1.1. The impact of word characteristics on L2 vocabulary knowledge

Research has identified several word-related factors that predict the effort needed to learn a word (Peters, Reference Peters and Webb2020). In this study, we focus on word use characteristics such as frequency, word meaning characteristics (e.g., concreteness), word form characteristics (e.g., word length, part of speech) and interlanguage effects (e.g., cross-linguistic similarity).

Although more factors (e.g., polysemy, L1 frequency) have been identified to date (see Peters, Reference Peters and Webb2020, for a recent overview), we discuss only the factors that are most relevant in the context of receptive vocabulary learning in beginner adult migrant L2 learners with diverging educational and L1 backgrounds.

Word frequency – as inferred from linguistic corpora – is perhaps the most-researched word-level variable in vocabulary research. There is general consensus that vocabulary knowledge is built primarily on word occurrence statistics (Durrant et al., Reference Durrant, Siyanova-Chanturia, Kremmel and Sonbul2022). To learn a word, learners typically need sufficient encounters with that word (Pellicer-Sánchez, Reference Pellicer-Sánchez2016). As such, the more frequent a word is (i.e., the more commonly used it is), the more likely it is that learners will encounter it more frequently. These words are thus more likely to be acquired quickly (Durrant et al., Reference Durrant, Siyanova-Chanturia, Kremmel and Sonbul2022) and to be known to learners (Crossley et al., Reference Crossley, Kyle and Salsbury2016). This word frequency effect may be even more pronounced for L2 words than it is for L1 words (Durrant et al., Reference Durrant, Siyanova-Chanturia, Kremmel and Sonbul2022) and has been found in university-educated learners (e.g., Crossley et al., Reference Crossley, Kyle and Salsbury2016), primary school students (e.g., Puimège & Peters, Reference Puimège and Peters2019a) and preliterate children (e.g., Verhagen et al., Reference Verhagen, Van Stiphout and Blom2022). It is important to note, however, that although many studies have confirmed the impact of word frequency on vocabulary knowledge, its role has been nuanced in recent years. Schmitt et al. (Reference Schmitt, Dunn, O’Sullivan, Anthony and Kremmel2021), for instance, found very low correlations between frequency rankings and knowledge rankings of individual words, albeit only for form recall and not for form or meaning recognition. De Wilde’s study (2023), too, revealed that word frequency was not a significant predictor of English vocabulary knowledge among Dutch-speaking primary school pupils. She hypothesized that the high prevalence of English-Dutch cognates may mitigate frequency effects (see below). Additionally, the frequency effect seems to be much stronger for words that are acquired through written input (i.e., reading) compared to spoken input (i.e., listening) (Peters & Webb, Reference Peters and Webb2018; Vidal, Reference Vidal2011).

Other word-level factors that can predict vocabulary knowledge are semantic or phonological in nature. One of the semantic factors affecting L2 vocabulary knowledge is word concreteness (Laufer, Reference Laufer, Schmitt and McCarthy1997). Brysbaert et al. (Reference Brysbaert, Stevens, De Deyne, Voorspoels and Storms2014, p. 5) define concreteness as ‘the degree to which a concept denoted by a word refers to a perceptible entity’. The main theory guiding this principle is the dual coding theory (Paivio, 2013, in Brysbaert et al., Reference Brysbaert, Stevens, De Deyne, Voorspoels and Storms2014). It is assumed that upon hearing a concrete word, such as ‘apple’, perceptual memory of the object that the word refers to, is activated parallel to the word-meaning mapping. Because of dual coding, concrete words are considered easier to remember and to learn than more abstract words. Various empirical studies with both university-educated and primary school students appear to confirm the dual coding hypothesis (e.g., Brysbaert et al., Reference Brysbaert, Stevens, De Deyne, Voorspoels and Storms2014; Keuleers, Brysbaert & New, Reference Keuleers, Brysbaert and New2010; Puimège & Peters, Reference Puimège and Peters2019a; Verhagen et al., Reference Verhagen, Van Stiphout and Blom2022).

Another factor that influences vocabulary knowledge is word length. Typically, longer words take longer to articulate, which results in a disadvantage in phonological rehearsal and thus storage of the novel word (Nishiyama, Reference Nishiyama2020). In general, word length seems to impact learning, with learners being typically more likely to know more short (i.e., fewer phonemes, or letters) rather than long words (e.g., Ellis & Beaton, Reference Ellis and Beaton1993; Willis & Ohashi, Reference Willis and Ohashi2012). However, not all studies confirm these findings, suggesting a modality effect or a differential effect of word length on different aspects of vocabulary knowledge. In an incidental vocabulary learning experiment from multimodal input with young EFL learners and form recall as outcome measure, Puimège & Peters (Reference Puimège and Peters2019b) found that longer words were more likely to be acquired than shorter words. They argued that the salience of longer words in spoken input could explain this result (Crossley et al., Reference Crossley, Kyle and Salsbury2016). Similar to the effect of word frequency (cf. Peters & Webb, Reference Peters and Webb2018), there might also be a modality effect on the relationship between word length and vocabulary knowledge. Additionally, Barclay & Pellicer-Sánchez (Reference Barclay and Pellicer-Sánchez2021) investigated the impact of L2 word length on learning burden among 48 EFL university students with different L1’s. Their findings indicated that learners encountered greater difficulty in recalling the form of longer words compared to shorter ones, although the effect of word length on form recognition was less prominent. Notably, their study did not assess meaning recognition.

Part of speech (i.e., a word’s grammatical category or word class) (PoS) is also mentioned as a factor that influences L2 vocabulary knowledge, but empirical evidence is relatively thin (Durrant et al., Reference Durrant, Siyanova-Chanturia, Kremmel and Sonbul2022; Peters, Reference Peters and Webb2020). Typically, young children acquire nouns before adjectives, verbs and adverbs (Gentner, Reference Gentner1982; Goodman et al., Reference Goodman, Dale and Li2008; Crossley et al., Reference Crossley, Kyle and Salsbury2016). This trend is especially pronounced in Indo-European languages, as recent studies indicate that languages in which verbs are more salient, like Turkish, Korean, Mandarin or Tseltal, children show a verb rather than a noun bias (e.g., Casillas et al., Reference Casillas, Foushee, Méndez Girón, Polian and Brown2024; Frank et al., Reference Frank, Braginsky, Yurovsky and Marchman2021; Setoh et al., Reference Setoh, Cheng, Bornstein and Esposito2021 Yee, Reference Yee2020). Supporting this pattern, both van Zeeland & Schmitt (Reference van Zeeland and Schmitt2013) and Reynolds et al. (Reference Reynolds, Wu, Liu, Kuo and Yeh2015) found that nouns are more easily acquired than verbs in English as L2. Puimège and Peters (Reference Puimège and Peters2019a) explained this effect by the fact that verbs can occur in many more different forms than nouns, which may hamper acquisition. Additionally, nouns typically refer to more concrete entities, which may also facilitate their acquisition (cf. dual coding principle, see above; Puimège & Peters; Reference Puimège and Peters2019a). However, a recent study that examined the acquisition of pseudowords and German words by English speakers, while manipulating both concreteness and grammatical class, found no evidence to support the latter claim (Martin & Tokowicz, Reference Martin and Tokowicz2020). This suggests a more distinct separation between the effects of concreteness and grammatical class.

A final word-level characteristic that is not necessarily learner-independent, but which has also been found to affect word processing and L2 acquisition is cross-linguistic similarity (e.g., Schepens et al., Reference Schepens, van der Slik and van Hout2016). Since both L1 and L2 words are probably stored into the same mental lexicon (Durrant et al., Reference Durrant, Siyanova-Chanturia, Kremmel and Sonbul2022), cross-linguistic influence may affect vocabulary knowledge. Cognates are a well-known representation of such similarities (cf. the cognate facilitation effect, Lemhöfer, Dijkstra & Michel, Reference Lemhöfer, Dijkstra and Michel2004). In this study, we consider a word a cognate when it is similar in terms of orthography and pronunciation in both the L1 and L2, following the definition by Zhang & Zhang (Reference Zhang and Zhang2022). Several studies with primary school students of English as a foreign language have shown that overlap, or even partial overlap between L1 and L2 words, benefits vocabulary learning (e.g., De Wilde et al., Reference De Wilde, Brysbaert and Eyckmans2020; Puimège & Peters, Reference Puimège and Peters2019a). The cognate effect is typically larger for knowledge at the level of meaning recall compared to word recognition and in spoken texts compared to written texts (Peters & Webb, Reference Peters and Webb2018; Peters, Reference Peters and Webb2020).

It is important to note that all studies reviewed thus far have primarily focused on either university-educated students (e.g., Barclay & Pellicer-Sánchez, Reference Barclay and Pellicer-Sánchez2021; Crossley et al., Reference Crossley, Kyle and Salsbury2016; Martin & Tokowicz, Reference Martin and Tokowicz2020; Peters & Webb, Reference Peters and Webb2018; Reynolds et al., Reference Reynolds, Wu, Liu, Kuo and Yeh2015; Willis & Ohashi, Reference Willis and Ohashi2012; van Zeeland & Schmitt, Reference van Zeeland and Schmitt2013) or primary school students (De Wilde et al., Reference De Wilde, Brysbaert and Eyckmans2020; De Wilde, Reference De Wilde2023; Puimège & Peters, Reference Puimège and Peters2019a). Consequently, it remains uncertain whether these findings can be generalized to adult L2 learners with no or less extensive schooling in their L1.

1.2. L2 vocabulary learning and the role of educational background

Educational background, defined as the level of formal schooling an L2 learner has received in their L1 (Tarone, Reference Tarone2010), is inherently linked to literacy skills, given that schooling is a literate activity (Huettig & Mishra, Reference Huettig and Mishra2014; UNESCO, 2022). Despite the sparse research on the influence of educational background and L1 literacy on L2 vocabulary acquisition, emerging studies underscore the significant impact these factors have. Deygers & Vanbuel (Reference Deygers and Vanbuel2022) revealed that learners with limited education, or LESLLA learners, exhibit markedly lower receptive vocabulary in L2 compared to their more educated peers, even when accounting for other variables such as L1, age, or out-of-school exposure.

As has been argued, this discrepancy is related to the lower metalinguistic abilities of LESLLA learners (Kurvers, Reference Kurvers2015). In order to learn words, learners rely on the phonological loop, a working memory component which helps to decode, store and maintain novel words (Baddeley et al., Reference Baddeley, Logie, Nimmo-Smith and Brereton1985). Yet, the capacity to utilize phonological information effectively when encountering new spoken words varies significantly across individuals with different levels of schooling (Baddeley et al., Reference Baddeley, Logie, Nimmo-Smith and Brereton1985; Demoulin & Kolinsky, Reference Demoulin and Kolinsky2016). Literacy fosters the development of explicit visual language representations, enhancing word recognition and the connection between word forms and their meanings (Kolinsky, Reference Kolinsky, Pollatsek and Treiman2015). This advancement in visual representation aids not only in written language processing but also improves oral language processing (Morais, Alegria & Content, Reference Morais, Castro, Scliar-Cabral, Kolinsky and Content1987; Huettig & Mishra, Reference Huettig and Mishra2014). Schooling and literacy, rather than age, are pivotal in developing the ability to segment words from spoken language, as seen in emergent literate children. Initially processing language in larger units, children begin to discern individual words as their literacy develops (Justino & Kolinsky, Reference Justino and Kolinsky2023; Havron & Arnon, Reference Havron and Arnon2017). Similar patterns have been found in non- and low-literate adults. For example, in a visual world eye-tracking experiment, Huettig, Singh and Mishra (Reference Huettig, Singh and Mishra2011) explored the extent to which adults with high (i.e., 15 years on average) and low levels (i.e., 2 years) of schooling relied on phonological or semantic information to map words onto objects. The study revealed that individuals with higher levels of education and literacy exhibit greater proficiency in utilizing phonological information for word-object mapping, whereas lower-educated adults struggle to employ this information effectively and primarily relied on semantic cues in the input. Moreover, Kurvers’ (Reference Kurvers2015) comparative research on the metalinguistic skills of non-literate L2 learners, L2 low-educated literates and young children illuminated larger differences between unschooled and reading adult L2 learners than in unschooled adult L2 learners and preliterate L1 children. In particular, non-literate L2 learners appeared to have difficulty judging word length, syllogisms and segment sentences into isolated words and words into phonemes.

Furthermore, research indicates that non-literate and low-literate adults face challenges not only in processing spoken language but also in storing and retrieving linguistic information. A series of elicited imitation and lexical decision experiments indicated that L1 adults with reduced schooling face difficulties with phonological tasks, such as repeating pseudowords, compared to tasks involving familiar words, where semantic knowledge can offset phonological limitations (Kosmidis et al., Reference Kosmidis, Folia, Lahou and Kiosseoglou2004, Reference Kosmidis, Tsapkini and Folia2006). Similarly, Deygers (Reference Deygers2020) found that LESLLA learners struggle with repeating phonological structures or pseudowords, whereas they performed similarly to university-educated learners on tasks that required the repetition of existing words. In summary, these findings illustrate the profound influence of formal education and literacy on metalinguistic abilities and the capacity of the phonological loop, which are crucial for learning new words (Godfroid et al., Reference Godfroid, Boers and Housen2013; Nation, Reference Nation2001).

1.3. Differential effects of word-level variables

The reviewed studies in 1.1 suggest that various word-level factors significantly influence the learnability of words, thus serving as reliable predictors of vocabulary knowledge in L2 learners. However, these word-based effects have been shown to vary in magnitude depending on learner-level variables. Indeed, individual differences in these effects (i.e., word frequency, cognateness, concreteness, PoS) have been reported in several studies.

Puimège & Peters (Reference Puimège and Peters2019a) investigated the effects of age, exposure to L2 output outside the classroom and word features on L2 receptive vocabulary knowledge in 10–12-year-old Dutch learners of English as an L2, using the Peabody Picture Vocabulary Test (PPVT). They discovered that both the frequency effect and the effect of cognateness increased with age, suggesting that L2 vocabulary knowledge in terms of meaning recognition is more influenced by these factors in older learners. These findings were attributed to the expanded vocabulary size of the learners, who were still beginning L2 learners at the time of data collection. Consistent with psycholinguistic studies, the word frequency effect first increases with increased language exposure and then decreases again as the learner becomes more proficient in the language (Brysbaert et al., Reference Brysbaert, Stevens, Mandera and Keuleers2016).

Similarly, De Wilde et al. (Reference De Wilde, Brysbaert and Eyckmans2020) explored the relationship between word-related factors, proficiency level and L2 vocabulary knowledge in young EFL learners. They identified a significant and positive interaction effect between word frequency and L2 proficiency, indicating a stronger frequency effect in more proficient L2 learners. Drawing parallels with findings in L1 adults (e.g., Brysbaert et al., Reference Brysbaert, Mandera and Keuleers2018), De Wilde et al. (Reference De Wilde, Brysbaert and Eyckmans2020) proposed that this effect may be attributed to increased exposure to high-frequency words. Additionally, they uncovered a significant interaction effect between word concreteness and L2 proficiency, suggesting that concreteness is a better predictor of vocabulary knowledge in low-proficient L2 learners, aligning with the idea that concrete words are easier to learn. Furthermore, a reversed interaction effect between L2 proficiency and cognates was observed, indicating that less proficient learners tend to rely on their L1 to infer the meaning of L2 words. No differential effects were found of PoS or word length.

What remains less clear is whether these effects vary depending on learners’ educational background in the L1. Given the crucial role of phonological abilities in acquiring novel words (Baddeley et al., Reference Baddeley, Gathercole and Papagno1998) and their apparent enhancement with schooling and literacy acquisition (e.g., Hu, Reference Hu2013; Demoulin & Kolinksy, Reference Demoulin and Kolinsky2016; Kurvers, Reference Kurvers2015), the interaction effect between word-level variables and educational background warrants further examination (Martin & Tokowicz, Reference Martin and Tokowicz2020). To date, differential effects of word-level variables in LESLLA learners and their higher-educated counterparts have not yet been explored. However, research with L1 children and non- or low-literate L1 adults provides intriguing insights into the word length effect and the word frequency effect.

For instance, Morra & Camba (Reference Morra and Camba2009) demonstrated that phonological sensitivity predicts the learning of long nonwords in 10–11-year-old children. Additionally, children with phonological processing difficulties did not exhibit sensitivity to the word length effect in experiments conducted by Palladino & Ferrari (Reference Palladino and Ferrari2008). They conducted experiments with EFL children who encounter challenges in learning foreign languages, all of whom experienced difficulties with phonological processing and phonological memory, indicating a connection between phonological abilities and the word length effect.

The frequency effect appears to be less influenced by learners’ phonological abilities (Brysbaert et al., Reference Brysbaert, Stevens, Mandera and Keuleers2016). However, given that LESLLA learners possess a smaller vocabulary (Deygers & Vanbuel, Reference Deygers and Vanbuel2022) and that the frequency effect is generally weaker in less proficient learners due to reduced exposure, it might operate differently across L2 learners with varying educational backgrounds. For example, it might not impact LESLLA learners as much as it does tertiary-educated L2 learners, who likely have additional exposure to language through written input. Moreover, as suggested by Brysbaert et al. (Reference Brysbaert, Stevens, Mandera and Keuleers2016), meanings of some low-frequency words can be inferred by analyzing the word forms since these words are derivations, inflections or compounds of high-frequency words. Consequently, this may give an advantage to higher-educated L2 learners with stronger L1 literacy skills and more robust metalinguistic abilities.

3.1. Participants

Participants of this study were adult migrants who were enrolled in a formal Dutch L2 course. Data were collected from 1020 adult L2 learners from 11 schools in Flanders, Belgium. Depending on their educational background, estimated by means of a cognitive ability test which provides an indication of level of schooling (Verschueren et al., Reference Verschueren, Buyse, Germeijs, Janssen, Magez, Van Nijlen, Buysse, Vangoetsenhoven, Arkens and Doumen2011) and L1 literacy skills in a Latin script, which are estimated based on thorough intake conversations by experts in the learners’ L1 (De Niel et al., 2016), participants enroll in a so-called extended, standard and accelerated track. Extended tracks are specifically organized for students who have had limited or interrupted experience with formal schooling in their home country and obtained at most a primary education degree (i.e., LESLLA learners). These learners are emergent functional literates in an alphabetic script, meaning they can likely read and understand isolated words or even sentences, but they struggle to comprehend extended written text in any language they speak. Accelerated tracks typically target students with a tertiary education degree, and standard tracks typically address learners with a degree of lower or upper secondary education (see Deygers & Vanbuel, Reference Deygers and Vanbuel2022, for more details).

Table 1 provides an overview of the background variables of participants in this study, ranked by track type (i.e., educational background and L1 literacy level). Half of the students were taking an A1-level Dutch course at the time of data collection; the other half was enrolled in an A2-level course. Data from 134 students were omitted since no information on their educational background, length of residence or employment was available. Additionally, data from 123 learners who did not pass the level of the Dutch L2 course were removed in order to control for L2 proficiency.

Table 1.

Descriptives of the participants by educational background

3.2. Measures

Participants’ vocabulary knowledge in Dutch was measured by means of the Peabody Picture Vocabulary Test in Dutch (PPVT-NL, Dunn & Dunn, Reference Dunn and Dunn2005). The PPVT-NL measures students’ receptive knowledge at the level of meaning recognition by means of recorded single-word items with multiple-choice questions. Participants hear a word and select one of four drawings that best matches the prompt. Sets 1 through 9 were administered, resulting in a total number of 108 items per participant. Four words were excluded from the analyses because they were multiword expressions. For these words, no information on the word-level parameters was available in the corpora from which we subtracted this information. Cronbach’s alpha for the remaining 104 items was .86.

Student background information was gathered by means of a brief written questionnaire with questions formatted in Dutch, French, English, Spanish, Russian, Polish, Turkish, Pashto, Arabic and Mandarin. In the extended tracks, all questionnaires were administered one-on-one to avoid difficulties with reading and interpretation of the questions. Participants were asked about their L1(s) and home languages, their age, gender, length of residence in Belgium, exposure to Dutch tv and radio, educational background and employment.

3.3. Procedure

Students completed the test using paper and pencil. The test was administered during regular classes in 2017 (Deygers, Reference Deygers2023; Deygers & Vanbuel, Reference Deygers and Vanbuel2022). Given the scale of this study, the test was administered in a group rather than individually.

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. Participants were told about the research goals and gave oral informed consent, a procedure that was approved by the Faculty’s Ethical Committee. For more details about the test administration, see Deygers & Vanbuel (Reference Deygers and Vanbuel2022).

3.3. Independent variables

This study operationalizes four word-level measures and eight student-level measures (i.e., educational background and control variables). Cross-linguistic similarity was not operationalized at the word level since the students in the dataset had 69 different L1s. Instead, it was operationalized as a student-level variable. At the word level, we included:

• - Frequency: For each word, we included a frequency score which we obtained from the Subtlex corpus (Keuleers et al., Reference Keuleers, Brysbaert and New2010). This corpus is based on subtitles in movies and series and is thus based on oral input, which is similar to the test format that we used to measure vocabulary knowledge in this study. Frequency scores were included as Zipf scores, which range from 1 (very low-frequency words) to 6 (very high-frequency content words). The frequency scores are log-transformed with base 2 (i.e., frequency per billion words) (Mean = 3.75, SD = .85, min = 1.36, max = 5.90).

• - Part of speech (PoS): For each word, we included whether it was a verb, adjective or noun. Nouns made up the largest part of the words in the test (N = 70, 67.31%), followed by verbs (N = 24, 23.08%) and adjectives (N = 10, 9.62%).

• - Word length: For each word, word length was measured by the number of letters and computed automatically via T-scan (Pander Maat et al., Reference Pander Maat, Kraf, van den Bosch, van Gompel, Kleijn, Sanders and van der Sloot2014). As one reviewer pointed out, operationalizing word length by the number of sounds might be more appropriate given the oral nature of the vocabulary test. However, we chose to use the number of graphemes over phonemes to maintain consistency with similar studies (e.g., De Wilde et al., Reference De Wilde, Brysbaert and Eyckmans2020). Additionally, the correlation between the number of letters and the number of sounds in this study was very high (r = .92, p < .001), indicating that the two measures are closely aligned. On average, the words in our study consisted of 6.66 letters (SD = 2.29) and 5.90 (SD = 2.22) sounds, with a minimum length of 3 letters and 2 sounds and a maximum length of 13 letters and 11 sounds.

• - Concreteness: To determine the concreteness of the items, we used Brysbaert et al.’s (Reference Brysbaert, Stevens, De Deyne, Voorspoels and Storms2014) concreteness values. These values are retrieved from a survey in 75 university students who were mother tongue speakers of Dutch, who rated a total of 6000 lemma for concreteness from 1 (abstract) to 5 (concrete). The mean score for concreteness for the words in our dataset is 4.17 (SD = .77), indicating that they are relatively concrete.

Word-level variables showed slight to moderate correlations. Specifically, word length and frequency exhibited a moderate negative correlation (r = −.420, p < .001), while concreteness and frequency were positively but weakly correlated (r = .095, p < .001). Additionally, word length and concreteness were slightly negatively correlated (r = −.260, p < .001). Although these correlations were statistically significant, none surpassed a threshold of .8, indicating an absence of overall strong correlations that could pose a risk of multicollinearity.

PoS did not show a significant relationship with word frequency (F(2) = .053, p > .05, η² = .001) or word length (F(3) = 2.057, p > .05, η² = .06) either. However, there were significant and large differences observed among words from various grammatical classes in terms of concreteness (F(2) = 46.132, p < .001, η² = .48). These differences will be taken into account in the analyses.

Information on students’ educational background and other background variables were gathered by means of a questionnaire and cross-checked with school data and included the following:

• - Educational background: For each student, a categorical variable was included referring to their educational background (i.e., maximal educational attainment). 21.1% of the students were LESLLA learners (that is no degree or degree of primary education alone), 43.1% had a degree of secondary education and 35.8% had a degree of tertiary education).

• - CEFR level: 49.4% (N = 377) of the participants were enrolled in an A1-level track, 50.6% (N = 386) in an A2-level track.

• - Job: a binary variable was included for employment. 35.6% (N = 272) of the students in the dataset were employed at the time of testing, since out of school exposure to L2 input may affect L2 vocabulary knowledge (e.g., Puimège & Peters, Reference Puimège and Peters2019a).

• - L1–L2 distance: following Jeon and Yamashita (Reference Jeon and Yamashita2014), L1–L2 distance was measured by means of two indices, L1 Indo-European family and L1 Latin script:

  • ○ L1 Indo-European: a dichotomous variable for L1 was included. 52.7% (N = 402) of the students in our dataset had an Indo-European L1.

  • ○ L1 Latin writing system: a dichotomous variable for L1 writing system was also included. Research has shown that familiarity with the target script facilitates L2 word learning (see Shepperd, Reference Shepperd2024, for an overview). 44.8% (N = 342) of the students used an L1 with a Latin writing system. The correlation between Indo-European L1 and Latin writing system was low (Cramer’s V = .252, p < .001).

• - Time in Belgium: A continuous variable was included for the time period the student had lived in Belgium at the time of data collection. On average, participants had been in Belgium for 4.64 years (SD = 5.18) at the time of testing.

• - Age: A continuous variable indicating students’ age at the time of testing was included as well, since age differed significantly across educational background (F(2) = 6.793, p < .001) and the correlation with time in Belgium was low (Pearson’s r = .33, p < .001). On average, students were 34 (SD = 10) years old. The youngest student in the dataset was 14 years old, the oldest 68.

• - Exposure to Dutch-medium TV: A binary variable was included for media usage or exposure to Dutch outside of the classroom, since previous studies indicated that out-of-school exposure can be a strong predictor of L2 vocabulary knowledge (e.g., Puimège & Peters, Reference Puimège and Peters2019a). On average, 65.5% of the participants indicated that they watched television in Dutch.

3.5. Analysis

All participants were administered the same 108 items of the Dutch PPVT. Responses were scored dichotomously: a correct response received a score of 1, an incorrect response a score of 0. Items that were skipped were also scored as 0. To examine the impact of word- and student-level variables on vocabulary knowledge, logistic regression analyses were conducted. The dependent variable in this study was binary, i.e., whether a word is known. Since our data are hierarchically structured (i.e., words within students), we used two-level models for the analyses. This allows us to include random intercepts for words and students, and thus to take into account differences between words and individual learners (Hox, Reference Hox2010). Continuous variables (i.e., age, time in Belgium, frequency, word length, concreteness) were all centered around the mean.

Models were computed stepwise, relying on the Akaike Information Criterion (AIC) and deviance to find the model that best fitted the data. First, empty models only containing random effects for ‘word’ and ‘student’ were included in the models. Intraclass correlation coefficient (ICC) for word was .865 and .135 for students, meaning that the probability of knowing a word varied between 87% across words and 13% across students. As such, a two-level logistic regression model was suited to analyze our data.

Next, word-level variables ‘frequency’, ‘concreteness’, ‘part of speech’ and ‘word length’ were added to the model to answer RQ1. All variables were centered around the mean. The model improved significantly when ‘frequency’, ‘concreteness’ and ‘word length’ and ‘PoS’, were added to the model. Yet, since there was a significant relationship between PoS and concreteness (F(2) = 42085.697, p < .001) and since the number of adjectives in the dataset was limited, PoS was omitted from further analyses. Table S1 in the Supplementary Materials Online provides an overview of the model fit comparison.

As a third step, ‘educational background’ was added to the model, together with other relevant student-level variables (i.e., level, age, length or residence, L1, employment, exposure to Dutch). Only those variables that significantly contributed to the model, were retained in order to avoid overfitting. The table S2 with an overview of model comparisons can be found in the Supplementary Materials.

In a fourth step, interaction effects were computed between word-level variables and educational background to examine a differential impact of these variables for students with diverging educational backgrounds (RQ2). Data were analyzed in R Version 4.2.2 (R Core Team, 2022) using the ‘glmer’ function from the package ‘lme4’ (Bates et al., Reference Bates, Maechler, Bolker and Walker2015).