoA approached significance (β = 0.059, t = 2.009, p = 0.053). Participants with more balanced proficiency showed a reduced B:A ratio and thus, greater flexibility. Earlier L2 AoA also resulted in increased flexibility (but only marginally significantly).

• Intra-Extra Dimensional (IED) Set Shift

• Total Errors

The best model for Total Errors in the IED task included Code-Switching, Single-Language context use, and Age [AIC = 349.23, F(3,32) = 3.632, p = 0.023, R² = 0.25, R²_adjusted = 0.18]. Only the effect of Code-Switching was significant (β = 15.705, t = 2.753, p = 0.010). Importantly, our model suggests that individuals who engage in more Code-Switching make more errors in the IED task (see Figure 9).

After visual observation of the effect of Code-Switching on total errors, we thought that the effect may be driven by extreme outliers. We therefore conducted the analysis again, after removing the outliers (i.e., scores that fell into the 1.5 inter-quartile ranges (IQR) below/above the 1st/3rd quartile). The effect of Code-Switching disappeared. The best model [AIC =177.47, F(1,29) = 6.024, p = 0.020, R² = 0.17, R²_adjusted = 0.14], after removing the outliers, included VIQ only (β = 0.224, t = 2.454, p = 0.02).

• Extra-Dimensional Shift Errors

The best model for Extra-Dimensional Shift Errors included Entropy, Age, and NVIQ [AIC = 238.48, F(3,32) = 8.618, p = 0.0002, R² = 0.45, R²_adjusted = 0.40]. All coefficients were found significant: Entropy (β = 11.366, t = 4.090, p = 0.0002), Age (β = 1.362, t = 2.823, p = 0.008), and NVIQ (β = −11.077, t = −2.432, p = 0.021). Higher Entropy scores led to more errors, as did higher age. Higher NVIQ led to reduced number of Extra-Dimensional Shift Errors (see Figure 10).

2.

Inhibition

• Stroop Task

• RT cost

For the Stroop RT cost, the best model included Balanced Proficiency, Single-Language context use, and Code-Switching [AIC = 428.86, F(3,32) = 6.748, p = 0.001, R² = 0.39, R²_adjusted = 0.33]. The effect of Balanced Proficiency was significant (β = 29.57, t = 2.192, p = 0.036), with an increase in Stroop RT cost for participants with less balanced proficiency. The effects of Code-Switching (β = 43.94, t = 2.505, p = 0.018) and Single-Language context (β = 56.83, t = 3.850, p = 0.0005) were both significant. Interestingly, despite the two types of contexts negatively correlating with each other, they both led to increased RT costs in the Stroop task. In other words, both engaging in dense code-switching and using language(s) in a single-language mode seemed to negatively impact inhibitory control, as measured by the Stroop task, yet the effect of single-language use was stronger (see Figure 11).

3.

Working Memory (WM)

• Spatial Working Memory (SWM)

• Total Errors

The best model for SWM total errors included Entropy and SES [AIC = 318.22, F(2,33) = 3.21, p = 0.053, R² = 0.16, R²_adjusted = 0.11]. In fact, the model only approached significance (p = 0.053). The coefficients were not significant, but Entropy negatively correlated with the number of errors; that is, there was a tendency for participants with more linguistically diverse profiles to make fewer errors, but the effect did not reach significance (β = −13.456, t = −1.638, p = 0.11).

• Strategy score

The best model for SWM Strategy also included Entropy and SES [AIC = 205.07, F(2,33) = 6.005, p = 0.006, R² = 0.27, R²_adjusted = 0.22]. The model was significant, as was the coefficient for SES (β = 9.473, t = 2.461, p = 0.019), with a higher SES correlating with a higher Strategy score, that is, worse strategy. The effect of Entropy observed on total errors was clearer here, as higher Entropy scores almost significantly correlated with better SWM Strategy (β = −3.426, t = −2.007, p = 0.053).

• Verbal Working Memory (VWM)

• Backward Digit Span (BDS)

For the BDS task, the best model included Code-Switching, SES, and NVIQ [AIC = 181.81, F(3,32) = 4.56, p = 0.009, R² = 0.30, R²_adjusted = 0.23]. Only the intercept for NVIQ was significant (β = 6.677, t = 3.233, p = 0.003). NVIQ significantly predicted length of backward digit span.

Structure Learning (SL)

• SL Performance Index (PI)

The best model for SL PI in the Singapore sample included Entropy and SES [AIC = −51.40, F(2,33) = 12.75, p < 0.0001, R² = 0.44, R²_adjusted = 0.40]. Both factors had a significant and positive effect on SL PI improvement (Entropy: β = 0.179, t = 3.693, p = 0.0008; SES: β = 0.437, t = 3.999, p = 0.0003). This means that participants with more diverse linguistic profiles showed greater improvement in the SL task.

• SL Strategy ICD

The model for SL Strategy ICD including Entropy approached significance [AIC = −34.66, F(1,34) = 3.589, p = 0.067, R² = 0.10, R²_adjusted = 0.07]. Higher Entropy scores correlated with higher Strategy ICD scores, reflecting a maximisation strategy (Entropy: β = 0.116, t = 1.894, p = 0.0667).

The significant models for the Singapore sample are summarised in

Table 7. Summary of the significant multiple regression models for the Singapore sample: background variables as predictors of performance in EF and SL tasks.

Variable	TMT B:A Ratio	IED Total Errors	IED ED Errors	Stroop RT Cost	SWM Strategy	BDS	SL PI
	β	β	β	β	β	β	β
(Intercept)	0.86 ***	0.39	−24.93 *	−35.41	11.84 ***	10.66 ***	−0.23 ***
L2 AoA	0.06	–	–	–	–	–	–
Balanced Proficiency	0.12 *	–	–	29.57 *	–	–	–
Language Entropy	–	–	11.37 ***	–	−3.43	–	0.18 ***
Single-Language Context	–	–	–	56.83 ***	–	–	–
Code-Switching	–	–	–	43.94 *	–	−0.94	–
VIQ	–	0.22 **	–	–	–	–	–
NVIQ	–	–	−11.08 *	–	–	6.68 **	–
SES	–	–	–	–	9.47 *	−4.50	0.44 ***
Age	–	–	1.36 **	–	–	–	–
R2	0.26	0.17	0.45	0.39	0.27	0.20	0.44
F	5.90 **	6.02 *	8.62 **	6.75 **	6.01 **	4.56 **	12.75 ***

Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

.

4. Discussion

The aim of this study was twofold. First, after operationalising linguistic experience as a multifactorial continuum, we assessed differences in the linguistic profiles of participants from the UK and Singapore. Second, we explored whether these multilingualism experience factors correlate with cognitive performance in several tasks, and if the relationship varies in the different contexts.

4.1. Linguistic Profiles

In terms of linguistic profiles, the UK and Singapore participants differed in three important aspects. The Singapore group acquired their L27 at 2.98 years of age on average, significantly earlier than did the UK group, which, on average, acquired an L2 at 8.46 years. Participants from Singapore also exhibited a more balanced language usage, measured as the difference between an individual’s first and second most used languages (weighted against each other). Higher scores on the Balanced Usage metric reflect a greater difference between a participant’s first and second most used languages, hence less balance, and the average score for UK participants was 0.74 vs. 0.62 for Singapore participants. Finally, the participants from Singapore reported a higher engagement in dual-language use and dense code-switching contexts compared to the UK participants.

The above patterns align with the sociolinguistic realities of the UK and Singapore and the expectations that come with them. The environment of Singapore is highly multilingual, with the various ethnic and linguistic groups interacting frequently.8 Due to Singapore’s bilingual education system, people are at least bilingual, but often trilingual or even multilingual. Studies on the linguistic environment of Singapore suggest that languages are acquired very early (if not from birth), partly through immersion and partly through instruction (Leimgruber 2013; Ooi et al. 2018; Siemund et al. 2014). This is in line with our findings, as most participants reported acquiring two languages from birth, while the rest report acquiring a second language before the age of 8. Therefore, the significantly earlier L2 AoA for the Singapore group is expected. Notably, the UK group included participants who were not UK nationals but who resided in the UK to attend university. These individuals learned English at a young age, usually at the beginning of formal education. If our sample was confined to UK nationals, the difference in L2 AoA would likely be more pronounced.

The difference in Balanced Usage was also expected. In recent years, English has become more prominent in Singapore and is endorsed as the main working, educational, administrative, and governmental language of the country. Still, three additional languages are listed as official in the Constitution, providing recognition to the three major ethnic groups (Chinese, Malays, and Indians) (Leimgruber 2013). It is therefore not surprising that our Singapore sample exhibited more balanced language use (i.e., greater use of a second language) than our UK sample. Despite our inclusion of non-UK nationals, who were usually highly proficient English-as-an-L2 speakers, the monolingual environment of the UK does not provide many occasions in which these speakers could use their L1. These were restricted to communications with family and friends from home (as many stated in the questionnaire). Interestingly, while the Singapore sample also had higher Entropy scores than did the UK sample, the difference between the two groups did not reach significance. Entropy was calculated based on the active use of all an individual’s languages (not just the two most used, as in the Balanced Usage measure). This suggests that despite exposure to diverse languages, Singaporeans might primarily engage in bilingual rather than multilingual active language use. This is line in with the data from recent literature (Cavallaro and Chin 2014; Siemund et al. 2014) and underscores the need for more refined instruments for capturing both active use and passive exposure.

Greater engagement in dual-language and dense code-switching contexts among Singapore participants is consistent with the linguistic diversity and code-switching prevalence in the country (Ong and Zhang 2010; Xie and Cavallaro 2016). In addition, the use of Singlish is widespread. While our questionnaire did not include a question explicitly regarding Singlish, Singaporeans could consider speaking Singlish as speaking one language, with frequent code-switching. This could further explain the significant difference between Singapore and the UK participants. In any case, opportunities for code-switching for bi-/multilinguals residing in the UK are far more limited.

4.2. Relationship between Language Experience and Cognitive Control in Different Contexts

Our second aim was to assess the relationship between different multilingualism experience factors (L2 AoA, Balanced Proficiency, Balanced Usage, Language Entropy, Context of Use) and cognitive performance (on EF and SL tasks), taking the sociolinguistic context (UK vs. Singapore) into account.

In the UK sample, multilingualism factors had a limited impact on cognitive control, with significance observed only in the Structure Learning (SL) task. Specifically, Balanced Proficiency significantly predicted SL Performance Index (PI) improvement. This means that participants with more balanced proficiency had a greater capacity for learning the stochastic sequences presented in the SL task. Prior literature supports the link between bilingual proficiency and statistical learning (Bartolotti et al. 2011; Onnis et al. 2018; Potter et al. 2017). The interesting element of our SL task is that it involves learning stochastic, that is, non-deterministic, sequences of stimuli, without explicit feedback. Participants were, in some sense, required to impose order on chaos and to use subtle cues in their input to form rules. This process is very similar to natural language learning, so it is not surprising that people who have had extensive experience in mastering two languages perform better in such a task. The SL task was the only one to show sensitivity to multiple-language experience in the UK sample, indicating its potential to unveil the so-far undeciphered relationship between multilingualism and cognition. This task has been shown to correlate with the executive and memory-related networks of the brain (Giorgio et al. 2018; Karlaftis et al. 2018), and our analysis shows that it may also relate to linguistic experience. It is therefore a very interesting avenue to explore in future research.9

The image emerging in the Singapore sample is quite different, as more tasks were influenced by multilingualism-related factors. Out of the EF tasks, two measuring CF, the Trail Making Test (TMT) and the Intra-Extra Dimensional (IED) Shift task, revealed significant effects, as did the Stroop task, a task of inhibition, and the SWM task. Significant effects were found in the SL task as well.

In the TMT, a significant effect of Balanced Proficiency emerged, revealing that individuals with greater balance between their two most proficient languages exhibited a reduced B:A ratio, indicative of enhanced flexibility. Literature examining the performance of bilinguals on the TMT is limited, but studies usually do not report an effect, owing to the fact that the switching and visual attention involved in the TMT are arguably not trained through multiple language control (Filippi et al. 2022; Stasenko et al. 2017; Torres et al. 2022). Nevertheless, there are a few studies reporting benefits for bilinguals in different age groups (Bialystok 2010; Estanga et al. 2017), but these usually evaluate the performance in Trail A and Trail B separately, while the B:A ratio is in fact a more accurate measure of CF (Kopp 2011). Notably, our study revealed a significant effect on B:A ratio in the Singapore sample, warranting further exploration to determine generalisability across bi-/multilingual populations. An intriguing hypothesis would be that people in Singapore are often literate in more than one language, and that these languages also have distinct writing systems, thus, these participants may have benefitted from exposure to multiple scripts, positively impacting the visual attention and switching skills involved in the TMT (Yin et al. 2022). Research on the effects of biliteracy and different scripts (alphabetic vs. logographic) on EF is, however, extremely scarce, so this is only speculation.

The effects noted from the results of the IED task lean in a different direction. In the two measures we evaluated, total number of errors and extra-dimensional errors (the latter considered a more accurate measure of CF), increased Code-Switching and higher Entropy (i.e., frequent use of more than two languages) had a negative impact, leading to more errors. Such an effect is not unprecedented. A number of studies suggest that higher contextual or everyday language-switching predicts higher switch costs (Jylkkä et al. 2017, 2021), while other research that found reduced switch costs for bilinguals failed to link them to language-switching frequency or fluency (Woumans et al. 2019). It could be argued that frequent switching or increased exposure to multiple languages are not sufficient, in their own right, to lead to cognitive benefits. In fact, there is evidence that the type of switching (mandatory vs. voluntary) poses different cognitive demands (Jevtović et al. 2020), so freely code-switching between languages, without restrictions (e.g., when both interlocutors speak both languages), may not require significant cognitive effort, thus not leading to cognitive benefits. Nevertheless, inspection of the data revealed a few extreme outliers, so we reran our analysis without them. This made the effect of Code-Switching disappear. We therefore argue that frequent code-switching is not necessarily detrimental, but that this may be the case for a few individuals. This calls for a more refined evaluation of code-switching, as it may not in fact be a uniform practice.

The model for the Stroop RT cost in the Singapore sample was significant, with Balanced Proficiency, Single-Language use, and Code-Switching all reaching levels of significance. Regarding the balanced proficiency effect, greater balance between languages, equivalent to greater proficiency in L2, led to a reduced Stroop effect. This finding is interesting, as literature regarding the effects of L2 proficiency on inhibition is highly controversial. In a study on elderly bilinguals, no effect of L2 proficiency was found, either on a numerical or on a verbal Stroop task (Antón et al. 2016). Neuroimaging evidence suggests that higher L2 proficiency is associated with more automatic and efficient inhibitory control on a Simon task (Jia 2022). A positive effect on ERPs has also been reported in a non-verbal Stroop task, but this effect did not stand up when behavioural measures were considered (Jiao et al. 2019). Crucially, the two studies most similar to ours, employing comparable methodology and focusing on young adults in Singapore, show dissimilar results. In the study using a verbal Stroop task, the same as that used in our study, the effects of L2 AoA and Balanced Usage were found to be significant. The effect of Balanced Proficiency was in the same direction as that noted in our study but did not reach significance (Yow and Li 2015). On the other hand, the study using a numerical Stroop task reported no effect of Balanced Proficiency whatsoever (Li et al. 2021). It is difficult to extract a common thread from this plethora of results, but a trend that seems to emerge is that Balanced Proficiency may be more relevant for versions of the Stroop task with a verbal or linguistic component, rather than for the numerical versions. While not entirely corroborated by the reported literature, this hypothesis is worth exploring further, as it is highly likely that the effects may be, to some extent, task-specific. It is interesting, and curious at the same time, that our results do not fully align with those of other studies assessing very similar samples. We should, however, bear in mind that all our participants reported quite high proficiency in both their languages; hence, there was not much variation in their profiles. In addition, the studies by Li et al. (2021) and Yow and Li (2015) used English–Mandarin bilinguals only, whereas we did not apply such restrictions to our sample. This allows for greater variation in the combinations of languages, which may be closer or further from each other, potentially influencing the magnitude of the effects.

The Single-Language and Dense Code-Switching contexts both led to detrimental effects in the Stroop task; that is, using one’s languages in a single-language mode, as well as frequent code-switching, were both correlated with a greater Stroop effect. Both effects are, in principle, plausible. Operating in a single-language context means that the need to inhibit one’s other language is not regular, so inhibitory control is not exercised frequently. Similarly, unrestricted code-switching, as mentioned, may not pose great cognitive demands. Nevertheless, the fact that the effect of both contexts reached significance is interesting and poses certain challenges for the adaptive control hypothesis (ACH) (Green and Abutalebi 2013). Conceptually, these two contexts are opposing, as frequently operating in single-language mode should equate to infrequent code-switching. While the two measures negatively correlated in our sample, the correlation did not reach significance. In addition, careful observation of our data showed that the individuals who drove the negative effect of code-switching also drove the negative effect of single-language context use. This hints towards two possible explanations: (1) these two contexts are not as easily distinguishable as the ACH posits; (2) our operationalisation of the different contexts was not entirely successful, as participants may not have understood their difference, or may have been unable to place themselves in the most appropriate context, perhaps influenced by their thinking thinking of Singlish. A recent study conducted with young English–Mandarin bilinguals in Singapore reports results very similar to ours, as Single-Language and Dual-Language contexts were not found to be diametrically opposed, and Dual-Language and Dense-Code-Switching contexts correlated highly (Lai and O’Brien 2020). Evidently, the contexts proposed by the ACH may not be completely distinguishable, especially in the case of multilingual populations. The authors posit that three clearly distinct interactional contexts may lack ecological validity in multilingual populations due to the fluidity of communicative environments and the general overlap of language experiences. The ACH may provide a more nuanced approach than a simple comparison of monolinguals and bilinguals, but a categorical operationalisation of linguistic contexts can still obscure intricacies that arise in the natural linguistic ecology of bilinguals (and multilinguals).

Moving to SWM Strategy, the best significant model included SES and Entropy, with Entropy nearly reaching significance (p = 0.053). While there is abundant literature reporting a positive relationship between bilingualism and SWM in multiple age groups (Antón et al. 2019; Blom et al. 2014; Kerrigan et al. 2017; Luo et al. 2013; Morales et al. 2013; Sullivan et al. 2016), the Entropy measure used here more broadly encompasses linguistic diversity, taking all of a participant’s languages into account. A study with young multilingual adults from South Africa found that the addition of a third language had detrimental effects on an extensive set of WM tasks (Cockcroft 2022). This finding is not in line with our results, but it highlights the importance of taking sociolinguistic context into account when examining the relationship between language experience and cognitive processing. Differences in the linguistic realities of Singapore and South Africa may generate different cognitive demands, hence training different cognitive aspects. Importantly, the effect of Entropy in our sample only approached significance, and the Singapore sample was relatively small.

Finally, in the Singapore group, Entropy reached significance in the SL PI model and approached significance in the SL Strategy ICD model (p = 0.067). We have already discussed the importance of the SL task, and our results here further demonstrate its potential in uncovering the relationship between language and cognitive control. Interestingly, in the Singapore sample, the model including Entropy rather than Balanced Proficiency reached significance, unlike the results for the UK sample. This is noteworthy because Entropy encompasses the use of all a participant’s languages, while Balanced Proficiency accounts for proficiency in only the two most proficient languages. The role of sociolinguistic context, therefore, once again becomes apparent, as different multilingual experience factors seem to be pertinent in different contexts. We do not overestimate the effect of Entropy on SL Strategy, as it only approached significance. However, we must highlight that different strategies in the SL task have been associated with the activation of different areas of the brain (Giorgio et al. 2018; Karlaftis et al. 2018), so it is intriguing to consider that linguistic diversity could correlate with distinct brain activation patterns.

5. Conclusions

This study resonates with previous work suggesting that cognitive control is influenced by an interplay of multiple language experience factors (Ooi et al. 2018). We add that this interplay varies as a function of sociolinguistic context. The effects of multilingual experience on cognitive performance were much more evident in our Singapore sample. This underscores the significance of diversifying language research, particularly in the Global South, to avoid a skewed representation based primarily on Western samples (Bak and Alladi 2016; Makoni et al. 2022). In our study, the effects of multilingual experiences on cognitive performance were very limited in the UK sample, but were quite dispersed in the Singapore sample. We therefore support the suggestion of Ooi et al. (2018) that we should be talking of “bilingualisms” (or even “multilingualisms”) rather than simply bilingualism. Further investigations in Singapore, an exceptionally rich multilingual environment, may reveal relationships between language and cognition that have so far been masked.

Another important insight is the great potential of the Structure Learning (SL) task for research on the potential influence of multilingualism on cognitive functioning. Importantly, this was the only task in which language-related effects surfaced for both the UK and Singapore samples. As illustrated, this task relates to both executive function (EF) and language learning. The prospect of SL acting as an intermediary link should be further investigated.

Of course, our study includes certain limitations. Firstly, our UK and Singapore samples were not restricted in terms of their linguistic profiles. This means that our participants varied in the combinations of languages they spoke, which led to less homogeneous samples. While this increased ecological validity, future studies targeting specific linguistic profiles could eliminate the confounding effects of typological divergence in the participants’ languages.

Secondly, the cross-sectional design, though mitigated by controlling for confounding variables, such as SES, age, and intelligence, remains subject to inherent limitations and restricts causal inferences (Laine and Lehtonen 2018). Finally, between-group differences in background variables (e.g., SES, age, intelligence) restricted our ability to perform direct comparisons between Singapore and the UK. It would be interesting to directly compare participants from the UK and Singapore, but there were too many parameters which were out of our control to allow for the performance of this comparison.

Finally, challenges with the operationalisation of the adaptive control hypothesis (ACH) and the measurement of multilingual experience highlight areas for refinement, such as the use of more detailed and continuous measures like the Contextual Linguistic Profile Questionnaire (CLiP-Q), which takes both individual and contextual diversity into account (Wigdorowitz et al. 2020).