2.1 Adding a multilingual component to the existing body of monolingual age-related research in SLA

So far there has been a ‘monolingual bias’ in age-related classroom research in that scholars have predominantly focused on monolingual early vs late learners, which is regrettable for several reasons. Over 50% of the population across the European Union report being functionally bilingual (European Commission 2012), which means that we have increasing numbers of multilingual children in our schools. In consequence, primary-school teachers in Europe see their classes becoming more heterogeneous (Meijer, Soriano & Watkins Reference Meijer, Soriano and Watkins2003) in terms of cultural backgrounds; also multilingual schools are becoming increasingly common, following the European Union recommendation that children be introduced to two additional languages from an early age (European Commission 2012). Furthermore, in many schools more and more students’ home language differs from the school languages (Muñoz & Singleton Reference Muñoz, Singleton, Singleton and Aronin2018). The paucity of research on the age factor in multilingual populations is also regrettable, since the profile of skill development that has been obtained for monolingual early and late starters may differ in crucial respects from that of children who develop two languages in childhood and establish basic cognitive competencies through the mediation of two languages (for a review, see Bialystok et al. Reference Bialystok, Abutalebi, Bak, Burke and Kroll2016).

Evidence for an interaction between age of onset (AO), bilingualism and biliteracy is provided by Pfenninger & Singleton's (Reference Pfenninger and Singleton2018) longitudinal study, in which we assessed the L2 English development of 636 secondary-school students, who had all learned Standard German and French at primary school, but only half of whom had had English from third grade (age 8) onwards, the remainder having started five years later at secondary school. In addition, 200 of them were monolinguals, 144 were simultaneous bilinguals with biliteracy skills, 107 were simultaneous bilinguals without biliteracy skills, and 185 were sequential bilinguals. These findings suggest that different learner populations (monolinguals, simultaneous bilinguals, sequential bilinguals) are differentially affected by third language (L3) AO effects, due partly to individual differences (e.g. bi/literacy skills and family circumstance), partly to contextual effects that mediate successful L3 outcomes (e.g. language environment at home, classroom effects, influence of peers and siblings). Such studies aim at a deeper understanding of the potential benefits of early foreign language (FL) instruction by taking account of the multilingual situation in schools, the hybridity of multilingual experiences and the multilingual policies relating to students’ starting age.

Pfenninger & Singleton's (Reference Pfenninger and Singleton2018) study sought to address two research questions: (1) How does FL age of onset affect different learner populations in the short and the long run? (2) What is the relationship between parents’ behaviours, attitudes to and beliefs about English as a foreign language (EFL) and their children's bilingualism/biliteracy as well as English learning? Both of these questions and their associated research design could well be adapted to focus on other language configurations and cultural backgrounds. The aim would be to provide the basis of a finer-grained distinction between monolinguals, simultaneous and sequential bilinguals, and earlier and later bilinguals, in order to establish whether different learner populations are differentially affected by age of FL acquisition onset. For instance, while in Pfenninger & Singleton (Reference Pfenninger and Singleton2018) language typology is not a variable under analysis, one might conceivably wish to analyse the same variables in a comparable socio-educational context that involves typologically different languages. This would take account of the open empirical hypothesis that the relationship among the languages involved might also account for differences in the results of research on the effects of bi/literacy on L3 learning. Furthermore, Pfenninger & Singleton's study does not operationalize level of bilingualism either. While we included only children who could be considered functionally bilinguals, the latter undoubtedly had slightly different levels of oral proficiency in their different languages. It would be fruitful to analyse whether a more balanced level of proficiency and use of two languages within the bilingual groups affects L3 learning outcome. Yow & Li (Reference Yow and Li2015), for instance, found a significant effect of balanced usage and proficiency in two languages on certain components of executive functioning – in addition to the impact of age of L2 acquisition.

2.2 Looking at the respective roles of AO and typological difference in the acquisition of two additional languages

It has long been argued in some quarters that age is less important for proficiency-attainment level than the fact of prior knowledge of another language (Grosjean Reference Grosjean and Harris1992; Cook Reference Cook, Singleton and Lengyel1995). It has also been noted that the degree of distance between the first language (L1) and L2 in question may play a role – not only relative to diverse aspects of L2 performance (see e.g. Ó Laoire & Singleton Reference Laoire, Singleton, Aronin and Hufeisen2009) but also with respect to starting age outcomes. In regard to differences between the effects of starting age on attainment in respect of typologically different additional languages, McDonald (Reference McDonald2000) found, for example, that learners of English from a Spanish-speaking background exposed to the language before age five were able to perform to native levels on an English grammaticality judgement test, whereas Vietnamese speakers with pre-age-five experience of English were not. Also, simultaneous acquirers of two very distinct languages have been found to acquire certain features of one language before they acquired these features in the other language. Thus, a number of studies (e.g. Schelleter, Sinka & Garman Reference Schelleter, Sinka and Garman1997) looked at children acquiring, respectively, Latvian and German (both highly inflected) alongside English (not a very inflected language). They found evidence of the development of functional categories in Latvian and German from the earliest stages, but not in English.

To explore the starting age question while at the same time taking account of the typological difference question a study-design is required involving (1) a set of learners with the same mother tongue (say, English), who have started two similar additional languages (e.g. Spanish and Italian) at different ages (say, Spanish at 8; Italian at 13) and (2) a set of learners with the same L1 as the first group, who have started two very different languages (e.g. Mandarin Chinese and Russian) at similarly different ages (say, Chinese at 8; Russian at 13). The rhythm of learning in terms of weekly exposure time would need to be constant across all groups. Subject search would not be unproblematic, but the essential point is that the pairs of languages should be of different degrees of typological/lexical relatedness. If the earlier-learned language were the same in both cases, this would not be a problem. There are certainly countries (like Ireland) where one L2 (Irish in this case) is introduced in the primary school and others (of a range of typological/lexical distances from Irish) in the secondary school. Any long-term advantages discovered for attainment in the earlier-started languages might be attributable to AO and consequent length of exposure. Any major differences discovered between the starting age effect in the two pairs of languages could be related to language differences; in other words, if in the above instance Italian and Spanish proficiency were markedly closer in terms of attainment than Russian and Chinese proficiency, typological/lexical differences would be obvious explanations to invoke.

2.3 The interaction between AO and intensity of instruction

Analysing whether age effects vary across different educational programmes is at the heart of the third research task. The question has arisen as to what crucially influences the FL mastery of young classroom learners, and how we can exploit an earlier starting age more effectively. Attention has focused on intensity of instruction: myriad classroom studies (see e.g. Muñoz Reference Muñoz2012) have demonstrated that intensive exposure in the classroom is a more determinant predictor of FL outcome than starting age and regular drip-feed instruction with a few, spaced lessons. However, time is one of the most contested pedagogical resources; accordingly, it is difficult to increase the student allocation of hours for FLs. This is where bilingual language instruction such as CLIL comes in. CLIL is a method in which school subjects are taught in a target language different from the mainstream school language, thereby extending FL exposure and providing motivation for purposeful communication (see e.g. Dalton-Puffer & Smit Reference Dalton-Puffer and Smit2013). The amount of FL exposure in CLIL programmes is much greater than in traditional formal instruction, with between 50% and 100% of the content being taught via the target language (see Cenoz, Genesee & Gorter Reference Cenoz, Genesee and Gorter2014). The expansion of the range of types of CLIL (pre-)primary schools calls for critical discussion of the role of the age factor in early bilingual education of this type. Both starting age and CLIL have been explored, albeit not in interaction. The optimal starting age question is of particular interest in immersive school contexts, in which school subjects are taught in a language different from the mainstream school language (with varying amounts of target language exposure), since more intensive CLIL programmes constitute a hybrid between naturalistic L2 acquisition (where we find a general ‘earlier = better’ trend) and instructional FL contexts (where no age effects in favour of early starters emerge).

This research task can include CLIL programmes from a continuum of learning contexts, with minimal CLIL standing at the instructed end of the continuum, total immersion towards the naturalistic end and partial immersion somewhere in between. Participants with different starting ages within a CLIL programme need to be recruited, the main interest being when the optimal moment to introduce a CLIL programme may be. It is important to keep in mind that such a study should not aim to focus on comparisons across programmes (e.g. CLIL vs non-CLIL students). Inter-programme comparability is known to be very problematic, since intensive CLIL students receive considerably more hours of instruction – a well-known confounding variable (see Tragant et al. Reference Tragant, Marsol, Serrano and Llanes2016) – and entering the more intensive CLIL programmes is often voluntary. The goal would be to focus on age effects within programmes (e.g. within a full CLIL programme). Early vs late CLIL students can be assessed quantitatively at different points in the programme. In such a study, where sample size is likely to be limited, it might be useful to conduct a dynamic analysis of linguistic data with the goal of testing whether our set of predictors (AO, CLIL) has a significant effect on the trajectories under investigation (see also Research task 4). The goal would therefore not be to recruit as many participants as possible but to collect learning trajectories through dense time serial measurements at equal intervals (e.g. 30 measurements per participant) before using Generalized Additive (Mixed) Modeling (GA(M)M) – e.g. by means of the mgcv package (Wood Reference Wood2006) in R (R Development Core Team 2016) – which can deal with any kind of data structure (continuous, categorical or count) and can model any type of non-linearity (Winter & Wieling Reference Winter and Wieling2016), thus being applicable to many different kinds of datasets within applied linguistics in general and SLA in particular. If we want to make general claims about L2 development with such a research design, the time-series data becomes the target of inferential statistics. That is, we may consider each set of clusters in the research design as a sample from a population of clusters that we wish to generalize upon, rather than generalizing from sample to population (the traditional basis of inferential statistics). Another advantage of GAM is that it runs on unaveraged data und thus works around the problem of discretizing continuous variables such as age by carving it up into several levels (e.g. children aged 3–6 in one category and 7–9-year-olds in another category). In order to take account of the hypothesis that early bilingual instruction will equip students with key socio-affective skills, such as intercultural awareness (see e.g. Wode Reference Wode2004), a qualitative approach would also be useful – focusing on e.g. cultural awareness, internationalization, and autonomy.

Age-related research to date has lacked such a systematic analysis – and this study could add to our understanding of the influence of different degrees of immersion. It would also help to disentangle age effects from learning context. Its findings would also inform educators and parents about when children should optimally begin bilingual education without compromising outcomes or wasting resources.

3.1 Dynamic analysis of linguistic and socio-affective data from third-age L2 learners

While there is increasing research on L2 acquisition in prime-of-life learners, little attention has been paid to older adults and individual differences in their ability to acquire new languages. There are also doubts about whether ‘a phenomenon as complex as adult learning will ever be explained by a single theory, model or set of principles’ (Merriam Reference Merriam and Merriam2001, quoted in Knowles, Holton & Swanson Reference Knowles, Holton and Swanson2005: 1). One reason for this might be the persistence of the ‘younger is better’ view despite overwhelming evidence that (1) there is much more to age than maturation, and that age-related social, psychological and contextual factors may play as significant a role as strictly maturational factors (see e.g. Moyer Reference Moyer2013; Pfenninger & Singleton Reference Pfenninger and Singleton2017); (2) an understanding of the age factor must take into account the complexities of individual differences that interact with, and mitigate, age effects – and potentially increase with age (see e.g. DeKeyser Reference DeKeyser2013); and (3) the brain preserves much of its plasticity even at an advanced age (see e.g. Peltzer-Karpf Reference Peltzer-Karpf and Ahrens2003), allowing it to acquire new skills, such as additional language capacities.

Specifically, little research has been conducted on individual differences among older L2 learners, and this has concentrated on executive function, field dependency, educational level, working background, hearing loss and hearing loss treatment (Giroud et al. Reference Giroud, Lemke, Reich, Matthes and Meyer2017; Pot, Keijzer & de Bot Reference Pot, Keijzer, de Bot and Gabrys-Barker2018; Ramírez Gómez Reference Ramírez Gómez2016). Very few studies – notably Mackey & Sachs (Reference Mackey and Sachs2012), Bak et al. (Reference Bak, Long, Vega-Mendoza and Sorace2016), Cox (Reference Cox2017), Ramos et al. (Reference Ramos, García, Casaponsa and Duñabeitia2017), Ware et al. (Reference Ware, Damnee, Djabelkhir, Cristancho, Wu, Benovici and Rigaud2017) and Kliesch et al. (Reference Kliesch, Giroud, Pfenninger, Meyer and Gabrys-Barker2018) – have investigated the influence of L2 development in the elderly. Therefore, a critical component of the future agenda outlined by Mackey & Sachs (Reference Mackey and Sachs2012) concerns the need to replicate these intervention studies with larger sample sizes and different methodologies to obtain a clearer picture of cognitive capacities that vary as a function of age, but are also influenced by the extent to which intellectually stimulating exercises are included in an individual's lifestyle. We would argue that there is an even greater need to investigate further learner dimensions – in particular, what learners bring to the process in terms of their individual attributes. Our understanding of the L2 learning process in older adults should also take into consideration how previous L2 learning experience, motivational curves, expectancy of success, goal relevance, volition and agency as they affect motivation to learn and to re-learn languages, self-assessment of L2 skills and learning strategies change over the lifespan and their effect on ability to acquire new knowledge, such as a new language.

With research on L2 learning in later life only beginning to establish itself, the first item in this section arises from the observation that (1) there is particularly great inter-variability and heterogeneity among older adults, which makes statistical generalizations across subjects almost impossible, and (2) L2 learning is a complex, non-linear process, i.e. snapshots of states will not tell the full story of the process of adult L2 learning.

In contrast to the intervention studies mentioned above, in which L2 learners in later life were tested only twice or thrice (usually at the beginning and end of L2 training), it is vital to make use of denser time serial measurements in order to focus on micro-development and idiodynamic trajectories, capturing and capitalizing on individuality in the L2 learning process, homing in on variability, non-linear development, and progression in iterative steps (see Larsen-Freeman & Cameron Reference Larsen-Freeman and Cameron2008; Verspoor, de Bot & Lowie Reference Verspoor, de and Lowie2011). This would ideally involve collecting a relatively small time-series data set containing 5+ learning trajectories of L2 learners between ages 65 and 75, each represented by at least 30 measurements taken at equal intervals (e.g. every week). The observations in the data set are the individual measurements, which means that there are numerous data-points altogether – although these are not completely independent measurements due to the temporal structure in the data. Importantly, studies on the impact of L2 training programmes should include active control groups as well as passive control groups (who do not undergo the training but participate in the testing) to be able to test the efficacy of the training and also to dissociate the effects of different cognitively demanding types of training (Ramos et al. Reference Ramos, García, Casaponsa and Duñabeitia2017). Participants, who will voluntarily sign up for the language course, will be selected after a careful examination of their language background, physical fitness and hearing ability, among other criteria (see Kliesch et al. Reference Kliesch, Giroud, Pfenninger, Meyer and Gabrys-Barker2018). Those in the treatment groups will receive a period of intensive English training – see e.g. Knowles, Holton & Swanson (Reference Knowles, Holton and Swanson2011, quoted in Ramírez-Gómez Reference Ramírez Gómez2016), who discuss eight variables significant in teaching adult/third-age learners. Language tests will be used to assess skills addressed in the course and therefore expected to improve with time. Ideally, they will tap into language production and reception as well as aspects that seem susceptible to age effects (e.g. morpho-syntax) and those that appear less age-vulnerable (e.g. lexico-semantics). In order to target the socio-affective dimension of L2 learning in a classroom setting, an open-ended questionnaire (or alternatively oral interviews) will be administered so as to evaluate learning experiences, motivation, attitudes, overall well-being, personal goal-setting, different life histories and learning stories, autonomy, present day conditions, general communicative skills, sense of purpose, L2 awareness, anxiety, expectations, self-confidence and learning strategies. Questions that can then be asked of statistical models are (1) when development is statistically significantly increasing (or decreasing) with respect to L2 developmental and socio-affective dimensions, and (2) whether age has a significant effect on the trajectories under investigation. Data from such iterated learning studies will be crucial in establishing learner profiles and will inform educators about such profiles as well as the design of individualized language training.

3.2 Investigate whether starting a language in late adulthood yields cognitive benefits

As discussed in the previous section, the question of how good older adults are at additional language learning has often been addressed rather anecdotally (see e.g. Brändle Reference Brändle1986; Johnstone Reference Johnstone2002). In recent times the investigation of this issue has become more systematic and has prompted another line of investigation, namely, how good additional language learning is for older adults. A not uncommon notion – and indeed motivating factor – among older language learners is that learning other languages is an ‘anti-ageing activity’ (Ryan & Dörnyei Reference Ryan, Dörnyei and Berndt2013: 93).

There is some evidence that challenging mental activities help older people maintain their cognitive functions and also reduce the incidence of negative psychological changes such as depression (cf. Lövden et al. Reference Lövden, Bäckman, Lindenberger, Schaefer and Schmiedek2010). Language learning is frequently referred to in this context. Other studies suggest that people who have experienced L2 learning benefit from maintaining brain plasticity (Bialystok & Craik Reference Bialystok and Craik2010; Bak et al. Reference Bak, Nissan, Allerhand and Deary2014). There are also suggestions that the experience of L2 learning later helps with the retention of memory skills (Lapkin, Swain & Shapson Reference Lapkin, Swain and Shapson1990). Bilingualism has also been referred to in this context as staving off the cognitive decline that often comes with aging and with providing protection against illnesses that hasten this decline, like Alzheimer's disease (see Schweizer et al. Reference Schweizer, Ware, Fischer, Craik and Bialystok2012; Bialystok et al. Reference Bialystok, Abutalebi, Bak, Burke and Kroll2016).

We appear then to have plenty of suggestive clues indicating that additional language learning experience may benefit older individuals from a cognitive perspective, but we need more research, more evidence. We lack, in particular, studies focusing on the effects of L2 learning on older adults who begin the relevant language learning process in old age. One area where it might be of particular interest to look for evidence of the cognitive effects of late L2 learning would be that of working memory capacity. Working memory is thought of as a mechanism responsible for the temporary manipulation and maintenance of relevant information during cognitive operations such as language comprehension. Many see it as the key to the question of the role of ageing in what happens to memory. The issue of whether memory declines in old age is controversial (see e.g. Wingfield et al. Reference Wingfield, Stine, Lahar and Aberdeen1988; Salthouse & Babcock Reference Salthouse and Babcock1991). An interesting question would be whether it can be improved at this stage.

Working memory ‘has traditionally been thought of as a “trait” – a relatively fixed capacity that increases in a predictable, maturationally constrained way as children grow’ (Mitchell, Myles & Marsden Reference Mitchell, Myles and Marsden2013: 155). Recently, however, the view has emerged that working memory is, in fact, amenable to modification through relevant experience and training (Holmes, Gathercole & Dunning Reference Holmes, Gathercole and Dunning2009; Klingberg Reference Klingberg2010). As Singleton (Reference Singleton2017) notes, this change of view has impinged on the way working memory capacity is now perceived and researched within the SLA domain (see e.g. Williams Reference Williams, Gass and Mackey2012), resulting in something of a paradigm shift. For example, Gass & Lee (Reference Gass, Lee, Schmid and Lowie2011) found that two L2 groups at different stages in their university L2 study (first and third year respectively) evinced significantly different L2 working memory scores, which points to a shaping, changing role for formal learning experience in relation to working memory capacity. MacDonald & Christiansen (Reference MacDonald and Christiansen2002) earlier demonstrated a correlation between working memory capacity and L2 practice and experience in general. It would be useful to discover whether similar effects for L2 learning were discernible among ‘third agers’!

The design of this study is simple, but subject search may pose some problems since older learners, at least according to anecdotal evidence, are often shy about participating in research. As large as possible a group of older adults beginning the learning of an L2 in their autumn years need to have administered to them every six months (to minimize practice effects) for two years (or more) a test, which will establish whether working memory is declining, remaining stable or improving. Two years would seem sufficient to at least begin to indicate whether any change is occurring, and not too long to risk too many drop-outs among participants. The age-range for the learners could perhaps be set at 65–70 years. Older adults beginning to learn additional languages at that age are not hard to find. In today's world there is little chance of finding participants who have never been taught an L2 previously, but we should look for learners who are starting L2 learning again after a very long break – say 20 years –, who have not used the L2(s) they learned earlier in the meantime, and who are embarking on the learning of an L2 they have not previously had dealings with. In order to be confident that the teaching is having an appropriate impact we should also look for learners who are being instructed in the L2 at a rate of at least four hours per week but who are not receiving an amount of training which is overwhelmingly strenuous.

As for the test, an instrument such as the Trail Making Test could be used. This test requires the testee to connect sequentially circles containing numbers – in Part A – and containing numbers and letters in Part B (where the participant is required to follow the pattern 1–A–2–B–3–C etc.). The test is scored in terms of the number of seconds taken to complete it. Scores achieved indicate speed of cognitive processing and executive functioning, and provide information on visual search, perceptual speed, working memory and ability to maintain two lines of thought simultaneously (see Reitan Reference Reitan1992; Strauss, Sherman & Spreen Reference Strauss, Sherman and Spreen2006; Sánchez-Cubillo et al. Reference Sánchez-Cubillo, Periáñez, Adrover-Roig, Rodríguez-Sánchez, Ríos-Lago, Tirapu and Barceló2009). Scores on Part B of the test are thought to be a particularly good indicator of executive functioning, since Part B requires multiple abilities to complete it. One of the advantages of the test is that it takes only a short time to complete. The difference between this study and most other studies focusing on third-age cognitive benefits of L2 learning (such as Bak et al. Reference Bak, Nissan, Allerhand and Deary2014) is that in the proposed study the relevant L2 learning will take place in the third age, whereas in the majority of previous studies the operative L2 experience has occurred in childhood or earlier in adulthood. It should be noted, however, that some steps in the direction proposed here have already been taken – see Bak et al. (Reference Bak, Long, Vega-Mendoza and Sorace2016), Ramos et al. (Reference Ramos, García, Casaponsa and Duñabeitia2017) and Ware et al. (Reference Ware, Damnee, Djabelkhir, Cristancho, Wu, Benovici and Rigaud2017).

4.1 Age-within-context

In our project ‘Beyond Age Effects’ (Pfenninger & Singleton Reference Pfenninger and Singleton2017), we were impressed by the extent to which AO was affected as a factor by environmental influences, notably ‘classroom effects’. Classroom effects refer to a complex interplay of individual characteristics – including self-confidence, personality, emotion, motivational resources, degree of control over learning, perceived opportunity to communicate, willingness to communicate, time to learn, and memory capacity – and environmental conditions, such as school location and resources, school socio-economic status, teacher education/experience, topic, task, receptivity to the teacher and pedagogical approach, classroom dynamics and group size, parental beliefs/expectations, time invested by the environment to support learning, the language used by the environment, reinforcement by the environment and material resources (books etc.) (see e.g. Cao Reference Cao2011; Kozaki & Ross Reference Kozaki and Ross2011). It became clear in our project that students under different school context and school climate conditions demonstrate differences in educational attainment not attributable to differences in AO. From this it follows that students within one class and school are more similar to each other than to students in other classes and schools. In the case of age-related research, it could be that one class of early starters may have a teacher that makes performance in that class different from other classes. The crux is that we need to know whether between-group and within-group differences are due to age effects or to something else such as effects of class/school context and climate. It is thus important in institutional environments that age effects are considered in the light of macrocultural and microcultural phenomena which may have a bearing on interpersonal relations that influence, shape, strengthen, or weaken variables, such as motivation, which interact with age (see Pfenninger & Singleton Reference Pfenninger and Singleton2016). Although age-related research has made great strides in tackling these issues, still more sensitivity needs to demonstrated to the contexts of research and to the characteristics and diversity of research participants. We suggest there are two main ways of achieving this.

Taking a bird's eye view, three ‘micro agendas’ seem particularly noteworthy as quantitative age classroom research moves forward:

1. 1. We need to find ways to reconcile large-sample methods with individual-level analyses – to maximize generalizability without sacrificing too much variability.

2. 2. We need to acknowledge the force of models that focus on the learner as embodying a developing autonomous system rather than as simply a generalized hypothetical representative of a larger sample (see de Bot, Verspoor & Lowie Reference De Bot, Lowie, Verspoor, Verspoor, de Bot and Wander2011).

3. 3. We need to increase the ecological validity of our statistical analyses and employ contextually sensitive methodology that takes school context and classroom climate into account.

The question, of course, is how to operationalize such an ecological perspective on the age factor in FL classrooms.

In traditional statistics, variability is smoothed over to reveal general trends, on the assumption that random variability should be ignored (van Dijk, Verspoor & Lowie Reference Van Dijk, Verspoor, Lowie, Verspoor, de Bot and Lowie2011: 60). General linear models – the default foundation for analyses in SLA (Plonsky Reference Plonsky2014) – cannot directly provide information about individual change or capture the complexity of contextual effects on individual learning. Thus, if we run an ANOVA to compare differences between early and late starters in a school context, the model will automatically assume that measurements for a given age group have ‘uncorrelated errors’, e.g. as if all early starters are in the same class, school, family, ethnolinguistic group, neighbourhood, etc. Similarly, all late starters will be regarded as being in the same ‘bin’. Variability is considered ‘noise’ and the result of ‘environmental error’. The problem is that smoothing over variability in this way, by averaging, obliterates the very information that we may need to shed light on L2 development (Larsen-Freeman Reference Larsen-Freeman2006). In addition (see Pfenninger & Singleton Reference Pfenninger and Singleton2017), we may, in so doing, obtain age effects which are illusory.

Cunnings and colleagues (e.g. Cunnings Reference Cunnings2012; Cunnings & Finlayson Reference Cunnings, Finlayson and Plonsky2015; Linck & Cunnings Reference Linck and Cunnings2015) have discussed in detail how multilevel modelling – a subgroup of linear mixed-effects regression models – can offer a fruitful way of examining how individual differences in general may affect L2 acquisition. Multilevel models accommodate the intra- and inter-individual correlation and the variation discussed above, and thus help to protect the validity of statistical conclusions drawn from our analyses. We would argue that age research can be greatly enhanced by such an approach. Since these models can be used for assessing the impact of context-varying factors on individual difference variables, we are able to address external factors and environmental influences that interact with age effects and possibly mediate them.

In a statistical analysis of such data, AO, biological age and/or time are modelled as fixed effects (it being assumed that these effects do not vary randomly within the population at hand), while random variation in the sample is modelled using random effects (see examples about subject-specific and task-specific differences below), which take account of the fact that every participant in the study has slightly different characteristics. This is an idiosyncratic factor that affects all responses from the same subject, thus rendering these different responses inter-dependent rather than independent.

In a first step the fixed effects (or explanatory variables or predictors) are specified, which follows precisely the pattern of the traditional ANOVA. What about interactions? If we predict one of our fixed effects (e.g. AO) to have an effect on the L2 outcome (or whatever else is measured) which is somehow modulated through another variable, say motivation, we should include this inter-dependence, too. For example, it could be that motivation has opposite effects in early starters and late starters. Or it could be that older learners react differently to a certain intervention (e.g. a new teaching method) than younger learners.

Besides specifying the fixed effects, we then identify all possible upper-level variance-covariance structures relevant to the relationships among the lower-level units (e.g. participants sampled from a larger population of L2 learners), the language stimuli sampled from a larger population of linguistic materials (i.e. items in a task), and/or repeated measures in a longitudinal design. Samples drawn from classes within schools can also vary in a non-nested fashion. Some of the students within a class may come from the same family, and students from the same family could be in different classes. In this case, families and classes are not nested hierarchically but are instead crossed at the same level of sampling; mixed-effects analyses can model such crossed random effects as well (Raudenbush Reference Raudenbush1993).

We include the random components as ‘random intercepts’, which allow mean values for each participant and each item to vary. ‘Random slopes’, on the other hand, provide an account of the fact that different participants and different items may vary with regard to how sensitive they are to the manipulation involved (Cunnings Reference Cunnings2012: 374). For instance, it could be that age effects are restricted to certain items, in line with the idea that the age factor represents an individual difference variable, or that certain early starters profit more from an early start than others. However, if participants are either in an early AO group or late AO group, as is typical of age studies, one cannot estimate a by-participant slope for AO from a quantitative perspective. The participants either belong to the early starters or the late starters. By contrast, for within-group variables (e.g. L2 development within one AO group over time), random slopes are required to account for the random variation in the repeated measures. Likewise, a random slope model is needed if the effects of AO are expected to be different for different items.

The output section, then, not only provides parameter estimates for the fixed effects (t-values and standard errors) but also information on the random components (variance and standard deviation). This enables the answering of a number of interesting questions: how much of the variance do the random effects account for? How does the impact of AO change if we do not include random effects? To what extent do individual differences and environmental factors (school, class, family, neighbourhood) mediate, and perhaps mitigate, age effects? Are some items and tasks more susceptible to age effects than others? Are early and late starters, older and younger learners, similarly affected by the FL input?

4.2 Attitudes of L2 teachers working with different age groups

One reason sometimes suggested for the relatively poor level of L2 proficiency attained by primary-school language learners – the fact, for example, that their L2 proficiency tends to be swiftly caught up with by secondary-school beginners – is that the attitudes of primary-school teachers towards the teaching of the L2 in question may be lacking in enthusiasm or commitment. Thus, much was made in Ireland of the results of a large-scale survey conducted in the 1980s (INTO 1985), which found that 22% of primary-school teacher respondents expressed themselves as ‘indifferent’ to their role of teaching the Irish language at primary level (mostly as an L2). In Pfenninger & Singleton's recent long-term longitudinal study of the progress of early and later beginners in English in Switzerland (Pfenninger & Singleton Reference Pfenninger and Singleton2017), there was often implied criticism in the early starters’ accounts of their experience of their primary-school teachers’ attitudes and approach in comparison with those of their secondary-school teachers (see also Pfenninger & Lendl Reference Pfenninger and Lendl2017).

Teachers’ assumptions and approach (cf. Richards & Lockhart Reference Richards and Lockhart1994: 34) may be based on many factors, such as their training and their experience as language learners as well as teachers. In regard to teaching at large, attitudinal differences have sometimes been observed between primary and secondary-school teachers with respect, for instance, to student participation and involvement (e.g. Karanezi Reference Karanezi2014). Concerning L2 teaching, there is, in fact, plenty of evidence of enthusiasm on the part of L2 teachers – including, of course, primary-school L2 teachers (see Al Darwish Reference Al Darwish2017). Interestingly, there are also some indications that teachers catering to different age groups may diverge with respect to what they think it is important to teach (see e.g. Tsukamoto Reference Tsukamoto2001). The fact is, though, that language teachers’ attitudes to their task of teaching a given L2 is a somewhat neglected research area, and that comparative studies of primary- and secondary-school teachers’ attitudes are quite rare – especially in the L2 area.

This study requires the recruitment of a rough balance of primary and secondary language teachers teaching the same L2 in the same type of school (state, private, etc.), where the socio-economic profile of the students is similar and where the L2 has the same kind of place on the curriculum (limited hours, immersion, etc.). The objective will be to explore the attitudes of the teachers towards their L2 teaching task, to identify the (problem of) articulation between levels in the educational system, and to discover whether or not there are sharp attitudinal differences between the two categories of teachers. We suggest a mixed methods approach, with the quantitative dimension being catered for via an attitudes questionnaire (Linn, Bermel & Ferguson Reference Linn, Bermel and Ferguson2015). The relevant questionnaire should be devised using insights from similar questionnaires focusing on various aspects of teaching, but its questions should (a) be relevant to both primary and secondary level; (b) focus on the specific L2 being taught and its features; (c) explore the respondents’ perspective on and commitment to early L2 instruction and their notions about multilingualism; (d) gauge degree of enthusiasm for different ways of pedagogically approaching different aspects of the language; and (e) probe the perceived reasons for the expression of the attitudes enunciated. The qualitative aspect can be taken care of by randomly selecting a small subsample from among the set of primary teachers and another from the set of secondary teachers, which can then function as focus groups who can – both separately and in interaction with each other – discuss the matters raised by the questionnaire and shed more light on the consensus or differences between them. We are confident that the use of a simple mixed methods approach such as the above stands a fair chance of beginning to shed some light on attitudinal differences (or their absence) among language teachers dealing with different age groups.