Research demonstrates that the linguistic system is not a static unit but rather a malleable and adaptive organization that integrates novel entries at different levels (de Bot et al., 2007; Kroll et al., 2014; Kaan et al., 2019) and the context of sentence processing provides a broad but complex ground to investigate how various linguistic representations interact in the bilingual mind. In the present study, we present a systematic review and meta-analysis study to explore the current evidence on how the bilingual experience changes the processing of sentences in the native language to unravel different patterns and the factors that underlie them.

The search was conducted in Web of Science (WOS), PsycINFO, PubMed, and Scopus from database inception to March 3, 2021, and the strategy comprised keywords and text words related to bilingualism, attrition, and sentence processing and comprehension. We first piloted the search strategy in WOS and then adapted it to run across PsycINFO, PubMed, and Scopus (see Supplementary Material). We also screened reference lists of included studies and previous reviews on the topic. There were no language or year restrictions.

Following PICO criteria (population, intervention, comparator, and outcome), our inclusion criteria were as follows: the population was bilinguals, here broadly defined as participants tested in their native tongue but proficient in a second language; individuals suffering from any linguistic deficit, children, and heritage speakers were excluded; regarding intervention or exposures, we included cross-linguistic influence from the second language (L2) to the first language (L1) in morphosyntactic processing, and the comparator was morphosyntactic native language processing in bilingual participants contrasted to monolinguals’ (participants with minimal to no experience with a language different from their native tongue) in both sentence comprehension and sentence production assessed by behavioral measures. Finally, the outcome of this study was the influence of the L2 on the L1, as seen through different psycholinguistic tasks and behavioral measures.

Two independent reviewers screened the title and abstract (LG and IGG) and full text to assess the eligibility of the studies. A third reviewer made the final decision (PR) in case of disagreements. The software system used for recording decisions was Rayyan (Ouzzani et al., 2015).

Again, two independent reviewers (PR and IGG) extracted the data and resolved discrepancies by consensus. We collected information related to the first author, publication year, total sample size, bilingual sample size, monolingual sample size, target population (pair of languages and language experience), cognitive processes studied (comprehension and production), task, and measures. More specifically, behavioral measures such as reaction times, accuracy, acceptability ratings in comprehension and pauses, errors, diversity, and complexity in production were included. Seven articles were discarded because of unavailable data, and information from groups of heritage speakers or children was not included in the analyses.

Quantitative data were analyzed using the Comprehensive Meta-Analysis (CMA) software package, V.3 (CMA) (Borenstein et al., 2013) and Stata release 14.2. (StataCorp, 2015). First, standardized mean differences (SMDs) were calculated from each group’s mean and SD, and when unavailable, we calculated SMDs from sample size and p-value. Then, CMA was used to obtain the equivalent SMDs. Finally, the pooled SMDs for all studies and their 95% CIs were estimated. Some studies only reported p-values corresponding to interactions or group effects that included heritage speakers groups and were discarded from the analyses.

The SMD between bilingual and monolingual participants was used as effect size using Hedges’ g formula (Hedges, 1981) for small sample sizes. A negative value indicated higher scores in morphosyntactic tasks for the monolingual group in sentence comprehension and sentence production. Following previous evidence, the sign of the SMD was inverted for reaction time, errors, and pause measures so that the SMD went in the same direction as other measures such as accuracy in which the higher the mean, the better the performance (e.g., Beaussart et al., 2018). Interpretation of the resulting SMD followed Cohen’s proposal: 0.20 as a small effect size, 0.50 as a medium effect size and, 0.80 as a large effect size (Cohen, 1988). In addition, random-effects models were used for pooling effect sizes assuming studies included heterogeneous populations that may differ from each other.

Inspection of heterogeneity was carried out through visual inspection of the forest plot, Q Cochran’s statistic, and p-value. Additionally, heterogeneity was quantified by the I² index and its 95% CI, and interpretation of the I² index was subject to the following level and percentages: unimportant heterogeneity (0–40%), moderate heterogeneity (30–60%), substantial heterogeneity (50–90%), and considerable heterogeneity (75–100%) (Higgins and Green, 2011; Borenstein et al., 2017).

We employed the Duval and Tweedie trim-and-fill procedure (Duval and Tweedie, 2000), the Begg and Mazumdar rank correlation (Begg and Mazumdar, 1994), and the Egger test (Stuck et al., 1998) to explore publication bias.

Furthermore, sensitivity analyses were conducted to explore SMD when Cohen’s d and the fixed-effects model were used. Finally, possible SMD variations were examined independently for studies focused on sentence comprehension, sentence production and for those exploring syntactic and morphological processing.

Finally, subgroup analyses were performed using a mixed-effects model for the categorical moderators AoA (early or late bilinguals), immersion context (either L1 or L2), length of residence (LoR) (short, long or no immersion), task modality (visual, auditorily and audiovisual), and structure congruence (similar and dissimilar across languages).

There was more than one effect size for all studies, and, therefore, the data might be considered dependent. Because of such dependence, we performed a meta-regression analysis on the continuous data with robust variance estimates (RVE; Hedges et al., 2010; Tanner-Smith and Tipton, 2014) using the Robumeta command in Stata. When fewer than four degrees of freedom were present, the results were considered unreliable (Tanner-Smith et al., 2016). Following previous research, covariates included in the RVE meta-regression were LoR, AoA, and L2 proficiency (e.g., Schmid and Dusseldorp, 2010). We performed both bivariate meta-regression analysis (including only one of the three independent variables in each meta-regression analysis) and multivariate meta-regression analysis (including all covariates in the same meta-regression analysis).

A total of 817 articles were identified through searching databases and other sources. After eliminating duplicates, 521 published studies and unpublished doctoral dissertations remained, and we screened the title and abstract. Fifty-seven studies initially met the inclusion criteria and were subject to a full-text inspection. This procedure yielded 16 articles that were included in the systematic review (see Figure 1 for details about the exclusion criteria). We finally included 14 studies with 14 independent group comparisons and 81 effect sizes for the meta-analysis calculations. Two studies (Castro et al., 2017; Dragoy et al., 2019) were excluded from the meta-analysis because they did not provide the needed statistics to calculate SMD.

A total of 1,044 participants were included across 16 studies in the systematic review (see Table 1 for study characteristics). Of them, 486 were bilinguals, and 412 were monolinguals. Ten studies included late bilinguals (studies IDs: 2, 5, 8, 9, 10, 11, 12, 14, 15, and 16), three studies included only early bilinguals (studies 3, 7, and 13), another two studies included both late, and early bilinguals (studies 1 and 6) and one study did not provide information regarding the age of acquisition (study 4). Regarding the pair of languages under scrutiny, two studies investigated Spanish–Swedish bilinguals (studies 2 and 3). The following pairs were investigated in one study each: Turkish–German (study 1), English–Spanish (study 12), Spanish–English (study 10), Italian–English (study 16), German–English (study 14), Brazilian Portuguese–European Portuguese (study 4), Chinese–Korean (study 5), Russian–German (study 6), Turkish–English (study 7), Greek–Swedish (study 8), Russian–Hebrew (study 9), German–Dutch (study 11), German–Italian (study 15), German–Spanish (study 15), Greek–English (study 16), and Spanish–Catalan (study 13). Ten studies measured sentence comprehension (studies 1, 2, 3, 4, 6, 7, 8, 9, 10, and 11) while five studies explored sentence production (studies 5, 12, 14 and 15). One study collected data from both sentence production and comprehension (study 16). All studies used behavioral measures.

The meta-analysis was calculated based on 81 effect sizes reported in 14 articles. The pooled SMD was −0.155 (95% CI, −0.301 to −0.009; p < 0.001). There was substantial heterogeneity across studies (I² = 79.1%; 95% CI, 74% to 83%), and it reached significance (Q₈₀ = 382.99; p < 0.001). These results showed a statistically significant difference between monolinguals and bilinguals in L1 morphosyntactic processing, and according to Cohen’s proposal, the effect size was small.

Results of the Egger’s test (bias, −0.907; 95% CI, −4.082 to 2.268; p = 0.571) and Begg and Mazumdar’s test (z = 1.02; p < 0.314) indicated no publication bias. The Duval and Tweedie procedure did not impute any missing studies, and the SMD did not change [SMD, −0.155; (95% CI, −0.301 to −0.009); p < 0.038]. As a result, there is no evidence of publication bias. The funnel plot is shown in Figure 2.

The pooled SMD based on the 81 effect sizes revealed little change when Cohen’s d [SMD, −0.157 (95% CI, −0.305 to −0.008; p = 0.039)] and the fixed-effects model [SMD, −0.144 (95% CI, −0.210 to −0.078; p = 0.000)] were used. The effect size remains small in both cases, and the differences between monolingual and bilingual speakers remain statistically significant. If studies on morphological processing were independently subject to analyses, the SMD increased considerably, and the differences between monolingual and bilingual speakers remain statistically significant [SMD, −0.882 (95% CI, −1.167 to −0.598; p < 0.001)]. However, when studies on syntactic processing [SMD, −0.011 (95% CI, −0.155 to 0.133; p = 0.879)], comprehension tasks [SMD, −0.165 (95% CI, −0.339 to −0.008; p = 0.061)], and production tasks [SMD, −0.124 (95% CI, −0.397 to 0.148; p = 0.372)] were analyzed separately, the SMD varied with no significant differences between monolingual and bilingual groups. Table 2 summarizes the results of the sensitivity analysis.

Table 3 depicts the results for the subgroup analysis. The studies that did not report information about some of the variables of interest (AoA, immersion context, immersion duration, modality of presentation, structure similarity) comprised a no-information group included in the corresponding subgroup analysis to explore their role in the outcome.

We inspected subgroups associated with characteristics of bilingual experience. There were statistically significant differences between subgroups investigating early, late bilinguals, and a no-information group (Q₂ = 18.05, p < 0.001) with greater effect sizes observed in studies without information related to AoA (k = 11), while studies including late AoA (k = 45) presented a greater effect compared to those comprised of bilinguals with early acquisition of the L2 (k = 25). However, none of the individual effects in each group was significant (all p’s > 0.05). When considering immersion, no differences appeared between articles exploring bilinguals immersed in their L1 (k = 1) and those exploring bilinguals immersed in their L2 (k = 80), but this last group of studies presented a significant effect size (p = 0.035). Finally, when looking at LoR in the L2 environment, we found statistically significant differences between short and long LoR studies (Q₁ = 17.76, p < 0.001), and the effect was significant in the short LoR (k = 2), but it did not reach significance in the long LoR subgroup (k = 79; p = 0.080).

Exploration of the similarity of the structures between languages revealed that effect sizes significantly differed when we compared studies using similar and dissimilar features between languages (similar, dissimilar, collapsed, n/a, and no-information, Q₄ = 109.21, p < 0.001), but differences between monolinguals and bilinguals were not significant in the group of studies that used either similar or dissimilar structures. As seen in Table 3, effect sizes were larger if the languages’ characteristics were collapsed (k = 1).

Investigating the modality of stimuli presentation, research dealing with auditory, visual, and audiovisual material did not significantly differ between modalities, and the effect size was statistically significant only in the group with visual presentation (k = 42; p = 0.043).

Meta-regression analyses with RVE (Hedges et al., 2010; Tanner-Smith and Tipton, 2014) were performed. We considered the effect of each continuous moderator individually (LoR, L2 proficiency, or L2-AoA; bivariate meta-regression) over the SMD, and the results indicated a significant difference in the effect size magnitude associated with LoR (β = 0.040 [95% CI, 0.013 to 0.066]; p < 0.016). However, according to Tanner-Smith et al. (2016), the p-value was untrustworthy because the degrees of freedom were less than four. No significant differences were observed for proficiency in L2 and L2-AoA variables (see Table 4).

When the effect of all covariates (LoR and L2-AoA) were considered simultaneously (multivariate meta-regression) in the same RVE meta-regression model, LoR and age of acquisition did not predict the effect size magnitude. L2 proficiency had to be excluded from the analysis due to missing values.

Sentence production is less explored than comprehension despite the stress that recent approaches give to considering both production and comprehension and its interdependence to feed language models (Pickering and Garrod, 2013; Dell and Chang, 2014). Our work reflects this trend in research, with a higher number of studies targeting comprehension.

Within comprehension, most of the reviewed studies used acceptability/grammaticality judgments (12 studies; 10 included in the meta-analysis, see Table 1). In an investigation by Schmid and Dusseldorp (2010), German monolinguals outperformed German–English bilinguals and German–Dutch bilinguals in verbal fluency, the C-test, and a film re-telling task; however, their behavior was similar in an auditorily grammatical judgment task. Although their results may reveal stable L1 knowledge at this linguistic skill in bilinguals, the authors do not discard the hypothesis that the sentences could be easy for the readers. Also, the use of brain markers as participants read sentences may reveal differences that the offline acceptability judgments do not, even in low complexity sentences (see Román et al., 2015). For example, Italian–English bilinguals and Italian monolinguals in Kasparian et al. (2017) had to rate the grammaticality of eight-word sentences in Italian that could include local or non-local agreement violations while neurophysiological activity was recorded (for example, “Il lavatore torna dalla fabbrica sporco di grasso,” The workers_plural returns_singular from the factory dirty_singular with grease). They did not find a group effect or group interaction at the acceptability ratings, but groups differed in reaction times and the neurophysiological patterns. Together with neurophysiological methodologies, other online measures like eye-tracking or self-paced reading may be more sensitive than offline acceptability judgments to catch differences between groups (e.g., self-paced listening, Kaltsa et al., 2015, in our review; for a deeper discussion on online measures in the field, see Marinis, 2003; Roberts, 2012). The prevalence of the acceptability-ratings task and behavioral measures within our sample of studies might be behind the lack of a subgroup effect.

Research dealing with sentence production, on the other hand, employed spontaneous speech/writing sampling (speech, Schmid, 2014; Schmitz et al., 2016; writing, Chunpeng and Hee-Don, 2017), storytelling (Ribes and Llanes, 2015), and a headlines task, which provides a verb, a noun phrase and an adverbial expression that participants have to use to produce a sentence describing a picture (Tsimpli et al., 2004). When present, differences in these studies showed that bilinguals tended to be slower and employed syntactical structures that were permitted in their L1 but preferred in their L2, suggesting that variation, rather than attrition or loss is a term better suited for L1 changes in individuals that speak more than one language (Schmid, 2014; Schmitz et al., 2016). Although a separate analysis of bilinguals vs. monolinguals in sentence production studies did not show a significant effect size, the production pattern reveals co-activation and transfer of L2 features in the bilingual grammar, as we will explain later.

The magnitude of effect sizes differed between groups within AoA, and LoR, as seen in the subgroups analysis, but were only statistically significant when looking at studies that included immersion in L2 and short LoR. Nevertheless, LoR was the only continuous variable predicting differences between monolinguals and bilinguals in the bivariate meta-regression. In Schmid and Dusseldorp’s (2010) study mentioned above, one of their objectives was investigating variables that predicted L1 attrition. Only LoR predicted attrition in free speech (in lexical diversity and errors) using a film re-telling task. Some articles in the present review that did not investigate the relation of attrition with LoR used a similar procedure. For example, Schmid (2014) had a subgroup of 20 late German/English bilinguals with a minimum LoR in Canada of 9 years but up to four decades (mean and standard deviation not provided) that did not differ from their monolingual peers in morphosyntactic variables. In our meta-regression, lower LoR is associated with larger differences between bilinguals and monolinguals and grants the need for further research, including an earlier (and shorter) range of LoR in sentence production.

More recently, Schmid and Karayayla (2020) explored the role of LoR in comprehension. They collected sentence production data from 92 Turkish-English bilinguals (collapsed including heritage speakers and therefore excluded from our review and meta-analysis) about their L1 maintenance and acquisition within a wide range of age at onset of bilingualism (AaO; from birth to adulthood). The data indicated that LoR predicted morphosyntactic complexity in L1, but in a direction opposite to our results, the longer the residence in a context where the second language is used, the lower the proficiency in L1. Importantly, they observed that the effects were more evident in early than late bilinguals. Because in the present work 11 of the 14 studies included late bilinguals, it is necessary to assume that variations in the L1 are qualitatively different between bilingual experiences and likely the result of different cognitive processes.

In spite of that, the continuous variables LoR, L2 proficiency, and AoA did not predict variation in the L1 in the multivariate meta-regression. One potential explanation is that bilinguals can use cues of different nature to compensate for differences, masking the effect of the variables of interest at the behavioral level examined here. For example, in a study collecting neurophysiological data, Kasparian and Steinhauer (2017) used an acceptability judgment task in Italian with relative clause structures that could be temporarily ambiguous (garden-path as in “Il poliziotto che i ladri arresta registra i nomi,” The policeman that the thieves arrests registers the names). The sentences were grammatical in Italian, but some were ungrammatical in English (as in the example) and less preferred garden-path structures in Italian. As expected, bilinguals found the grammatical sentences in Italian but ungrammatical in English as less acceptable than monolinguals. Moreover, the authors predicted a P600 in both groups (greater in bilinguals) to the verb in the relative clause (arresta), commonly found in garden-path sentences as an index of syntactic difficulty, and an N400 related to difficulties in semantic integration because they introduced strong semantic cues (policeman-thieve-arrest) that did not conflict with the structure. Italian monolinguals, which rely more on semantic information, evinced an N400, while the bilinguals did not show this component but a greater P600, as anticipated if they used, like English monolinguals, the strict word order preferably than semantic cues.

One problem in our study that may prevent us from finding a stronger impact of L2 proficiency is using the bilingual term and the collection of L2 proficiency data broadly to cover as many studies as possible. Some authors have warned about the implications that the way we conceive, and measure bilingualism have on the diversity of outcomes we obtain in our growing field of knowledge (Green and Abutalebi, 2013; Luk and Bialystok, 2013; Ooi et al., 2018; Anderson et al., 2020; Kremin and Byers-Heinlein, 2021). In the pool of articles reviewed here, proficiency tests go from subjective reports, including those with questions about how much effort it takes to use a language (Liceras and Senn, 2009), to objective measures and placement tests (for example, TOEFL for English proficiency, Chang, 2009; Gürel, 2015; telc for Turkish, Arslan et al., 2015). In addition, some works did not contain proficiency information (e.g., Kaltsa et al., 2015), or provided ranges (e.g., Bylund and Ramírez-Galan, 2014). Under such circumstances, there was enough data to consider proficiency as a categorical variable but impeded the continuous data to be used as recommended for the meta-regression (Bialystok, 2018; De Cat et al., 2018; Gunnerud et al., 2020).

As mentioned in the introduction, researchers have considered three leading causes behind the patterns observed in bilingual L1 processing: language co-activation in bilinguals, cross-linguistic transfer, and a reduced frequency of L1 use (Costa and Sebastián-Gallés, 2015; Cunnings, 2017a; Schmid and Köpke, 2017).

Cunnings (2017a) proposes that retrieval interference/facilitation lies behind differences in bilingual comprehension. Retrieval interference appears because working memory is a capacity-limited entity (Baddeley, 2013), and co-activation of languages increases the demands when bilinguals have to select one representation from those retrieved from long-term memory and integrate it with the incoming information. In agreement with this idea, Chunpeng and Hee-Don (2017) tested Chinese/Korean bilinguals reporting the use of both languages daily, and therefore more prone to co-activation. Their sample presented difficulties retrieving words in written composition and spent a longer time completing the task compared to Chinese monolinguals. Apart from that, in their research, grammar differences were mainly related to transfer from their L2 (Korean word order, punctuation, among others), similar to English/Spanish bilinguals in Ribes and Llanes (2015) that produced more subordinate constructions allowed in English but preferred in Spanish, and more pauses. Bilinguals may then become “opportunistic” speakers under co-activation contexts selecting representations that alleviate their cognitive load and languages cooperate rather than interfere (Beatty-Martínez et al., 2020).

A prediction derived from the above is that similarities across languages may influence bilingual processing, increasing co-activation, interfering or facilitating processing, or making bilinguals opting for the common structures in sentence processing. However, we have not observed differences between groups with similar and dissimilar structures, and none of the analyses that independently tested similarity and dissimilarity in bilingual vs. monolingual performance was significant. In this respect, co-activation might happen not only when the use of specific structures sharing properties across languages spread activation to nodes in the L2, but also when using the L1 in broader contexts such as L2 immersion (as seen confirmed in the meta-analysis) or when co-activation is locally induced (e.g., watching a movie in the L2; Elston-Güttler et al., 2005). This indicates that any context that prompts co-activation may increase the chances of altered processing even in the absence of shared properties.

In association with this idea, Green and Abutalebi (2013) proposed the adaptive control hypothesis. According to this approach, bilinguals find themselves in different contexts, and each posits specific demands to which the reader/speaker adapts. That is, bilinguals’ L1 (and L2) is not only subject to differences in the long run but varies depending on the demands of the environment (e.g., the accent of the interlocutor, unilingual workplace, bilingual community, discourse complexity, etc.). Whenever co-activation creates competition between language schemas, there will be a need for processes that handle interference and select the desired linguistic representations (Green and Abutalebi, 2013). Such processes, on the one hand, use limited resources that may be unavailable for efficient processing (Perfetti and Stafura, 2014; Li and Clariana, 2019) and, on the other hand, make competing representations less accessible through inhibition (e.g., Levy et al., 2007). These processes engaged in controlling interference might explain the direction of LoR effects in our data; at short LoR, the native language is dominant in late bilinguals and L2 usage is expected to trigger inhibitory processes to overcome the interference and facilitate retrieval of the weaker L2 representations. When bilinguals try to retrieve their L1 later, it takes time to access the suppressed representations in the L1. As asymmetry between languages decreases, inhibition is no longer needed (Levy et al., 2007).

Also, predictions from this approach are that reading and producing sentences will show either an impairment (for example, an increasing number of pauses, reduced processing speed) or compensation (for example, using simpler structures, using a limited number of cues, repetition of more accessible structures). As discussed earlier, bilinguals in some of the studies reviewed here present more pauses and longer time to produce written or verbally sentences in their L1 (Ribes and Llanes, 2015; Chunpeng and Hee-Don, 2017). Additionally, it is expected that bilinguals (and monolinguals) will have to overcome co-activation and inhibition, and that succeeding in doing so must be more difficult in constructions that demand resources allocation, such as complex grammatical forms (e.g., evidentiality marking in Turkish, Arslan et al., 2015) or pronominal referential relations in sentence comprehension (Li and Clariana, 2019). In fact, pronominal resolutions have been studied in the context of language attrition, considering that difficulties in this type of structure may arise due to the need to coordinate different interfaces (grammatical and discourse information; Sorace, 2011; Chamorro et al., 2015). Giving support to this view, we have found research in which bilinguals do not have a strong preference toward a pronominal assignment in the presence of more than one antecedent noun as monolinguals have (Kaltsa et al., 2015; Castro et al., 2017; but see Liceras and Senn, 2009; Schmitz et al., 2016) and show preferences that place less cognitive demands on the reader (low-attachment in L1 in both English–Spanish and Spanish–English bilinguals, Dussias, 2003, 2004), thereby supporting the notion of bilinguals as strategists that adapt to the requirements of their linguistic and cognitive context.

Another factor that may alter L1 processing is the cross-linguistic transfer (Costa and Sebastián-Gallés, 2015). In this regard, we have seen that bilinguals show in their L1 a bias to structures and interpretations that appear frequently in their L2 (examples in production are Ribes and Llanes, 2015; Chunpeng and Hee-Don, 2017; in comprehension, Dussias and Sagarra, 2007). Changes in the L1 sentence processing can appear as a lack of preference for both L1 and L2 biases too, similar to what Kaltsa et al. (2015) and Castro et al. (2017) observed. Specifically, in two experiments, Kaltsa et al. (2015) employed a self-paced listening task with a sentence-picture matching to test pronoun resolution in Greek. They presented sentences such as “I γiaγia xeretise tin kopela otan afti pernuse to δromo” (The old lady greeted the girl when she crossed the street) where the antecedent of the overt pronoun in the subordinate clause “afti” (she) could be the subject (old lady) or the direct object (the girl). In Greek, the null subject is the default, and the subject antecedent is preferred under such conditions, while a non-subject antecedent (the girl in the example) is preferred in overt subject pronoun as the one in the example. In this research, Greek–Swedish bilinguals with long LoR were expected to choose more subject antecedents in overt pronoun conditions than monolinguals, but their performance should be similar in null pronoun conditions (Swedish is a non-null subject language). Monolinguals showed differences between null and overt pronoun conditions, but bilinguals did not. Despite this outcome, Kaltsa and colleagues consider that cross-linguistic transfer is not causing their results because a group of older monolinguals presented a performance closer to bilinguals.

Finally, a frequency-based account highlights a reduced L1 input that weakens and biases activation toward L2 properties (Gollan et al., 2005). One way to explore this effect would have been to include relative frequency of use across languages in our analyses, but only a few had information about it (Liceras and Senn, 2009; Castro et al., 2017), and they did not explore its impact. Immersion length may switch the frequency balance between L1 and L2, and studies exploring parsing preferences that differ across languages are helpful since language biases may progressively shift with exposure. However, the results in our meta-regression with shorter LoR showing larger effects say otherwise. In two studies, Spanish–English bilinguals with long L2 immersion (Dussias and Sagarra, 2007) and Spanish–English bilinguals with shorter L2 immersion (Dussias, 2004) but similar proficiency change from L1 to L2 preferences while highly proficient Spanish–English bilinguals immersed in their L1 maintained the monolingual routines (Dussias and Sagarra, 2007). Such results could suggest that frequency-related modifications occur rapidly and appear in readers with shorter exposure to L2 but also that there is more than one mechanism underlying variations across bilingual groups in agreement with evidence that shows a non-linear L1 variation (Schmid and Karayayla, 2020). Other processes such as the need for L1 inhibition in an L2 setting cannot be ruled out, and ERPs studies point to a more significant role of competition in grammar differences instead of a reduced frequency in L1, at least in the early stages of immersion in an L2 context (for a review, see Steinhauer and Kasparian, 2020).

Although these explanations do not exclude each other and may act simultaneously, they do predict different outcomes under different conditions. Therefore, more research is necessary to separate their respective effects and the relative weight they have on bilinguals’ L1 at several stages and situations.