The narrative coherence of witness transcripts in children on the autism spectrum

Background and Aims: Autistic children often recall fewer details about witnessed events than typically developing children (of comparable age and ability), although the information they recall is generally no less accurate. Previous research has not examined the narrative coherence of such accounts, despite higher quality narratives potentially being perceived more favourably by criminal justice professionals and juries. This study compared the narrative coherence of witness transcripts produced by autistic and typically developing (TD) children (ages 6–11 years,

The current study extended the story grammar approach to look at narrative coherence in witness transcripts produced by children on the autism spectrum. An interview comprising only open-ended questions was used, given that such methods are most likely to elicit story grammar elements (Feltis et al., 2010;Snow, Powell, & Murfett, 2009). Children between the ages of 6 and 11 years with and without an autism diagnosis were included, as at least some markers of story grammar are present in witness accounts throughout this age range (Brown et al., 2018;Westcott & Kynan, 2004). We did not look at developmental increases in the inclusion of story grammar elements as these are well-established (e.g., Brown et al., 2018;Feltis et al., 2011); the aim was to compare groups of autistic and TD children matched on age (as well as IQ and receptive language).
Importantly, this variability in results may depend on the tasks used. Open-ended tasks, such as describing personal narratives, recalling orally presented fairy tales, or making up a story to go with an emotionally ambiguous picture, tend to reveal greater difficulties for autistic children and adults than more structured and guided tasks such as narrating the story to a 'wordless picture book' (King et al., 2013;Lee et al., 2018;Losh & Capps, 2003;Losh & Gordon, 2014;Tager-Flusberg & Sullivan, 1995). Wordless picture book tasks often reveal almost no group differences at all (e.g., Losh & Capps, 2003;Norbury & Bishop, 2003), and may reduce the cognitive demands of storytelling by scaffolding memory, attention, story organisation and language production via providing temporally sequenced visual cues. Importantly, although in their meta-analysis Baixauli et al. (2016) reported no effects of narrative task, they could only assess two combined categories of narrative task given the limited number of available studies. Results also differ depending upon the nature of the comparison group. There are relatively few (or no) differences in narrative ability on some tasks between well-matched autistic and non-autistic children (e.g., Capps et al., 1998Capps et al., , 2000Diehl et al., 2006;Losh & Capps, 2003;Norbury & Bishop, 2003;Tager-Flusberg & Sullivan, 1995;Young et al., 2005). Such matching, most commonly for verbal ability and age, ensures that group differences -if found -are not just a function of language skills or developmental level.
When looking at results for the inclusion of the global elements of story grammar, previous research also presents a mixed picture. Norbury and Bishop (2003), for example, compared autistic and non-autistic children matched for age and non-verbal ability, but who showed differences on language measures including receptive vocabulary and grammar. There were no group differences in the inclusion of key story grammar elements. Diehl et al. (2006) included samples of autistic and TD children who all had ability levels in the 80 + IQ range and who were matched on age, IQ and language ability. They found no group differences in the inclusion of gist elements or the proportion of basic story elements recalled (although there were group differences in causal connectivity). Similarly, Lee et al. (2018) found no group differences in the inclusion of story grammar elements between autistic and TD adults (the groups differed on verbal IQ so this variable was controlled in the analyses). Nonetheless, others have described differences in the inclusion of story elements between autistic and non-autistic groups (Banney et al., 2015;Goldman, 2008;Losh & Capps, 2003; see also Pearlman-Avnion & Eviatar, 2002, although groups in this study were not matched for language); found differences in the types of information produced (e.g., reductions in the preponderance or amount of 'gist' rather than 'detail' information: Barnes & Baron-Cohen, 2012;McCrory et al., 2007); or reported group differences in a combined meta-analysis of relevant studies (Baixauli et al., 2016).
In evaluating previous research into narrative coherence in individuals on the autism spectrum, several difficulties emerge. First, studies have used different elicitation methods, ranging from wordless picture books (Losh & Capps, 2003), to describing autobiographical memories (King et al., 2013), to recalling sections from television programmes (Barnes & Baron-Cohen, 2012) or making up a story to an ambiguous picture (Lee et al., 2018). Only one study assessing narrative coherence in autistic individuals has looked at eyewitness memory (Pearlman-Avnion & Eviatar, 2002, using a narrated slide show), which is a more realistic everyday remembering task. In the current study, a secondary analysis of the story grammar elements produced in witness transcripts collected in a previous study of autistic and non-autistic children (Henry, Messer et al., 2017) is presented. Second, the approach to matching has differed. Some studies have matched for chronological age, intelligence (IQ -usually verbal but sometimes non-verbal), and one or more aspects of language ability (e.g., Banney et al., 2015); others have used more than one comparison group (e.g., with matching for chronological age and IQ and language ability and IQ, King et al., 2013); and others have matched on some but not all of these variables (e.g., Capps et al., 1998Capps et al., , 2000Norbury & Bishop, 2003;Tager-Flusberg & Sullivan, 1995;Young et al., 2005). It may be difficult to match on multiple indices, particularly if language ability is included, given the heterogeneity and variability of language skills found in autistic individuals (Kwok, Brown, Smyth, & Oram Cardy, 2015;Taylor, Maybery, & Whitehouse, 2012). Indeed, some studies do not attempt to do so (Goldman, 2008) or control key variables statistically (Lee et al., 2018). In the present study, all available participants were included from the larger study who could be matched on age, IQ and one aspect of language (receptive vocabulary), although other aspects of language ability varied significantly between groups. This provided a test of narrative skill in autistic children who were similar in age, intellectual ability and receptive vocabulary to a comparison group of non-autistic children. Finally, there are few studies with larger sample sizes. The current sample size (52 autistic children, 52 TD children) was larger than in previous studies, and a conservative significance level was used to account for multiple group comparisons (see Banney et al., 2015).
We also measured the length of the transcripts in terms of total number of words, and their semantic diversity in terms of number of different words. This is important, as although many previous studies have found no significant differences between autistic and typical individuals in story length or semantic diversity, albeit using slightly varying measures (Banney et al., 2015;Diehl et al., 2006;Losh & Capps, 2003;McCabe et al., 2013;Norbury & Bishop, 2003;Tager-Flusberg & Sullivan, 1995;Young et al., 2005), there are some reports of differences (Baixauli et al., 2016;Capps et al., 2000;King et al., 2013;Lee et al., 2018). Finally, we included preliminary correlational analyses between story grammar measures and other cognitive and language variables, although few relationships were expected given limited findings in previous research (Banney et al., 2015).
Matching TD and autistic groups on several indices including age, IQ and receptive language level may minimise differences in narrative performance. Therefore, it was hypothesised that there would be no significant differences between the groups with regard to the number and diversity of 'story grammar' elements included in their accounts. We also tentatively predicted no significant differences in overall length or number of different words.

Participants
The data used in this project were taken from interviews carried out as part of a larger study conducted by Henry, Messer et al. (2017), which compared eyewitness memory skills in children with and without autism diagnoses. The original study included 272 participants (162 boys and 110 girls, aged between 76 months and 142 months). Of these children, 71 had a formal diagnosis of autism, obtained independently of the research study by a suitably qualified professional. Children were recruited from mainstream primary schools or special educational needs schools in Greater London or South-East England.
Selecting a meaningful comparison group for studies of children on the autism spectrum is not straightforward when discrepant 'peaks and valleys' represent a common cognitive profile (Burack, Iarocci, Flanagan, & Bowler, 2004). We tried to avoid 'overmatching' where statistical bias is accidentally caused by matching for what is believed to be a confounding variable but is actually a core feature of the autism profile. Burack et al. (2004) advise matching on a subset of features, but caution against matching 'out' the key features of autism. In line with previous literature, matching was carried out for age, IQ and one measure of language (receptive vocabulary). Receptive vocabulary was selected following Goldman (2008), who reported non-significant differences in receptive vocabulary scores in their samples, together with substantial differences on two subtests from a broader language battery (the Clinical Evaluation of Language Fundamentals, CELF-4 UK, Semel, Wiig, & Secord, 2006). All children who recalled at least three items of correct information in their interview accounts of the witnessed event were eligible for inclusion (three children were excluded at this stage: two autistic, one TD). Individual matches within +/-10 points for both IQ and receptive vocabulary and +/-6 months for age were hand selected.
The resulting sub-sample comprised: 104 children (52 in the TD group, and 52 in the autism group) matched on age, IQ and receptive language ability. There were 83 boys and 21 girls (TD group: 38 boys and 14 girls; autism group: 45 boys and 7 girls). It was not possible to match the groups exactly on gender, as the autism group in the original sample included a high proportion of males (62 boys and 9 girls). The composition of the current sample reflects the wider autistic population, where males are estimated to outnumber females (Loomes, Hull, & Mandy, 2017). Table 1 provides mean scores on all background variables. Mervis and Klein-Tasman (2004) suggest that for groups to be well-matched, the group distributions on the control variable in question should overlap, with a p-level of at least .50 on the test of mean differences. An independent samples t-test was used to compare the groups, as these data were normally distributed. On all three indices, age, IQ and receptive vocabulary, the groups were closely matched (see Table 1). Data are also reported on other measures of language, memory and attention (not matched between groups), as well as the total number of correct items of information recalled in the initial interview (see Table 1).
Mirroring the original study findings (Henry, Messer et al., 2017), the groups differed in the number of correct details recalled about the witnessed event (Table 1), with autistic children (M = 25.71, SD = 14.37) recalling fewer correct details than TD children (M = 33.42, SD = 14.24), t(102)=-2.75, p = .007. However, the analyses of story grammar elements and length/diversity of narratives did not focus solely on correct details, but included the full transcripts for all children.

Measures
Participants' intellectual abilities were assessed in the original study using the Wechsler Abbreviated Scale of Intelligence (WASI-II; Wechsler & Zhou, 2011). An estimate of full-scale IQ was obtained using 'Vocabulary' from the Verbal Comprehension Index and 'Matrix Reasoning' from the Perceptual Reasoning Index. As noted, following matching, the groups did not differ on this measure.
Language level of participants was measured using several tasks from the original study. Receptive vocabulary was assessed using the British Picture Vocabulary Scale -Third Edition (BPVS-3; Dunn, Dunn, & Styles, 2009), which requires children to select the correct picture from a choice of four upon hearing a word spoken. Two subtests from the Expressive Language Test -2 (ELT-2; Bowers, Huisingh, Logiudice, & Orman, 2010) included: 'Sequencing', to test narrative ability, and 'Grammar and Syntax', to test grammatical morphology. A further two subtests were taken from the Clinical Evaluation of Language Fundamentals -4th edition (CELF-4 UK; Semel et al., 2006). 'Recalling Sentences' assesses the ability to accurately repeat sentences using internalised grammatical structure, and 'Formulated Sentences' assesses the ability to produce full, grammatically accurate, and meaningful sentences about picture stimuli.
When compared on the language measures, the two participant groups were well-matched on receptive vocabulary ability (BPVS), as expected (Table 1), with no significant difference between the autism and TD groups. However, significant group differences (or marginal differences) remained in scores on the other language measures (ELT-2 'Sequencing'; ELT-2 'Grammar and Syntax'; CELF-4 'Recalling Sentences'; CELF-4 'Formulated Sentences').
An assessment of general memory ability from the Test of Learning and Memory (2 nd Edition, TOMAL-2, Reynolds & Voress, 2007) included: Memory for Stories; Paired Recall; Facial Memory; and Visual Sequential Memory. Table 1 includes composite scores for the Verbal Memory Index, on which groups did not differ, and the Nonverbal Memory Index, on which the autistic children obtained lower scores than the TD children (although no differences were present on the combined Composite Memory Index). Three assessments of focused, sustained and sustained-divided (dual task) attention completed the test battery (Sky Search; Score!; and Sky Search Dual Task from the Test of Everyday Attention for Children, TEA-Ch, Manly, Robertson, Anderson, & Nimmo-Smith, 1999). There were no group differences on the attention measures.

Materials and procedure
Ethical approval for the original study and the present secondary data analysis study was granted by the Research Ethics Committees of the universities at which the research took place. All parents or guardians provided informed written consent for their children to take part. The original research study involved data being collected from both TD and autism groups across several phases, in a manner intended to replicate the processes involved in gathering evidence from eyewitnesses in the criminal justice system (evidence gathering statements, investigative interviews, identification line-ups, cross-examinations). Only the first set of interview data (evidence gathering statements, or 'Brief Interviews') administered on the same day as the event was witnessed, are relevant to the current paper. These initial questions about the event were included to mimic a response officer's initial contact with a * Groups were matched on these variables and, as expected, no significant group differences were observed. 1 Standardised scores (mean 100 SD 15). 2 scaled scores (mean 10 SD 3). 3 non-standardised raw scores.
witness. The procedure used in this phase of the study is summarised briefly below. For a more detailed description, please see Henry, Messer et al. (2017).

Experimental procedure
Participants were shown a staged event, featuring two actors who gave a short presentation about what school was like in Victorian times. For practical and logistical reasons, this was either viewed live during a school assembly, or a video of the performance was shown to the children. There were no recall differences between live and video presentation for autistic or TD children (Henry, Messer et al., 2017) so these data were combined. The talk consisted mainly of facts about Victorian schools, but involved a staged minor crime, where either a phone or a set of keys was 'stolen' by one of the actors (the 'theft' was later explained as a misunderstanding). Participants were randomly assigned to one of two versions of the talk, which were virtually identical but used different materials and gave alternative names to the actors. The rationale was to give some indication of the generalisability of the results. As there were no differences between versions, data were collapsed over this variable (Henry, Messer et al., 2017).
After viewing the staged event, participants were interviewed individually on the same day. Interviewers used a standard protocol with an initial question designed to elicit a free recall account ('Tell me what you remember about what you just saw'). This was followed by a series of open-ended questions that could be used to prompt further information depending on the child's response (e.g. 'Who was there?'; 'What did they do?'). Finally, participants were asked if they could remember anything else. This final prompt was repeated until the child could no longer provide any new items of information.
Interviews were audio-taped and transcribed, then coded by two independent coders according to the narrative frameworks described below. The coders were blind to group status.

'Story grammar' narrative analysis
The story grammar framework developed by Murfett et al. (2008) was used to analyse participants' accounts by tallying the number of narrative features included. 'Correctness' of information in the transcripts was not taken into account, as this would not be possible in a real case when jurors and legal professionals are unlikely to know what actually happened to a victim/witness. Further, story grammar is a 'framework' for organising event details and, as such, is scored independently of accuracy (Feltis et al., 2011).
Seven narrative elements were scored: Setting, Initiating event, Internal response, Plan, Action/attempt, Direct consequence, and Resolution (see Table 2). Elements were scored based on the number of times they appeared in the narrative, regardless of their accuracy (e.g., a point would be awarded for setting if the child mentioned the location of the event, even if this was incorrect). The total number of story grammar elements and the number of different story grammar elements (maximum 7) used were noted.

Additional structural measures
Two structural measures were hand coded: the total number of words in each account, and the total number of different words used (excluding non-word utterances such as 'um'), to obtain a broad comparison of the overall length and semantic diversity of the narratives produced by each group.

Reliability analyses
All transcripts were coded independently by two coders, following discussion to agree the coding rules. Intra-class correlation coefficients were calculated to assess the reliability of coding, and these indicated good to excellent reliability for all but one story grammar element (Direct consequence): Setting: ICC = .94; Initiating event: ICC = .83; Internal response: ICC = .81; Plan: ICC = .92; Action/attempt: ICC = .82; Direct consequence: ICC = .08; Resolution: ICC = .80. For the aggregated story grammar element scores,

Table 2
Examples of scored story grammar elements from the witness transcripts.

Story grammar element
Examples from the witness transcripts Setting "This morning at about quarter to ten…" "It happened straight after assembly when we was doing our test and we got called downstairs…" Initiating event "There was two men, who came to a school… they came to tell them about how it was like in the Victorian times a hundred years back." "This morning we had a talk…" Internal response "They were being quite funny so we sort of, we had to try and keep our laughter in…" "It was very quick…" Plan "They were videoing, uh, videoing, yeah, what they were doing to show other schools…" "Mark said, I'm going to leave my, my jumper on the floor and, um, my keys on the, on the chair…" Action/attempt "The other man, um, took a chalkboard out of the red bag…" "He turned it round and wrote the answer…" Direct consequence "One of them forget, um, forgot their, um, one of the bags so they had to come and pick it up…" "Then he said, where's those keys and we all, and we all, um, pointed at Alex…" Resolution "And then we just went back to music…" "And then they just said good-bye and they went…" reliability was also good: Total number of grammar elements, ICC = .94; Number of different grammar elements, ICC = .93. Intercoder agreement for the word count measures was excellent: Total number of words, ICC = 1.00; Number of different words, ICC = 1.00. Given the poor agreement for Direct consequence, the first coder looked at the discrepancies in coding (coder 2 noted more instances of this element than coder 1) and recoded data based on commonalities in coding criteria. After recoding, agreement was much higher, ICC = .98. Nevertheless, given the initial discrepancies, we suggest caution in interpreting the results for Direct consequence. First coder ratings were used for all data analyses. Table 3 provides means, medians and ranges for each story grammar element and the structural measures (total number of words, number of different words). As group comparisons across several areas were conducted, a more stringent significance level (p < .01) was adopted to limit the chances of incorrectly rejecting the null hypothesis (recommended by Banney et al., 2015).

Results
Data for the structural measures of total number of words and number of different words were normally distributed, so independent samples t-tests were performed to compare groups. 'Total number of words' showed no significant difference between groups [t (102) Given the groups did not differ on overall length or diversity of their narratives, data on story grammar elements were analysed using raw scores. Table 3 shows that means and medians for many story grammar elements were low, with medians of zero in one or both groups for initiating event, internal response, plan, direct consequence, and resolution. Kolmogorov-Smirnov normality tests indicated that scores for the story grammar elements and their totals violated the parameters of normality (either for both groups or one group) on every measure, therefore, non-parametric tests were used (Mann-Whitney U).
There were no significant differences in the distributions for any of the story grammar elements, or for the total scores ]. In many cases, medians across the two groups were identical (see Table 3).
To investigate whether any background variables related to Total number of story grammar elements or to Number of different story grammar elements, preliminary Spearman non-parametric correlations were carried out. For the TD and autism groups, Total number of story grammar elements correlated with Total correct details recalled in the interview: Autism group r = .71; TD group r = .79 (ps < .001). Number of different story grammar elements also correlated with Total correct details recalled in the interview: Autism group r = .76; TD group r = .64 (ps < .001). There were correlations between Total number of story grammar elements and age in the TD group, which failed to reach significance in the autism group: Autism group r = .25, p = .07; TD group r = .41, p = .003, and between Number of different story grammar elements and age in the TD group, which again failed to reach significance in the autism group: Autism group r = .27, p = .055; TD group r = .41, p = .003. No other correlations were significant.

Table 3
Mean scores (SDs), medians and ranges for each story grammar element for the autism and typical development (TD) groups.

Discussion
This study investigated the narrative coherence of eyewitness accounts in autistic and non-autistic children (6-11 years) matched on age, IQ and receptive language ability. We replicated a previous finding that autistic children recalled fewer correct items of information than non-autistic children in a brief interview about a staged event involving a minor crime, witnessed earlier that day (Henry, Messer et al., 2017). Despite this difference in volume of correct recall, no significant group differences emerged when a 'story grammar' framework was used to assess the presence and number of key story elements in participants' accounts (incorporating all information mentioned, whether correct or not). The findings confirm the utility of previous work using the story grammar framework to evaluate eyewitness accounts of children with developmental differences (Brown et al., 2018;Gentle et al., 2013;Murfett et al., 2008), and provide support for claims that there are relatively few (or no) differences in narrative ability on some tasks between well-matched groups of autistic and non-autistic children (e.g., Capps et al., 1998Capps et al., , 2000Diehl et al., 2006;Losh & Capps, 2003;Norbury & Bishop, 2003;Tager-Flusberg & Sullivan, 1995;Young et al., 2005).
On metrics assessing the length in words and semantic diversity (number of different words) of the accounts, there were also no group differences in performance. This supports much previous literature (e.g., Banney et al., 2015;Diehl et al., 2006;Losh & Capps, 2003;McCabe et al., 2013;Norbury & Bishop, 2003;Tager-Flusberg & Sullivan, 1995;Young et al., 2005). The present findings also suggest that, although children on the autism spectrum may recall fewer items of correct information about an event, their overall narrative accounts are not shorter or less semantically diverse, nor are they less coherent in terms of narrative structure. Taken together with findings that eyewitness accuracy in autistic children is generally as high as in comparable non-autistic children (2017b, Almeida et al., 2019;Bruck et al., 2007;Henry, Messer et al., 2017;McCrory et al., 2007; although accuracy levels may vary with interview type -see Mattison et al., 2018), the present results provide further evidence for the overall reliability of autistic child witnesses. Narrative coherence adds important information, because it may affect the degree to which members of the criminal justice system understand a child's evidence. This could be via impacting on how meaningful and credible the child appears to jurors ( , Brown et al., 2018Feltis et al., 2010;Gentle et al., 2013;Murfett et al., 2008), helping the police to establish relevant 'points to prove', and/or contributing to decisions by the Crown Prosecution Service about whether or not to authorise charges.
The most commonly included story grammar element in children's narratives related to descriptions of what happened in the event (i.e., Action/attempt details, Mdn = 5.5 autism group; Mdn = 6.0 TD group). This confirms previous observations that Action/ attempt details are the most commonly elicited story grammar element. Feltis et al. (2011) found about 50 % of story grammar elements in immediate recall were Action/attempt details, close to the 60-65 % recorded in the current study. Nearly all other story grammar elements had median scores at or close to zero, indicating that children in the current age range (6-11 years) often omitted several story grammar elements. For Setting details, the median score of 2 items in both groups indicated that many children included information about where and when the event took place (see also Westcott & Kynan, 2004). However, the median number of types of story grammar elements included was only 3 out of a possible 7, underlining the fact that maturation in narrative coherence was not complete in the current sample of 6-11 year old children. Further research should assess narrative coherence in witness accounts of older children who include more story grammar elements, to negate the possibility that low scores on some measures reduced our ability to detect group differences. Nevertheless, group differences on story grammar measures with higher scores (e.g. total scores, action/attempt) were still absent, increasing confidence in the findings.
Although autistic (relative to matched non-autistic) children produced narratives of similar coherence (based on an analysis of story grammar elements), length (based on number of words) and semantic diversity (based on number of different words), the type of narrative task is important. Losh and Capps (2003) noted that experimental findings may not reflect children's narrative competence "within the less structured and more socially demanding contexts of daily life" (p.248). In the same way that wordless picture book tasks may reduce cognitive demands of storytelling by scaffolding memory, attention, story organisation and language production using temporally sequenced visual cues, adult-led interviews are structured interactions in which specific narrative features may be prompted through direct questions (e.g., 'Where did this happen?'). This is not necessarily typical of how children's narratives are produced in real conversational exchanges, although it does reflect how forensic interviews might proceed (although note that we did not use full investigative interviews, which would usually occur later than the 'same-day' brief evidence gathering statements used in the current study). Further, since autistic children often experience difficulties with emotional regulation (Samson et al., 2014), they might perform less well when describing a 'real' witnessed event, particularly if it induced heightened emotions or distress. Therefore, caution is required when generalising the current findings to more emotionally charged situations such as are likely to be encountered in criminal investigations. Finally, as the current groups were matched on age, general ability and one aspect of language (receptive vocabulary), it is important to note that the findings could be different in less closely matched samples.
Preliminary correlational analyses assessed whether background age and cognitive measures related to the number or diversity of story grammar elements included in children's accounts. Correlations between both of these measures and the number of correct details recalled in the interview suggested that volume of correct recall is a good indicator of the extent to which both autistic and TD children include key story elements. Higher volume of recall, better structure and greater narrative coherence could all have an impact on the credibility of a witness before jurors (see also Henry, Ridley, Perry, & Crane, 2011). The relationships between number and diversity of story grammar elements and age (significant only in the TD group) confirmed the known developmental improvements in the use of story grammar elements (Feltis et al., 2011).

Summary
This study investigated the narrative skills of 52 autistic children aged 6-11 years (IQs 70+), comparing them to 52 TD peers L.A. Henry, et al. Research in Developmental Disabilities 96 (2020) 103518 matched on age, IQ and receptive language ability. Participants had been interviewed using open-ended prompts shortly after viewing a staged mock crime event, and interview transcripts were coded using existing analytical frameworks for narrative features (story grammar elements; narrative length; semantic diversity). There were no significant differences between the autism and TD groups in terms of narrative skills on any of these measures. The findings add to, and extend, previous reports that when children on the autism spectrum are well-matched to peers with TD on cognitive and some language measures, they exhibit a similar level of narrative skill. The current findings also attest to the comparability of autistic and non-autistic child witnesses in terms of accuracy, account length and narrative coherence, if not in volume of correct information recalled.

Declaration of Competing Interest
None.