Processing of code-switched sentences by bilingual children: Cognitive and linguistic predictors

Abstract

Production studies of language switching have identified costs in the speed and/or accuracy of word production, but it is unclear whether processing costs are experienced by listeners as well. A related question is whether language control during comprehension recruits domain-general cognitive control. The current study examined processing of code-switching in Spanish-English bilingual children (ages 6;0–11;10) using an auditory moving window paradigm. Cognitive control was indexed by the Dimensional Change Card Sort. Children exhibited significant costs in processing speed when listening to code-switched sentences, but no costs in a measure of offline comprehension. The extent to which cognitive control skills moderated processing costs depended on the robustness of the language system: children with higher language skills exhibited a greater moderating effect of cognitive control. Taken together, the findings provide limited support for a role of cognitive control in children’s code-switching processing and suggest that the processing costs incurred may be transitory.

Bilinguals frequently code-switch, or alternate between languages, in conversation with other bilinguals. These language switches can occur between sentences (inter-sentential code-switching) or within a single sentence (intra-sentential code-switching). Studies across communities suggest that both intra- and inter-sentential code-switching may be a frequent component of bilingual parents’ input to their children (e.g., Byers-Heinlein, 2013), including among Spanish-English bilingual children in the United States (e.g., Bail, Morini, & Newman, 2015; Place & Hoff, 2011, 2016). Despite the frequency of code-switching as a conversational strategy by bilingual adults, parents are often advised to use only one language at a time and to avoid code-switching around their children, especially for children with language difficulties (e.g., Kohnert, Yim, Nett, Kan, & Duran, 2005). However, the available empirical evidence does not paint a clear picture regarding potential costs associated with processing code-switched input.

During bilingual production, language switching has been repeatedly shown to exert costs in the form of slower production when switching into a different language compared to staying in the same language. These costs have been identified in bilinguals of varying proficiency and in a variety of tasks, including both cued and voluntary switching (e.g., Costa, Santesteban, & Ivanova, 2006; Declerck, Stephan, Koch, & Philipp, 2015; Gollan & Ferreira, 2009; Heikoop, Declerck, Los, & Koch, 2016; Meuter & Allport, 1999; Olson, 2016; Tarlowski, Wodniecka, & Marzecova, 2013). The presence of costs during production has been proposed to reflect the operation of cognitive control processes (e.g., task engagement/disengagement, response inhibition, interference control) that help to regulate the activation of the target and non-target languages (e.g., Green, 1998; Green & Abutalebi, 2013). However, studies in adults have yielded mixed results as to whether comprehension of a language switch also imposes costs (i.e., more effortful or slower processing) and recruits domain-general cognitive control (e.g., Abutalebi et al., 2007; Adamou & Shen, 2017; Blanco-Elorrieta & Pylkkanen, 2016, 2017; Jylkkä et al., 2017; Liao & Chan, 2016).

Furthermore, the theoretical models and findings from studies in adults cannot necessarily be generalized to children, whose cognitive control skills are still developing (e.g., Best & Miller, 2010; D’Souza & D’Souza, 2016; Davidson, Amso, Anderson, & Diamond, 2006). Previous work with children has yielded inconsistent findings regarding the effects of mixed-language input on children’s language outcomes (e.g., Bail et al., 2015; Byers-Heinlein, 2013; Place & Hoff, 2011, 2016), the presence of costs when children listen to language switching (e.g., Byers-Heinlein, Morin-Lessard, & Lew-Williams, 2017; K Kohnert & Bates, 2002; Peynircioglu, Durgunoglu, Heredia, & Altarriba, 2002; Potter, Fourakis, Morin-Lessard, Byers-Heinlein, & Lew-Williams, 2019), and the underlying mechanisms reflected by such costs (e.g., Byers-Heinlein et al., 2017; Potter et al., 2019). It is of theoretical importance to develop models of language control that apply to individuals at different developmental points and that account for individual differences in the ability to process and comprehend mixed-language input. With this goal in mind, in the current study, we examined children’s ability to process code-switched sentences and considered the cognitive and linguistic skills that may support this ability.

Comprehension of Language Switching in Adults

Production studies have robustly demonstrated costs associated with language switching (e.g., Costa, Santesteban, & Ivanova, 2006; Declerck, Stephan, Koch, & Philipp, 2015; Gollan & Ferreira, 2009; Heikoop, Declerck, Los, & Koch, 2016; Meuter & Allport, 1999; Olson, 2016; Tarlowski, Wodniecka, & Marzecova, 2013), and whether or not a language switch imposes costs on comprehension has been the subject of a parallel line of research. Comprehension-to-production studies have yielded costs in speed when participants were required to produce a word in one language after listening to input in a different language, suggesting that language control mechanisms may be shared between comprehension and production (e.g., Gambi & Hartsuiker, 2016). However, other studies have suggested that the mechanisms of language control during comprehension may be different from those employed during production (e.g., Blanco-Elorrieta & Pylkkanen, 2016; Mosca & de Bot, 2017). Comprehending a language switch has been described as a “passive” process (e.g., Blanco-Elorrieta & Pylkkanen, 2016), and it is possible that it may be less costly than the more active process of producing a language switch.

Most comprehension studies of language switching in adults have been conducted in the visual modality using single written words or self-paced reading of sentences. Some studies have found costs in terms of slower response speed or event-related potential (ERP) components that reflect processing difficulties (e.g., N400, LAN, LPC; see Liao & Chan, 2016 for a review). These costs have been observed both in single-word paradigms (e.g., Jylkkä et al., 2017; Macizo, Bajo, & Paolieri, 2012; Mosca & de Bot, 2017; von Studnitz & Green, 2002) and sentence contexts (e.g., Altarriba, Kroll, Sholl, & Rayner, 1996; Bultena, Dijkstra, & Van Hell, 2015; Litcofsky & Van Hell, 2017; Moreno, Federmeier, & Kutas, 2002; Ng, Gonzalez, & Wicha, 2014; Philipp & Huestegge, 2014; Wang, 2015). However, in other studies, bilinguals were able to process language switches without costs when the code-switches followed patterns frequently documented in production corpora and/or when participants had extensive experience with code-switching (e.g., Beatty-Martínez & Dussias, 2017; Chan, Chau, & Hoosain, 1983; Ibanez, Macizo, & Bajo, 2010; Johns, Valdés Kroff, & Dussias, 2018). However, code-switching is largely a phenomenon of spoken language. Findings documented in the visual modality may not be representative of what occurs when bilinguals listen to code-switching. Furthermore, auditory input may provide subtle cues about a speaker’s bilingual vs. monolingual status and help listeners to predict an upcoming code-switch in ways that would not be possible during reading.

Fewer studies have examined language switching costs in comprehension using auditory stimuli. As in the visual modality, results have been mixed. When listening to sentences that contained a language switch, bilinguals have exhibited processing costs compared to listening to single-language sentences, both in processing speed (e.g., Adamou & Shen, 2017; Fricke, Kroll, & Dussias, 2016; Olson, 2016a) and in evoking ERP components associated with processing difficulties (e.g., Liao & Chan, 2016). However, subtle acoustic cues that distinguish code-switched sentences from single-language sentences, including F0 contours (Piccinini & Garellek, 2014) and slowed speech rate and shifts in voice-onset time (Fricke et al., 2016), have been shown to reduce costs for listeners. Code-switches that were more likely to occur in spontaneous code-switching in a given community were also less likely to elicit costs (Adamou & Shen, 2017). In addition, Olson (2016a) found no costs in response to a language switch within a sentence if another language switch had already occurred earlier in the sentence, creating a “bilingual mode.”

What are the proposed mechanisms behind these variable costs? Language control models associated with both production (e.g., Inhibitory Control Model (ICM), Green, 1998) and comprehension (e.g., BIA, Dijkstra & van Heuven, 2002; BIA+, Dijkstra & van Heuven, 1998; BIA-d, Grainger, Midgley, & Holcomb, 2010) posit that accessing the target language after a language switch involves modulation of language activation levels, which requires effort and results in a processing cost. In the BIA models, this modulation of activation is driven by bottom up processes within the language system, while in the ICM it requires domain-general control processes outside the language system. However, extensions of the ICM (e.g., Adaptive Control Hypothesis, Green & Abutalebi, 2013; Control Process Model of Code-Switching, Green & Wei, 2014) suggest that a cooperative mode of language control also exists in dense code-switching contexts in which both languages are equally accessible and modulation of activation is not necessary. If listening to language switching involves this cooperative mode, then there may not be costs (e.g., Olson, 2016a). Another interpretation, drawing on usage-based models of comprehension (e.g., Production - Distribution - Comprehension account, MacDonald, 2013), is that processing a language switch is only costly if it violates listeners’ predictions based on their own production habits (e.g., Adamou & Shen, 2019; Beatty-Martínez & Dussias, 2017; Fricke et al., 2016; Johns et al., 2019). Taken together, past work in adults suggests that listening to code-switched input may impose costs, but there are a variety of potentially mitigating circumstances.

Comprehension of Language Switching in Children

It is unclear how adult models and findings from adult language switching studies may apply to children, whose cognitive, linguistic, and metalinguistic skills are still developing. While there is evidence from studies of cued and voluntary language switching that children experience language switching costs in both the speed and accuracy of word production (e.g., Gross & Kaushanskaya, 2015, 2018; Jia, Kohnert, Collado, & Aquino-Garcia, 2006; Kohnert, 2002; Kohnert, Bates, & Hernandez, 1999), studies that have examined processing of language switching in children have yielded varying outcomes.

A few studies have suggested that children can process mixed-language input without costs. At the single word level, Kohnert and Bates (2002) did not find any costs in accuracy or reaction time when Spanish/English bilingual children (ages 5–16) performed a picture-word verification task in a mixed-language condition. Scaling up to a higher linguistic level, Peynircioglu and colleagues (Peynircioglu et al., 2002) found that Spanish/English bilingual preschool children were as accurate in answering comprehension questions after listening to a mixed-language story as after listening to a single-language story. However, in an eye-tracking study at the sentence level, Byers-Heinlein, Morin-Lessard, and Lew-Williams (2017) found that French/English bilingual 20-month olds, as well as bilingual adults, exhibited lower accuracy in looking toward objects labeled in a mixed-language sentence than a single-language sentence, particularly when the switch occurred into their non-dominant language. Potter and colleagues (Potter et al., 2019) also identified costs for switches into the non-dominant language using a similar paradigm with Spanish/English bilingual toddlers. Although these findings for offline comprehension vs. online processing came from separate studies and different age groups, they could suggest that costs incurred as the language switch unfolds do not necessarily translate to down-stream effects on sentence comprehension. To further investigate this possibility, the current study included measures of both offline comprehension and online processing within the same task.

Byers-Heinlein and colleagues (2017) suggested that the costs observed in online processing of a language switch may reflect a “side effect of efficient bilingual processing” (Byers-Heinlein et al., 2017, p. 9035) rather than a sign of processing difficulties. Based on parallel findings of increased pupil dilation, they further associated these online costs with increased cognitive load and in turn as evidence that even very young children exert control over their languages as they listen. However, Potter and colleagues (2019) suggested an explanation of costs based on prediction and statistical regularities rather than cognitive control. As in the adult literature, this raises the question of the extent to which domain-general cognitive control processes are recruited to process language switches during comprehension.

The Role of Cognitive Control in Processing Code-Switching

Production models of language control (e.g., Adaptive Control Hypothesis, Green & Abutalebi, 2013; Inhibitory Control Model, Green, 1998; Control Process Model of Code-switching, Green & Wei, 2014) incorporate control processes outside the language system as part of a set of domain-general control mechanisms used to regulate a variety of behavior (also known as executive functions, e.g., Miyake & Friedman, 2012; Miyake et al., 2000). Building on the framework of Miyake and colleagues, Green and Abutalebi (2013) list eight domain-general control processes proposed to be involved in language control (goal maintenance, conflict monitoring, interference suppression, salient cue detection, selective response inhibition, task disengagement, task engagement, opportunistic planning). In the current paper, we use the broad term “cognitive control” to refer to this set of domain-general control processes. Behavioral studies attempting to link language control during production with measures of domain-general cognitive control have yielded mixed findings, with some studies identifying significant associations (e.g., Festman & Münte, 2012; Festman, Rodriguez-Fornells, & Munte, 2010; Gollan, Sandoval, & Salmon, 2011; Prior & Gollan, 2011; Woumans, Ceuleers, Van der Linden, Szmalec, & Duyck, 2015) and other studies identifying non-significant associations or different patterns of linguistic vs. non-linguistic switching costs (e.g., Branzi, Calabria, Boscarino, & Costa, 2016; Calabria, Branzi, Marne, Hernández, & Costa, 2015; Gollan, Kleinman, & Wierenga, 2014; Jylkkä, Lehtonen, Lindholm, Kuusakoski, & Laine, 2018; Weissberger, Wierenga, Bondi, & Gollan, 2012).

Comprehension models of language control (e.g., BIA, Dijkstra & van Heuven, 1998; BIA+, Dijkstra & van Heuven, 2002; BIA-d, Grainger et al., 2010) have generally situated control over language activation levels entirely within the linguistic system without the involvement of domain-general control processes. However, studies with mixed-language stimuli interleaved within the trials of a non-linguistic cognitive control task have provided some support for a role for cognitive control in the processing of code-switching. For both single words (Wu & Thierry, 2013) and intra-sentential code-switching (Adler, Valdés Kroff, & Novick, 2019), adult bilinguals were better able to resolve conflict in a nonverbal task when trials were presented with mixed-language than with single-language stimuli. This effect of the language context on cognitive control performance has been interpreted to suggest that cognitive control was being recruited to process the mixed-language input. However, Jylkkä and colleagues (2018), who measured cognitive control and mixed-language processing separately, found no consistent relationship.

Given these mixed results, it is possible that the involvement of cognitive control in the processing of language switches may depend on participant and task characteristics. For example, in an ERP study of reading code-switched sentences, Beatty-Martinez & Dussias (2017) found an early positive component associated with modulation of attentional control in adult bilinguals who did not code-switch, but they did not find such a response in habitual code-switchers. With regard to task factors, a semantic categorization task (Jylkkä, Lehtonen, Kuusakoski, et al., 2018) may not require modulation of language-specific activation levels to the same extent as processing an intra-sentential code-switch (Adler et al., 2019).

Further, the role of cognitive control in the processing of mixed-language input may vary depending on the maturity of the linguistic and the cognitive system. Based on their pupillometry findings, Byers-Heinlein and colleagues (2017) argued that control processes modulating access to each language were active as 20-month olds processed mixed-language input. Conboy, Sommerville, Wicha, Romo, and Kuhl reported in a conference paper (2011) that two-year olds with higher cognitive control skills were more efficient at processing words in a mixed-language condition. The children in these studies were very young and in the very early stages of developing inhibitory control (e.g., Best & Miller, 2010). The current study examines the role of cognitive control in the processing of intra-sentential code-switching in school-age children, whose cognitive control is more developed but still not adult-like (e.g., Davidson et al., 2006; Huizinga et al., 2006).

The Role of Language Ability in Processing Code-Switching

If language control during comprehension is at least partially situated within the language system, as predicted by the BIA model and its variants (e.g., BIA, Dijkstra & van Heuven, 1998; BIA+, Dijkstra & van Heuven, 2002; BIA-d, Grainger et al., 2010), then the robustness of a child’s linguistic system may play a role in modulating the ability to process mixed-language input. In addition, when discussing individual differences in reference to their Adaptive Control Hypothesis of language control, Green and Abutalebi (2013) note that their model assumes high proficiency in both languages; thus, it is unclear how variability in language skill may affect the mechanisms of language control. Studies of the linguistic factors contributing to language control have generally focused on relative proficiency in each language and asymmetries in switching costs (e.g., Liao & Chan, 2016; Litcofsky & Van Hell, 2017; Wang, 2015) rather than effects of overall language ability (i.e., the ability to grasp the meaning and structure of language, regardless of which language it is) on language control.

In adults, Moreno and colleagues (2002) found that participants who were more proficient in Spanish showed an overall benefit in language control such that they noticed a code-switch earlier and showed smaller processing costs in the late positive complex (LPC) ERP component associated with sentence integration. In young Spanish/English bilingual children (20 months), Conboy and Mills (2006) found that the timing and localization of ERP responses (negative component to unknown vs. known words in the 200–600 ms range) in a mixed-language paradigm depended both on children’s language dominance and on their overall conceptual vocabulary, which incorporated skills in both languages. The role of children’s overall language skills in their ability to process mixed-language input is important because parents who have children with language challenges are often advised to keep their languages separate (e.g., Kohnert et al., 2005), but it is unclear whether low language ability may exacerbate any costs associated with processing mixed-language input.

Current Study

The current study examined processing of code-switched sentences by school-age Spanish-English bilingual children (ages 6;0–11;10) and considered the extent to which factors outside the language system (i.e., domain-general cognitive control skills) and factors within the language system (i.e., overall language ability) modulated any costs associated with processing a language switch. An auditory moving window paradigm (Ferreira, Henderson, Anes, Weeks, & McFarlane, 1996) was used to examine children’s processing of code-switched sentences as they unfolded. To assess the cognitive and linguistic predictors of children’s ability to process code-switched sentences, children also completed the Dimensional Change Card Sort (DCCS) as a broad measure of cognitive control and the Clinical Evaluation of Language Fundamentals in English (Semel, Wiig, & Secord, 2003) and Spanish (Wiig, Secord, & Semel, 2006) as measures of language ability. The current study sought to answer the following questions:

1. How does a language switch within a sentence affect bilingual children’s processing speed relative to single-language sentences? What are the effects on their offline comprehension of the sentence’s meaning?

2. How are the effects of the language switch on processing speed and offline comprehension modulated by differences in language ability and cognitive control?

Method

Participants

The current study included ninety Spanish-English bilingual children (36 boys) between the ages of 6;0 and 11;10 (M_Age = 9;0 years;months, SD = 1;3). Participants were defined as bilingual based on Kohnert’s (2013) functional, needs-based definition of an individual who uses or needs both Spanish and English to succeed in their home and school environments. Children reported by their parents to speak both English and Spanish were included in the study as long as they were not subject to the exclusionary criteria described below. Participants represented diverse acquisition experiences: 60% of children (n = 54) were exposed to Spanish from birth, and 40% (n = 36) were native English speakers who acquired Spanish through dual-immersion programs with 50–90% of school instruction in Spanish. Among the native Spanish speakers, children varied in whether they were exposed to English from infancy (n = 34) or after age 2 (n = 20). Table 1 presents the background characteristics of the sample. Because we were interested in the effects of language ability on processing of code-switching, children with a broad range of language skills were included in the study. None of the children had an official diagnosis of language impairment, but some children demonstrated low skills in both languages that may reflect sub-clinical or undiagnosed language difficulties.

Table 1.

Language background characteristics for participants (n=90) based on parent report.

Characteristic	Mean (SD)
Age of First English Exposure (months) Age of First Spanish Exposure (months)	9.69 (17.12) [Range: 0–60] 23.42 (29.53) [Range: 0–84]
Current English Exposure (% waking hrs / week) Current Spanish Exposure (% waking hrs / week) Average Exposure to Code-Switching (0–10 scale)a	61.77% (16.93) [Range: 12–99] 38.23% (16.93) [Range:1–88] 3.07 (1.63) [Range: 0–6.78]
Language of Instruction at School b	Eng-only: 29% / Eng+Span: 71%
Language Currently Heard by Child in Home b	Eng: 52% / Span: 30% / Both: 18%
Maternal Education (1–5 scale) c	3.29 (1.42) [Range: 1–5]
Nonverbal Intelligence d	108.37 (14.00) [Range: 81–135]
English Core Language Standard Score e	98.60 (19.65) [Range: 40–136]
English Receptive Language Standard Score e	101.26 (18.32) [Range: 51–135]
English Expressive Language Standard Score e	98.44 (19.73) [Range: 45–136]
Spanish Core Language Standard Score e	86.67 (12.69) [Range: 50–114]
Spanish Receptive Language Standard Score f	96.74 (11.72) [Range: 68–126]
Spanish Expressive Language Standard Score f	82.38 (12.89) [Range: 51–108]
Highest Core Language Standard Score g	103.65 (13.62) [Range: 73–136]
Dimensional Change Card Sort Mixed Phase (prop. correct)	0.79 (0.15) [Range: 0.3–1.0]

^aAverage of parent ratings for how frequently (0–10 scale) the child hears code-switching from the following people in the environment: mother, father, siblings, grandparents, other relatives, friends during play, classmates at school, adults at school, and unfamiliar individuals in the community.

^bPercentages reflect the percent of the sample in each category.

^cScale: 1 = < HS, 2 = HS/GED, 3 = some college, 4 = BA/BS, 5 = at least some graduate school

^dNonverbal Intelligence was indexed by the Perceptual Reasoning Index from the Weschler Intelligence Scale for Children (WISC-IV) or the Nonverbal Matrices subtest of the Kaufman Brief Intelligence Test (KBIT-2).

^eClinical Evaluation of Language Fundamentals, 4th Edition (CELF-4)

^fClinical Evaluation of Language Fundamentals, 4th Edition Spanish (CELF-4 Spanish)

^gThe Core Language Standard Score from the CELF-4 English (n = 66) or the CELF-4 Spanish (n=23), whichever was higher. Core Language scores were not available for one participant.

Exclusionary criteria included learning disabilities, psychological/behavioral disorders, neurological impairment, hearing impairment, other developmental disabilities, nonverbal intelligence scores below 80, current exposure to a language other than English or Spanish (>5% of waking hours in a typical week), and significant past exposure to a language other than English and Spanish (e.g., daycare provider spoke another language to the child). All children passed a bilateral pure tone hearing screening at 25 dB at 1000 Hz, 2000 Hz, and 4000 Hz in the testing room. An additional eight children completed the experimental tasks but could not be included in the analysis due to an existing diagnosis of language impairment (n = 1), nonverbal intelligence scores below 80 (n = 2), knowledge of a third language (n = 2), failed hearing screening (n = 1), and experimental error (n = 2).

Procedure

Data were collected during 2–3 sessions in a research laboratory. All procedures were approved by the Institutional Review Board. Parents provided informed consent in their preferred language, and children provided assent in their preferred language prior to being escorted to a separate room for testing. Children completed standardized assessments of language and cognition, the Auditory Moving Window paradigm to measure processing of code-switched sentences, and the Dimensional Change Card Sort (DCCS) to measure cognitive control. English and Spanish language assessments were conducted in separate sessions by research assistants proficient in the target language. The auditory moving window paradigm occurred either at the beginning of the third session prior to language testing or at the end of the second session, preceded by nonlinguistic cognitive control tasks (including the DCCS). Task instructions were generally provided in English, unless the child demonstrated a preference for Spanish. Parents completed an interview and a background questionnaire in their preferred language to provide information about their child’s development, education, language use and exposure, relevant medical history, and family background. In addition, parents completed the Language Experience and Proficiency Questionnaire (LEAP-Q, Marian, Blumenfeld, & Kaushanskaya, 2007) about their own language background and educational history. Maternal education level was scored using a Likert scale: 1=less than high school; 2 = high school or GED; 3 = some college; 4 = college; 5 = at least some graduate school.

Auditory Moving Window Task

An auditory moving window paradigm (Ferreira et al., 1996) indexed children’s processing of code-switched sentences as they unfolded. Ten sentences were constructed to be similar in length (7–10 words) and complexity, such that the sentences could be divided into three segments: a noun phrase introducing the subject, a verb phrase containing an object, and an additional clause or phrase introduced by a coordinating or subordinating conjunction or a preposition. The language of the three segments was manipulated to create 40 sentences in four conditions (see Appendix): single-language English with all three segments in English (e.g., The mom | sings to the boy | that is sleeping), single-language Spanish with all segments in Spanish (e.g., La mamá | le canta al niño | que está durmiendo), code-switched English with the first two segments in Spanish and a switch into English on the third segment (e.g., La mamá | le canta al niño | that is sleeping), and code-switched Spanish with the first two segments in English and a switch into Spanish on the third segment (e.g., The mom | sings to the boy | que está durmiendo). The language switches occurred at a natural boundary within the sentence and did not violate the grammatical rules of either language.

For each stimulus sentence, different yes/no questions in English and Spanish indexed children’s understanding of the sentence content. English questions (e.g. Does the mom sing?) followed sentences that began in English (i.e., single-language English and code-switched Spanish) and Spanish questions (e.g., ¿Baila la mama?) followed sentences that began in Spanish (i.e., single-language Spanish and code-switched English). The distribution of yes/no responses was balanced across English and Spanish questions. In addition, four practice sentences and questions were developed, one in each of the four conditions. A Spanish-English bilingual young adult female, who was a native speaker of both languages, recorded all sentences and comprehension questions in a sound proof booth. Acoustic analysis software (Praat, Boersma & Weenink, 2014) was used to normalize the sentences to 70 dB and splice them into the three segments. English and Spanish segments were re-combined to create sentences in the four conditions. To maintain experimental control for the comparison between single-language and code-switched sentences within each language, the same final segment recordings appeared in both conditions.¹

The sentences were presented to the children one segment at a time using SuperLab 4.5 (Cedrus Corporation, 2012). Children were instructed to press the space bar as soon as they were done listening to each part to hear the next part or the comprehension question. Children provided a yes/no response to the question by pressing a green button positioned over the “/” key on a keyboard or a red button positioned over the “z” key. During four practice trials, the children were coached, as needed, to press the space bar as soon as they were done listening and to press the red or green button to answer the question. They received feedback on their yes/no responses during the practice phase. The 40 test sentences were presented in a pseudo-randomized sequence to prevent consecutive repetitions of different versions of the same sentence (e.g., sentence 6 in the single-language English condition could not be followed immediately by sentence 6 in the code-switched Spanish condition).

Processing speed was indexed by children’s reaction time (RT) to press the space bar after the third segment, which is where the language switch occurred in the code-switched conditions. The effect of the language switch was indexed by comparing RTs to the third segment in the code-switched condition and the corresponding single-language condition. For example, for “La mamá le canta al niño that is sleeping,” the comparison would be between children’s RT to “that is sleeping” when it is preceded by “La mamá le canta al niño” vs. when it is preceded by “The mom sings to the boy.” The third segments varied in duration across sentences (595 – 1611 ms), with the Spanish final segments tending to be longer (M = 1107 ms, SD = 315) than the English final segments (M = 934 ms, SD = 169), W = 31, p = .162. Although the Wilcoxon rank sum test comparing English and Spanish durations was not significant, to account for this variability, segment duration was included as a covariate in the analyses. Trials in which children pressed the space bar before they heard the content of the third segment (RT < 400 ms) were excluded from both RT and accuracy analyses as impulsive (n = 7 out of 3600 trials). RTs more than 3.0 standard deviations slower than each child’s individual mean (n = 64 trials) were also excluded from the RT analysis. As a result, 1.97% of RTs were excluded from the final processing speed analysis. Trials in which children responded incorrectly to the comprehension question were still included in the analysis of processing speed, as the two outcome measures may index different processes.

Children’s accuracy in answering the yes/no questions indexed their offline comprehension of the sentence content. Trials in which children responded within 400 ms (n = 18) were excluded because they would not have heard enough of the question to provide a valid response. Trials in which children accidentally pressed the space bar instead of the “yes” or “no” keys (n = 126 out of 3600 trials) or pressed other stray keys (n = 2) were also excluded. Combining impulsive responses to the third segment or the question and stray key presses, 4.25% of the total yes/no responses were removed from the final comprehension analysis. Key presses within a one-key radius of the target “z” and “/” keys (n = 8) were assumed to reflect a finger slip and were accepted as valid responses.

Dimensional Change Card Sort (DCCS) Task

The children completed a Dimensional Change Card Sort (DCCS) task as an omnibus measure of domain-general cognitive control that incorporates multiple processes, including switching, interference suppression, and monitoring. Because all of these processes are hypothesized to be involved in language control (e.g., Green & Abutalebi, 2013), the DCCS is ideally suited to capturing variability in children’s broad domain-general cognitive control skills. The version of the DCCS administered in the current study integrated components of the “colour-shape game” (Bialystok & Martin, 2004) and the NIH toolbox DCCS (Zelazo et al., 2013), while minimizing language demands as much as possible. The stimuli were simple red circles and blue squares, initial verbal instructions were presented in the child’s preferred language with photographs that illustrated what to do, and sorting cues were presented nonverbally at the top of the screen (a row of amorphous color patches for sorting by color, and a row of grey circles and squares for sorting by shape). To reduce working memory demands, the cues remained on the screen throughout each trial.

The DCCS task was presented on a desktop computer monitor using E-Prime 2.0 (build 2.0.10.242, Psychology Software Tools, 2012). Throughout the task, grey response buckets marked with a red square and a blue circle were present at the left and right bottom corners of the screen, and children had a corresponding response box with left and right buttons marked with a red square and a blue circle. For each trial, the sorting cue appeared at the top of the screen, and, after 500 ms, the stimulus (a red square or blue circle) appeared in the center of the screen while the cue remained at the top. Children were instructed to put the stimulus into one of the buckets by pressing the corresponding button as quickly as possible without making mistakes. The cue, stimulus, and response buckets remained on the screen until the child responded or for up to 10 seconds. Following the child’s response or the end of the response window, the next trial began after an inter-trial interval of 800 ms.

The trials were organized into three phases: pre-switch (i.e., before presenting a switch in sorting rules), post-switch (i.e., after presenting the rule switch), and mixed (i.e., switching back and forth between sorting rules). During the pre-switch phase, the children were introduced to the “color game” and taught to sort the blue square into the bucket marked with the blue circle and to sort the red circle into the bucket marked with the red square by pushing the corresponding buttons. They completed four practice trials with feedback. To ensure that children understood the task, the instructions and practice were repeated if a child responded incorrectly on more than one practice trial. Then the child completed the five pre-switch trials with no feedback. In the post-switch phase, children were introduced to the “shape game” with an example of how to sort the red circle into the bucket with the blue circle and the blue square into the bucket with the red square. The children completed the five post-switch trials with no practice to avoid diluting the effect of the rule shift. Finally, the children were told they would play both games together in the mixed phase (30 trials) and were instructed to look at the cues at the top of the screen each time to know which game to play. The mixed phase followed the design of the NIH toolbox version (Zelazo et al., 2013) such that there were more shape trials (n = 23) than color trials (n = 7), trials were presented in a fixed pseudorandomized sequence with 2–5 shape trials between each color trial, and color trials never repeated consecutively. Accuracy and reaction time data were collected for each trial. Following previous work with a similar version of the DCCS task (Gross & Kaushanskaya, 2018), overall accuracy during the mixed phase was selected to index children’s cognitive control.

Standardized Measures

To measure language skills, children were administered the Clinical Evaluation of Language Fundamentals, 4^th Edition, in English (CELF-4, Semel et al., 2003) and in Spanish (CELF-4 Spanish, Wiig et al., 2006). The higher of the two Core Language Index scores (English for 66 children; Spanish for 23 children) was used to index overall language ability, or the robustness of children’s language skills regardless of the specific language. This approach is conceptually similar to taking children’s best performance across languages in the Bilingual English Spanish Assessment (BESA, Peña, Gutierrez-Clellen, Iglesias, Goldstein, & Bedore, 2014) and in experimental measures of semantic skills (Sheng, McGregor, & Marian, 2006). To ensure that nonverbal intelligence was within normal limits, children completed the Nonverbal Matrices subtest of the Kaufman Brief Intelligence Test (KBIT-2, Kaufman & Kaufman, 2004; n = 34) or the Perceptual Reasoning Index of the Wechsler Intelligence Scale for Children (WISC-IV Wechsler et al., 2003).

Analyses

Mixed effects linear regression models and mixed effects logistic regression models (lme4 package, Bates, Maechler, Bolker, & Walker, 2015) examined the effects of a language switch on processing speed and offline comprehension. The reaction time for children’s button press after the third segment of each sentence indexed processing speed. Reaction time values were log₁₀₋ transformed to better approximate a normal distribution. The accuracy of children’s responses to the yes/no questions after each sentence indexed offline comprehension. Because accuracy is a binary variable (0/1), logistic regression was used to examine these data.

In each set of analyses, the base model included the task-level categorical predictors of sentence type (single-language vs. code-switched) and the language of the stimulus. Because children varied in their language dominance, the language of the stimulus was coded in terms of each child’s dominant vs. non-dominant language (based on their performance on the English vs. the Spanish CELF-4), rather than English vs. Spanish. A deviation coding scheme (−0.5/0.5) measured main effects of each categorical predictor collapsed across levels of the other predictor (Mirman, 2014). For the analysis of processing speed, segment duration in the log₁₀ scale was included as a trial-level covariate. Following Barr’s “keep it maximal” approach (Barr, Levy, Scheepers, & Tily, 2013), these base models included random intercepts for both participants and items and random by-participant slopes and by-item slopes for the effects of sentence type, language, and their interaction, given that both variables were manipulated within participants and within items.

Next, moderation models containing child-level predictors were constructed to examine whether the effect of sentence type was moderated by children’s overall language ability and/or cognitive control skills. Because the research questions focused on sentence type, the moderation models included only the task-level variable of “sentence type,” collapsing across languages.² The child-level predictors of interest were language ability (indexed by the higher of the Core Language Index scores on the CELF-4) and cognitive control (indexed by accuracy in the DCCS mixed phase), and their interactions with the effects of sentence type. The interaction of language ability and cognitive control with each other and the three-way interaction among cognitive control, language ability and sentence type were also included to test whether the effects of cognitive control depended on children’s level of language ability. In addition, control variables (age, maternal education) that may explain additional variance in task performance were tested in an initial regression, and those with significant effects on the outcome variable were retained as covariates in the target models with the predictors interest. All continuous variables were centered around the group mean. All models included random intercepts for both participants and items, a random by-participant slope for the within-participant variable (i.e., sentence type), and random by-item slopes for within-item variables (i.e., sentence type, language ability, cognitive control, and interactions among them). For all analyses, effects with a t-value (or z-value for logistic regression) greater than 1.96 were considered significant (p < .05).

Results

Effects of Language Switch on Processing Speed

Processing speed.

The analysis of processing speed included 3529 observations for 90 participants and 10 sentence stimuli. The base model revealed a significant effect of sentence type, such that children were slower to respond to the third segment of sentences when the previous segments were in a different language than the same language (t = 3.60). After controlling for segment duration, children did not exhibit an overall difference in response times to final segments in their dominant vs. non-dominant language (t = 0.27), but they did exhibit a tendency (t = 1.78) toward a larger effect of sentence type for their dominant language than their non-dominant language. Descriptive data are presented in Table 2, and the statistical model is presented in Table 3.

Table 2.

Mean (SD) for processing speed and offline comprehension in the auditory moving window task

	Dominant Language		Non-Dominant Language
	Single-Lang.	Code-Switched	Single-Lang.	Code-Switched
Processing Speeda	1745 ms (329)	1901 ms (360)	1833 ms (355)	1912 ms (383)
Cost (CS-Single)b	152 ms (231)		68 ms (263)
Comprehension Y/N (Prop. Correct)	0.94 (0.10)	0.94 (0.11)	0.90 (0.12)	0.87 (0.12)

^aThese values represent raw RTs, aggregated per subject, in response to the third segments of sentences in each condition. Note that these raw values are not adjusted for differences in segment duration, but segment duration was entered as a covariate in the statistical model.

^bAverage costs, aggregated per subject, calculated by subtracting the RT for each single-language trial from the RT for its code-switched equivalent, thereby accounting for differences in segment durations across languages and sentences

Table 3.

Mixed-effects model for processing speed (base model)

	Processing Speed (ms in log10 scale)
Variable	Estimate	SE	t
Intercept	3.245	0.010	331.38*
Segment Duration	0.351	0.028	12.50*
Sentence Type (Single-Lang vs. CS)	0.029	0.008	3.60*
Language (Dominant vs. Non-Dominant)	0.002	0.007	0.27
Sentence Type X Language	−0.017	0.010	−1.78

^*p < .05

Offline comprehension.

The analysis of the accuracy of children’s responses to the yes/no questions included 3447 observations from 90 participants and 10 sentence stimuli. The base model revealed a significant effect of the language of the question (z = −3.02), such that children were more accurate when responding to questions presented in their dominant language than in their non-dominant language. There was no significant effect of sentence type (z = −0.04) and no significant interaction between sentence type and question language (z = −0.30). Descriptive data are presented in Table 2, and the statistical model is presented in Table 4.

Table 4.

Mixed-effects model for offline comprehension (base model)

	Accuracy (log-odds)
Variable	Estimate	SE	z
Intercept	3.06	0.19	15.72*
Sentence Type (Single-Lang vs. CS)	−0.01	0.36	−0.04
Language of Question (Dom vs. NonDom)	−1.01	0.33	−3.02*
Sentence Type X Language	−0.14	0.47	−0.30

^*p < .05

Cognitive Control and Language Ability as Moderators of Language Switch Effects

Processing speed.

The analysis of cognitive control and language ability as moderators of the sentence type effect on processing speed included 3490 observations from 89 participants (one participant had missing data for language ability) and 10 sentence stimuli. Of the subject-level control variables (age, maternal education), only age accounted for a significant amount of variance in processing speed and was included in the model; older children were faster to respond overall (t = −4.67). The main effect of sentence type persisted (t = 4.03), but there were no main effects of language ability (t = −0.29) or cognitive control (t = 1.13). Crucially, the three-way interaction among language ability, task switching, and sentence type was significant (t = −2.51). Cognitive control had a stronger mitigating effect on the processing costs associated with the language switch as children’s language ability increased. That is, for children with higher language ability, processing costs decreased with higher accuracy on the DCCS. This decrease in processing costs reflected slightly faster responses on code-switched trials, but also slower responses on single-language trials. For children with lower language ability, improved performance on the DCCS was associated with slower responses, especially for code-switched sentences, and slightly increased processing costs. Figure 1 presents a model plot of the three-way interaction. Children with lower language ability also showed a tendency toward smaller processing costs than children with higher language ability (t = 1.85), but none of the two-way interactions were significant. The specifics of the statistical model are presented in Table 5.³

Figure 1.

Model plot showing the three-way interaction among the effects of sentence type, cognitive control, and language ability. Although language ability was analyzed as a continuous variable in the model, for graphing purposes three language groups were created (lowest third [std scores 73–96], middle third [std scores 97–111], and highest third [std scores 112–136]). The grey ribbons show one standard error of the mean within the model predictions.

Table 5.

Mixed-effects model for processing speed with child-level predictors

	Processing Speed (ms in log10 scale)
Variable	Estimate	SE	t
Intercept	3.245	0.010	326.67*
Age (years)	−0.027	0.006	−4.67*
Segment Duration	0.343	0.026	13.15*
Sentence Type (Single-Lang vs. CS)	0.034	0.008	4.03*
Cognitive Control (Mixed Accuracy)a	0.067	0.059	1.13
Language Abilityb	−0.0002	0.0006	−0.29
Type X Cognitive Control	−0.034	0.033	−1.03
Type X Language Ability	0.0006	0.0003	1.85
Cognitive Control X Language Ability	−0.0014	0.0045	−0.30
Type X Cognitive Control X Language Ability	−0.0064	0.003	−2.51*

^*p < .05

^aProportion correct in the mixed phase of the Dimensional Change Card Sort (DCCS) task.

^bCore Language Index Standard Score in stronger language from the Clinical Evaluation of Language Fundamentals (CELF-4 or CELF-4 Spanish)

Offline comprehension.

The analysis of child-level predictors of offline comprehension included 3411 observations from 89 participants and 10 sentence stimuli. There was a significant main effect of language ability (z = 3.91), such that children with higher language ability were more accurate overall. However, the effect of sentence type remained non-significant and was not moderated by language ability, cognitive control, or their interaction (all z < 0.82). Age was included as a control variable to be consistent with the processing speed analysis, but age did not have a significant effect on offline comprehension (z = 1.50). Table 6 shows the full model.

Table 6.

Mixed-effects model for offline comprehension with child-level predictors

	Accuracy (log odds)
Variable	Estimate	SE	z
Intercept	2.90	0.21	13.53*
Age (years)a	0.12	0.08	1.50
Sentence Type	−0.05	0.35	−0.13
Cognitive Control (Mixed Accuracy)b	0.20	0.96	0.21
Language Abilityc	0.04	0.01	3.91*
Type X Cognitive Control	0.49	1.34	0.37
Type X Language Ability	0.01	0.01	0.79
Cognitive Control X Language Ability	−0.01	0.07	−0.17
Type X Cognitive Control X Language Ability	0.08	0.10	0.81

^*p < .05

^aAge was included as a covariate in the model, even though it was not a significant predictor of offline comprehension, in order to be consistent with the processing speed model.

^bProportion correct in the mixed phase of the Dimensional Change Card Sort (DCCS) task.

^cCore Language Index Standard Score in stronger language from the Clinical Evaluation of Language Fundamentals (CELF-4 or CELF-4 Spanish)

Discussion

The current study examined the effects of code-switching on language processing in bilingual children with a broad range of language abilities. Overall, children were slower to process sentences containing a language switch compared to single-language sentences. However, the presence of an intra-sentential code-switch did not affect children’s accuracy in responding to yes/no questions about the content of the sentence. An analysis of child-level predictors revealed that processing costs associated with code-switched sentences were moderated by a combination of cognitive control and language ability. In children with higher language ability, higher accuracy on the DCCS was associated with reduced processing costs, but this relationship decreased in strength and reversed in direction as children’s language skills decreased. Thus, children with the lowest language ability demonstrated an effect of cognitive control in the opposite direction, with slightly increased processing costs associated with higher DCCS accuracy. For offline comprehension, children with higher language ability were more accurate in answering the yes/no questions, but child-level predictors did not moderate the effects of the language switch.

Language Switching Costs in Comprehension

The findings of the current study are consistent with previous studies that have found online processing costs associated with listening to intra-sentential code-switching in adults (Adamou & Shen, 2019; Fricke et al., 2016; Liao & Chan, 2016; Olson, 2016a) and in children (Byers-Heinlein et al., 2017; Potter et al., 2019). However, there were some differences in the pattern of costs identified. Byers-Heinlein, Potter, and colleagues found costs in children only for switches into the non-dominant language. The current study revealed costs across both languages with no significant asymmetry, although there was a tendency toward larger costs for switches into the dominant language. A variety of methodological differences could have accounted for these different findings, including the method of determining dominance, blocked vs. intermixed presentation of stimuli in each switching direction, the outcome measure (proportion of looks vs. reaction times), and the age of the participants. Studies in adults have also varied in whether larger costs are identified for switches into the non-dominant language (e.g., Liao & Chan, 2016) or switches into the dominant language (e.g., Olson, 2016a). Larger costs in the dominant language are often interpreted in terms of top-down inhibitory control mechanisms (e.g., Inhibitory Control Model, Green, 1998). Larger costs for switches into the non-dominant language are often interpreted in terms of bottom-up mechanisms based on linguistic context, frequency, and prediction (e.g., BIA+, Dijkstra & van Heuven, 2002; Potter et al., 2019). However, inferences about the role of inhibitory control based on dominance asymmetries are not straight-forward (e.g., Bobb & Wodniecka, 2013). The current study examined the role of cognitive control through associations with a cognitive control task.

Although intra-sentential code-switching may impose online processing costs, the current study did not identify costs in offline comprehension. This finding is consistent with the study by Peynircioglu and colleagues (2002), who found that children did not make more errors when answering comprehension questions after listening to a mixed-language story compared to a single-language story. The findings of the current study are consistent with the interpretation of Byers-Heinlein and colleagues (2017) that online processing costs as children listen to intra-sentential code-switching may just be a normal feature of bilingual processing rather than a sign of difficulty, as children in the current study were still able to answer comprehension questions about the content of the code-switched sentences. However, it is possible that the yes/no questions presented in the current study did not reflect costs because they were too easy. At the same time, the finding that language ability predicted comprehension accuracy indicates that there was still variability in children’s comprehension performance, and therefore the lack of costs in the comprehension data cannot necessarily be attributed to ceiling effects. Further work with different methods of measuring offline comprehension would be necessary to verify these findings.

Moderators of Language Switching Costs in Comprehension

Neither cognitive control, nor overall language ability served as independent predictors of switching costs in processing speed or offline comprehension as children listened to code-switched sentences in the current study. However, the significant three-way interaction among sentence type, cognitive control, and overall language ability suggested that the effect of cognitive control on children’s online processing costs depended on their level of language ability. For children with higher overall language skills, more accurate performance on the DCCS predicted smaller processing costs. Visual inspection of Figure 1 suggests that high-language children with higher accuracy on the DCCS may have used a proactive (monitoring) control strategy (e.g., Braver, 2012) during the sentence processing task of keeping both languages at a similar level of activation to be better prepared for a language switch (e.g., Gollan & Ferreira, 2009). This proactive control would result in slower processing of single-language sentences but less of a cost when encountering a language switch. In contrast, reactive control would involve adjustments of activation levels only upon encountering the language switch, resulting in slower responses to code-switched sentences than single-language sentences. A similar proactive approach to monitoring sorting cues throughout the DCCS may have helped these children to be more accurate in responding to changes in sorting rules.

Thus, children with more robust language systems exhibited the findings that would be expected if domain-general cognitive control plays a role in auditory processing of intra-sentential code-switching, as suggested by some studies in children (e.g., Byers-Heinlein et al., 2017; Conboy et al., 2011) and adults (e.g., Abutalebi et al., 2007). However, other findings from adults were not consistent with a role for cognitive control (e.g., Blanco-Elorrieta & Pylkkanen, 2016, 2017). One explanation for these divergent results may be the nature of the stimuli. Both the current study and that of Abutalebi and colleagues (2007) created the code-switched stimuli by splicing together single-language recordings produced by bilingual speakers, while Blanco-Elorrieta and Pylkkanen used natural code-switching from spontaneous speech (2017) or single-word stimuli (2016), which may be easier to process without imposing switching costs or recruiting domain-general cognitive control. Further work with children using more naturalistic stimuli would be necessary to verify this interpretation.

A key question raised by the findings of the current study is why the relationship between cognitive control and language switching costs changed with decreasing levels of language ability. One possibility is that children with lower language skills may not have the resources to use the same proactive control strategy during sentence processing employed by their peers with higher language skills. Another possibility is that an additional variable, such as degree of dominance, extent of balanced exposure, or experience with code-switching, may have been driving the observed effects of language ability. If children with higher language ability were highly dominant in one language or less familiar with code-switching, they may have relied more on cognitive control to process this unfamiliar input. However, language ability had only weak to moderate associations with dominance, exposure balance, and parent ratings of exposure to code-switching (see Appendix, Table A.2). Furthermore, including these alternate variables in the model instead of language ability did not yield a significant three-way interaction with cognitive control and sentence type (ts < 1.7). Finally, yet another interpretation stems from the trend toward smaller processing costs in children with lower language. It may be that children with lower language skills found the auditory moving window task too difficult and thus were less sensitive to the presence of the language switch. The significant effect of language ability on overall accuracy on the offline comprehension questions (i.e., lower language skills predicting lower accuracy) further supports the possibility that children with lower language skills were not fully processing or attending to the sentences, even in the single-language condition.

Effects of domain-general cognitive control and the presence of switching costs appear to go hand in hand. Switching costs were present during online processing and there was an effect of cognitive control (at least for children with more robust language). In contrast, there were no switching costs for offline comprehension and there were no effects of cognitive control. Taken together, these findings suggest that switching costs during comprehension may reflect domain-general involvement and slower processing as children process the code-switch, but these costs may not ultimately interfere with accessing the meaning of the sentence. Instead, the ability to comprehend the sentence may be more affected by children’s language skills, regardless of their cognitive control skills and regardless of whether the sentence contains a code-switch.

However, when comparing online processing and offline comprehension, it is important to consider the position of the codeswitch. The current study, like many other studies of intra-sentential language switching, situated the code-switch at the end of the sentence, making it difficult to tease apart online processing of the switched words themselves from sentence wrap-up effects. Further work that varies the position of the code-switch and that incorporates a more challenging measure of offline comprehension would be necessary to further uncover the mechanisms, both within and outside the language system, by which language switching affects online processing and offline comprehension.

Limitations

Studies in adults have identified factors that may help to mitigate costs of processing intra-sentential code-switching during reading and listening, including the naturalness of the code-switching (e.g., Beatty-Martínez & Dussias, 2017; Chan et al., 1983), the linguistic context (e.g., Johns et al., 2019; Olson, 2016a), the presence of acoustic cues that might alert the listener of an upcoming code-switch (e.g., Fricke et al., 2016; Piccinini & Garellek, 2014), and participants’ experience with code-switching (e.g., Beatty-Martínez & Dussias, 2017). The design of the current study did not allow for an examination of all of these factors in children.

To increase experimental control for the comparison between single-language and code-switched sentences, the same ten sentences and the same final segment recordings were used in each condition, removing any acoustic cues that might have signaled the upcoming code-switch. In future work, it will be important to examine whether children are able to use the same acoustic and structural cues as adults to anticipate an upcoming code-switch and facilitate processing. Other aspects of the design, such as breaking the sentences into segments, placing the code-switch at the end of the sentence, and using yes/no questions to assess comprehension may also have impacted the findings and should be considered in future studies.

While the current study did include a parent rating of how often their children heard code-switching from various people in their environment, this measure was not associated with children’s processing of code-switching in the study. More detailed methods of assessing children’s experience with code-switching may be necessary, as parents often expressed uncertainty in their ratings. However, the lack of relationship is consistent with findings by Potter and colleagues (2019) that parent ratings of their own use of intra-sentential code-switching were not associated with toddlers’ processing of mixed-language sentences.

Measures of cognitive control and language ability also warrant further consideration. The current study used only one task (the DCCS) to measure cognitive control. The goal of using the DCCS was to examine broad relationships between domain-general cognitive control and potential processing costs when listening to code-switching, rather than to isolate specific control processes. In future work, multiple measures of cognitive control could be used to extract latent constructs and to examine the effects of more specific processes. To index language ability, the current study used the higher of the Core Language Index scores from the English or Spanish version of the CELF-4. A best-score approach was selected over creating a composite from the two languages because a child with mid-low scores in both languages and a child with a high score in one language and a very low score in the other would achieve a similarly low composite, even though the latter may have stronger underlying language skills than the former. The goal was to use a measure that would reflect the robustness of the child’s language system, regardless of language dominance. However, the English and Spanish versions of the CELF were not co-normed on a bilingual sample, and thus scores in the two languages cannot be considered directly comparable, even though they are on the same standardized scale. Taking the higher core language score also did not allow for mixed dominance, as bilingual children often have some skills that are stronger in one language and other skills that are stronger in the other (e.g., Peña et al., 2014). An assessment that allows children to use either language to obtain a single score or that has co-normed English and Spanish versions would have been ideal, but such an assessment of broad language (and not just vocabulary) was not available for the age range in the current study.

Implications

From a theoretical standpoint, the switching costs identified in the current study suggest at least some overlap between the control mechanisms engaged as children listen to language switches and those engaged in production. Based on associations with children’s cognitive control skills, language control mechanisms in production and in comprehension appear to draw on domain-general cognitive control, but perhaps to different degrees. These findings would be consistent with models such as Green’s (1998) Inhibitory Control Model, but not necessarily with BIA models (e.g., BIA, Dijkstra & van Heuven, 1998; BIA+, Dijkstra & van Heuven, 2002; BIA-d, Grainger et al., 2010) that situate language control mechanisms within the language system. The use of a cooperative method of control during intra-sentential code-switching, as delineated in the Adaptive Control Hypothesis (Green & Abutalebi, 2013) and the Control Process Model of Code-switching (Green & Wei, 2014) would have predicted no switching costs. However, the dense code-switching environment described in these models may not have been evoked in the current paradigm. Further work with a different paradigm and more natural code-switching cues would be needed to test the applicability of these models to the comprehension of code-switching in children. In addition, more detailed information about children’s own patterns of code-switching would be necessary to evaluate the applicability of usage-based models that explain processing costs based on prediction error (e.g., Adamou & Shen, 2019; Fricke et al., 2016).

With regard to implications for children growing up in a bilingual home, the current study suggests that children may process code-switched sentences more slowly than single-language sentences, but this may not ultimately affect their ability to understand the content of a mixed-language sentence. Consistent with Byers-Heinlein (2013) and Byers-Heinlein and colleagues (2017), we interpret these small costs in speed as a typical component of bilingual processing that has rather transitory effects on language comprehension. The novel finding here is that children with high language skills and good cognitive control may be able to exercise strategies to reduce those costs. It is worthy of note that children with lower language skills did not exhibit increased costs when listening to code-switched sentences, suggesting that code-switched input may not pose greater challenges for children with less robust language. However, we are cautious in over-interpreting this finding, given the possibility that the paradigm may have been too difficult for these children.

Conclusion

The findings of the current study suggest that children experience processing costs when listening to sentences containing a language switch. These processing costs may reflect the recruitment of domain-general cognitive control processes to monitor and control language activation levels as the target language changes. However, the involvement of domain-general cognitive control appears to depend on the robustness of the child’s language system. Despite the costs obtained during online processing of code-switching, analyses of offline comprehension suggest that these costs may not have lasting effects on children’s ability to understand the content of the sentences.

Highlights.

• Bilingual children listened to code-switched sentences using self-paced listening

• A Dimensional Change Card Sort task indexed domain-general cognitive control

• Code-switched sentences elicited significant online processing costs

• The role of cognitive control in moderating costs depended on language ability

• Code-switching costs in offline comprehension were not observed

Appendix

Table A.1.

Sentence stimuli and comprehension questions.

Item	Segments 1–2	Segment 3 Single-language	Segment 3 Code-switch	Comprehension Question
1E	The boy | is looking at the elephant |	that is stomping.	que está pisando.	Is the boy looking at the elephant?
1S	El niño | está mirando el elefante |	que está pisando.	that is stomping.	¿Está mirando la familia el elefante?
2E	The boy | talks to the girl |	that is eating.	que está comiendo.	Does the boy talk to the man?
2S	El niño | habla con la niña |	que está comiendo.	that is eating.	¿Habla el niño con la niña?
3E	The mom | sings to the boy |	that is sleeping.	que está durmiendo.	Does the mom sing?
3S	La mamá | le canta al niño |	que está durmiendo.	that is sleeping.	¿Baila la mamá?
4E	The teacher | waved to the children |	that were walking.	que estaban caminando.	Did the teacher wave to the parents?
4S	El maestro | saludó a los niños |	que estaban caminando.	that were walking.	¿Saludó el maestro a los niños?
5E	The girl | watched a movie |	while she was eating.	Mientras que estaba comiendo.	Did the girl watch a movie?
5S	La niña | vio una película |	Mientras que estaba comiendo.	while she was eating.	¿Vio el niño una película?
6E	The dog | chased the ball |	and ran in circles.	y corrió en círculos.	Did the dog chase the car?
6S	El perro | persiguió la pelota |	Y corrió en círculos.	and ran in circles.	¿Persiguió el perro la pelota?
7E	The girl | sits with her friends |	and talks.	y habla.	Does the girl sit with her friends?
7S	La niña | se sienta con sus amigas |	y habla.	and talks.	¿Se sienta la niña con sus los adultos?
8E	The farmer | raises cows |	and drinks milk.	y toma leche.	Does the farmer raise chickens?
8S	El granjero | cría vacas |	y toma leche.	and drinks milk.	¿Cría el granjero vacas?
9E	The girl | likes to dance |	with her friends.	con sus amigas.	Does she like to dance?
9S	A ella | le gusta bailar |	con sus amigas.	with her friends.	¿Le gusta a ella cantar?
10E	The teacher | gives the students homework |	to complete.	para completar.	Does the teacher give the students prizes?
10S	La maestra | da tarea a los estudiantes |	para completar.	to complete.	¿Da la maestra tarea a los estudiantes?

Note: Segment divisions are marked with a “|”. The item column lists the 10 different sentences. E = English, S = Spanish, indicating the language of the first two segments and the comprehension question. The language of the third segment was manipulated to create the single-language and code-switched conditions. The effect of the language switch on processing speed was examined by analyzing the reaction time of the button press in response to the third segment in each condition (i.e., E-E single-language vs. S-E code-switch; S-S single-language vs. E-S code-switch). The effect of the language switch on offline comprehension was examined by analyzing the accuracy of the responses to the yes/no questions (e.g., Spanish question following Spanish single-language sentence vs. Spanish question following S-E code-switched sentence).

Table A.2.

First-order correlations among child variables, controlling for age.

	LangAbility	CogControl	Dominance	ExpBalance	CSExposure
Language Abilitya	1.0
Cognitive Controlb	0.42*	1.0
Dominancec	0.43*	0.11	1.0
Exposure Balanced	−0.24	−0.14	−0.25	1.0
CS Exposuree	−0.24	−0.14	0.04	0.36*	1.0

^*significant after correcting for multiple tests

^aCore Language Index Standard Score in stronger language from the Clinical Evaluation of Language Fundamentals (CELF-4 or CELF-4 Spanish)

^bProportion correct in the mixed phase of the Dimensional Change Card Sort (DCCS) task

^cAbsolute value of difference between Core Language scores for CELF-4 English and CELF-4 Spanish, indexing extent of dominance in one language

^dProportion of waking hours exposed to less-heard language (0-.50), where higher values reflect more balanced exposure

^eAverage exposure to code-switching from parents, siblings, grandparents, other relatives, friends, classmates, adults at school, and unfamiliar people in the community, based on parent ratings on a 0–10 frequency scale