# paper.pdf ## Metadata - PDFFormatVersion=1.4 - Language=en-US - IsLinearized=false - IsAcroFormPresent=false - IsXFAPresent=false - IsCollectionPresent=false - IsSignaturesPresent=false - CreationDate=D:20260601200907+02'00' - Creator=Adobe InDesign 18.3 (Windows) - Custom.CrossMarkDomains[1]=springer.com - Custom.CrossMarkDomains[2]=springerlink.com - Custom.CrossmarkDomainExclusive=true - Custom.CrossmarkMajorVersionDate=2010-04-23 - Custom.doi=10.1007/s10802-026-01469-4 - ModDate=D:20260611035105+02'00' - Producer=Adobe PDF Library 17.0; modified using iText® 5.5.13.4 ©2000-2024 iText Group NV (SPRINGER SBM; licensed version) - Trapped=False - crossmark:crossmarkdomains=springer.comspringerlink.com - pdfx:crossmarkdomains=springer.comspringerlink.com ## Contents ### Page 1 Received: 24 December 2025 / Accepted: 15 May 2026 © The Author(s) 2026 Alessandra Giuliani Tamsin Sharp Yeukai Chideya Katie Spurgin Richard Meiser-Stedman Mark Tomlinson Sarah L. Halligan 1 Department of Psychology, University of Bath, Bath, BA,UK 2 Bristol Medical School, University Of Bristol, Bristol, UK 3 Institute for Life Course Health Research, Department ofGlobal Health, Stellenbosch University, Stellenbosch, SouthAfrica 4 Department of Psychology, Institute of Psychiatry,Psychology and Neuroscience, King’s College London,London, UK 5 School of Nursing and Midwifery, Queen’s UniversityBelfast, Belfast, UK # Abstract # Cognitive models of post-traumatic stress disorder (PTSD) propose that traumatic events are encoded in fragmented,sensory-rich, and poorly contextualised memories that maintain PTSD symptoms. Although this hypothesis has receivedsupport in adults, evidence in children and adolescents remains limited, methodologically heterogeneous, and largelyrestricted to high-income, English-speaking settings. We pooled data from 407 children and adolescents exposed to trauma(6–17 years) recruited from emergency departments in South Africa and the United Kingdom. A month after their trau-matic experience, participants were asked to recall their traumatic event and completed self-report measures of traumamemory quality and PTSD symptoms. Trauma narratives were transcribed, and indices of coherence and sensory valencewere derived using three narrative-based approaches. We examined the validity of narrative-derived indices with self-reported memory quality, their associations with concurrent PTSD symptoms, and whether these patterns varied acrosscultural context and age. Across both cultural contexts, self-reported trauma memory qualities were significantly associ-ated with PTSD symptoms. By contrast, narrative-derived indices showed no convergence with the corresponding self-report measures and did not meaningfully predict PTSD symptoms. Limited age-related differences were observed: in theSouth African sample, self-reported trauma memory characteristics were associated with PTSD only among older youth,whereas in the UK sample, they were predictive across age groups. In children and adolescents exposed to trauma, com-monly used narrative-based indices of memory disorganisation and sensory content have no construct validity nor clinicalutility for predicting PTSD. In contrast, self-report measures of trauma memory features emerge as a reliable and cross-culturally valid indicator of PTSD symptomatology. Keywords Post-traumatic stress · Trauma memory · Child and adolescent · Cross-cultural Cross-Cultural Validation of Manual and Automated Methods forExtracting Trauma Memory Features and Predicting PTSD in YoungPopulations Alessandra Giuliani1[ ![http://orcid.org/0009-0002-4199-3611]() ](http://orcid.org/0009-0002-4199-3611) · Tamsin Sharp2 · Yeukai Chideya3 · Katie Spurgin1 · Richard Meiser-Stedman4 ·Mark Tomlinson3,5 · Sarah L. Halligan1 Research on Child and Adolescent Psychopathology (2026) 54:75 [ ![http://crossmark.crossref.org/dialog/?doi=10.1007/s10802-026-01469-4&domain=pdf&date_stamp=2026-6-1]() ](http://crossmark.crossref.org/dialog/?doi=10.1007/s10802-026-01469-4&domain=pdf&date_stamp=2026-6-1) ![Springer Logo]() ### Page 2 Introduction Many children experience traumatic events, such as seri-ous accidents, interpersonal violence, or acute medicalepisodes (Lewis et al.,). For most, post-traumaticreactions are transient (Hiller et al.,). However, aminority go on to develop persistent post-traumatic stresssymptoms (PTSS), including re-experiencing, avoidance,hyperarousal, and negative alterations in cognition andmood, that may meet diagnostic criteria for post-traumaticstress disorder (PTSD) in approximately 15% of cases(Alisic et al.,; American Psychiatric Association,). Early identification of children at heightened riskfor developing PTSD is critical for guiding timely and tar-geted interventions, and psychological theories can pro-vide valuable insights into the factors associated with theemergence of post-traumatic psychopathology. Cognitive models of PTSD highlight trauma memorycharacteristics as central to symptom development andmaintenance (Brewin et al.,; Ehlers & Clark,).According to these models, trauma memories are encodedand stored differently from typical autobiographical memo-ries, exhibiting greater fragmentation, sensory richness, andcontextual disconnection (Brewin et al.,; Ehlers &Clark,). These theories have been primarily developedand tested in adult populations. However, evidence sug-gests that similar aetiological factors are relevant in youth,supporting the application of these theoretical frameworksto children and adolescents (Reed et al.,). However,when specifically examining trauma memory characteris-tics, developmental differences in children’s language andautobiographical memory skills may confound measure-ment (Fivush,). It is unclear whether observable fea-tures of how children recount their trauma can meaningfullyindex memory processes or predict PTSD symptomatology.It is therefore critical to evaluate the suitability of differ-ent trauma memory assessment methodologies in youngerpopulations and to determine their relevance for identifyingchildren at elevated clinical risk. Measuring Trauma Memory Characteristics A range of methodologies have been developed to assesstrauma memory characteristics, resulting in inconsistenciesacross studies and limited comparability. Broadly, thesemethods fall into two categories: self-report questionnairesand narrative-based approaches. Self-Report Measures of Trauma Memory The most widely used self-report measures in PTSDresearch are the Trauma Memory Quality Questionnaire(TMQQ; Meiser-Stedman et al.,) and its AdaptedChild version (ACTMQQ; Hiller et al.,). These scalesindex children’s own perceptions of how disorganised orsensory-rich their trauma memories are. Given that thesememory qualities are inherently subjective, these self-reportmeasures allow direct insight into children’s appraisals oftrauma memories. That is, they provide access to internal,phenomenological qualities of memory which are not easilyobservable. From a cognitive perspective, such appraisalsmay reflect how accessible, intrusive, or poorly contextu-alised the memory feels to the individual, processes thatare central to leading cognitive models of PTSD (Brewinet al.,; Ehlers & Clark,). Across multiple studies,self-reported trauma memory qualities have shown moder-ate to strong associations with concurrent PTSD symptoms(Reed et al.,). This suggests that children’s subjectiveappraisals of their memories may capture clinically relevantaspects of memory processing that are closely related tosymptom development. Self-report measures may also beless affected by developmental differences in expressivelanguage ability compared to narrative-based approaches,making them potentially more suitable for assessing traumamemory characteristics in younger populations. Narrative-Based Approaches Narrative methods, in contrast, focus on the observablestructure and content of an oral or written trauma account.Compared to self-report measures, narrative methods offerthe advantage of capturing more spontaneous and nuancedmemory expressions, potentially reflecting the natural struc-ture and organisation of the memory more objectively. How-ever, it has been argued that the extent of disclosure maydiffer between age groups and trauma type (McElvaney etal.,). Moreover, narrative methodologies require vali-dation in children, as children tend to show poor integrationof autobiographical memories because of less-developedlanguage abilities (Fivush,; Nelson & Fivush,).Specifically, three main narrative-based approaches havebeen employed in the literature: manual coding, word-count-based methods, and Natural Language Processing (NLP). Manual Coding A common method for extracting memory characteristicsfrom trauma narratives is manual coding, in which trainedraters code trauma narratives for features such as repeti-tion, disorganised or unfinished thoughts, and sensory refer-ences (Foa et al.,; Halligan et al.,; Salmond etal.,). This approach allows for nuanced, context-sensi-tive judgements. However, it is resource-intensive, vulner-able to rater subjectivity, and has produced mixed findings ### Page 3 in youth: while some studies have found no significantassociations between narrative disorganisation or sensorydetail and PTSS (e.g., McKinnon et al.,), others havereported modest positive relations for specific narrative fea-tures only (e.g., McGuire et al.,), even when similarcoding schemes have been used. Word-Count-Based Methods Word-count-based methods quantify the presence of particu-lar linguistic categories within a narrative by counting wordsbelonging to predefined dictionaries or semantic classes. Awidely used dictionary-based text analysis tool is the Lin-guistic Inquiry and Word Count (LIWC; Pennebaker et al.,), which applies a validated dictionary to text to iden-tify linguistic and psychological features, producing scoresthat represent the percentage of words falling into relevantcategories. In adult trauma research, LIWC indices such ascognitive mechanism words (e.g., “because”) or analyticthinking have been interpreted as indicators of more coher-ent and organised trauma memories (Jelinek et al.,) andhave shown associations with PTSD severity (Jaeger et al.,; Kleim et al.,). However, evidence in childrenand adolescents is far more limited and notably inconsistent.For example, Miragoli et al. () reported associationsbetween LIWC-derived emotional language and children’strauma responses following an earthquake, whereas otherstudies of youth have failed to identify reliable links betweenLIWC indices and PTSD symptoms or narrative organisa-tion. Importantly, there is little evidence that LIWC capturesstructural features such as fragmentation or disorganisationin young populations, raising concerns about its sensitiv-ity to the trauma memory properties emphasised in cogni-tive models of PTSD. Moreover, because LIWC does notaccount for linguistic context or higher-order narrative struc-ture, its applicability may be particularly constrained wherechildren’s developing language abilities limit the complexityand coherence of trauma recall. Natural Language Processing (NLP) NLP approaches offer a more sophisticated alternative toword-count-based approaches by analysing the relationshipsbetween words and ideas across the entire narrative. Tech-niques such as Latent Semantic Analysis, Latent DirichletAllocation, and computational coherence algorithms learnfrom large text corpora and can quantify semantic related-ness, topic structure, and global coherence without relyingon predefined dictionaries (Landauer et al.,; Li,).Early work suggests that NLP-derived features can classifyPTSD status from adult trauma narratives with promisingaccuracy (Bartal et al.,; Zhang,). However, NLPtools are typically trained on English-language data fromadult populations (Hovy & Prabhumoye,), and theirperformance with child narratives, particularly in non-Eng-lish or translated narratives, remains unexplored. This limita-tion reflects broader structural biases in NLP development,whereby only a small subset of the world’s languages aresufficiently resourced to support robust model training (Blasiet al.,; Joshi et al.,; Udeze,). Crucially, thesechallenges extend beyond cross-language differences. Evenwithin English, substantial sociolinguistic and cultural varia-tion exists, including differences in dialect, narrative style,and culturally shaped modes of emotional expression. Suchvariation can influence how experiences are encoded in lan-guage, yet is often underrepresented in the standardised cor-pora used to train NLP models (Zampieri et al.,). As aresult, models developed on dominant or “standard” formsof English may inadequately capture the linguistic featuresof narratives produced in diverse cultural contexts. Current Gaps In sum, both narrative-based and self-report approaches offerdistinct strengths and limitations. However, several impor-tant gaps remain in the trauma narrative literature. First, themajority of work examining trauma memory characteristicshas been conducted with adults, and the limited researchinvolving children and adolescents has typically treatedthem as a homogeneous group (McGuire et al.,).Yet, substantial developmental changes in language, auto-biographical reasoning, and metacognitive abilities occuracross middle childhood and adolescence (Fivush,;Nelson & Fivush,). These evolving capacities affectboth how trauma memories are encoded and how they canbe expressed, meaning that methods appropriate for olderadolescents may be poorly suited to younger children andvice versa. Second, existing work on trauma memory char-acteristics in youth has been conducted almost exclusivelyin high-income, English-speaking contexts, despite PTSDbeing a global condition (Atwoli et al.,; Figueira etal.,; Pham et al.,). It is unclear whether narrative-based indices derived from English-language tools (e.g.,LIWC dictionaries, NLP models) perform adequately whenapplied to narratives translated from other languages. Dif-ferent languages and cultural contexts may exhibit distinctstructural and lexical patterns (Schnell & Schiborr,),which may affect how we interpret indices such as ‘cognitivemechanism words’ or ‘analytic thinking’. Finally, very fewstudies have directly compared multiple memory measures,manual coding, automated narrative indices, and self-reportwithin the same sample. Direct comparisons are essentialbecause they allow assessment of convergent and discrimi-nant validity across methods, clarifying whether different ### Page 4 tools are capturing the same underlying memory constructsor instead reflecting distinct processes such as language abil-ity or subjective appraisal. One notable exception is Rubinand colleagues (), who reported the predictive value ofnarrative features in adults. However, that study focused ona small sample and did not include children or adolescents,limiting its relevance for developmental applications. Current Study The current study addresses these gaps by providing a sys-tematic, cross-cultural comparison of manual coding, word-count and NLP approaches, and self-report questionnairesfor assessing trauma memory characteristics in children andadolescents. These methods are evaluated across two largesamples of youth exposed to trauma: one from a low- andmiddle-income country (LMIC) context (South Africa) andthe other from two studies conducted in the United King-dom, a high-income country (HIC). Cultural factors, suchas willingness to discuss traumatic experiences and aware-ness of mental health difficulties, may well differ betweenthese samples (Monnapula-Mazabane & Petersen,). Forexample, UK contexts often place greater emphasis on indi-vidual emotional awareness and verbal expression of internalstates, which may facilitate more elaborated narratives andhigher endorsement of subjective memory qualities in self-report measures. In contrast, in some South African contexts,cultural norms may place greater emphasis on behaviouralcoping, social resilience, or reduced verbal disclosure ofdistress, which could result in less elaborated narratives orlower endorsement of internal states despite similar underly-ing experiences (Monnapula-Mazabane & Petersen,).By comparing findings across these settings, the study exam-ines the extent to which cultural factors influence the valid-ity of each approach and their capacity to identify childrenat heightened risk of PTSD. We operationalise validity asthe correspondence between narrative-based and self-reportindicators of memory qualities and predictive utility, reflectedin associations with post-traumatic stress symptoms, and weexamine developmental differences by comparing younger(6–12 years) and older (12–17 years) participants. Specifically, the study aims to: - 1. Evaluate the validity of manual coding, word-count,NLP, and self-report methodologies for assessing traumamemory characteristics in children across different cul-tural and language contexts. - 2. Assess the extent to which memory characteristicsderived from each method predict concurrent PTSDsymptom severity. - 3. Examine whether the predictive utility of these methodsdiffers between younger and older children. Methods Participants This study pooled secondary data from three closely alignedcohorts of children and adolescents who had been exposed totrauma, recruited from hospital emergency departments aftera traumatic incident. In total, the combined dataset comprises407 participants from South Africa and the United King-dom (UK). The South African sample was drawn from theSinethemba study (Sharp et al.,), a longitudinal projectinvestigating post-trauma adaptation among youth living insocioeconomically disadvantaged communities in Cape Town.Participants (N = 222; 154 male participants; 61.6%) wereaged 8–17 years (M = 12.41, SD = 1.29) and were all identifiedas Black African. Caregiver characteristics were as follows:caregivers (M = 41.16, SD = 10.96) were predominantly moth-ers (68.0%), followed by other biological relatives such asgrandparents or uncles (20.1%), fathers (4.7%), stepmothers(3.0%), other caregivers (1.7%), and sisters (1.3%). Regard-ing marital status, 54.6% were married or in a relationship.Caregivers’ educational attainment was predominantly at theschool level: 88% had completed up to Grade 12, 11% held apost-matric certificate or diploma, and 1% held a degree. The UK sample consisted of data from two studies: PRO-TECT (N = 128; Hiller et al.,) and ASPECTS (N = 57;Meiser-Stedman et al.,), yielding a combined sampleof 185 children and adolescents exposed to trauma (118male participants; 63.8%) from the UK, aged 6–17 years(M = 11.78, SD = 1.17). The UK sample was predominantlyWhite (93.8%). Caregiver characteristics for the UK sampleincluded a mean caregiver age of 39.7 years (SD = 7.0), with90% being mothers. The majority were married or cohab-iting (74%), and educational attainment was distributedacross school-level qualifications (27%), further education(38%), and higher education (35%). Across cohorts, trauma exposure differed by sample.In the UK sample, trauma types included road traffic acci-dents (49.4%), accidental injuries (31.5%), assault (6.7%),acute medical emergencies (5.6%), dog bites (2.2%), andother events (7.9%). In the South African sample, traumaexposure comprised accidental injuries (39.2%), road trafficaccidents (24.8%), assault (14.4%), other events (18.5%),and acute medical emergencies (0.5%). Procedure Across studies, procedures followed similar proto-cols. Detailed recruitment procedures and samples aredescribed in the original publications of their studies.Across all studies, written informed consent was obtainedfrom caregivers, and age-appropriate assent from all child ### Page 5 and adolescent participants was obtained at the point ofrecruitment. Trained research staff explained the studyaims, procedures, confidentiality, and the voluntary natureof participation to both caregivers and participants, ensur-ing comprehension and providing opportunities to askquestions. Participants were informed that they couldwithdraw at any time without consequence. At any follow-up assessments, consent and assent were reconfirmed priorto data collection. Approximately a month following thetraumatic event, participants and their caregivers com-pleted an interview and a battery of self-report measures.During the interviews, participants were asked to recalltheir traumatic incident using the following question: “Iwould now like you to tell me about the event that ledto your recent hospital visit in as much detail as you can.You can do this in any way that you want; there are norules. Try to tell me everything you know about what hap-pened and include anything about what has happened sincethat you think is important. I will leave you to talk for aslong as you like”. Prompts (e.g., “Do you remember anymore?”) were given by the interviewer when the child dis-closed little information. All interviews were audio-recorded. For the SouthAfrican cohort, interviews were conducted in isiXhosa orEnglish, depending on the participant’s preference. Forthe UK cohort, interviews were conducted in English.Audio recordings were transcribed verbatim by trainedtranscribers. For the South African cohort, isiXhosa nar-ratives were translated into English, prior to analysis,following a multi-step procedure designed to preservemeaning and minimise semantic loss. This included for-ward translation by bilingual researchers, review anddiscussion within the research team, and resolution of dis-crepancies through consensus. All subsequent narrativecoding and automated text analyses were conducted onthe English-language transcripts. Narrative Pre-Processing Before applying manual and automated methods for traumamemory feature extraction, narrative transcripts underwentstandardised pre-processing to ensure consistency across stud-ies. Narratives containing fewer than 50 words (2 in the SouthAfrican Sample and 1 in the UK sample) were excluded, con-sistent with prior methodological conventions, which indicatethat very short texts provide insufficient linguistic content forreliable coding and computational analysis (Pennebaker et al.,). Transcripts were then cleaned to retain only the par-ticipants’ speech and remove interviewers’ questions and com-ments. Participants’ responses to specific follow-up questionswere excluded to preserve the narrative’s natural flow andavoid artificially inflating coherence scores. No further pre-processing was performed (e.g., stop-word removal, lemmati-sation, or stemming) in order to preserve the original linguisticproperties of the narratives (Khurana et al.,). Materials Demographic information, trauma-related details (includ-ing type of event and basic injury characteristics), and self-reported measures of trauma memory quality and PTSDsymptoms were obtained from children and adolescents dur-ing the interview. For the South Africa sample, questionnaireswere translated into isiXhosa prior to administration. Trans-lation followed established forward–backward translationprocedures. Materials were first translated from English intoisiXhosa by bilingual research staff and then independentlyback-translated into English by a separate bilingual speaker.Discrepancies were reviewed and resolved through discussionwithin the research team to ensure conceptual equivalence. PTSD Symptoms PTSD severity was measured using established age-appropri-ate clinical measures of PTSD. Specifically, the PROTECTstudy used the University of California at Los Angeles Reac-tion Index (UCLA-RI; Steinberg et al.,). The ASPECTand Sinethemba studies used the Children’s PTSD SymptomScale Self-Report (CPSS-5-SR; Foa et al.,). Both instru-ments assess the frequency and severity of PTSD symptomsover the past month. The UCLA-RI yields total scores rang-ing from 0 to 80, whereas the CPSS-5-SR yields total scoresranging from 0 to 60. To account for this, total scores fromboth scales were divided by their respective maximums,yielding standardised scores ranging from 0 to 1. Self-Reported Trauma Memory Quality Self-reported trauma memory characteristics wereassessed using the Trauma Memory Quality Question-naire (TMQQ; Meiser-Stedman et al.,) in theASPECTS cohort and the Adapted Child Trauma MemoryQuality Questionnaire (ACTMQQ; Hiller et al.,) inthe PROTECT and South African cohorts. The TMQQcomprises 11 items that assess two theoretically deriveddimensions of trauma memory: disorganisation (e.g.,confusion, gaps, difficulty piecing the memory together)and sensory qualities (e.g., vivid perceptual details). TheACTMQQ retains all TMQQ items and includes an addi-tional seven items to broaden developmental suitabilityand enhance coverage of both domains, while maintain-ing the same two-factor structure. ### Page 6 To ensure direct comparability across cohorts, only the 11items common to both versions were included in the pres-ent analyses. Each item is rated on a 4-point Likert scale(1 = strongly disagree; 4 = strongly agree). Subscale scores fordisorganisation and sensory qualities were calculated as themean of their respective items, generating scores from 1 to 4,with higher values indicating more fragmented and more sen-sory-vivid trauma memories. Manual Coding Manual coding produced three variables: a global disor-ganisation rating, an elements-of-disorganisation score, anda sensory score. It was carried out by three trained ratersusing a coding scheme adapted from Foa et al. () andpreviously employed by McGuire et al. (). The manualscheme is available in the supplementary materials (MaterialA). Raters were blind to the study aim, PTSD severity, andparticipants’ demographics, and each rater coded all narra-tives. To minimise coder drift (Myford & Wolfe,), thecoding team held regular calibration meetings throughout thecoding period and jointly reviewed a subset of narratives atpredefined intervals. Inter-rater reliability was assessed usingthe intraclass correlation coefficient (ICC), based on a mean-rating (K = 3), two-way random-effects model. Across codingcategories, reliability was good to excellent, with ICC val-ues ranging from 0.76 to 0.95. Discrepancies were resolvedthrough consensus discussion; where consensus could notbe reached, disagreements were resolved by an independentthird-party reviewer, which was required for approximately10% of coded instances, consistent with best-practice guide-lines (Syed & Nelson,). Following the manual, narratives were divided intochunks coded as repetitions, organised thoughts, disorgan-ised thoughts, unfinished thoughts, pain utterances, andsensations. Based on the manual instructions, raters alsoprovided a global rating score (0–10) for the degree of mem-ory disorganisation of the narrative as a whole, with higherscores indicating greater disorganisation. Repetition, organ-ised thoughts (reversed), disorganised thoughts, and unfin-ished thoughts were transformed into z-scores and combinedto comprise ‘elements of disorganisation’. Pain and sensorycategories were transformed into z-scores and combined toproduce the ‘sensory’ variable. LIWC Linguistic Inquiry Word Count is used to produce the twooutcome variables of Disorganisation and Sensory. The Dis-organisation variable was computed as the composite scorefor Cognitive Processes (the percentage of words related tothe category containing 797 words, such as “cause”, “know”,and “ought”) and Analytical Thinking. Analytical Thinkingis a summary measure representing the degree to which thelanguage in the text suggests formal, logical, and hierarchi-cal thinking patterns. The Sensory variable was computedthrough the Perception category (sum of the percentage ofterms related to “sight”, “hearing”, and “touch”/ “feeling”). NLP Several metrics of overall text semantic overlap extractedusing TAACO (Tool for the Automatic Analysis of Cohesion;Crossley et al.,), a free automatic text analysis software,were aggregated to compute a disorganisation score. TAACOcomputes semantic overlap (lexical cohesion within a text)using three computational models: latent semantic analy-sis (LSA) (Landauer et al.,), Latent Dirichlet Alloca-tion (LDA) (Blei et al.,), and Word2vec (Mikolov etal.,). All three scores range between 0 and 1. LSA andWord2vec are computed using the cosine similarity betweensegments. LDA scores are calculated from the Jensen-Shan-non divergence between the normalised combined vectorweights of words within each segment. Therefore, the LDAscore indicates dissimilarity between adjacent segments andwas reversed. Each narrative’s similarity score was computedby averaging the three measures. Each narrative was assigneda final score ranging from 0 (low cohesion) to 1 (high cohe-sion). Aggregating text cohesion metrics has been shown tocapture text cohesion reliably (Miani et al.,). We employed the concreteness norms from Brysbaert etal. (2014) to capture the narratives’ sensory content. This isa norms document for 40,000 English lemmas, scored ona concreteness scale from 1 to 5. More concrete words areeasier to imagine, visualise, and experience; hence, a con-creteness score can be utilised to capture the overall sensorycontent of a text (Paivio). We extracted valence andconcreteness scores for each narrative by computing meanscores of the two variables of all the words present in thetext. Higher valence values indicate more positive senti-ment; therefore, the final score was reversed. Power Analysis An a priori power analysis was conducted using G*Powerversion 3.1.9.6 (Faul et al.,) to estimate the sample sizerequired to detect a medium-sized effect of trauma memorycharacteristics on PTSD, while adjusting for three covari-ates (age, sex, and narrative length) in a multiple regres-sion framework. The analysis determined that a minimumsample size of 99 was required to detect an effect size of ² =0.15, with 90% power, at an alpha level of 0.05. ### Page 7 Data Analysis All analyses were conducted separately for the South Africaand UK cohorts to evaluate cross-contextual generalisabilitywhile preserving site-specific variability in narrative and clini-cal characteristics. Prior to modelling, data were examined foroutliers, distributional assumptions, missingness, and multi-collinearity. Missing data on questionnaire items were mini-mal (< 10%). When fewer than 20% of items within a scalewere missing, mean substitution within that participant’s com-pleted items was applied to compute prorated total scores. Allanalyses relied exclusively on robust linear regression modelsto mitigate the impact of heteroskedasticity and outliers. To evaluate the validity of each feature extractionmethod (Aim 1), we examined whether narrative-basedindices (manual coding composites, word-count-basedvariables, and NLP-derived measures) corresponded toself-reported trauma memory disorganisation and sensoryqualities. This was tested using robust linear regressionmodels in which each narrative feature served as a pre-dictor of the relevant TMQQ subscale. To assess predic-tive utility (Aim 2), equivalent robust regression modelswere estimated with standardised PTSD severity scoresas the dependent variable. Narrative-based indices andself-reported memory qualities were evaluated in separatemodels to determine their unique associations with PTSDsymptom severity. To examine developmental differences(Aim 3), all predictive models were repeated following apriori age-based stratification (< 12 years vs. ≥12 years),reflecting key developmental transitions in autobiographi-cal memory and language ability. Differences in effectpatterns were evaluated descriptively due to subgroupsample size constraints. Age and narrative length wereincluded as covariates in all models. All analyses wereconducted using R 4.3.1 (R Core Team,). Results Descriptive Statistics Descriptive statistics for demographic characteristics, narra-tive features, and trauma memory variables for each sampleare presented in Table. Corresponding descriptive statis-tics stratified by age group (< 12 vs. ≥ 12 years) are providedin Supplementary Tables S1–S2. As shown in Table, par-ticipants in the UK sample produced longer narratives onaverage than those in the South Africa sample, whereasmean self-reported memory disorganisation and sensoryscores were higher in the UK sample. Correlations betweentrauma memory measures, narrative length, and age are inthe Supplementary Tables S3 and S4. Validity of Narrative Feature Extraction Methods Across both samples, narrative-based measures showed min-imal evidence of convergent validity. In the South Africansample (N = 222; df = 218), none of the manual, word-count,or NLP indicators significantly predicted either traumamemory disorganisation or sensory qualities (Table). Acomparable pattern emerged in the UK sample (N = 185;df = 181), with the exception that the LIWC sensory score | Table 1 Descriptive statistics for demographics, word count, narrative-derived features, and self-reported trauma memory scores. | | - | | Variables | South Africa | UK | | M | SD | Range | M | SD | Range | | Age | 12.55 | 2.53 | 8.04–16.95 | 10.91 | 2.77 | 6–17.75 | | Children | 9.97 | 1.15 | 8.04–11.92 | 9.78 | 1.85 | 6–12 | | Adolescents | 14.38 | 1.39 | 12.05–16.95 | 14.8 | 1.66 | 12.04–17.75 | | PTSD scores | 0.32 | 0.16 | 0–1 | 0.28 | 0.23 | 0–1 | | Narrative word count | 263.48 | 297.21 | 19–2262 | 422.79 | 351.86 | 26–1841 | | Manual coding | | Disorganisation features* | 2.12 | 2.34 | -4.84–14.48 | 7.45 | 2.47 | -1.08–16.42 | | Overall disorganisation | 4.29 | 1.29 | 0–10 | 4.46 | 2.33 | 0–10 | | Sensory features* | 4.28 | 1.01 | -0.56–5.38 | 6.17 | 3.12 | -1.43–15.22 | | Word-count based | | Disorganisation | 63.24 | 18.36 | 0–86.46 | 70.94 | 17.44 | 0–102.18 | | Sensory features | 3.05 | 1.90 | 0–12.95 | 2.75 | 1.71 | 0–15.56 | | NLP-derived | | Coherence | 0.59 | 0.09 | 0.40–0.95 | 0.64 | 0.11 | 0.34–0.92 | | Sensory features | 2.64 | 0.12 | 2.35–3.04 | 2.48 | 0.10 | 2.02–2.84 | | Self-reported | | Disorganisation | 1.58 | 0.59 | 1–4 | 2.25 | 0.85 | 1–4 | | Sensory features | 1.65 | 0.65 | 1–4 | 2.12 | 0.68 | 1–3.89 | | *Note: The `Disorganisation features’ and ‘Sensory features’ variables are determined by summing the z-scores across coding categories | | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | ### Page 8 showed a weak but statistically significant association withself-reported sensory qualities (β = 0.14, SE = 0.05, t = 3.14,p = .002, f2 = 0.05; Table). According to Cohen’s ()benchmarks, this represents a small effect. No method dem-onstrated significant predictive power for self-reported dis-organisation in either context. Prediction of Post-Traumatic Stress SymptomSeverity The second aim of this study was to assess the extent towhich memory characteristics derived from each methodpredict the severity of PTSD symptoms. Robust regressionswere estimated with PTSD severity as the dependent vari-able. Narrative-based measures did not significantly predictPTSD scores in either the South African or UK samples(Table). In contrast, self-reported trauma memory quali-ties consistently predicted PTSD scores. In the South Afri-can sample, both sensory qualities (β = 0.35, SE = 0.05,t = 7.64, p < .001, f2 = 0.27) and disorganisation (β = 0.22,SE = 0.07, t = 3.11, p = .002, f2 = 0.04) were significant posi-tive predictors. These represent medium-to-large and smalleffect sizes, respectively (Cohen,). A similar patternwas observed in the UK sample, where sensory qualities(β = 0.37, SE = 0.06, t = 6.64, p < .001, f2 = 0.24) and dis-organisation (β = 0.29, SE = 0.06, t = 4.52, p < .001, f2 =0.11) showed robust associations with PTSD scores, reflect-ing medium effect sizes (Cohen,). Thus, only self-report measures demonstrated meaningful predictive utilityacross cultural contexts. The correlations between memory | Table 2 Regression summaries of narrative-derived variables predicting self-reported memory characteristics in the South African sample | | - | | Models, predictors | Self-reported sensory qualities | Self-reported disorganisation | | Manual coding | Disorganisation features | | Memory characteristic | β = –0.22, SE = 0.33, t = − 0.67, p = .504, f2 < 0.01 | β = –0.07, SE = 0.06, t = − 1.20, p = .230, f2 < 0.01 | | Age | β = 0.06, SE = 0.02, t = 2.93, p = .004**, f2 = 0.04 | β = –0.03, SE = 0.02, t = − 1.38, p = .171, f2 < 0.01 | | Narrative length | β ≈ 0.00, SE = 0.01, t = 1.19, p = .235, f2 < 0.01 | β ≈ 0.00, SE = 0.00, t = − 1.30, p = .196, f2 < 0.01 | | Overall disorganisation score | | Memory characteristic | – | β = 0.04, SE = 0.11, t = 0.37, p = .320, f2 < 0.01 | | Age | – | β = –0.12, SE = 0.12, t = − 1.02, p = .267, f2 < 0.01 | | Narrative length | – | β = 0.02, SE = 0.02, t = − 1.03, p = .196, f2 = 0.01 | | Word count approach | | Memory characteristic | β = –0.04, SE = 0.04, t = − 1.08, p = .259, f2 = 0.01 | β = –0.03, SE = 0.05, t = 0.61, p = .541, f2 < 0.01 | | Age | β = –0.06, SE = 0.02, t = 2.64, p = .005**, f2 = 0.03 | β = –0.04, SE = 0.03, t = 1.16, p = .053, f2 = 0.01 | | Narrative length | β ≈ 0.00, SE = –0.01, t = 0.08, p = .976, f2 < 0.01 | β ≈ 0.00, SE = 0.01, t = 1.68, p = .467, f2 = 0.01 | | NLP approach | | Memory characteristic | β = 0.02, SE = 0.01, t = 4.12, p = .906, f2 < 0.01 | β = –0.35, SE = 0.53, t = − 0.66, p = .906, f2 < 0.01 | | Age | β = 0.01, SE = 0.00, t = 1.67, p = .788, f2 = 0.01 | β = 0.054, SE = 0.02, t = 2.79, p = .788, f2 < 0.01 | | Narrative length | β ≈ 0.00, SE = 0.01, t = − 1.61, p = .108, f2 = 0.01 | β ≈ 0.00, SE = 0.00, t = − 1.61, p = .548, f2 < 0.01 | | *p<.05; **p<.01; ***p<.001 | | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | | Table 3 Regression summaries of narrative-derived variables predicting self-reported memory characteristics in the UK sample | | - | | Models, predictors | Self-reported sensory qualities | Self-reported disorganisation | | Manual coding | Disorganisation features | | Memory characteristic | β = 0.50, SE = 0.31, t = 1.62, p = .107, f2 =0.01 | β = 0.27, SE = 1.08, t = − 0.25, p = .092, f2 < 0.01 | | Age | β = –0.01, SE = 0.02, t = − 0.37, p = .710, f2 < 0.01 | β = –0.06, SE = 0.04, t = − 1.49, p = .137, f2 = 0.01 | | Narrative length | β ≈ 0.00, SE = 0.00, t = 1.07, p = .286, f2 < 0.01 | β ≈ 0.00, SE = 0.00, t = − 0.005, p = .770, f2 < 0.01 | | Overall disorganisation score | | Memory characteristic | — | β = 0.03, SE = 0.04, t = 0.73, p = .464, f2 < 0.01 | | Age | — | β ≈ 0.00, SE = 0.01, t = − 1.02, p = .492, f2 < 0.01 | | Narrative length | — | β = –0.02, SE = 0.02, t = − 1.03, p = .494, f2 < 0.01 | | Word count approach | | Memory characteristic | β = 0.14, SE = 0.04, t = 3.14, p = .002**, f2 = 0.05 | β = –0.02, SE = 0.03, t = − 0.75, p = .130, f2 < 0.01 | | Age | β = –0.01, SE = 0.02, t = − 0.41, p = .680, f2 < 0.01 | β = –0.00, SE = 0.00, t = − 0.79, p = .948, f2 < 0.01 | | Narrative length | β ≈ 0.00, SE = –0.01, t = 0.49, p = .619, f2 < 0.01 | β ≈ 0.00, SE = 0.01, t = − 0.85, p = .554, f2 < 0.01 | | NLP approach | | Memory characteristic | β = 0.59, SE = 0.66, t = 0.89, p = .543, f2 < 0.01 | β = 0.93, SE = 1.15, t = 0.81, p = .893, f2 < 0.01 | | Age | β = 0.12, SE = 0.20, t = 0.61, p = .591, f2 < 0.01 | β = –0.03, SE = 0.02, t = − 1.73, p = .234, f2 = 0.02 | | Narrative length | β ≈ 0.00, SE = 0.01, t = 0.82, p = .412, f2 < 0.01 | β ≈ 0.00, SE = –0.01, t = 0.73, p = .398, f2 < 0.01 | | *p < .05; **p < .01; ***p < .001 | | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | ### Page 9 characteristics derived from each method and PTSD sever-ity are presented in Fig. for both the South African andUK samples. Developmental Differences The third aim of this study was to examine whether the utilityof each memory characteristic extraction methodology forpredicting PTSD differs between children and adolescents.To assess age-related differences, analyses were repeatedseparately for younger (< 12 years) and older (≥ 12 years)participants. None of the narrative-based approaches signifi-cantly predicted PTSD in either age group in the South Afri-can or UK sample (see Supplementary Tables S4 and S5). In the South African sample of younger children (N = 92;df = 88) aged 8.04–11.92 (M = 9.97, SD = 1.15), neither thesensory nor the disorganisation subscales of the TMQQsignificantly predicted PTSD severity scores (respectively,β = 0.14, SE = 0.33, t = 0.42, p = .205, f2 < 0.01 and β = 0.05,SE = 0.12, t = 0.42, p = .675, f2 < 0.01). However, amongadolescents (N = 130; df = 126) aged 12.05–16.95 (M = 14.38,SD = 1.39), both the sensory and disorganisation subscaleswere found to predict PTSD (respectively, β = 0.52, SE = 0.08,t = 6.26, p < .001, f2 = 0.31 and β = 0.34, SE = 0.08, t = 4.09, p< .001, f2 = 0.13), representing large and medium effect sizes,respectively (f2 > 0.15; Cohen,). Graphical depictionsof the estimated robust linear regression model are shown inFig.a and b for children and adolescents, respectively. In the UK sample of younger children (N = 143; df = 139)aged 6–12 (M = 9.78, SD = 1.85), both the sensory and dis-organisation subscales of the TMQQ significantly predictedPTSD severity scores (respectively, β = 0.23, SE = 0.04,t = 6.45, p < .001, f2 = 0.30 and β = 0.03, SE = 0.002, t = 3.98,p < .001, f2 = 0.11). Equally, for adolescents (N = 42; df = 38) | Table 4 Regression summaries for trauma memory characteristics predicting PTSD severity | | - | | Models, Predictors | South Africa | UK | | Manual coding | | Sensory | β = –0.26, SE = 0.42, t = − 0.62, p = .536, f2 < 0.01 | β = 0.17, SE = 0.12, t = 1.43, p = .154, f2 = 0.01 | | Disorganisation | β = 0.62, SE = 0.37, t = 1.67, p = .096, f2 = 0.01 | β = 0.06, SE = 0.05, t = 1.22, p = .224, f2 < 0.01 | | Overall disorganisation score | β = 0.32, SE = 0.24, t = 1.32, p = .073, f2 = 0.01 | β = 0.12, SE = 0.12, t = 0.99, p = .492, f2 < 0.01 | | LIWC | | Sensory | β = 0.54, SE = 0.53, t = 1.01, p = .492, f2 < 0.01 | β = 0.00, SE = –, t = 1.17, p = .137, f2 = 0.01 | | Disorganisation | β = 0.01, SE = 0.01, t = 0.91, p = .587, f2 < 0.01 | β = 0.28, SE = 0.28, t = 1.01, p = .126, f2 = 0.01 | | NLP | | Sensory | β = 0.58, SE = 0.41, t = 1.98, p = .059, f2 = 0.02 | β = 0.12, SE = 0.13, t = -1.98, p = .491, f2 = 0.02 | | Coherence | β = − 0.11, SE = − 0.12, t = -2.06, p = .312, f2 = 0.02 | β = − 0.17, SE = 0.11, t = -2.47, p = .175, f2 = 0.03 | | Self-reported | | Sensory | β = 0.35, SE = 0.05, t = 7.64, p < .001***, f2 = 0.27 | β = 0.37, SE = 0.06, t = 6.64, p < .001***, f2 = 0.24 | | Disorganisation | β = 0.22, SE = 0.07, t = 3.11, p = .002**, f2 = 0.04 | β = 0.29, SE = 0.06, t = 4.52, p < .001***, f2 = 0.11 | | *p < .05; **p < .01; ***p < .001 | | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | ### Page 10 aged 12.04–17.75 (M = 14.79, SD = 1.66) in this sample,both the sensory and disorganisation subscales were foundto predict PTSD (respectively, β = 0.28, SE = 0.06, t = 4.50, p< .001, f2 = 0.53 and β = 0.23, SE = 0.05, t = 3.94, p < .001, f2= 0.41). Graphical depictions of the estimated robust linearregression model are shown in Fig.a and b for children andadolescents, respectively. Discussion Across two socio-cultural contexts, three main findingsemerged. First, narrative-based methods demonstrated a lackof validity, showing no convergence with self-reported traumamemory qualities and little to no predictive value for PTSDsymptom severity. Second, self-reported disorganisation andsensory properties of trauma memory emerged as reliable cor-relates of PTSD. Third, developmental patterns differed bycontext, such that self-reported trauma memory qualities pre-dicted PTSD from early childhood in the UK, but only fromearly adolescence onward in South Africa. The most central and unexpected finding was the system-atic failure of any narrative-based approach to show con-vergence with self-reported trauma memory characteristics.This finding offers important new insight. Narrative-derivedfeatures have been used in the past literature as measuresof memory characteristics to test models of PTSD in bothadult and youth research (Halligan et al.,; McGuire etal.,; Miragoli et al.,; Salmond et al.,), withseveral studies reporting null or inconsistent associationswith PTSD symptoms, especially in youth. Yet, in our data,narrative-derived indices did not capture any variance inchildren’s subjective experience of trauma memory, indicat-ing a fundamental lack of construct validity, at least in pae-diatric samples. However, it is critical to consider that theseindices were largely developed and validated in English-speaking adult populations, and may not adequately capturehow trauma is narrated in different linguistic and culturalcontexts. In particular, narratives produced in isiXhosa andsubsequently translated into English may not preserve themorphosyntactic, pragmatic, or culturally embedded fea-tures that these tools are designed to detect. In addition, all ### Page 11 narrative-based methods failed to predict participants’ con-current PTSD severity. Together, these findings challengethe assumption that the cognitive mechanisms outlinedin cognitive PTSD models manifest reliably in narrativestructure in youth. Unlike adults, children may be unableto express core memory processes into coherent verbaldescriptions, due to limitations in language development,narrative organisation, and metacognitive insight (Nelson &Fivush,). These developmental constraints may be fur-ther amplified in contexts where narrative elaboration andverbal emotional expression are less emphasised in every-day communication or educational settings. To our knowledge, only one study has attempted to eval-uate multiple trauma-memory measures within the sameparticipants (Rubin et al.,). Although that study foundno reliable differences in coherence across trauma versusnon-trauma memories or between adults with high versuslow PTSD symptoms, its modest sample size meant thatonly large effects could be detected, and its focus was lim-ited to coherence rather than other theoretically relevantdimensions such as sensory vividness. Moreover, adult find-ings remain mixed overall: while some researchers reportthat disorganisation or fragmentation is elevated amongthose with PTSD (e.g., Halligan et al.,; Salmond et al.,), others have observed null associations (e.g., Jelineket al.,; McGuire et al.,). The present study advances this literature in severalimportant ways. First, we broaden the conceptual scope toinclude both disorganisation and sensory memory vivid-ness, a dimension central to cognitive models (e.g., Brewin,). Second, we test these associations during childhoodand adolescence, periods when autobiographical memorystructure and narrative skills are still developing and mayaffect the detectability of memory-based processes. Third,we draw on substantially larger and culturally diverse sam-ples than previous studies, including youth from a low- andmiddle-income country context, thereby increasing statisti-cal power and evaluating the cross-cultural generalisabilityof narrative-based assumptions. Taken together, our find-ings extend and qualify existing adult-based conclusions: ### Page 12 narrative-derived indicators show not only a lack of clinicalutility, but also poor convergent validity with self-reportedmemory qualities, indicating that, in youth, they may notreliably index the cognitive-memory mechanisms they areintended to represent. In contrast, in our study, self-reported trauma memoryfragmentation and sensory qualities showed strong andconsistent associations with PTSD symptoms in bothcountries. This aligns with prior evidence (Meiser-Sted-man et al.,; McGuire et al.,) and supports theview that children’s subjective appraisal of their memoryis a more proximal marker of clinical risk than objectivenarrative structure. However, the utility of self-reportappears to depend on educational, cultural, and socio-environmental experiences. In the UK sample, self-report trauma memory characteristics were correlatedto PTSD symptoms in both younger children and olderadolescents, whilst in the South African sample, onlyadolescents above 12 years old showed this association.This pattern may reflect contextual differences in howyounger children interpret and respond to introspectivequestionnaire items. For example, differences in educa-tional practices, exposure to self-reflective questioning,and familiarity with verbally articulating internal statesmay influence children’s ability to engage with self-report measures. In line with recommendations to situatepsychological findings within broader structural and his-torical contexts (Buchanan et al.,), these differencesshould not be interpreted as inherent group differencesbut rather as reflecting inequities in education, health-care access, and trauma exposure. Specifically, variationamong samples in linguistic framing, educational con-text, and familiarity with self-reflective reporting formatsmay have shaped how children in each sample appraisedand reported trauma memory characteristics (Van et al.,). Equally, cultural norms around emotional disclo-sure and mental health may explain this pattern. Stud-ies in South Africa describe pervasive stigma, limitedmental-health literacy and substantial barriers to help-seeking among youth and families, including concernsabout judgment, confidentiality and appropriateness ofservices (Mokitimi et al.,; Monnapula-Mazabane &Petersen,; Morris et al.,). Although stigma-related concerns about social evalua-tion typically intensify during adolescence, contemporaryadolescents in urban South African contexts may also beincreasingly empowered to recognise and challenge stigma,creating a more complex interaction between social expec-tations and disclosure. These intersecting developmentaland cultural influences may discourage younger childrenfrom openly describing psychological distress or memoryexperience. Early PTSD risk may therefore be underesti-mated in such groups, underscoring the need for culturallysensitive screening tools. Strengths and Limitations This current study has several strengths. By combining large-scale cohorts from two distinct sociocultural contexts, evalu-ating multiple trauma memory measurement approaches, andspanning both children and adolescents, the research offers novelinsight into the developmental and cultural generalisability oftrauma memory assessment methodologies in new demograph-ics. However, several limitations should be noted. In the SouthAfrican sample, despite rigorous translation procedures, subtledifferences in meaning across translations may have affectedthe validity of narrative-based measures. The cross-languageand cross-cultural context complicates the direct comparabilityof linguistic indices originally developed for English-speakingadults. Second, although the methods and procedures across thethree secondary studies were closely aligned, important con-textual differences remain. For instance, interpersonal violenceand cumulative trauma are more prevalent in South Africanperi-urban low-income communities than in the UK emergencydepartment context (Mitra et al.,; Seedat et al.,), andour focus on the index hospitalisation event under-captured thebroader trauma histories of South African participants. Thosewho have experienced more frequent interpersonal trauma havebeen shown to be more at risk of impairments to autobiographi-cal memory functioning (Borrelli et al.,). For these chil-dren, who may be disproportionately represented in the SouthAfrican sample, unaccounted trauma experiences may haveimpacted trauma memory processing. Future designs wouldbenefit from considering how previous trauma history inter-plays with trauma memory characteristics for indexed traumaevents. Third, the design was cross-sectional with respect toour aims and the primary analyses reported here, limiting ourability to draw conclusions about the prediction of longer-termPTSD trajectories. Longitudinal work is needed to determinewhether the patterns we observe extend to chronic symptomsand treatment response. Finally, we did not collect matched con-trol narratives (e.g., non-traumatic negative or positive events)in this study, preventing us from testing whether any observednarrative features are specific to trauma memories versus auto-biographical memories more broadly (Reed et al.,). Future Work Future work should examine whether more fine-grained,locally focused narrative coding (e.g., focusing on the mostemotionally intense or distressing moments within thetrauma narrative), as recommended in adult work (Brewin, ### Page 13 ), yields stronger links with PTSD than the global indi-ces examined here. Second, research should prioritise thedevelopment and validation of developmentally appropriate,culturally grounded measures of trauma memory qualitiesfor younger children in LMICs. This might include picto-rial response formats, interactive digital tools, and play-based paradigms that reduce literacy demands and relianceon sophisticated metacognitive language, alongside formaltests of measurement invariance and qualitative work toensure conceptual equivalence. This should also considerpotential confounds which may be introduced throughtranslation processes. This would determine whether thelack of validity and predictive utility of narrative-basedmethodologies can be attributed to population-specific fac-tors or to translated isiXhosa. Finally, intervention studies,particularly those evaluating trauma-focused cognitive ther-apies in children exposed to multiple and chronic trauma,should routinely assess trauma memory qualities over timeto clarify whether perceived changes in memory character-istics mediate symptom improvement, and whether suchmediation differs across cultural contexts. Conclusions This study evaluated multiple approaches to assessing traumamemory in children and adolescents following acute trauma,comparing self-report indices with manually and automaticallycoded narratives across youth in the UK and South Africa.Across methods, settings, and age groups, narrative-based indi-ces showed minimal convergence with self-reported traumamemory qualities and failed to predict PTSD. In contrast, chil-dren’s and adolescents’ own ratings of disorganisation and sen-sory vividness were robustly associated with PTSD symptomsin both countries, although clinically informative from earlychildhood in the UK, but only from early adolescence in SouthAfrica. These findings demonstrate that, in youth, the linguisticstructure of trauma narratives is a poor proxy for the cognitivemechanisms emphasised in leading PTSD models. In contrast,subjective appraisals of trauma memories are reliable mark-ers of clinical risk. Practically, brief self-report tools emergeas a scalable, low-burden strategy for early identification ofPTSD, with particular promise for stepped-care triage. At thesame time, the developmental and cross-cultural differencesobserved highlight the need for culturally sensitive, develop-mentally tailored assessment approaches in LMIC contexts.Overall, this work advances theory and practice by clarifyinghow trauma memory should, and should not, be measured inyoung people following trauma. Supplementary InformationThe online version contains supplementarymaterial available at[h t tp s : // d oi . org / 1 0 . 1 0 0 7 / s 1 0 8 0 2 - 0 2 6 - 0 1 4 6 9 - 4](https://doi.org/10.1007/s10802-026-01469-4). Authors’ Contributions AG conceived the study, conducted the analy-ses, and drafted the manuscript. MT, RMS and SLH contributed to thestudy design, theoretical framing, and interpretation of findings. KScontributed to the implementation and extraction of narrative-basedanalytic methods. TS and KS contributed to data preparation andmanuscript feedback. YC and MT contributed to data collection, studycoordination, and contextual interpretation of findings in the SouthAfrican cohort. All authors critically reviewed the manuscript and ap-proved the final version for submission. Funding This research was conducted as part of a doctoral programmesupported by the University of Bath. Funding for the individual stud-ies contributing data to this manuscript was obtained through awardedgrants and institutional funding, as described in detail in the originalpublications for each cohort. Data Availability The data analysed in this study are not publicly avail-able due to ethical restrictions and participant confidentiality agree-ments. Anonymised data may be made available from the correspond-ing author upon reasonable request. Declarations Ethical Approval Ethical approval for the UK studies was obtainedfrom the relevant UK National Health Service (NHS) Research EthicsCommittees and affiliated university ethics boards. Ethical approvalfor the South African study was obtained from the Health ResearchEthics Committee of Stellenbosch University. Written informed con-sent was obtained from parents or legal guardians, and assent was ob-tained from children and adolescents prior to participation. Conflict of interest The authors declare that they have no conflicts ofinterest that are directly or indirectly related to the work submitted forpublication. Open Access This article is licensed under a Creative CommonsAttribution 4.0 International License, which permits use, sharing, adap-tation, distribution and reproduction in any medium or format, as longas you give appropriate credit to the original author(s) and the source,provide a link to the Creative Commons licence, and indicate if changeswere made. The images or other third party material in this article areincluded in the article’s Creative Commons licence, unless indicatedotherwise in a credit line to the material. If material is not included inthe article’s Creative Commons licence and your intended use is notpermitted by statutory regulation or exceeds the permitted use, you willneed to obtain permission directly from the copyright holder. To view acopy of this licence, visit. References Alisic, E., Zalta, A. K., van Wesel, F., Larsen, S. E., Hafstad, G. S.,Hassanpour, K., & Smid, G. E. (2014). Rates of post-traumaticstress disorder in trauma-exposed children and adolescents: Meta-analysis. The British Journal of Psychiatry: The Journal of MentalScience, 204, 335–340. American Psychiatric Association, D (2013). Diagnostic and statisti-cal manual of mental disorders: DSM-5 (Vol. 5, Issue 5). Ameri-can psychiatric association Washington, DC. Atwoli, L., Stein, D. J., Williams, D. R., Mclaughlin, K. A., Petuk-hova, M., Kessler, R. C., & Koenen, K. C. (2013). Trauma andposttraumatic stress disorder in South Africa: Analysis from the ### Page 14 South African Stress and Health Study. Bmc Psychiatry, 13(1),182. Bartal, A., Jagodnik, K. M., Chan, S. J., Babu, M. S., & Dekel, S.(2022). Identifying Women with Post-Delivery PosttraumaticStress Disorder using Natural Language Processing of PersonalChildbirth Narratives. medRxiv: The Preprint Server for HealthSciences, 2022.08.30.22279394. Blasi, D., Anastasopoulos, A., & Neubig, G. (2022). Systematicinequalities in language technology performance across theworld’s languages. In Proceedings of the 60th Annual Meetingof the Association for Computational Linguistics (Volume 1:Long Papers) (pp. 5486–5505) Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet alloca-tion. Journal of Machine Learning Research, 3, 993–1022. Borrelli, G., Lamberti Zanardi, A., Scognamiglio, C., Cinquegrana,V., & Perrella, R. (2024). The relationship between childhoodinterpersonal and non-interpersonal trauma and autobiographi-cal memory: a systematic review. Frontiers in Psychology, 15,1328835. Brewin, C. R. (2014). Episodic memory, perceptual memory, and theirinteraction: Foundations for a theory of posttraumatic stress dis-order. Psychological Bulletin, 140(1), 69–97. Brewin, C. R., Dalgleish, T., & Joseph, S. (1996). A dual representationtheory of posttraumatic stress disorder. Psychological Review,103(4), 670–686. Buchanan, N. T., Perez, M., Prinstein, M. J., & Thurston, I. B. (2021).Upending racism in psychological science: Strategies to changehow science is conducted, reported, reviewed, and disseminated.American Psychologist, 76(7), 1097–1112. Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sci-ences (2nd ed.). Lawrence Erlbaum Associates,. Crossley, S. A., Kyle, K., & McNamara, D. S. (2016). The tool for theautomatic analysis of text cohesion (TAACO): Automatic assess-ment of local, global, and text cohesion. Behavior Research Meth-ods, 48(4), 1227–1237. Ehlers, A., & Clark, D. M. (2000). A cognitive model of posttraumaticstress disorder. Behaviour Research and Therapy, 38(4), 319–345. Faul, F., Erdfelder, E., Buchner, A., & Lang, A. G. (2009). Statisti-cal power analyses using G*Power 3.1: Tests for correlation andregression analyses. Behavior Research Methods, 41(4), 1149–1160. Figueira, I., da Luz, M., Braga, R. J., Cabizuca, M., Coutinho, E., &Mendlowicz, M. V. (2007). The increasing internationalization ofmainstream posttraumatic stress disorder research: A bibliometricstudy. Journal of Traumatic Stress, 20(1), 89–95. Fivush, R. (2011). The Development of Autobiographical Memory.Annual Review of Psychology, 62(1), 559–582. Foa, E. B., Asnaani,Anu, Z., Yinyin , Capaldi, Sandra,and, & Yeh,R. (2018). Psychometrics of the child PTSD symptom scale forDSM-5 for trauma-exposed children and adolescents. Journal ofClinical Child & Adolescent Psychology, 47(1), 38–46. Foa, E. B., Molnar, C., & Cashman, L. (1995). Change in rape narra-tives during exposure therapy for posttraumatic stress disorder.Journal of Traumatic Stress, 8(4), 675–690. Halligan, S. L., Michael, T., Clark, D. M., & Ehlers, A. (2003). Post-traumatic stress disorder following assault: The role of cognitiveprocessing, trauma memory, and appraisals. Journal of Consult-ing and Clinical Psychology, 71(3), 419–431. Hiller, R. M., Meiser-Stedman, R., Elliott, E., Banting, R., & Halli-gan, S. L. (2021). A longitudinal study of cognitive predictors of(complex) post-traumatic stress in young people in out-of-homecare. Journal of Child Psychology and Psychiatry, 62(1), 48–57. Hiller, R. M., Meiser-Stedman, R., Fearon, P., Lobo, S., McKinnon, A.,Fraser, A., & Halligan, S. L. (2016). Research Review: Changesin the prevalence and symptom severity of child post-traumaticstress disorder in the year following trauma - a meta-analyticstudy. Journal of Child Psychology and Psychiatry and AlliedDisciplines, 57(8), 884–898. Hiller, R. M., Meiser-Stedman, R., Lobo, S., Creswell, C., Fearon, P.,Ehlers, A., Murray, L., & Halligan, S. L. (2018). A longitudinalinvestigation of the role of parental responses in predicting chil-dren’s post-traumatic distress. Journal of Child Psychology andPsychiatry, 59(7), 781–789. Hovy, D., & Prabhumoye, S. (2021). Five sources of bias in naturallanguage processing. Language and Linguistics Compass, 15(8),e12432. Jaeger, J., Lindblom, K. M., Parker-Guilbert, K., & Zoellner, L. A.(2014). Trauma Narratives: It’s What You Say, Not How You SayIt. Psychological Trauma: Theory Research Practice and Policy,6(5), 473–481. Jelinek, L., Stockbauer, C., Randjbar, S., Kellner, M., Ehring, T., &Moritz, S. (2010). Characteristics and organization of the worstmoment of trauma memories in posttraumatic stress disorder.Behaviour Research and Therapy, 48(7), 680–685. Joshi, P., Santy, S., Budhiraja, A., Bali, K., & Choudhury, M. (2020).The state and fate of linguistic diversity and inclusion in the NLPworld. In Proceedings of the 58th annual meeting of the associa-tion for computational linguistics (pp. 6282–6293). Khurana, D., Koli, A., Khatter, K., & Singh, S. (2023). Natural lan-guage processing: State of the art, current trends and challenges.Multimedia Tools and Applications, 82(3), 3713–3744. Kleim, B., Horn, A. B., Kraehenmann, R., Mehl, M. R., & Ehlers, A.(2018). Early Linguistic Markers of Trauma-Specific ProcessingPredict Post-trauma Adjustment. Frontiers in Psychiatry, 9, 645. Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction tolatent semantic analysis. Discourse Processes, 25(2–3), 259–284. Lewis, S. J., Arseneault, L., Caspi, A., Fisher, H. L., Matthews, T.,Moffitt, T. E., Odgers, C. L., Stahl, D., Teng, J. Y., & Danese,A. (2019). The epidemiology of trauma and post-traumatic stressdisorder in a representative cohort of young people in Englandand Wales. The Lancet Psychiatry, 6(3), 247–256. Li, H. (2024). Latent Dirichlet Allocation. In H. Li (Ed.), MachineLearning Methods (pp. 439–471). Springer Nature. McElvaney, R., Greene, S., & Hogan, D. (2014). To tell or not to tell?Factors influencing young people’s informal disclosures of childsexual abuse. Journal of Interpersonal Violence, 29(5), 928–947. McGuire, R., Hiller, R. M., Ehlers, A., Fearon, P., Meiser-Stedman,R., Leuteritz, S., & Halligan, S. L. (2021). A Longitudinal Inves-tigation of Children’s Trauma Memory Characteristics and TheirRelationship with Posttraumatic Stress Disorder Symptoms.Research on Child and Adolescent Psychopathology, 49(6), 807–816. ### Page 15 McKinnon, A., Brewer, N., Meiser-Stedman, R., & Nixon, R. D. V.(2017). Trauma memory characteristics and the development ofacute stress disorder and post-traumatic stress disorder in youth.Journal of Behavior Therapy and Experimental Psychiatry, 54,112–119. Meiser-Stedman, R., McKinnon, A., Dixon, C., Boyle, A., Smith, P.,& Dalgleish, T. (2019). A core role for cognitive processes in theacute onset and maintenance of post-traumatic stress in childrenand adolescents. Journal of Child Psychology and Psychiatry,60(8), 875–884. Meiser-Stedman, R., Smith, P., Yule, W., & Dalgleish, T. (2007). TheTrauma Memory Quality Questionnaire: Preliminary develop-ment and validation of a measure of trauma memory characteris-tics for children and adolescents. Memory (Hove England), 15(3),271–279. Miani,A., Hills, T., & Bangerter, A. (2022). Interconnectedness and(in)coherence as a signature of conspiracy worldviews. ScienceAdvances, 8(43), Article eabq3668. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient esti-mation of word representations in vector space. arXiv. Miragoli, S., Camisasca, E., & Di Blasio, P. (2019). Investigating lin-guistic coherence relations in child sexual abuse: A comparison ofPTSD and non-PTSD children. Heliyon, 5(2), e01163. Mitra, A., Okafor, U. B., Kaswa, R., & Adeniyi, O. V. (2022). Epidemi-ology of interpersonal violence at a regional hospital emergencyunit in the Eastern Cape, South Africa. South African FamilyPractice, 64(1), 5511. Mokitimi, S., Jonas, K., Schneider, M., & de Vries, P. J. (2019).Child and Adolescent Mental Health Services in South Africa-Senior Stakeholder Perceptions of Strengths, Weaknesses,Opportunities, and Threats in the Western Cape Province.Frontiers in Psychiatry, 10, 841. Monnapula-Mazabane, P., & Petersen, I. (2023). Mental health stigmaexperiences among caregivers and service users in South Africa: Aqualitative investigation. Current Psychology (New Brunswick N J),42(11), 9427–9439. Morris, M. C., Hellman, N., Abelson, J. L., & Rao, U. (2016). Cor-tisol, heart rate, and blood pressure as early markers of PTSDrisk: A systematic review and meta-analysis. Clinical PsychologyReview, 49, 79–91. Myford, C. M., & Wolfe, E. W. (2009). Monitoring Rater PerformanceOver Time: A Framework for Detecting Differential Accuracyand Differential Scale Category Use. Journal of EducationalMeasurement, 46(4), 371–389. Nelson, K., & Fivush, R. (2004). The Emergence of AutobiographicalMemory: A Social Cultural Developmental Theory. Psychologi-cal Review, 111(2), 486–511. Paivio, A., Yuille, J. C., & Madigan, S. A. (1968). Concreteness, imag-ery, and meaningfulness values for 925 nouns. Journal of Experi-mental Psychology, 76(1, Pt. 2), 1–25 Pennebaker, J., Booth, R., Boyd, R., & Francis, M. (2015). (PDF) Lin-guistic inquiry and word count (LIWC). ResearchGate. Pham, P. N., Weinstein, H. M., & Longman, T. (2004). Trauma andPTSD symptoms in Rwanda: Implications for attitudes toward jus-tice and reconciliation. Journal Of The American Medical Associa-tion, 292(5), 602–612. R Core Team (2023). R: A language and environment for statisticalcomputing (Version 4.3.1). R Foundation for Statistical Comput-ing Reed, J., Meiser-Stedman, R., Dalgleish, T., Goodall, B., Wright, I.,Boyle, A., Burgess, A., Murphy, F., Hitchcock, C., Schweizer,S., Travers-Hill, E., Dixon, C., Mul, C. L., Smith, P., Newby, J.,& McKinnon, A. (2024). Trauma Memory Characteristics andNeurocognitive Performance in Youth Exposed to Single-EventTrauma. Research on Child and Adolescent Psychopathology,52(6), 997–1008. Reed, J., Taylor, J., Randall, G., Burgess, A., & Meiser-Stedman, R.(2023). Associations between the Trauma Memory Quality Ques-tionnaire and posttraumatic stress symptoms in youth: A system-atic review and meta-analysis. Journal of Traumatic Stress, 36(1),31–43. Rubin, D. C., Deffler, S. A., Ogle, C. M., Dowell, N. M., Graesser, A. C., &Beckham, J. C. (2016). Participant, Rater, and Computer Measuresof Coherence in Posttraumatic Stress Disorder. Journal of AbnormalPsychology, 125(1), 11–25. Salmond, C. H., Meiser-Stedman, R., Glucksman, E., Thompson, P.,Dalgleish, T., & Smith, P. (2011). The nature of trauma memoriesin acute stress disorder in children and adolescents. Journal ofChild Psychology and Psychiatry and Allied Disciplines, 52(5),560–570. Schnell, S., & Schiborr, N. (2022). Crosslinguistic Corpus Studies inLinguistic Typology. Annual Review of Linguistics, 8, 171–191. Seedat, S., Nyamai, C., Njenga, F., Vythilingum, B., & Stein, D. J.(2004). Trauma exposure and post-traumatic stress symptoms inurban African schools: Survey in Cape Town and Nairobi. TheBritish Journal of Psychiatry, 184(2), 169–175. Sharp, T. H., Chideya, Y., Giuliani, A., Hunt, X., Tomlinson, M.,Seedat, S., Creswell, C., Fearon, P., Hamilton-Giachritsis, C.,Hiller, R., Meiser-Stedman, R., Toit, D., Stewart, S., J., & Hal-ligan, S. L. (2024). Post-traumatic stress disorder symptoms fol-lowing exposure to acute psychological trauma in children aged8–16 years in South Africa: Protocol for the Sinethemba longi-tudinal study. British Medical Journal Open, 14(7), e085129. Steinberg, A. M., Brymer, M. J., Kim, S., Briggs, E. C., Ippen, C. G.,Ostrowski, S. A., Gully, K. J., & Pynoos, R. S. (2013). Psychomet-ric properties of the UCLA PTSD Reaction Index: Part I. Journalof Traumatic Stress, 26(1), 1–9 Syed, M., & Nelson, S. C. (2015). Guidelines for establishing reli-ability when coding narrative data. Emerging Adulthood, 3(6),375–387. Udeze, N. (2025). Designing culturally inclusive nlp models for low-resource languages and communities. Van, A. S., Swart, E., & Oswald, M. M. (2019). Developing meta-cognition among young learners by using stories. South AfricanJournal of Education, 39(2), 1–12. Zampieri, M., Nakov, P., & Scherrer, Y. (2020). Natural language pro-cessing for similar languages, varieties, and dialects: A survey.Natural Language Engineering, 26(6), 595–612. Zhang, Y. (2022). A Computational Approach to Understanding EarlyTrauma and Paediatric PTSD [Doctoral, University of EastAnglia. Norwich Medical School]. Publisher’s Note Springer Nature remains neutral with regard to juris-dictional claims in published maps and institutional affiliations.