The topic of cross-linguistic influence at the level of morphosyntax has been extensively investigated in child bilingualism (see Serratrice (2013) for an overview and van Dijk et al. (2021) for a recent meta-analysis evaluating cross-linguistic influence across 26 experimental studies). Research has shown that one language can have an effect on the other language at a morphosyntactic level (Hulk and Müller, 2000) and can lead to differences between monolingual and bilingual children which can be either quantitative, qualitative, or both. Quantitative differences stem from the frequency with which a certain structure is accepted or used by bilingual compared to monolingual children (Serratrice et al., 2004; Argyri and Sorace, 2007; Nicoladis and Gavrila, 2015). In other words, a phenomenon also present in monolingual development is reinforced in bilingual development under the influence of one language over the other. Qualitative differences stem from the presence of different language patterns in bilingual children’s production and comprehension relative to monolinguals (Nicoladis, 2006, 2012; Strik and Pérez-Leroux, 2011). Recently, Bosch and Unsworth (2020) investigated cross-linguistic influence in the production and acceptability of V2 word orders in English-Dutch bilingual children and found both quantitative and qualitative differences. Bilinguals accepted V2 orders with auxiliary verbs significantly more than monolingual children, but also accepted V2 with main verbs, contrary to monolinguals.

According to Hulk and Müller (2000) and Müller and Hulk (2001), cross-linguistic influence holds when the child’s two languages overlap at the surface level. If one language (language A) displays two structural options and the other language (language B) only makes one of these options available, then the option shared by the two languages may be reinforced in language A under influence from language B. In other words, “there has to be a certain overlap of the two systems at the surface level” (Hulk and Müller, 2000, p. 228–229). However, there is mixed evidence from the literature showing that cross-linguistic influence does not hold even in the presence of such structural overlap (Argyri and Sorace, 2007) or that cross-linguistic influence occurs in the absence of structural overlap (Nicoladis, 2006, 2012). Importantly, cross-linguistic influence does not seem to occur all the time and one of the factors that has been proposed to influence cross-linguistic influence is language dominance, which refers to the language that the child uses more frequently or the language in which the child has higher proficiency (Yip and Matthews, 2006). Here the prediction is that cross-linguistic influence goes from children’s dominant language into their weaker language (van Dijk et al., 2021), although there are also studies which found no relation between cross-linguistic influence and language dominance (Blom, 2010; Serratrice et al., 2012), showing that cross-linguistic influence can occur independently of language dominance.

While the majority of studies on early bilingual acquisition has investigated children’s offline comprehension, judgements, and production, only a few have examined real-time sentence processing in bilingual children. These have mainly focused on early L2 learners and compared children’s real-time processing of L2 morphosyntactic properties to that of their monolingual peers (Marinis, 2007; Chondrogianni and Marinis, 2012; Marinis and Saddy, 2013; Chondrogianni et al., 2015a,b) and generally report qualitatively similar processing patterns in bilinguals and monolinguals. Lemmerth and Hopp (2019) and van Dijk et al. (2022) specifically tested the effects of cross-linguistic influence on bilingual children’s on-line sentence processing. van Dijk et al. (2022), for example, tested English-Dutch and German-Dutch bilinguals aged 5 to 9 on a self-paced listening task assessing processing of word order in Dutch sentences. They found similar listening patterns in the V2 and V3 condition in Dutch in both monolinguals and bilinguals, but also report effects of cross-linguistic influence in the German-Dutch group in the condition instantiating a structural overlap between the two languages. In other words, the German-Dutch bilinguals slowed down when listening to V2 structures in Dutch and this slowdown was more pronounced in children who were more German dominant.

In contrast to the substantial literature on L2 acquisition, comparatively fewer studies investigated the acquisition of morphosyntax in HL development and how this is affected by cross-linguistic influence from the societal language. Some studies found no effects of cross-linguistic, suggesting that language-external factors shape child HL development (Daskalaki et al., 2019; Rodina et al., 2020). Other studies linked the differences in performance between child heritage speakers and monolinguals to the properties of the societal language (Meir et al., 2017; Meir and Janssen, 2021).

The acquisition of wh-dependencies in the HL has also received little attention. Cuza (2016) used an elicited production task to assess subject-verb inversion in matrix and embedded questions in Spanish heritage children aged 5;0 to 13;3 born and raised in the US. The results showed that Spanish-English bilingual children produce subject-verb inversion in Spanish to a significantly lower rate that their monolingual peers and that they also use subject-verb inversion less in embedded compared to matrix questions. Cuza (2016) argues that this pattern of performance arises from the interplay between cross-linguistic influence from English, the societal language, language dominance and issues of structural complexity. In a similar vein, Strik and Pérez-Leroux (2011) assessed Dutch-French bilinguals aged 5 to 7 and living in France on the production of wh-questions in Dutch, their L1. Although Strik and Pérez-Leroux (2011) do not use the label Dutch heritage speaker for their bilingual group, the children included in their study match the criteria used to define HSs (see Kupisch and Rothman (2018) for a discussion on HL terminology and early child bilingualism). Strik and Pérez-Leroux (2011) found that some of the wh-questions that bilingual children produced in Dutch differed qualitatively from those produced by Dutch monolingual children and followed a French-like structure. These were questions with a fronted wh-phrase and without subject-verb inversion, like *Wat jij doe giraffe? (lit. What you do giraffe?), and also wh-in-situ questions as in *Jij doe wat giraffe? (lit. You do what giraffe?). According to Strik and Pérez-Leroux, complexity is a trigger for cross-linguistic influence such that structures involving less derivational complexity in one language (e.g., in-situ questions) may impact structures which are derivationally more complex in the other language (e.g., wh-fronting with subject-verb inversion).

These previous works reporting different performance patterns in heritage compared to monolingual children assessed only children’s productive skills in their heritage language/L1. Various studies with monolinguals and bilinguals have revealed asymmetries between comprehension and production (Hendriks and Koster, 2010; Grimm et al., 2011) and although there are studies showing that production outpaces comprehension (see Hendriks (2014), Martinez-Nieto and Restrepo (2022) for the acquisition of pronouns), other studies report better performance in comprehension compared to production. Chondrogianni and Marinis (2012), for example, examined the on-line processing and production of tense and non-tense morphemes in L2 English children and children with Developmental Language Disorder (DLD). While the DLD children manifested difficulties with both comprehension and production, the typically-developing L2 children showed on-line sensitivity to the omission of tense morphemes, similarly to the L1 English children, despite variable production rates. Haiden et al. (2009) compared the comprehension and production of wh-questions in French by English- speaking children with L2 French and found high accuracy rates for their comprehension of questions with wh-fronting, on a par with those. In this study we compare HL children’s production to their real-time comprehension of multiple wh-questions and use both off-line and on-line methods. This can reveal whether HL children show qualitatively similar processing patterns to monolinguals but also whether asymmetries appear in the comprehension and production of questions with multiple wh-movement.

Full acquisition of multiple wh-questions involves various aspects that are subject to cross-linguistic variation. We will briefly outline the properties of multiple wh-questions that Romanian-speaking children need to acquire, by putting emphasis on differences with English. (1) illustrates multiple who-questions and (2) exemplifies which-questions in Romanian.

In terms of lexical properties of wh-words, wh-objects in Romanian are marked with a differential object marker pe, similar to a in Spanish. Although in the prescriptive use of English, who shows overt case-assignment in the form of whom, increasingly native English-speakers use who instead in informal spoken contexts (Aarts, 1994). Additionally, care (‘which’)-phrases in Romanian are doubled by a co-indexed clitic pronoun îl (‘him’) for masculine and o (‘her’) for feminine.

In terms of movement properties, wh-words move overtly, the difference with respect to English being that multiple wh-words in Romanian move together to a clause-initial position, as shown by (1) and (2) above. This is a property that Romanian shares with Bulgarian and other Slavic languages. According to Alboiu (2002), multiple wh-constructions in Romanian are derived by first moving the closest candidate (the subject), defined in terms of c-command, to a Spec,XP position. The remaining phrases then move via a ‘tucking in’ mechanism (see Richards, 1997) below the specifier created by the moved subject and this ‘tucking in’ movement of the following wh-phrases can take place in any order. On the other hand, fronting a who-object over a who-subject is ungrammatical (in both Romanian and English), as indicated by the asterisk in example (1b). Movement of wh-words in both languages obeys Superiority (Chomsky, 1973), a condition that limits the ordering of wh-words and blocks one wh-word from moving over another wh-word occupying a hierarchically higher position in the structure. Alboiu (2002) suggests that Superiority is observed in Romanian under her proposed analysis. Given that the subject occupies a structurally higher position and is the closest candidate, it should move first. This requirement does not hold for which-questions, as evidenced by the grammaticality of the example in (2b; see Pesetsky (2000) for an explanation). Laenzlinger and Soare (2005) and Soare (2009) convincingly argue for Romanian that which-expressions always appear clause-initially, preceding who-phrases. By adopting a split-CP analysis (Rizzi, 1997) and a cartographic approach to syntactic structures (Rizzi, 2004), Soare (2009) shows that which-phrases target the specifier position of a Topic head above the specifier Focus position which they postulate as the landing site of who-phrases.

The semantic properties of multiple wh-questions require establishing a pairing relation between the wh-phrases: a felicitous answer for a question like (1a) is “The girl covers the dog and the boy covers the cat.” in which the exhaustive sets of who and which are pairwise linked.

Children’s experience with such sentences is extremely reduced. Grebenyova (2005, 2011) showed that there are only five instances of such questions in the English CHILDES database. A search through the two corpora on Romanian in CHILDES (MacWhinney, 2000) yielded no instances of multiple wh-questions. The acquisition of multiple wh-questions has received relatively little attention in the literature. Grebenyova (2011) elicited multiple interrogatives from 20 monolingual English-speaking children (aged 3;07–6;02), 20 monolingual Russian-speaking children (aged 3;05–6;05) and 18 Malayalam-speaking children (aged 4;05–5;04). The three languages differ with respect to the movement properties of wh-words. Russian allows multiple wh-fronting, while English fronts one wh-phrase and Malayalam is a wh-in-situ language. Grebenyova’s findings demonstrate that English- and Malayalam-acquiring children have adult-like knowledge of the syntax of multiple wh-questions, whereas Russian-speaking children allow fronting of only one of the wh-phrases, following an English-like structure.

To our knowledge, three studies so far investigated the acquisition of multiple wh-questions in Romanian and they all looked at how Romanian-speaking monolingual children ranging in age from 4 to 9 years old comprehend this type of question. Bentea (2010) examined how 4- to 6-year-old English, French and Romanian children (24 in total) interpret multiple wh-questions (i.e., whether they assign pair-list readings to multiple interrogatives). Bentea (2010) was also interested in whether children assign an adult-like structure to multiple interrogatives in their language and whether cross-linguistic differences appear between English, French and Romanian children regarding the interpretation and structure of multiple questions. Bentea’s (2010) results showed similar performance in the English and French groups, while Romanian-speaking children were more likely to answer only the lower wh-element present in the question. In the same vein, Măniță (2017) addressed the question of exhaustivity in the comprehension of Romanian multiple interrogatives. Măniță (2017) tested 42 monolingual Romanian-speaking children (age range 4;0–6;10) and found that the rate with which children give exhaustive answers increases with age, although it does not reach ceiling performance at the age of 6. Furthermore, her results show that children preferentially answered the highest wh-word, which was also the subject.

In a recent study on the processing of Romanian multiple who and which-questions, Bentea and Marinis (2021) show that both monolingual children (6 to 9-year-olds) and adults slow down when processing who- compared to which-phrases, as measured by reaction times (RTs) in a self-paced listening task. However, only adults seem to show an online sensitivity to the ordering constraints in who-questions illustrated in (1b) above. Bentea & Marinis also report higher accuracy scores with multiple who- than which-questions and show that the latter pose more difficulties for comprehension, particularly in the object-subject order (1,2b), where participants (especially children) show a preference to interpret the first wh-element as agent, along the lines of what has been reported for the processing and comprehension of simple which-questions. Bentea & Marinis also found that children even at the age of 6 and 7 answered only one of the wh-phrases, similarly to Bentea (2010) and Măniță (2017), but provided exhaustive lists of referents either for the wh-subject or the wh-object. This suggests that Romanian children have difficulties with pairing the two wh-elements and that this difficulty persists until around the age of 8 when they are able to exhaust the question domain and also pair the two wh-elements. Therefore, the question that arises is whether bilinguals, who receive less input than monolinguals and are often not tutored in the L1, converge on the correct syntactic structure for multiple wh-questions and attain knowledge of the grammaticality distinctions among who and which-multiple questions, especially when such wh-dependencies display a different structure in the L2, the societal language.

To sum it up, multiple wh-interrogatives allow to explore the extent to which bilingual children’s language comprehension and production are affected by cross-linguistic influence, as they vary across languages and display language-specific syntactic and semantic properties that children need to acquire although these structures are not frequent in the parental input. In this study, we compare for the first time both the on-line/off-line comprehension and the production of these structures in Romanian heritage children in order to get a clearer picture of the way in which the societal language (here English) influences the acquisition of morphosyntax in the HL.

The present study investigates the early stages in the acquisition of the HL to examine whether the differences that emerge between HL children and monolinguals hold not only for production, as has been shown by the previous studies examining the acquisition of simple wh-questions in HL children, but also for comprehension. We postulated that the use of a more sensitive and implicit on-line comprehension task, like the self-paced listening task used in this study, might offer a more straightforward glimpse into underlying language representations that are accessed for real-time processing. Together with production tasks, on-line comprehension might help to better understand what differentiates between HL children and monolingual children. The study focused on Romanian as heritage language and addressed the following research questions:

Previous studies with L2 children looking at real-time sentence processing report qualitatively similar processing patterns in bilinguals and monolinguals for tense (Chondrogianni and Marinis, 2012), articles (Chondrogianni et al., 2015a), articles and clitics (Chondrogianni et al., 2015b), word order (van Dijk et al., 2022). Therefore, we expected Romanian HL children to show similar processing patterns to Romanian monolingual children. On the other hand, if there is cross-linguistic influence from English, the societal language, on Romanian HL children’s processing of multiple wh-questions, regardless of surface overlap (Nicoladis, 2006, 2012), then heritage children should slow down when they hear the second wh-phrase immediately following the first wh-word.

This is the first study to examine the production of multiple wh-interrogatives in Romanian, as previous studies have only looked into how Romanian-speaking children comprehend this type of questions (Bentea, 2010; Măniță, 2017; Bentea and Marinis, 2021). If the Romanian-speaking children tested in Romania have fully acquired the syntax of multiple interrogatives, they should mainly produce questions with multiple wh-fronting. As far as the heritage group is concerned, we base our predictions on the previous studies on the production of wh-dependencies in child HL (Strik and Pérez-Leroux, 2011; Cuza, 2016) which show qualitative differences between HL and monolingual children in the production of wh-questions. We thus expected Romanian heritage children to be more likely to produce multiple wh-questions with one fronted wh-phrase and one in-situ, under cross-linguistic from English, the majority language. Moreover, if asymmetries arise between production and comprehension, then we expect the results to show a similar pattern to that reported for other bilingual populations in which comprehension of multiple interrogatives outpaces their use in production (Haiden et al., 2009; Chondrogianni and Marinis, 2012; Chondrogianni et al., 2015a,b).

Eighteen 6- to 9-year-old Romanian heritage children with English as L2 (6 boys; age range = 6;0–9;3; mean age = 96.6 months; SD = 13.7 months) living in the United Kingdom (Greater London area and South-East England) and 30 Romanian monolingual children aged six to nine (15 boys; age range = 6;11–9;8; mean age = 99.1; SD = 11.2) living in Romania, participated in the study¹. None of the monolingual children had a history of speech and/or language delay or impairment, while one bilingual child had mild expressive language delay diagnosed at the age of three and for which she underwent Speech and Language Therapy until the age of six. As this participant’s results at the time of testing did not differ from those of other children, they were included in all subsequent analyses.

Details regarding the bilingual children’s language history, including information about their current use of and exposure to both Romanian and English, were collected using a modified version of the Questionnaire for Parents of Bilingual Children (PABIQ; Tuller, 2015). Three parents did not complete the questionnaire. The language background data obtained show that all children were exposed to Romanian from birth, but had a different age of onset (AoO) of English: one child was a simultaneous bilingual, nine children were exposed to English before the age of two (between 5 and 18 months), and five children were exposed to English after the age of three (between 3 and 5 years). The mean age of onset (AoO) of English was 1;10 years (SD = 18 months, range = 0–5;0 years) and the mean length of exposure (LoE) to English was 6;0 years (SD = 21 months, range = 2;1–8;0 years). The language background data for the heritage children illustrated in Table 1. Also show that Romanian is used more at home than English, as determined by a paired t-test [t(14) = 3.61, p = 0.002], whereas English is the dominant language outside the home [t(14) = 3.62, p = 0.003], as well as when it comes to children’s current expressive language skills, as reported by the parents [t(14) = 4.45, p < 0.001].

The study was approved by the Ethics Committee of the University of Reading. Informed parental consent was obtained for each child prior to the testing sessions, as well as the oral consent of the child. Each participant was tested individually in a quiet room either at their school or in their home. The experimenter gave oral instructions for both tasks. These were administered at least 1 week apart and in different orders, such that half of the children first saw the SPL task, followed by the elicited production task, while for the other half we first assessed production and then comprehension. Each testing session lasted around 30–40 min, but children could take breaks whenever they felt tired. They received stickers and certificates of participation after each task and at the end of the study they received as well a voucher which they could use to buy books at local bookstores.

The SPL task was programmed and administered using PsychoPy (Peirce et al., 2019). The task started with an introduction in which an image of Paddington first appeared on the screen, telling children what they need to do in the task. Children were also familiarized with the images and names of each of the three superheroes and princesses, although almost all them already knew these characters from cartoon movies. Before starting the test phase, the children were presented with four practice items, two of which contained simple what-questions (e.g., What is Batman reading?) and two multiple wh-questions with an animate subject and an inanimate object (e.g., Who is drinking what?). Each participant then listened to a total of 50 sentences during the test phase and these were randomized in PsychoPy. The task instructions and all the sentences were pre-recorded by a native speaker of Romanian.

The elicited production task was administered as a PowerPoint presentation. The task started by introducing each child to the Paddington puppet. The experimenter explained that Paddington wanted to become a magician and for this he had to improve his guessing skills. The child’s task was to help Paddington by playing a game with him in which the child asked Paddington questions about various images and the puppet had to guess the correct answer. For each correct guess, the child gave Paddington a smiley face sticker and if Paddington had at least 20 correct guesses by the end of the task, the child gave Paddington his Magician Diploma. Given the complexity of the target questions, these were presented in four blocks and interspersed with fillers. Each block contained 6 items of the same type (e.g., SO who questions) and the order of the blocks was randomized across four lists, each containing 30 items. The task began with four practice trials which elicited multiple wh-questions with a mismatch in animacy (e.g., Who ate what? or Which boy hid where?). The audio presentation of the items was also pre-recorded by a native speaker of Romanian. The questions that the children produced were recorded on answer sheets and then coded for analysis.

Figure 1 presents the descriptive results for comprehension accuracy of Romanian multiple wh-questions with two who and two which-phrases and with distinct subject-object orders. The results indicate that both the Romanian-speaking monolinguals and the heritage Romanian children comprehend multiple who-questions well (above 0.80), but show lower accuracy for multiple which-questions.

The two groups also show similarities in the errors they produce, which are (i) over-exhaustive answers, (ii) list answers, and (iii) reversed role answers. For over-exhaustive answers children name all the pairs in the image, even when one pair performs a different action. For example, children would answer a question like in (3) or (4) above with The fireman is splashing Superman, the elephant is splashing Batman, and the boy is pulling Spiderman although the question only refers to the action of splashing. List answers are cases in which children only answer one of the two wh-words by exhaustively listing all the characters involved in the action. That is, they would answer question (5) with Jasmine and Elsa. Role reversals are answers in which the Agent-Patient roles are reversed. For instance, children would answer the question in (6) with Anna is covering the black monkey and thus assign the wrong thematic role to which princess. There were also very few instances in which children simply identified the wrong action, for example the pulling action when the question was about splashing. Table 2 summarizes the number and type of errors for monolingual and heritage children and indicates that the most common errors children make with multiple who-questions are providing list answers to only one of the wh-words, while role reversal errors are the most frequent with multiple which-questions. This suggests that children find it harder to assign the correct thematic roles in a patient-before-agent structure when the question contains two which-phrases.

Given the binary nature of the data (Correct/Incorrect), we analyzed the accuracy results using a binomial generalized linear mixed model with group (monolingual vs. heritage), wh-type (who vs. which), wh-order (SO vs. OS), and their interaction as fixed factors. The fixed factors were coded using repeated contrast coding which tests the difference between the mean of the dependent variable for one level of the categorical variable and the mean of the dependent variable for the adjacent level (Schad et al., 2020). The random-effects structure included intercepts for both participants and items, as well as random slopes for wh-type by participant. Alternative models with a more complex random-effect structure either failed to converge or were not retained as their goodness-of-fit resulted in increased Akaike information criterion (AIC)-value. The analysis was conducted using the lme4 package (Bates et al., 2015) in R (R Core Team, 2022). Planned comparisons, if justified, were done using the emmeans package (Lenth, 2022).

The three-way interaction group*wh-type*wh-order did not significantly improve the fit of the model as indicated by model comparison using the anova function (p = 0.988). The summary of the final model is given in Table 2. The results showed a significant effect of wh-type (who-questions were comprehended better than which-questions) and a significant wh-type*wh-order interaction. No differences in performance emerged between the monolingual and the heritage group (Table 3).

We followed-up on the significant interaction between wh-type and wh-order with a pair-wise comparison. This showed that response accuracy for questions with two who-phrases did not differ significantly for the SO and OS orders (β = −0.402, SE = 0.271, z = −1.480, p = 0.138), while there was a significant difference between the two wh-orders in the case of questions with two which-words, with SO which questions yielding significantly better comprehension scores than OS which questions (β = 0.405, SE = 0.150, z = 2.703, p < 0.01).

The segment-by-segment residual reaction times (RTs) for the online processing of multiple wh-questions are illustrated in Figure 2 (for the monolingual group) and Figure 3 (for the heritage group). We plot raw residual RTs for readability, but the analyses were done on log-transformed RTs. Only items with correct responses to the comprehension questions were included in the RT analyses. Residual RTs were calculated to control for the difference in length between segments. Extreme values were calculated based on boxplots and were excluded from the final analyses. These were residual RTs below −600 ms and above 2000 ms. RTs of 2 SD above or below the mean per condition per participant and per item were considered as outliers and therefore replaced with the mean per condition per participant and per item. The total proportion of extreme values and outliers was 3.2% of all data points.

We report the results from six segments, starting with the segment containing the first wh-word and including the segments after the verb, as these can reveal spill-over effects. The log-transformed RTs for each segment were analysed using linear mixed-effects models fit with the maximal random effects structure that converged. This included by-participant and by-item random intercepts and slopes for wh-type by-participant. Group (Monolingual vs. Heritage), wh-type (who vs. which), wh-order (SO vs. OS), as well as their interaction, were included in the fixed effects structure for each model. All fixed effects were coded using repeated contrast coding. Values of p were calculated by way of Satterthwaites’s approximation to degrees of freedom with the lmerTest package (Kuznetsova et al., 2017).

Segment 1 (the first wh-word) Results attested to a significant effect of group (β = 0.178, SE = 0.057, t = 3.121, p = 0.003), indicating that the monolingual children had overall longer RTs than the heritage children. There was also a significant effect of wh-type (β = 0.237, SE = 0.042, t = 5.522, p < 0.003), with longer RTs for who-phrases (M = 482 ms) than which-phrases (317 ms), as well as a significant effect of wh-order (β = −0.178, SE = 0.057, t = 3.121, p = 0.003), with shorter RTs for subject (M = 371 ms) compared to object wh-words (458 ms). None of the interactions was significant.

Segment 2 (the second wh-word) Results revealed a significant effect of group (β = 0.143, SE = 0.055, t = 2.564, p = 0.013), with the monolingual children displaying longer RTs than the heritage children, and a significant effect of wh-type (β = 0.167, SE = 0.034, t = 4.845, p < 0.001), with longer RTs for who-phrases (M = 484 ms) than which-phrases (352 ms). No other effect or interaction were significant.

Segment 3 (the verb in multiple who-questions and the clitic in multiple which-questions) There was a significant effect of group (β = 0.190, SE = 0.084, t = 2.266, p = 0.028), with longer RTs for the monolingual group compared to the bilingual group, and a significant effect of wh-type (β = −0.436, SE = 0.046, t = −9.291, p < 0.001) showing that the verb segment in the multiple who-conditions yielded shorter RTs (M = 386 ms) than the clitic segment in multiple which-conditions (M = 638). There was also a significant interaction Group*WhType (β = 0.151, SE = 0.051, t = 2.910, p = 0.003). As a follow-up on the significant interaction, pair-wise comparisons revealed that the heritage group showed significantly shorter RTs than the monolingual group for the who-conditions, (β = −0.266, SE = 0.086, t = −3.065, p = 0.003), while no significant differences surfaced between the two groups in the which-conditions.

Segment 4 (the adverb in multiple who-questions and the verb in multiple which-questions) No effect was significant.

Segment 5 (in the morning) Results attested to a significant effect of group (β = 0.149, SE = 0.072, t = 2.057, p = 0.045), as the monolingual children had longer RTs than the heritage children. No other effect was significant.

Segment 6 (at the zoo) No effect was significant.

To summarize, the results of the comprehension task show similar response accuracy and a similar pattern during online comprehension of multiple wh-questions in Romanian-speaking monolingual and heritage Romanian children.

The results of the elicitation task showed that children do not only produce questions with multiple wh-fronting, the expected target structure based on the syntax of Romanian multiple wh-questions, but that they often produce other structures as well. To reflect this variability in the children’s answers, four scoring categories were used, each corresponding to the four main question types that children produced and which we classified as follows:

The distribution of responses differs between the Romanian monolingual and Romanian heritage children, as can been seen in Figure 4. Monolingual children produce three types of questions at similar rates: MWH_MULTIPLEMOVE (0.34), MWH_SINGLEMOVE (0.28) and SIMPLE_WH (0.31). They also produce COORDINATED_WH to a lesser extent (0.07). The most frequent type question that the heritage children produce is MWH_SINGLEMOVE (0.68), followed by SIMPLE_WH (0.24). There are also a few instances of COORDINATED_WH questions (0.04), as well as instances of MWH_MULTIPLEMOVE questions (0.04). However, a closer look at the data reveals that the majority of multiple wh-questions with two fronted wh-phrases are produced by one child and that there are six other children who only produce one question with multiple wh-movement throughout the whole task.

A finer characterization of the results (Figure 5) gives an indication of the response distribution within the two groups of participants for each type of elicited question (see examples 7 to 10 above). As the children mainly produced questions with a subject-object order, we collapsed the results of SO which and OS which questions and thus report the data for three types of multiple wh-questions with two who-phrases (who), with two which-phrases (which), and with one which and one who phrase (which-who). Figure 5 shows that MWH_SINGLEMOVE questions (with one wh-phrase fronted and one in-situ) represent the preferred produced structure for the heritage group across all types of elicited questions, irrespective of whether these contained only who, only which, or both which and who phrases. The monolingual group produce more MWH_SINGLEMOVE structures of the type illustrated in (12) with questions containing which-phrases, whereas they produce more MWH_MULTIPLEMOVE structures like in (10) in the presence of two who-elements.

In order to uncover whether differences emerge between monolingual and heritage children in their productions and whether these differences are modulated by the type of elicited question (who vs. which vs. which-who), we fitted a generalized liner mixed model for each of the four scoring categories outlined above, namely MWH_MULTIPLEMOVE, MWH_SINGLEMOVE, SIMPLE_WH, COORDINATED_WH. The dependent variable in each model was response accuracy, that is, the correct production of questions within each scoring category. We analyzed the production data using the lme4 package (Bates et al., 2015) in the R environment (R Core Team, 2022), specifying the optimizer ‘bobyqa’ for our models. Each model included the same fixed factors (Group: Monolingual vs. Heritage and QuestionType: who vs. which vs. which-who) as well as their interaction. We used a repeated contrast specification for the fixed factors. The random effect structure included by-participant random slopes. We only report the significant effects and interactions.