Language experiences, evaluations and emotions (3Es): analysis of structural models of multilingual identity for language learners in schools in England

ABSTRACT Researchers have traditionally relied on institutionally-defined, native speaker conceptions of linguistic competence to define whether or not individual learners are multilingual. However, to better understand the relationship between language and identity in the context of additional language learning, there is a need to widen the definition of multilingualism by including a larger repertoire of semiotic competences and a wider range of levels of linguistic competence. To date, little systematic research has investigated the relationship between different key variables influencing language learners’ construction of their multilingual identity (MI). Based on surveys completed by 1338 learners in seven secondary schools in England, this paper reports on a structural equation modelling of key latent variables influencing MI. The analysis suggests that MI is more strongly correlated with direct experiential contact with languages and with social interaction than with explicit beliefs about language learning. Similarly, the influence of parental beliefs is greater than those of teachers or friends. The paper also considers the significance of potential differences between EAL and non-EAL respondents and urban and rural school contexts. Having conducted SEM we offer support for our theorisation of MI as an outcome latent variable composed of three main components, namely, experience, evaluation and emotion.


Introduction
Linguistic identity research is notoriously complex.Unravelling the multilingual identity construct is arguably all the more challenging as it is a latent construct that encompasses a number of linguistic identities sitting within an overarching multilingual identity umbrella.The aim of the work presented here is to theorise the dimensions of multilingual identity and develop an integrated and holistic theoretical model of multilingual identity that can be used for systematic data collection and comparative analyses across groups.
Drawing on data from surveys completed by 1338 secondary school-age learners in England, we use Structural Equation Modelling (SEM) to test the complex relationships between a number of latent variables.A sub-interest is how well a model of multilingual identity fits certain groups of language learners, such as those who identify as having languages other than English in their profile and those whose schools are in urban versus rural contexts.
Having conducted SEM we offer support for our theorisation of multilingual identity as an outcome latent variable composed of three main components, namely, experience, evaluation and emotion.Our findings suggest that, in this context in particular, learners' experiential contact with languages and with language-focussed social interaction, as well as their self-efficacy beliefs and emotional dispositions towards other languages, are key factors influencing their identifications as multilinguals.Considering the evidence that identification as multilingual is likely to affect all round academic attainment (Rutgers et al., 2021), we conclude that educators should pay attention to helping learners develop the agency to identify as multilingual, should they so wish, and should pay particular attention to the role of parents and to learners' self-beliefs.

Modelling multilingual identity: theoretical background
Introduction: from identity to multilingual identity Identity research in SLA has flourished in the last twenty years, with a growing body of work exploring the close link between language and identity within the sociocultural and poststructuralist traditions.Concerned with the relational and socially-and temporally-constructed nature of linguistic identity, scholars (e.g.Block, 2007;Darvin & Norton, 2015;Norton, 2013;Norton & Toohey, 2011;Norton Peirce, 1995) have emphasised the complexity and fluidity of people's identifications with their languages.Indeed, Norton Pierce's concept of investment (1995) was groundbreaking insofar as it was a theory of social identity which foregrounded the role of context and addressed how power relationships affected language learner perceptions and actions.As learner experiences have been of central importance for researchers working within such sociocultural and poststructuralist frameworks, linguistic identity in the field of second language acquisition (SLA) has been operationalised largely through qualitative methods that allow for the first person perspective implicit in an epistemology that centres the subject as 'knower' and the identity space as a site of struggle, and have mainly generated case studies focusing on individuals' lived experience and on learning stories.
Social psychologists, on the other hand, have generated a large body of work on the self construct where beliefs and emotions have been researched using, in the main, survey instruments.Mirroring moves in SLA over the last few decades, however, research on the self in relation to motivation has seen a shift from psycholinguistic towards sociocognitive approaches.Dornyei's L2 Motivational Self System, for example, built on work by Markus and Nurius (1986) to make strides in modelling and measuring the impact of mental representations of possible ideal and ought-to selves on motivated learning behaviour and, more importantly, attempted to understand their relationship with the L2 Learning Experience (Dörnyei, 2009).This addition of situation-specific motives related to the learning environment has begun to answer the call for a person-in-context relational view (Ushioda, 2009), where context is no longer merely an independent variable external to the individual but integral to and shaped by the person.Methodologically, this entailed a move away from motivation as measured almost always by survey instruments to more mixed methods, including interview and observation (see Boo et al., 2015, for an overview).Moreover, moves to pay more attention to the relationships between individuals and their social contexts have resulted in the growth of interest in complexity theory, where some researchers have begun to consider the self in relation to a complex nested system of beliefs and emotions situated in contexts which are integral to the system (e.g.Henry, 2017;Mercer, 2016aMercer, , 2016b)).
In line with the recent focus on multilingualism as a phenomenon (see Aronin & Laoire, 2003;Jessner, 2006Jessner, , 2008)), identity research has begun to be extended to considerations of the multilingual self and multilingual identity (Fisher et al., 2020;Henry, 2017), moving away from the monolingual mindset of previous research (Henry, 2017).Researchers have typically ascribed the term 'multilingual' or indeed 'bilingual' to the individuals under consideration, though how such labels are arrived at is often opaque and varies greatly.While the term is usually linked to proficiency or experience, researchers acknowledge that it is more than these variables (Aronin & Laoire, 2003;Henry, 2017).On the other hand, as researchers we can identify our participants in line with Blommaert's conception of linguistic repertoire (2010,2013), and ascribe the term 'multilingual' not only to proficient bilingual or multilinguals, but also to 'monolingual' speakers who are beginning to learn a foreign language in school (and who, consequently, may have relatively low levels of proficiency in this language), or who have knowledge of dialects and varieties of language and non-verbal forms of communication such as sign languages (see Fisher et al., 2020).It is clear that a number of variables come into play when people decide to ascribe or claim for themselves the label of multilingual.Which other variables are implicated is as yet unclear, though as we outline below, certain constructs such as emotions, beliefs and attitudes are likely implicated and we turn to these now to explicate the choice of latent variables for testing in our study.

Modelling multilingual identity: latent variablesExperience
Language experience is concerned with repertoires, affiliation (Cook, 2008) and with a variety of historical, contextual and social factors (Block, 2009;Norton Peirce, 1995).Clearly we can presume that, past and present, learners' exposure to and interaction with languages in the home, classroom, on travels, in the community or digitally are highly likely to influence their identification as multilingual.Indeed, authors have noted that the opportunities learners have for engaging with different languages in their repertoires affect their understanding of and identification with their own multilinguality (e.g.Ceginskas, 2010;Fielding, 2016) .We argue that, furthermore, past experience will also play a significant role.As can be seen in Figure 1, we therefore include Experience as a latent variable to test its relationship to multilingual identity.

Evaluations
Also potentially important in any identification as multilingual are learners' beliefs and attitudes towards their languages, towards language learning and about themselves as language learners.Beliefs in general have been judged to have an important influence on learning (Chik, 2018;Mercer & Williams, 2014) and self-beliefs, self-efficacy, for example, is a key component of learner identity (e.g.Trujillo & Tanner, 2014), and relates closely to learning and learning outcomes (Alivernini & Lucidi, 2011).Based on research suggesting that learners understand the importance of languages, but that this makes little difference to their learning choices when they are able to choose whether to continue studying them or not (Coleman et al., 2007;Fisher, 2001), we test the idea that beliefs about language (for example, about their importance) will be different from beliefs about self in relation to language (for example, self-efficacy beliefs).We therefore include two separate latent variables: beliefs about language and beliefs about self in relation to language.
Moreover, given the situated and sociocultural dimension of identity, all individuals' conceptions are liable to be influenced by those around them.Taylor (2013), for example, in her work with adolescent language learners found that family, friends, classmates and teachers were all important in shaping adolescents' identities.Operating within a number of communities of practice (Lave & Wenger, 1991;Wenger, 1998), learners' interactions with others and how they internalise those interactions will influence their own implicit and explicit beliefs about languages and about themselves in relation to their languages.Essentially, and related to social identity theory (e.g.Tajfel & Turner, [1986] 2004), learners are engaged in 'self categorisation' and 'social reality tests' (Turner, 1991, p. 16), where they are trying to work out their own beliefs in relation to those around them and evaluating whether what they think is endorsed by others.To this end we include a latent variable of 'community beliefs' comprised of teachers', friends' and parents' beliefs as they relate to learners' own beliefs and attitudes.

Emotions
This brings us finally to our last major construct of Figure 1, that of the affective dimension of language learning.In particular, we view learners' beliefs about themselves as language learners to be inherently linked with emotions, a construct that has gained more research attention in the last few years in SLA, following the 'affective turn' (Pavlenko, 2013).While researchers such as Dewaele and MacIntyre (2016) and Dewaele et al. (2018) have noted the importance of positive psychology in language learning, emotion is an important construct for researchers in all areas of identity research too.For example, Weedon's (1987) definition of subjectivity, 'the conscious and unconscious thoughts and emotions of the individual, her sense of herself and her ways of understanding her relation to the world' (p.32), sees emotions as having a central role in our constructions of the world.Findings on emotions and multilinguals have revealed a variety of emotional responses to the languages in one's repertoire (Dewaele, 2013;Pavlenko, 2005Pavlenko, , 2006)).Usually this is in relation to shifts in 'language selves' and whether individualsthey feel they become different people in each of their languages (Pavlenko, 2006).How emotion relates to oneself as a user and learner of a wider range of semiotic resources and how this is related to an identity as multilingual remains to be explored.
To conclude, we conceive of multilingual identity as an orientation that reflects a closer affinity to being or becoming multilingual, and is more concerned with an individual's relationship with their linguistic repertoire than it is with the exact constitution of this repertoire, though the two may well correlate.We theorise that it is shaped: by one's language experiences (past and present); by one's evaluations (beliefs about language, beliefs about oneself as a user and learner of language, perception of others' values); and by emotions.This is expressed contextually through attitudes, behaviours, emotions and future life scenarios.In this 3Es conceptualisation (Experience, Evaluations and Emotion) we bring together the historical/contextual, the social/relational and psychological/intramental dimensions of identity development discussed elsewhere (see Fisher et al., 2020) forming the basis of the theoretical model of multilingual identity, presented in Figure 1, which we have tested on our data.
Our research questions therefore are: . Which latent variables comprise the theoretical construct of multilingual identity?
. To what extent is the proposed 3Es model (Experience, Evaluation and Emotion) appropriate for various sub-groups of learners?
Block (2013) notes that most work on language and identity is inspired by poststructuralism and adopts a social constructivist perspective, where identity is concerned with 'the multiple ways in which people position themselves and are positioned', but with a focus on the social, not on the psychological (Block, 2013, p. 129).This is reflected in methodologies where case studies built from interviews and reflective writing are the norm.Surveys do not lend themselves to current trends in identity research as they are felt to be a form of essentialism, that is, 'the philosophy behind labelling any number of normative characteristics or practices as constituting the core of an individual or group which are then used to define them and held to be true of all members of the group' (Omoniyi, 2006, p. 16).The perceived danger for identity research is that of conceptualising identity as static, as product rather than process.However, strategic essentialism (Bucholtz, 2003) can be useful, for instance, to identify undescribed groups and so effect change, in line with Block's (2013) call for more research into the identity inscriptions of social class and socioeconomic stratification, for example.
We have noted elsewhere the need for research in multilingual identity to draw on a variety of fields of scholarship (Fisher et al., 2020) with their own epistemologies and variety of methodological approaches, according to the demands of the task.Here, we see the development of a model for multilingual identity, based on survey data, as serving different purposes, such as helping researchers in systematic data collection and comparison across groups of learners or for teachers to help students develop agency to choose such an identity and possibly to develop a shared identity that did not exist before.In their parallel study on Norwegian learners' multilingual identity, Haukås et al. (2021) have similarly developed a questionnaire exploring language use habits, beliefs about multilingualism, future multilingual self, open-mindedness and other factors.To agree, therefore, with Joseph, ' … there must remain space for essentialism in our epistemology, or we can never comprehend the whole point for which identities are constructed'.(Joseph, 2004, p. 90).

Context and participants
The data presented in this paper stem from a larger 2-year-long research project involving over 2000 students in Years 8-9 (age 12-14) and Years 10-11 (age 15-16) from seven state-funded secondary schools across the East of England and London.These were selected to represent a range of geographical locations (urban/rural) and student demographics (e.g.first language background and socioeconomic status).For this research project, we collected student demographic and attainment data from schools, as well as student self-report data through questionnaires, interviews and drawings designed to capture students' identification along the three interrelated dimensions of multilingual identity outlined above; that is to say, students' experiences with languages, and their evaluations of and emotions relating to language learning and themselves as users and learners of languages.Here we focus on the questionnaire data collected from 1338 Year 8 students at the start of the longitudinal project to evaluate our theoretical model of Multilingual Identity.These Year 8 students are in their second year of secondary school language learning, with generally patchy experiences in the primary school.Written informed consent via letter was sought from participants and their parents.The characteristics of the sample are presented in Table 1.
We used the total sample from all seven schools for most of our structural modelling procedures; however, the complexity of a model capturing a multifaceted phenomenon like multilingual identity, as well as the comparison across groups and contexts inherent to our analyses, means that the exact sample sizes and compositions differed for the models explored (see data analysis and results below).We assigned the first four schools (1-4 in Table 1) to the category of 'Rural' and schools 4-7 as 'Urban' based on their geographical location and the diversity of their intake.Schools in the rural/homogeneous group tended to have a less ethnically and linguistically diverse intake of students.To assign students to the category of EAL or non-EAL we transformed students' responses on the self-reported first/native language item into two scores, namely, zero for 'English only' (i.e.non-EAL) and one where students identified either a language other than English or both English and another language as their first or native language (i.e.EAL).To assign students to groups according to socio-economic status (SES) we drew on two indicators: the participating schools' data on Pupil Premium (PP) and the parental education levels reported by students on the questionnaire.PP is a yearly sum of money the UK Government has since 2011 given to schools to improve the attainment of disadvantaged children, and so is an important SES indicator.We awarded a score of zero (i.e.'No') if (a) the student did not receive PP and (b) had at least one parent/carer with a university degree; however, they received a score of one (i.e.'Yes') for either PP receipt or having parents/carers without a university degree.While neither indicator is perfect, a composite variable of these measures is likely to be a more accurate SES indicator than each alone.

Instruments
The inherent difficulty in researching multilingual identity is in part due to the need to conceptualise multilingual identity as a latent construct; that is to say, a phenomenon that is not directly observable or measurable.As outlined above, we conceptualise multilingual identity as shaped by students' experiences of languages (i.e.their exposure to and use of languages both in and out of the home), by both their own and others' evaluations of languages and of themselves as language learners, and by students' emotional responses to their own language learning.The data used for our structural modelling were collected through a questionnaire designed to collect information on these three 'Es' of Multilingual Identity.
Each of the latent variables in our theoretical model was measured by several fivepoint Likert scale questionnaire items, with the exception of the latent variable of 'language emotions', which was measured by three distinct question types, due to the inherent complexity of researching and measuring emotions (Zembylas, 2007).In addition to a five-point Likert item in which students were asked to respond to the statement 'I like learning other languages', we measured students' sense of pride in relation to each of their additional languages.This was done by first inviting students to list any languages that they knew, and then asking them to indicate (by ticking a box) whether they were proud to be able to speak this language.These data were then used to calculate an overall 'multilingual pride score' by dividing the number of additional languages that students were proud of by the number of additional languages listed.Finally, the questionnaire captured students' emotions relating to their language learning through a metaphor elicitation task and two sentence completion tasks, as more indirect and open-ended means of accessing learners' emotions (Fisher, 2013).Students were asked to complete the metaphor statement 'Learning a foreign language is like … because … ', and the two sentences 'When I speak in a foreign language I feel like … because … ' and 'When I'm in the foreign language classroom I feel like … because … ', in an open-ended way.Students' responses on these open-ended items were coded as either positive, negative or neutral, resulting in a 3-point measurement scale to complement the other scale items (Table 2).
For our criterion measure we used a multilingual scale in the form of a simple 'Visual Analogue Scale' (VAS) item.Students were asked to put a cross on a continuous 100 mm straight line, where the labels 'monolingual' and 'multilingual' were situated at the endpoints (see Figure 2) with space for students to explain why they placed their cross where they did.VASs have been shown to be useful for measuring perceptions, attitudes and constructs with affective components (De Boer et al., 2004) and avoid the clustering found with Likert scales (Hayes et al., 2013).The scale does not capture students' multilingual identity per se, but their willingness to state that they consider themselves to be multilingual.Here we hypothesised a strong positive correlation between the latent variable of multilingual identity and students' scores on the scale.

Procedures
The aim of the analysis was to test the fit of our theoretical model to the empirical data, as well as to establish whether the model holds: for learners of different language backgrounds, in this case speakers of English as an additional language (EAL) who are also learning foreign languages at school; for English as first language speakers learning a foreign language (non-EAL); and for learners situated in linguistically and culturally different school environments.The analysis procedure consisted of three main steps.
Firstly, structural equation modelling (SEM) was employed to test the fit of our theoretical model to the overall data set, using the lavaan package in RStudio version 1.2.1335.Indices most often advised in the SEM literature were used to assess model fit, and we report the Comparative Fit Index (CFI), the root square mean error of approximation (RMSEA), the SRMS, as well as chi-square and CMIN/df statistics (Hooper et al., 2008;Kline, 2016).An analysis of the modification indices was then used to make any modification to increase the fit of our theoretical model, only adding suggested parameters when these were theoretically sound and meaningful.ANOVA analyses were conducted

Language Experience (latent variable)
. Multilingual international experience: aggregate score of (a) the number of non-English speaking countries lived in and (b) visited on holidays .Multilingual social ties: whether other languages were spoken within (a) their family and (b) friend circles .Number of languages being learned at school .Usage of multiple languages outside the school, family and international context, via (a) music, (b) books, and (c) other media (aggregate score for all languages divided by number of languages listed).
Language Beliefs (Evaluationslatent variable) -5-point Likert scale items . I think that learning other languages is important .I think that learning other languages is pointless because everyone speaks English .I think that learning another language helps me understand more about other cultures .I think that learning another language helps me understand more about my own culture .If I were to travel to another country, I would like to be able to speak to people in their language .I think that it's cool to be able to speak other languages

Language Self
Language Self-beliefs (Evaluationslatent variable) -5-point Likert scale items on the fit statistics of the different models (i.e.without and with modifications) to establish whether each of the modifications added to the model increased model fit significantly.
Secondly, multi-group SEM was then used to test (a) whether learners in the subsets of data (i.e.'EAL' versus 'non-EAL', and 'rural/homogeneous' versus 'urban/heterogeneous' catchment areas) ascribe the same meaning to the theoretically-specified multilingual identity constructs, and (b) whether the relationships between multilingual identity constructs and variables have the same strengths for EAL and non-EAL learners, as well as for learners in more urban/heterogeneous and more rural/homogeneous catchment areas.This second analysis was done using the R packages lavaan and semTools.This analysis followed a systematic, step-by-step statistical procedure to test for measurement invariance, and to identify exactly at what level the model diverges for the groups of interest.The specific levels analysed were: the latent variable structure, the latent variable loadings, the intercepts, the residuals or spread on the manifest and latent variables, the means on the latent variables, and the regressions specified in the model.The significance of any decrease in model fit when constraining the parameters to be equal at each level were assessed using the CFI, adopting the rule that a drop of Delta(CFI) > .01 between consecutive models is indicating a substantial deterioration in model fit (Cheung & Rensvold, 2002).By adopting this procedure, it was possible not only to test the fit of our theoretical model to data obtained from students with different language learning backgrounds as well as from different school environments, but also to understand exactly where differences between students from different language backgrounds and environments may lie.

Results
The final multilingual identity models As a first step, our initial theoretical model was evaluated against the full Year-8 data set.Figure 3 presents the schematic representation of the final model with the standardised estimates.
It was found that the theoretical model described above provided acceptable model-fit data (CFI = 0.842, RMSEA = 0.058 (90% CI = +/-5), SRMR = 0.078), but that fit could be improved by three modifications at the measurement level of the model (i.e. the latent variable measurements).ANOVA analysis on the fit statistics of the two models (i.e.without and with the regression between 'language self-beliefs' and 'language emotions') revealed that a regression between self-efficacy and emotions significantly increased the model fit (Table 3).
Next, the fit of the modified model was evaluated simultaneously for the two grouped samples (i.e.EAL/non-EAL and Urban/Rural), and it was found that our theoretical model decreased substantially in fit for the joint model-data for each grouping of interest (see Table 4).
These decreases in fit suggest that, while providing an acceptable fit when taking an all-inclusive conceptualisation of multilingual identity, the model seems to less accurately capture the complexity of multilingual identities associated with more specific and narrower notions of multilingualism.Fit measures for the model as applied to the separate data sets for each grouping (results not shown) revealed that our model had a considerably better fit for non-EAL students and for more homogeneous school settings.This suggests that identifying and understanding the multilingual identities of students with a home language background may require a more comprehensive or more nuanced instrument in order to accurately capture the nature and dimensions of their multilingual identities.The multi-group SEM analysis was able to shed further light on this, revealing that the difference in model fit between the overall model and joint EAL/non-EAL model was related to a greater spread on the questionnaire items and latent variables within the EAL group (i.e. the residual level) as well as differences in the mean scores for the latent variables.There were no statistically significant differences between the two groups in relation to latent structure, latent loadings or regression paths.In other words, the overall structure of the model and the path loadings are statistically the same for EAL and non-EAL students.However, EAL students displayed much greater divergence in their answers, suggesting that they are a significantly less homogeneous group than non-EAL students, and that identifying and understanding the multilingual identities of students with a home language background may require a more nuanced instrument.
Interestingly, the model has a slightly better fit for the grouping by school environment than by home language background, suggesting that there are other socio-cultural factors that go beyond home language and language experience and that may make a school population see themselves as more or less multilingual, even if their multilingual identities are different in nature.Alternatively, there may be an interaction between EAL population and School population, whereby rural and urban schools attract EAL students of a different social and cultural background.With regard to the poorer fit of the model for urban/heterogeneous and rural/homogeneous groups, the multi-group SEM analysis reveals that significant differences between urban/heterogeneous and rural/homogeneous school populations are located at the level of the latent means only.In other words, while the two groups differed significantly in their mean scores on the latent variables, there is no difference in the clustering of the variables: the same structure emerged for both groups (i.e.related to the path coefficient).
Figure 4(a,b) present the schematic representation of the final models for the EAL(a) and non-EAL(b) students respectively.
Comparing the paths of the structural models for the EAL and non-EAL students reveals a stronger loading of language experience onto multilingual identity for EAL students than for non-EAL students.Non-EAL students seem to be more strongly influenced by community beliefs in their language beliefs and language self, with non-EAL students' beliefs also being predicted more so by friends' beliefs.Nevertheless, in both groups, parental beliefs seemed to have the strongest influence on students' own beliefs and their language self.Our data thus suggest that, with regard to the social influence on an individual's beliefs and feelings about languages and themselves as language learners, parents have by far the strongest influence within a learner's direct community.For both EAL and non-EAL students, multilingual identity is most strongly expressed in students' language self (i.e.self-beliefs and emotions).
Figure 5 (a,b) present the schematic representation of the final models for students attending schools in urban/heterogeneous catchment areas and in rural/homogeneous catchment areas respectively.
When comparing the paths of the structural models for the rural and urban catchment areas students, there are no obvious differences between the two groups on the correlations and loadings.This was further supported by the measurement invariance analysis: while it found significant differences at the level of latent means, it found no statistically significant differences at the level of latent structure, latent loadings, intercepts, and residuals.

Discussion
In our study Multilingual Identity is neither viewed as a fixed state nor as pertaining to specific linguistic identities (whether relating to the participants' mother tongues or to the particular foreign languages they were learning).Rather, we have argued that the phenomenon is an effect of the dynamic interaction of sources of beliefs, experiences, and affective dispositions towards linguistic identity-related diversity.Our analysis aimed to distinguish the relative degrees of strength of a number of key latent variables in mediating learners' construction of their multilingual identity.The model we arrived at offers a holistic representation of the network of influences that inform learners' perception of their multilingual identity.
We believe that the following interpretations are supported by the results of our analysis.An important caveat to bear in mind is that while the first three findings listed below are supported by our SEM analysis, the fourth and fifth are to be taken more as suggested trends since the statistical differences between the groups were not significant.These interpretations, which in some cases relate to the overall data and in others to a comparison of the EAL and non-EAL data, reflect important issues in theoretical and educational approaches to the relationship between language learning and multilingual identity.
(1) Multilingual Identity is mainly expressed through the influence of Language Experience and Language Self The definition of multilingual identity through an analysis of the relative strengths of sources of influence is a complex endeavour and further constrained by the inevitable limitations of any modelling instrument.We do not claim that our model is based on a comprehensive account of all the potential factors involved in the construction of learners' multilingual identity.However, within the framework of the latent variables we have focused on in our modelling, the indications are that Language Experience and Language Self are more strongly associated with Multilingual Identity than is the construct of Language Beliefs.
Direct experiential contact with languages and with language-focussed social interaction, whether in the classroom or on holiday or in other personal and social contexts, seems to have a stronger correlation with the expression of multilingual identity than explicit beliefs, whether one's own or those of others, about language learning.While we are not claiming that the latter have no impact, the relative degree of correlation is, however, noteworthy.This suggests that beliefs about other languages and explicit positioning regarding the learning of other languages (which, the model also tells us, are largely informed by 'community' beliefs in the shape of parental, peer, and teacher beliefs) are less decisive than exposure to the language and beliefs and feelings about one's own linguistic behaviour and competence.It may be that the contrast here is also partly between the impact of cognition-oriented belief, or 'understandings', (e.g.community views about the relative importance of multilingualism), on the one hand, and emotion-oriented beliefs, or 'dispositions', (e.g.how one feels about one's multilingual potential), on the other.Learners' encounters with other languages (through, for instance, their international experience, social ties, or language usage and learning) as well as their self-efficacy beliefs and emotional dispositions towards other languages are key contributory factors in shaping their multilingual identity.Our analysis therefore echoes Lantolf and Swain's (2020) underlining of the interdependence of experience and emotion encapsulated by the Vygotskyan concept of 'perezhivanie' in L2 development.Applying this concept to the experience of the languages classroom, the authors comment that 'the emotional component of perezhivanie is the affective reaction of the person refracted through a particular environment/person dialectic, while the intellectual component is the evaluation of the reaction.Together they potentially result in new practical behavior ' (p. 121).Transferred to the domain of expression of multilingual identity, our analysis suggests that this combination of variables is likely to result in the development of learners' 'possible identities' (Oyserman & James, 2011) through their positive emotional self-perceptions as future multilingual speakers.
(2) Community Beliefs are strongly correlated to Language Beliefs and therefore have a weaker correlation with Multilingual Identity (which is more strongly tied to selfefficacy and emotional disposition represented by the construct of Language Self) One of the implications of this finding is that, for this age group at least, the dynamic is more than just a reversal of the perspective of a 'sociolinguistics of identity', which focuses on 'the ways in which people position or construct themselves and are positioned or constructed by others in socio-cultural situations through the instrumentality of language' (Omoniyi & White, 2006, p. 1).In the context of multilingual identity construction, it does not seem, on this evidence, that community beliefs and practices alone are influential but that it is more likely to be informed by the subjective play of reflexivity and affective disposition.In self-efficacy conceptual terms, a parallel contrast is between the 'solipsistic self' ('construed as autonomous and attuned to internal goals, thoughts and motives') and the 'contextualized self' (for whom 'in-group goals take priority over personal goals') (Pilotti & El Alaoui, 2018, p. 102), although in our model the perspectives are not mutually exclusive but defined by differences of degree.
(3) Parental Beliefs have a stronger correlation with Multilingual Identity than do Teachers' or Friends' Beliefs.
The general education literature has researched the effect of parental influence on their children's academic achievement (Paulson, 1994;Schlechter & Milevsky, 2010;Spera, 2006) and on their educational aspirations (Kirk et al., 2011).In the context of foreign language learning, Bartram (2006) examined parental influence on their children's attitudes to language learning, and Iwaniec (2020) has investigated parental influence on children's motivation to learn a language.Researchers have pointed to socio-economic variables affecting positive or negative parental influences on children's language learning.Equally, national contexts may also be a determining factor: for instance, parental beliefs in urban China about the value of learning English as a foreign language will broadly contrast with those of parents in rural England with regard to their children's study of French.However, to the best of our knowledge, no empirical study has as yet focused on the impact of parental beliefs on children's perceptions of their multilingual identity.Yet the strong correlation between the variable of perceptions of parental beliefs and the learners' own beliefs about languages and language learning in our study supports the claims in the literature (e.g.Chen et al., 2005;Spera, 2006) regarding parental impact on motivation and attitudes.However, what our structural model does not show is that this affects their multilingual identity perceptions in a salient way.
(4) Language Experience has a stronger correlation with Multilingual Identity in the case of EAL students than non-EAL students.
Researchers in the field of multilingualism and identity in the context of migrant background schoolchildren have pointed to the heightened experience of negotiation of identity (Pavlenko & Blackledge, 2004) through the development of 'transcultural identities' (Suárez-Orozco, 2004) or, for instance, through 'ascribed identities' whereby they are categorised as belonging to particular identity affiliations either by their home community or by members of the host community.This experience of indexing of identity heightens the individual's consciousness of the diversity of identities which can potentially be assumed (Evans et al., 2020).Associated to this, the 'ability to code-switchto move fluidly between languages and cultures' (Suárez-Orozco, 2004, p. 1) enables the process of negotiating new identities, as the migrant student increases his or her mastery of the host language.The stronger correlation between language experience and multilingual identity with our EAL participants than with our non-EAL participants aligns with this view of the role that language plays in EAL children's identity construction.
(5) Community Beliefs have a stronger influence on Language Beliefs and Language Self in the case of non-EAL students than of EAL students.
While showing equal dependence on the influence of parental beliefs, EAL and non-EAL responses revealed a noticeable divergence in the impact of friends' beliefs, with a stronger influence in the case of non-EAL participants in the study.This would suggest that in terms of their views about languages and language learning, EAL students are more independent in the formulation of their declarative thinking and evaluation of this issue.This could be due to their stronger biographical anchoring in a lived experience of bilingualism and therefore less in need of external direction for their beliefs about the value of language and language learning.The comparative correlation of non-EAL students' language beliefs with the influence of friends' beliefs echoes the conclusions of research into peer influence on learning, attitudes and motivation of foreign language learners in the UK and other western countries (Bartram, 2006;Kormos et al., 2011).While researchers have argued that peer and parental influences play a strong role on learners' self-related beliefs and L2 motivation, what our model indicates is that the trajectory towards the formulation of their multilingual identity is not so much via the pathway of the learners' language beliefs but more through the reflexive and affective prism of their self-beliefs and emotional disposition towards languages.

Conclusion and further research
The research reported in this paper represents a rare attempt to use structural equation modelling in order to test a theoretical model of multilingual identity and to examine the complex relationship of constituent latent variables, in the context of foreign language learning in schools in England (Henry and Thorsen (2018) have applied the method to measure the validity of the ideal multilingual self construct in relation to pupils in schools in Sweden).Our paper also represents the first attempt to model the relations between language experience, evaluations and emotional dispositions in this context, with specific reference to EAL/non-EAL, and urban/rural sub-groups in the sample.
The validation of our theoretical model of the 3Es (Experience, Evaluation and Emotion) legitimises the development of intervention studies and other forms of empirical research on multilingual identity construction in the foreign language learning context.More specifically, the findings of our analysis showing that experience and emotions are particularly influential in multilingual identity construction have clear implications for teachers and policy-makers in terms of opportunities for experiencing languages, whether on visits abroad or in the classroom, and pedagogical approaches which foster learners' positive emotional responses to linguistic diversity.For a discussion of the pedagogical effects of such an approach see our earlier papers (Forbes et al., 2021;Rutgers et al., 2021).The stronger correlation between multilingual identity and language experience in our EAL sample than in our non-EAL sample further reinforces the connection between these two variables.
However, while our model shows some clear divergences in the spread of the data between the EAL and non-EAL subjects in our sample (but not in the structure of the model and magnitude of the correlation between the latent variables), the same is not true of our analysis of the urbanrural data.It may be that, in the seven schools in our study, the school variable over-rode broader environmental influences.Further research drawing on data from a larger number of schools in both groups would be useful to explore the potential effect of this variable on multilingual identity.

Figure 1 .
Figure 1.Schematic representation of a theoretical model of Multilingual Identity.

Figure 2 .
Figure 2. The multilingual VAS item on the Multilingual Identity questionnaire.

Figure 3 .
Figure 3.The final Multilingual Identity model for UK Year-8 students with standardised estimates.

Figure 4 .
Figure 4. (a) The final Multilingual Identity model for EAL students with standardised estimates.(b) The final Multilingual Identity model for non-EAL students with standardised estimates.

Figure 5 .
Figure 5. (a) The final Multilingual Identity model for students attending secondary schools in more urban and heterogeneous catchment areas, with standardised estimates.(b) The final Multilingual Identity model for students attending secondary schools in more rural and homogeneous catchment areas.

Table 1 .
Sample characteristics.For School 2 we were unable to obtain student-level data on this parameter specific to this Year-8 cohort.

Table 3 .
Selected fit measures for the final Multilingual Identity model for UK Year-8 students.

Table 4 .
Joint selected fit measures for the final models for the full data set and for each grouping of interest.