Person-specific connectivity mapping uncovers differences of bilingual language experience on brain bases of attention in children

Abstract

Bilingualism influences children’s cognition, yet bilinguals vary greatly in their dual-language experiences. To uncover sources of variation in bilingual and monolingual brain function, the present study used standard analysis and innovative person-specific connectivity models combined with a data-driven grouping algorithm. Children (ages 7–9; N = 52) completed a visuo-spatial attention task while undergoing functional near-infrared spectroscopy neuroimaging. Both bilingual and monolingual groups performed similarly, and engaged bilateral frontal and parietal regions. However, bilinguals showed greater brain activity than monolinguals in left frontal and parietal regions. Connectivity models revealed two empirically-derived subgroups. One subgroup was composed of monolinguals and bilinguals who were more English dominant, and showed left frontal-parietal connections. The other was composed of bilinguals who were balanced in their dual-language abilities and showed left frontal lobe connections. The findings inform how individual variation in early language experiences influences children’s emerging cortical networks for executive function, and reveal efficacy of data-driven approaches.

1. Introduction

Bilingualism is a common early-life experience that influences children’s cognition and brain development. For instance, in searching for word meaning, bilinguals often consider lexical items in both languages upon hearing the onset of a word (pear and perro [dog in Spanish]; Blumenfeld & Marian, 2013). Some suggest that this bilingual-specific experience in language selection may increase the demand for Executive Function mechanisms, and in turn enhance cognitive control performance and their associated neural networks (Kroll & Bialystok, 2013; Kroll, Dussias, Bice, & Perrotti, 2015; Marian et al., 2014; 2017). Nevertheless, there is much inconsistency and debate on whether bilinguals differ from monolinguals in task performance and brain processes (Duñabeitia & Carreiras, 2015; García-Pentón, Fernández García, Costello, Duñabeitia, & Carreiras, 2015; Paap, 2016, 2019; Paap, Johnson, & Sawi, 2015). One core possibility for this inconsistency is the wide heterogeneity in bilinguals’ language experiences (Berken, Chai, Chen, Gracco, & Klein, 2016; Luk & Bialystok, 2013; DeLuca, Rothman, Bialystok, & Pliatsikas, 2019, 2020; Sulpizio, Del Maschio, Del Mauro, Fedeli, & Abutalebi, 2020). Fortunately, person-specific functional connectivity methods can shed light on the potential sources of individual differences in brain function (e.g., Grant, Fang, & Li, 2015). The present study examines the impact of language experiences on neuro-cognitive development by comparing bilingual and monolingual children using standard and innovative person-specific analytical approaches.

1.1. Impact of bilingualism on cognition & brain development

Recent work suggests that bilingualism may incur a set of changes in the responsiveness of functional brain networks (Abutalebi & Green, 2016; Zhang, Wu, & Thierry, 2020). The Neuroemergentism account posits that these neural changes emerge gradually in the bilingual brain (Hernandez et al., 2018, 2019): first, in sub-cortical regions (e.g., basal ganglia) during infancy periods of language acquisition (Stocco, Yamasaki, Natalenko, & Prat, 2014), and over time in broader structures (e.g., frontal cortex) known to support top-down domain-general mechanisms (Grundy, Anderson, & Bialystok, 2017). With increasing age (maturation and experiences), there is greater involvement of top-down domain-general mechanisms, such as during lexical retrieval, and therefore brain adaptations emerge from bilingual experiences in broader structures such as the prefrontal cortex (Hernandez et al., 2018; Grundy, Anderson, & Bialystok, 2017). With prolonged bilingual experience, changes in the brain may give rise to a distinct set of cognitive abilities and/or a distinct set of brain networks in cognitive processing. These differences may stem from the increased attentional demand in dual-language usage, thereby expanding their influence on children’s development of domain-general cognitive mechanisms (Hernandez et al., 2018, 2019).

A number of studies document bilinguals’ better task performance than monolinguals on attentional control tasks prior to age 5 (Bialystok, 1999; Kovács & Mehler, 2009; Singh et al., 2015; Tran et al., 2015, 2019; Yang et al., 2011), yet inconsistent results with children ages 5 and older have fueled a debate on the reliability of a bilingual advantage in task performance (Antón et al., 2014; de Bruin, Treccani, & Della Sala, 2015). Attentional control (i.e., the ability to focus selectively, ignore unnecessary information, shift and re-orient focus) is one Executive Function mechanism facilitating the processing and selection of linguistic items that require conflict resolution. A growing number of neuroimaging studies document brain differences in how bilinguals engage non-linguistic attentional control systems, relative to monolinguals, even when there are no differences in task performance between the language groups (Arredondo, Aslin, & Werker, 2021; Arredondo, Hu, Satterfield, & Kovelman, 2017; Mohades et al., 2014; Garbin et al., 2010; DeLuca et al., 2020; Rodríguez-Pujadas et al., 2013; Costumero et al., 2015). Using non-linguistic cognitive control tasks, these studies find that bilingual school-age children engage left frontal regions (known to activate for language processes), while monolinguals activate right frontal regions (Arredondo et al., 2017; Mohades et al., 2014). Importantly, bilinguals’ left frontal brain activity is associated with second-language competence, such that those with greater competence show reduced activation in left frontal regions (Arredondo et al., 2017). Greater left frontal activity in attention processes is exhibited in young bilingual-learning infants as well, and their left frontal activity is associated with the extent of the biligual parent’s reported language switching behavior (Arredondo et al., 2021). However, contradictory evidence with preschoolers (ages 3–5) attending second-language immersion programs reveals monolingual-like brain activity (Moriguchi & Lertladaluck, 2019; Li et al., 2020). One possibility is that the discrepancy in results is due to differences in bilingual experiences and depend on the extent of the bilingual environment (e.g., immersion school vs. at-home bilingual environments). Thus, an important limitation of these studies is the lack of assessments on aspects of the bilingual experience that likely yield changes in the child brain.

1.2. Understanding individual differences in bilingual adaptations

No two bilingual children are the same, and the field recognizes the wide heterogeneity of bilingual experiences by advocating for a move away from categorical comparisons (Luk & Bialystok, 2013; Pliatsikas, 2020). Researchers are emphasizing the critical need for understanding how individual differences in language experiences, such as the amount of time a child spends using each of their languages or experiences in bilingual versus monolingual social settings, combine to modulate children’s brain and cognition (Berken et al., 2016; Luk & Bialystok, 2013; DeLuca et al., 2020; Sulpizio et al., 2020). Consistent with this, several recent studies report differences in neural responses when bilingualism is treated as a continuous versus a categorical variable (Arredondo et al., 2021; Grundy, Pavlenko, & Bialystok, 2020; Dash, Berroir, Joanette, & Ansaldo, 2019; DeLuca, Rothman, Bialystok, & Pliatsikas, 2019, 2020). For instance, variance in patterns of brain activity in bilinguals’ language processing and production have been shown to depend largely on individuals’ age of second language exposure and/or dual-language proficiency (see review by Li, Legault, & Litcofsky, 2014). Patterns of functional brain activity have also been shown to vary during an attentional control task, with differences associated with duration of bilingual experience (such as length of immersion in the second language, age of acquisition) and with the extent of second language use at home or socially within the community (DeLuca et al., 2020). Despite the importance of this work for highlighting variability in non-linguistic cognitive functions, it has largely been conducted with adult participants, and therefore it is unclear whether these bilingual cognitive adaptations are present – or even more variable – during childhood when the brain is still undergoing rapid developmental changes.

The emerging architecture of bilingual children’s neural organization can provide new insight into which bilingual experiential variables have the greatest impact on the developing brain. In other words, some bilingual experiences might be especially potent at exerting neurodevelopmental changes that may go unnoticed in the adult brain or in dichotomized group comparisons. For instance, during language tasks, children exposed to two languages from birth show greater activity in brain regions associated with language (frontal and temporal), relative to later-exposed (ages 7–10) bilinguals and monolinguals (Jasinska & Petitto, 2013; see also Archila-Suerte, Zevin, Ramos, & Hernandez, 2013). While age of acquisition or first exposure to a new language is an important aspect of the bilingual experience, it is also difficult to assess in Spanish-English bilingual children growing up in the United States, because the amount of exposure to the two languages varies across development. That is, young bilingual heritage speakers are increasingly exposed to the majority language (i.e., English) with increasing age. In addition, Bedore et al. (2016) found that variance in language task performance varied between 1st and 3rd graders: age of English exposure explained a large portion of the variance in performance for 1st graders, while amount of Spanish exposure explained a large portion of the variance for 3rd graders (Bedore, Peña, Griffin, & Hixon, 2016). Thus, experience-based variables in non-linguistic cognitive functions may also differ in their sensitivity over the lifespan, while also depending on the characteristics of the sample.

1.3. The present study

It is generally agreed that early life experiences can have a substantial impact on children’s cognitive development and brain function; yet there is ongoing disagreement on the specific impacts of bilingualism and the ways they vary across individual children. Analytical procedures that compare averages in monolinguals and bilinguals assume homogeneity, and obscure heterogeneity and individual differences within a group (Molenaar, 2004). Person-specific functional connectivity mapping can overcome the limitations of analyses based on averages and provide insight into individual differences by capturing the dynamic relations among brain regions for an individual child. Person-specific functional connectivity mapping has major advantages over traditional approaches since it captures the dynamic relations among brain regions for an individual, however, person-specific functional connectivity work on bilingualism is still scarce. Fortunately, Group Interactive Multiple Model Estimation (GIMME) can provide such a mapping through data-driven approaches, as it generates person-specific connectivity maps that identify unique patterns of functional connectivity in individuals, but also neural similarities among subsets of individuals that will ultimately facilitate group inference-making (Gates & Molenaar, 2012; Beltz & Gates, 2017). In the present study, we apply GIMME to understand the brain activity of a narrow-age child sample and link it to sources of variability within the bilingual experience.

Using functional near-infrared spectroscopy (fNIRS) neuroimaging, we recorded brain activity on early-exposed Spanish-English bilinguals and English monolinguals (ages 7–9) while they completed the Attentional Network Task (ANT; Fan et al., 2005; Rueda et al., 2004). The ANT (Fan et al., 2002) is a common visuospatial measure for studying children’s (Zelazo et al., 2013) and bilinguals’ attentional control, because it is devoid of language and it is thought to tap into key components of attentional networks (cf. Posner, 2011; see bilingual work Arredondo et al., 2017; Costa et al., 2008; Kapa & Colombo, 2013; Tran et al., 2015; Yang, Yang, & Lust, 2011). In the present study, we use a combination of standard contrast analyses and a functional connectivity approach (i.e., GIMME) to understand the effects of bilingualism – and its heterogeneity – on children’s neural representation for attentional control. The group-level GIMME map accounts for homogeneity across participants, while the subgroup-level GIMME map accounts for differences that are specific to a smaller group of (or even single) individuals. Therefore, we were especially interested in whether a data-driven approach could identify monolingual and bilingual children based solely on their patterns of neural activity and connectivity. We hypothesized that given the inconclusive evidence on task performance, bilingualism may not necessarily alter children’s behavior (accuracy and/or response time) during the ANT. However, the advanced attentional demands of dual-language contexts may yield greater engagement of the left hemisphere “language” regions and influence the functional organization of frontal-parietal networks for attention, in comparison to monolingual peers. Group differences in functional organization would be revealed by both standard and connectivity analytical approaches. In addition, any differences in the connectivity maps that may emerge in the form of subgroups may be compared with performanceoin the ANT, cognitive and language assessments (e.g., proficiency), as well as bilingual experiences or demographic characteristics. To the best of our knowledge, this is the first study to use GIMME to explore the bilingual brain phenomena and therefore these analyses are ultimately exploratory in nature, helping to provide novel insights and to guide future, confirmatory research.

2. Methods

2.1. Participants

Fifty-two children (age range = 7.1 – 9.9 years-old) took part in the study: 26 Spanish-English bilinguals, and 26 age- and gender-matched English monolinguals (both groups included 12 females and 14 males). Participants were recruited from southeast Michigan vicinities (United States), where Hispanics are a minority (5% of the area population) and speaking a language other than English at home is not common (10–20% of the area population; U.S. Census, 2021). Selection criteria for all children included: right-handed, no history of language, cognitive or motor development difficulties or brain injury, and no current regimen of medication affecting brain functioning. Selection criteria for monolingual children included: none-to-little exposure to a second language prior to testing (<2-hours a week). Selection criteria for bilingual children included: Spanish exposure from birth at home from at least one parent, English exposure by age five (on average, 1 ½ years old), current daily exposure to both languages (on average, Spanish in the home and English outside of the home), and three years minimum of English exposure prior to testing (on average, six years). The sample size in the present study was based on prior work using the ANT task with young bilingual and monolingual children (e.g., Tran et al., 2015; Kapa & Colombo, 2013), including neuroimaging studies with children of the same age (e.g., Arredondo et al., 2017) and adult bilinguals (DeLuca et al., 2020). An additional 10 children were tested, but excluded from data analysis due to noisy neuroimaging data (two monolinguals and five bilinguals; see section 2.4.2 for details) and technical issues yielding incomplete neuroimaging data (two monolinguals and one bilingual). Data collection took place between January 2015 and September 2016.

Table 1 provides group demographics, as well as performance in English and Spanish language measures. Regarding demographics, monolingual children were from households of higher income and education levels, than bilingual children (p < 0.05). For all children, English was the language of school instruction, and children typically received a 30-to-60 minutes foreign language class (Spanish or Chinese) per week at their school. Monolinguals and bilinguals did not differ (p > 0.05) in their English language abilities and literacy skills. Bilinguals’ language abilities were better in English than Spanish. At the time of testing, half of the bilingual sample reported attending a Spanish-heritage language learning Saturday school for 2-to-3 hours, which included daily Spanish literacy homework. Most bilingual children were born in the United States, except five who were born in a Spanish-speaking country (three of them moved to the U.S. prior to age 2, one at age 3, and one at age 5). Most bilingual children’s parents, except 3 fathers, were native Spanish speakers and all families reported consistent use of Spanish at home with their child(ren).

Table 1.

Means (with standard deviations) for monolingual and bilingual children’s demographics and task performance for language and cognitive measures.

Measures	Monolinguals	Bilinguals	T-valuese
Demographics
Incomea	8.24 (1.64)	6.54 (2.36)	2.94**
Mother’s educationb	6.96 (1.28)	5.81 (2.55)	2.06*
Father’s educationb	6.65 (1.70)	5.35 (2.81)	2.03*
English Behavioral Measures
Phonological Awarenessc	10.65 (2.56)	11.00 (3.09)	0.44
Vocabularyd	115.23 (11.05)	110.00 (11.75)	1.65
Morpho-syntax (%)	93.00 (5.89)	88.69 (9.40)	1.79
Readingd	119.81 (8.94)	115.69 (10.72)	1.50
Naming Speed – Numbersd	105.31 (10.72)	107.35 (15.31)	0.56
Spanish Behavioral Measures e
Vocabularyd	–	93.23 (9.98)	6.46***e
Morpho-syntax (%)	–	68.63 (22.74)	4 39***e

Notes.

^*p < 0.05,

^**p < 0.01,

^***p < 0.001.

^aOptions for demographic responses on yearly household income were the following: (1) less than $5,000; (2) $5,000 - $11,999; (3) $12,000 - $15,999; (4) $16,000 - $24,999; (5) $25,000 - $34,999; (6) $35,000 - $49,999; (7) 50,000 - $74,999; (8) $75,000 - $99,999; (9) $100,000 and greater. Three families (2 bilingual, 1 monolingual) voluntarily skipped this question.

^bOptions for responses on education were the following: (1) primary school, (2) some secondary school, (3) High school diploma or equivalent (GED), (4) some college, (5) Associate’s degree, (6) Bachelor’s degree, (7) Master’s degree, (8) Doctorate degree [Ph.D], (9) Professional degree [MD, DD, DDS, etc].

^cScores are standardized at a mean of 10 (SD = 3).

^dScores are standardized at a mean of 100 (typical average scores range between 85 and 115).

^eT-values represent independent-samples t-tests that compare measures between bilinguals and monolinguals, except for Spanish behavioral measures. T-values in the rows that include the Spanish behavioral measures represent paired-sample t-tests that compare bilinguals’ Spanish assessments relative to bilinguals’ performance in English during comparable tasks (i.e., English vocabulary vs. Spanish vocabulary, English Morpho-syntax vs. Spanish Morpho-syntax).

2.2. Procedure

Following parental consent and child assent procedures, parents completed questionnaires while their child took part in the testing session. Children first completed English language and reading assessments with an English-speaking researcher; specifically, the assessments measured receptive vocabulary, expressive morpho-syntax, phonological awareness, word reading, and naming speed. Children then completed the fNIRS portion of the study. During the fNIRS session, experimenters set the head-cap and optodes in place, photographed cap placement, and instructed children on the task. Following the fNIRS session, bilinguals completed Spanish language assessments with a native Spanish-speaking researcher; specifically, the assessments measured receptive vocabulary and expressive morpho-syntax. The study was approved by the University of Michigan ethical review boards. Families received monetary compensation, children received a Frisbee and token prizes as a thank you for their participation.

2.3. Measures

2.3.1. Parent questionnaire

Parents completed a questionnaire about their child’s age of first/second language exposure, literacy development in one or both languages (formal instruction, hours per week), language preference (“which language does your child feel most comfortable speaking?”, “if your child is playing alone or talking to him/herself, what language would he/she use?”), and others’ perception about the child’s language abilities (“How do your family and friends perceive your child?” options included “Bilingual,” “English speaker,” and “Spanish speaker”). The questionnaire also requested information on the family’s educational level, household income, the parents’ bilingualism status as well as language use at home and with their child (for more detail see Arredondo, 2017).

2.3.2. Child language and cognitive assessments

2.3.2.1. English receptive vocabulary (KBIT-2 verbal knowledge subtest; Kaufman & Kaufman, 2004).

During testing, the experimenter presented the child with a matrix of six images along with a question or a word, and the participant was asked to point to the picture that represented the work or answered the question the best. Basal and ceiling levels were established as recommended; standard scores based on a mean of 100 (SD = 15) are reported and used in the analyses.

2.3.2.2. English expressive morpho-syntax (CELF-4 word structure subtest; Semel, Wiig, & Secord, 2003).

During testing, the child was asked to complete a sentence that was read by the experimenter. A figure accompanied the prompt. The assessment measured participants’ ability to apply morphology and syntactic rules. Participants earned 1 point for correct items and 32 testing items were presented; proportions are reported and used in the analyses.

2.3.2.3. English word reading (Woodcock reading mastery tests revised 2nd edition, word identification subtest; Woodcock, 1998).

During testing, the experimenter presented the child with a word to read aloud. Basal and ceiling levels were established as recommended; standard scores based on a mean of 100 (SD = 15) are reported and used in the analyses.

2.3.2.4. English phonological awareness (CTOPP elision subtest; Wagner, Torgesen, & Rashotte, 1999).

During testing, the experimenter asked the child to say a word, and repeat it without a portion (e.g., “Say winter, now say winter without saying /t/,” “winner”). Participants earned 1 point for correct items; the task included 6 practice items and 20 testing items. Testing stopped when the ceiling item was reached, or 3 consecutive errors. Standard scores based on a mean of 10 (SD = 3) are reported and used in the analyses.

2.3.2.5. English naming speed (rapid automatized naming, numbers subtest; Wolf & Denckla, 2005).

During testing, children were asked to name 50 numbers on a card as fast as possible; the numbers included: 2, 6, 9, 4, and 7. Standard scores based on a mean of 100 (SD = 15) are reported and used in the analyses.

2.3.2.6. Spanish receptive vocabulary (receptive one-word picture vocabulary test Spanish bilingual edition; Brownell, 2000).

Only bilingual children completed this assessment. The experimenter presented the child with four images and a word, and participants pointed to the representing picture. This measure is normed with Spanish-English bilinguals (over 2000 individuals, ages 2 through 80 +), specifically heritage Spanish speakers, residing in the United States. Basal and ceiling levels were established as recommended; standard scores based on a mean of 100 (SD = 15) are reported and used in the analyses.

2.3.2.7. Spanish expressive morpho-syntax (CELF-4 word structure subtest; Semel et al., 2006).

Only bilingual children completed this assessment. The assessment is similar to the English version (see description 2.3.2.2 above); 29 items were presented and proportions are reported and used in the analyses. This measure is normed with bilingual children, specifically heritage language speakers of Spanish residing in the United States, whose families immigrated from the Caribbean (25%), Central or South America (28%), or Mexico (46%).

2.3.3. Neuroimaging attentional network task (ANT)

Children completed a modified child-friendly version of the ANT (Rueda et al., 2004). The ANT uses a combination of cuing events along with a flanker paradigm, to assess three attentional networks: Alerting, Orienting, and Executive attention (Posner, 1980; Eriksen & Eriksen, 1974). The ANT requires that participants monitor their attention and solve trials with conflicting and non-conflicting information, by attending to the direction (left or right) of a Target and ignoring the direction of surrounding flankers (Fan et al., 2005). A fixation point (+) in the center of the screen was present throughout the task. Each trial first consisted of a cuing event (150-ms), either a Spatial cue (signals a forthcoming trial and upcoming location of the target; top or bottom), a Central cue (signals a forthcoming trial and no information on location of the target), or No cue (no information about a forthcoming trial). A blank screen was presented (400-ms) following the cue. Next, a Congruent (non-conflicting condition) or an Incongruent (conflicting condition) trial was presented either on the top or bottom of the fixation point (1700-ms); see Fig. 1. During Congruent trials, the Target points to the same direction as the flankers (→→→→→or ←←←←←). During Incongruent trials, the Target points to the opposite direction as the flankers (→→←→→or ←←→←←); this condition introduces visuo-spatial conflict.

Fig. 1.

Example trials for the Attentional Network Task.

The task was designed as a rapid event-related experiment. A 6-second rest period was introduced at the beginning and end of each run. Inter-stimulus intervals (ISI; jittered rest periods) were randomized within each run; ISIs were a minimum of 1-second and a maximum of 6-seconds, and their total sum in a run corresponded to 48-seconds (which is 1/3 of the time in a run with no jitter). Trials and ISIs were randomized using OptSeq2 (Dale, 1999). Each run lasted about two minutes, and consisted of 48 randomized trials (24 Congruent and 24 Incongruent). Each cue type was presented for 1/3 of the trials in each run. The entire task (five runs) lasted about 10 min, and consisted of 240 trials (120 Congruent and 120 Incongruent). The task was presented using E-Prime 2 (Psychology Software Tools, Inc.) on a 30-inch HP Z30i LED monitor connected to a Dell Optiplex 780 desktop computer. A two-button control box (Current Designs, Inc.) connected to the computer recorded children’s responses. The children were seated approximately ~ 60 cm from the monitor, and kept their head stable while wearing the fNIRS cap. Performance was assessed by accuracy and response time. Response times were collected at the child’s response and after the onset of the Target and flanker stimuli. If the child did not respond during a trial, the trial was deemed a “miss” and not included in the analysis.

2.4. Data analysis

2.4.1. ANT behavior analysis

Accuracy was examined using a Generalized Linear Mixed Model (GLMM), which can handle data that is not normally distributed such as binomial distributions; specifically, we carried out a repeated-measures binomial logistic regression. The model specified Cuing events (Spatial, Center, No cue) and Congruence conditions (Incongruent, Congruent) as within-subjects repeated factors, language groups (monolinguals, bilinguals) were specified as a between-subjects factor, and participant IDs were specificed as a random factor in the structure. The model included all individual trials for each child in which a response was recorded; that is, 97.5% of the data (12159 trials; 97.9% correct and 2.1% incorrect trials). A total of 2.5% trials were considered missing due to a non-response.

Response time data were positively skewed; thus, the data were transformed to natural logarithm values. Response time was examined using a Linear Mixed Model (LMM) with a normal distribution; specifically, we carried out a repeated-measures multinomial regression. Similar to the accuracy model, Cuing events (Spatial, Center, No cue) and Congruence conditions (Incongruent, Congruent) were specified as within-subjects repeated factors, language groups (monolinguals, bilinguals) were specified as a between-subjects factor, and participant IDs were specified as a random factor in the structure. The model included all individual trials for each child in which a response was recorded. The advantages of using GLMM and LMM include: the handling of non-normally distributed data, the use of all individual data (i.e., trials) for each participant, and the estimation of random effects (i.e., variation within participants across trials) as part of the linear predictor. For more information on the use of GLMM and LMM, see Boisgontier and Cheval (2016), Krueger and Tian (2004), as well as Baayen, Davidson, and Bates (2008).

2.4.2. Neuroimaging data acquisition and preprocessing

We used a TechEN-CW6 fNIRS system to collect data from 44 channels (source-detector connection) spaced 2.7-cm apart and sampled at 50-Hz (22 channels per hemisphere). The cap (see Fig. 2) was applied consistently on each child using the international 10–10 transcranial system positioning (Jurcak, Tsuzuki, & Dan, 2007) for the following points: inion, nasion, auricular left and right, Fz, FpZ, Cz, T7/8, and F7/8. Data quality control was carried out on the 690 and 830 nm wave-lengths timeseries separately, using Homer2 v2.3 (retrieved from the NITRC database; Huppert, Diamond, Franceschini, & Boas, 2009). This approach yields exclusion criteria for children whose signal was below 3 M units and did not reveal cardiac signal for over 50% of 690 data channels, likely due to a large amount of motion artifacts and/or hair obstruction (Arredondo et al., 2017, 2019a; 2019b; Hu et al., 2015).

Fig. 2.

Functional NIRS probe configuration (left hemisphere shown). (A) Dots correspond to optode placements at a distance of ~2.7 cm, over an average brain template (blue circles = sources/emitters of light; green circles = detectors). (B) Probe and channel configuration for left hemisphere, numbers denote connections (channels) between sources and detectors. (C) Brain regions covered by the fNIRS measurement as maximally overlaid by the probe arrangement in the order of greatest probability for each channel (BA = Brodmann Area).

Data were analyzed in MATLAB (2017b, MathWorks) using functions from NIRS Brain AnalyzIR Toolbox (Santosa, Zhai, Fishburn, & Huppert, 2018) and Homer2^′s motion detection (hmrMotionArtifactByChannel) and Spline functions (hmrMotionCorrectSpline). The following preprocessing steps were completed: optical density change data conversion, motion artifact detection and correction via spline interpolation, and concentration change data conversion yielding oxygenated (HbO) and deoxygenated hemoglobin (HbR) values. First, the raw time courses were converted into units of optical density (OD) change. The OD data went through a quality control step for integrity and presence of signal and motion artifacts, on a channel-by-channel basis (Scholkmann, Spichtig, Muehlemann, & Wolf, 2010). Motion artifacts were defined as signal changes with an amplitude greater than 1 threshold of a standard deviation of 50 within half a second and masked for an additional 1 s; motion detection values were specified as suggested by prior methodological work for analyzing fNIRS child data (Hu et al., 2015; Brigadoi et al., 2014). These artifacts were corrected with a spline interpolation set to the 0.99 parameter. Spline interpolation replaces the motion segments by reconstructing the motion and replacing it with a combination of the mean value of the identified segment and the mean value of the previous segment to ensure a continuous signal (Scholkmann et al., 2010). The artifact-corrected OD data were then converted into hemoglobin concentration data using the modified Beer-Lambert law, yielding HbO and HbR values.

2.4.3. Neuroimaging data analyses

2.4.3.1. Standard contrast analyses.

Group-level analyses were conducted to confirm task-related brain activation, and to serve as a comparison to GIMME results. These statistical analyses were evaluated at a False Discovery Rate (FDR) threshold correction of p < 0.05 (Benjamini & Hochberg, 1995). Specifically, the FDR correction was applied to all 44 channels, and in accordance with the NIRS Brain AnalyzIR Toolbox (Santosa et al., 2018) which considers the surrounding fNIRS source-detector connecting pairs for each channel’s data. Below we provide results for HbO, see Appendix for HbR results.

The first step was to examine brain response for each attentional network (alerting, orienting, executive). We applied a General Linear Model (GLM; Barker et al., 2013; Poline & Brett, 2012) ordinary least squares (OLS) fit, and modeled the dual-gamma canonical hemodynamic response function. The GLM estimated beta values for the Cuing conditions (Spatial, Central, and No cue), and for Congruence conditions (Congruent and Incongruent). The following subtractions assess the three attentional networks (Fan et al., 2002, 2005): (i) No cue trials from Center cue trials (Center > No Cue) for Alerting, (ii) Center cue trials from Spatial cue trials (Spatial > Center Cue) for Orienting, and (iii) Congruent trials (across all cue types) from Incongruent trials (Incongruent > Congruent) for Executive.

2.4.3.2. Person-specific connectivity analyses with subgrouping (GIMME).

We used the GIMME data-driven connectivity mapping approach which uses a set of a priori regions of interest (ROIs) and in an iterative manner builds networks at the group-level with connections that are common to everyone in the sample, shared among members of a data-determined portion of the sample (subgroup-level), or unique to a child (individual-level). GIMME is a person-specific method and each participant’s model is based on their own fNIRS time series. To fit these person-specific networks, GIMME uses Lagrange multiplier equivalents tests to show the extent to which the model for that participant would improve if a given connection were estimated. In the present work, we chose a group-level criterion of 90% (see below for more information on this threshold) which considers if the improvement for the same connection in 90% of the participants is significant (i.e., on a chi-square distribution with 1 degree of freedom), then it is estimated in every participant’s network, and the process iterates.

After the group-level network is determined, subgrouping is conducted. Individuals with common network features (based on individual estimates in similarity matrices from the group-level connections) are detected using a Walktrap algorithm, and then, further commonalities among those individuals are detected by estimating any connections that would improve model fit in a majority of individuals’ models (again using Lagrange multiplier equivalents tests). After subgrouping iterations complete, then individual-level connections that would improve a single participant’s network (based on Lagrange multiplier equivalents tests), are iteratively estimated until the model (containing group-, subgroup-, and individual-level connections) fits well.

Using this person-specific approach, GIMME can accurately detect small groups of individuals with unique networks, especially when the number of subgroups in the network is small (see initial simulations in Gates et al., 2017). Moreover, empirical data using subgrouping with GIMME regularly reveals small groups of participants that have distinct neural networks, often in ways related to behavior (see Lane et al., 2019; Price et al., 2017; Price et al., 2020). In addition, largescale simulations have yielded valid and reliable GIMME results for sample sizes of 52 (see Gates & Molenaar, 2012; Gates et al., 2017), and many empirical studies have also utilized the person-specific networks of GIMME in samples of 60 or smaller (see Beltz et al., 2013; Lane et al., 2019; Price et al., 2020; Stout et al., 2021; Weigard et al., 2020).

In order to apply GIMME to the dataset, we defined regions of interest (ROIs) using a different standard analysis than the contrast analyses specified above in Section 2.4.3.1. Specifically, a restricted/residual maximum likelihood (REML) multivariate linear mixed-effects model (LMM; a first-order autoregressive (AR[1]) repeated covariance structure with homogeneous variances) was used to identify brain regions engaged for the Executive Attention Network. We focused on the Executive Attention network, as this was the focus of prior work (e.g., Arredondo et al., 2017). FDR corrections were applied in the REML LMM, in accordance with the NIRS Brain AnalyzIR Toolbox (Santosa et al., 2018).

In GIMME, connections among ROIs at the group-, subgroup-, and individual-level are directed and can be contemporaneous (that is, reflecting directed relations among regions at the same time point), or connections can be lagged (that is, reflecting regions that predict activity in themselves or in other regions at the following time point). Given that the fNIRS temporal resolution is still much higher than prior fMRI work using GIMME, we do not distinguish directionality or lagged nature of the connections, instead we focus on contemporaneous connections. We focus only on contemporaneous (i.e., time-locked) relations to emphasize cross-relations in the person-specific networks. Data with large temporal dependencies, such as fNIRS data, have autocorrelations in individual that and do not provide insight into the neural networks underlying task performance; that information is best provided by cross-relations between ROIs (see Smith et al., 2011; 2012).

The HbO time course from the ANT for each ROI and child (without noting who belonged to the monolingual or bilingual group) were submitted to the freely accessible R-version of GIMME (Lane et al., 2017; https://cran.r-prooject.org/web/packages/gimme). The raw data for GIMME was first down-sampled to 2-hz (as recommended by Beltz & Gates, 2017), prior to preprocessing the data in a similar manner as specified above. GIMME begins with a null model (see Beltz & Gates, 2017 for a detailed review) then adds connections at the group-level (for both monolinguals and bilinguals) that would significantly improve the networks for 90% of the sample; this is a much stricter criterion than most studies using GIMME (i.e., 75%), and ensures that group-level connections represent substantial homogeneity across individuals. The program iterates (i.e., adds a connection, re-estimates the model, adds another connection) until the 90% criterion is not met, and then prunes any connections below the 90% cut-off. A conservative threshold of 90% ensures more homogeneity in the resulting networks than a lower threshold, and thus, permits more straight-forward comparisons with the present mean-based methods used for the standard contrast analysis. Next, GIMME attempts to identify subgroups using information from the group-level maps and the Walktrap community detection algorithm. If subgroups are detected, then connections that would significantly improve the networks for 50% of each subgroup are iteratively added (and pruning is conducted if needed); this is a typical and valid criterion (Beltz & Gates, 2017; Gates, Lane, Varangis, Giovanello, & Guiskewicz, 2017; Price et al., 2017). Subgroups share connectivity patterns, including the directions of connections. Finally, individual-level connections are added if the group-level and subgroup-level connections do not adequately explain the covariation among ROIs for a child.

2.4.3.3. Analyses of connectivity features and links to behavior.

If subgroups were detected, we planned to conduct non-parametric tests to reveal whether one subgroup included a greater proportion of bilinguals or monolinguals than expected by chance. We also planned to examine subgroup differences in features of the connectivity maps, such as in network density (i.e., the number of connections in frontal, parietal, frontal to parietal, and parietal to frontal channels), using independent samples t-tests. Finally, we planned to characterize differences between the subgroups by comparing children’s performance in the ANT, cognitive and language assessments (e.g., proficiency), as well as bilingual experiences or demographic characteristics from the parent reports using parametric and non-parametric statistical tests. All analyses are exploratory, and thus, we do not correct for multiple comparisons.

3. Results

3.1. ANT performance

Independent-sample t-tests did not reveal bilingual versus monolingual group differences in accuracy or response time for any of the Cuing or Congruence conditions, as well as the attentional networks (see Table 2). Accuracy was high for all children (above 90%). The GLMM examined main effects, all two-way interactions and a three-way interaction (language group × Cuing × Congruence). The model for accuracy only revealed a significant main effect of Congruence (χ²(1, 52) = 43.16, p < 0.001) that stemmed from better accuracy in the Congruent than Incongruent trials (see Table 2).

Table 2.

Attentional network task performance means (with standard deviations) for monolingual and bilingual children.

Condition	Monolinguals	Bilinguals
Accuracy (%)
Cuing conditions
No cues	98.08 (2.16)	97.91 (2.44)
Center cues	97.96 (2.24)	97.53 (2.38)
Spatial cues	98.38 (1.85)	97.88 (1.69)
Congruence conditions
Congruent trials	98.82 (1.21)	98.70 (1.25)
Incongruent trials	97.32 (2.50)	96.90 (2.68)
Networks
Alerting Network	−0.12 (2.32)	−0.38 (2.52)
Orienting Network	0.42 (2.32)	0.36 (2.36)
Executive Network	−1.51 (2.00)	−1.80 (2.51)
Response time (ms)
Cuing conditions
No cues	869.70 (80.53)	892.27 (88.10)
Center cues	855.40 (84.03)	871.36 (95.49)
Spatial cues	800.22 (95.70)	834.21 (89.97)
Congruence conditions
Congruent trials	802.75 (85.06)	830.73 (88.21)
Incongruent trials	880.60 (86.05)	904.01 (92.97)
Cue by Congruence conditions
No cue, Congruent	838.98 (85.99)	866.15 (81.98)
No cue, Incongruent	902.24 (82.29)	920.70 (101.10)
Center cue, Congruent	813.78 (86.65)	836.24 (95.78)
Center cue, Incongruent	898.36 (89.31)	908.55 (101.05)
Spatial cue, Congruent	755.03 (92.83)	791.43 (97.11)
Spatial cue, Incongruent	840.77 (99.79)	878.01 (95.17)
Networks
Alerting Network	−14.31 (27.45)	−20.90 (35.67)
Orienting Network	−55.18 (34.87)	−37.15 (37.57)
Executive Network	77.84 (38.05)	73.28 (39.52)

The LMM for reaction time revealed significant effects of Congruence (χ²(1, 52) = 207.29, p < 0.001) and Cue (χ²(2, 52) = 169.27, p < 0.001), as well as a Cue by Congruence interaction (χ²(2, 52) = 22.07, p < 0.001). See Table 2 for reaction time performance. Regarding Congruence trial types, children were faster in Congruent trials than Incongruent trials. Regarding Cue trial types, children were faster in Spatial cue trials, followed by Center cue trials and then No cue trial types (all Bonferroni ps ≤ 0.001). The interaction stemmed from children showing similar response time performance between the most difficult Congruent trial type (i.e., No Cue) and the easiest Incongruent type trial (i.e., Spatial), and showed similar response time between the Center Incongruent trials and No Cue Incongruent trials (all Bonferroni ps ≤ 0.001). The model did not reveal a significant group effect or any other interactions. In sum, bilinguals and monolinguals performed similarly in accuracy and response time.

3.2. Functional neuroimaging

3.2.1. Standard contrast analyses

Results of the standard Attentional Network analyses for oxygenated hemoglobin (HbO) are depicted in Fig. 3 and Table 3; see Appendix Table A.1 for deoxygenated hemoglobin (HbR) results. For the Alerting Network (Center cues > No cues), monolinguals engaged multiple channels on left frontal regions, as well as right frontal and right superior parietal/occipital. Similarly, bilinguals engaged multiple channels on left frontal regions, as well as a channel on left temporal lobe. For the Orienting Network (Spatial cues > Center cues), both groups (monolinguals and bilinguals) did not show greater activity by the Spatial cues. For the Executive Network (Incongruent > Congruent), monolinguals’ brain activity did not pass the FDR threshold at 0.05. In contrast, bilinguals engaged multiple channels on left frontal regions, as well as right frontal and left posterior temporal regions.

Fig. 3.

Contrasts for the Alerting (Center > No cues), Orienting (Spatial > Center cues), and Executive (Incongruent > Congruent) attentional networks for monolinguals and bilinguals. Color bar reflects t-values that passed a false discovery rate (FDR) threshold for multiple comparisons. The Alerting network was corrected at p < 0.01; the Orienting and Executive network were corrected at p<0.05.

Table 3.

Oxygenated hemoglobin results for the attentional networks in child bilinguals and monolinguals.

Hemisphere	Channel	Region (approximation)	Beta (SE)	T
Alerting network (Center cue > No cue)
Monolinguals
R	8	Middle/Superior Frontal Gyrus	30.93 (8.13)	3.80**
R	22	Superior Parietal/Occipital	33.25 (7.38)	4.50***
L	6	Middle Frontal Gyrus	31.41 (9.38)	3.35**
L	7	Middle Frontal Gyrus	55.77 (16.50)	3.38**
L	8	Middle/Superior Frontal Gyrus	39.31 (9.56)	4.11***
L	9	Middle/Superior Frontal Gyrus	38.22 (9.82)	3.89**
L	10	Middle/Inferior Frontal Gyrus	33.60 (9.47)	3.55**
Bilinguals
L	2	Inferior Frontal Gyrus	48.37 (10.82)	4.48***
L	3	Inferior Frontal Gyrus	27.88 (7.55)	3.69**
L	7	Middle Frontal Gyrus	52.40 (15.66)	3.35**
L	8	Middle/Superior Frontal Gyrus	35.92 (9.82)	3.66**
L	9	Middle/Superior Frontal Gyrus	32.71 (10.30)	3.18**
L	10	Middle/Inferior Frontal Gyrus	34.54 (10.37)	3.33**
L	11	Middle/Superior Frontal Gyrus/Precentral	35.79 (10.98)	3.26**
L	14	Superior/Middle Temporal	44.21 (13.08)	3.38**
Orienting network (Spatial cue > Center cue)
Monolinguals
		No significant suprathresholds
Bilinguals
		No significant suprathresholds
Executive network (Incongruent > Congruent)
Monolinguals
		No significant suprathresholds
Bilinguals
R	6	Middle Frontal Gyrus	20.54 (6.50)	3.16**
L	3	Inferior Frontal Gyrus	19.12 (5.78)	3.31**
L	6	Middle Frontal Gyrus	15.96 (6.39)	2.49*
L	8	Middle/Superior Frontal Gyrus	19.53 (7.39)	2.64*
L	12	Superior/Middle Temporal	24.15 (10.09)	2.39*
L	13	Temporo-parietal	23.39 (10.18)	2.30*

Notes. False discovery rate (FDR) thresholds correcting for multiple comparisons at p < 0.05 are reported.

^*p < 0.05

^**p < 0.01

^***p < 0.001.

3.2.2. Defining ROIs

The first step to conducting person-specific GIMME analyses is to identify ROIs that are specific to the sample. Here, we focused on regions associated with the Executive Attentional Network. The LMM revealed that children (bilinguals and monolinguals) showed condition effects (Incongruent > Congruent) in a left frontal channel (Ch 11) and a left posterior parietal channel (Ch 18). The LMM also revealed a main effect of group, with greater activity for bilinguals than monolinguals, in two left frontal channels (Ch 7 and 9) and a left posterior channel (Ch 20). There were no channels showing greater activity for monolinguals as compared to bilinguals. There was also a group by condition interaction in two left frontal channels (Ch 6 and 9), driven by the presence of Congruency effects (Incongruent > Congruent) in these channels in the bilingual group, but not in the monolingual group. See Appendix Figure A.1. In sum, the ROIs used in GIMME person-specific connectivity analyses include four left frontal channels (Ch 6–7, 9 and 11) and two left posterior parietal channels (Ch 18 and 20).

3.2.3. Person-specific connectivity analyses with subgrouping (GIMME)

Person-specific connectivity maps generated by GIMME fit the data well, given that three out of four indices met standards of excellent fit (average: X² = 2287.02, df = 128.84, CFI = 0.97, NNFI = 0.95, RMSEA = 0.09, SRMR = 0.019). A number of models (n = 15) were manually reviewed using LISREL (Jöreskog & Sörbom, 1992) to ensure connection magnitudes were accurately estimated when error variance estimates were high (similar to procedures in Beltz & Molenaar, 2015).

The group-level map revealed two channels in left frontal (Ch 6 and 7) and two channels in left parietal (Ch 18 and 20) that were common across children; see Fig. 4a. As specified above and for clarity purposes, these connections were contemporaneous and show presence of connections. Next, GIMME detected two subgroups and generated corresponding maps. The subgroup-level map for Subgroup 1 revealed three common connections, all of which connected left parietal (Ch 18) to left frontal channels (Ch 6, 9 and 11); see Fig. 4b and 4e. The subgroup-level map for Subgroup 2 revealed six connections, all connected within the left frontal (Ch 7 to Ch 6, Ch 7 to 9, Ch 7 to and 11, Ch 6 and Ch 11, Ch 9 and Ch 11), and one connection between a left frontal channel (Ch 6) and a left parietal channel (Ch 18); see Fig. 4c and 4f. Both subgroups had at least one channel from left frontal to posterior channel 18. Most connections involved left parietal channel 18 for Subgroup 1, and left frontal channels 6, 7, and 11 for Subgroup 2. Similar to the group-level, these connections are contemporaneous and do not distinguish directionality.

Fig. 4.

GIMME was carried out on six regions of interest (ROIs: 4 frontal channels, 2 parietal channels). (a) Group-level brain figure reflects two common contemporaneous connections among monolinguals and bilinguals – channels 6 and 7 are connected in frontal lobe, channels 18 and 20 are connected in parietal lobe. (b) Subgroup 1 brain figure reflects three contemporaneous connections between ROIs – channel 18 in parietal lobe is connected to frontal lobe channels 6, 9 and 11. (c) Subgroup 2 brain figure reflects six contemporaneous connections between ROIs – channels 6, 7, 9 and 11 are connected in frontal lobe, and channel 6 in frontal lobe is connected to channel 18 in parietal lobe. (d) Bar graph reflects split between monolingual and bilingual participants across the two subgroups that emerged from the subgrouping procedure in GIMME. (e) Circles reflect regions of interest and bold lines reflect connections between the ROIs in Subgroup 1. (f) Circles reflect regions of interest and bold lines reflect connections between the ROIs in Subgroup 2.

There were few individual-level connections in Subgroup 1; specifically, two children had an additional connection beyond the group- and subgroup-level connections (i.e., one child had an additional connection in frontal lobe and another child had an additional connection in parietal to frontal). Children in Subgroup 2, however, were more variable in their individual-level connections. Specifically, most children (n = 9, out of 14) had a total of seven connections in the frontal lobe; however, two children had two fewer connections and three children had one fewer connection.

3.2.4. Analyses of connectivity features and links to behavior

See Fig. 4d. Subgroup 1 included the most children (n = 38), and it identified 24 out of 26 monolinguals (92.31%) and 14 out of 26 bilinguals (53.85%). Subgroup 2 was smaller in size (n = 14) and it identified 2 out 26 monolinguals (7.69%) and 12 out of 26 bilinguals (46.15%). There was a statistically significant difference in the proportion of monolingual and bilingual children in the subgroups (χ²(1, 52) = 11.08, p = 0.001); Subgroup 1 included more monolinguals than expected, and Subgroup 2 included more bilinguals than expected.

Participants in Subgroup 2 had networks with greater density (i.e., more connections at the total group-, subgroup-, and individual-level) than Subgroup 1 (Subgroup 1: M = 5.05, SD = 0.22; Subgroup 2: M = 8.50 connections, SD = 0.76; t(50) = 25.44, p < 0.001). Participants in Subgroup 2 also had greater connection density in left frontal regions than Subgroup 1 (Subgroup 1: M = 1.03, SD = 0.16; Subgroup 2: M = 6.5, SD = 0.76; t(50) = 42.53, p < 0.001). Despite the overall lower number of connections in Subgroup 1, participants in that group had significantly more connections from parietal regions to frontal regions than participants in Subgroup 2 (Subgroup 1: M = 2.03, SD = 0.16; Subgroup 2: M = 1.00, SD = 0.00; t(50) = 46.45, p < 0.001). See Fig. 4b–f.

Next, we examined subgroup differences on measures of language, reading and cognitive abilities, as well as ANT performance. Independent samples t-tests did not reveal differences among the subgroups on these variables. Given that most monolinguals (except 2) belonged in Subgroup 1, a sensible next step was to explore differences among the bilinguals in Subgroup 1 versus those in Subgroup 2. Despite all participants performing at ceiling levels during the ANT, analyses revealed that bilinguals in Subgroup 2 were more accurate than bilinguals in Subgroup 1 during Congruent trials (Subgroup 1: M = 98.26, SD = 1.45; Subgroup 2: M = 99.22, SD = 0.68; t(24) = 2.10, p = 0.046).

Further, we explored differences in language, reading and cognitive abilities. We found no differences in children’s performance between the subgroups. Next, we explored children’s balance of dual-language abilities, in which a score for balance was generated by subtracting scores in Spanish and English separately for vocabulary and morpho-syntax measures. A balance score of or closer to ‘0’ suggests similar proficiency in both languages, and a score farther away from ‘0’ suggests some unbalanced knowledge in one language over the other, such that a positive score indicates greater proficiency in English and a negative score indicates greater proficiency in Spanish. This analysis revealed that bilinguals in both subgroups showed greater proficiency in English vocabulary; however, bilinguals in Subgroup 2 were closer to having balanced vocabulary abilities in both languages (M = 11.16, SD = 12.60) than bilinguals in Subgroup 1 (M = 21.57, SD = 12.19), t(24) = 2.14, p = 0.043). The morpho-syntax balance scores did not reveal significant differences between the Subgroups.

Finally, we examined the parent questionnaires for differences that may be reflected in children’s language environments, experience and usage. These analyses revealed that the subgroups did not differ in their age of acquisition (Subgroup 1: M = 1.36 years old, SD = 1.82; Subgroup 2: M = 2.00 years old, SD = 2.00). However, the subgroups differed in the parents’ own bilingualism: children in Subgroup 1 were more likely to have parents who identified as bilingual, and children in Subgroup 2 were more likely to have parents who identified as monolingual Spanish speakers; Mann-Whitney U(14, 12) = 31, z = −2.97, p = 0.003. These analyses also revealed that the subgroups differed in parents’ report of how their child was perceived by others (bilingual, Spanish dominant, or English dominant). Although relatively equal proportions of parents from both subgroups reported that their child was perceived as bilingual (Subgroup 1: 65%, 9 out of 14; Subgroup 2: 66%, 9 out of 12), differences emerged in the proportions of parents who reported how their child was perceived in regards to dominance: 35% (5 out of 14) of the children in Subgroup 1 were perceived as English dominant, whereas 33% (3 out of 12) of the children in Subgroup 2 were perceived as Spanish dominant; Mann-Whitney U(14, 12) = 47.50, Z = −2.23, p = 0.026. In sum, three key characteristics can describe differences across bilinguals in Subgroup 1 and 2: bilinguals in Subgroup 2 (i) had better accuracy in Congruent trials, (ii) were more balanced in their dual-language vocabulary abilities, and (iii) were perceived as bilinguals/Spanish dominant by their parents. Instead, bilinguals in Subgroup 1 were more English-dominant, and perceived as bilinguals/English dominant by their parents.

Finally, given the socio-economic differences between bilinguals and monolinguals, we examined whether the subgroups differed on socio-economic status (parent educational levels and family income). Independent samples t-tests revealed that bilinguals in Subgroup 1 were of higher socio-economic background (education: M = 6.61, SD = 1.73; income: M = 7.38, SD = 2.06) than bilinguals in Subgroup 2 income: (education: M = 4.38, SD = 2.86;income: M = 5.00; SD = 2.67); t(24) = 2.45, p = 0.02, t(22) = 2.52, p = 0.02, respectively).

4. Discussion

Cognitive control mechanisms support language processing and selection in bilinguals (Kroll & Bialystok, 2013). Though highly debated (e.g., Paap, 2019), bilingualism possibly enhances attentional control (Arredondo et al., 2017; Bialystok, 2017; DeLuca et al., 2020). Contradictory neuroimaging evidence has emerged in studies of bilingual children’s brain function for cognitive control (e.g., Li et al., 2020 vs. Arredondo et al., 2017). However, the discrepancy in prior results may be due to the heterogeneity in the bilinguals’s context (e.g., immersion vs. heritage speakers). In the present study, we focused on school-age children (ages 7–9) and investigated whether patterns of brain activity in bilinguals and monolinguals differed in standard between-group analyses (which assume homogeneity) as well as in person-specific connectivity analyses (which assume heterogeneity). Similar to prior bilingual studies in this age range (Antón et al., 2014), we did not find a bilingual advantage in ANT performance. We also hypothesized that, in spite of task performance similarity, bilingual experience will influence the emerging functional organization underlying attention networks. Consistent with this hypothesis and in accordance with prior evidence (Arredondo et al., 2017; Mohades et al., 2014; DeLuca et al., 2020), the results revealed differences in brain activity and connectivity among monolinguals and bilinguals.

First, both monolingual and bilingual children activated multiple channels in the left frontal region for the Alerting network (Center > No Cue): monolinguals also showed activity in right frontal and right posterior near superior parietal regions, and bilinguals in the left temporo-parietal region. Second, neither group showed significant activity for the Orienting network (Spatial > Center). In prior work, monolingual adults engaged a bilateral fronto-parietal network, with right frontal and left parietal activity being the most representative across the literature, during both Alerting and Orienting (Corbetta et al., 2000; Corbetta & Shulman, 2002; Fan et al., 2005; Konrad et al., 2005; Nee, Wager, & Jonides, 2007; Petersen & Posner, 2012). The present results on the Orienting network are similar to prior neuroimaging findings with monolingual children (ages 8–12; Konrad et al., 2005), which did not reveal greater activity during Spatial cue trials, except in Occipital regions which were not directly measured in the present study. The present study’s analyses, however, revealed that Center cues, which are associated with the Alerting network, involved greater brain activity in fronto-parietal regions for children. One possibility is that Spatial cues supported participants’ performance by providing information on the upcoming location of the Congruence stimuli (hence, their performance was faster and more accurate during these trials), while Center cues recruited additional cognitive resources to compensate for not having information about the location of the upcoming stimulus. Prior work with infants suggests the Orienting attentional system is in place by 4-to-6 months of age (Arredondo et al, 2021; Richards, 2000, 2001, 2005); therefore, it is not the case that children in the present study have not yet developed an Orienting attentional system. Instead, children are likely relying on brain regions that are either not covered by the fNIRS probe in the present study, or on subcortical regions that can not be measured using fNIRS.

The Executive attentional network was of special interest, given prior findings showing a bilingual cognitive advantage (Costa et al., 2008; Yang et al., 2011; Kapa & Colombo, 2013; Tran et al., 2015). Using the standard averaging approach, we found that bilingual children showed brain activity across left frontal channels and left posterior temporo-parietal regions for the Executive attentional network. In contrast, monolinguals’ brain activity for this network did not pass FDR thresholds. Similar to bilinguals’ standard averaging approach, the GIMME group-level result revealed common frontal and parietal connections for both monolingual and bilingual children.

We then turned to data-driven person-specific functional connectivity mapping which detected two subgroups of children in the data, and generally mapped them onto bilingual and monolingual language groups. One empirically-derived subgroup included almost all monolinguals (Subgroup 1), and the other subgroup included mostly bilinguals (Subgroup 2). Specific to brain activity, we originally hypothesized greater engagement within the brain’s left hemisphere “language” regions for bilinguals, and in GIMME connectivity maps, differences between the subgroups can be explained by differences in bilingual experience. Person-specific network models revealed data-driven subgroup differences, in which bilinguals who were closer to proportionate abilities in both of their languages showed different patterns of brain activity from bilingual children who tended to have a greater disproportionate balance in their two languages. Bilinguals in the two subgroups differed in neural network density (i.e., number of connections which tend to reflect greater integration and communication among regions). More specifically, bilinguals who showed robust engagement of the left frontal region by having a greater number of connections within this region (Subgroup 2), showed better Congruent accuracy, were perceived as bilinguals with a greater dominance in their first heritage language by their parents, and importantly, their English language abilities did not differ from bilinguals in Subgroup 1. In contrast, bilinguals with a more “monolingual-like” network (Subgroup 1) showed a left fronto-parietal network and unbalanced bilingual abilities. These expoloratory results point to changes in attentional networks that may largely depend on bilinguals’ balance of vocabulary knowledge in both languages, which may consequently reflect parents’ perceptions of their children as bilingual speakers (e.g., Spanish-dominant vs. English-dominant). These results are novel and suggest that monolingual-bilingual neural signatures emerge in mid-childhood – and they can be detected by an algorithm blind to language experience that parses individuals into groups that differ in the extent of their bilingual experience.

Bilinguals in the present study are growing up in a context in which English is the majority language, and they lack access to large bilingual communities or dual-language schools. Thus for bilingual children growing up in this region, maintaining a positive developmental trajectory for both languages may be more effortful, since they are at a higher risk for attrition and loss of their heritage language and becoming “monolingual-like”. Previous neuroimaging work with adults shows that stronger activity in frontal and parietal regions is consistent with mature attentional networks (Abundis-Gutíerrez et al., 2014; Konrad et al., 2005). In the adult brain, however, differences across hemispheres (whether left or right) vary by task difficulty levels. For instance, left frontal regions support better performance during increasingly difficult conditions for attention tasks (Swick, Ashley, & Turken, 2008). Developmentally, children appear to have a left lateralized or a more bilateral brain response, than adults who often show robust responses in right inferior frontal gyrus (Bunge et al., 2002; Moriguchi & Hiraki, 2013; Abundis-Gutíerrez et al., 2014; Konrad et al., 2005). These hemispheric differences are typical of early cognitive development across a variety of Executive Function tasks, suggesting that earlier in development, children’s left hemisphere is more efficient at extracting, re-evaluating rules and flexibly applying them (Bunge & Zelazo, 2006; Moriguchi & Hiraki, 2013, 2014; Zelazo, Carlson, & Kesek, 2008). Given that bilinguals show stronger brain activity within the left hemisphere, one possibility is that bilinguals’ increasing abilities in extracting the linguistic rules in both languages supports a more mature specialization of this region for a more efficient engagement in the non-linguistic domain. Indeed, bilinguals (Subgroup 2), who showed robust engagement of the left frontal region by having a greater number of connections within this region, also performed better during Congruent trials than bilinguals in Subgroup 1. This is commensurate with prior work showing that balanced bilingual children exhibit greater brain activity than unbalanced bilinguals during language processing tasks (Archila-Suerte, Munson, & Hernandez, 2016; Archila-Suerte, Woods, Chiarello, & Hernandez, 2018).

4.1. Limitations and future directions

The present research has several limitations. Brain activity measurements were restricted to the frontal, temporal and parietal lobes, which likely limited the imaging of the Orienting attentional network. Future work in visual orienting of attention should include measurements of posterior regions that include the occipital lobe. In addition, variability in head sizes across child participants ranged between 52 and 56 cm and possibly inserted a degree of variability in the fNIRS measurements. Recent work, however, suggests that there is some consistency in fNIRS measurements covering the frontal brain regions in children of a similar age as those in the present study and in adults who vary in head sizes (Hu et al., 2020). However, Hu et al. (2020) also showed that there is greater variability in fNIRS measurements covering posterior brain regions (Hu et al., 2020). Thus, one possibility is that individual variability in head sizes affected the results in posterior channels, and to a lesser degree in frontal channels. Nevertheless, the validity and reliability of the present findings, especially those in the frontal lobe, are supported by converging prior work showing that bilingualism alters the functional organization of attentional networks (Arredondo et al., 2017, 2021).

The GIMME analysis draws power from the entire timecourse for each participant (which included 10-minutes of data collection and over 1700 observations), yet the present study is still limited by its small sample size included in the GIMME subgroups. The present study included monolingual participants from households of higher income and educational levels, and the GIMME subgrouping revealed that bilinguals in Subgroup 1 (the subgroup including most monolinguals) were also of higher income and educational levels than bilinguals in Subgroup 2. Socio-economic environments can impact brain development and cognition (Brito & Noble, 2014; Leonard et al., 2019; Noble et al, 2012), and they are a confound in the present work. Recent work suggests that young children, who engage in turn-taking during conversations (an ability that requires cognitive control mechanisms similar to bilingualism), show greater activity from left frontal regions, and these are independent of socio-economic background and general cognition (Romeo et al., 2018). Furthermore, this effect was replicated following an intervention enhancing parent–child conversations and language environments (Romeo et al., 2021), suggesting that the extent of language experiences impact children’s neuroplasticity in left frontal regions. Due to the small sample, the effects of socio-economic background and bilingual experiences cannot be distinguished from the results. Given that bilinguals (Subgroups 1 and 2) and monolingual children performed similarly in the ANT task, as well as in a variety of language and cognitive assessments, it is possible that differences in socio-economic backgrounds may not have affected their behavior performance, and thus, functional connectivity differences may be due to their dual-language experiences. Nevertheless, future research with larger samples – and methods that characterize child-specific variabioity in brain functional connectivity – is needed to advance the understanding of these results.

The connectivity analyses, however, should also be interpreted in context. First, their links to behavior were exploratory and did not include Bonferroni corrections. Second, and to the best of our knowledge, this is the first fNIRS study that applied GIMME, and so, some reasonable assumptions were made during analyses, but they require explicit examination in future work. For instance, the temporal resolution in fNIRS data is greater than in fMRI data. We, therefore, down-sampled the data and used a 90% group-level criterion to ensure more homogeneity in cross-ROI connections within the resulting networks, and to provide a better comparison with the field’s standard (mean-based) contrasts that we also provide. However, future work using GIMME should consider a less conservative threshold (e.g., 75%) similar to that used in fMRI research, which would detect greater heterogeneity. Nevertheless, the present work serves as a vital methodological demonstration of the utility of GIMME and person-specific network methods for the advancement of fNRIS data analysis, and subsequent conceptual insights.

Lastley, another caveat is that bilinguals completed the Spanish assessments at the end of the 2-hour session and it is possible that some children performed more poorly on these language assessments due to fatigue. Future work should include a greater number of language assessments for both languages and randomize the language presentation to participants.

5. Conclusions

A Neuroemergentist perspective suggests cascading neurodevelopmental effects of early bilingual experience on children’s emerging brain organization and function (Hernandez et al., 2019). Consistent with this idea, prior work has shown that children who use two languages from early life show advanced neural specificity for language function with stronger left and reduced right frontal activation during syntactic tasks (Arredondo et al., 2019b) and increased left hemisphere engagement during attention-demanding lexical selection tasks (Arredondo et al., 2019a). The present study used an innovative person-specific functional connectivity method to elucidate how the heterogeneity of bilingual experiences guide differences in brain function and connectivity within bilinguals, as well as between bilinguals and monolinguals. This new data-driven approach revealed that it can detect monolingual and bilingual brain patterns, and the extent of the bilingual impact may depend on the heterogeneity of the bilingual experience (i.e., balanced vs. unbalanced dual-language abilities). Though the study is limited by a small sample size and the inclusion of one attentional control task, the present results are broadly consistent with recent fMRI findings (Kwon et al., 2021) from the large-scale ABCD Study. These results extend findings by showing how variation in language and bilingual experience can also influence the neurodevelopmental trends for executive function during non-verbal attention tasks. The present findings fill significant knowledge gaps in models of bilingual neuro-cognitive development, highlighting that heterogeneity in bilingual experience matters for brain function in children (Berken et al., 2016; DeLuca et al., 2020; Sulpizio et al., 2020). Taken together, these findings suggest that despite a lack of a bilingual advantage in task performance, bilingual experience impacts the dynamics of brain activation and connectivity for attentional control within the left fronto-parietal network in childhood.

Appendix A

See Table A1.

Table A1.

Deoxygenated hemoglobin results for the attentional networks in child bilinguals and monolinguals.

Hemisphere	Channel	Region	Beta (SE)	T
Alerting network (Center > No Cue)
Monolinguals
R	16	Occipitotemporal	−21.37 (6.32)	−3.38**
L	7	Middle Frontal Gyrus	−26.30 (7.65)	−3.44**
L	10	Middle/Inferior Frontal Gyrus	−15.51 (4.62)	−3.36**
L	12	Temporal	−19.30 (4.84)	−3.98**
L	14	Temporal, Occipitotemporal	−19.04 (5.61)	−3.39**
L	16	Occipitotemporal	−16.60 (3.60)	−4.61***
Bilinguals
R	12	Temporal	−18.89 (5.67)	−3.33**
L	6	Middle Frontal Gyrus	−24.63 (3.94)	−6.25***
L	7	Middle Frontal Gyrus	−32.44 (7.47)	−4.34***
L	9	Middle/Superior Frontal Gyrus	−23.85 (5.88)	−4.06***
L	10	Middle/Inferior Frontal Gyrus	−21.37 (6.19)	−3.45**
L	11	Middle/Superior Frontal Gyrus/Precentral	−25.55 (7.35)	−3.48**
Orienting network (Spatial > Center Cue)
Monolinguals
R	7	Middle Frontal Gyrus	17.03	2.82*
R	8	Middle/Superior Frontal Gyrus	13.82	3.04*
R	10	Middle/Inferior Frontal Gyrus	9.91	2.92*
L	7	Middle Frontal Gyrus	22.48	2.80*
L	8	Middle/Superior Frontal Gyrus	15.42	3.18*
L	9	Middle/Superior Frontal Gyrus	5.33	2.54*
L	10	Middle/Inferior Frontal Gyrus	4.83	2.75*
L	11	Middle/Superior Frontal Gyrus/Precentral	16.64	3.16*
L	16	Fusiform Gyrus	10.28	2.70*
Bilinguals
R	6	Middle Frontal Gyrus	11.16	2.77*
R	7	Middle Frontal Gyrus	16.42	3.54**
R	9	Middle/Superior Frontal Gyrus	16.23	2.83*
R	11	Middle/Superior Frontal Gyrus/Precentral	12.38	2.65*
L	6	Middle Frontal Gyrus	9.49	2.42*
L	9	Middle/Superior Frontal Gyrus	17.71	2.92*
L	20	Inferior Parietal Lobe	32.06	2.47*
Executive network (Incongruent > Congruent)
Monolinguals
R	18	Inferior parietal, supramarginal gyrus	9.12 (3.26)	2.79*
Bilinguals
R	2	Inferior Frontal Gyrus	−10.37 (3.62)	−2.86*
R	3	Inferior Frontal Gyrus	−10.19 (3.47)	−2.93*

Notes. False discovery rate (FDR) thresholds correcting for multiple comparisons at p < 0.05 are reported.

^*p < 0.05

^**p < 0.01

^***p < 0.001.