Identification of fluency and word-finding difficulty in samples of children with diverse language backgrounds

Symptoms indicative of fluency difficulty and the role of WWR

According to Brocklehurst (2013) and Wingate (2001), WWR are not symptoms that indicate fluency difficulty. Support for this observation is that the most widely used instrument for assessing fluency, Riley's (2009) Stuttering Severity Instrument (SSI), does not include them but uses symptoms that involve word fragmentation (e.g., prolongation, part-word repetition or word break) instead. Other support for the view that WWR are not indications of fluency difficulty is reviewed in Howell (2010). Subsequently Jiang et al. (2012) examined whether or not WWR should be grouped with the fragmentary symptoms of fluency difficulty. People who stuttered heard an incomplete sentence, which they completed whilst lying in an MRI scanner. Each trial was classified according to any symptoms of non-fluency that occurred, and the associated brain activation patterns for different types of non-fluencies were examined. The trials where there had been fragmentary dysfluencies activated different areas of the brain from those where there were other common non-fluencies not associated with fluency difficulty (multi-syllable word repetitions, phrase repetitions and pauses). The brain activation patterns from trials where there had been a WWR were then automatically assigned to either the fragmentary or the common non-fluency class. The patterns that occurred in WWR were placed with the common non-fluencies, supporting the view that they are not symptoms of fluency difficulty.

The symptom measure used in SSI (% syllables with fragmentary dysfluencies/all syllables, %SS) distinguishes those 4–5-year-old children who do and do not have fluency difficulty (Howell 2013, Mirawdeli and Howell 2016). Identification of fluency difficulty using fragmentary symptoms worked even when samples contained approximately 40% CwEAL (Mirawdeli and Howell 2016). Howell's (2013) procedure does not classify children with high rates of WWR as having fluency difficulty. If WWR had been included as symptoms of fluency difficulty, symptom rate would have been approximately 25% higher (Yairi and Ambrose 2005), more children's symptom counts would have exceeded the threshold, and some children could have been misclassified as having fluency difficulty. The prediction that inclusion of WWR in symptom counts would result in false alarms about fluency difficulty was confirmed by Howell (2013) who showed that models to identify children who stuttered that included WWR along with fragmentary symptoms had lower sensitivity and specificity than models that excluded them.

Howell's (2013) procedure allows the WWR symptoms to be used to indicate WFD. Consistent with this, WWRs are used as fillers when words are not known (Bada 2010, Clark and Clark 1977, MacWhinney and Osser 1977) and high rates of WWR symptoms alone are indicative of WFD (Bergmann et al. 2015, Fathman 1980, Fox et al. 1996, German 1991, Hilton 2008). Moreover, WWR rate indicates WFD irrespective of a child's language background (Lennon 1990, Rydland and Aukrust 2005). Hence, the proposal to separate off children with fluency difficulty using %SS and then to use rate of WWR in the remainder as an indication of WFD should be applicable to CwEAL as well as CwEO. Although the 40% CwEAL in Mirawdeli and Howell (2016) had high rates of WWR because of WFD, they were not misclassified as having fluency difficulty in comparison with teachers’ indications.

NWR performance in children with fluency difficulty and WFD

Non-word repetition (NWR) performance could be used to validate the symptom-based approach in samples where there is diversity in language backgrounds. NWR performance is affected when a child has fluency difficulty (Anderson and Wagovich 2010, Bakhtiar et al. 2007, Hakim and Ratner 2004), These studies have reported repetition of two- and three-syllable non-words is poorer in children who stutter relative to fluent control children probably because of impaired phonological processing (Gathercole et al. 1994).

Also, in the UK, whilst NWR performance should be sensitive to fluency difficulties, it should not be affected when there is WFD in CwEAL: They often have to produce phoneme sequences in English words that they are not familiar with and this makes similar demands to NWR (Bialystok et al. 2003). The same should apply to CwEO who have WFD if the problem is due to poor vocabulary (Ellis Weismer et al. 2000). Consistent with this position, no studies reporting NWR deficits in children with WFD were returned in literature searches. In summary, NWR performance is: (1) an indication of fluency difficulty; (2) not impaired if children only have WFD; and (3) at least as good for fluent CwEAL as it is for fluent CwEO.

Design requirements of UNWR

Masoura and Gathercole (1999) reported that Greek children who learned English at school performed better on a Greek than an English NWR test. Also, Windsor et al. (2010) showed that children whose first language was Spanish were more accurate than CwEO on a Spanish NWR test, but the opposite was true when an English NWR test was used. Therefore, NWR tasks designed for one language cannot be applied to other languages and a task that applies to more than one language is required when NWR ability is assessed in heterogeneous language samples. To provide such a test, a common core of syllabic phonotactic constraints was identified that apply to the languages spoken by the children who were tested. Non-words generated according to these constraints are phonologically well-formed for all these languages and are appropriate for testing children who speak any of the targeted languages.

UNWR tests non-words that are between two and five syllables in length. Tests that are used in schools need to be short (to minimize disruptions to children's learning time). To keep UNWR brief, the number of non-words at each syllable length was set at seven. A child had to get all seven items correct to progress to the next syllable length (p < .05, by Sign test). UNWR has to be straightforward to administer so that teachers can conduct it rather than employing outside professionals. Equipment requirements should be minimal, again for reasons of efficiency. Consequently, Conti-Ramsden et al.’s (2001) procedure was adopted, where the experimenter spoke the material. These considerations led to differences between the UNWR and Gathercole et al.’s (1994) Child test of NWR (CNRep) which is widely used to test English-speaking children. CNRep has ten non-words per syllable length, delivers all syllable lengths to a child irrespective of performance on shorter non-words and works from recordings.

UK schools insist that testing is conducted in English as they do not have the staff to deal with all the languages they encounter (NALDIC, 2013) and teaching English is a focal educational goal.2 Together these considerations make UNWR a brief, easily administered test that can be used to assess all children who speak one of the 20 languages.

The UNWR non-words were checked to ensure they were not words in any of the other languages (by native respondents for all 20 languages and computationally using lexicons for the five languages used in this study). Non-words were also checked for word-likeness for the five targeted languages. Non-words that obey the phonotactic constraints of an individual's native language are high in word-likeness and this influences accuracy of repetition (Dollaghan et al. 1995, Gathercole 1995, Munson et al. 2005). Luce's (1986) neighbourhood density measure was used to determine how many phonologically similar words there are to a target non-word.

Current study

Short samples of speech were used to determine participants’ degree of fluency difficulty and degree of WFD (%SS, and %WWR instances out of all syllables, respectively). Multiple regression analyses were conducted that tested whether: (1) %SS and language group (CwEO versus CwEAL) were predictors of UNWR; (2) %WWR and language group were predictors of UNWR; and (3) whether UNWR scores varied across languages. The corresponding predictions based on Howell (2013) were: (1) %SS should predict UNWR; (2) %WWR should not predict UNWR; (3) UNWR should not favour any of the languages and, therefore, it should be immune to differences in English language ability of the CwEAL. Therefore, language group was not expected to be a significant predictor in any regression. The children were also tested on CNRep for comparison with UNWR

Participants

All children from reception classes in five mainstream primary schools were tested (three in the London borough of Merton and two in Ipswich. Ipswich had a population of 133,400 in 2014, of which 82.9% were White British. The average weekly pay for men in Ipswich was £456.3 This was lower than that of England overall (£513). Approximately a quarter of the Ipswich population (26.6%) lived in the most deprived conditions in England.4 The population of Merton was 199,700 in 2015 and 75.0% of the population were white British. The median gross weekly pay in Merton was £535.50, which was fourteenth out of 32 London boroughs.5 Based on these statistics, both regions are in mid to low socio-economic areas.

None of the children had neurological deficits. Language spoken was reported by the schools. Thirteen children were excluded because they spoke a language not in the set of 20 UNWR languages and 18 more were excluded because a lexicon was not yet available for that language for checking word status and word-likeness properties computationally. Five children were excluded because their %SS was more than three standard deviations from the mean. The 17% of children outside that limit were considered outliers where random factors led to variation in %SS (e.g., a child had an emotional issue at home or at school or was starting with an illness). The remaining 96 children spoke English (79.17%), Urdu (7.29%), Polish (5.21%), Portuguese (7.29%) or Romanian (1.04%). There were 49 males (41 of whom were CwEO) and 47 females (35 of whom were CwEO). The mean age of the CwEO group was 4.49 years (SD = 0.50 years) and of the CwEAL was 4.60 years (SD = 0.50 years). This difference in age between the two groups was not statistically significant using Wilcoxon rank-sum test with continuity correction (p = .3730). Ethics approval was granted by UCL's IRB (4374/001).

UNWR stimuli

The UNWR stimuli varied in the number of syllables within a non-word (syllable length) and the number of consonants within a syllable. Syllable-internal complexity was varied according to two main parameters that have the same settings in all 20 target languages (English included): (1) the number of consonants in the syllable onset (one versus two) and (2) whether or not a syllable was closed by a coda consonant. Two lower thresholds of syllabic complexity were fixed: (1) every syllable contained a vowel (some of the languages lack syllabic consonants), and (2) every onset contained at least one consonant (some of the languages lack onset-less syllables). Two upper thresholds were also fixed: (1) onsets contained at most two consonants (some of the languages lack onset clusters larger than this), and (2) codas contained at most one consonant (some of the languages lack complex coda clusters).

Non-word candidates were selected for each syllable length as follows. First, all permitted syllable combinations which could serve as templates (e.g., [CV.CV] for two-syllable combinations) were created for each syllable length, and 100 of these templates per syllable length were selected at random. Second, consonant and vowel phones were selected randomly, apart from the above constraints, and entered into the template for each syllable length. Third, individual syllables were combined. The stimulus set featuring vowel reduction was generated by converting all vowels to schwa. The phonetic transcriptions in table 1 show the difference between two versions of UNWR: a language-independent (‘universal’) form in which stress, vowel length and reduction are absent and an English implementation in which these properties are present.

Computational checks that UNWR non-words are not words in any of the five languages and evaluation of word-likeness of UNWR and CNRep stimuli

Word-likeness

Luce's (1986) phoneme edit metric was used to determine the number of phonologically similar words there are to a target. A target non-word has a word as a neighbour if that word results when a single operation at a phonemic level is made. The operations are insertion, deletion or substitution of a phoneme, and these are made at all positions in the target phone string. The measure indicates the number of words differing by a single phoneme from the target non-word.

The word-likeness estimate of the target non-word was the total number of word neighbours found after all operations were applied (the more neighbours, the higher the word-likeness). Each non-word was checked against all the words in the lexicons of the five languages using Levenshtein distance (Levenshtein 1966).

Using full non-word target transcriptions to compute word-likeness scores weights vowel and consonant phone changes equally. Vowels vary more across languages than do consonants and this can impact on the number of neighbours the same non-word has in different languages. Vowels were reduced to schwa (the reduced set) and word-likeness scores were recalculated This normalized comparisons across languages that have different numbers of vowels, avoided problems that arise because vowels have different qualities across the languages and gave vowels less weight than consonants when determining word neighbours of non-word strings. Word-likeness scores were computed for the two non-word repetition sets (UNWR versus CNRep), two types of vowel information (full versus. reduced), four syllable lengths (length ranged from two to five syllables in UNWR and CNRep) and the five selected languages.

Procedure

Children were tested individually in a quiet room in a 15-min session, which was audiotaped on a Zoom H4n recorder using an X/Y stereo microphone. A speech sample was elicited using probe questions and Riley's (2009) picture material. The mean length of the samples was 333.21 syllables. Finally, the two NWR tests were performed. The order of presentation of UNWR and CNRep was counterbalanced across children. Children were told prior to the non-word tests that they would hear made-up words that they should repeat. For UNWR, the non-words were spoken by a female experimenter using English stress patterns (table 1). Two- and three-syllable items had strong–weak and strong–weak–weak patterns respectively. The stress pattern for four- and five-syllable English words varied (their stress patterns are indicated in table 1). The stress patterns for CNRep stimuli are given in table 3 (Gathercole et al. 1994). Each NWR test began with the two-syllable stimuli, and syllable lengths were increased successively. Three stimuli per syllable-length were used as practice material. The seven test stimuli for a particular syllable length were randomized. Each child was allowed as much time as was necessary to respond after each non-word was presented. The accuracy of the child's production was evaluated as each word was tested. When evaluating a child's repetition of a UNWR stimulus, only consonants were scored; any differences in stress, vowel length or quality were ignored. All stimuli at a given syllable length were tested even when a child made an error, but a child only progressed to the next syllable length when all seven non-word stimuli at the present syllable length were correct. The total score for UNWR and CNRep was the number of non-words that were correct across all syllable lengths that were attempted.

Analysis and reliability estimates

The spontaneous speech samples were analyzed to obtain %SS (Riley 2009). For %WWR, all the repetitions on the monosyllabic word were counted as a single non-fluent event, which is how Riley (2009) treated stuttering symptoms to obtain SS. The total syllable count was adjusted by deducting the number of WWR repeats from the total syllable count in the %SS analysis in which each repeated word was counted. %WWR was the percentage of WWR non-fluent events out of all syllables.

The speech recordings were reanalyzed for %SS and %WWR so that intra- and inter-judge reliability could be calculated. For intra-judge reliability the original judge reassessed 10 randomly selected samples. Agreement about non-fluent events for the 10 samples varied between 87 and 95% for %SS and between 90 and 95% for %WWR. All κ coefficients were above .75, ‘excellent’ according to Fleiss (1971). For inter-judge reliability, a second judge estimated %SS and %WWR on ten randomly selected children's speech samples. Agreement, calculated as above, across the judges for the ten samples varied between 81 and 90% for %SS and between 85 and 92% for %WWR. All κ coefficients were again above .75.

Ten UNWR and ten CNRep tests were reassessed by the same judge and by a second judge. Inter- and intra-judge reliability for UNWR and CNRep consonant accuracy was assessed as percentage agreement across the two overall correct scores. κ coefficients indicated excellent agreement in all cases (Fleiss 1971).

Evaluation of whether UNWR is appropriate for equitable testing across languages and comparison with CNRep

The average number of neighbours is an index of word familiarity. Row one of figure 1 shows the data where there was no vowel reduction: None of the UNWR non-words had any word neighbours for any of the five lexicons; Row two shows that CNRep non-words had neighbours for two-syllable non-words across all lexicons, and for three-syllable non-words for the British English lexicon and there were few neighbours for any lexicon for the four- and five-syllable non-words. Comparison across languages showed that the CNRep stimuli had more neighbours for the English lexicon than for the non-English lexicons. The third and fourth rows in figure 1 show the data when the vowels were reduced. Again, CNRep non-words had more neighbours than UNWR non-words and only the British English, Polish and Romanian two-syllable UNWR non-words had neighbours. Overall, CNRep had more neighbours than UNWR, particularly when non-words had few syllables. Statistical models evaluated this pattern of variation in the number of neighbours to establish whether UNWR is appropriate to assess the children who participated in the study.

Analysis of these data is complicated because 88.1% of the data points from figure 1 had zero counts (figure 2). Consequently, a Hurdle model (Mullahy 1986) was used. There are two components to hurdle models: the hurdle component itself which models zero counts versus non-zero counts. A binomial distribution was used to model this component (Zeileis et al. 2008); the truncated count component which models positive counts. A negative binomial distribution was used to model the truncated component (Zeileis et al. 2008).

A saturated model with fully crossed predictors (main effects and all interaction terms) was fitted first as a baseline against which improvement in subsequent models could be ascertained. This did not converge. Consequently, the model was simplified by basing the order in which factors were excluded on the hierarchy from most complex (the biggest interaction term) to least complex (single factor terms). The marginality principle was adhered to, meaning that models containing an interaction term were not allowed unless the respective main effects and all lower-order interactions with this factor were significant.

This model converged and provided separate regression terms for the count and hurdle components. Starting with the Converged model, a series of nested model comparisons was performed using a χ² likelihood ratio test (lrtest) with α = .05 (Zeileis and Hothorn 2002). Retention or omission of factors was based on whether or not model fit changed significantly: A term was retained or omitted depending on whether fit reduced or improved significantly respectively. Again the principle of marginality was adhered to. A series of nested model comparisons was then performed on superset and subset model pairs. A superset model contained more regression terms than a subset model, and all the terms in a subset model had to be present in its superset model. The best-path algorithm was used for model construction where going backward (exclusion) determined whether there were multiple subset models that resulted in p-values that exceeded the α-level in their nested model comparisons with the superset model. This indicated which subset model had the strongest evidence (the highest p-value). When going forward (inclusion), if there were multiple superset models that resulted in p-values below the α-level in their nested model comparisons with the subset model, the superset model with the strongest evidence (the lowest p-value) was selected. The direction of comparisons was first backward (exclusion) then forward (inclusion), and this pattern was repeated until no further terms were included or excluded. The order in which terms were excluded went from the most complex (the largest interaction terms) to the least complex (single terms). The reverse was the case when terms were considered for inclusion. The process was performed for the count and hurdle components.

This procedure provides the optimal combination of the selected factors in the final prediction model (summarized in table 4). There were main effects of Repetition Set, Vowel Reduction and Syllable Size for the count component, and of Repetition Set, Vowel Reduction, Language and Syllable Size for the hurdle component. There were no interaction terms for either component. All the predictors (Repetition Set, Vowel Reduction, Language and Syllable Size) discussed in the descriptive analysis contributed significantly to the prediction of word-likeness except that the Syllable Size contrast [5 versus {2, 3, 4}] was not significant in the count component. This suggests that the syllable size effect decreased with length (an effect only occurred up to four syllables). Language was present in the hurdle component showing that non-words were more word-like in English than the non-English languages. Language was absent in the count component, showing that non-words that have at least one neighbour were not significantly more word-like in English than non-English languages. UNWR had lower word-likeness (fewer neighbours) than did CNRep overall and for each of the five languages with either no difference (Full vowels), or small differences (Reduced vowels), across the languages. The virtual absence of a language effect on word-likeness with UNWR reduced and full stimuli allows comparison of performance across children who speak one or more of the target languages for which it has been comprehensively evaluated.

Speech measures and language group as predictors of UNWR or CNRep scores

For the children's performance data, mean %SS was 3.39 (SD = 4.25) for CwEO and 1.90 (SD = 1.37) for CwEAL. Mean %WWR was 3.33 (SD = 2.53) for CwEO and 2.65 (SD = 2.88) for CwEAL. CwEO tended to show higher rates of both types of difficulty than did CwEAL. The mean %SS and %WWR were not significantly different between the EO and EAL groups by Wilcoxon tests (p = .1593 and p = .1998 respectively).

Statistical analyses employed one of the two speech measures (%SS or %WWR) and language group (EO versus EAL) to predict either UNWR or CNRep scores giving four models in total. The modelling procedure was equivalent in all cases. First a saturated model with fully crossed predictors (single factors and all interaction terms) was fitted. A series of nested model comparisons was then made using a χ² likelihood ratio test (α = .05). This showed which factors (main effects and any of their interaction terms) could be left out without model fit being affected. Only those factors left after non-significant factors were pruned were significant predictors of the respective NWR scores.

All four continuous variables (UNWR scores, CNRep scores, %SS and %WWR) were logarithmically transformed (base 10, with Laplace smoothing), and then converted into z-scores. The categorical language variable was coded with contrast values of 1 and –1 for EAL and EO respectively.

%SS and language group were assessed as predictors of UNWR scores in the first model. The model only included %SS as a main effect. The model is formulated as ‘UNWR scores ∼ %SS’ where ‘∼’ = predicted by. The fit statistics are given in the first section of table 5 (model 1). Children who had high %SS tended to show lower UNWR scores.

The second model used %WWR and language group to predict UNWR scores. None of the factors were significant as main effects nor interactions. The model is formulated as ‘UNWR scores ∼ Intercept’ and statistics are summarized in the second section of table 5.

The next two models looked at either %SS (model 3) or %WWR (model 4) along with language group as predictors of CNRep scores. None of the factors was significant as main effects nor interactions in either model. Both models are formulated as ‘CNRep scores ∼ Intercept’. The statistics for these models are labelled model 3 and model 4 in table 5. The fit statistics for models 3 and 4 are identical because they are null models predicting the same thing—CNRep scores.

Language group and %WWR were not significant predictors in any of the models. However, %SS predicted UNWR scores (p = .0044) but not CNRep scores. Thus, a relationship between fluency difficulty (indicated by %SS) and NWR performance was only found for the UNWR test.

Given the findings from these four models, another regression was performed with both %SS and %WWR included (‘UNWR scores ∼ %SS + %WWR’) to see if such a model predicted UNWR performance better than %SS. alone A nested model comparison showed that the model fit was significantly improved when %SS was included (χ²(1) = 6.9822, p = .0082) but not when WWR was included (χ²(1) = 0.17376, p = .6664). Therefore, WWR was dropped and the resultant model is identical to model 1 in table 5. The hypothesis that NWR scores can discriminate between fluency difficulty and WFD may apply to CNRep for the EO children that the test was designed for. To assess this, a further model (model 5) was fitted that only used data from the EO group. The model examined whether %SS predicted CNRep scores (‘CNRep scores ∼ %SS’). %SS was not a significant predictor of CNRep scores and it did not significantly improve the model fit (χ²(1) = 1.3523, p = .2449, p > .1), and was therefore dropped. Hence, CNRep scores were not sensitive to fluency difficulty.

Patterns of non-fluencies in children with fluency difficulty and in CwEAL

One explanation of the relationship between SS and UNWR is that they both reflect phonological planning whereas WWR do not (WWR are an articulatory response used to hold the floor whilst generating or reformulating the following speech output). According to MacWhinney and Osser (1977), 75% of WWR in 5-year-olds occur on function words. Their explanation was that function-word repetition allows pre-planning of upcoming utterances as opposed to them being prone to errors during production, since function words tend to be phonologically simple. Howell (2010) adapted this idea to account for different symptom patterns within phonological word (PW) units (Selkirk 1984). PW units have a content word (C) as the obligatory nucleus and function words (F) as optional affixes and suffixes. Howell assumed that the content-word in a PW has properties that can lead to its plan not being ready when prior words have been produced. For example, in the PW He spilt it (FCF) spilt is complex and its plan may not be ready after he has been uttered. The speaker can either; (1) repeat the prior function word he to buy planning time which would lead to a WWR (MacWhinney and Osser 1977); or (2) continue the utterance and rely on the remainder of the plan being completed whilst the part which is available is produced. If the plan is completed, the speech will be fluent. Otherwise fragmentary non-fluencies on the content word would occur. In this account, WWR and the fragmentary non-fluencies are different ways of tackling the same problem, i.e., that the plan for the content word has not been generated in time. Two factors jointly determine whether non-fluencies will arise in connected speech in this explanation: planning difficulty associated with the content word and the articulation rate on the stretch of function words that precede the content word. If the speech prior to the content word is produced rapidly, there is pressure on when the content-word plan needs to be ready. From this perspective, WWRs are regarded as a way of effectively reducing articulation rate that influences planning indirectly whereas SS indicate that speech has been attempted based on an incomplete plan. Many young children use function-word repetition in preference to fragmentary non-fluencies and this is not a barrier to fluency whereas children who produce fragmentary non-fluencies can experience long-term fluency difficulty (Howell 2010). Therefore, function-word repetition may allow the children to avoid fluency failure on content words.

Function-word repetition could also allow children to recover fluent speech control when there is WFD, as happens when, for instance, CwEAL do not know a word. Advancing to the content word is not an option as the word is absent so fragmentary non-fluencies cannot arise. WWR would allow time for these children to employ alternative ways out of such blocks (Howell 2010). For instance, a word from their native language could be used (code switching) or they could circumlocute using an alternative English word. To summarize: (1) SSs arise during planning and are an indication of fluency difficulty (Howell 2010, Mirawdeli and Howell 2016); and (2) WWRs operate at the articulation stage, which adjusts articulation rate to allow more time for planning when there is fluency difficulty (Howell 2010) and to allow a child with WFD time to determine an alternative formulation of the utterance (Bergmann et al. 2015, Fathman 1980, Fox et al. 1996, Hilton 2008).

Why should %SS predict UNWR scores whereas %WWR does not?

It follows that if fluency difficulty results from planning problems (Howell 2010), a relationship between %SS and UNWR would occur if UNWR also assays planning problems. Gathercole et al. (1994) argued that NWR is related to phonological rehearsal (which sets up plans). Evidence in support of this is that NWR scores correlate with auditory digit span in typically developing participants (Gathercole et al. 1994, Gray 2003, Gupta 2003) and with performance on memory tests in neuropsychological patients (Romani 1992). Rehearsal would be impaired in children with fluency difficulty because a plan can only be rehearsed once it has been generated and these individuals have planning impairments. Thus, the planning problem gives rise to SS and UNWR deficits in people with fluency difficulty and would offer one explanation as to why %SS predicts UNWR.

NWR performance also requires non-words to be segmented and perceived accurately (Gathercole et al. 1994). Wolk et al. (1993) proposed that phonological analysis is a problem for children with fluency difficulty and this could lead to fragmentary disfluencies if these representations are involved in generation of speech plans for output. UNWR would be affected if the proposal that phonological analysis problems lead to NWR deficits (Gathercole et al. 1994) applies to individuals with fluency difficulty. Thus, if people with fluency difficulty have problems in segmentation and perceptual analysis, this would affect NWR ability and provide another route by which %SS predicts UNWR.

The account of how WWR arise, which was extended here to CwEAL, argued that this type of disruption to speech does not arise from planning problems per se. Rather WWR solve the planning problem in an indirect way by slowing prior speech rate. The lower speech rate allows more planning time that avoids fragmentary non-fluencies. WWR would not impact on UNWR performance, which is a result of phonological planning, rather than articulation. Therefore, WWR should be independent of UNWR and would not be affected across language groups who have different levels of WFD (as was found).

Practical applications of UNWR

UNWR could provide an alternative method from that of Howell (2013) for identifying children with fluency difficulty. Separating children into those with WFD who are otherwise fluent and those with fluency difficulty raises interesting possibilities for interventions. Group-based procedures could be developed for tackling WFD whereas SLTs could continue with the individually tailored interventions they currently give to children with fluency difficulties. The present results imply that interventions that tackle WFD should have an impact on WWR (i.e., they should reduce as the child learns more vocabulary) even though speech was not treated directly.