MLU as a measure of child language development

The idea that average sentence length might be a way of assessing children’s language development was first proposed by Nice (Reference Nice1925). R. Brown’s (Reference Brown1973) use of MLU as a simple index of the development of constructional complexity in children’s early language was the impetus for wider uptake. He proposed five stages of children’s early morphosyntactic development using MLU and the longest sentence produced at specific ages. Since then, MLU has been extensively applied in child language research and clinical practice (Parker & Brorson, Reference Parker and Brorson2005).

Although developed for English, studies using MLU have been done on a variety of different languages, including Afrikaans (Oosthuizen & Southwood, Reference Oosthuizen and Southwood2009); Basque (Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018); Bengali (Gouda et al., Reference Gouda, Kumar, Sarkar, Rashmi, Chatterjee and Pani2020); Canadian Inuktitut (Allen & Dench, Reference Allen and Dench2015); Greek (Voniati, Reference Voniati2016); Icelandic (Thordardottir & Weismer, Reference Thordardottir and Weismer1998); Irish (Hickey, Reference Hickey1991); Mandarin Chinese (Wu, Reference Wu2020); Spanish (Jackson-Maldonado & Conboy, Reference Jackson-Maldonado, Conboy, Centeno, Anderson and Obler2007); and Turkish (Ege, Reference Ege, Topbaş and Yavaş2010). MLU has also been used to address questions about bilingual children’s development (see Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018; Marchman et al., Reference Marchman, Martínez‐Sussmann and Dale2004; Meisel, Reference Meisel2011; Thordardottir, Reference Thordardottir2005).

Overall, MLU has been found to be a reliable measure correlating with age and other measures of general language development (Dethorne et al., Reference Dethorne, Johnson and Loeb2005; Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018; Rice et al., Reference Rice, Redmond and Hoffman2006). MLU has also been usefully applied to children with language impairments (Dethorne et al., Reference Dethorne, Johnson and Loeb2005; Parker & Brorson, Reference Parker and Brorson2005; Rice et al., Reference Rice, Redmond and Hoffman2006; Wieczorek, Reference Wieczorek2010; Wu, Reference Wu2020). However, MLU may not be a valid measure of development at all ages (see Eisenberg et al., Reference Eisenberg, Fersko and Lundgren2001, for a discussion). R. Brown (Reference Brown1973) suggests that after Stage 5, at an MLU of four morphemes, the nature of an interaction will shape children’s responses more than their underlying linguistic knowledge of a language system (see Southwood & Russell, Reference Southwood and Russell2004; Tolentino, Reference Tolentino2022).

Several scholars also argue that, in older children, further linguistic complexity may involve internal embedding rather than additional length (Oosthuizen & Southwood, Reference Oosthuizen and Southwood2009) and a decrease in morphological growth (Parker & Brorson, Reference Parker and Brorson2005). Various studies have reported different ages at which MLU may no longer be a sensitive measure (see Allen & Dench, Reference Allen and Dench2015; Cheung, Reference Cheung1998; Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018; Oosthuizen & Southwood, Reference Oosthuizen and Southwood2009; Rice et al., Reference Rice, Redmond and Hoffman2006, Reference Rice, Smolik, Perpich, Thompson, Rytting and Blossom2010; Scarborough et al., Reference Scarborough, Rescorla, Tager-Flusberg, Fowler and Sudhalter1991; Tomas & Dorofeeva, Reference Tomas and Dorofeeva2019; Wu, Reference Wu2020).

When using spontaneous language samples, the number of children’s utterances needed to accurately measure MLU has also been debated (see Eisenberg et al., Reference Eisenberg, Fersko and Lundgren2001, for a discussion). Pragmatic variation means that children may respond with single-word answers in interactions because it is not necessary for them to use additional words. Moreover, a single interactive event may not comprehensively represent a child’s language abilities (Oosthuizen & Southwood, Reference Oosthuizen and Southwood2009; Southwood & Russell, Reference Southwood and Russell2004). Although R. Brown (Reference Brown1973) recommended using samples of 50 to 100 utterances, some studies have cautioned against using samples containing fewer than 175 utterances (Gavin & Giles, Reference Gavin and Giles1996; Oosthuizen & Southwood, Reference Oosthuizen and Southwood2009; Tomas & Dorofeeva, Reference Tomas and Dorofeeva2019).

Another variant of MLU is MLU3, based on the longest sentences a child currently produces. Studies that use MLU3 have found it to be a useful and valid tool for calculating utterance length as a measure of child language development (Allen & Dench, Reference Allen and Dench2015; Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018; Fenson et al., Reference Fenson, Dale, Reznick, Bates, Thal, Pethick, Tomasello, Mervis and Stiles1994; Heilmann et al., Reference Heilmann, Weismer, Evans and Hollar2005; Jackson-Maldonado & Conboy, Reference Jackson-Maldonado, Conboy, Centeno, Anderson and Obler2007). Parent report studies find significant correlations between MLU3, age and other indices of language growth such as vocabulary and syntactic development (Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018). Jackson-Maldonado and Conboy (Reference Jackson-Maldonado, Conboy, Centeno, Anderson and Obler2007) suggest that parent report is more efficient as it can “bypass the performance limitations inherent in spontaneous and structured language sampling” (p. 147).

During early childhood, children show considerable variation in language skills when considered by age (Frank et al., Reference Frank, Braginsky, Yurovsky and Marchman2021). MLU correlates significantly with age, but there is also considerable variability in MLU at specific ages (Dethorne et al., Reference Dethorne, Johnson and Loeb2005; Tomas & Dorofeeva, Reference Tomas and Dorofeeva2019; Vasilyeva et al., Reference Vasilyeva, Waterfall and Huttenlocher2008). R. Brown (Reference Brown1973) has argued that MLU is a more accurate tool for predicting children’s expressive language abilities than age because MLU permits the identification of children at similar levels of constructional complexity regardless of chronological age. He holds that “two children matched for MLU are much more likely to have speech that is, on internal grounds, at the same level of constructional complexity than are two children of the same chronological age” (R. Brown, Reference Brown1973, p. 55).

Some studies based on natural language samples report significantly high correlations between MLU-m and age, supporting the use of MLU-m as a general index of language development (Miller & Chapman, Reference Miller and Chapman1981, r = .88; Ege, Reference Ege, Topbaş and Yavaş2010, r = .81; Thordardottir & Weismer, Reference Thordardottir and Weismer1998, r = .84). However, other studies report that MLU-m is significantly but not necessarily always strongly correlated with age, with r scores between .3 and .8 (see Allen & Dench, Reference Allen and Dench2015). Several studies also show that MLU-m has higher correlations with other indices of development than age does (Dromi & Berman, Reference Dromi and Berman1982; Thordardottir & Weismer, Reference Thordardottir and Weismer1998).

Significant correlations are found between MLU and other indices of language development. For English, Dethorne et al. (Reference Dethorne, Johnson and Loeb2005) found that MLU-m is associated with both lexical and morphosyntactic development, with lexical development – measured as the number of different words – strongly correlating with MLU-m and accounting for 51% variance in MLU-m. Ezeizabarrena and Garcia Fernandez (Reference Ezeizabarrena and Garcia Fernandez2018) also report high correlations between MLU and other scales of communicative development, suggesting that MLU is a general measure of early development rather than a measure of one specific component such as semantics or morphosyntax. However, in some studies, MLU correlated better with measures of grammatical development and therefore may be a more reliable indicator of grammatical ability when compared to other indices (Allen & Dench, Reference Allen and Dench2015; Miller & Chapman, Reference Miller and Chapman1981; Thordardottir & Weismer, Reference Thordardottir and Weismer1998).

Measurement of MLU in words versus morphemes

A central issue has been whether to measure MLU in words or morphemes. R. Brown (Reference Brown1973) argued that counting morphemes is a more accurate method as it takes inflectional complexity into account. With a language like English that has lower inflection and is morphologically sparse, there may be little difference between calculating MLU in words or morphemes (Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018). For instance, Parker and Brorson (Reference Parker and Brorson2005) found high correlations (r = .998) between MLU scales (MLU-m and MLU-w) and age (r = .69 for both) in English-speaking children (3;0–3;10), concluding that MLU-w is as effective a measure as MLU-m. Similarly, there appears to be minimal difference between measurements in words versus morphemes in Cantonese and Mandarin Chinese – isolating languages with little inflection (Allen & Dench, Reference Allen and Dench2015; Cheung, Reference Cheung1998; Klee et al., Reference Klee, Stokes, Wong, Fletcher and Gavin2004; Wu, Reference Wu2020). Similar results have also been found for more inflectional languages such as Dutch (Arlman-Rupp et al., Reference Arlman-Rupp, van Niekerk de Haan and van de Sandt-Koenderman1976), Icelandic (Thordardottir & Weismer, Reference Thordardottir and Weismer1998) and Irish (Hickey, Reference Hickey1991), with correlations of between .98 and .99 for MLU-w and MLU-m. Basque, an agglutinative and morphologically rich language, also reports a high correlation between MLU3-m and MLU3-w (r = .97) and strong correlations between vocabulary, nominal and verbal morphology, and MLU3-m and MLU3-w with an r range .81-.97 (reduced to .66 - .95 when controlling for age (Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018)).

Based on significantly high correlations between MLU-m and MLU-w, several studies conclude that MLU-w provides a better alternative to MLU-m for tracking children’s language development trajectories because it is easier to count words (Jackson-Maldonado & Conboy, Reference Jackson-Maldonado, Conboy, Centeno, Anderson and Obler2007; Parker & Brorson, Reference Parker and Brorson2005; Thordardottir & Weismer, Reference Thordardottir and Weismer1998). However, whereas measurements using words or morphemes might be similar in some languages, MLU-w is always likely to be equal to, or slightly lower than, MLU-m (Gouda et al., Reference Gouda, Kumar, Sarkar, Rashmi, Chatterjee and Pani2020, for Bengali; Oosthuizen & Southwood, Reference Oosthuizen and Southwood2009). Wieczorek (Reference Wieczorek2010) challenges the conclusion that high correlations between MLU-w and MLU-m mean either can be used, arguing that MLU-w is related to lexical development whereas MLU-m measurements are better indicators of grammatical development.

Some studies claim that a morpheme count is more accurate in morphologically rich languages such as synthetic languages like Russian (Tomas & Dorofeeva, Reference Tomas and Dorofeeva2019) and Hebrew (Dromi & Berman, Reference Dromi and Berman1982); polysynthetic languages like Canadian Inuktitut (Allen & Dench, Reference Allen and Dench2015); or agglutinative languages like Turkish (Ege, Reference Ege, Topbaş and Yavaş2010). Allen and Dench (Reference Allen and Dench2015) found that MLU-w had no significant correlation with age in Inuktitut, a polysynthetic language. They found that MLU-m and Mean Length of Words measured in morphemes and syllables, respectively, correlated significantly with age. However, some studies on morphologically rich languages (such as Basque) report that both MLU-m and MLU-w indicate children’s expressive language skills equally well although MLU-w is a consistently lower measure (Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018).

Identifying and counting morphemes has its challenges. In synthetic languages like Spanish, Finnish, Dutch, Italian and Icelandic, one morpheme may contain more than one grammatical feature such as tense and person (see Allen & Dench, Reference Allen and Dench2015; Arlman-Rupp et al., Reference Arlman-Rupp, van Niekerk de Haan and van de Sandt-Koenderman1976; Jackson-Maldonado & Conboy, Reference Jackson-Maldonado, Conboy, Centeno, Anderson and Obler2007; Parker & Brorson, Reference Parker and Brorson2005, for discussions). Consequently, Allen and Dench (Reference Allen and Dench2015) argue that we may underestimate morpheme counts in synthetic languages compared to agglutinative ones. Researchers need to decide whether to include zero, fused and suppletive morphemes in the morpheme count, and how to handle multimorphemic words such as portmanteaus and compounds (Ezeizabarrena & Garcia Fernandez, Reference Ezeizabarrena and Garcia Fernandez2018). Researchers may find counting morphemes for MLU-m difficult – particularly for cross-linguistic studies – as decisions need to be made morpheme by morpheme and are language dependent (see, e.g., Allen & Dench, Reference Allen and Dench2015, for Inuktitut; Dromi & Berman, Reference Dromi and Berman1982, for Hebrew; Eisenberg et al., Reference Eisenberg, Fersko and Lundgren2001, general, with reference to American English; Hickey, Reference Hickey1991, for Irish; Tomas & Dorofeeva, Reference Tomas and Dorofeeva2019, for Russian; Voniati, Reference Voniati2016, for Cypriot Greek).

Whereas studies discuss the operationalization of morphemes and how to ensure consistent computation of an MLU-m index, we have found that most research does not question the definition of a ‘word’ or how to operationalize a word count, especially cross-linguistically. In fact, almost all studies conflate words with orthographic conventions. The only studies that question this issue are Allen and Dench’s (Reference Allen and Dench2015) study of Inuktitut, and the Bengali study by Gouda et al. (Reference Gouda, Kumar, Sarkar, Rashmi, Chatterjee and Pani2020). Allen and Dench (Reference Allen and Dench2015) suggest that word counts are less arbitrary than morpheme counts, and that operationalizing a word count is not a challenge for most languages. However, they point out that in morphologically rich polysynthetic languages like Inuktitut – the focus of their study – much of the grammar is realized within the morphology word boundaries, which they then identify orthographically (Allen & Dench, Reference Allen and Dench2015, p. 384). Their solution is to introduce morphemes per word counts as a word can contain ten or more morphemes.

E. K. Brown and Miller (Reference Brown and Miller2013, p. 473) conceive of a ‘word’ as an uninterruptible linguistic unit consisting of a stem with or without affixes, where no constituents may be inserted between a stem and its affixes. This definition aligns closely with orthographic word boundaries in most languages where MLU-w has been employed as a measure. Typically, orthographic boundaries (i.e., spaces) between morphemes have been taken to indicate the boundaries of a ‘word’, and each item orthographically separated by a space to be equal to one ‘word’. Given that orthographies are products of convention and consensus, orthography should be treated with caution if used as a basis for linguistic analysis. However, orthographic boundaries may serve as good common-sense dividing lines between ‘words’ (especially for clinical purposes) in languages whose orthographies closely align word boundaries with morpheme boundaries (e.g., English, Dutch). For example, Allen and Dench (Reference Allen and Dench2015) view counting words as more practical if, and only if, a language’s writing system “leaves spaces between words,” (p. 384) again equating the concept of ‘word’ with orthographic boundaries. For such languages, researchers and clinicians may perceive MLU-w to be more convenient, faster, easier, more reliable, simpler to implement, more adaptable and less arbitrary in nature than MLU-m (Hickey, Reference Hickey1991; Parker & Brorson, Reference Parker and Brorson2005). However, when orthographic word boundaries do not align with linguistic boundaries as an uninterruptible stem + affixes unit, the accuracy of MLU-w as a measure is in doubt.Footnote ³ In agglutinative Bantu languages, we note that orthographic standards may determine the segmentation of words rather than linguistic criteria. Using orthographic word boundaries for counting words presents challenges in agglutinative languages such as Basque (Ezeizabarrena, personal communication December 22, 2022) and polysynthetic languages such as Inuktitut (Allen & Dench, Reference Allen and Dench2015).

Southern Bantu Languages

The languages in this study belong to the Bantu language family (Atlantic-Congo) and represent three Southern Bantu language groups: Nguni group (isiXhosa), Sotho-Tswana group (Sesotho and Setswana), and Tswa-Ronga group (Xitsonga). These four languages are agglutinative. They have high inflectional paradigms, strong agreement, high derivation and morpheme-to-word ratio, and a very low number of free roots (Nurse & Philippson, Reference Nurse and Philippson2006). Both nouns and verbs have abstract roots. Bantu languages categorise nouns using a noun class system – comparable to grammatical genders in some other languages – where each noun class uses a particular morphemic prefix, or ‘noun class prefix’ (Nurse & Philippson, Reference Nurse and Philippson2006). The noun classes are numbered using an internationally recognized system, and some Bantu languages have up to 19 noun classes (Nurse & Philippson, Reference Nurse and Philippson2006). Adjectives, verbs, and function words show agreement with the class of the noun they are referring to, overtly expressed with a morphemic prefix (Nurse & Philippson, Reference Nurse and Philippson2006). Tense, aspect, mood, causativity and negation are marked on the verb with affixes (Nurse & Philippson, Reference Nurse and Philippson2006). The default word order is Subject-Verb-Object (SVO), but free word order and extraposition for discourse purposes are common (Demuth, Reference Demuth and Slobin1992; Nurse & Philippson, Reference Nurse and Philippson2006). The syllable structure is typically Consonant-Vowel (CV). These aspects of Bantu languages contrast with analytic languages like English, that has a weak inflectional paradigm and agreement, high derivation, and a high number of free roots with a low morpheme-to-word ratio (Nurse & Philippson, Reference Nurse and Philippson2006).

The orthographies of South African Bantu languages have not been standardized to the same degree as those of European languages (Taljard & Bosch, Reference Taljard and Bosch2006). The Nguni languages use a conjunctive orthography, whereas Sotho-Tswana languages employ a disjunctive orthography. Xitsonga orthography is less clear-cut but is more disjunctive than conjunctive (Lee & Hlungwani, Reference Lee and Hlungwani2017). The different orthographic conventions, decided by missionaries, can be seen more as historical accidents than as indicative of underlying differences in phrase construction in these languages. The disjunctive orthographies of Sesotho and Setswana regularly insert word boundaries between affixal and stem morphemes that do not constitute uninterruptible words (see examples 1a and 1b). In languages with a conjunctive orthography, one orthographic ‘word’ may correspond to two or more linguistic elements (see examples 1c and 2c), whereas in languages with disjunctive orthography, several orthographic ‘words’ may correspond to one linguistic ‘word’ (see examples 3a and 3b).

The following examples demonstrate these orthographic differences across the four languages and the relative consistency of phrase construction between them.Footnote ⁴

Whereas morpheme boundaries are relatively clear, examples 1 to 3 illustrate the challenges that the different orthographic conventions for these morphologically rich, agglutinative languages pose, and that operationalizing word boundaries is not as simple as in analytic or even synthetic languages with relatively disjunctive orthographies. However, following most other studies, we operationalize ‘word’ using the definition of an orthographic word: all linguistic elements that are written separately in the practical orthography and are separated by spaces (Louwrens & Poulos, Reference Louwrens and Poulos2006). Since some of the languages in this study are written conjunctively and some disjunctively, we will see whether the outcome of using morphemes or words in languages with disjunctive and conjunctive orthographies is substantially different. These comparisons will have clinical implications in a multilingual situation when comparing children speaking typologically different languages with different orthographic conventions as is the case in South Africa where different Bantu and Germanic languages are spoken.

The MB-CDI instrument

The MB-CDIs for toddlers (1;4-2;6) were adapted by mother tongue speaking linguists and speech-language therapists of the four languages following the MB-CDI Board’s guidelines (http://mb-cdi.stanford.edu).Footnote ⁵ These adaptations were informed by the wordlists in the UK and USA versions of the MB-CDI, and by the grammatical items in the Kenyan Kilifi versions of the MB-CDIFootnote ⁶ that focus on two Bantu languages as well as by spontaneous speech samples from six children per language collected in their homes. These versions underwent further adaptation in focus groups followed by two pilot studies with participants from communities in which these languages are spoken. At each stage, lexical and grammatical items were evaluated and retained or removed depending on their suitability. A total of 1200 words were tested per language in the first pilot and between 733 and 748 in the second. Forty grammatical items were tested in the first pilot and reduced to 37 items in the second pilot. Where possible, we retained lexicosemantic and grammatical equivalence across the four languages. This study is based on data from the second pilot.

The toddler MB-CDI measures lexical production and early grammatical development. It presents caregivers with a word list grouped into 21 semantic domains and asks them to identify the words their child produces. Then caregivers are asked about their children’s early grammatical development. The first two grammar sections focus on word complexity relating to noun class prefixes and verbal affixes. Caregivers are then asked whether their child is combining words and, if they are, to provide their child’s three longest utterances of the past week (see Fenson et al., Reference Fenson, Dale, Reznick, Bates, Thal, Pethick, Tomasello, Mervis and Stiles1994; Frank et al., Reference Frank, Braginsky, Yurovsky and Marchman2021). Sentence length and complexity are then measured by presenting the caregivers with a series of items (e.g., inflected words, phrases, and sentence constructions) with three to four options reflecting different levels of morphosyntactic development. Caregivers are asked to identify which option reflects best what their child currently produces. (Examples of items testing word and sentence complexity and how they are scored can be found in Statistical Analysis below).

Procedure

Whereas CDIs can be answered by caregivers independently in many contexts, in the South African environment, low literacy levels and lack of familiarity with answering questionnaires – especially online – necessitated that we recruit and train fieldworkers from local communities to administer the CDIs either in person or (during COVID) telephonically or online. Fieldworkers recruited participants through local social networks and childcare organizations in their communities. Fieldworkers used the Qualtrics (2019–2020) online survey tool to capture participants’ responses.

In addition to completing the CDIs, participants were asked to complete a family background questionnaire that collected information on the mother’s pregnancy, the parents’ education and employment status, the socio-economic status of the child’s family, family composition and housing, the child’s medical history (e.g., ear infections), known or suspected areas of communication difficulties (speech-language difficulty, hearing impairments, developmental disability), and the child’s level of exposure to other languages.

Participants

Data were collected from caregivers of 472 toddlers (1;4–2;8). The isiXhosa, Sesotho and Setswana toddlers were recruited from urban and rural areas where these languages predominate (See Table 1). Previous studies on language development in African contexts point to differences between rural and urban children (Alcock et al., Reference Alcock, Rimba, Holding, Kitsao-Wekulo, Abubakar and Newton2015; Vogt et al., Reference Vogt, Mastin and Aussems2015). However, monolingual Xitsonga toddlers could only be recruited from rural and semi-urban areas. (Xitsonga is a minority language, and Xitsonga-speaking toddlers in urban areas acquire other languages from an early age.) We sampled isiXhosa urban speakers in the city of Cape Town in the Western Cape Province and rural speakers in Ilinge in the Eastern Cape Province. Urban Sesotho speakers were drawn from the city of Bloemfontein in Free State Province and rural participants from a rural area 300 km north-east of Bloemfontein in the same province. Urban Setswana speakers were sampled in the town of Rustenburg and rural speakers in the Taung area (394 km from Rustenburg), both in North-West Province. We recruited Xitsonga speakers from semi-rural Giyani and rural Malamulele in Limpopo Province; young Xitsonga-speaking children have little exposure to another language in these areas.

Table 1. Child Demographic Information by Language

The combined monthly income of households of participants ranged from 0 to 80,000 ZAR (0–4,661 USD). The mean monthly income per household was in the range 2,401 to 5,000 ZAR (140–291 USD), whereas the mean monthly food expenditure per household was in the range 1,201 to 2,000 ZAR (70–117 USD). Regarding parental education, for 29% of children both parents had completed high school or studied beyond, for 21% neither parent had completed high school or studied beyond, and for 38% one parent had completed high school or studied beyond (11% were fathers and 27% mothers). In 12% of the cases, there was incomplete information on parental education. Regarding employment status, for 12% of children, both parents were employed; for 20%, neither parent was employed; and for 41%, one parent was employed (26% were fathers and 15% mothers). Twenty seven percent had incomplete information on employment status. The mean number of adults per household in urban areas was lower (M = 2.8, SD = 1.5) than in rural areas (M = 3.5, SD = 1.8). The mean number of children younger than 18 in each household in urban areas was also lower (M = 1.5, SD = 2.3) than in rural areas (M = 4.3, SD = 5.1).

All children with suspected language difficulties, as reported by participants, were excluded. We found differences in vocabulary size for children exposed to another language for more than four hours per day (see Southwood et al., Reference Southwood, White, Brookes, Pascoe, Ndhambi, Yalala, Mahura, Mössmer, Oosthuizen, Brink and Alcock2021); thirteen children were excluded for this reason (isiXhosa = 2, Sesotho = 2, Setswana = 8, Xitsonga = 1). Out of the remaining 459 toddlers, 11 were excluded due to (a) missing vocabulary scores (i.e., only the background questionnaire was completed, not the CDI too), or (b) being outside the target age range. The 448 children included in the study (males: n = 222; females: n = 224; not specified n = 2) were aged 1;4–2;8. We included the children whose gender was not specified as gender comparison was not the focus of the study, and as we do not yet have norms that would allow us to consider gender differences. Table 1 gives the number of children per language, the sample distribution per age in months, gender, and whether rural or urban.

Informed consent was sought from the parents before interviews, after the details of the study had been explained to them. They were informed of their right to withdraw from the study at any point without any negative repercussions. Data were collected anonymously and stored securely. Participating caregivers received a supermarket voucher of approximately 10 USD to thank them for participating. The study was reviewed and approved by the Ethics Committee of the Linguistics Section, and the Faculty of Health Sciences Human Research Ethics Committee, both at the University of Cape Town.

Calculating MLU3-w and MLU3-m

The data for Sesotho, Setswana and Xitsonga were analysed by two linguists and a first language speaker with formal linguistic training at university level. For isiXhosa, a linguist who is a second language speaker of isiXhosa worked with a first language speaker with no formal linguistic training to analyse the data. One of the linguists participated in the analysis of all four languages to ensure consistency across languages.

Uniform inclusion and exclusion criteria were applied across all four languages. As there are morphological differences between the languages – such as the use of pre-prefixes in isiXhosa but not in the other three languages – another set of morpheme-counting rules was also applied, some of which were language-specific.

Statistical analysis

To establish whether MLU3 is a valid indicator of language development, and whether MLU3-w or MLU3-m is a better measure, we compared both MLU3-w and MLU3-m scores, based on the three longest sentences caregivers reported for each child, with their vocabulary (i.e., lexical production) and grammar scores, based on the wordlist and grammatical items, respectively, in their completed MB-CDIs. To obtain their MLU3-w and MLU3-m scores, we counted the words and morphemes in each child’s reported sentences and calculated simple averages. Their vocabulary scores were obtained by counting the number of words from the CDI wordlist that each caregiver said their child produced. For the four languages, the maximum possible vocabulary score ranged from 733 to 748 lexical items (isiXhosa = 748, Sesotho = 741, Setswana = 734, Xitsonga = 733).

For grammar, caregivers were asked whether their child had started using grammatical features such as plurals and noun class prefixes. These items were scored one for yes and zero for no. Grammatical complexity at the word level was then tested in detail with each option minimally more complex or more grammatically correct than the last. For example, noun prefixes were tested by presenting a word stem with no prefix, then the word with a shadow prefixFootnote ⁷ followed by the word with its full and correct prefix. For example, in Sesotho, the caregiver could choose between shemane, a-shemane, or the fully correct form bashemane ‘boys.’ These were scored one, two and three, respectively. For sentence complexity, each item had between two and four options for the participant to choose from, set in order of increasing length and complexity. For example, we asked the caregiver to choose which option best reflected what the child would say for phrases such as ‘the ball is on top of the table.’ In Xitsonga, the choice was for instance between (a) bolo tafula ‘ball table’, (b) bolo henhla tafula ‘ball on table’, and (c) bolo yile henhla ka tafula ‘the ball is on top of the table’. These options were scored 1, 2, and 3, respectively, and a fourth option would be scored 4 if the item had one. The maximum possible scores for the entire grammar section ranged from 93 to 103 across the languages (Setswana = 103, Sesotho = 102, isiXhosa = 98, Xitsonga = 93).

Underlying distributions of responses were described through empirical distribution plots, and consideration of skewness and kurtosis parameters. Internal consistency of scales was assessed using Cronbach’s alpha. Association between measures was illustrated using pairwise scatterplots and measured using Spearman correlation coefficients, as opposed to Pearson correlation coefficients, to guard against bias due to possible deviations from normality. Regression models were used to further estimate and compare (using adjusted multiple correlation squared [R²] and Akaike Information Criterion [AIC] statistics) associations between grammar and vocabulary scales with age, MLU3-m, and MLU3-w measurements. Regression models were validated through an examination of the model residuals.