Bilingual effects on lexical selection: A neurodevelopmental perspective

Abstract

When a listener hears a word, multiple lexical items may come to mind; for instance, /kæn/ may activate concepts with similar phonological onsets such as candy and candle. Acquisition of two lexicons may increase such linguistic competition. Using functional Near-Infrared Spectroscopy neuroimaging, we investigate whether bilingualism impacts word processing in the child’s brain. Bilingual and monolingual children (N=52; ages 7–10) completed a lexical selection task in English, where participants adjudicated phonological competitors (e.g., car/cat vs. car/pen). Children were less accurate and responded more slowly during competing than non-competing items. In doing so, children engaged top-down fronto-parietal regions associated with cognitive control. In comparison to bilinguals, monolinguals showed greater activity in left frontal regions, a difference possibly due to bilinguals’ adaptation for dual-lexicons. These differences provide insight to theories aiming to explain the role of experience on children’s emerging neural networks for lexical selection and language processing.

1. Introduction

When listening to speech, multiple words and meanings may come to mind, so that /kændi/ may trigger “cane” (a syntagmatic associative concept), “sugar” (a semantically related concept), or /kændəl/ (a phonologically associated concept) (Postman & Keppel, 1970). An individual’s mental lexicon is essential for spoken word comprehension (Jackendoff, 2002): it contains information about a word’s meaning, phonological information, and likelihood of invoking other words (MacDonald, 1997; Markman & Hutchinson, 1984; Tanenhaus, Spivey-Knowlton, Eberhard, & Sedivy, 1995). The initial phonemes of a spoken word may particularly trigger other words that begin with similar sounds, and these lexical items will compete for selection in the mind (Marslen-Wilson & Tyler, 1980). Knowing multiple languages can double such competition, so that an English-Spanish bilingual could also trigger concepts like [kara’melo] “caramelo” (sweets or candy in English) or [‘keso] “queso” (cheese in English) (Blumenfeld & Marian, 2013; Jared & Kroll, 2001; Marian & Spivey, 2003). All speakers develop mechanisms to help them select the target concept and inhibit the competitors during language processing and production (Righi, Blumstein, Mertus, & Worden, 2010; Rigler et al., 2015). However, models of dual-language processing suggest that the potential increase in lexical competitors may influence how the bilingual mind and brain resolves lexical selection (Hernandez, Ronderos, & Claussenius-Kalman, 2018; Shook & Marian, 2013; Kroll, Dussias, Bice, & Perrotti, 2015). Childhood is a critical time of brain development; thus, it is an ideal period for researchers to study both the development of these mechanisms and the experiential influences on the process.

1.1. Theoretical Models of Bilingualism and Word Selection

There are three models typically used to explain lexical access in bilingual speakers. First, the Bilingual Interactive Activation plus Model (BIA+; Dijkstra & van Heuven, 2002) offers a comprehensive account for written word identification. The model consists of a set of perceptual/bottom-up mechanisms that cue the listener to a given language via language-specific properties (phonological, prosodic, and orthographic), and attentional/top-down mechanisms support the selection of the appropriate linguistic feature. Second, the Inhibitory Control model (IC; Green, 1998) suggests a heavy reliance on top-down mechanisms to adjudicate linguistic competitors and to inhibit representations from the other language. These models are not mutually exclusive; rather, they vary in the valiance allocated towards the top-down controls that select the linguistic units.

Finally, the Bilingual Language Interaction Network for Comprehension of Speech (BLINCS, Shook & Marian, 2013) extends both of these perspectives to spoken word recognition. BLINCS posits that a shared phono-lexical network emerges where both languages interact, specifically when lexical items co-occur but also when items share a phonemic overlap across both languages. For instance, in English and Spanish bilinguals, “pear” /pɛr/ may activate “pan” /pæn/ to the same degree as the Spanish word for dog “perro” [ˈpɛ.ro] via their shared phonemic onset. In contrast, lexical items that rarely co-occur, are semantically unrelated, and do not share an acoustic phonetic form or phonological representation (e.g., juggler and dress) are less likely to overlap within the shared network (see also Apfelbaum, Blumstein, & McMurray, 2011). These self-organizing networks, however, largely depend on the individual’s language dominance, dual-language proficiency and their maturational age (Blumenfeld & Marian, 2007).

1.2. Brain Systems Supporting Dual-Language Lexical Selection

In support of the BLINCS framework, eyetracking studies show that bilinguals’ languages are relatively co-active and top-down mechanisms support the resolution of inter-language conflict (Blumenfeld & Marian, 2007, 2011, 2013; cf. Kroll et al., 2015). For instance, Spanish-English bilinguals examine a Target image matching a spoken word but also examine competitors (more than non-competing fillers) when presented with images in which another item’s direct translation in the other language matches the phonological onset of the Target, such as “pool” as the Target and “thumb” as the competitor which translates to “pulgar” [pul.ˈɣaɾ] in Spanish (Blumenfeld & Marian, 2013; Marian & Spivey, 2003). Neuroimaging research using the same word-selection paradigm shows bilinguals engage brain regions in left frontal cortex and anterior cingulate (regions associated with language processing and attention/error monitoring respectively), when presented with cross-language competition (Marian, Bartolotti, Rochanavibhata, Bradley, & Hernandez, 2017). During single-language competition (e.g., candy vs. candle), however, bilinguals’ brain activity in left frontal cortex and anterior cingulate regions is reduced; while monolinguals exhibit greater brain activity in these regions (Marian Chabal, Bartolotti, Bradley, & Hernandez, 2014). Bilinguals’ brain activity when managing single- and dual-language competition is also associated with their top-down Executive Function performance, but not for monolinguals (Marian et al., 2014, 2017). Of importance is to note that this work is correlational, yet suggestive of a relationship between the amount of experiences in managing linguistic competition and the result of altered top-down Executive Function mechanisms.

While infants engage left superior temporal regions for bottom-up lexical retrieval processes (Friedrich & Friederici, 2010; Travis et al., 2011), the top-down mechanisms for selecting and categorizing lexico-semantic concepts continue to mature until age 10 (Brauer & Friederici, 2007; Skeide et al., 2014). This is likely due to the protracted myelination of the inferior fronto-occipital fasciculus (IFOF) white matter tracts that connect parietal and temporal regions to inferior frontal gyrus (IFG, Brodmann area 44; Dubois et al., 2015; Mohades et al., 2015; Skeide et al., 2014). Between the ages of 7–10, children’s neural organization for language becomes increasingly adult-like within the left anterior frontal and perisylvian temporal network, with differentiated systems for phonological, semantic, and syntactic processes (cf. Skeide et al., 2014; Arredondo, Ip, Hsu, Tardif, & Kovelman, 2015; Arredondo, Hu, Seifert, Satterfield, & Kovelman, 2018; Ugolini, Wagley, Ip, Hsu, Arredondo, & Kovelman, 2016). At the same time, this period is also marked by developmental changes in the fronto-parietal system supporting gains in attention, verbal working memory, and other executive function mechanisms (Arredondo, Hu, Satterfield, & Kovelman, 2017; Davidson, Amso, Anderson, & Diamond, 2006; Ofen, Chai, Schuil, Whitfield-Gabrieli, & Gabrieli, 2012). Recent work suggests bilingual children in this age range are re-organizing brain networks for attention differently than monolinguals, by showing greater reliance on the left frontal regions for non-linguistic attentional processes, while monolinguals show greater reliance on right frontal regions (Arredondo et al., 2017). Yet, it remains largely unknown how these higher-level changes in attention are beginning to emerge over time in relation to dual-language processing. In the present work we ask, what role does bilingual experience play on the formation of attention-demanding word identification processes during mid-childhood (ages 7–10), a key period of language and cognitive development where these mechanisms are beginning to become established.

The newly proposed Neurocomputational Emergentism (or Neuroemergentism) framework challenges researchers to see the effects of bilingualism as ‘greater than the sum of its parts’ (Hernandez, Claussenius-Kalman, Ronderos, & Vaughn, 2018; Hernandez et al., 2019). In this framework, low-level changes can give rise to a mass of gradual, new capacities and otherwise qualitative changes in cognition and the brain (Hernandez et al., 2018^a, 2019; Hernandez, Ronderos, & Claussenius-Kalman, 2018^b). In the case of childhood bilingualism, Neuroemergentism suggests that while young bilinguals can follow monolingual-like acquisition and typical trajectories for each of their languages (Petitto & Kovelman, 2003), the two linguistic systems interact and compete for shared resources for language and attention processing (Hernandez et al., 2018^{a, b}, 2019). Early in development, the competition might be predominantly taking place at the basic phonological level and then expanding to include competition at the semantic and syntactic levels as language acquisition and brain development progress (Krizman & Marian, 2015). During childhood, experience with bilingualism may, therefore, incur changes in basal ganglia, auditory cortex, and posterior temporal/parietal regions supporting these initial phonological processes, and manifest in greater recruitment of these regions (Hernandez et al., 2018^{a, b}, 2019). Mid-childhood (ages 7–10) is marked by some of the final developmental changes towards a more adult-like language system, as well as active maturation of attentional/Executive Function mechanisms contributing to top-down processes in language (Arredondo et al., 2015, 2017, 2018; Davidson et al., 2006). At this developmental intersection, the Neuroemergentism framework guides us to hypothesize that early bilingualism may exert its influence by mid-childhood on the top-down word selection processes. Further, understanding whether these changes are already in place for developing cortical systems will offer a foundation for the emergence of biologically-based changes that come into effect and are altered across lifelong bilingualism.

1.3. Present Study: Research Question and Hypothesis

In the current study, we ask what is the impact of bilingual experience on children’s neuro-cognitive organization for word processing? To investigate the question of emergent linguistic systems in young bilinguals, we use functional near-infrared spectroscopy (fNIRS) neuroimaging. fNIRS is silent and motion-tolerant, making it a neuroimaging technique optimally suited for auditory experiments with children (Quaresima, Bisconti, & Ferrari, 2012); akin to functional magnetic resonance imaging (fMRI), fNIRS also detects changes in hemodynamic response, yet it provides better temporal sampling.

The Neuroemergentism framework considers developmental changes in the bilingual mind and brain, cascading from changes driven by the individual’s maturation and interaction with the linguistic environment (Hernandez et al., 2018^{a, b}, 2019). Within this framework, we hypothesize that during the developmental period when the brain’s neural mechanism of lexico-semantic processing undergoes the near-final stages of neural specialization (Skeide & Friederici, 2016), while at the same time Executive Functions and the frontal lobe regions that support those functions are in the midst of active development, bilingualism will exert influence on children’s top-down mechanisms of language processing. To test this hypothesis, Spanish-English bilingual children and English monolinguals completed an attention-demanding word-selection task in English. We offer the following predictions relative to age- and English-proficient matched monolingual children: First, bilingual children will show reduced activation in left frontal regions that support attentional/top-down mechanisms for lexical selection. Second, this reduction in frontal lobe activation might be paired with increased activation in left temporo-parietal regions reflecting a more automated word processing system in bilinguals. Both predictions are based on the Neuroemergentism perspective suggesting that early-life experiences with increased levels of linguistic competition, may increase the efficiency of neural resources in bilinguals for automated word selection processes and thus incur reduced reliance of neural resources on anterior top-down mechanisms for lexical selection (Hernandez et al., 2017, 2018^{a, b}). These findings will inform theories of bilingualism, as well as language and brain development across all learners, by offering evidence on how early-life bilingualism may influence attention-demanding language processes.

2. Materials and Methods

2.1. Participants

Fifty-two children participated (24 females, 28 males; age range = 7.1–9.9 years): 26 Spanish-English bilinguals and 26 gender- and age-matched English monolinguals. From these children, 5 bilinguals and 4 monolinguals yielded noisy neuroimaging data (see fNIRS section below for details); thus, the final set for data analysis includes 21 Spanish-English bilinguals (9 females, 12 males; M_age = 8.09) and 22 English monolinguals (11 females, 11 males; M_age = 8.12). All children were right-handed, neurotypical, and raised and educated in southeast Michigan (United States). English was the language of school instruction for all children.

Selection criteria for bilingual participants were as follows: Spanish exposure from birth, English exposure prior to age 5, which is the period demonstrating that children can be classified as ‘bilingual L1s’ where they are simultaneously acquiring two first languages (Genesee & Nicoladis, 2006); current daily exposure to both languages (Spanish in the home and English outside the home); 3 years minimum of English exposure prior to testing; and dual-language competence including a standard score above 85 in English and Spanish receptive vocabulary abilities. At the time of testing, 11 bilingual children were attending a Spanish-heritage language-learning Saturday school for 2–3 hours per week, in addition to daily Spanish literacy homework. Three bilingual children were born in a Spanish-speaking country; the remaining bilingual children were born in the United States. Aside from one mother and three fathers, the bilingual children’s parents were native Spanish speakers. All bilingual families reported consistent use of Spanish at home with their child(ren) by at least one parent, while monolingual families reported English was the only language spoken at home.

The study was approved by institutional review boards; parents and children completed respective informed consent and assent forms. Families received monetary compensation and children received a Frisbee as a thank-you for their participation.

2.2. Behavioral Measures

Parents were asked to complete a demographics and language background questionnaire (see Arredondo, 2017), which provided information regarding their child’s language development, as well as information regarding the family’s educational level and household income.

Given that language competence among school-age children is mainly assessed via vocabulary (lexical) knowledge and morpho-syntactic information, we included measures for these abilities in both languages for bilinguals (vocabulary: Verbal Knowledge subtest from the Kaufman Brief Intelligence Test [KBIT-2; Kaufman & Kaufman, 2004], Receptive One-Word Picture Vocabulary Test Spanish Bilingual Edition [Brownell, 2000]; morpho-syntax: Word Structure subtest from the Clinical Evaluation of Language Fundamentals [Semel, Wiig, & Secord, 2003, 2006]. Since children were attending English-only schools, we additionally included literacy measures to ensure comparable academic language development (reading: Word Identification subtest from the Woodcock Reading Mastery Tests [Revised 2^nd edition; Woodcock, 1998]; phonology: Elision subtest from the Comprehensive Test of Phonological Processing [Wagner, Torgesen, & Rashotte, 1999]). Children also completed measures of general cognition, including: nonverbal intelligence (Matrices subtest from KBIT-2), naming speed (Numbers subtest from the Rapid Automatized Naming [Wolf & Denckla, 2005]), and attentional control (Pair Cancellation subtest from the Woodcock-Johnson III Tests of Cognitive Abilities [Woodcock, McGrew, & Mather, 2001]).

2.3. Neuroimaging Measure: Phonological Linguistic Competition Task

In order to make the study maximally parallel between bilinguals and monolinguals, children completed a task where phonological onsets for English concepts compete; the task is based on prior adult work using eye-tracking and neuroimaging methods (Marian & Spivey, 2003; Marian et al., 2014). The present design offers a principled extension that improves reproducibility for neuroimaging methods with children and allows a more transparent interpretation of the findings. Participants were presented with three conditions (see Figure 1): an experimental condition (Phonologically Related) and two control conditions (Phonologically Unrelated and Baseline).

Figure 1.

Example trials for the Language Competition task. During the trial, participant was presented with two images, a target image and a competitor, and heard a target word. (A) Phonologically Related experimental condition, “bed” and “bell”. (B) Phonologically Unrelated control condition, “ant” and “tie”. (C) Baseline control condition, “ear” along with a scrambled indecipherable image.

Each trial consisted of a display of two stimuli images, a Target and a Competitor, along with a sound clip of a Target word. At trial onset, both images appeared on the center-left and center-right side of the screen. Following 500-ms, participants heard the Target word and had 2500-ms to respond; thus, each trial was 3000-ms in length. Using a two-button pressing box, children pressed left/right buttons to indicate whether the Target word matched the picture on the left or right.

In the Phonologically Related (PhR) condition, the onset for the names of the Target and Competitor overlapped in acoustic phonetic form and phonological representation (e.g., bed /bɛd/ and bell /bɛl/). In the Phonologically Unrelated condition, the onset for the names of the Target and Competitor did not overlap in acoustic phonetic form and phonological representation (e.g., ant /ænt/ and tie /taɪ/). In the Baseline condition, the Target image is displayed alongside a scrambled image that acted as the Competitor. The scrambled image was designed from an image whose phonological onset for the Target did not overlap (e.g., ear /ɪr/ and nine /naɪn/). The goal of the Baseline condition was to show children’s performance when there was little verbal competition. Although the Baseline condition may be simpler to respond, it might still require some visual cognitive load (see Cano, Class, & Polich, 2009).

The task was designed as a rapid event-related neuroimaging experiment, with 21 trials per condition and a total of 63 trials. A 6-second rest period was introduced at the beginning and end of the task. Additionally, inter-stimulus intervals (ISI; jittered rest periods) were randomized across the task and their total length corresponded to the total length of a condition (i.e., 63-seconds); these intervals were a minimum of 1 second and a maximum of 6 seconds. Thus, the task was comprised of 25% PhR trials, 25% PhU trials, 25% Baseline trials, and 25% ISI. We used OptSeq2 (Dale, 1999) to randomize trials and ISIs, which is a widely used program in fMRI specialized in the order and timing of rapid-presentation events. The total length of the task was about 4-1/2 minutes. 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; sound played via two Creative Inspire T12 2.0 multimedia speakers. A two-button box (Current Designs, Inc.) was connected to the desktop computer to record participants’ responses. Task performance was assessed by accuracy and response time. Trials were deemed correct when the participant pressed the button corresponding to the side (left/right) matching the Target word and image. Response times were collected starting trial onset, that is at the onset of the images.

2.3.1. Stimuli.

All stimuli were high-frequency words, chosen from the MacArthur-Bates Communicative Development Inventory (MB-CDI; Fenson et al., 1993) Words and Sentences English version, which is a checklist of vocabulary words appropriate for 16–36 month-olds. First, we created PhR trials by grouping MB-CDI words by phonological onset. Then created trials by pairing those with a similar number of phonemes and syllables. From these pairs, we chose the Target as the lower frequency from the two. Next, we designed PhU and Baseline trials by selecting items that matched the PhR Target and Competitors in frequency, the number of phonemes and syllables, phonological neighborhood size, concreteness, and imageability. A total of 37 trials were designed for each condition. Then, we obtained black and white line drawings for each item from Microsoft Office Clip Art or Google Images. We piloted our stimuli at the Ann Arbor Hands-On Museum in partnership with the University of Michigan’s Living Lab program. Each image was first pre-tested by at least ten children (ages 6–9 years) to show high naming consistency to the stimuli words (at least 80% accuracy for the present stimuli; M= 97.05%, SD= 6.03). Next, a female speaker native to the region of testing recorded all stimuli words. Finally, we put together the trials (images and sound clips) and piloted a behavioral version of this task at the Ann Arbor Hands-On Museum with 21 English-monolingual children (11 girls; M = 8.25 years, SD = 1.22), using a similar experimental design where trials were randomized, and children used a two-button pressing box to respond. Due to time constraints, children only completed PhR and PhU trials. For the final PhR and PhU stimuli, children were in average 97% accurate (SD = 2.82), and they took longer to respond to PhR than PhU trials, t(20) = 4.58, p < 0.001.

The final set of stimuli showed high naming consistency (see above), high word-to-picture matching accuracy (>90%; PhR and PhU trials only), matched across in the number of phonemes and syllables (Target words: phonemes M = 3.30, SD = 0.80; syllables M = 1.11, SD = 0.32; Competitors phonemes: M = 2.98, SD = 0.82; syllables: M = 1.10, SD = 0.81), and did not share a phonological onset overlap to a Spanish translation that would create additional competition between the Target and Competitor words. Following these requisites, the final PhR set included 21 trials. Thus, we selected PhU trials with items that closely matched the frequency of both PhR Target and Competitors on an item-specific basis (Target: PhR M = 1,499.05, SD = 2,202.05 and PhU M = 1,149.24, SD = 1,180.78; Competitors: PhR M = 4,942.90, SD = 6,262.93 and PhU M = 7,648.19, SD = 12,806.25; SUBTLEXUS; Brysbaert & New, 2009); these were not statistically different (p > .05). Next, in addition to the earlier requisites mentioned above, we selected Baseline stimuli with frequency that neared the total average of PhR and PhU Targets and Competitors (PhR M = 3,220.97, SD = 4,953.44 and PhU M = 4,398.71, SD = 9,565.39; Baseline Target M = 3,789.19, SD = 6,263.95, Baseline Competitors: M = 3,256.57, SD = 5,537.02). We used the Scramble filter in Adobe Photoshop to scramble Baseline Competitor images into a 4×4 mosaic tile pattern. All stimuli and task may be requested for download here, https://osf.io/frst2/?view_only=2dec759d8566405381525b626b8448a9.

Across all three conditions, there was no statistical difference (ps > .05) in spoken word duration (PhR M = 820ms, SD = 137; PhU M = 805ms, SD = 119; Baseline M = 747ms, SD = 152), total word frequency (SUBTLEXUS; Brysbaert & New, 2009), phonological neighborhood size (CLEARPOND; Marian, Bartolotti, Chabal, & Shook, 2012), concreteness and imageability (MRC Psycholinguistic Database; Coltheart, 1981). See the list of stimuli words in Appendix A.

2.4. Procedure

Parents completed questionnaires while their child took part in the testing session. Children completed English assessments with a native English-speaking experimenter and Spanish assessments with a native Spanish-speaking experimenter. During the fNIRS brain-imaging portion of the study, experimenters set the cap and optodes in place, photographed cap placement, and instructed children to press buttons from a response box as quickly as possible (left or right). Prior to testing, children completed a practice session with nine trials (3 per condition) that were not part of the testing session.

2.5. Neuroimaging Data Acquisition and Analysis

We used a TechEN-CW6 fNIRS system to collect data from 44 channels (i.e., source-detector connection) spaced 2.7 cm apart and sampled at 50-Hz (22 channels per hemisphere; see Figure 2). The cap was applied consistently for each participant 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.

Figure 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-set 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 visualized using Homer2, a MATLAB-based software retrieved from the NITRC database (Huppert, Diamond, Franceschini, & Boas, 2009). Data quality control included separate examination of the 690 and 830nm wavelengths timeseries; since the 830 wavelength allows deeper penetration of light than the 690, we used the 690 wavelength as an indicator for data signal quality. We excluded participants whose 690 signal was above 3 molar units and did not reveal cardiac signal for over 50% of the data channels (4 monolinguals, 5 bilinguals; these artifacts are typically caused by head motion and/or hair obstruction).

The remaining data (22 monolinguals, 21 bilinguals) were analyzed using NIRS Toolbox (Santosa, Zhai, Fishburn, & Huppert, 2018), and customized MATLAB scripts that included Homer2’s motion detection (hmrMotionArtifactByChannel) and Spline (hmrMotionCorrectSpline) functions (Huppert et al., 2009). The following pre-processing steps were completed in the following order: optical density change data conversion, motion artifact detection, motion artifact correction via spline interpolation, and concentration change data conversion yielding oxygenated (HbO) and deoxygenated hemoglobin (HbR) values. First, the raw time course data was converted into units of optical density change. Next, the optical density change 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 one threshold of a standard deviation of 50 within half a second and masked for an additional 1-second. 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 (Brigadoi et al., 2014; Scholkmann et al., 2010). The artifact-corrected optical density data were then converted into hemoglobin concentration data using the modified Beer-Lambert law. Hemoglobin data can be contaminated by physiological noise, especially when sampled at a temporal resolution greater than 10-Hz, leading to serially correlated error terms (Barker et al., 2013). A multiple regression General Linear Model (GLM) approach accounts for statistical parametric mapping assumptions (Friston, Ashburner, Kiebel, Nichols, & Penny, 2006), and is one way of correcting for autocorrelations (Barker et al., 2013; Poline & Brett, 2012). Thus, we applied a GLM approach at the first-level analysis in each participant’s hemoglobin concentration data were via an ordinary least squares (OLS) fit. The GLM OLS assumed the dual-gamma canonical hemodynamic response function peaking at 8-seconds after trial onset (Friston et al., 2006; Hu, Hong, Ge, & Jeong, 2010). The first-level GLM analysis estimated beta values, which are indices of percent signal change, for each condition (Phonologically Related, Phonologically Unrelated, and Baseline).

Next, second-level group analyses were carried out using a restricted/residual maximum likelihood (REML) multivariate linear mixed-effects model (LMM) for each data channel. The model included conditions (Baseline, PhU, PhR) and groups (monolingual, bilingual) as fixed factors, and participants were defined as a random effect variable. For each channel, we carried out two LMMs, one predicting HbO and another for HbR. Although HbO is more commonly reported in the fNIRS literature, the present study presents both HbO and HbR. Prior work suggests HbR is less robust in comparison to HbO, yet exhibits more spatial specificity (Strangman et al., 2002). The present results are compatible with typical hemoglobin responses (see below) so that HbO shows increasing activity and HbR shows decreasing activity (Villringer, Planck, Hock, Schleinkofer, & Dirnagl, 1993). All statistical analyses were evaluated at a False Discovery Rate (FDR) threshold correction of p < .05 (see Benjamini & Hochberg, 1995).

3. Results

3.1. Participant Descriptives

Participants did not differ in age and English language abilities, (p > .05); see Table 1. Bilinguals’ language abilities were greater in English than Spanish, as is typical for bilinguals living in an English-dominant environment. Participants did not differ in cognitive abilities, except in nonverbal intelligence where monolinguals performed better than bilinguals, although all participants had scores within the normal range (above 85). Monolingual children’s families were from households of higher income than bilinguals’, even though parents’ education levels were not significantly different between the groups.

Table 1.

Monolingual and bilingual participants’ average scores (and standard deviation) in language and cognitive task performances.

Measures	Monolinguals’ English (n = 22; 11 girls, 11 boys)	Bilinguals’ English (n = 21; 9 girls, 12 boys)	Bilinguals’ Spanish	T-values between groups’ English	T-values between bilinguals’ languages (English/Spanish)
Age	8.12 (0.80)	8.09 (0.77)	--	0.12	--
IQ†	118.68 (9.21)	108.95 (13.45)	--	2.78**	--
Demographics
Income‡	8.14 (1.77)	6.32 (2.43)	--	2.74**	--
Mother’s education§	6.91 (1.31)	5.76 (2.49)	--	1.91	--
Father’s education§	6.55 (1.77)	5.29 (2.87)	--	1.74	--
Language measures
Phonological Awareness¶	10.86 (2.55)	11.29 (3.08)	--	0.49	--
Vocabulary†	114.36 (10.90)	110.95 (10.68)	92.14 (10.03)	1.04	7.46***
Morpho-syntax (%)	92.53 (6.33)	89.06 (10.10)	89.06 (10.10)	1.36	4.15**
Reading†	119.45 (9.47)	115.52 (10.03)	--	1.32	--
Cognitive tasks
Naming Speed – Numbers†	105.14 (11.34)	108.81 (15.75)	--	0.88	--
Pair Cancellation Score	45.41 (7.68)	47.33 (10.38)	--	0.69	--
Neuroimaging task: Phonological linguistic competition
Accuracy (%)
Baseline (B)	98.92 (2.52)	99.05 (3.40)	--	--	--
Phonologically Unrelated (PhU)	99.57 (1.40)	97.18 (4.02)	--	--	--
Phonologically Related (PhR)	93.70 (6.96)	87.88 (11.88)	--	--	--
Reaction Time (ms)
Baseline (B)	1452.30 (171.14)	1487.68 (109.66)	--	--	--
Phonologically Unrelated (PhU)	1528. 72 (148.85)	1577.49 (148.85)	--	--	--
Phonologically Related (PhR)	1719.08 (128.30)	1743.44 (145.06)	--	--	--

Notes.

^*p < .05

^**p <.01

^***p < .001

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

^‡Options 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.

^§Options 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.].

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

3.2. Linguistic Competition Task – Behavioral performance

To examine task accuracy, we conducted a Generalized Linear Mixed Model (GLMM) analysis, specifically a repeated-measures binomial logistic regression in IBM SPSS Statistics 25. The task conditions (Baseline, PhU, and PhR) were specified as a within-subject repeated factor, language groups (monolinguals, bilinguals) as a fixed factor, and defined participants as a random factor in the structure. The model included all individual trials for each participant where a response was recorded, that is 99.2% of the data (2687 trials; 96.1% correct and 3.9% incorrect trials). The model (χ² = 114.66, df = 5, p < .001) revealed a significant effect of condition (χ² = 63.94, df = 2, p < .001) that stemmed from participants’ better performance in the Baseline (M = 99%, SE=.4) and PhU (M = 99%, SE=.3) conditions than the PhR condition (M = 91%, SE=1.0, Bonferroni ps < .001). Accuracy performance between the Baseline and PhU conditions was not significantly different. The model also revealed a significant effect of group (χ² = 4.98, df = 1, p =.026) that stemmed from monolinguals’ better performance (M = 99%, SE=.4) than bilinguals’ (M = 97%, SE=.6, Bonferroni p = .023), although both groups had ceiling performance (>90%). The model did not reveal a significant group by condition interaction (χ² = 3.65, df = 2, p = .16).

To examine RT, we converted the data into natural logarithm values since descriptive proportion-plots revealed the data was negatively skewed. Using these converted values, we conducted a GLMM, specifically a repeated-measures multinomial regression in IBM SPSS Statistics 25. Since the accuracy GLMM revealed that monolinguals were more accurate than bilinguals, the model included all individual trials for each participant where a correct response was recorded, that is 95.3% of the data (2582 trials). The model (χ² = 276.13, df = 5, p < .001) revealed a significant effect of condition (χ² = 285.27, df = 2, p < .001) that stemmed from participants’ faster performance in the Baseline (M = 1473-ms, SE=10.35) followed by the PhU (M = 1547-ms, SE=10.41) and PhR (M = 1742-ms, SE=10.84) conditions (all Bonferroni ps < .001). The model also revealed a significant effect of group (χ² = 7.27, df = 1, Bonferroni p =.007) that stemmed from monolinguals’ better performance (M = 1569-ms, SE=8.43) than bilinguals’ (M = 1606-ms, SE=8.78). The model did not reveal a significant group by condition interaction (χ² = .12, df = 2, p = .94).

3.3. Neuroimaging Results - Linear Mixed-Effect Models

Using IBM SPSS Statistics 25, we modeled each participant’s first-level GLM results into second-level group REML LMMs as specified above. First, we added children’s age as a covariate to investigate age-related changes in brain activity. Given group differences in IQ, family income, and task performance, we planned to include these variables as covariates in the REML LMMs. An exploratory correlation between IQ and income turned out non-significant, r=.30, p =.62; thus, these variables were entered as separate covariates rather than nested. Task accuracy and RT were entered separately, as well as nested in separate models. We entered covariates as follows: first, all covariates were entered individually, and then each was added across all variations (e.g., IQ only, IQ and income, IQ and RT, income and RT, and so on) until all four covariates were part of the model. We compared the AIC goodness-of-fit indices across all models and found that models of best fit consistently included IQ, income, and RT across channels, and omitted age and task accuracy. Notably, even when age-only was entered as a covariate, this variable did not predict brain activity in any of the channels. Thus, age and task accuracy were removed as covariates from the following analysis. It is possible that the narrow age range of our sample did not provide enough variance for these variables. Likewise, it is possible that response time was a better predictor of the brain data because it better showcased how task conditions varied in difficulty. Next, we present two models for each hemoglobin: one that includes IQ and income as covariates, and a second one for comparison that includes IQ, income, and RT as covariates. See Appendix B for results of models without covariates.

3.3.1. Oxygenated Hemoglobin

First, brain activity results for models including IQ and family income as covariates are presented. From data in 44 channels, 23 of them were significant and reliably modeled, 20 passed an FDR threshold, and 13 had significant effects (11 condition effects and 2 group effects); see Figure 3, Tables 2 and 6. Channels showing a condition effect covered bilateral anterior temporal/inferior fontal regions (Ch 2 in left and right hemisphere), left frontal (Ch 6), as well as bilateral parietal regions (left: Ch 18 and 20; right: Ch 16–19, 21–22); see Figure 3a. Post-hoc analyses using a Bonferroni correction revealed participants had the most robust levels of increasing HbO activity during the PhR condition, in comparison to Baseline (all channels) and PhU conditions (left Ch 20; right: Ch 2, 16, 17, 21, 22), as well as one left channel (Ch 20) showing greater HbO activity during the PhU than Baseline condition; see Table 2. Additionally, IQ was a significant predictor in right hemisphere channels covering anterior temporal/inferior fontal (Ch 2) and parietal lobe (Ch 17, 19, 21). Family income was not a significant predictor of the data in these models beyond IQ. Channels showing a group effect approximately covered bilateral inferior/middle frontal gyri (Ch 10); see Figure 3b. Specifically, these channels showed that monolinguals had the most robust levels of increasing HbO activity than bilinguals, and IQ was a significant predictor in the right hemisphere channel; see Table 6. Post-hoc analysis revealed that monolinguals showed the largest amount of brain activity in the PhR condition than control (PhU and Baseline) conditions (see Table 6).

Figure 3.

Brain images depict significant effects by (A) task conditions for oxygenated and deoxygenated hemoglobin, and by (B) group for oxygenated hemoglobin. Top row effects include IQ and income as covariates. Bottom row effects include IQ, income, and task response time as covariates. (A) Oxygenated hemoglobin effects by task conditions showed the greatest increase of activity for phonologically Related (PhR) trials, followed by Unrelated (PhU), and Baseline; see Tables 2 and 4. Deoxygenated hemoglobin effects by task conditions showed the greatest decrease of activity for Phonologically Related (PhR) trials, followed by Unrelated (PhU), and Baseline; see Tables 3 and 5. (B) Oxygenated hemoglobin effects by group showed the greatest increase of activity for monolinguals as compared to bilinguals; see Table 6. Color bar reflects F-values mapped for comparison of brain activation on approximate regions covered by the fNIRS probeset. IQ=non-verbal intelligence quotient; SES=family income; RT=task response time.

Table 2.

Condition effect results for brain activity models of oxygenated hemoglobin (HbO) during the linguistic competition task, include IQ and family income as covariates.

Channel	AIC	F-value Condition Effect	Wald Z Condition All ps≦.001	Wald Z IQ	PhR coefficient (CI)	PhU coefficient (CI)	Baseline coefficient (CI)	Post-hoc, Bonferroni p-values
Right hemisphere
2	1548.97	6.20**	6.14	3.96***	154.74 (77.30, 232.19)	65.43 (−12.02, 142.88)	34.41 (−43.04, 111.86)	PhR vs. B, .002 PhR vs PhU, .031
16‡	1461.74	3.82*	5.98	n.s.	125.74 (82.10, 169.38)	58.12 (14.48, 101.76)	56.75 (13.11, 100.39)	PhR vs. B, .04 PhR vs PhU, .046
17†	1539.77	7.30**	5.47	3.91***	189.62 (101.71, 277.54)	103.10 (15.19, 191.02)	85.83 (−2.09, 173.74)	PhR vs. B, .001 PhR vs PhU, .007
18	1485.45	5.28**	6.36	n.s.	93.78 (38.97, 148.59)	57.49 (2.68, 112.30)	28.21 (−26.6, 83.02)	PhR vs. B, .01
19‡	1362.71	3.20*	3.42	2.75**	66.41 (34.30, 98.53)	39.99 (7.87, 72.10)	32.02 (−.10, 64.14)	PhR vs. B, .05
21‡	1465.36	6.15**	4.25	2.91**	101.29 (53.04, 149.54)	38.62 (−9.63, 86.87)	21.39 (−26.86, 69.64)	PhR vs. B, .003 PhR vs PhU, .027
22‡	1564.88	10.41***	5.68	n.s.	207.84 (129.51, 286.17)	102.80 (24.48, 181.13)	36.42 (−41.90, 114.75)	PhR vs. B, <.001 PhR vs PhU, .023
Left hemisphere
2‡	1566.45	3.61*	3.13	n.s.	124.20 (53.84, 194.55)	53.24 (−17.12, 123.59)	26.42 (−43.94, 96.77)	PhR vs. B, .037
6†	1471.36	6.67**	5.11	n.s.	122.99 (77.81, 168.17)	68.22 (23.04, 113.40)	24.16 (−21.02, 69.34)	PhR vs. B, .002
18	1498.93	6.62**	6.72	n.s.	148.68 (94.90, 202.45)	92.16 (38.39, 145.93)	58.63 (4.86, 112.41)	PhR vs. B, .003
20	1625.17	13.95***	5.94	n.s.	242.38 (122.76, 362.00)	129.57 (9.96, 249.19)	21.51 (−98.11, 141.13)	PhR vs. B, <.001 PhR vs PhU, .027 PhU vs. B, .009

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for condition effects are presented. When significant, Wald Z estimates are presented for covariate effects, Wald Z condition effects are presented for comparison. Family income was not a significant predictor of brain activity differences above and beyond IQ.

AIC = Akaike’s Information Criterion; n.s. = not significant; CI = 95% confidence interval (lower bound, upper bound)

Task conditions - PhR= phonologically related condition, PhU= phonologically unrelated condition, B= baseline condition.

^*p < .05

^**p <.01

^***p < .001

^†Indicates a channel that is significant in both HbO and HbR comparable models, see Table 3.

^‡Indicates a channel that is also significant in the HbO model when task response time is added as a covariate, see Table 4.

Table 6.

Group effect (M=monolinguals, B=bilinguals) results for brain activity models of oxygenated hemoglobin (HbO) during the linguistic competition task, include IQ, family income and task response time (RT) as covariates.

					Coefficients by group		Coefficients for each condition showing group effects: M = Monolinguals’ / B = Bilinguals’
Channel	AIC	F-value Group Effect	Wald Z Condition All ps≦.001	Wald Z Covariates (IQ, income, RT)	Monolinguals coefficient (CI)	Bilinguals coefficient (CI)	PhR coefficient (CI)	PhU coefficient (CI)	Baseline coefficient (CI)
Model with IQ and income as covariates
Right hemisphere
10	1381.57	10.20**	4.55	IQ = 2.73**	98.28 (60.16, 136.41)	12.53 (−24.55, 49.61)	M= 135.45 (89.33, 181.56) B= 9.16 (−36.87, 55.20)	M= 82.16 (36.04, 128.29) B= 16.09 (−29.94, 62.12)	M= 77.24 (31.13, 123.36) B= 12.34 (−33.70, 58.37)
Left hemisphere
10	1473.84	5.42*	3.14	n.s.	96.71 (41.84, 151.57)	6.48 (−47.21, 60.18)	M= 121.48 (53.90, 189.05) B= −22.63 (−90.48, 45.21)	M= 81.87 (14.29, 149.44) B= 7.43 (−60.42, 75.28)	M= 86.78 (19.20, 154.35) B= 34.65 (−33.20, 102.50)
Model with IQ, income, and RT as covariates
Right hemisphere
10	1372.29	9.89**	n.s.	RT = 3.69***	97.58 (61.87, 133.28)	14.31 (−24.12, 52.73)	M= 135.26 (89.47, 181.05) B= 12.23 (−36.41, 60.87)	M= 80.64 (38.04, 123.24) B= 16.97 (−28.80, 62.73)	M= 76.83 (35.53,118.13) B= 13.73 (−30.43, 57.88)
16	1462.65	3.97*	n.s.	n.s.	110.72 (63.33, 158.11)	45.27 (0.65, 89.90)	M= 148.22 (85.17, 211.27) B= 99.09 (36.67, 161.50)	M= 84.65 (23.00, 146.31) B= 27.17 (−33.87, 88.21)	M= 99.28 (38.17, 160.40) B= 9.57 (−50.67, 69.81)
Left hemisphere
10	1466.47	6.20*	n.s.	RT = 3.62***	105.09 (51.41, 158.77)	5.94 (−51.83, 63.72)	M= 131.24 (62.19, 200.28) B= −21.55 (−94.88, 51.78)	M= 89.44 (25.19, 153.69) B= 4.44 (−64.57, 73.45)	M= 94.58 (32.28, 156.89) B= 34.94 (−31.66, 101.54)

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for group effects are presented. When significant, Wald Z estimates are presented for significant covariate effects (IQ, family income, response time), Wald Z condition effects are presented for comparison. Family income was not a significant predictor of brain activity differences above and beyond IQ or RT. Coefficients for each condition by group are presented for comparison.

AIC = Akaike’s Information Criterion; n.s. = not significant; CI = 95% confidence interval (lower bound, upper bound)

^*p < .05

^**p <.01

^***p < .001

Next, brain activity results for the second set of models including RT as a covariate, as well as IQ and income, were carried out. From data in 44 channels, 15 of them were significant and reliably modeled, 10 passed an FDR threshold, and 7 had significant effects (5 condition effects and 3 group effects); see Figure 3, Tables 4 and 6. Channels showing a condition effect covered left anterior temporal/inferior fontal regions (Ch 2), and right parietal regions (Ch 16–19, 21–22); see Figure 3a. Post-hoc analyses using a Bonferroni correction revealed participants had the most robust levels of increasing HbO activity during the PhR condition, in comparison to Baseline (all channels) and PhU conditions (right Ch 16, 21, 22). These post-hoc analyses also showed greater HbO activity during the PhU than Baseline condition on one right channel (Ch 22). Finally, RT was a significant predictor in channels covering the right parietal (Ch 19, 21) and left frontal (Ch 2) regions; see Table 4. IQ and family income were not significant predictors of the data in these models above and beyond RT. Channels showing a group effect covered bilateral inferior/middle frontal gyri (Ch10) and right inferior temporo-occipital (Ch 16), and revealed monolinguals engaged these regions for the PhR condition to a greater extent than control conditions (see Figure 3b, Table 6). Overall, results between models with and without RT as a covariate provide strong evidence that left frontal and right parietal regions are engaged during linguistic competition at the phonological onset.

Table 4.

Condition effect results for brain activity models of oxygenated hemoglobin (HbO) during the linguistic competition task, include IQ, family income and task response time (RT) as covariates.

Channel	AIC	F-value Condition Effect	Wald Z Condition All ps≦.001	Wald Z RT	PhR coefficient (CI)	PhU coefficient (CI)	Baseline coefficient (CI)	Post-hoc, Bonferroni p-values
Right hemisphere
16‡	1462.65	3.78*	n.s.	n.s.	123.65 (79.29, 168.01)	55.91 (12.53, 99.29)	54.43 (11.52, 97.33)	PhR vs. B, .036 PhR vs PhU, .041
19‡	1355.80	3.33*	4.95	3.65***	63.27 (29.78, 96.75)	36.56 (5.4, 67.67)	28.75 (−1.2, 58.70)	PhR vs. B, .042
21‡	1460.75	6.46**	5.19	3.55***	102.79 (51.33, 154.26)	39.63 (−8.29, 87.55)	23.2 (−23, 69.41)	PhR vs. B, .002 PhR vs PhU, .019
22‡	1548.73	11.93***	n.s.	n.s.	198.94 (119.20, 278.69)	95.64 (21.89, 169.40)	30.5 (−40.33, 101.33)	PhR vs. B, <.001 PhR vs PhU, .011 PhU vs. B, .038
Left hemisphere
2‡	1566.41	3.61*	n.s.	3.56***	129.94 (55.88, 203.99)	58.74 (−10.82, 128.30)	32.39 (−35.01, 99.79)	PhR vs. B, .027

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for condition effects are presented. When significant, Wald Z estimates are presented for covariate effects, Wald Z condition effects are presented for comparison. IQ and family income were not significant predictors of brain activity differences above and beyond RT.

AIC = Akaike’s Information Criterion; n.s. = not significant; CI = 95% confidence interval (lower bound, upper bound)

Task conditions - PhR= phonologically related condition, PhU= phonologically unrelated condition, B= baseline condition.

^*p < .05

^**p <.01

^***p < .001

^‡Indicates a channel that is also significant in the HbO model when task response time is not included as a covariate, see Table 2.

3.3.1. Deoxygenated Hemoglobin

First, brain activity results for models including IQ and family income as covariates are presented. From data in 44 channels, 14 of them were significant and reliably modeled, 4 passed an FDR threshold, and 3 had significant effects (3 condition effects, no group effects); see Figure 3a and Table 3. Channels showing a condition effect covered left frontal regions (Ch 6 and 7) and right inferior parietal (Ch 17); see Figure 3a. Post-hoc analyses using a Bonferroni correction revealed participants had the most robust levels of decreasing HbR activity during the PhR condition, in comparison to Baseline (all channels). One left channel (Ch 6) showed robust level of decreasing HbR activity during the PhR in comparison to PhU (left Ch 6), and PhU in comparison to Baseline; see Table 3. Family income and IQ were not significant predictors of HbR data.

Table 3.

Condition effect results for brain activity models of deoxygenated hemoglobin (HbR) during the linguistic competition task, include IQ and family income as covariates.

Channel	AIC	F-value Condition Effect	Wald Z Condition All ps≦.001	Wald Z IQ	PhR coefficient (CI)	PhU coefficient (CI)	Baseline coefficient (CI)	Post-hoc, Bonferroni p-values
Right hemisphere
17†	1452.80	6.31**	5.97	n.s.	−101.40 (−150.47, −52.33)	−78.04 (−127.11, −28.97)	−33.87 (−82.94, 15.20)	PhR vs. B, .002
Left hemisphere
6†	1306.47	23.90***	6.15	n.s.	−81.52 (−107.55, −55.49)	−42.05 (−68.08, −16.02)	−2.03 (−28.06, 24.00)	PhR vs. B, <.001 PhR vs. PhU, .004 PhU vs. B, <.001
7‡	1418.86	10.79***	3.06	n.s.	−108.55 (−146.18, −70.92)	−65.83 (−103.46, −28.19)	−19.55 (−57.18, 18.08)	PhR vs. B, <.001

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for condition effects are presented. When significant, Wald Z estimates are presented for covariate effects, Wald Z condition effects are presented for comparison. Family income was not a significant predictor of brain activity differences above and beyond IQ.

AIC = Akaike’s Information Criterion; n.s. = not significant; CI = 95% confidence interval (lower bound, upper bound)

Task conditions - PhR= phonologically related condition, PhU= phonologically unrelated condition, B= baseline condition.

^*p < .05

^**p <.01

^***p < .001

^†Indicates a channel that is significant in both HbO and HbR comparable models, see Table 2.

^‡Indicates a channel that is also significant in the HbR model when task response time is added as a covariate, see Table 5.

Next, brain activity results for the second set of models including RT as a covariate, as well as IQ and income, were carried out. From data in 44 channels, 10 of them were significant and reliably modeled, 10 passed an FDR threshold, and 2 had significant effects (2 condition effects, no group effects). Channels showing a condition effect covered the right inferior fontal (Ch 3) and left middle/inferior frontal (Ch 7) regions; see Figure 3a and Table 5. Post-hoc analyses using a Bonferroni correction revealed participants had the most robust levels of decreasing HbR activity during the PhR condition, in comparison to Baseline (both channels) and PhU conditions (right Ch 3). Both channels also showed the most robust level of decreasing HbR activity during the PhU than Baseline condition, and RT was a significant predictor in both channels; see Table 5. IQ and family income were not significant predictors of the data in these models above and beyond RT. Overall, results between HbR models with and without RT as a covariate in comparison to HbO results, provide strong evidence that left frontal regions are engaged during linguistic competition at the phonological onset.

Table 5.

Condition effect results for brain activity models of deoxygenated hemoglobin (HbR) during the linguistic competition task, include IQ, family income and task response time (RT) as covariates.

Channel	AIC	F-value Condition Effect	Wald Z Condition All ps≦.001	Wald Z RT	PhR coefficient (CI)	PhU coefficient (CI)	Baseline coefficient (CI)	Post-hoc, Bonferroni p-values
Right hemisphere
3	1370.11	17.31***	5.96	4.14***	−92.31 (−130.37, −54.25)	−52.37 (−87.47, −17.26)	−6.61 (−40.27, 27.04)	PhR vs. B, <.001 PhR vs. PhU, .05 PhU vs. B, <.001
Left hemisphere
7‡	1419.40	11.10***	6.31	3.49***	−109.45 (−149.95, −68.96)	−65.52 (−103.43, −27.61)	−19.86 (−56.52, 16.81)	PhR vs. B, <.001 PhU vs. B, .026

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for condition effects are presented. When significant, Wald Z estimates are presented for covariate effects, Wald Z condition effects are presented for comparison. IQ and family income were not significant predictors of brain activity differences above and beyond RT.

AIC = Akaike’s Information Criterion; n.s. = not significant; CI = 95% confidence interval (lower bound, upper bound)

Task conditions - PhR= phonologically related condition, PhU= phonologically unrelated condition, B= baseline condition.

^*p < .05

^**p <.01

^***p < .001

^‡Indicates a channel that is also significant in the HbR model when task response time is not included as a covariate, see Table 3.

4. Discussion

The present work examined the impact of bilingualism on attention-demanding language processes, during a critical period in brain development of neural specificity for semantic processing and attentional processes supporting top-down lexical selection. We investigated children’s processing of linguistic competitors at the phonological onset in a single-language (e.g., car-cat vs. car-pen), and how early and systematic bilingual exposure may influence this process. Language acquisition is mostly complete during ages 7–10, but top-down processes known to support lexical selection are still undergoing active development (Arredondo et al., 2015, 2017, 2018; Davidson et al., 2006; Skeide & Friederici, 2016). Guided by the Neuroemergentist framework (Hernandez et al., 2018^{a, b}, 2019), we hypothesized that the impact of early-life bilingual experiences would manifest in changes in how children engage top-down mechanisms in the fronto-parietal network. Specifically, we hypothesized bilingual children would show reduced activation in left frontal regions that support attentional/top-down mechanisms for lexical selection, and possibly greater activation in posterior temporo-parietal regions that support more automated lexical recognition. Our findings support this hypothesis by showing that monolingual children had stronger activation in bilateral frontal and right parietal regions, while bilinguals instead showed reduced activation in these regions. In sum, the present study offers evidence that, by mid-childhood, bilingual children have begun to engage attentional/top-down mechanisms differently for linguistic processes than monolingual-raised children.

The Neuroemergentism framework posits that the bilingual brain adapts to dual-language demands as early as childhood, where small adaptations across development cascade into broad non-linear changes (Hernandez et al., 2018^{a, b}, 2019). In particular, bilinguals dynamically adapt sub-cortical and posterior cortical neural systems in order to support linguistic processes across their lifetime. For instance, monolinguals’ language systems undergo a posterior-to-anterior shift during language development, while bilinguals continue to show greater reliance on subcortical and posterior neural systems for language and possibly other types of cognitive processes (Hernandez et al., 2018b; Marian et al., 2014, 2017). Our findings suggest that during lexical selection tasks, bilingual children exhibit less brain activity in bilateral frontal and right temporo-parietal regions in comparison to monolinguals. Within frontal regions, monolingual children showed greater brain activity for linguistic competitors (PhR condition) than non-competing conditions (PhU and Baseline), suggesting reliance of anterior frontal regions for linguistic processes. Prior research shows monolingual adults activate anterior left frontal regions for single-language linguistic competitors to a greater extent than do bilingual adults (Marian et al., 2014), while bilingual adults activated these left frontal regions when competitors across both languages were presented (Marian et al., 2017). Considering prior results with the present findings, we suggest that young bilinguals are allocating neural resources in a more automated manner by showing less engagement of anterior frontal regions during single-language competitors and principally relying on posterior regions for this lexical process. Taken together, our findings suggest that these bilingual effects emerge early in life and persist well into adulthood.

Developmental frameworks on lexico-semantic processing and selection suggest a brain network that relies on the IFOF white matter tract connections between the left frontal, temporal, and parietal brain regions (Mohades et al., 2015; cf. Skeide & Friederici, 2016). In the present study, both monolingual and bilingual children’s brain activity showed greater brain activity in bilateral anterior frontal and posterior temporo-parietal regions during lexical selection. Children’s brain activity was most pronounced in left anterior frontal and right posterior parietal regions across both HbO and HbR. More specifically, children showed task-related incremental activity across conditions: less brain activity for non-competing conditions baseline (the less difficult condition) followed by PhU, and greater for lexical competitors (PhR, the most challenging condition). Overall, these findings are consistent with two neuro-cognitive networks based on monolingual adult data: (1) the left anterior fronto-temporal network (that shifted from posterior reliance early in development) to support top-down mechanisms during the extraction of phonological and lexico-semantic processes in the mental lexicon (Friederici, 2012; Friederici & Gierhan, 2013); and (2) the bilateral fronto-parietal network known to be involved in Executive Functions, including monitoring, planning, and execution of behavior (Bunge et al., 2002^a,b; Davidson et al., 2006). We also found that non-verbal IQ ability and task-related RT predicted brain activity in bilateral frontal and right parietal regions. This result suggests that cognitive ability is associated with children’s broader brain activity in regions associated with top-down Executive Function processes. Importantly, covariates did not predict brain activity in the left posterior regions, supporting the notion that this region supports automated language processes for school-age children. In sum, the present findings suggest that when bilingual and monolingual children are grouped together, they engage an anterior-posterior network for lexical selection. However, bilingual children engage anterior brain regions to a lesser extent than monolinguals and instead show principal reliance on left posterior regions.

Remarkably, these findings are commensurate to multiple studies now showing that children in this age range who experienced early bilingualism are beginning to optimize brain structures associated with lexico-semantic processing. More balanced bilingual children have thinner left frontal and temporal cortices and thicker subcortical regions (Archila-Suerte, Woods, Chiarello, & Hernandez, 2017; Rodriguez, Archila-Suerte, Vaughn, Chiarello, & Hernandez, 2018; Olulade et al., 2015). Bilingual children show higher IFOF white matter tracts connecting lexico-semantic left anterior-posterior regions by ages 8–10 (Mohades, Struys, Van Schuerbeek, Mondt, Van De Craen, & Luypaert, 2012), but stronger by ages 10–13 (Mohades et al., 2015). The present findings are also in line with Pierce et al. (2015) where a phonological working memory task revealed bilinguals relied on posterior regions, while monolinguals relied on anterior regions (Pierce, Chen, Delcenserie, Genesee, & Klein, 2015). Thus, the present findings are in line with the Neuroemergentism framework and suggest that in mid-childhood, proficient young bilinguals rely on posterior (and likely subcortical regions such as the basal ganglia; Hernandez et al., 2018a, 2019) for automated lexical selection. Although a limitation of the current study is the low number of trials and that fNIRS has a limited spatial resolution (~3 cm of the cortex), we urge future research to cross-validate these findings by further investigating subcortical differences in children (e.g., basal ganglia, Stocco & Prat, 2014; Stocco, Yamasaki, Natalenko, & Prat, 2014).

Nevertheless, the Linguistic Competition task was successful in demanding conflict resolution among lexical competitors, by revealing that children were less accurate and took longer to respond when linguistic competitors were presented (PhR condition) followed by unrelated (PhU) and baseline trials. Overall, participants’ language and cognitive abilities, as well as task performance, were within typical ranges. Although participating families were recruited from similar neighborhoods/school districts and parents’ education levels were similar, monolingual children came from households of higher income than bilinguals. The groups differed in nonverbal intelligence, where monolinguals performed better than bilinguals, although all participants had scores within the normal range (above 85). Research suggests that differences in performance on standardized IQ tests between majority and minority (Latino) groups may stem from socio-cultural factors, more so than underlying intelligence aptitude variables (Nguyen & Ryan, 2008; Valencia & Suzuki, 2000; Wicherts & Dolan, 2010). Nevertheless, bilingual and monolingual children did not differ in their English and academic abilities. Both groups performed at ceiling (see Table 1) in the Linguistic Competition task, yet monolingual children were more accurate and responded faster than bilinguals. Since bilingual children’s English language abilities are on par with monolinguals’, these differences in task performance may suggest that bilingual children’s dual-language lexicon may have been induced, though the task was maximally designed to control for such variability (as shown by Marian et al., 2014). This was likely the case since the present study’s bilinguals were highly proficient in Spanish, their language abilities were within a typical range and parents reported consistent daily use of both languages. Nevertheless, bilinguals were more proficient in English than Spanish, as is typical for children whose school curriculum is mainly in English.

5. Conclusions

Dual-language experiences during early cognitive and brain development may influence the neural representations of cognitive systems (Hernandez et al., 2018^a,b; Kroll et al., 2015). In line with prior adult findings (Marian et al., 2014), monolingual children recruit left frontal regions to a greater degree than bilinguals during the adjudication of competing linguistic input. Here we provide developmental evidence that these top-down processes differences among bilinguals and monolinguals are likely due to managing two lexicons, to the extent that language competitors impose greater cognitive demands (Kroll et al., 2015). In support of the Neuroemergentist framework (Hernandez et al., 2018^a,b, 2019), early childhood bilingualism effects likely emerge early in life and cascade into broader and more pronounced effects in the bilingual mind and brain, across the lifetime. Importantly, these capabilities may be related to bilinguals’ (superior, yet inconclusive; Bialystok, 2017; Paap, 2018) performance in domain-general Executive Functions (Blumenfeld & Marian, 2011, 2013; Kroll & Bialystok, 2013).

Highlights.

• When processing spoken language, multiple lexical items may compete for selection.

• We used fNIRS to measure children’s brain activity when processing linguistic competitors.

• Bilingual and monolingual children activated fronto-parietal brain regions.

• Monolinguals engaged greater brain activity in left frontal regions than bilinguals.

• Bilingualism adapts top-down frontal lobe functions, likely to process dual-lexicons.

Appendix A.

List of stimuli used in trials for the Language Competition task. Parentheses represent transcriptions according to the International Phonetic Alphabet (IPA).

Phonologically Related (PhR)		Phonologically Unrelated (PhU)		Baseline
Target	Competitor	Target	Competitor	Target	Competitor - Scrambled Image
Bed (bɛd)	Bell (bɛl)	Ant (ænt)	Tie (taɪ)	Ear (ɪr)	Nine (naɪn)
Belt (bɛlt)	Bear (bɛr)	Bread (brɛd)	Star (stɑr)	Bunny (ˈbʌni)	Nest (nɛst)
Brick (brɪk)	Bridge (brɪʤ)	Carrot (ˈkærət)	Pillow (ˈpɪloʊ)	Truck (trʌk)	Bag (bæg)
Candle (ˈkændəl)	Candy (ˈkændi)	Tooth (tuθ)	Chair (ʧɛr)	Hand (hænd)	Flower (ˈflaʊər)
Cane (keɪn)	Cake (keɪk)	Clock (klɑk)	Ice (aɪs)	Fish (fɪʃ)	Crab (kræb)
Card (kɑrd)	Car (kɑr)	Leaf (lif)	Deer (dɪr)	Eight (eɪt)	Lock (lɑk)
Cheek (ʧik)	Cheese (ʧiz)	Desk (dɛsk)	Girl (gɜrl)	Foot (fʊt)	Pear (pɛr)
Cloud (klaʊd)	Clown (klaʊn)	Dress (drɛs)	Eye (aɪ)	Boy (bɔɪ)	Tent (tɛnt)
Couch (kaʊʧ)	Cow (kaʊ)	Flag (flæg)	Cat (kæt)	Apple (ˈæpəl)	Knife (naɪf)
Doll (dɑl)	Dog (dɔg)	Spoon (spun)	Frog (frɑg)	Whale (weɪl)	Train (treɪn)
Fork (fɔrk)	Four (fɔr)	Grape (greɪp)	Bus (bʌs)	Bow (boʊ)	Crown (kraʊn)
Goat (goʊt)	Ghost (goʊst)	Hat (hæt)	Ball (bɔl)	Knight (naɪt)	Lamp (læmp)
Hanger (ˈhæŋər)	Hammer (ˈhæmər)	Mop (mɑp)	Head (hɛd)	Soap (soʊp)	Tree (tri)
Horn (hɔrn)	Horse (hɔrs)	Jar (ʤɑr)	Shoe (ʃu)	Sled (slɛd)	Ring (rɪŋ)
Key (ki)	King (kɪŋ)	Nail (neɪl)	Bone (boʊn)	Cage (keɪʤ)	Bike (baɪk)
Knob (nɑb)	Knot (nɑt)	Pencil (ˈpɛnsəl)	Box (bɑks)	Book (bʊk)	Boat (boʊt)
Moose (mus)	Moon (mun)	Pig (pɪg)	Two (tu)	Brush (brʌʃ)	Mask (mæsk)
Mouse (maʊs)	Mouth (maʊθ)	Purse (pɜrs)	Juice (ʤus)	Wood (wʊd)	House (haʊs)
Onion (ˈʌnjən)	Oven (ˈʌvən)	Shark (ʃɑrk)	Door (dɔr)	Boot (but)	Scarf (skɑrf)
Swing (swɪŋ)	Swim (swɪm)	Sock (sɑk)	Bat (bæt)	Egg (ɛg)	Drums (drʌmz)
Wash (wɑʃ)	Watch (wɑʧ)	Worm (wɜrm)	Duck (dʌk)	Cup (kʌp)	Nose (noʊz)

Appendix B.

Condition effect results for brain activity models of oxygenated hemoglobin (HbO) during the linguistic competition task, no covariates.

Channel	AIC	Condition effect	PhR coefficient (CI)	PhU coefficient (CI)	Baseline coefficient (CI)	Post-hoc, Bonferroni p-values
Right hemisphere
2‡	1678.09	4.61*	137.21 (61.02, 213.4)	52.63 (−23.57, 128.82)	35.25 (−40.94, 111.44)		PhR vs. B, .013 PhR vs PhU, .05
15	1626.03	5.44*	120.3 (49.28, 191.31)	45.73 (−25.29, 116.74)	27.64 (−43.37, 98.66)		PhR vs. B, .005 PhR vs PhU, .032
17†‡	1646.36	8.75***	181.87 (104.37, 259.38)	90.96 (13.45, 168.47)	68.99 (−8.51, 146.5)		PhR vs. B, <.001 PhR vs PhU, .005
18‡	1590.92	4.93*	91.41 (39.88, 142.93)	52.93 (1.4, 104.46)	30.35 (−21.17, 81.88)		PhR vs. B, .009
19‡	1468.81	4.95*	80.66 (48.99, 112.33)	46.32 (14.65, 77.99)	37.25 (5.58, 68.92)		PhR vs. B, .01 PhR vs PhU, .05
21‡	1569.65	6.46*	97.24 (51.46, 143.03)	36.8 (−8.99, 82.58)	21.35 (−24.44, 67.13)		PhR vs. B, .003 PhR vs PhU, .023
22‡	1676.66	10.21***	188.78 (115.48, 262.08)	90.26 (16.96, 163.56)	29.51 (−43.79, 102.81)		PhR vs. B, <.001 PhR vs PhU, .022
Left hemisphere
2‡	1692.84	4.46*	134.82 (64.52, 205.12)	62.2 (−8.1, 132.51)	28.28 (−42.03, 98.58)		PhR vs. B, .016
6‡	1623.55	7.75**	150.03 (98.88, 201.19)	77.05 (25.9, 128.21)	24.59 (−26.57, 75.74)		PhR vs. B, .001
7	1706.97	3.99*	178.22 (103.91, 252.54)	87.23 (12.92, 161.55)	76.87 (2.55, 151.19)		PhR vs. B, .039
17	1675.76	10.83***	160.08 (75.07, 245.08)	68.07 (−16.94, 153.08)	13.29 (−71.72, 98.3)		PhR vs. B, <.001 PhR vs PhU, .016
18†‡	1606.37	6.68*	148.73 (95.03, 202.43)	89.14 (35.44, 142.84)	58.67 (4.97, 112.37)		PhR vs. B, .002
20‡	1769.34	15.18***	285.5 (158.32, 412.67)	155.54 (28.37, 282.72)	39.23 (−87.94, 166.41)		PhR vs. B, <.001 PhR vs PhU, .02 PhU vs. B, .007

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for condition effects are presented.

AIC = Akaike’s Information Criterion; CI = 95% confidence interval (lower bound, upper bound)

Task conditions - PhR= phonologically related condition, PhU= phonologically unrelated condition, B= baseline condition.

^*p < .05

^**p <.01

^***p < .001

^†Indicates a channel that is significant in both HbO and HbR comparable models, see Table below.

^‡Indicates a channel that is also significant in the HbO model when task response time and/or IQ, income are added as covariates, see Tables 2–5 in the text.

Condition effect results for brain activity models of deoxygenated hemoglobin (HbR) during the linguistic competition task, no covariates.

Channel	AIC	Condition effect	PhR coef (CI)	PhU coef (CI)	Baseline coef (CI)	Post-hoc, Bonferroni p-values
Right hemisphere
6	1440.88	12.76***	−69.86 (−98.14, −41.58)	−36.56 (−64.84, −8.28)	−3.48 (31.76, 24.8)	PhR vs. B, <.001 PhR vs. PhU, .05 PhU vs. B, .011
11	1461.49	5.76**	−46.79 (−75.05, −18.52)	−18.04 (−46.31, 10.22)	−4.8 (−33.06, 23.47)	PhR vs. B, .008
17‡	1557.53	6.54**	−100.15 (−149.64, −50.66)	−77.32 (−126.82, −27.83)	−34.79 (−84.28, 14.71)	PhR vs. B, .002
Left hemisphere
8	1444	14.43***	−69.75 (−99.86, −39.65)	−37.53 (−67.63, −7.42)	−2.91 (−33.01, 27.2)	PhR vs. B, <.001 PhR vs. PhU, .04 PhU vs. B, .011
11	1485.64	4.81*	−64.56 (−95.66, −33.46)	−31.68 (−62.79, −.58)	−15.3 (−46.41, 15.8)	PhR vs. B, .009
18	1408.8	3.67*	−34.45 (58.95, −9.95)	−24.21 (−48.71, .29)	−5.44 (−29.94, 19.06)	PhR vs. B, .028

Notes. Models presented here passed the False Discovery rate (FDR) threshold at p<.05, F-values for condition effects are presented.

AIC = Akaike’s Information Criterion; CI = 95% confidence interval (lower bound, upper bound)

Task conditions - PhR= phonologically related condition, PhU= phonologically unrelated condition, B= baseline condition.

^*p < .05

^**p <.01

^***p < .001

^†Indicates a channel that is significant in both HbO and HbR comparable models, see Table above.

^‡Indicates a channel that is also significant in the HbR model when task response time and/or IQ, income are added as covariates, see Tables 2–5 in the text.