Does code-switching influence novel word learning?

Abstract

Code-switching occurs regularly in the input to bilingual children. Yet, the effect of code-switched input on language development is unclear. To test whether word learning would be affected by code-switching, Spanish-English bilingual children (N=45, 19 boys, Mean_Age=5.05 years; Ethnicity: 37 Hispanic/Latino, 6 Non-Hispanic/Latino, 2 Unreported) were taught English-like novel words in two conditions. In the English-only condition, definitions for novel words were provided entirely in English. In the Code-Switch condition, definitions for novel words were provided in English and Spanish, incorporating code-switches. Children required fewer exposures to retain novel words in the Code-Switch than the English-only condition and this effect was not moderated by children’s language ability or exposure to code-switching, suggesting that code-switched input does not pose word-learning risks to bilingual children, including children with lower levels of language ability.

Code-switching (alternation between languages) is a common practice among bilinguals, and code-switching both within and across sentences occurs regularly in the input to young bilingual children (e.g., Bail, Morini, & Newman, 2015; Byers-Heinlein, 2013; Nicoladis & Secco, 2000; Tare & Gelman, 2011). Yet, laboratory studies of code-switches suggest that comprehending information that involves code-switches can be more cognitively taxing for children than engaging in single-language comprehension (e.g., Byers-Heinlein, Morin-Lessard, & Lew-Williams, 2017; Gross et al., 2019; Morini & Newman, 2019; Potter et al., 2019). If these costs flow downstream, presence of code-switches in the input may negatively affect bilingual children’s language outcomes. A potent counterargument to this theorizing is provided by a growing literature indicating that naturalistic code-switching behaviors that conform to community norms may carry little processing cost (e.g., Adamou & Shen, 2019; Blanco-Elorrieta & Pylkkänen, 2017). These findings form a foundation for an alternative hypothesis, namely, that bilingual children, at least those who routinely experience code-switching in their environment, would be unaffected by such input.

Crucially, there has been very little empirical work attempting to examine whether exposure to code-switching is associated with children’s language outcomes, and the findings have been remarkably inconclusive. One study reported an inverse relationship between parental self-ratings of language mixing and bilingual children’s vocabulary performance (Byers-Heinlein, 2013), one study reported a positive relationship between parents’ within-sentence code-switching and a parent-report-based measure of bilingual children’s vocabulary (Bail et al., 2015), and two studies reported an absence of a relationship between parents’ language mixing behaviors (documented in language diaries) and children’s language skills (Hoff et al., 2012; Place & Hoff, 2011). To complicate matters, Kaushanskaya and Crespo (2019) reported an inverse relationship between exposure to code-switching and language outcomes in children with weaker working memory skills, but a positive relationship in children with stronger working memory skills.

Although correlational studies linking exposure to code-switching to children’s language skills are useful in shedding light on the possible relationship between the two, any interpretation of such a relationship, if observed, must remain non-directional. Furthermore, parent reports of code-switching exposure are yet to be fully verified for validity and reliability, and objective counts of parental code-switches necessarily provide only a keyhole view of the child’s overall exposure to code-switching in the home and the community. Careful experimental manipulations of code-switching exposure are necessary to begin pinpointing causal linkages between exposure to code-switching and language outcomes in bilingual children. In the current study, we did precisely this, testing whether code-switched input would influence word-learning performance in bilingual children. In focusing on word-learning, we were cognizant of the possibility that learning performance may differ from processing performance in its sensitivity to input manipulations.

One long-standing finding in the study of learning is that variability is a desirable property when the goal is for the learner to home in on the relevant dimensions of the to-be-learned information (Estes & Burke, 1953; Munsinger & Kessen, 1966; Posner & Keele, 1968). For instance, variability in vocal affect has been shown to enhance infants’ learning and generalization of newly-acquired words (e.g., Singh, 2008). Similarly, exposure to multiple exemplars during category learning facilitates children’s ability to extend the newly-learned words to untaught exemplars and to retain the novel word-referent mappings over time (e.g., Twomey et al., 2014). Exemplar variability also accelerates children’s vocabulary development outside the lab (Perry et al., 2010), and can enhance word learning (e.g., Aguilar, Plante, & Sandoval, 2018) and morphosyntactic treatment outcomes (Plante et al., 2014) in children with clinically significant weaknesses in language. One explanation for why variability may be helpful to word learning is that learners (especially children, Vlach and Sandhofer, 2011) encode all cues to the novel word, including those that are non-contrastive. Variability in the non-contrastive cues can promote learning of the contrastive (target) information. In the context of bilingual word learning, could code-switched input act to promote word learning because variability in the linguistic context would enable children to home in on the word itself?

While contextual variability has received relatively less attention than exemplar variability in shaping learning outcomes, classic work on the learning of nonadjacent dependencies has demonstrated that increasing the variability of intervening elements (effectively, increasing variability of surrounding linguistic context) serves to improve learning of such dependencies in both adults and children (Gomez, 2002). Similarly, learning and generalization of new phonotactic patterns is enhanced under conditions of increased contextual variability (e.g., Denby et al., 2018). Relying on learning theory, and on studies indicating benefits of contextual variability for learning, it is logical to predict that for bilingual children, hearing novel words in language frames that are different from the language of the target novel word (i.e., in code-switched contexts) can serve to highlight the invariant structure of the novel words.

In the present study we tested the effect of code-switched input on bilingual children’s word learning with three hypotheses in mind. Our first hypothesis was grounded in the literature indicating processing costs associated with code-switched input. We hypothesized that if processing costs associated with code-switched input have negative downstream consequences for learning, then bilingual children should be less successful learning novel words in code-switched contexts than in single-language contexts. Our second hypothesis was grounded in learning and variability literature indicating that variable context at encoding enhances learning outcomes. We hypothesized that if code-switching in the input enhances contextual variability, then bilingual children may be more successful learning novel words in code-switched contexts than in single-language contexts. Our third, null hypothesis was that children would demonstrate similar learning patterns in single-language vs. code-switched contexts. Such a finding was possible if, in line with recent studies, more ecologically-valid instances of code-switching (as implemented here) carry minimal processing costs. Beyond these main hypotheses, we also examined whether the effect of code-switched input on word-learning (if observed) would be moderated by children’s language skills and their experience with code-switched input.

Method

Participants

Fifty Spanish-English bilingual participants (20 boys) ages 4-5 were recruited. One participant withdrew from the study and four participants were excluded for failing the hearing screening (n=2) and for demonstrating very low levels of Spanish proficiency (n=2). A total of forty-five Spanish-English bilingual participants (19 boys) were included in the analyses (Mean_Age = 5.05 years; SD_Age = 0.57). Ethnicity of the children was reported as 37 Hispanic/Latino, 6 Not Hispanic/Latino, 2 Unreported. Race was reported as 24 White, 2 American Indian, 2 Black, 1 Asian, and 16 Unreported. Seven children were identified as having clinically-low language skills based on the following criteria: standard scores below 85 on the Bilingual English Spanish Assessment (BESA) Index Composite (n=5), and standard scores below 85 in either the morphosyntax or semantics subtests in both languages (n=2). However, the sample as a whole was characterized by highly variable language skills. Participants were exposed to English and Spanish by 48 months of age, with the majority of participants (40/45) exposed to both languages by 36 months. Inclusionary criteria included normal or corrected vision, normal hearing per parent report and passing a hearing screening. Exclusionary criteria included a history of abnormal hearing and neurodevelopmental medical diagnoses.

Procedure

Parent Interview and Questionnaires.

The Home survey from the Bilingual English Spanish Assessment (BESA) (Peña, Gutiérrez-Clellen, Iglesias, Goldstein, & Bedore, 2014) Bilingual Input-Output Survey (BIOS) was administered to parents to determine Spanish and English use at home. The BIOS is a self-report measure that presents parents with a timetable split into half-hour intervals, representing a typical weekday, a typical Saturday, and a typical Sunday. The timetable spans hours between 6am and 11pm, and allows parents to designate half-hour intervals when children are awake, and the language children are exposed to (input) and produce (output) during each half-hour interval – English, Spanish, or Both. Weekly totals for language exposure (English vs. Spanish) are aggregated and translated into proportions. For our study, we focused on the input measure, since the experimental task was receptive in nature.

Information about bilingual participants’ language dominance and language preference was collected via parent interviews. A measure of children’s exposure to code-switching were also collected during parent interviews. Specifically, parents were asked to assess how often family members living with the child at home code-switched around their child using a scale from 0 (never) to 10 (always). Primary caregivers’ socioeconomic status was operationalized as level of education. See Table 1 for participant characteristics.

Table 1:

Participant Characteristics with Means (SD; Ranges)

	Whole Sample	Children who met learning criteria in at least 1 condition
N	45 (19 boys)	34 (14 boys)
Age	5.05 (0.57; 4.00-5.90)	5.14 (0.54; 4.1-5.90)
Nonverbal IQ a	102.04 (11.53; 85-135)	102.97 (12.24; 85-135)
Mother’s Years of Education	14.64 (4.68; 6-24)	15.09 (5.00; 6-24)
English Age of Acquisition (months)	10.67 (16.78; 0-48)	7.76 (14.90; 0-48)
Spanish Age of Acquisition (months)	4.22 (10.20; 0-46)	5.21 (11.44; 0-46)
Overall Language Ability b	103.22 (12.06; 74-119)	107.18 (9.02; 84-119)
Exposure to Code-Switching c	3.65 (2.30; 0-9.33)	3.42 (2.05; 0-9)
Total Input Spanish d	0.54 (0.18; 0.19-0.90) e	0.65 (0.41; 0.19-0.79)
Total Input English d	0.46 (0.18; 0.10-0.81) e	0.35 (0.41; 0.21-0.81)
Dominant Language (n)
English	24	19
Spanish	21	15

^aVisual Matrices subtest of the Kaufman Brief Intelligence Test (standard score)

^bBilingual English Spanish Assessment (BESA) Index Composite (standard score)

^cParent report of how often family members living with the child at home code-switched around their child on a scale from 0 (never) to 10 (always)

^dBilingual Input-Output Survey of the Bilingual English Spanish Assessment (BESA)

^en= 44; 1 datapoint was omitted because of an error on the BIOS

Standardized Measures.

The Bilingual English Spanish Assessment (BESA) was administered to measure children’s overall language abilities in both English and Spanish. The BESA includes semantics and morphosyntax subtests for English and Spanish, focusing on items that reflect language-specific markers, structure, and cultural references. The BESA Index Composite score that combines the best score in each domain across languages was calculated. The Visual Matrices subtest of the Kaufman Brief Intelligence Test (KBIT-2) was used to assess each participant’s non-verbal intelligence.

Word-Learning Task

Stimuli.

All stimuli were presented on a 23” Dell UltraSharp U2312HM desktop monitor with a 1920 x 1080 x 60 Hz resolution. Twelve 1-syllable English-like nonwords and twelve black and white novel objects matched on semantic set size (Storkel & Adlof, 2009) were used. The novel words had a CVC structure to resemble common, early-acquired English words. The novel words were not real words in English or Spanish and contained only phonemes realized in both Spanish and English. We targeted shared phonemes so that reduced familiarity with English-specific phonemes could not impact children’s performance. However, the novel words were realized as phonetically-English, with all consonants and vowels pronounced with English-specific phonetic detail. Bilingual research assistants in the lab uniformly identified the novel words as English-like rather than Spanish-like. Each word was paired with an object and was assigned to one of three semantic categories (i.e., foods, tools, animals). The 12 novel word-referent pairings were split across two lists of 6 words-referents each (see Supplementary Appendix). The two lists of novel words matched on developmental sound class of the initial and final consonants, English and Spanish average biphone probability, and English and Spanish neighborhood density. To reduce interference during learning, no two words within the same list started with the same consonant. The novel words were yoked to pictures and definitions, but the assignment of list (A or B) to condition (English-only vs. Code-Switch) was counterbalanced across participants.

Conditions.

In the English-only condition, participants heard novel English-like words embedded in English sentences. In the Code-Switch condition, participants were exposed to novel English-like words embedded in alternating English and Spanish sentences. Language switching occurred both within a sentence (intrasentential code-switching; El nem…) and between sentences (intersentential code-switching; Look at this nem. El nem es un tipo de comida. El nem es muy crujiente. You have to wash a nem to eat it). The structural aspects of the Code-Switch condition were carefully considered, and the particular phrasal structure of the intrasentential switch (Spanish determiner-English novel word) was chosen based on the available corpus data (Pfaff, 1979; Timm, 1975) which indicate that the vast majority of switches (70%-99%) produced by Spanish-English bilingual adults on the determiner phrase involve the Spanish determiner and the English noun (e.g., el car). Such single-word insertions are also the most frequently documented types of intra-sentential (within-sentence) code-switching observed in parental speech to bilingual children (Bail, Morini, & Newman, 2015).

Participants were exposed to a list of 6 word-object pairs in an English-only condition and a different list of 6 word-object pairs in a Code-Switch condition. Participants completed each condition in separate sessions on different days, scheduled at least three days apart. Order of presentation (English only vs. Code-Switch) and lists were counterbalanced. The presentation of stimuli in each condition consisted of cycles of exposure and testing that continued until the child reached learning criteria or until the child maxed out (10 cycles without reaching the learning criterion).

Exposure Phase.

During the exposure phase, participants were instructed to look at pictures of novel objects and learn their names. In each trial, one novel object appeared in the center of the screen and was labeled four times. Each production of the target label was embedded in a descriptive sentence. The first sentence labeled the object (Look at this nem!); the second sentence identified the sematic category (A nem is a type of food); and sentences three and four described the object (A nem is very crunchy. You have to wash a nem to eat it).

Testing Phase.

After presentation of all six novel items, children’s memory for the novel words was tested. Each novel word was tested once. In each test trial, participants were required to select the target object in a 3-Alternative Forced-Choice display. A response was required to progress to the next trial. Participants completed 1-10 cycles of exposure and testing in each condition. In Cycle 1, participants did not receive feedback during the testing phase. This enabled us to test short-term retention of novel words immediately after learning, akin to most word-learning studies. In Cycles 2-10, participants received feedback in each testing trial. The learning-to-criterion approach was implemented to assess children’s ultimate learning of novel words. Feedback consisted of a smiley face when the correct object was selected and a frowning face when the incorrect object was selected. The task terminated when learning criteria were met, or when 10 cycles were completed without reaching the learning criterion. Learning criteria were defined as a score of 6 out of 6 in a single cycle (i.e., all 6 novel objects were accurately identified) or a score of 5 out of 6 in any two cycles.

Analyses

All data and scripts have been uploaded to Open Science Framework (https://osf.io/xdpr2/). Two sets of analyses were implemented. First, logistic mixed effect models were constructed in R, version 3.2.2 using the lme4 package to analyze item-level dichotomous accuracy data (0, 1) in Cycle 1 (completed without feedback). Each model examined whether predictors increased or decreased the likelihood (log-odds) of making a correct response during this first cycle of testing. The fixed effect of Condition (contrast-coded) was included in all models, and fixed effects of overall language ability (i.e., BESA Index Composite scores) and exposure to code-switching, and their interaction with Condition, were estimated in separate models. Models with full random-effect structures yielded singularity issues, suggesting that models were overfitted for the data. Model comparisons revealed that inclusion of a by-subject random slope for the effect of condition, and by-item random slope for the effect of condition, language skills, and their interaction, did not significantly account for any variance. Therefore, logistic mixed effect models only included by-subject and by-item random intercepts.

Second, the maximum number of cycles required to reach learning criteria was modeled using linear mixed effects models. Models were constructed in the same way as in the item-level analyses. By-subject random intercepts were included in each model. Because each participant contributed only one observation per condition (i.e., number of cycles), there was insufficient data to estimate by-subject random slopes for the effect of condition.

Nonverbal IQ, mother’s years of education, and age of first exposure to English were entered as covariates in the preliminary analyses. Nonverbal IQ and age of first exposure to English were retained as covariates but mother’s years of education was omitted, because it did not improve model fit and because it significantly correlated with children’s nonverbal IQ (r = .35, t(43)= 2.43, p = .02). For all analyses, effects with a t-value or z-value greater than 1.96 were considered significant (p < 0.05).

Results

Cycle 1 analyses

Three children abandoned task in the English-only condition and three different children abandoned task in the Code-Switch condition. As the result, forty-two participants contributed Cycle 1 data in the English-only condition and forty-two participants contributed Cycle 1 data in the Code-Switch condition (note that there is not a complete overlap between participants who contributed data in the two conditions). Cycle 1 performance in the English-only condition (M = 52%, SD = 50%; Range: 17%-100%; t(41) = 6.10, p < .0001) as well as in the Code-Switch condition (M_Accuracy = 50%, SD_Accuracy = 45%; Range: 0-100%; t(41) = 4.28, p < .0001) was significantly above chance (i.e., 33% , given the three-forced choice task).

A logistic mixed effects model examining the effects of Condition and overall language ability on Cycle 1 accuracy and controlling for nonverbal IQ and age of first exposure to English, included 504 observations and 45 participants. The effect of Condition was not significant (z = −0.69, p =.49). However, results revealed a significant main effect of language skills, (B = 0.27, SE = 0.13, z = 2.04, p = .04) such that children with more robust language skills were significantly more likely to identify target words than children with lower language skills (OR = 1.31 95% CI = 1.01 – 1.69).

A logistic mixed effects model examining the effects of condition and exposure to code-switching in the home, and controlling for nonverbal IQ and age of first exposure to English, on Cycle 1 accuracy included 480 observations and 43 participants. A significant main effect of exposure to code-switching was revealed (B = − 0.25, SE = 0.11, z = − 2.17, p = .03) such that, overall, children with greater exposure to code-switching in the home were less likely to be accurate across conditions (OR = 0.78, 95% CI = 0.62 – 0.98). Once again, the effect of Condition was not significant (z = − 0.30, p = .76). See Supplementary Materials Table 1 for full logistic mixed effects model results.

Cycles to criterion analyses

Thirty-four participants met learning criteria in at least one condition; twenty-nine participants met criteria in the English-only condition (85%) and twenty-seven participants met criteria in the Code-Switch condition (79%). Comparisons between children who met and did not meet learning criteria are presented in Supplementary Table 3 (English-only condition) and Supplementary Table 4 (Code-switch condition). Across both conditions, children who met learning criteria had significantly stronger language skills than children who did not. In addition, children who met criteria in the English-only condition were characterized by younger English age of acquisition than children who did not. No other differences between children who met and did not meet learning criteria were observed.

A linear mixed effects model examining the effects of Condition and overall language ability on number of Cycles to criterion, and controlling for nonverbal IQ and age of first exposure to English, included 56 observations and 34 participants. A significant main effect of Condition was observed (B = −0.99, SE = 0.48, t = − 2.06, p = .04), such that children required fewer cycles to meet learning criteria in the Code-Switch condition (M_{No.of Cycles} = 4.18, SD_{No.of Cycles} = 0.15) than in the English-only condition (M_{No.of Cycles} = 5.03, SD_{No.of Cycles} = 2.29) (Figure 1). A significant main effect of overall language ability (B = − 0.86, SE = 0.32, t = −2.73, p = .04) was also observed, such that children with more robust overall language skills required significantly fewer cycles to meet learning criteria. Notably, when analyses focusing on Cycle 1 accuracy were rerun on the subsample of children who met learning criteria, the original findings were confirmed, with the results once again yielding a non-significant main effect of Condition.

Figure 1. Cycles to Criterion by Condition.

Note. Dots note maximum number of cycles required for participants to meet learning criteria in each condition. Lines show performance across conditions for an individual participant.

A linear mixed effects model examining the effects of Condition and the effect of exposure to code-switching in the home on number of Cycles to criterion and controlling for nonverbal IQ and age of first exposure to English, included 55 observations and 33 participants. A significant main effect of Condition was observed (B = − 0.92, SE = 0.44, t = − 2.11, p = .04), such that children required fewer cycles to meet learning criteria in the Code-Switch condition than in the English-only condition. A main effect of code-switching exposure (B = 0.84, SE = 0.33, z = 2.57, p = .02) was observed, such that children who were exposed to more code-switching required a greater number of cycles to meet learning criteria than children with less exposure.

In both models, the interactions with condition were not significant, ps > .05. See Supplementary Materials Table 2 for full linear mixed effects model results.

Discussion

The goal of the present study was to examine whether code-switching in the input may influence word-learning performance in bilingual children. We observed a null effect of code-switching on word-learning performance when word retention was indexed by children’s accuracy on the recognition measure administered immediately after the children were exposed to all the novel words for the first time. We observed a positive effect of code-switching on word-learning performance when retention was indexed by the number of learning-testing cycles required to reach mastery when recognizing novel words. The effect of code-switching on word-learning was not moderated by either language skills or exposure to code-switching.

A unique feature of our experiment was inclusion of two recognition testing measures: a measure that indexed short-term retention of newly-learned words and a measure that indexed a longer-range process that indexed ultimate competence with newly-learned words. The inclusion of both measures enabled us to test whether the effect of code-switching in the input would influence word-learning performance when learning was taken to its highest level (per individual child), or whether the effect of code-switching (if observed) would be short-lived.

The immediate recognition measure we implemented resembles measures most frequently implemented in psycholinguistic word-learning studies to index retention immediately after learning. Our findings indicate that the effect of code-switching exposure had no impact on this measure. Processing studies indicate that code-switches that are implemented in the lab carry few processing costs when these switches closely align with code-switches that are common in the bilingual community (Adamou & Shen, 2019; Blanco-Elorrieta & Pylkkänen, 2017). Our findings are consistent with this body of literature. Our finding that code-switching in the input enhances word-learning performance when indexed by learning-to-criterion indicates that repeated exposures to code-switching during a learning task can facilitate novel word learning. We predicted that should code-switched input function similarly to variable context for the novel words, it may help learners to isolate the novel words and to retain them. The presence of Spanish in the input, and especially the combination of the Spanish determiner with the novel word, may have enhanced children’s awareness of the novel word-form, and facilitated its retention over time. There are a few points of caution regarding this interpretation that are important to note. First, some children were not able to reach criterion at all, and therefore, analyses of learning-to-criterion included only a sub-sample of participants, those with relatively higher language skills. This is not surprising, given the high criterion that was set for learning and the wide spread of language skills in the sample. Even with the reduced numbers, the sample size remained quite robust (n=27 in the code-switch condition); however, it will be important for future studies to consider relaxing the learning criteria and to confirm our findings. At this time, we can only say that ultimate success with learning of novel words may be enhanced by code-switched context in children with higher language skills.

Second, while the code-switched stimuli included both inter- and intra-sentential switches, prior literature suggests that inter-sentential switches carry little processing cost (Byers-Heinlein, Morin-Lessard, & Lew-Williams, 2017), and therefore, it is the intra-sentential switches that would be hypothesized to cause processing difficulty. Intra-sentential switches in our study were switches at a single-word level, between the Spanish determiner (el) and the English-like novel noun (nem), and it is possible that children did not process these as switches, but instead, treated the novel words as Spanish cognates of the English novel words. This interpretation is unlikely, given the English-specific phonetic realization of the novel words. However, two sets of follow-up studies would be important to test for this possibility: One, where the novel words contain English-specific phonemes, and two, where the intra-sentential switches are presented on familiar words preceding or following the target novel word.

We observed a general effect of language skills on word-learning performance, but it was not the case that weaker language skills were associated with a particular difficulty learning novel words in the code-switched condition. Such an interaction would be hypothesized if code-switched input were more cognitively taxing. However, in the present study, code-switched input did not prove to be cognitively taxing, and therefore, an absence of an interaction between condition and children’s language skills is entirely logical. Alternatively, it is possible that fluctuations in language skills were not sensitive enough to effectively moderate the effect of code-switched input on learning. In prior non-experimental work, Kaushanskaya and Crespo (2019) found that verbal working memory skills moderated the association between exposure to code-switched input and children’s language skills, and it is possible that verbal working memory tasks would better capture fine-grained variability in processing capacity than an omnibus language measure used in the present study. Future studies will need to test the role of working memory skills (in addition to language skills) in children’s learning from code-switched input.

Paradoxically, while the experimental manipulation of code-switched context in the present study was associated with more successful word-learning outcomes (vs. single-language context), the parent report measure of code-switched exposure was associated with overall lower word-learning performance. The finding that children’s exposure to code-switched input in the home was associated with less successful word-learning is in line with prior work (Byers-Heinlein et al., 2013), but critically, it was not the case that children who experienced more code-switched input in the home were less affected by code-switched input when learning novel words. Our not finding an interaction between condition and children’s exposure to code-switching is likely explained by the broad absence of costs associated with code-switched input in this study. However, it is difficult to reconcile the two distinct patterns of results associated with the two code-switching constructs – the experimental manipulation and the parent report. As always when using a self-reported measure of exposure to code-switched input (especially a single measure, as in the present study), we must caution that such self-reported measures may or may not be related to actual levels of exposure to code-switched input. We therefore would eschew over-interpreting this finding, and reiterate that empirical manipulations of code-switched input and its effect on learning and processing are more likely to provide a reliable insight into the possible effect of code-switched input on language outcomes.

In conclusion, the findings of the present study indicate that code-switched input, at least as manipulated here, does not carry costs to word-learning performance, at least as instantiated in the present study. In fact, repeated exposure to code-switched input may enhance word-learning outcomes over time, especially in children with stronger language skills. Our findings highlight the importance of including learning paradigms when building a theoretical framework for code-switching processes. A crucial practical finding is that children with lower levels of language ability learn just as well from code-switched input as from single-language input. These findings should reassure parents, educators, and clinicians who may be concerned about the possibility that code-switched input may be non-optimal for children with language difficulties. Instead, our findings suggest that code-switched input may in fact be optimal for bilingual children, and it will be the goal of future studies to understand why this may be the case, and whether it is the case for children from other bilingual communities.

Research Highlights.

• We tested the effect of code-switched input on Spanish-English bilingual children’s word learning.

• Immediate recognition testing revealed comparable learning in the English-only and code-switched condition, while learning-to-criterion testing revealed faster learning in the code-switched than the English-only condition.

• These effects were not moderated by children’s language ability or code-switching exposure.

• The findings indicate that code-switched input does not pose word-learning risks to bilingual children, including children with lower levels of language ability.

Data availability statement:

All raw data, analyses, and scripts can be accessed via Open Science Framework via https://osf.io/xdpr2/