غربالگری کودکان در سن مدرسه برای خطر لکنت زبان
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|33561||2013||22 صفحه PDF||سفارش دهید||18610 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Journal of Fluency Disorders, Volume 38, Issue 2, June 2013, Pages 102–123
Objectives Howell and Davis's (2011) model that predicts whether stuttering in eight-year old children will persist or recover by teenage was adapted for screening school-aged children for risk of stuttering. Stuttering-severity scores were used to predict whether children belonged to fluent or stuttering groups. Predicted group assignments were compared for models in which severity measures were made with whole-word repetitions excluded or included. The best model for distinguishing children who stutter (CWS) from fluent children was validated across a wide range of ages. Design Stuttering-severity scores from CWS (222 for development, and 272 for validation, of the models) and fluent children (103 for development, and 25 for validation, of the models) were employed. Models were developed that predicted prognosis and screened CWS and fluent children. All these analyses were conducted both with whole-word repetitions excluded and included in the stuttering-severity scores. The model that screened fluent children from all CWS which excluded whole-word repetitions was validated for children across a range of ages. Results All models achieved around 80% specificity and sensitivity. Models in which whole-word repetitions were excluded were always better than those which included them. Validation of the screening for fluency with whole-word repetitions excluded classified 84.4% of fluent children, and 88.0% of CWS, correctly. Some of these children differed in age from those used to develop the model. Conclusion Howell and Davis's risk factor model for predicting persistence/recovery can be extended to screen school-aged children for fluency. Educational objectives: After reading this article, participants will be able to: (1) describe the difference between finding group differences and risk factor modeling in stuttering research; (2) summarize the strengths and weaknesses of stuttering severity instrument version three; (3) discuss how validation of diagnostic and screening models for fluency can be performed; (4) see how risk models have potential applications for screening for communication disorders in general
The first contact a speech language pathologist usually has with a child who stutters (CWS) is when the child attends a clinic for confirmation of diagnosis of the disorder and to decide on a course of treatment. There has been a period prior to the child's appearance at clinic where the child and his or her family had little or no professional advice about stuttering. Early clinical intervention is constrained when there is such a delay between when the disorder started and consultation at the clinic. This is potentially a problem because early intervention is usually considered to be more effective than later intervention (Yairi & Ambrose, 2005). The delay would be reduced if there were convenient methods for screening large unselected groups of children in order to identify stuttering at key stages in development (e.g. at school entry). However, no such screening instrument is currently available. Research, that may aid development of a screening instrument, has shown that children diagnosed as stuttering differ in many ways from fluent children (Yairi & Ambrose, 2005). Any of the factors that show significant differences are potentially useful for screening children for stuttering. The success with which such a factor correctly classifies the two groups of children as stuttering or fluent can be established by a statistical procedure such as logistic regression (Reed & Wu, in press). A clinical cohort does not have a supply of fluent children, and this is one reason why screening procedures have not been developed to date. A related question, that has been the focus of much research, is whether the prognosis of children in a clinical cohort (identification of CWS who will go on to recover or persist) can be predicted at their initial examination (Howell, 2010 and Yairi and Ambrose, 2005). Children who subsequently recover differ from the CWS who persist on the majority of the same measures found to differ between fluent control children and CWS (Howell, 2010). To establish whether any of these factors plays a role in long-term prognosis for stuttering, first measures on factors at the initial examination where prognosis is not known need to be obtained. Then the status of stuttering at an age at which stuttering has resolved into its recovered or persistent form has to be established. Finally, measures on factors obtained at the initial examination need to be correlated with persistence or recovery established at the later age at which prognosis was determined (Howell & Davis, 2011). The risk factors that predict persistence of the disorder may provide valuable information that help clinicians target resources on those who are most susceptible to long-term fluency problems (Reilly et al., 2009, p. 271; Yairi, Ambrose, Paden, & Throneburg, 1996, p. 74). There are two important implications of this discussion: first, just showing that two groups (fluent children versus CWS, or CWS who will persist versus CWS who will go on to recover) differ when some factor is measured, does not establish that what was measured is a risk factor for either stuttering in general or for its persistent form. In risk factor analysis a measure is taken (independent variable) and it is established how well it predicts group membership (dependent variable). When groups are tested to see whether they differ on some measure, the groups are selected (independent variable) and the measure is examined to see whether it differs between the groups (dependent variable). As Reed and Wu (in press) pointed out, relative to risk factor analysis, studies that look for differences between groups reverse “the relationship between the outcome and the predictor variables, making the outcome of interest into the independent variable, and the predictors into the dependent variable”; Second, measures that increase the risk of starting to stutter are not necessarily the same as those that increase the risk of persistence (Howell, 2010). Similarly, measures that predict onset or prognosis of stuttering may or may not apply to screening. The current study developed and assessed models for screening for stuttering that can be administered to unselected cohorts of children at selected ages, such as when they start school. The screening can be done in different ways, all of which are suitable for different clinical purposes. First, the screen might require children to be classified as fluent, likely to recover, or likely to persist (screen for stuttering types). This would be used if there is graded health care provision (e.g. parents of fluent children do not need to do anything about their child's fluency, parents monitor their child if he or she is considered likely to recover, whereas children likely to persist attend clinic). Second, the screen at the time of the first examination may need to separate the CWS who will go on to persist from fluent children and CWS who will later recover (screen for persistence). This may be appropriate if health services want to focus attention on children likely to have long-term fluency problems (i.e. the CWS who will persist). Third, the screen may be required to separate the children who are fluent from both those CWS who will go on to recover and those who will go on to persist (screen for fluency). This would be appropriate if health services want to examine all CWS irrespective of the expected path their stuttering will take. All three types of screen are addressed in this study. The next question is what factor or factors to measure (the independent variables). As mentioned above, the risk factors that are successful when examining one topic (e.g. prognosis that has been worked on) may or may not be useful for other topics (here screening that has not been addressed previously). The factor that was examined as potentially useful for screening was that used by Howell and Davis (2011) in their investigation into prognosis in a heterogeneous sample of CWS who were followed up longitudinally between the ages of eight years and teenage. The essential detail about Howell and Davis's study needed at this point (full details are given in Section 1.3) is that although they examined a wide range of risk factors that were obtained on CWS around the age of eight years, only one of them predicted whether the CWS would recover or persist at teenage (teenage is the age at which most childhood stuttering has resolved into recovered or persistent form). This was stuttering severity measured according to version three of Riley's (1994) instrument, SSI-3 (see Appendix A for a description and an appraisal of SSI-3 and Section 1.1 for details about SSI-3 that are particularly pertinent for the current work). Consequently, SSI-3 was examined as the potential predictor factor for each form of screen in the study reported below. SSI-3 may be useful for screening for stuttering because it incorporates a measure of the symptoms of stuttering and these would be expected to be rarer in fluent children. SSI-3 has a precise way of measuring severity. Probably the most notable aspect of SSI-3 is that it does not consider whole-word repetitions (WWR), as in “my, my, my friend”, to be symptoms of stuttering when percentage of syllables stuttered (%SS) are calculated. There is debate about whether WWR should or should not be included in %SS counts. In the work below, SSI-3 was calculated both with WWR excluded and included and the performance of the prognosis and screening models that resulted was compared. The remainder of the introduction starts by giving background information about why SSI-3 may be successful in prognosis and screening (Section 1.1). Evidence about the general role of WWR with respect to stuttering and the potential roles of WWR in screening and prognosis of stuttering are considered in Section 1.2. Section 1.3 compares the proposed modeling approach with approaches adopted in other risk factor studies on stuttering (prospective work on an unselected cohort, work that reports retrospectively whether adults had stuttered, and Howell and Davis's work on prognosis). 1.1. Properties of SSI-3 that are potentially relevant for prognosis and screening As indicated in Appendix A, the severity score supplied by SSI-3 is based on three components obtained on one or more speech samples: (1) the percentage of syllables stuttered (%SS) where stutters are defined in a precise way; (2) the duration of the three longest stutters in the samples; and (3) a score based on observed physical features extraneous to speech shown by the speaker (called physical concomitants). The conversion tables to produce values for each of the component measures so that they can be summed to give the total SSI-3 score are also given in Appendix A. The following issues about use of SSI-3 scores are singled out for further discussion here because they have relevance to performance of SSI-3 as a risk factor for prognosis and screening in the current study. (1) Exclusion of WWR in counts of %SS: The general debate about whether WWR are, or are not, a symptom of stuttering is an important one (see the discussion in Section 1.2). Here, two alternative ways of making symptom counts that have been used in SSI-3 are considered (with WWR excluded or included). Riley (1994) stated that WWR are usually not counted as stutters (see Appendix A). He also specified when words in WWR may be counted as stutters, namely when they are “shortened, prolonged, staccato, tense, etc.” However, when repeated words have the latter properties, they would be classified as part-word repetitions, prolongations or word breaks respectively. Because of this, they would be counted as stutters in any case, so this advice regarding when to count such WWR as stutters is not necessary (Howell, Soukup-Ascencao, Davis, & Rusbridge, 2011). Conversely, repeated words with these properties are not WWR. Other authors have a different point of view about WWR and whether they should be included in %SS for obtaining SSI-3 scores. For example, Anderson and Wagovich (2010) included all WWR in their counts of frequency of stuttered disfluencies (not just the short ones, etc.) that they then used to obtain SSI-3 scores. Instead of taking one position about whether or not WWR should be counted as stutters, in the study reported below empirical comparison was made between SSI-3 scores obtained with WWR excluded (Howell et al., 2011) and with them included (Anderson & Wagovich, 2010). For the stutter counts with WWR excluded, WWR were not counted as stutters unless they had properties that led them to be classified as other types of symptom allowed by Riley in counts of %SS. A further point to note is that leaving WWR out of SSI-3 calculations can affect duration scores and overall syllable counts as well as %SS (see Section 2). For the stutter counts with WWR included, all WWR were counted as stutters. The comparison of SSI-3 with WWR excluded and included should offer some indication about which way of counting %SS in SSI-3 is appropriate based on the impact this has on prognosis and screening. (2) Threshold cutoff value for fluent speakers: A fluent threshold is required when SSI-3 is used in screening unselected samples of children. Thresholds are given in the SSI-3 manual for very mild, mild, moderate, severe and very severe stuttering. However, although Riley (1994) indicated that SSI-3 scores are useful for diagnosis, there is no cutoff below which a child would be called fluent and above which a child would be said to stutter. Howell and Davis (2011) estimated an approximate cutoff value of 8 (the lowest score of CWS) for fluent children (the current study revises and extends this). The precise value may be procedure-dependent (see point 4 below, which describes how SSI-3 scores vary across the procedures that are permitted in the manual). As discussed in Appendix A, the original standards were obtained from audio recordings, and spontaneous and read samples alone were used for readers. These conventions about format and speech sample type are adhered to in the current study so that Riley's conversion tables and standards can be applied. (3) Scaling of scores across the severity range: Examination of the conversion tables for the %SS and duration components of SSI-3 show that the converted scores increase rapidly at the lower end of the scale (see Appendix A). Skewing components of the severity measure so their scores make proportionately more contribution for low-scoring individuals may be important for detecting children with low, but significant, levels of stuttering (further details about scaling of SSI-3 scores are given in Appendix A). The sensitivity of SSI-3 in situations where fluency problems are mild or absent may make the instrument particularly suited to separating fluent children from CWS. (4) Flexibility and constraints on how SSI-3 assessments are made: There is some flexibility on how SSI-3 scores are obtained insofar as different procedures can be used. Riley (1994) offered this flexibility so that his instrument could serve diverse needs ranging from use in clinics to research laboratories, and this has obvious advantages. However, the way that the different procedures affect scores and the impact these have on prognostic and screening predictions need to be ascertained. This study makes a partial contribution by comparing SSI-3 with WWR excluded or included as risk factors for prognosis and screening. Other than this, the procedures employed were constrained so that all severity measurements were obtained in the same way because it is known that different procedures lead to differences in SSI-3 scores (Howell et al., 2011 and Jani et al., submitted for publication). In summary, SSI-3 is potentially useful for screening because it excludes WWR and as it includes three different component measures (both of these may allow SSI-3 scores which may help distinguish fluent children from CWS). It was observed that SSI-3 has no fluency cutoff value and that SSI-3 scores are skewed so that mild fluency problems have more impact than severe problems in determining whether a child stutters. Although procedural flexibility is useful for ensuring that SSI-3 can be used in diverse situations, a single procedure has to be adhered to when data are pooled for analysis (Reed & Wu, in press). 1.2. Role of WWR in relation to stuttering The current study provides some data as to whether WWR are symptoms of stuttering or not and their role in distinguishing fluent children from CWS in that the performance of the screening models with WWR excluded and included were compared. This section reviews other evidence about whether WWR should or should not be considered stutters. Some studies have found differences in the nature and frequency of WWR between fluent and stuttering pre-schoolers that suggest WWR may have a different role in the two groups of children. For instance, pre-schoolers who stuttered produced monosyllabic WWR 3.5–4 times more frequently than non-stuttering pre-schoolers (Hubbard and Yairi, 1988 and Yairi and Lewis, 1984), and fluent children tended to pause longer than CWS between the spoken segments of a WWR (Throneburg, Yairi, & Paden, 1994). Whilst these studies might suggest that WWRs differ between young CWS and fluent children, their role as a risk factor in screening for stuttering has not been established nor has their use in predicting risk for long-term stuttering (prognosis) been addressed. In contrast, there is more evidence on older children and adults that suggests WWR are not symptoms of stuttering. There are several pieces of empirical evidence that indicate WWR have a different role to symptoms that are more typical of stuttering, such as prolongations, part-word repetitions and word breaks. Longitudinal work by Howell, Bailey, and Kothari (2010) showed that CWS who recovered had a higher rate of WWR relative to more typical symptoms than did persistent CWS when they were first examined (i.e. when both groups were stuttering), and these differences between the rates of these two types of symptoms across the groups of CWS increased up to teenage, at which age stuttering had resolved into recovered or persistent forms. This suggests that the decrease in proportion of WWR relative to more typical stuttering symptoms might be a risk factor for persistence of stuttering. One empirical study that supports this suggestion is that operant conditioning to increase the rate of WWR gave temporary improvement in fluency (Reed, Howell, Davis, & Osborne, 2007). A second study also supported the general implication that WWR are associated with recovery. Japanese is a highly inflected language. Consequently, since most WWR occur on monosyllabic words, Japanese speakers do not have many opportunities to produce WWR in their language. The more typical symptoms predominate in Japanese CWS (Ujihira, 2011). As a result of the low rate of WWR, the chance of recovering from stuttering in Japanese children is lower than that for English-speaking children (Ujihira, 2011). A final compelling piece of evidence that WWR have different roles to more typical stutters was from a scanning study (Jiang, Lu, Peng, Zhu, & Howell, 2012). First of all, it was established that more typical stutters have different brain activity patterns from other disfluencies that are commonly seen in fluent children's speech (WWR were not included in either group of symptoms at this stage). Subsequently, automatic classification procedures categorized the brain activity pattern of WWR as a member of the other disfluency class rather than the more typical class. Thus, the brain activity of WWR is different from more typical stutters and similar to that of the other disfluencies that are frequently seen in fluent speech. In summary, opinion is divided about whether or not WWR should be considered as symptoms of stuttering. Rather than take a particular stance on this issue, the current study examined whether SSI-3 scores calculated with WWR excluded or included (as indicated, there are precedents for both of these in the literature) affected prognostic and screening classifications. If different classification patterns occur in prognosis and screening when SSI-3 is used with WWR excluded as opposed to included, then these would provide additional evidence about the status of WWR. 1.3. Comparison of the current approach with other approaches to risk factor modeling in stuttering A recent study by Reilly et al. (2009) used a cohort of young English-speaking Australian children to examine whether a range of factors predicted which children would start to stutter. The risk factors examined were based upon studies that had established differences between CWS and fluent children. They then used logistic regression in a prospective study of a large sample of children and determined which children subsequently started to stutter. A few of the factors that were measured before stuttering started correlated significantly with stuttering outcome (indicating they may be risk factors for stuttering). However, the overall fit of the logistic regression model was poor. This arose because the sample consisted predominantly of fluent children, which restricted the ability of the logistic regression model to identify CWS as such (hits) and increased the chance of calling CWS fluent (misses) (Howell, 2009). This problem is common to most multivariate techniques when they are used to develop a model for data where one class dominates. In these situations, a satisfactory model is difficult to develop because there is a tendency to place all cases in the class with maximum members as this ensures an accuracy rate of at least that of the most frequent category (Howell and Davis, 2011 and Reed and Wu, in press). To illustrate, if the chance of stuttering is 5%, five children in a sample of 100 will stutter. Automatic classifiers using data from one or a number of measures obtained from the 100 children would achieve 95% correct performance by calling all children fluent. This performance looks reasonable, but no CWS is correctly classified. The problem does not apply when the number of cases in each class is balanced. Other aspects to note about Reilly et al.’s (2009) study were that there was significant selective attrition (as more mothers without degrees withdrew their children from the study than mothers with degrees), and that they considered WWR as a symptom of stuttering. Mothers with degrees might have particular interest in their child's language development, so the attrition of mothers without degrees is of concern. As has been seen in Section 1.2, WWR is a questionable symptom whose role with respect to stuttering has been debated for many years amongst experts. The inclusion of WWR would lead to misdiagnosis of stuttering in very young fluent children starting to speak if they are not symptoms of stuttering. This would explain the exceptionally high rates of stuttering reported by Reilly et al. (2009). Ajdacic-Gross et al. (2009), was a large-scale retrospective and opportunistic study. Swiss army conscripts self-reported stuttering (there was no attempt to distinguish recovered from persistent forms). This revealed some factors that corresponded with reports in other work, and other unexpected factors. The strengths of the paper were that a large number of participants was involved and they underwent detailed psychiatric examinations. As it was a retrospective study about factors that increased the chance of any type of stuttering, it suggested some risk factors for stuttering in general, but it did not address prognosis. The limitations in the study were due to the nature of retrospective studies (reported stuttering in the past cannot be verified) and, as the data were obtained as part of a general screening for conscripts (not specifically for stuttering), the authors did not include certain factors that are usually considered pertinent to stuttering. Howell and Davis's (2011) model for prognosis offers the possibility of addressing screening (neither of the previous studies would allow this). Howell and Davis (2011) used children who were confirmed to stutter at age eight who were known to have either continued stuttering or not (persistent/recovered) at teenage. The number of CWS who subsequently recovered was roughly equal to the number of CWS who persisted so development of the model avoided the imbalance problem discussed earlier in this section. Seven risk factors were available at age eight. These were: (1) Head injury; (2) Age at onset of stuttering; (3) Family history; (4) Handedness; (5) Speaking two or more languages in the preschool years; (6) Gender; (7) Stuttering severity. Logistic regression can automatically determine which factors from a group like this are significant predictors of the outcome and which are not (Reed & Wu, in press). Howell and Davis (2011) used this procedure and found that the SSI-3 scores were the only factor that predicted whether the CWS at age eight would recover or persist by teenage and they did so with around 80% sensitivity (called persistent CWS, persistent) and specificity (did not call CWS who would later recover, persistent). These findings have been replicated by Cook, Howell, and Donlan (2012) for a sample of German CWS. Howell and Davis (2011) pointed out that their model could be adapted for use with fluent children and that it could be extended to younger ages, which together potentially allow its use in screening unselected samples of children over a broad age range. The steps in the argument are: that SSI-3 can be obtained from fluent children as well as CWS; data are required on all participants that ensure that a child is fluent or stuttering when first examined and where there are longitudinal measures on CWS over the age range eight to teenage so that the form of stuttering at teenage (recovered/persistence) can be confirmed; that the models for different types of screen need to be validated and, if children from outside the age range that was used when the model was developed are included in this and performance is acceptable, it can be assumed that the models are applicable across these ages; similarly when the models are validated with samples from an unselected sample of children where there are imbalances between the numbers of fluent children and CWS and between the number of CWS who later recover and CWS who persist, it can be determined whether the performance of the models established on balanced samples is maintained. If so, the models developed on balanced data automatically adapt to any imbalances between fluent and CWS classes such as those that occur in the population of children at large. These points are each considered in turn in more detail. As mentioned, SSI-3 is a measure that can be obtained for fluent children. This is not the case for some of the other factors Howell and Davis (2011) obtained, such as age of stuttering onset. This unavailability issue would not apply to all of the other factors Howell and Davis (2011) examined. However, measurements on certain of the remaining factors are biased in other ways depending on whether they are obtained from fluent children or CWS, making them unsuitable for screening to distinguish these groups. As one example, Reilly et al. (2009) noted that history of reports of stuttering in the family changed dramatically as a child changed from not stuttering to stuttering, so family history measures depend on whether or not stuttering has been diagnosed. Similar influences might apply to factors like head injury if parents seek an explanation for why their child started to stutter. There are no such problems with SSI-3, which can be measured in the same way on fluent children as on CWS. Consequently, fluent children can be included in screening models that use SSI-3 as the sole predictor (Howell & Davis, 2011). One aspect of stuttering in the eight-teenage age range is that there are roughly equal numbers of CWS who will persist or recover at teenage (the current study adds approximately equal numbers of fluent children that are used in model development). This distribution allows optimal models to be developed for separating fluent and stuttering groups. Models for all the screens require confirmation that the children are either fluent or stutter when they are first seen. A further requirement is that longitudinal data are needed for screens for stuttering types and persistence, so that CWS can be classified as recovered or persistent at teenage (as mentioned earlier, teenage is the age at which most developmental stuttering has resolved into recovered or persistent form). After the above steps have been attended to, a model can be developed that predicts class membership of different combinations of fluent, CWS who will later recover and CWS who persist (depending on the type of screen). Any such model can be used to see which of the specified classes newly-assessed individuals are assigned to. These results can be used to validate the model against the available clinical assessment classifications for these new individuals where classifications depend on the particular screening type. Such validations can be done for children of any age (not just those in the age range that the model was developed on). As well as validating the selected model, the data on children outside the age range for which the model was developed show how well the model generalizes over age groups. Finally, the models can be examined with unselected samples of children to see whether their performance scales up appropriately when samples with natural imbalance are examined. For instance, if the screen for fluency was used to assess a sample of school children (i.e. with a different distribution of fluent children versus CWS from that in the sample on which the model was developed), does it maintain the ability to correctly classify 80% of the fluent children and 80% of the CWS? In general, whether the screening models scale appropriately across data with class imbalances can be established empirically by determining whether the classification performance of each group involved in the screen is maintained in the samples used for validation (around 80% correct for each type of classification). 1.4. Summary and predictions In summary, Howell and Davis (2011) found that the only risk factor needed to predict prognosis was SSI-3. An SSI-3 score can be obtained for fluent children, so SSI-3 is used as the factor in the risk factor models developed for prognosis and screening in this study. Performance of the models with SSI-3 calculated with and without WWR included as symptoms of stuttering are compared empirically to see whether WWR should be considered symptoms of stuttering. All three types of screen are examined (for persistence, for fluency, for types of stuttering). Satisfactory models for risk of stuttering cannot be developed for data that have imbalances between fluency classes (Reilly et al., 2009). Imbalances between numbers of CWS who will recover and CWS who persist were avoided here by using data from a similar age range to that used by Howell and Davis (2011) for model development. Roughly equal numbers of fluent children were added when models for screening were developed again to minimize the imbalance problem. The models can be validated outside the range they were developed for when data are available on children who have been independently classified as fluent or stuttering and, for some models, classification of CWS as will recover later or persist. The assumption that the models automatically scale when the distribution across classes changes (e.g. towards a high number of fluent children compared to CWS in unselected samples of school children) can be tested. The current study addressed the following questions concerning childhood stuttering: (1) was SSI-3 successful in predicting prognosis in the Howell and Davis (2011) study because WWR were excluded in its severity counts? This was assessed by seeing how well a model that uses SSI-3 at age eight to predict persistent or recovered outcome at teenage performed when WWR were excluded and included in the SSI-3 scores; (2) are SSI-3 scores useful for screening for: (a) persistence; (b) fluency; (c) stuttering types? As was the case when prognosis was examined, assessments were made with and without WWR included in SSI-3 scores; (3) at present, cases that can be used for validation are only available that indicate whether a child stutters or is fluent when first examined (not, in the case of CWS, whether the child persisted or recovered at teenage). Consequently, validation within and beyond the age range used in model development can only be performed for screening for fluency that does not require prognostic information. If this validation is successful, the model can be used to extrapolate prediction to cases outside the age range for which it was developed and to situations where the CWS are in a minority (e.g. at school intake).
نتیجه گیری انگلیسی
The findings show that Howell and Davis's (2011) model for predicting risk of stuttering can be successfully adapted to classify fluent children. This offers the possibility of it acting as a screening instrument for stuttering with children at the age of school intake. The model is successful at classifying CWS outside the age range for which it was originally developed.