1. Introduction
Barbara, a trilingual woman from Germany who is married to an English speaker and is living in the UK, pointed out in Dewaele (Reference Dewaele2013) that despite the fact of being equally proficient in her German L1 and English L2, emotion words in both languages felt very different: “The German words have (…) much more of a physical connection (…) like I tell off my children and if I do it in German, I get involved but if I do it in English it's a purely rational disciplinary thing. Expressing emotions in English to my English friends often feel as if I'm only pretending to have these emotions” (p. 94). Barbara's experience reflected a broad pattern in the data of 1,579 adult multilinguals from all over the world collected through the Bilingualism and Emotions Questionnaire (BEQ, Dewaele & Pavlenko, Reference Dewaele and Pavlenko2001-2003): emotional resonance of L1 emotion words was significantly stronger than that of emotion words from languages acquired later in life (LX), even among participants who rated themselves as being equally proficient in their L1 and LX (Dewaele, Reference Dewaele2010).
The difference in emotional resonance between L1 and LX has been linked to the fact that they are acquired in very different contexts of acquisition: the L1 being absorbed in the multimodal and emotionally rich home context versus the LX typically being learnt in a relatively de-contextualised way between classroom walls (Harris et al., Reference Harris, Gleason, Ayçiçeği and Pavlenko2006). Age of onset of acquisition, often hypothesised to be causally related to reduced emotional resonance in LX, shows some overlap with the emotional context of language learning but does not seem to be the cause of the L1-LX difference in itself (Harris et al., Reference Harris, Gleason, Ayçiçeği and Pavlenko2006).
L1 feels more strongly embodied because it is acquired during a period of intense affective socialisation (starting from birth) during which the child goes through “a process of integration of phonological forms of words and phrases with information from visual, auditory, olfactory, tactile, kinesthetic, and visceral modalities, autobiographical memories, and affect” (Pavlenko, Reference Pavlenko2012, p.421). All the connected information is stored in the child's implicit memory. Emotion words and expressions thus gain a “physical connection” within the individual (cf. Barbara's personal account quoted above). Multilinguals who grow up with multiple L1s link the words swirling around them with rich autobiographical memories. They acquire the socio-pragmatic skills that allow them to gauge the exact impact of emotion words and to use these words appropriately. They might remember the startled look on the face of a teacher when using a taboo word in class, or the grin on the face of their friends when using that same word away from adults. This integration of emotion words and the emotional reactions they elicit will be much looser in the LX (typically acquired after the age of 3), and especially for LXs acquired through formal school instruction only. The main reason is “the decontextualized nature of the language classroom, which does not provide many opportunities for integration of all sensory modalities and verbal conditioning (other than foreign language anxiety) and thus leads to development of ‘disembodied’ words, used freely by speakers who do not experience their full impact” (Pavlenko, Reference Pavlenko2012, p.421). LX words are more likely to be stored in the learner's declarative memory, meaning the user can access them and translate them, but multimodal connections remain sparse and weak. This has consequences for the way L1 and LX words are processed.
The processing of emotion words is automatic in L1(s) while in LXs there is semantic processing but no affective processing (Pavlenko, Reference Pavlenko2012), which may explain why Barbara (quoted above) felt that telling her children off in English was a “purely rational disciplinary thing”, in contrast to the extra depth when expressing her heartfelt anger in German. In other words, Barbara had a deep emotional connection with her German emotion words, and a much more superficial and intellectual one with her English emotion words.
Following on from more qualitative research into the matter, recent years have also seen a growing number of more quantitative attempts to measure differences in emotional resonance between L1 and LX (see Toivo & Scheepers, Reference Toivo, Scheepers, Mavrou, Pérez Serrano and Dewaele2022 for a review). Here, we will briefly outline some of the key findings.
1.1. (Neuro-) physiological and behavioural measures
Physiological measures utilise techniques that are sensitive to the activation of the Autonomous Nervous System (ANS); pupillometry and the recording of skin-conductance responses are two prominent examples. Both techniques have provided solid support for the notion of reduced emotional resonance in LX. A general finding is that when perceivers encounter emotionally arousing stimuli, there will be greater activation in the ANS, as is measurable via increases in pupil size or skin-conductance, respectively. In participants’ LX, the difference in the magnitude of physiological activation to arousing vs. neutral stimuli is typically smaller (for example, Harris et al., Reference Harris, Ayçíçeğí and Gleason2003; Toivo & Scheepers, Reference Toivo and Scheepers2019), which can be interpreted as an indication of reduced emotional resonance in LX.
Caldwell-Harris et al. have conducted skin-conductance response studies across a number of different language combinations, modalities of stimuli, and types of bilinguals (e.g., Caldwell-Harris et al., Reference Caldwell-Harris, Tong, Lung and Poo2010; Harris, Reference Harris2004; Harris et al., Reference Harris, Ayçíçeğí and Gleason2003). They have found strong support for a reduced physiological response in LX. However, the effect seemed to depend on the types of stimuli used; for example, Caldwell-Harris et al. (Reference Caldwell-Harris, Tong, Lung and Poo2010) found that participants who spoke Mandarin as their L1 responded more strongly to endearments in their LX. Further, late English L2 learners had stronger skin-conductance responses for childhood reprimands in their L1, but for early bilinguals, responses were comparable between L1 and L2 (Harris, Reference Harris2004).
Pupillometry studies have established reduced emotional resonance using designs that involve single words (Toivo & Scheepers, Reference Toivo and Scheepers2019) and sentences (Iacozza et al., Reference Iacozza, Costa and Duñabeitia2017). Iacozza et al., for example, tested Spanish (L1)–English (L2) late bilinguals and found that the difference in pupillary responses to high vs. low arousing stimuli was smaller in L2 as opposed to L1. Similarly, Toivo and Scheepers (Reference Toivo and Scheepers2019) found that the difference in pupillary responses to high vs. low arousal words was smaller in participants’ L2 than in L1, both for German (L1)–English (L2) and for Finnish (L1)–English (L2) late bilinguals.
Behavioural measures, on the other hand, utilise standard paradigms from cognitive psychology, such as the Lexical Decision Task (e.g., Conrad et al., Reference Conrad, Recio and Jacobs2011; Ponari et al., Reference Ponari, Rodríguez-Cuadrado, Vinson, Fox, Costa and Vigliocco2015), different variations of the Stroop Task (e.g., Fan et al., Reference Fan, Xu, Wang, Zhang, Yang and Liu2016; Winskel, Reference Winskel2013), or other measures such as the Implicit Association Task (Segalowitz et al., Reference Segalowitz, Trofimovich, Gatbonton and Sokolovskaya2008), or the Affective Simon Task (Altarriba & Basnight-Brown, Reference Altarriba and Basnight-Brown2010). These paradigms are perhaps most commonly used in studying bilingual emotions, as they are widely validated and relatively easy to administer. In the context of emotional resonance in bilinguals, such studies are built on the assumption that emotional activation is more automatic in speakers’ L1. Correspondingly, reduced emotional resonance in LX would be reflected in an L1 advantage effect (faster responses to emotional stimuli in L1) or in an L2 advantage effect (faster responses to emotional stimuli in L2), depending on the type of paradigm used.
For example, testing Chinese English speakers with the Emotional Stroop Task, Fan et al. found that the Stroop effect was stronger in L1 as opposed to L2 (Fan et al., Reference Fan, Xu, Wang, Zhang, Yang and Liu2016, Reference Fan, Xu, Wang, Xu, Yang and Lu2018). Winskel (Reference Winskel2013) found a similar pattern in Thai speakers, where the expected interference effect was again stronger in L1. On the other hand, several studies have failed to detect differences in emotional interference/facilitation between L1 and L2 and concluded that they did not find evidence for reduced emotional resonance in L2 (e.g., Dudschig et al., Reference Dudschig, de la Vega and Kaup2014; Eilola & Havelka, Reference Eilola and Havelka2010; Kazanas & Altarriba, Reference Kazanas and Altarriba2016; Sutton et al., Reference Sutton, Altarriba, Gianico and Basnight-Brown2007). Interestingly, Eilola and Havelka (Reference Eilola and Havelka2010) found differences in skin-conductance responses between L1 and L2, but the interference effect was equal between L1 and L2 in a Stroop Task.
Lastly, there is also neuro-physiological evidence in support of reduced emotional resonance in L2. An ERP study by Wu and Thierry (Reference Wu and Thierry2012), for example, suggested that words associated with more negative emotional valence tend to block simultaneous activation of L1 and L2. Similarly, Jończyk et al. (Reference Jończyk, Boutonnet, Musiał, Hoemann and Thierry2016) showed lower N400 amplitudes for negative valence sentences in L2 than in L1. Combining ERP with a Lexical Decision Task, other studies suggested weaker or delayed affective processing in L2 compared to L1 (Conrad et al., Reference Conrad, Recio and Jacobs2011; Opitz & Degner, Reference Opitz and Degner2012). These findings suggest that there is a difference in how bilinguals process affective words, and that this may in turn affect access to such words in the mental lexicon. In an fMRI study where participants read passages of Harry Potter in L1 and L2, Hsu et al. (Reference Hsu, Jacobs and Conrad2015) found stronger hemodynamic responses in the amygdala and the left pre-frontal cortex to happy vs. neutral passages, but this was only found when passages were read in L1. Again, this indicates stronger emotional involvement when reading emotional texts in L1 rather than L2.
While the different experimental paradigms – on the whole – appear to provide compelling evidence for the notion of reduced emotional resonance in bilinguals’ LX, their feasibility in terms of operationalising and psychometrically quantifying this construct (e.g., to capture related inter-individual differences) is naturally somewhat limited. Many of these paradigms require specialist equipment and/or procedures that can only reasonably be applied in a laboratory setting. Moreover, while such paradigms tap into specific behavioural or physiological manifestations of affective processing in various languages, they are hardly able to capture emotionality in LX across a wider range of contexts and experiences. This is where self-report methods come into their own, which we will discuss next.
1.2. Self-report methods
A key finding that emerged from studies using the BEQ or interviews was the significantly stronger emotional investment in L1 compared to LX which became manifest in a self-reported preference to express emotions in either L1 or LX, depending on emotional content (e.g., a preference for expressing endearments in L1, but a preference to talk about embarrassing topics in LX).
The emotional power of L1 swearwords was positively correlated with their frequency of use (Dewaele, Reference Dewaele2004, Reference Dewaele2010, Reference Dewaele2011). One sub-group of bilingual participants that deviated from this pattern were Asian and Arab (especially female) participants, who were more likely to report that they preferred swearing in LX, to avoid the taboo of swearing in their L1. The use of LX for swearing allowed them to overcome this social constraint (Dewaele, Reference Dewaele2013). Shakiba and Dewaele (Reference Shakiba, Dewaele, Mavrou, Pérez Serrano and Dewaele2022) found a similar pattern among 204 Persian–English multilingual immigrants in Canada. Female participants reported swearing significantly less frequently in Persian than their male peers, but no gender difference existed for English. One female participant pointed out: “I cannot swear in Persian, I feel the heaviness of using those words and bad reaction from Persian culture to a woman who swears” (p. 15). Other studies that used the BEQ such as Resnik (Reference Resnik2018), who collected data from 167 multilinguals, confirmed the overall preference of L1 for swearing and the superior emotional resonance of swear and taboo words in participants’ L1 compared to their LX.
The closed items about language preference for swearing and for estimation of emotional resonance of L1 and LX swearwords were formulated very broadly in the BEQ. This sufficed for the identification of general patterns: “If you swear in general, what language do you typically swear in?” (with a 5-point Likert scale ranging from “never” to “all the time”) and “Do swear and taboo words in your different languages have the same emotional weight for you?” (with a 5-point Likert scale ranging from “not strong” to “very strong”) (Dewaele, Reference Dewaele2013, p.230).
In order to gain a more granular view of multilinguals’ swearing habits and perceptions, Dewaele (Reference Dewaele2016) made a list of 30 English negative emotion-laden words ranging from relatively mild levels of offensiveness such as “daft” and “fool” to much more offensive words and expressions such as “bitch” and “prick” which were integrated in a short sentence to provide some context. A total of 1,159 English L1 users and 1,165 English LX language users rated their understanding of each word, its offensiveness and frequency of use on a 5-point Likert scale (Dewaele, Reference Dewaele2016). The analysis of differences between L1 and LX users yielded some expected and some unexpected findings. For instance, L1 users reported a better understanding of the words and expressions and more frequent use of the highly offensive ones. Also, English LX users reported swearing significantly less in English than L1 users. Unexpectedly, however, LX users tended to significantly overestimate the offensiveness of 29 out of the 30 words and expressions. This could have been the result of foreign language teachers warning them that these were “red flag” words that could get them into trouble and that it was safer to stick to more neutral words. Lacking sufficient exposure to authentic use of the negative emotion-laden words, LX users may have struggled to differentiate between them and calibrate their affective power accurately. The expected pattern of underestimation of the emotional weight of LX swearwords was found only for a single word, starting with “c” and ending in “t”, the most offensive word in the list. This could be due to the fact that this word is truly taboo, and as such used much less frequently by L1 users than a word like (say) “fuck”, which in turn may make it harder for LX users to pick up its offensiveness (Dewaele, Reference Dewaele2018).
Emotional resonance is not just a feature of negative emotion words. Positive emotion words and declarations of love can be particularly resonant depending on the language in which they are uttered. Dewaele (Reference Dewaele2008) found that half of the participants who filled out the BEQ felt that the words “I love you” were most powerful in their L1, around a third of participants rated them equally strong in their L1 and an LX, while the remaining participants described the expression as being stronger in their LX – the language used with their partner. Many participants admitted that even after years of using LX they still remained a little unsure about the emotional resonance of “I love you” in LX.
Caldwell-Harris et al. (Reference Caldwell-Harris, Kronrod and Yang2013) focussed on the Mandarin translation of “I love you” (“Wo ai ni”) among 66 bilingual Chinese university students and the perception and use of “I love you” by 71 monolingual English American students. The expressions were found to be pragmatically different in both languages. Compared to the American students, the Chinese students were significantly less expansive in expressing love, and they actually preferred nonverbal expressions of love. Ożańska-Ponikwia (Reference Ożańska-Ponikwia2017) focused on the emotional resonance of “I love you” and the equivalent Polish expression “Kocham cię” among 72 Polish–English bilinguals living in the UK and Ireland. “Kocham cię” was judged to be emotionally stronger that “I love you”, but socialisation in English culture, a high self-perceived English proficiency and a high frequency of English use, were linked with a stronger emotional resonance of “I love you”. Pursuing the investigation of love in an LX, Dewaele and Salomidou (Reference Dewaele and Salomidou2017) investigated whether expressing love to a partner in an LX caused linguistic and psychological challenges. A third of 429 participants from all over the world claimed not to have experienced any difficulty, while half mentioned lexical and conceptual limitations in the LX which they claimed had hampered their communication of emotion. An apparent lack of emotional resonance in LX was also mentioned by participants who were in an intercultural romantic relationship (Dewaele & Salomidou, Reference Dewaele and Salomidou2017). Many complained about a lack of emotional intensity when communicating with their partner in LX. Around a quarter of participants linked this to a lack of ‘genuineness’, at least at the start of the relationship. Most participants went through a process of emotional socialisation in LX and for a majority, their LX eventually became their ‘language of the heart’.
Dewaele et al. (Reference Dewaele, Lorette, Rolland and Mavrou2021) developed a composite measure reflecting participants’ self-reported emotional reactions when watching television news and films in L1 or LX. These include self-reported frequency and intensity of feeling emotional, frequency of laughter and degree of trust. Participants were 271 British English L1 users, 282 Greek and 271 Hungarian English LX users living in their home country. The study revealed an expected pattern of significantly weaker emotional reactions for English LX users compared to English L1 users. Unexpected differences emerged between the two LX user groups: the Greek–English bilinguals reported significantly stronger emotional reactions than the Hungarian–English bilinguals, despite similar levels of English (LX) proficiency. This could be linked to the quality of the foreign-language English teaching in the Greek education system, combined with a tradition of watching foreign films in the original version with subtitles rather than dubbing like in Hungary.
The reduced emotional resonance in LX may cause LX users to feel inauthentic, but it can also allow them to disclose and process traumatic experiences that would be too painful to discuss in their L1. Switching to the LX allows multilingual clients in psychotherapy to either zoom in or out depending on whether they need to distance themselves from the traumatic event they are discussing (Costa & Dewaele, Reference Costa and Dewaele2014, Reference Costa and Dewaele2019; Rolland et al., Reference Rolland, Dewaele and Costa2017, Reference Rolland, Costa and Dewaele2021). Cook and Dewaele (Reference Cook and Dewaele2021)'s multiple case study on the language preferences of three refugees who had been tortured in their homeland because of their sexual orientation before starting therapy in the UK showed that some experiences were just too traumatic to retell in their L1 but that the use of English (LX) enabled them “to visit their pain” as one participant put it.
1.3. The present research
As the above illustrates, quantitative studies that rely on self-reports have the potential to reach a larger number of participants with more different language combinations and cultural backgrounds than would be feasible in a laboratory experiment. Self-report methods are also able to capture a wider range of emotional experiences than an experiment with a narrow focus on behavioural or physiological manifestations of affect would ever allow to measure.
To this date, however, there exists no self-report survey instrument that would demonstrably satisfy key psychometric standards for a reliable and valid measurement of reduced emotional resonance in LX. This is precisely the gap that the current study is going to address. Specifically, we aim to operationalise reduced emotional resonance in bilinguals’ LX and to develop a psychometrically validated measurement tool to quantify this construct for the purpose of resolving inter-individual differences among bilinguals with diverse language combinations and cultural backgrounds.
2. Study 1
2.1. Method
Conceptualisation of reduced emotional resonance in LX
The purpose of the scale we develop here is to measure reduced emotional resonance in the LX of bilingual speakers. The scale is aimed at speakers who use two or more languages in their daily lives, but consider their L1 to be dominant. They will be referred to as bilingual speakers here. The term LX is used to describe the “later learnt language” of these speakers. This term is chosen to reflect bilingualism as a spectrum of different language backgrounds. This is to move away from describing bilingualism in terms of a clear “order” of languages (Dewaele, Reference Dewaele2017). Moreover, this idea allows for the concept of first and consequent languages to be more dynamic, based on the context of learning, context of current use and which language the speaker feels to be more dominant or their “main language”.
Bilinguals often feel “less” when speaking in their LX. Despite being proficient and having an excellent command of the language, it may not feel the same as one's “first” language (L1) does (Pavlenko, Reference Pavlenko2005). In some cases, the LX can feel detached or even fake (Pavlenko, Reference Pavlenko2005). Bilinguals may also refer to an emotional distance in LX (Degner et al., Reference Degner, Doycheva and Wentura2011), as opposed to their L1, which usually feels more like the language of emotions (Dewaele, Reference Dewaele2010, Reference Dewaele2013).
This emotional distance may make it easier to discuss embarrassing topics (Bond & Lai, Reference Bond and Lai1986) or swear in one's LX. Bilinguals may also feel that emotional words and phrases (such as the phrase “I love you” Ożańska-Ponikwia, Reference Ożańska-Ponikwia2017) have less weight in LX, as opposed to L1. Bilinguals may prefer to switch to their L1 when angry, so as to be able to express their anger (Dewaele, Reference Dewaele and Pavlenko2006). Indeed, emotional distance in L2 might actually be useful in therapy, when discussing emotional content or accessing memories (Rolland et al., Reference Rolland, Dewaele and Costa2017).
Here, we will refer to this emotional distance as reduced emotional resonance of LX. This term is chosen to reflect bilinguals’ feeling of emotional intensity in a language, rather than expressing or experiencing different emotional states. We aim to capture bilingual speakers’ perceptions of this emotional intensity, and how it might be reduced in their LX, in comparison to their L1. This provides flexibility in terms of which languages the researcher wants to measure and allows for the use of the scale in multilingual settings and with participants who have complex language backgrounds.
Candidate item generation
We began the process by generating candidate items. Each author contributed items to the candidate list, and prior to the first phase of data collection, all items were reviewed and discussed with all the authors. Most candidate items were inspired by previous literature on reduced emotional resonance in LX, attempting to capture multiple modalities and situations. The pilot scale consisted of 22 candidate items. The full list of candidate items and supporting references for each can be found in Appendix 2.
Participants
Participants were recruited through social media, bilingualism/multilingualism-related mailing lists, and the subject pools at Glasgow University and Birkbeck, University of London. Participants were not paid for their participation, but 74 participants who were recruited through subject pools and completed the study as a course requirement were given course credit for their participation. In total, 1120 participants started the survey, of which 688 (61%) completed the study in full. Only the latter were included in subsequent analyses.
Of the final 688 participants, 527 (77%) identified themselves as female, 134 (19%) as male, and 17 (2.5%) as non-binary. Ten participants (1.5%) preferred not to reveal their gender. The average reported age was 32.4 years (SD = 13.5 years), ranging from 18 to 80 years.
Participants were from very diverse L1 backgrounds, encompassing 87 different languages (Appendix 1a). All participants reported to actively speak between 2 and 10 different languages; 213 (31%) reported to use two languages and the remaining 475 (69%) to use more than two languages regularly. Eighty-four different languages were listed among participants’ LXs (Appendix 1b).
Participants’ average self-reported length of stay in the country where their LX is spoken was 10.98 years (SD = 12.2 years), ranging from 0 to 65 years.
Procedure
All data collection took place online through the Experimentum data collection platform (DeBruine, Reference DeBruine, Lai, Jones, Abdullah and Mahrholz2020). Participants were sent a link to access the study. First, participants were asked to provide demographic information (age, gender, languages acquired before the age of 3 [L1s] and languages acquired after the age of 3 [LXs], length of stay in a country where LX is spoken, and which LX they use the most frequently). After completing the demographic questions, participants were asked to respond to all the candidate items (see Appendix 2), the order of which was randomised individually for each participant. Participants were asked to think of the L1 and LX that they use the most frequently. Each candidate item was presented with a Likert scale ranging from 1 (strongly disagree) to 6 (strongly agree), and participants were asked to use this scale to indicate, as best as they could, their degree of agreement with the statement in each given item.
2.2. Results
For the analyses reported in this section, all data and scripts are available at https://osf.io/v4k32/.
Test reliability and internal consistency
Analyses of test reliability were performed using the alpha() function of the R package psych (Revelle, Reference Revelle2022). By setting the check.keys argument to TRUE, the procedure automatically reverse-scored items with negative loadings on the first principal component. It emerged that the initial candidate set of 22 items (Appendix 2) already obtained a ‘good’ Cronbach's alpha of 0.815, with a 95% CI of [0.795, 0.832]Footnote 1. The median inter-item correlation (MedIIC) was 0.178, with a 95% CI of [0.160, 0.205]. However, the output suggested that dropping some of the items would further improve reliability.
We therefore successively removed individual items from the scale (at each step dropping the item whose removal yielded the largest improvement in alpha) until alpha did not increase any further. Following this procedure, items Q2, Q3, Q6, Q8, Q9, Q11, and Q13 were excluded.
The resulting final scale includes 15 items. It obtains a Cronbach's alpha of 0.866 (95% CI = [0.849, 0.880]) and a MedIIC of 0.302 (95% CI = [0.269, 0.332]). Compared to the initial 22-item scale, the final scale significantly improves on alpha by 0.051 units (95% CI = [0.030, 0.056]) and on MedIIC by 0.124 units (95% CI = [0.089, 0.141]).
Composite test scores
Composite test scores were calculated via averaging across the final 15 items after reverse-scoring item Q7 (as suggested in the previous analyses). Higher values on the composite score indicate greater emotional resonance reduction in LX relative to L1. The 688 bilingual participants in our sample were roughly normally distributed on this metric (Figure 1), with an overall mean of 3.575 (95% CI = [3.506, 3.643]) and a standard deviation of 0.933 (95% CI = [0.887, 0.977]).
Figure 1. (TOP) Density distribution of the mean composite test scores from the final 15-item scale. Test scores can theoretically range from 1 to 6, in line with the 6-point Likert scales per item. Also shown is a hypothetical normal distribution curve for this value range, as predicted from the sample mean and SD. (BOTTOM) Normal Q-Q plot for the mean composite test scores, with robustly estimated prediction line and 95% confidence bands. Only 20 of the 688 observations (3%) are not within the confidence envelope of a hypothesised normal distribution, due to somewhat under-dispersed test scores towards the upper end of the predicted quantile range.
Test reliability for smaller samples
The previous figures for Cronbach's alpha and MedIIC suggest good test reliability of the final 15-item scale, based on a sample size of 688 participants. However, in more practical terms, it is also useful to know how the scale will perform when sample sizes are much smaller – say, if one wants to measure emotional resonance reduction in LX for use as a participant-specific control variable in an experiment with only 20-60 participants.
Indeed, since sampling error is expected to increase with smaller Ns (higher likelihood of “extreme” samples) and assuming that sampling distributions for alpha and MedIIC are skewed (both measures can only take values between 0 and 1), there will likely be a bias in the average alphas and MedIICs for smaller Ns, combined with wider distributional spreads around those averages.
To examine this issue, and to provide some benchmarks for future applications of our scale, we performed a Monte Carlo analysis (based on bootstrapping) in which we simulated the sampling distributions for alpha and MedIIC at various ‘small’ settings of N, ranging from 15 to 60 participants by increments of 5 participants. Fifteen participants (same as the number of items) was considered the minimum feasible sample size for determining test reliability. For each setting of N, we took 10K resamples from our data, each time by randomly drawing (with replacement) N participants from the pool of 688 available. For each of these resamples, we then determined our test statistics of interest (Cronbach's alpha and MedIIC).
The results are summarised in Table 1. Across the 10K resamples per N, we calculated the means as well as 5% and 95% quantiles per test statistic. These quantiles represent the lower and upper bound for the central 90% of the bootstrapped sampling distribution per measure and N (or if one prefers, estimates for the upper and lower bound of the two-tailed 90% CI per measure and N).
Table 1. Bootstrap results for small(ish) participant samples ranging from N = 15 to N = 60. Shown are the means, 5% quantiles (Q .05), and 95% quantiles (Q .95) for Cronbach's alpha and median inter-item correlation (MedIIC). Figures are based on 10,000 resamples per N.
The table shows that even with relatively small Ns, mean alphas and MedIICs stay reasonably close to their expected values of alpha = 0.866 and MedIIC = 0.302, respectively (cf. 2.2. Test Reliability and Internal Consistency). However, it is also evident that as N gets smaller, the mean alpha tends to deviate more negatively and the mean MedIIC more positively from the relevant expected value. These biases are due to increasing sampling error for smaller Ns combined with differently-skewed sampling distributions per test statistic (negative skew for alpha, positive skew for MedIIC).
The bootstrapping results also confirm that larger Ns lead to systematically narrower spreads in the sampling distributions for each test statistic (distances between the 5% and 95% quantiles become consistently smaller as N increases). Again, this is predicted from the fact that larger Ns lead to a decrease in sampling error.
Given that the Q .05 figures in the table also indicate estimated lower bounds for the one-tailed 95% CIs per measure and N, we can be confident that our scale will achieve at least ‘acceptable’ test reliability (alpha ≥ 0.7) even with as few as only 15 participants per study. With between 45 to 50 participants, we can be 95% confident that the scale achieves at least ‘good’ reliability (alpha ≥ 0.8). Additional Monte Carlo runs (not shown in the table) indicated that, from around 150 participants per study, confidence in alphas of 0.8 or better increases to at least 99.9%.
Dimensionality
High test reliability of the final scale reveals little about its underlying factor structure. Indeed, Cronbach's alpha must not be interpreted as a test of unidimensionality. We therefore also conducted an Exploratory Factor Analysis (EFA) over the final 15 items, using data from all 688 participants. As with the previous analyses, we employed the R package psych (Revelle, Reference Revelle2022) for this kind of modelling.
Given the ordinal nature of the original item responses (6-point “agreement” ratings), factors were determined on the basis of polychoric correlations using the maximum likelihood factoring method. To estimate how many factors are needed, we performed a Parallel Analysis (Horn, Reference Horn1965) over 1000 iterations. In addition, we considered Bayesian Information Criteria (BIC and SABIC) for unrotated factor solutions. All of these indices converged on five as the optimum number of factors to extract.
To interpret the five extracted factors, we explored a range of different oblique rotation methods. Two of these (simplimax and biquartimin) failed to converge; the remaining ones (promax, oblimin, bentlerQ, geominQ, and cluster) all converged on the same factor structure, but with minor differences in the magnitudes of the loadings and factor correlations. Figure 2 shows the promax-rotated solution.
Figure 2. Polychoric EFA diagram illustrating how the 5 extracted factors (coloured ellipses on the left) load on the final 15 items (coloured rectangles on the right) after promax rotation. The corresponding arrows only show ‘substantial’ loadings with absolute values ≥ 0.3. The numbered arcs on the left represent between-factor correlations. Item Q7 was entered into the EFA using its original (non-reversed) scoring, hence the negative loading from factor F1. Model fit was good (RMSEA = 0.037, fit = 0.924, off-diagonal fit = 0.998), and so was factoring reliability (TLI = 0.976).
As can be seen from the figure, factor F1 (Eigenvalue = 1.921) loads negatively on item Q7 (“I find it easier to talk about sex in my L1 than in my LX”), but positively on items Q20 (“I emotionally connect with my conversation partner better and faster in my L1 than in my LX”) and Q21 (“I have a better sense of what my conversation partner is thinking or feeling when using my L1 than my LX”), thus lending itself to a Reduced Emotional Connection in LX interpretation.
Factor F2 (Eigenvalue = 1.803) is interpretable as Reduced Vulnerability in LX. It loads positively on items Q5 (“Swear words affect me more in my L1 than in my LX”), Q10 (“Being criticised feels more unpleasant in my L1 than in my LX”), and Q15 (“Insults feel more hurtful in my L1 than in my LX”).
Factor F3 (Eigenvalue = 1.676) loads positively on items Q16 (“I prefer my L1 over my LX when reading for pleasure”), Q17 (“A sad film is more likely to make me cry when I watch it in my L1 rather than my LX”), Q19 (“Poetry in my LX has less of an effect on me than poetry in my L1”), and Q22 (“Romantic songs feel more intense in my L1 than in my LX”). We therefore interpret it as Reduced Emotional Engagement with Media/Art in LX.
Factor F4 (Eigenvalue = 1.667) loads positively on items Q1 (“I feel less emotional when using my LX than when using my L1”), Q12 (“Saying the equivalent of “I love you” has more weight in my L1 than in my LX”), and Q14 (“Compared to L1, I feel like there is more of an emotional distance when I use my LX”). Hence, we label this factor Perceived Emotional Distance in LX.
Finally, factor F5 (Eigenvalue = 1.167) loads positively on items Q4 (“When I'm really angry, I tend to use my L1 more than my LX”) and Q18 (“When I want swear words to have real weight, I use my L1 rather than my LX”), therefore lending itself to a Reduced Offensiveness in LX interpretation.
Very importantly, all between-factor correlations (numbered arcs in Figure 2) are significantly positive (Bonferroni-corrected ps < .05) and the majority of them quite substantial (≥ 0.3) – a pattern that is consistent across all the oblique rotation methods considered (promax, oblimin, bentlerQ, geominQ, and cluster). This suggests that the five extracted factors represent more detailed (but related) facets of a single underlying meta-factor, which we may call Reduced Emotional Resonance in LX. Indeed, additional parallel analyses using factor scores from the different oblique rotation methods as input vectors consistently suggested only a single principal component for the five extracted factors. Our scale therefore measures one general construct which is composed of five interrelated sub-domains (or facets) that are more specific.
3. Study 2
The second study was designed to assess the validity of our final 15-item scale, which we will henceforth refer to as RER-LX (for Reduced Emotional Resonance in LX). Specifically, we aim to evaluate its discriminant validity in relation to LexTALE (a measure of linguistic proficiency) as well as its convergent validity in relation to an item set that was used to measure reduced emotional resonance in LX in previous research.
3.1. Method
Materials: Reduced Emotional Resonance in LX (RER-LX)
The final version of our RER-LX scale included 15 items, as shown in Appendix 3. Composite scores (for correlation with other measures) were computed as described in section 2.2 Composite Test Scores.
Materials: Proficiency in LX.
The brief LexTALE test (Lemhöfer & Broersma, Reference Lemhöfer and Broersma2012) was used to assess participants’ proficiency in English (the target LX in this study). The LexTALE is a 60-item lexical decision test, which takes about 5 minutes to complete. Participants are asked to distinguish real English words from English-looking words that do not exist in English. Following the authors’ suggestions, the test scores from this scale were calculated as the percentage of correct responses, adjusted for unequal proportions of words and non-words in the test. In our final sample of participants, the mean score was 85.75% (SD = 11.20%), ranging from 35% to 100% (fourteen participants in our sample, i.e., 6%, achieved the top proficiency score).
Materials: BEQ sub-scale
There are no existing scales that directly measure reduced emotional resonance in LX and return a single score. To test the convergent validity of RER-LX, three questions from the Bilingualism and Emotions Questionnaire (BEQ, Dewaele & Pavlenko, Reference Dewaele and Pavlenko2001-2003) were chosen for comparison, as they had been used as a proxy for emotional resonance in LX in previous research (Dewaele, Reference Dewaele2013; Pavlenko, Reference Pavlenko2005).
1. BEQ-swearing: If you swear in general, what language do you typically swear in? For this question, participants were asked to estimate how often they swear, both in their L1 and in their LX, on a scale from 1 (never) to 5 (all the time).
2. BEQ-feelings: What language do you express your deepest feelings in? Here, participants were asked to estimate how often they express their feelings, both in L1 and in LX, on a scale from 1 (never) to 5 (all the time) across four situations (when alone; with parents/partner; when talking to friends; in letters and emails). Their responses to the four situations were summed up when calculating the score for BEQ-feelings.
3. BEQ-anger: If you are angry, what language do you typically use to express your anger? For this question, participants were asked to estimate how often they express their anger, again both in L1 and in LX, on a scale from 1 (never) to 5 (all the time) across five situations (when alone; with parents/partner; when talking to friends; in letters and emails; when talking to strangers). Their responses to the five situations were summed up when calculating the score for BEQ-anger.
For each of the three questions, we then calculated the difference between L1 and LX, with a higher difference score reflecting a stronger preference for L1 in the given item (interpretable as reduced emotional resonance in LX relative to L1).
Since the ratings for BEQ-feelings were summed up across four situations, we divided the relevant difference scores by 4; likewise, difference scores for BEQ-anger were divided by 5, in line with the five situations of assessment for this question. This way, we ensured that the three BEQ items contributed equally to the mean composite test scores for the BEQ subscale.
Finally, a small number of missing data points (6 for BEQ-swearing, 9 for BEQ-feelings, and 2 for BEQ-anger) were conservatively replaced with the relevant item means.
Participants
As before, participants were recruited through social media and the respective subject pools at the authors’ universities. In our advertisements, we specifically targeted L2 speakers of English. A total of 414 participants started the study, out of which 247 participants (60%) completed the study in full and were included in subsequent analyses. This sample size should give us sufficiently stable correlation estimates (see, in particular, Schönbrodt & Perugini, Reference Schönbrodt and Perugini2013).
All participants were LX speakers of English. Participants were not paid, but 103 participants were given course credit for their participation.
Participants’ mean age was 29.64 years (SD = 12.91 years), ranging from 17 to 86 years. One-hundred-forty-four (58%) identified themselves as female, 43 (17%) as male, 5 (2%) as non-binary, and 55 (22%) preferred not to reveal their gender identity.
The average self-reported age at which participants acquired LX (English) was 7.33 years (SD = 4.44 years), ranging from 0 to 40 years. The average self-reported length of stay in an English-speaking country was 8.54 years (SD = 9.14 years), ranging from 0 to 45 years.
There were 52 different languages among the spoken L1s reported by the participants in our sample (Appendix 1c).
Participants reported speaking between one and eight different LXs. The most commonly reported number of LXs spoken was two (80 participants) or three (80 participants).
Procedure
All data collection took place online through Experimentum (DeBruine et al., Reference DeBruine, Lai, Jones, Abdullah and Mahrholz2020). Participants were sent a link to access the study. First, participants were asked to provide demographic information (Age, Gender, L1, length of stay in an English-speaking country, age of acquisition for English). After completing the demographic questions, participants were asked to complete the LexTALE proficiency test, after which they completed a brief distractor task. Then they were asked to respond to all RER-LX scale items (presented in an individually determined random order), and finally to the three BEQ items. Each of the measures included in the study appeared on a separate page. Participants were instructed to think of English as their LX when completing RER-LX and BEQ.
3.2. Results
For the analyses reported in this section, all data and scripts are available at https://osf.io/v4k32/.
Test reliabilities for RER-LX and BEQ
For the new sample of participants (N = 247), our 15-item RER-LX scale obtained a Cronbach's alpha of 0.867 (95% CI = [0.837, 0.890]) and a MedIIC of 0.294 (95% CI = [0.253, 0.357]). These figures closely replicate those from Study 1.
The 3-item BEQ subscale obtained a Cronbach's alpha of 0.850 (95% CI = [0.805, 0.885]) and a MedIIC of 0.604 (95% CI = [0.523, 0.698]).
With a difference in alpha of only 0.017 (95% CI = [−0.023, 0.061]), test reliabilities for RER-LX and BEQ are roughly the same. This is in spite of considerably lower inter-item correlations for RER-LX than for the BEQ subscale (difference in MedIIC = −0.310, 95% CI = [−0.398, −0.214]). The much larger number of (more unique) items in RER-LX obviously compensates for the latter (cf. de Vet et al., Reference de Vet, Mokkink, Mosmuller and Terwee2017). Also note that the high MedIIC for BEQ (> 0.5) may actually indicate some undesirable redundancy among the three BEQ items.
Convergent and discriminant validity
To establish convergent validity, we correlated the mean composite scores from RER-LX with the mean composite scores from the BEQ subscale. Since the two scales are designed to measure the same construct and are scored in the same direction (higher meaning more emotional resonance reduction in LX relative to L1), we expected a positive correlation.
To establish discriminant validity, we correlated the RER-LX and BEQ measures with the scores from LexTALE, which is designed to measure linguistic proficiency in LX (in this case, English). To exhibit discriminant validity, we expected correlations with LexTALE to be around zero. We provide further theoretical motivation for this in the General Discussion.
Table 2 shows the inter-correlations between the three measures (NB, histograms for each measure suggested no obvious outliers), together with bootstrapped 95% confidence intervals.
Table 2. Inter-correlations between RER-LX, BEQ, and LexTALE, with two-tailed 95% CIs in square brackets. The latter were determined via bootstrapping over 10,000 resamples.
As becomes evident, RER-LX clearly exhibits convergent validity in relation to BEQ – the relevant correlation is positive, appreciable in size, and significant (the lower bound of the associated 95% CI is markedly above zero). At the same time, RER-LX is only weakly negatively related to LexTALE and the corresponding 95% CI encloses zero (meaning this correlation is not significant). In comparison, the BEQ subscale exhibits a more substantial negative correlation with LexTALE, which is indeed significant (the upper bound of the associated 95% CI is clearly below zero). A direct comparison between the two negative correlations confirms that the relation between RER-LX and LexTALE is significantly less negative (i.e., closer to zero) than the relation between BEQ and LexTALE (difference = 0.179, 95% CI = [0.053, 0.302]). RER-LX therefore yields better discriminant validity in relation to LexTALE than the BEQ subscale.
To further corroborate this point, we conducted a Principal Component Analysis across the three measures. We extracted two orthogonal components that together accounted for 83% of the total variance. Considering the loadings after Varimax rotation (Table 3 below), the first principal component (PC1, Eigenvalue = 1.414) clearly captures emotional resonance reduction in LX, as it receives strong positive loadings from both RER-LX and BEQ, but only a weak negative loading from LexTALE. The second component (PC2, Eigenvalue = 1.081) mostly reflects proficiency in LX, with a very strong positive loading from LexTALE, a weaker (but considerable) negative loading from BEQ, and only a very small positive loading from RER-LX. Again, this pattern supports the notion of RER-LX having better discriminant validity than the BEQ subscale. Loading on both PCs, the BEQ subscale obviously captures aspects of both emotional resonance and proficiency in LX. RER-LX, by contrast, is more orthogonal to LexTALE and therefore represents a ‘purer’ measure of emotional resonance reduction in LX than the BEQ subscale.
Table 3. Varimax loading matrix for the first two principal components (PC1 and PC2) extracted from RER-LX, BEQ, and LexTALE. Substantial loadings (absolute values ≥ 0.3) are highlighted with double asterisks.
4. General discussion
4.1. Summary
In this paper, we developed a psychometrically validated measurement instrument for the purpose of operationalising and quantifying emotional resonance reduction in bilinguals’ LX: the RER-LX scale. While there is an abundance of cross-disciplinary research studying bilingual emotions, a scale that explicitly quantifies how much bilinguals “feel less” in languages they acquired later in life has largely been missing in the field – a gap that the current paper tries to close.
As discussed in the introduction, there is ample support for the notion of reduced emotional resonance in bilinguals’ LX from a variety of cognitive-behavioural, psycho-physiological, and self-report paradigms. However, there is as yet no validated psychometric scale to quantify the concept directly. To this date, the Bilingualism and Emotions Questionnaire (BEQ, Pavlenko & Dewaele, Reference Dewaele and Pavlenko2001-2003) is the best and largest-scale attempt at quantifying and trying to understand the underlying reasons for reduced emotional resonance – however, the BEQ is too long to be used concurrently with other measures as a whole. Moreover, the complete BEQ instrument is less specific than RER-LX, includes multiple question types, does not come with explicit scoring instructions, and lacks validation. We have attempted to build on the questions captured in the BEQ and expand the measurement to include other situations and modalities, which have been established since the original BEQ project.
In the two online studies reported here (N = 688 and N = 247, respectively), we established our 15-item RER-LX scale to be a reliable and valid measure of emotional resonance reduction in LX. RER-LX achieves very good test reliability on large samples (alphas close to 0.87 in both studies). Even with smaller samples that are more typical for experimental research, alpha will likely stay well above 0.7 (see Test Reliability for Smaller Samples). Composite test scores from the scale are roughly normally distributed (see Composite Test Scores), which is desirable from a normative measurement perspective and makes the RER-LX scores reasonably suitable for parametric testing (t-test, ANOVA, linear regression, etc.).
RER-LX exhibits convergent validity in relation to the 3-item BEQ-subscale that has previously been used as a proxy for the same construct. Importantly, RER-LX improves on the latter by having significantly better discriminant validity in relation to LexTALE, a measure of proficiency in LX (see Convergent and Discriminant Validity). We will further elaborate on this point in section 4.2 below.
Lastly, the scale exhibits content validity by covering five related facets of emotional resonance in LX that seem to relate well with previous approaches to conceptualizing and measuring the construct: Reduced Emotional Connection in LX, Reduced Vulnerability in LX, Reduced Emotional Engagement with Media/Art in LX, Perceived Emotional Distance in LX, and Reduced Offensiveness in LX (see 2.2. Dimensionality and references in Appendix 2).
4.2. Proficiency and emotional resonance
In terms of discriminant validity, one may ask why we consider independence of proficiency a desirable property of a scale measuring emotional resonance reduction in LX. Indeed, the debate surrounding the relationship between proficiency and emotional resonance in LX is related to that between language and culture. Both constructs are entangled but are different types of skills and knowledge. Although they tend to co-occur, bilingualism and biculturalism are not isomorphous (Grosjean, Reference Grosjean2014). The capacity to speak an LX to some degree does not necessarily imply that the user possesses sociocultural and socio-pragmatic competence to the same degree. For example, foreign language learners who study in a school context will become increasingly proficient thanks to the linguistic input of the teacher and the learning materials, but they will only start developing sociocultural and socio-pragmatic competence through contact with the users of the language. In other words, by observing and participating in authentic interactions, LX users start to understand how the choice of specific words combined with volume, pitch, intonation, facial expression and body language contribute to the impact of emotion words. This will trigger a process of LX socialisation that could last a lifetime. An early start of that process is linked to both higher proficiency and higher emotional resonance in the LX. The degree of LX socialisation and the size of the LX networks were found to be strong predictors of LX emotional resonance in Dewaele (Reference Dewaele2013). LX users who had acquired their LX naturalistically and from a young age reported that the LX had higher levels of emotional resonance than those who had learnt it only in an instructed context and/or later in life.
Proficiency and emotional resonance in LX are therefore conceptually independent dimensions. Indeed, Dewaele et al. (Reference Dewaele, Lorette, Rolland and Mavrou2021) found a non-significant relationship (r = 0.14, p = 0.057) between self-reported emotional reactions and English proficiency in a group of adult British L1 users of English. The picture was different for proficient English LX users, where a small but significant positive correlation was found (r = 0.20). However, much stronger correlations were found between self-reported emotional reactions and both frequency of English use (r = 0.40) and frequency of watching television in English (r = 0.41) (p. 353). In conclusion, high proficiency does not automatically imply greater emotional resonance in LX, as the relationship between the two appears to depend on frequency of use and cultural exposure to LX, among other factors. We therefore deem it advantageous if a scale of emotional resonance in LX remains in principle orthogonal to proficiency in LX, the latter of which can be assessed separately.
4.3. Limitations and future uses of RER-LX
Being a self-report scale, RER-LX may suffer from some of the problems that are associated with such scales in general. For example, participants not only need to understand the questions, but also have sufficient motivation and introspective ability to answer them truthfully and accurately. There may also be cross-cultural differences in the social acceptability of certain questions (or the social desirability of certain answers, respectively) which in turn may bias participants’ responses. These issues clearly require additional empirical scrutiny in the future. All we can say at present is that, across the two independent studies reported here (each with participants from very diverse linguistic and cultural backgrounds), internal consistency of RER-LX stayed largely the same.
It is also worth noting that the present research was carried out under the restrictions imposed by the COVID-19 pandemic, which meant that we had to limit ourselves to the use of online questionnaires to evaluate the reliability and validity of the RER-LX scale. Future research would also need to further assess the scale's predictive usefulness, specifically in terms of how well variation in bilinguals’ RER-LX scores can predict variation in their cognitive-behavioural or (neuro-) physiological responses to affective stimuli. Indeed, our expectation is that bilinguals who score higher on RER-LX should exhibit a larger discrepancy between L1 and LX when processing emotional stimuli in those languages. Given the variety of tasks, designs and measures used in experimental research on the topic (see introduction and the review by Toivo & Scheepers, Reference Toivo, Scheepers, Mavrou, Pérez Serrano and Dewaele2022) it seems unlikely that such an assessment of the predictive validity of RER-LX can be achieved within a single ‘definitive’ study. Hence, we encourage interested members of the research community to make extensive use of RER-LX in their own future research on questions related to emotional resonance reduction in bilinguals’ LX. Over time, this should yield a more comprehensive sense of what the scale can and cannot predict, and in turn, may have useful implications for further theory development and refinement in this area.
5. Conclusion
Compared to previous psychometric attempts at measuring the same construct, RER-LX should allow for more reliable and accurate testing of theories about the underlying reasons for reduced emotional resonance in LX, and how it may dynamically develop over time as bilinguals may experience changes in daily exposure to and/or usage of their languages. The scale was intentionally designed to be used alongside other measures, to enable the testing of complex research questions about bilingual emotions; with 15 items (Appendix 3), the scale is reasonably brief, and it provides a single score to give an indication of participants’ emotional resonance reduction in LX relative to their L1. A higher composite score on the scale suggests that the participant feels a stronger emotional discrepancy between their L1 and LX, in the direction of reduced emotional resonance in LX.
The scale instructions can be modified to reflect participants’ L1 and a specific LX (see, e.g., Study 2 where all participants were asked to respond by comparing their L1 with English as their LX), or to allow for more flexibility by having participants to compare their L1 with an LX they determine themselves (e.g., Study 1 where participants were prompted to think of their L1 and the LX they use most frequently). This not only accounts for specific research questions, which compare two fixed languages, but is also suitable for larger-scale studies with participants from a variety of different language backgrounds and supports the idea of studying bilingualism as a dynamic concept.
On a final note, it is important to keep in mind that the fifteen RER-LX items (Appendix 3) always encourage participants to make a comparison between L1 and LX. This means that the scale does not measure emotional resonance in LX in an absolute sense, but rather in relation to L1. For any given participant, the overall RER-LX test score is therefore to be interpreted as the amount of emotional resonance reduction in their LX relative to their L1.
To conclude, use of the RER-LX scale is recommended for any study examining the emotional resonance of the languages of bi- and multilinguals. We hope that this new scale will facilitate the investigation of emotional resonance in psychology and applied linguistics.
Appendix 1
Appendix 1a. The 87 L1s reported by participants in Study 1:
Afrikaans, Arabic, Armenian, Asante Twi, Bahasa Indonesia, Bahasa Melayu, Balinese, Basque, Bengali, Bosnian, Brunei, Malay, Bulgarian, Cantonese, Catalán, Common Moroccan language, Croatian, Czech, Danish, Dhivehi, Dutch, English, Estonian, Fante, Faroese, Farsi, Filipino, Finnish, French, Frisian, Galician, German, Greek, Hakka, Haya, Hebrew, Hindi, Hokkien, Hungarian, Indonesian, Italian, Japanese, Java, Komi, Korean, Kurdish, Latvian, Lithuanian, Louisiana Creole, Makassarese, Malayalam, Malay, Maltese, Mandarin, Mauritian creole, Mongolian, Moroccan Darija, Norwegian, Odia, Paltienski, Persian, Pidgin English, Polish, Portuguese, Punjabi, Romanian, Russian, Sarawak, Scottish Gaelic, Serbian, Silesian, Sinhala, Slovak, Slovene, Spanish, Sundanese, Suzhou dialect, Swahili, Swedish, Swiss German, Tamil, Teochew, Tswana, Turkish, Turkmen, Ukrainian, Urdu, Valencian, Vietnamese, and Welsh.
Appendix 1b. The 84 LXs reported by participants in Study 1:
Afrikaans, American Sign Language, Arabic, Bahasa Indonesia, Bahasa Malaysia, Basque, Bengali, British Sign Language, Bulgarian, Cantonese, Castilian, Catalan, Catalan sign Language, Chilean Sign Language, Creole, Croatian, Czech, Danish, Dutch, English, Esperanto, Estonian, Farsi, Finnish, French, Gaelic, Galician, German, German Sign Sanguage, Greek, Haya, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Java, Kannada, Kiswahili, Korean, , Latvian, Lithuanian, Lule Sami, Malay, Mandarin, Minang, Nepali, Norwegian, Occitan, Persian, Polish, Portuguese, Punjabi, Putonghua, Romanian, Russian, Russian Sign Language, Sepedi, Serbian, Sign Language of the Netherlands, Sindhi, Sinhala, Slovak, South Sotho, Spanish, Spanish Sign Language, Standard Malay, Sundanese, Swahili, Swedish, Tamil, Telugu, Thai, Tsonga, Turkish, Ukrainian, Urdu, Vietnamese, Welsh, Zaza, and Zulu.
Appendix 1c. The 52 L1s reported by participants in Study 2:
Albanian, Arabic, Bangla, Bengali, Bulgarian, Cantonese, Catalan, Chinese, Croatian, Danish, Dutch, Farsi, Finnish, French, Scottish Gaelic, German, Greek, Gujarati, Hindi, Hungarian, Indonesian, Italian, Japanese, Khmer, Korean, Latvian, Lebanese, Lithuanian, Luganda, Malay, Malayalam, Mandarin, Mongolian, Nepali, Persian, Piedmontese. Polish, Portuguese, Punjabi, Romanian, Russian, Serbian, Shona, Slovak, Somali, Spanish, Tamil, Thai, Turkish, Ukrainian, Urdu, and Welsh.
Appendix 2
The initial 22 candidate items for Study 1, with cited sources of inspiration. Each item came with a 6-point Likert scale on which participants had to indicate their degree of agreement (1 = strongly disagree to 6 = strongly agree) with the relevant item statement.
Appendix 3
The final 15 RER-LX items. Each item comes with a 6-point Likert scale on which participants have to indicate their degree of agreement (1 = strongly disagree to 6 = strongly agree) with the relevant item statement. Ratings for item Q4 (Q7 from the pilot scale) need to be reverse scored (R). Z-standardisation of items is unnecessary – raw and standardized alphas for the scale were virtually the same in all analyses, small Ns included.
Note: In both of our studies, order of items was randomly determined for each participant.
