The provision of effective services to patients from diverse linguistic backgrounds is a big challenge within the South African healthcare system (Mosdell, Balchin & Ameen, 2010; Pascoe & Norman, 2011; Van Dulm & Southwood, 2014). Almost 20 years after becoming an officially multilingual country, the provision of speech-language therapist (SLT) services to a linguistically diverse population remains problematic. The majority of SLT professionals speak English or Afrikaans (or both) and have little or no knowledge of the official African languages (Bornman, Sevcik, Romski & Pae, 2010; Gerber, 2009; Van Dulm & Southwood, 2014). This is a particularly grave concern, given the fact that in SLT services ‘language’ is not only used as a means of communication, but is the reason why a person is seeking help.

Formal assessment of a patient’s linguistic profile and linguistic problems provide a baseline for intervention, but is compromised if a patient is assessed in a language that is not his or her first language. This is especially true for children. In South Africa, the majority of child language assessment measures used are in English and have been standardised in the United Kingdom or the United States of America. As a result, (ad hoc) translators or translations are often used by South African SLTs in clinical encounters. In a national survey conducted by Van Dulm and Southwood (2014), which investigated child language assessment and intervention practices in South Africa, less than half of SLTs who participated felt that commercially available or translated language assessment instruments are linguistically appropriate for the South African context. Only about one-third of the SLTs thought that the instruments they use are culturally appropriate. Thus, SLTs are aware that their lack of knowledge of the native languages of their patients, combined with the lack of standardised language tests in African languages, in fact jeopardise language assessment and intervention results. Barrett, Khoza-Shangase and Msimang (2012) note that ‘this awareness should range from case history taking to the actual assessment to the test material used in the assessment, and therefore more emphasis should to be placed on development of assessment and intervention materials in all languages’.

In response to the need for assessment materials in South African languages, the current study explored the usefulness of developing a non-word repetition task (NRT) in Northern Sotho (one of the official languages of South Africa). Internationally, the NRT is used widely to access both normal and impaired language acquisition. The NRT is particularly associated with diagnosing children with speech-language disorders such as specific language impairment (SLI) and dyslexia, but has become a widely used tool in the field of speech-language pathology (also in the population at large); the reason being that non-word repetition closely matches the phonological component of word learning and is associated with phonological working memory (PWM), a construct which is crucially important in the attainment of language and literacy skills (Dollaghan & Campbell, 1998; Gathercole, Willis, Baddeley & Emslie, 1994). In this study, data gathered with a Northern Sotho NRT (NS NRT) will be presented. Despite the routine use of the NRT as a clinical assessment instrument in other parts of the world, such tasks have not yet been standardised for any of the African languages spoken in South Africa. The aim of this study was to determine the effectiveness of a NRT in determining different levels of phonological processing in Northern Sotho-speaking children, and to investigate whether performance on a NS NRT reliably predicts reading performance in Northern Sotho children. Given the high occurrence of reading failure in South African schools, developing tests which measure linguistic skills associated with reading is critically important.

The broad construct phonological processing consists of ‘three separate but interrelated phonological abilities, namely phonological awareness, phonetic coding in working memory, and phonological recoding in lexical access’ (Claessen, Leitão, Kane & Williams, 2013, p. 472). In its broadest sense, phonological processing has been defined as the ability to encode, store, retrieve, and reproduce sounds (De Bree, 2007). More specifically, phonological processing involves the encoding of auditory (linguistic) signals, the temporary storage of these signals, the retrieval of stored signals, and the planning involved in reproducing these signals (Chiat, 2001; Snowling, Nation, Moxham, Gallagher & Frith, 1997).

The focus of this article is on PWM and therefore ‘phonological awareness’ and ‘phonological recoding’ will not be discussed further. PWM is a basic cognitive skill that underlies all successful auditory language processing. Specifically, it underlies the online cognitive processing that allows the segmentation of the incoming speech stream and the subsequent development of accurate phonological representations of sounds, and (later on) words. In infants and young children, the PWM has to store strings of sounds representing words that are unfamiliar sound sequences when the child hears them for the first time. The quality of a child’s PWM is thought to affect the accuracy and efficiency with which the phonological representations of words are converted from short-term to long-term memory. Thus, PWM is very closely related to vocabulary acquisition, whereas other measures of working memory may be more related to language comprehension and syntactic processing (Gathercole, 2006).

The best-known model of working memory has been proposed by Baddeley (1999, 2003). In this model, PWM is depicted as a subcomponent, i.e. the ‘phonological loop’, of working memory. The other subcomponents of working memory include the visuospatial sketchpad, the central executive, and the episodic buffer. The phonological loop consists of a phonological store (that retains all verbal information that has to be processed or produced for a short period) and an articulatory loop (that rehearses and invigorates the verbal information in the phonological store to ensure that it does not fade away before being processed accurately) (Baddeley, 1999). The phonological loop is active in the acquisition of new vocabulary, in general problem-solving, numerical problem solving, and in remembering instructions. The central executive incorporates information from the phonological loop, the visuospatial sketchpad, and the episodic buffer; and guides the conversion of information from the phonological loop to long-term memory. The central executive is further implicated in multitasking, selective attention, suppression of irrelevant information, and temporary activation of long-term memory. PWM is generally assessed with tasks that tap into the phonological loop and the central executive, such as the digit span forward and backward task and the NRT.

PWM is believed to also affect the acquisition and processing of written language (Catts & Kamhi, 1999; Geva & Siegel, 2000; Nithart et al., 2011), because the ability to decode a word (or syllable) into phonemes (which is necessary for learning to read) requires the PWM system to store the entire phonological representation of a word while the constituent phonemes of that word are first disjointed and then sequenced back into the correct order by the learner. Poor PWM may lead to unstable phoneme representations in a person’s long term memory, which in turn may affect the acquisition of the phoneme-grapheme correspondences of a language negatively.

Numerous studies suggest that deficits in PWM play a causal role in the phonological processing limitations that manifest in diverse clinical populations. PWM deficits have been documented in developmental genetic language disorders, such as SLI (Claessen, et al., 2013; Conti-Ramsden & Durkin, 2007; Gathercole, 2006), dyslexia (Laasonen, Lehtinen, Leppamaki & Tani, 2010; Ramus, 2003; Spironelli, Penolazzi, Vio & Angrilli, 2006) and in other genetic syndromes such as Down’s syndrome (Jarrold & Baddeley, 2001; Numminen, Service, Ahonen & Ruoppila, 2001). The cause(s) for a PWM deficit itself is the topic of much neurobiological research. Genetic-molecular studies using non-word repetition ‘have reported linkage to regions of interest on chromosomes 16 and 19 for children with language impairment’ (Shriberg, et al., 2009). Other genetic regions of interest are chromosome 3 and chromosome 6. So far, one gene, namely ROBO1 (situated on chromosome 3), has been linked specifically to the functioning of the phonological loop of the PWM system (Bates, et al., 2011).

Two decades ago Gathercole et al. (1994) observed that children who experience learning difficulties and who fail to progress in a formal schooling environment, often have difficulties in remembering spoken language even for short periods of time. This limitation manifests itself as the apparent inability of a child to attend to and carry out simple instructions. In an attempt to discover the precise role of phonological memory in children’s language development, Gathercole et al. (1994) developed the Children’s test of non-word repetition (CNRep). In their initial study, the CNRep was administered to more than 600 children, and the test developers found that the correlations between non-word repetition and language skills were consistently stronger than those between auditory digit span (which also contains a significant PWM component) and language skills.

In a NRT, a person has to repeat ‘pseudo’ words. Pseudo words are word forms that could have been lexical items in a particular language, in that they follow the morphophonological rules of that language, but which do not exist. Non-word repetition is more closely associated with linguistic abilities such as vocabulary knowledge, understanding of spoken language, reading ability, and later school performance than the digit span task (Gathercole, 2006; Gathercole & Baddeley, 1996; Gathercole et al., 1994). Because a NRT consists of novel items, which no child would have encountered prior to testing, it does not disadvantage children who received impoverished linguistic input (and thus have smaller vocabularies).

Even though NRTs are often said to measure PWM, such tasks measure a different aspect of phonological memory than digit or word span tasks. NRT is believed to entail the phonological analysis and temporary storage of novel phonological representations, with an individual being asked to recall a representation that at best mimics existing words (Nithart et al., 2011). Digit and word span, on the other hand, entail the processing of representations that already have a stored form in long-term memory – this stored information assists an individual in the analysis and temporary storage of phonological information. Thus, non-word repetition is thought to represent a clearer view of an individual’s phonological processing capacity in real time, and is more likely to detect inefficiencies in the phonological loop of PWM than digit or word span tasks.

NRTs are typically designed to contain both simple and complex words. The complexity of non-words is determined by two factors:

Clinical populations, such as dyslexics, children with SLI, people living with Down’s syndrome, and cochlear implant patients have all been found to be impaired on NRTs. These groups score significantly lower on non-word repetition than their age/language-matched normally developing peers. On the basis of this, and seeing that reading-disabled people also experience phonological awareness and phonological recoding impairments, the phonological deficit hypothesis (Ramus, 2001; Snowling, 2001; Vellutino, Fletcher, Snowling, & Scanlon, 2004) became one of the most influential theories in explaining reading disability in clinical populations. Generally speaking, non-word repetition deficits appear to remain present in reading-disabled individuals, even into adolescence and adulthood (De Bree, 2007). This is less true for digit span and word span tasks, which confirms that, although these tasks are related, they tap into different aspects of phonological processing.

There is no agreement as to whether the NRT can be used to predict reading ability in the population at large. Bishop, Adams and Norbury (2004, p. 96) are adamant that non-word repetition does not predict reading outcome in the general population at large. According to these scholars:

Bishop et al. (2004) are of the opinion that reading difficulties in non-clinical groups are caused by environmental factors such as poverty, inadequate education, and social factors such as the age and education of a child’s mother and the child’s position in the family. Although all of these environmental factors probably contribute to the low literacy levels of South African learners, the causal role of underdeveloped phonological processing skills and vocabulary skills (see Wilsenach, 2015), which are strongly associated with non-word repetition, remains unclear in the South African context.

Studies on the use of the NRT in African languages are very limited. Veii and Everatt (2005) investigated predictors of word reading among Grade 2–5 Herero-English bilingual children. They included Herero and English phoneme recognition, PWM, a Herero and English NRT, Herero and English rapid naming, and Herero and English word reading as measures. Phonological skills in Herero and English (such as phoneme awareness and non-word repetition) reliably predicted reading ability in both languages. Alcock, Ngorosho, Deus and Jukes (2010) studied the relationship between phonological awareness and literacy in Swahili-speaking children. They investigated implicit and explicit phonological awareness skills and literacy in Tanzanian school-going and non-school-going children between the ages of 7 and 10. Alcock et al. (2010) found that children who could not read had poorer phonological awareness skills than those who could, except in counting syllables, counting sounds, and in NRT (however, there might have been a ceiling effect in the NRT).

Wilsenach (2013) assessed phonological processing (including syllable awareness, non-word repetition, and digit span) in two groups of Northern Sotho-speaking third graders. One group received their schooling in their home language (Northern Sotho) from Grade 1–3, whereas the other group received their schooling in English from the beginning of Grade 1. Contrary to Bishop et al.’s (2004) prediction, Northern Sotho non-word repetition reliably discriminated between the two groups of normally developing Northern Sotho-speaking children. However, Wilsenach (2013) acknowledged that the study was limited – primarily because of the small sample (50 overall), and recommended that a larger sample needs to be tested.

The present study set out to measure phonological processing, specifically non-word repetition, memory for digits and reading in 120 Northern-Sotho third graders. In order to evaluate the effect of the Language of Learning and Teaching (LoLT) on the learners’ ability to process sounds in their mother tongue, half of the learners received their schooling in Northern Sotho, and the other half received their schooling in English.

The current study was conducted in two primary schools located in a high-poverty suburb of Pretoria (one of the capital cities of South Africa). Northern Sotho is the most common home language in this suburb, but several other African languages occur in the research area. Some schools within the area use Northern Sotho as LoLT from Grade R until the end of Grade 3, after which the curriculum is delivered via English. In the present study, Northern Sotho was used as LoLT in one of the schools; the other school followed a straight for English language policy.

One hundred and twenty Grade-3 learners¹ were selected to participate in the study, using a purposive and convenience sampling technique. In order to participate in the study, learners had to speak Northern Sotho as their home language. The participants were made up of two groups: Group 1 (N = 60; mean age 8.8) was instructed in English (with Northern Sotho being taught as an additional language), whereas Group 2 (N = 60; mean age 8.8) was instructed in Northern Sotho (with English being taught as an additional language). There was no significant difference between the groups in terms of age (t = -0.04; p = 0.96). Generally speaking, learners from both groups came from poor households, as evidenced by the fact that both schools have feeding schemes for learners (the children in Group 1 attended a quantile 2 school, whereas the children in Group 2 attended a quantile 1 school).² Group 1 contained 35 girls; Group 2 contained 29 girls.

The research was conducted as a cross-sectional quantitative study. The participants were tested individually during a single session. Various phonological skills (not all reported on here) and reading were tested. The research was conducted during a 5-week period in the second term of the school year. All the tasks were presented to the learners using a computer and headphones – this ensured that the auditory presentation happened accurately, and with consistent rate and consistent intonation. Performance was scored online, but a digital recording of each test session was made in order to be able to check a learner’s responses.

PWM were measured with a NS NRT, an English NRT, and an English digit span task.³ The NS NRT was similar to the task reported on in Wilsenach (2013). It was designed following most of the criteria set out by Dollaghan and Campbell (1998), including:

The NS NRT consisted of 20 test items, ranging from four syllables (e.g. sepokari) to seven syllables (e.g. nasibhekarabile) in length. The test comprised five items at each syllable length. The non-words were pre-recorded by a native speaker of Northern Sotho and were presented in the same order to each of the participants. The 20 test items are included in Appendix A.

English NRT was measured with subtest 5 of the Comprehensive Test of Phonological Processing (CTOPP) (Wagner, Torgesen & Rashotte, 1999). This NRT proceeded from 1-syllable non-words to 10-syllable non-words. PWM was further assessed with subtest four (memory for digits) from the CTOPP. The participants had to recall strings of digits in English, as presented on the CTOPP CD. The memory for digits test started with strings of two digits, and progressed to strings with nine digits. The English NRT and the memory for digits test were discontinued after three consecutive incorrect responses. The rationale for including the English NRT and the memory for digits task here is that the scores obtained on the NS NRT should correlate with these measures, if the NS NRT indeed taps into phonological processing/PWM abilities.

Word reading in Northern Sotho was tested with a tailor made word reading task, consisting of 30 items (in order of increasing length and difficulty; see Appendix B). Northern Sotho text reading was assessed by measuring the amount of words that learners could read accurately in one minute, using a grade appropriate reader. Word reading in English was assessed with the Diagnostic test of word reading processes (FRLL, Institute of Education, 2012), which consists of three subsections (non-word reading, exception word reading, and regular word reading). Only regular word reading will be reported on here, as this subsection of the test best compared with the Northern Sotho word reading test. English text reading was assessed by measuring the amount of words that learners could read accurately in one minute, using a grade appropriate reader.

Raw scores were calculated for every participant on each of the measures. The raw scores on the NS NRT and on the Northern Sotho word reading test were transformed into percentages, and based on these scores a mean raw score and mean percentage were calculated for the sample as a whole, as well as for the two groups. The raw scores on the English NRT and on the memory for digits test were converted to standard scores (SSs), using the age norms in Wagner et al. (1999). The raw scores on the English word reading test were used in the statistical analyses, as the GL assessment manual only provides SS for the composite word reading raw score (consisting of all three subsections of the test). Raw scores for Northern Sotho text reading and English text reading were calculated by counting the number of words that a child read aloud in 1 min, and subtracting the number of incorrectly read words from this total.

The distribution of the test scores on the NS NRT was tested with the Shapiro–Wilk test, and was further analysed via a visual inspection of the data. Seeing that the NS NRT is not a standardised test, it was important to confirm that the test scores were distributed normally. The effect of LoLT on Northern Sotho phonological processing ability, as well as between-group differences (in terms of NS NRT, English NRT, and PWM, and reading) were tested using a multivariate analysis of variance (MANOVA) (GLM option multivariate). Group was entered as an independent variable in the model, whereas NS NRT, English NRT, Digit span, NS Word reading, NS Text reading, English word reading, and English text reading were entered as dependent variables. Pearson correlations were conducted in order to determine the relationship between NS non-word repetition and related phonological constructs, such as memory for digits and English non-word repetition. Hierarchical multiple regressions (method enter) were conducted to determine the contribution of NS non-word repetition to NS reading.

The descriptive statistics for the NS NRT are given in Table 1.

A Shapiro–Wilk test, performed on the collapsed data, yielded a non-significant result, indicating that the NS NRT test scores were distributed normally (W(120) = 0.98, p = 0.058). Both skewness (-0.31; SE = 0.22) and kurtosis (-0.36; SE = 0.49) were negative, suggesting a build-up of high scores, but the associated z-scores for both skewness and kurtosis (calculated by dividing these values by their standard errors) were smaller than 1.96, indicating no significant problems with either skew or kurtosis. Further support for the claim that the NS NRT test scores were distributed normally is derived from the normal quantile-quantile (Q-Q) plot (Figure 1), in which the plotted quantiles fall very closely to the diagonal line (which represents a perfect normal distribution).

The histogram in Figure 2 confirms a (relatively) flat and light-tailed distribution of the NS NRT scores, but also indicates a slight build-up of higher scores; with a peak at raw score 13 (i.e. 65% correct responses to the test). The visual inspection of the data seems to suggest that a peak at 11 or 12 would fit the normal distribution curve better.

Nevertheless, for the purposes of this study, the non-significant result of the Shapiro–Wilk test was seen as sufficient evidence that the scores obtained in the NS NRT were normally distributed, and hence the rest of the statistical analyses were conducted using parametric tests.

The GLM model (option Multivariate) established a significant main effect for Group (F[3,116] = 9.67, p = 0.000). Following this significant effect, tests of between-subjects effect were conducted, which showed that Group had a significant effect on English NRT (F = 14.13, p = 0.000), but no effect on Digit span or NS NRT. Pairwise comparisons (Tukey–Bonferroni post hoc tests) confirmed that there were no significant differences between the groups in terms of Digit span or NS NRT, but that Group 1 significantly outperformed Group 2 in the English NRT. Furthermore, Group 1 performed significantly better on both English reading measures, whereas Group 2 performed significantly better on both the Northern Sotho reading measures. The mean scores and test statistics for the measures Digit span, NS NRT, English NRT, Northern Sotho word reading, Northern Sotho text reading, English word reading, and English text reading are presented in Table 2.

Pearson correlations (two-tailed) indicated that NS NRT was significantly correlated with Digit span, English NRT, NS word reading, NS text reading and English word reading. English NRT was significantly correlated with both English reading measures, but showed no correlation with the NS reading measures. Digit span was significantly associated with all the reading measures in both languages. The r-values of these correlations are given in Table 3 (significant correlations are flagged with an asterisk).

Multiple hierarchical regression models with NS word reading and NS text reading as dependent variables and NS non-word repetition and digit span as predictors, showed that performance on the NS NRT reliably predicted the outcome of NS reading ability (non-word repetition was entered at Step 1 of the model and digit span was entered at Step 2). The ability of the NS NRT to predict English reading skills was not of primary interest here, and the correlations mentioned in Table 3 suggest that there is no, or a very weak relationship between the NS NRT and English reading. Thus, no regression with English reading as dependent variable is reported here. The constant values, betas, standard errors, standardised betas, and R² values for each of these regression analyses are provided in Table 4.