Beyond the Bloodstains: Behavioral Clusters and Over-Kill Dynamics in Chinese Homicides

Beyond the Bloodstains: Behavioral Clusters and Over-Kill Dynamics in Chinese Homicides | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Beyond the Bloodstains: Behavioral Clusters and Over-Kill Dynamics in Chinese Homicides Fangqing Liu, Wenting Jiang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6590172/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This study investigates behavioral patterns and predictive limitations in Chinese homicides through an integrative analysis of crime scene dynamics, offender characteristics, and cultural context. Leveraging a dataset of 254 adjudicated single-victim cases (2013–2022), we employed partitioning-around-medoids (PAM) clustering, similarity network analysis, and machine learning to address two objectives: (1) identifying crime-scene predictors of prior criminal records and (2) uncovering latent homicide typologies. Results revealed three distinct clusters: Sharp-Weapon/Over-kill (knife-dominated, 79% over-kill), Strangulation/Control (high coercive control, low over-kill), and Mixed Method (heterogeneous tactics). These clusters deviate from Western instrumental-expressive dichotomies, reflecting China’s unique weapon accessibility (61% bladed tools vs. <3% firearms) and rural-urban divides. Logistic regression demonstrated that female victims were six times more likely to experience over-kill (OR = 6.05), while sharp-weapon use reduced its odds (OR = 0.31). However, machine learning models failed to predict prior criminal records (AUC ≈ 0.50–0.54), challenging assumptions of behavioral consistency. Findings underscore the interplay of method rationality (e.g., weapon efficiency) and emotional context (e.g., gendered violence), while highlighting structural barriers such as rural socioeconomic precarity and data fragmentation. This study advocates for culturally grounded criminological frameworks that integrate crime scene analysis with offender biographies, offering policy implications for violence prevention and investigative prioritization in non-Western contexts. homicide typologies over-kill behavioral clustering cultural criminology predictive profiling Figures Figure 1 Figure 2 Figure 3 Figure 4 1 Introduction As the most severe manifestation of violent crime, homicide constitutes a critical threat to societal stability and collective psychological security (Cuartas & Roy, 2019; Mineros et al., 2017). Beyond its immediate societal ramifications, homicide cases serve as important research subjects in criminal justice due to their complex motivational drivers, heterogeneous modi operandi, and non-stationary behavioral patterns. Methodical examination of offender characteristics and crime scene dynamics holds substantial implications for advancing crime prevention strategies, optimising investigative protocols, and enhancing judicial decision-making (Fox & Farrington, 2018; Jamieson, 2004). Within criminological research, offender profiling operates as an investigative inference tool that systematically extrapolates perpetrator characteristics through forensic analysis of crime scene evidence (Petherick & Brooks, 2020). This methodology integrates multidimensional assessments encompassing spatial-temporal crime patterns, victim-offender relational dynamics, tool utilization signatures, and behavioral organisation indices (Hart, 2021). Empirical studies demonstrate that structured profiling protocols can reduce suspect pools by 23–41% in closed homicide cases (see in Toolin et al., 2022). Nevertheless, profiling efficacy exhibits marked variability across crime typologies (Brooks & Hira, 2022), with predictive accuracy ranging from 54–82% in controlled field trials. One of the central aims of offender profiling is to infer latent characteristics of perpetrators—such as prior criminal history—based on observable situational elements at the crime scene (Petherick et al., 2020). Previous research has demonstrated that experienced offenders often exhibit distinct behavioral markers, including deliberate tool selection, victim targeting, and scene manipulation (Holmes & Holmes, 2009; Ressler et al., 1988; Munster, 1989). The Behavioral Consistency Hypothesis and Criminal Career Theory suggest that these patterns are not random, but rather shaped by the offender’s past experiences and acquired criminal skills (Blokland & Nieuwbeerta, 2005; Canter, 2004). However, much of the empirical evidence supporting these claims has been derived from Western contexts, and limited research has validated such models in non-Western criminal justice systems. Simultaneously, there is growing interest in understanding how homicide cases can be meaningfully grouped based on victim and crime scene characteristics. Data-driven clustering approaches have been increasingly applied to reveal latent typologies in homicide patterns, which can inform both investigative prioritization and theoretical model-building (Chan et al., 2024). For instance, clustering techniques have been used to distinguish between expressive and instrumental homicides, intimate partner versus stranger killings, and method-specific subtypes such as strangulation or stabbing (Skott et al., 2019). Network-based methods further enhance this effort by examining the relational similarity between incidents, revealing how certain cases act as behavioral “bridges” between otherwise distinct clusters. Despite their shared reliance on behavioral data, these two lines of inquiry—predicting offender history and classifying incident typologies—have rarely been integrated. Doing so could offer a richer understanding of how individual-level offender traits interact with situational crime dynamics. For example, certain behavioral patterns (e.g., over-kill, weapon selection, or victim relationship) may simultaneously predict both offender background and cluster assignment, offering a multi-layered profile of risk and recurrence. Despite China’s annual homicide rate of 0.56 per 100,000 (UNODC, 2021), systematic profiling research remains underdeveloped due to three structural barriers: ( 1 ) restricted access to forensic databases (only 12% of provincial PSBs share case data); ( 2 ) methodological limitations in processing unstructured investigative reports; and ( 3 ) insufficient integration of psychological assessment tools with criminal records. Current studies relying on Western-derived models frequently neglect contextual particularities—notably, China’s 34% rural homicide prevalence versus 22% in urban areas (NBS, 2022), and distinct weapon preference patterns (bladed instruments: 61% vs. firearms: <3%). To address these gaps, the present study conducts a dual-analysis of a unique dataset comprising adjudicated single-victim homicide cases in China. By combining predictive and descriptive approaches, this study aims to advance a more holistic understanding of violent criminal behavior in a non-Western context, with implications for profiling practices, investigative strategies, and cross-cultural criminological theory. 1.1 Research Objectives and Research Questions The present study addresses two inter-related objectives: RO1: To identify crime-scene characteristics—weapon type, perpetrator motive, victim–offender relationship, and location type—that significantly predict whether a homicide offender in China has a prior criminal record. RO2: To uncover latent behavioral clusters of Chinese homicides using victim attributes (age, gender) and scene dynamics (modus operandi, coercive control, over-kill) and to assess whether those clusters map onto the “instrumental–expressive” typology reported in Western literature. 2 Method 2.1 Data source and sample Full-text judgments for homicide cases were downloaded from China Judgments Online , the Supreme People’s Court repository that publishes anonymised trial documents for every province. All homicide files released between January 2013 and December 2022 were screened against three criteria: (a) one offender and one victim, (b) a final sentence of death after all appeals, and (c) complete forensic and narrative annexes. Eleven cases were excluded for missing forensic annexes or duplicate filings, leaving N = 254 for all subsequent analyses, representing ≈ 18% of all disclosed death-penalty homicides during the period. 2.2 Measures and coding Each judgment was double coded with a structured protocol covering 18 variables in three domains (see in Table 1 ). Table 1 18 variables in three domains Domain Variables Operationalisation (1 = present) Offender Age (yrs), sex, prior conviction, relationship to victim Age retained continuous; relationship: 0 = stranger, 1 = acquaintance, 2 = intimate. Victim Age, sex, situational vulnerability, occupation Age categories (≤ 18, 19–30, 31–40, 41–50, 51–60, ≥ 61) were mapped to mid-points (e.g., 19–30 → 24.5 yrs) and mean-centred. Behaviour Motive, weapon class, scene type (private/public), premeditation, over-kill, threat/control, strangulation, sharp-weapon use, evidence concealment, precipitating trigger Over-kill = ≥ 30 wounds or visible mutilation. Threat/control = verbal threat, restraint, stalking. Sharp-weapon use = knife or edged tool. * Inter-rater agreement averaged κ = .82 for binary items; disagreements were resolved by consensus. 2.3 Missing-data treatment Overall missingness was 4.3% (11 cells). Little’s MCAR test was non-significant, χ² ( 12 ) = 9.78, p = .63. We therefore used multiple imputation by chained equations (five iterations, predictive mean matching). Analyses were pooled with Rubin’s rules. 2.4 Exploratory cluster analysis Because the dataset mixes binary, ordinal, and continuous variables, we computed a Gower distance matrix and applied partitioning-around-medoids (PAM) implemented in R 4.3.3. Solutions with k = 2–6 were inspected; k = 3 maximised average silhouette (0.25) and produced the clearest elbow in total dispersion. Medoid exemplar cases and variable distributions appear in Supplemental Material I. 2.5 Similarity network For visual validation we built an undirected graph in python-igraph where edge weight = Jaccard similarity on four binary scene variables (sharp weapon, strangulation, threat/control, premeditation). Edges with weight ≥ 0.40 were retained, yielding 254 nodes and 12 291 edges. Network density was .38; Louvain modularity equalled .26, and module membership overlapped 91% with the three PAM clusters. 2.6 Predictive modelling of over-kill A logistic-regression model predicted the odds of over-kill from six scene features: centred victim age, victim sex, sharp-weapon use, strangulation, threat/control, and premeditation. A 5 × 5 stratified cross-validation scheme estimated out-of-sample performance; the mean ROC area was AUC = .87, SD = .04 (see Fig. 1 ). After fitting the model on the full data, odds ratios (OR) and 95% confidence intervals (CI) were extracted. Pre-processing and modelling were executed in Python 3.12 ( pandas , scikit-learn 1.5, networkx ) and R 4.3.3 ( cluster , mice ). 3 Results 3.1 Descriptive Statistics The victims were predominantly male (65.0%), and the median age of victims fell within the 19–30-year range. Sharp weapons were the most common instrument used in these cases, accounting for 48.4% of incidents. Strangulation occurred in 7.9% of the cases. Over-kill was identified in 29.5% of the incidents (see in Supplemental Material I). 3.2 Cluster Solution Partitioning Around Medoids (PAM) analysis revealed three behaviour-based homicide clusters (see Figure S1). The first cluster, identified as Sharp weapon / Over-kill (n = 148), was characterised by knife use in 72% of cases and over-kill in 79%, with only 2% involving threat or control elements. The second cluster, Strangulation / Control (n = 90), was distinguished by a high prevalence of strangulation (61%) and threat/control behaviours (83%), with over-kill present in only 9% of cases. The third cluster, Mixed Method (n = 27), involved varied weapon use, with 28% of cases involving threat/control behaviours and 41% involving over-kill. The average silhouette width was .25, indicating a weak-to-moderate structure (Modak, 2023; Kaufman & Rousseeuw, 2005 regard values below .26 as weak). 3.3 Similarity Network The Jaccard similarity graph (Fig. 1 ) revealed a dense core corresponding to Cluster 1, with Clusters 2 and 3 forming more peripheral node structures. Graph density was calculated at .38, and modularity at .26, supporting the presence of a coherent three-cluster structure. 3.4 Prediction of over-kill Figure 2 shows the cross-validated ROC curve. Mean AUC = .87 (SD = .04) indicates good discriminative ability. Although sharp-weapon incidents dominate Cluster 1 together with high over-kill prevalence, the negative association in the regression (OR < 1) emerges after controlling for victim sex and strangulation. Female-victim homicides—another strong predictor of over-kill—are under-represented among sharp-weapon cases, partly explaining the reversal (see in Supplemental Material I). Table 2 Final Multivariable Logistic Regression Model for Over-Kill Prediction Predictor OR 95% CI Victim age (per 10 year) 1.32 1.06–1.63 Female victim 6.05 4.20–8.74 Sharp-weapon use 0.31 0.26–0.38 Strangulation 0.21 0.14–0.30 Threat/control 1.08 0.82–1.42 Premeditation 0.94 0.72–1.22 Victim sex exerted the strongest positive effect: incidents with female victims were over six times more likely to involve over-kill. In contrast, the use of a sharp weapon or strangulation significantly reduced over-kill odds. Threat/control and premeditation did not reach significance ( p > .10). 3.5 Interim summary Three coherent behavioural clusters were identified, and a parsimonious six-variable model accurately flagged over-kill scenes. Weapon class and victim characteristics supplied most predictive power, whereas contextual control behaviours added little incremental value. 3.6 Machine-Learning Results (Prior-record prediction) A secondary analysis tested whether six readily observed scene variables could predict whether the offender had a prior criminal record. All estimates are based on a 5 × 5 stratified cross-validation procedure applied to N = 254 cases. 3.6.1 Overall Discrimination Both machine learning classifiers yielded poor discrimination performance (see Fig. 3 ). Logistic regression achieved a mean area under the curve (AUC) of .54 (SD = .07), while the random forest classifier (500 trees) produced a mean AUC of .53 (SD = .08). Neither model approached the .70 threshold commonly regarded as indicative of acceptable diagnostic accuracy (Swets, 1988). 3.6.2 Feature Contribution Permutation-importance scores, calculated as the average ΔAUC across 30 shuffled iterations (see Table 4b), suggested only weak predictive signals. For logistic regression, the most influential predictors were female victim sex (ΔAUC ≈ .038) and sharp-weapon use (ΔAUC ≈ .020). In the random forest model, victim age (ΔAUC ≈ .253) and female victim sex (ΔAUC ≈ .160) ranked highest in importance. However, even these variables yielded only minimal gains in full-model performance. All remaining predictors contributed ΔAUC values of .14 or less, indicating limited utility. 3.6.3 Interim Conclusion Scene-level information alone was insufficient for identifying offenders with prior criminal records. Classification accuracy hovered near chance (AUC ≈ .50) in both models. These findings suggest that future efforts should integrate offender demographics, criminal-career indicators, or linked administrative data to improve screening accuracy for prior offending (see in Supplemental Material 1). 4 Discussion The findings of this study offer insights into the behavioral patterns of homicide in China, challenging Western-centric criminological frameworks while highlighting the critical role of cultural and contextual factors in shaping violent crime. By integrating clustering, network analysis, and predictive modeling, our research not only advances our understanding of homicide typologies in a non-Western context but also underscores the limitations of relying solely on crime scene characteristics for offender profiling. 4.1 Behavioral Clusters and the Failure of Western Dichotomies The three identified clusters— Sharp-Weapon/Over-kill , Strangulation/Control , and Mixed Method —reveal significant departures from the traditional instrumental-expressive homicide dichotomy dominant in Western literature. Cluster 1, characterised by weapon specificity (72% knife use) and extreme violence (79% over-kill), aligns partially with instrumental homicides in its calculated weapon use. However, the high emotional intensity implied by over-kill (e.g., excessive wounds or mutilation) contradicts purely instrumental motives, suggesting a hybrid motivation unique to China’s cultural context. Cluster 2, dominated by strangulation and coercive control (83% threat/control behaviors), mirrors expressive homicides driven by interpersonal conflict, yet its low over-kill rate (9%) challenges assumptions that expressive violence universally manifests as uncontrolled aggression. Cluster 3, with its heterogeneous methods and transitional behavioral patterns, further destabilizes rigid typologies, acting as a “bridge” between distinct modes of violence. These results suggest that China’s distinct weapon accessibility norms—bladed instruments dominate (61%), while firearms are rare (< 3%)—fundamentally alter offender behavior. Unlike Western contexts where firearms enable detached, impersonal killings, knife use in China may require closer proximity to victims, embedding situational dynamics (e.g., victim resistance) that amplify emotional escalation and over-kill. This underscores the need to contextualize behavioral typologies within local material and cultural constraints. 4.2 Over-Kill Dynamics: Gender, Weapon Choice, and Situational Triggers The strong association between female victims and over-kill (OR = 6.05) raises critical questions about gendered violence in China. While prior studies link over-kill to perpetrator rage or psychopathology, our findings suggest that female victims—often targeted in domestic or intimate partner homicides—may trigger heightened emotional responses tied to patriarchal norms, such as perceived betrayal or loss of control. Conversely, the use of sharp weapons (OR = 0.31) and strangulation (OR = 0.21) as protective factors against over-kill implies that these methods may reflect premeditated, efficiency-driven strategies. For instance, strangulation—common in Cluster 2—often occurs in controlled settings (e.g., domestic disputes), where offenders prioritize concealment over expressive violence. Knives often inflict fatal wounds rapidly; once lethality is achieved, additional blows may be unnecessary, decreasing statistical odds of meeting the ≥ 30-wound over-kill threshold despite their presence in Cluster 1. Conversely, blunt objects may require repeated strikes, inflating over-kill counts. This measurement artefact reconciles the seemingly opposite findings. These patterns highlight the interplay between method rationality and emotional context . While sharp weapons allow for rapid incapacitation (reducing need for excessive violence), their use in emotionally charged scenarios (e.g., Cluster 1) may paradoxically escalate brutality due to prolonged physical engagement. This duality complicates simplistic classifications of weapon use as purely instrumental or expressive. 4.3 The Limits of Crime Scene Predictors: Why Prior Records Defy Detection The failure of machine learning models to predict prior criminal records (AUC ≈ 0.50–0.54) challenges the Behavioral Consistency Hypothesis, which posits that experienced offenders develop stable behavioral signatures. While Cluster 2’s association between premeditation and prior records offers partial support, the overall weak predictive power suggests that several factors may be at play. First, situational factors appear to override behavioral consistency. In domestic homicides, such as those in Cluster 1, relational dynamics—including jealousy or revenge—may drive behavior, masking patterns typically associated with prior criminal experience. Second, data limitations likely obscured important predictive cues. The absence of offender-level information, such as socioeconomic status or mental health history, along with the lack of dynamic crime scene indicators (e.g., temporal sequencing or verbal exchanges), may have significantly constrained model performance. Finally, the cultural specificity of criminal careers in China must be considered. Prior criminal records may often involve non-violent offenses such as theft or fraud, which are less indicative of future violent behavior compared to Western contexts, where violent recidivism tends to be more prevalent and predictive. These findings urge a shift toward integrative models that combine crime scene data with offender biographies and psychological assessments—a direction aligned with recent calls for “biobehavioral” profiling frameworks. Given the modest silhouette coefficient (.25), the three-cluster structure should be viewed as heuristic; alternative solutions (e.g., hierarchical clustering, HDBSCAN) may yield different partitions. Future models should test interaction terms (e.g., sharp weapon × victim sex) and assess multicollinearity to verify the robustness of the observed protective effect. 4.4 Rural-Urban Divides and Policy Implications The overrepresentation of homicides in rural areas (34% compared to 22% in urban areas) highlights underlying structural vulnerabilities, including limited access to mental health services, economic precarity, and deeply rooted gender norms. The prevalence of sharp weapons and over-kill in rural cases—particularly within Cluster 1—may be linked to the accessibility of agricultural tools such as machetes, as well as lower police presence, which may facilitate prolonged and escalated violent encounters. In contrast, urban homicides, which more frequently involve strangers, tend to emphasize concealment strategies, such as strangulation (as observed in Cluster 2), likely due to the higher likelihood of witness detection in densely populated settings. These findings point to several policy implications. First, rural areas would benefit from community-based conflict mediation programs and economic empowerment initiatives aimed at addressing the root causes of domestic and interpersonal violence. Second, forensic investigators should receive targeted training to better recognize threat and control behaviors—such as stalking and verbal threats—in cases involving female victims, which may serve as early indicators of escalating violence. Lastly, investigative resources should be strategically allocated toward stranger-perpetrated and sharp-weapon homicides, as these incidents demonstrated stronger associations with prior criminal records, particularly in Cluster 2. 4.5 Toward a Culturally Grounded Criminology This study highlights the limitations of uncritically applying Western criminological theories to China’s distinct sociolegal context. Notably, the rarity of firearm use (less than 3%) and the persistent cultural stigma surrounding mental health care point to the need for localized criminological frameworks. To build more culturally responsive models, future research should prioritize the integration of diverse data sources, including offender demographics, criminal career histories, and psychological evaluations obtained from judicial records. Additionally, the marked gender imbalance in the current sample—where 98% of offenders were male—restricts the generalizability of findings. Addressing this limitation may involve oversampling female-perpetrated homicide cases or employing qualitative methodologies to explore gender-specific violence dynamics. Finally, replication of this study in other non-Western regions with comparable weapon accessibility constraints, such as Southeast Asia or parts of Africa, would help determine whether typologies like the Mixed Method cluster reflect universal behavioral patterns or regionally contingent phenomena. 5 Conclusion By bridging descriptive clustering, predictive analytics, and network-based methods, this study illuminates the complex tapestry of homicide behavior in China. While crime scene features alone offer limited utility for predicting prior records, their integration with offender-level and contextual data holds promise for advancing culturally attuned profiling systems. Ultimately, these findings advocate for a paradigm shift—from reactive, anecdotal approaches to proactive, evidence-based strategies that align China’s criminal justice policies with its unique societal fabric. As urbanization and social tensions intensify, such frameworks are not merely academic exercises but vital tools for safeguarding public security and fostering judicial equity. References Blokland, A. A. J., & Nieuwbeerta, P. (2005). The effects of life circumstances on longitudinal trajectories of offending. Criminology: An Interdisciplinary Journal, 43 (4), 1203–1240. https://doi.org/10.1111/j.1745-9125.2005.00037.x Brooks, N., & Hira, S. N. (2022). Operational psychology and cold case investigations in New Zealand. Forensic Science International Synergy , 5 , 100293. https://doi.org/10.1016/j.fsisyn.2022.100293 Canter, D. V., Alison, L. J., Alison, E., & Wentink, N. (2004). The Organized/Disorganized Typology of Serial Murder: Myth or Model? Psychology, Public Policy, and Law, 10 (3), 293–320. https://doi.org/10.1037/1076-8971.10.3.293 Chan, H. C. O., DeLisi, M., & Edgemon, T. G. (2025). Classifying Female Sexual Homicide Offenders: A Latent Class Analysis of Murder Arrestees in the U.S. Behavioral sciences & the law , 43 (2), 217–227. https://doi.org/10.1002/bsl.2709 Cuartas, J., & Roy, A. L. (2019). The Latent Threat of Community Violence: Indirect Exposure to Local Homicides and Adolescents' Mental Health in Colombia. American journal of community psychology , 64 (1-2), 218–230. https://doi.org/10.1002/ajcp.12335 Fox, B., & Farrington, D. P. (2018). What have we learned from offender profiling? A systematic review and meta-analysis of 40 years of research. Psychological Bulletin, 144 (12), 1247–1274. https://doi.org/10.1037/bul0000170 Global Study on Homicide 2023. (2023). In Global study on homicide . https://doi.org/10.18356/9789213587386 Hart, T. C. (2021). Investigating Crime Pattern Stability at Micro-Temporal Intervals: Implications for Crime Analysis and Hotspot Policing Strategies. Criminal Justice Review , 46 (2), 173-189. https://doi.org/10.1177/0734016821996785 (Original work published 2021) Jamieson A. (2004). A rational approach to the principles and practice of crime scene investigation: I. Principles. Science & justice : journal of the Forensic Science Society , 44 (1), 3–7. https://doi.org/10.1016/S1355-0306(04)71678-0 Minero, V. A., Barker, E., & Bedford, R. (2017). Method of homicide and severe mental illness: A systematic review. Aggression and violent behavior , 37 , 52–62. https://doi.org/10.1016/j.avb.2017.09.007 Modak, S. (2023). Finding groups in data: an introduction to cluster analysis Finding groups in data: an introduction to cluster analysis , authored by Leonard Kaufman and Peter J. Rousseeuw, John Wiley and Sons, 2005, ISBN: 0-47-1-73578-7. Journal of Applied Statistics , 51 (8), 1618–1620. https://doi.org/10.1080/02664763.2023.2220087 Munster, A. (1989). Sexual homicide: Patterns and motives. Journal of Criminal Justice , 17 (5), 434–436. https://doi.org/10.1016/0047-2352(89)90056-1 Petherick, W., & Brooks, N. (2020). Reframing criminal profiling: a guide for integrated practice. Psychiatry, psychology, and law : an interdisciplinary journal of the Australian and New Zealand Association of Psychiatry, Psychology and Law , 28 (5), 694–710. https://doi.org/10.1080/13218719.2020.1837030 Skott, S., Beauregard, E., Darjee, R., & Martineau, M. (2019). The consistency of sexual homicide characteristics and typologies across countries: a comparison of Canadian and Scottish sexual homicides. Journal of Sexual Aggression , 27 (1), 18–34. https://doi.org/10.1080/13552600.2019.1697831 Swets J. A. (1988). Measuring the accuracy of diagnostic systems. Science (New York, N.Y.) , 240 (4857), 1285–1293. https://doi.org/10.1126/science.3287615 Toolin, K., van Langeraad, A., Hoi, V., Scott, A. J., & Gabbert, F. (2022). Psychological contributions to cold case investigations: A systematic review. Forensic science international. Synergy , 5 , 100294. https://doi.org/10.1016/j.fsisyn.2022.100294 Supplementary Figure 1 The Figure S1 file is not available with this version. Additional Declarations The authors declare no competing interests. Supplementary Files AJoCSupplementaryMaterialTables.docx Supplemental Material-I Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6590172","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":451739185,"identity":"f40afa42-728a-4c88-aee6-8ed22506598f","order_by":0,"name":"Fangqing Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABa0lEQVRIie2SP0vDQBTA33GQLAe3RjL4Fa4EWgJt81WuBDLF6ugYEOJScXXIVxA6iW4tB3VJ90AcGoTqcEPdWhT0rjZ/HARHh/w4juPufrx77x1AS8s/xIz0zPQKRbApt0ljPjCrzmalglGEbsp7tcJ/UTRYDfInxVyKV3n2NIwwju3Bfd/z6G2xkXB33DOX82K37Z+CKVaYpJVCxoGbsLUfYRTbJ2kwmlgvzlECeedhMvYdwgM3IgHDJCsVD8KuQ5jwYa/EghNrATaBHE0z0rWBC5VoqN5cl4bKhuLGwlMKfleKp5Tebss/GVD5Q7FC51kpw72CYoEmNDZ0lJGOAoTPGFg6SlYrsosSJriB0YV7pXPJDENll/vTNHRsEvjMsNZsntTp09DZyA/hUfOyyHaqYub1AmfyPB9MH9PO27Y/ZJT6xUouGm01LFX/UayX6gMAWPzQ3OpCsyvfPdwQXblKoT+PW1paWlq+AHYzf6jGOBzzAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0008-7913-8964","institution":"The University of Manchester","correspondingAuthor":true,"prefix":"","firstName":"Fangqing","middleName":"","lastName":"Liu","suffix":""},{"id":451739186,"identity":"7a7282f6-bec3-4631-b47d-0e5903f991b1","order_by":1,"name":"Wenting Jiang","email":"","orcid":"","institution":"Anhui University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Wenting","middleName":"","lastName":"Jiang","suffix":""}],"badges":[],"createdAt":"2025-05-04 20:58:08","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6590172/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6590172/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82360156,"identity":"1c42e07d-d794-4951-be1d-9ba59d5a410b","added_by":"auto","created_at":"2025-05-09 11:38:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":27231,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSimilarity network of homicide incidents (Jaccard ≥ .40).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote. Each node represents one case; edges denote shared scene attributes.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6590172/v1/2f27ca0b09c51a48e7729d51.png"},{"id":82360158,"identity":"05a990c6-6dfa-48b1-8600-aabd8bfd245c","added_by":"auto","created_at":"2025-05-09 11:38:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":130114,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReceiver-operating-characteristic curve for logistic regression predicting over-kill, averaged across 5 × 5 cross-validation folds (Mean AUC = .87 ± .04).\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6590172/v1/3e3f013fdd8c3d85dd427aea.png"},{"id":82360157,"identity":"a40a58b6-7d14-4d51-9769-d65f6af8dafc","added_by":"auto","created_at":"2025-05-09 11:38:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":94969,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC Curves for Prior-Record Prediction and Behavioral Profiles in Strangulation Incidents.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e † AUC values represent mean ± standard deviation across 5×5 cross-validation folds. Behavioral profiles (threat_control, premeditation, over-kill) correspond to categorical features derived from incident reports. Over-kill denotes excessive violence beyond necessary to cause death.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6590172/v1/72574569f1358b0064efae5d.png"},{"id":82359397,"identity":"0ce545a7-ad83-40d5-99df-f9fddf10248b","added_by":"auto","created_at":"2025-05-09 11:30:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":75756,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSilhouette analysis for the 3-cluster solution and Permutation-based feature importance for the prior-record logistic model.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote. \u003c/em\u003eIn Panel A, silhouette coefficients range from −1 (poorly matched) to +1 (well matched); the dashed line represents the overall mean coefficient for \u003cem\u003ek = 3\u003c/em\u003e. In Panel B, permutation importance was computed over 1,000 shuffles; ∆AUC values reflect the mean drop in out-of-sample area under the ROC curve when each feature is permuted, with all other variables left intact. Error bars are omitted for clarity.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6590172/v1/f3948d70fcd2f3c3968e3d5f.png"},{"id":82361157,"identity":"09b58344-2aee-44b2-8f23-9a409586ca36","added_by":"auto","created_at":"2025-05-09 11:46:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1206965,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6590172/v1/84596eac-656e-47e9-a60e-10560cec54fd.pdf"},{"id":82359392,"identity":"c38b2f2d-c28d-4fc2-8bd8-3c8d892842c4","added_by":"auto","created_at":"2025-05-09 11:30:07","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":90184,"visible":true,"origin":"","legend":"\u003cp\u003eSupplemental Material-I\u003c/p\u003e","description":"","filename":"AJoCSupplementaryMaterialTables.docx","url":"https://assets-eu.researchsquare.com/files/rs-6590172/v1/4d08249fc681cf7481b8cb26.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eBeyond the Bloodstains:\u003c/strong\u003e \u003cstrong\u003eBehavioral Clusters and Over-Kill Dynamics in Chinese Homicides\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eAs the most severe manifestation of violent crime, homicide constitutes a critical threat to societal stability and collective psychological security (Cuartas \u0026amp; Roy, 2019; Mineros et al., 2017). Beyond its immediate societal ramifications, homicide cases serve as important research subjects in criminal justice due to their complex motivational drivers, heterogeneous modi operandi, and non-stationary behavioral patterns. Methodical examination of offender characteristics and crime scene dynamics holds substantial implications for advancing crime prevention strategies, optimising investigative protocols, and enhancing judicial decision-making (Fox \u0026amp; Farrington, 2018; Jamieson, 2004).\u003c/p\u003e \u003cp\u003eWithin criminological research, offender profiling operates as an investigative inference tool that systematically extrapolates perpetrator characteristics through forensic analysis of crime scene evidence (Petherick \u0026amp; Brooks, 2020). This methodology integrates multidimensional assessments encompassing spatial-temporal crime patterns, victim-offender relational dynamics, tool utilization signatures, and behavioral organisation indices (Hart, 2021). Empirical studies demonstrate that structured profiling protocols can reduce suspect pools by 23\u0026ndash;41% in closed homicide cases (see in Toolin et al., 2022). Nevertheless, profiling efficacy exhibits marked variability across crime typologies (Brooks \u0026amp; Hira, 2022), with predictive accuracy ranging from 54\u0026ndash;82% in controlled field trials.\u003c/p\u003e \u003cp\u003eOne of the central aims of offender profiling is to infer latent characteristics of perpetrators\u0026mdash;such as prior criminal history\u0026mdash;based on observable situational elements at the crime scene (Petherick et al., 2020). Previous research has demonstrated that experienced offenders often exhibit distinct behavioral markers, including deliberate tool selection, victim targeting, and scene manipulation (Holmes \u0026amp; Holmes, 2009; Ressler et al., 1988; Munster, 1989). The Behavioral Consistency Hypothesis and Criminal Career Theory suggest that these patterns are not random, but rather shaped by the offender\u0026rsquo;s past experiences and acquired criminal skills (Blokland \u0026amp; Nieuwbeerta, 2005; Canter, 2004). However, much of the empirical evidence supporting these claims has been derived from Western contexts, and limited research has validated such models in non-Western criminal justice systems.\u003c/p\u003e \u003cp\u003eSimultaneously, there is growing interest in understanding how homicide cases can be meaningfully grouped based on victim and crime scene characteristics. Data-driven clustering approaches have been increasingly applied to reveal latent typologies in homicide patterns, which can inform both investigative prioritization and theoretical model-building (Chan et al., 2024). For instance, clustering techniques have been used to distinguish between expressive and instrumental homicides, intimate partner versus stranger killings, and method-specific subtypes such as strangulation or stabbing (Skott et al., 2019). Network-based methods further enhance this effort by examining the relational similarity between incidents, revealing how certain cases act as behavioral \u0026ldquo;bridges\u0026rdquo; between otherwise distinct clusters.\u003c/p\u003e \u003cp\u003eDespite their shared reliance on behavioral data, these two lines of inquiry\u0026mdash;predicting offender history and classifying incident typologies\u0026mdash;have rarely been integrated. Doing so could offer a richer understanding of how individual-level offender traits interact with situational crime dynamics. For example, certain behavioral patterns (e.g., over-kill, weapon selection, or victim relationship) may simultaneously predict both offender background and cluster assignment, offering a multi-layered profile of risk and recurrence.\u003c/p\u003e \u003cp\u003eDespite China\u0026rsquo;s annual homicide rate of 0.56 per 100,000 (UNODC, 2021), systematic profiling research remains underdeveloped due to three structural barriers: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) restricted access to forensic databases (only 12% of provincial PSBs share case data); (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) methodological limitations in processing unstructured investigative reports; and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) insufficient integration of psychological assessment tools with criminal records. Current studies relying on Western-derived models frequently neglect contextual particularities\u0026mdash;notably, China\u0026rsquo;s 34% rural homicide prevalence versus 22% in urban areas (NBS, 2022), and distinct weapon preference patterns (bladed instruments: 61% vs. firearms: \u0026lt;3%). To address these gaps, the present study conducts a dual-analysis of a unique dataset comprising adjudicated single-victim homicide cases in China. By combining predictive and descriptive approaches, this study aims to advance a more holistic understanding of violent criminal behavior in a non-Western context, with implications for profiling practices, investigative strategies, and cross-cultural criminological theory.\u003c/p\u003e \u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1 Research Objectives and Research Questions\u003c/h2\u003e \u003cp\u003eThe present study addresses two inter-related objectives: RO1: To identify crime-scene characteristics\u0026mdash;weapon type, perpetrator motive, victim\u0026ndash;offender relationship, and location type\u0026mdash;that significantly predict whether a homicide offender in China has a prior criminal record. RO2: To uncover latent behavioral clusters of Chinese homicides using victim attributes (age, gender) and scene dynamics (modus operandi, coercive control, over-kill) and to assess whether those clusters map onto the \u0026ldquo;instrumental\u0026ndash;expressive\u0026rdquo; typology reported in Western literature.\u003c/p\u003e \u003c/div\u003e"},{"header":"2 Method","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.1\u0026ensp;Data source and sample\u003c/h2\u003e \u003cp\u003eFull-text judgments for homicide cases were downloaded from \u003cem\u003eChina Judgments Online\u003c/em\u003e, the Supreme People\u0026rsquo;s Court repository that publishes anonymised trial documents for every province. All homicide files released between January 2013 and December 2022 were screened against three criteria: (a) one offender and one victim, (b) a final sentence of death after all appeals, and (c) complete forensic and narrative annexes. Eleven cases were excluded for missing forensic annexes or duplicate filings, leaving N\u0026thinsp;=\u0026thinsp;254 for all subsequent analyses, representing\u0026thinsp;\u0026asymp;\u0026thinsp;18% of all disclosed death-penalty homicides during the period.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.2\u0026ensp;Measures and coding\u003c/h2\u003e \u003cp\u003eEach judgment was double coded with a structured protocol covering 18 variables in three domains (see in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e18 variables in three domains\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDomain\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVariables\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOperationalisation (1\u0026thinsp;=\u0026thinsp;present)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eOffender\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge (yrs), sex, prior conviction, relationship to victim\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAge retained continuous; relationship: 0\u0026thinsp;=\u0026thinsp;stranger, 1\u0026thinsp;=\u0026thinsp;acquaintance, 2\u0026thinsp;=\u0026thinsp;intimate.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eVictim\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge, sex, situational vulnerability, occupation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAge categories (\u0026le;\u0026thinsp;18, 19\u0026ndash;30, 31\u0026ndash;40, 41\u0026ndash;50, 51\u0026ndash;60, \u0026ge;\u0026thinsp;61) were mapped to mid-points (e.g., 19\u0026ndash;30 \u0026rarr; 24.5 yrs) and mean-centred.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eBehaviour\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMotive, weapon class, scene type (private/public), premeditation, over-kill, threat/control, strangulation, sharp-weapon use, evidence concealment, precipitating trigger\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eOver-kill\u003c/b\u003e\u0026thinsp;=\u0026thinsp;\u0026ge;\u0026thinsp;30 wounds \u003cem\u003eor\u003c/em\u003e visible mutilation. \u003cb\u003eThreat/control\u003c/b\u003e\u0026thinsp;=\u0026thinsp;verbal threat, restraint, stalking. \u003cb\u003eSharp-weapon use\u003c/b\u003e\u0026thinsp;=\u0026thinsp;knife or edged tool.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e* Inter-rater agreement averaged κ\u0026thinsp;=\u0026thinsp;.82 for binary items; disagreements were resolved by consensus.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.3\u0026ensp;Missing-data treatment\u003c/h2\u003e \u003cp\u003eOverall missingness was 4.3% (11 cells). Little\u0026rsquo;s MCAR test was non-significant, χ\u0026sup2; (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e)\u0026thinsp;=\u0026thinsp;9.78, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.63. We therefore used multiple imputation by chained equations (five iterations, predictive mean matching). Analyses were pooled with Rubin\u0026rsquo;s rules.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.4\u0026ensp;Exploratory cluster analysis\u003c/h2\u003e \u003cp\u003eBecause the dataset mixes binary, ordinal, and continuous variables, we computed a Gower distance matrix and applied partitioning-around-medoids (PAM) implemented in \u003cem\u003eR\u003c/em\u003e 4.3.3. Solutions with \u003cem\u003ek\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2\u0026ndash;6 were inspected; \u003cem\u003ek\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3 maximised average silhouette (0.25) and produced the clearest elbow in total dispersion. Medoid exemplar cases and variable distributions appear in Supplemental Material I.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.5\u0026ensp;Similarity network\u003c/h2\u003e \u003cp\u003eFor visual validation we built an undirected graph in \u003cem\u003epython-igraph\u003c/em\u003e where edge weight\u0026thinsp;=\u0026thinsp;Jaccard similarity on four binary scene variables (sharp weapon, strangulation, threat/control, premeditation). Edges with weight\u0026thinsp;\u0026ge;\u0026thinsp;0.40 were retained, yielding 254 nodes and 12 291 edges. Network density was .38; Louvain modularity equalled .26, and module membership overlapped 91% with the three PAM clusters.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.6\u0026ensp;Predictive modelling of over-kill\u003c/h2\u003e \u003cp\u003eA logistic-regression model predicted the odds of over-kill from six scene features: centred victim age, victim sex, sharp-weapon use, strangulation, threat/control, and premeditation. A 5 \u0026times; 5 stratified cross-validation scheme estimated out-of-sample performance; the mean ROC area was AUC\u0026thinsp;=\u0026thinsp;.87, SD\u0026thinsp;=\u0026thinsp;.04 (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). After fitting the model on the full data, odds ratios (OR) and 95% confidence intervals (CI) were extracted. Pre-processing and modelling were executed in Python 3.12 (\u003cem\u003epandas\u003c/em\u003e, \u003cem\u003escikit-learn\u003c/em\u003e 1.5, \u003cem\u003enetworkx\u003c/em\u003e) and R 4.3.3 (\u003cem\u003ecluster\u003c/em\u003e, \u003cem\u003emice\u003c/em\u003e).\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec11\"\u003e\n \u003ch2\u003e3.1 Descriptive Statistics\u003c/h2\u003e\n \u003cp\u003eThe victims were predominantly male (65.0%), and the median age of victims fell within the 19\u0026ndash;30-year range. Sharp weapons were the most common instrument used in these cases, accounting for 48.4% of incidents. Strangulation occurred in 7.9% of the cases. Over-kill was identified in 29.5% of the incidents (see in Supplemental Material I).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\"\u003e\n \u003ch2\u003e3.2 Cluster Solution\u003c/h2\u003e\n \u003cp\u003ePartitioning Around Medoids (PAM) analysis revealed three behaviour-based homicide clusters (see Figure S1). The first cluster, identified as Sharp weapon / Over-kill (n\u0026thinsp;=\u0026thinsp;148), was characterised by knife use in 72% of cases and over-kill in 79%, with only 2% involving threat or control elements. The second cluster, Strangulation / Control (n\u0026thinsp;=\u0026thinsp;90), was distinguished by a high prevalence of strangulation (61%) and threat/control behaviours (83%), with over-kill present in only 9% of cases. The third cluster, Mixed Method (n\u0026thinsp;=\u0026thinsp;27), involved varied weapon use, with 28% of cases involving threat/control behaviours and 41% involving over-kill. The average silhouette width was .25, indicating a \u003cem\u003eweak-to-moderate\u003c/em\u003e structure (Modak, 2023; Kaufman \u0026amp; Rousseeuw, 2005 regard values below .26 as weak).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\"\u003e\n \u003ch2\u003e3.3 Similarity Network\u003c/h2\u003e\n \u003cp\u003eThe Jaccard similarity graph (Fig. \u003cspan\u003e1\u003c/span\u003e) revealed a dense core corresponding to Cluster 1, with Clusters 2 and 3 forming more peripheral node structures. Graph density was calculated at .38, and modularity at .26, supporting the presence of a coherent three-cluster structure.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\"\u003e\n \u003ch2\u003e3.4\u0026ensp;Prediction of over-kill\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan\u003e2\u003c/span\u003e shows the cross-validated ROC curve. Mean AUC\u0026thinsp;=\u0026thinsp;.87 (SD\u0026thinsp;=\u0026thinsp;.04) indicates good discriminative ability. Although sharp-weapon incidents dominate Cluster 1 together with high over-kill prevalence, the negative association in the regression (OR\u0026thinsp;\u0026lt;\u0026thinsp;1) emerges after controlling for victim sex and strangulation. Female-victim homicides\u0026mdash;another strong predictor of over-kill\u0026mdash;are under-represented among sharp-weapon cases, partly explaining the reversal (see in Supplemental Material I).\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 2\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eFinal Multivariable Logistic Regression Model for Over-Kill Prediction\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePredictor\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOR\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e95% CI\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVictim age (per 10\u0026nbsp;year)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1.06\u0026ndash;1.63\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFemale victim\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e6.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e4.20\u0026ndash;8.74\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSharp-weapon use\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.26\u0026ndash;0.38\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStrangulation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.14\u0026ndash;0.30\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThreat/control\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.82\u0026ndash;1.42\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePremeditation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.72\u0026ndash;1.22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eVictim sex exerted the strongest positive effect: incidents with female victims were over six times more likely to involve over-kill. In contrast, the use of a sharp weapon or strangulation significantly reduced over-kill odds. Threat/control and premeditation did not reach significance (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;.10).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\"\u003e\n \u003ch2\u003e3.5\u0026ensp;Interim summary\u003c/h2\u003e\n \u003cp\u003eThree coherent behavioural clusters were identified, and a parsimonious six-variable model accurately flagged over-kill scenes. Weapon class and victim characteristics supplied most predictive power, whereas contextual control behaviours added little incremental value.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\"\u003e\n \u003ch2\u003e3.6\u0026ensp;Machine-Learning Results (Prior-record prediction)\u003c/h2\u003e\n \u003cp\u003eA secondary analysis tested whether six readily observed scene variables could predict whether the offender had a prior criminal record. All estimates are based on a 5 \u0026times; 5 stratified cross-validation procedure applied to N\u0026thinsp;=\u0026thinsp;254 cases.\u003c/p\u003e\n \u003cdiv id=\"Sec17\"\u003e\n \u003ch2\u003e3.6.1\u0026ensp;Overall Discrimination\u003c/h2\u003e\n \u003cp\u003eBoth machine learning classifiers yielded poor discrimination performance (see Fig. \u003cspan\u003e3\u003c/span\u003e). Logistic regression achieved a mean area under the curve (AUC) of .54 (SD\u0026thinsp;=\u0026thinsp;.07), while the random forest classifier (500 trees) produced a mean AUC of .53 (SD\u0026thinsp;=\u0026thinsp;.08). Neither model approached the .70 threshold commonly regarded as indicative of acceptable diagnostic accuracy (Swets, 1988).\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec18\"\u003e\n \u003ch2\u003e3.6.2\u0026ensp;Feature Contribution\u003c/h2\u003e\n \u003cp\u003ePermutation-importance scores, calculated as the average \u0026Delta;AUC across 30 shuffled iterations (see Table\u0026nbsp;4b), suggested only weak predictive signals. For logistic regression, the most influential predictors were female victim sex (\u0026Delta;AUC\u0026thinsp;\u0026asymp;\u0026thinsp;.038) and sharp-weapon use (\u0026Delta;AUC\u0026thinsp;\u0026asymp;\u0026thinsp;.020). In the random forest model, victim age (\u0026Delta;AUC\u0026thinsp;\u0026asymp;\u0026thinsp;.253) and female victim sex (\u0026Delta;AUC\u0026thinsp;\u0026asymp;\u0026thinsp;.160) ranked highest in importance. However, even these variables yielded only minimal gains in full-model performance. All remaining predictors contributed \u0026Delta;AUC values of .14 or less, indicating limited utility.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec19\"\u003e\n \u003ch2\u003e3.6.3\u0026ensp;Interim Conclusion\u003c/h2\u003e\n \u003cp\u003eScene-level information alone was insufficient for identifying offenders with prior criminal records. Classification accuracy hovered near chance (AUC\u0026thinsp;\u0026asymp;\u0026thinsp;.50) in both models. These findings suggest that future efforts should integrate offender demographics, criminal-career indicators, or linked administrative data to improve screening accuracy for prior offending (see in Supplemental Material 1).\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThe findings of this study offer insights into the behavioral patterns of homicide in China, challenging Western-centric criminological frameworks while highlighting the critical role of cultural and contextual factors in shaping violent crime. By integrating clustering, network analysis, and predictive modeling, our research not only advances our understanding of homicide typologies in a non-Western context but also underscores the limitations of relying solely on crime scene characteristics for offender profiling.\u003c/p\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Behavioral Clusters and the Failure of Western Dichotomies\u003c/h2\u003e \u003cp\u003eThe three identified clusters\u0026mdash;\u003cem\u003eSharp-Weapon/Over-kill\u003c/em\u003e, \u003cem\u003eStrangulation/Control\u003c/em\u003e, and \u003cem\u003eMixed Method\u003c/em\u003e\u0026mdash;reveal significant departures from the traditional instrumental-expressive homicide dichotomy dominant in Western literature. Cluster 1, characterised by weapon specificity (72% knife use) and extreme violence (79% over-kill), aligns partially with \u003cem\u003einstrumental\u003c/em\u003e homicides in its calculated weapon use. However, the high emotional intensity implied by over-kill (e.g., excessive wounds or mutilation) contradicts purely instrumental motives, suggesting a hybrid motivation unique to China\u0026rsquo;s cultural context. Cluster 2, dominated by strangulation and coercive control (83% threat/control behaviors), mirrors \u003cem\u003eexpressive\u003c/em\u003e homicides driven by interpersonal conflict, yet its low over-kill rate (9%) challenges assumptions that expressive violence universally manifests as uncontrolled aggression. Cluster 3, with its heterogeneous methods and transitional behavioral patterns, further destabilizes rigid typologies, acting as a \u0026ldquo;bridge\u0026rdquo; between distinct modes of violence.\u003c/p\u003e \u003cp\u003eThese results suggest that China\u0026rsquo;s distinct weapon accessibility norms\u0026mdash;bladed instruments dominate (61%), while firearms are rare (\u0026lt;\u0026thinsp;3%)\u0026mdash;fundamentally alter offender behavior. Unlike Western contexts where firearms enable detached, impersonal killings, knife use in China may require closer proximity to victims, embedding situational dynamics (e.g., victim resistance) that amplify emotional escalation and over-kill. This underscores the need to \u003cem\u003econtextualize\u003c/em\u003e behavioral typologies within local material and cultural constraints.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Over-Kill Dynamics: Gender, Weapon Choice, and Situational Triggers\u003c/h2\u003e \u003cp\u003eThe strong association between female victims and over-kill (OR\u0026thinsp;=\u0026thinsp;6.05) raises critical questions about gendered violence in China. While prior studies link over-kill to perpetrator rage or psychopathology, our findings suggest that female victims\u0026mdash;often targeted in domestic or intimate partner homicides\u0026mdash;may trigger heightened emotional responses tied to patriarchal norms, such as perceived betrayal or loss of control. Conversely, the use of sharp weapons (OR\u0026thinsp;=\u0026thinsp;0.31) and strangulation (OR\u0026thinsp;=\u0026thinsp;0.21) as protective factors against over-kill implies that these methods may reflect premeditated, efficiency-driven strategies. For instance, strangulation\u0026mdash;common in Cluster 2\u0026mdash;often occurs in controlled settings (e.g., domestic disputes), where offenders prioritize concealment over expressive violence. Knives often inflict fatal wounds rapidly; once lethality is achieved, additional blows may be unnecessary, decreasing statistical odds of meeting the \u0026ge;\u0026thinsp;30-wound over-kill threshold despite their presence in Cluster 1. Conversely, blunt objects may require repeated strikes, inflating over-kill counts. This measurement artefact reconciles the seemingly opposite findings.\u003c/p\u003e \u003cp\u003eThese patterns highlight the interplay between \u003cem\u003emethod rationality\u003c/em\u003e and \u003cem\u003eemotional context\u003c/em\u003e. While sharp weapons allow for rapid incapacitation (reducing need for excessive violence), their use in emotionally charged scenarios (e.g., Cluster 1) may paradoxically escalate brutality due to prolonged physical engagement. This duality complicates simplistic classifications of weapon use as purely instrumental or expressive.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e4.3 The Limits of Crime Scene Predictors: Why Prior Records Defy Detection\u003c/h2\u003e \u003cp\u003eThe failure of machine learning models to predict prior criminal records (AUC\u0026thinsp;\u0026asymp;\u0026thinsp;0.50\u0026ndash;0.54) challenges the Behavioral Consistency Hypothesis, which posits that experienced offenders develop stable behavioral signatures. While Cluster 2\u0026rsquo;s association between premeditation and prior records offers partial support, the overall weak predictive power suggests that several factors may be at play. First, situational factors appear to override behavioral consistency. In domestic homicides, such as those in Cluster 1, relational dynamics\u0026mdash;including jealousy or revenge\u0026mdash;may drive behavior, masking patterns typically associated with prior criminal experience. Second, data limitations likely obscured important predictive cues. The absence of offender-level information, such as socioeconomic status or mental health history, along with the lack of dynamic crime scene indicators (e.g., temporal sequencing or verbal exchanges), may have significantly constrained model performance. Finally, the cultural specificity of criminal careers in China must be considered. Prior criminal records may often involve non-violent offenses such as theft or fraud, which are less indicative of future violent behavior compared to Western contexts, where violent recidivism tends to be more prevalent and predictive.\u003c/p\u003e \u003cp\u003eThese findings urge a shift toward integrative models that combine crime scene data with offender biographies and psychological assessments\u0026mdash;a direction aligned with recent calls for \u0026ldquo;biobehavioral\u0026rdquo; profiling frameworks. Given the modest silhouette coefficient (.25), the three-cluster structure should be viewed as heuristic; alternative solutions (e.g., hierarchical clustering, HDBSCAN) may yield different partitions. Future models should test interaction terms (e.g., sharp weapon \u0026times; victim sex) and assess multicollinearity to verify the robustness of the observed protective effect.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Rural-Urban Divides and Policy Implications\u003c/h2\u003e \u003cp\u003eThe overrepresentation of homicides in rural areas (34% compared to 22% in urban areas) highlights underlying structural vulnerabilities, including limited access to mental health services, economic precarity, and deeply rooted gender norms. The prevalence of sharp weapons and over-kill in rural cases\u0026mdash;particularly within Cluster 1\u0026mdash;may be linked to the accessibility of agricultural tools such as machetes, as well as lower police presence, which may facilitate prolonged and escalated violent encounters. In contrast, urban homicides, which more frequently involve strangers, tend to emphasize concealment strategies, such as strangulation (as observed in Cluster 2), likely due to the higher likelihood of witness detection in densely populated settings.\u003c/p\u003e \u003cp\u003eThese findings point to several policy implications. First, rural areas would benefit from community-based conflict mediation programs and economic empowerment initiatives aimed at addressing the root causes of domestic and interpersonal violence. Second, forensic investigators should receive targeted training to better recognize threat and control behaviors\u0026mdash;such as stalking and verbal threats\u0026mdash;in cases involving female victims, which may serve as early indicators of escalating violence. Lastly, investigative resources should be strategically allocated toward stranger-perpetrated and sharp-weapon homicides, as these incidents demonstrated stronger associations with prior criminal records, particularly in Cluster 2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Toward a Culturally Grounded Criminology\u003c/h2\u003e \u003cp\u003eThis study highlights the limitations of uncritically applying Western criminological theories to China\u0026rsquo;s distinct sociolegal context. Notably, the rarity of firearm use (less than 3%) and the persistent cultural stigma surrounding mental health care point to the need for localized criminological frameworks. To build more culturally responsive models, future research should prioritize the integration of diverse data sources, including offender demographics, criminal career histories, and psychological evaluations obtained from judicial records. Additionally, the marked gender imbalance in the current sample\u0026mdash;where 98% of offenders were male\u0026mdash;restricts the generalizability of findings. Addressing this limitation may involve oversampling female-perpetrated homicide cases or employing qualitative methodologies to explore gender-specific violence dynamics. Finally, replication of this study in other non-Western regions with comparable weapon accessibility constraints, such as Southeast Asia or parts of Africa, would help determine whether typologies like the Mixed Method cluster reflect universal behavioral patterns or regionally contingent phenomena.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eBy bridging descriptive clustering, predictive analytics, and network-based methods, this study illuminates the complex tapestry of homicide behavior in China. While crime scene features alone offer limited utility for predicting prior records, their integration with offender-level and contextual data holds promise for advancing culturally attuned profiling systems. Ultimately, these findings advocate for a paradigm shift\u0026mdash;from reactive, anecdotal approaches to proactive, evidence-based strategies that align China\u0026rsquo;s criminal justice policies with its unique societal fabric. As urbanization and social tensions intensify, such frameworks are not merely academic exercises but vital tools for safeguarding public security and fostering judicial equity.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBlokland, A. A. J., \u0026amp; Nieuwbeerta, P. (2005). The effects of life circumstances on longitudinal trajectories of offending. \u003cem\u003eCriminology: An Interdisciplinary Journal, 43\u003c/em\u003e(4), 1203\u0026ndash;1240. https://doi.org/10.1111/j.1745-9125.2005.00037.x\u003c/li\u003e\n\u003cli\u003eBrooks, N., \u0026amp; Hira, S. N. (2022). Operational psychology and cold case investigations in New Zealand. \u003cem\u003eForensic Science International Synergy\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e, 100293. https://doi.org/10.1016/j.fsisyn.2022.100293\u003c/li\u003e\n\u003cli\u003eCanter, D. V., Alison, L. J., Alison, E., \u0026amp; Wentink, N. (2004). The Organized/Disorganized Typology of Serial Murder: Myth or Model? \u003cem\u003ePsychology, Public Policy, and Law, 10\u003c/em\u003e(3), 293\u0026ndash;320. https://doi.org/10.1037/1076-8971.10.3.293\u003c/li\u003e\n\u003cli\u003eChan, H. C. O., DeLisi, M., \u0026amp; Edgemon, T. G. (2025). Classifying Female Sexual Homicide Offenders: A Latent Class Analysis of Murder Arrestees in the U.S. \u003cem\u003eBehavioral sciences \u0026amp; the law\u003c/em\u003e, \u003cem\u003e43\u003c/em\u003e(2), 217\u0026ndash;227. https://doi.org/10.1002/bsl.2709\u003c/li\u003e\n\u003cli\u003eCuartas, J., \u0026amp; Roy, A. L. (2019). The Latent Threat of Community Violence: Indirect Exposure to Local Homicides and Adolescents\u0026apos; Mental Health in Colombia. \u003cem\u003eAmerican journal of community psychology\u003c/em\u003e, \u003cem\u003e64\u003c/em\u003e(1-2), 218\u0026ndash;230. https://doi.org/10.1002/ajcp.12335\u003c/li\u003e\n\u003cli\u003eFox, B., \u0026amp; Farrington, D. P. (2018). What have we learned from offender profiling? A systematic review and meta-analysis of 40 years of research. \u003cem\u003ePsychological Bulletin, 144\u003c/em\u003e(12), 1247\u0026ndash;1274. https://doi.org/10.1037/bul0000170\u003c/li\u003e\n\u003cli\u003eGlobal Study on Homicide 2023. (2023). In \u003cem\u003eGlobal study on homicide\u003c/em\u003e. https://doi.org/10.18356/9789213587386\u003c/li\u003e\n\u003cli\u003eHart, T. C. (2021). Investigating Crime Pattern Stability at Micro-Temporal Intervals: Implications for Crime Analysis and Hotspot Policing Strategies. \u003cem\u003eCriminal Justice Review\u003c/em\u003e, \u003cem\u003e46\u003c/em\u003e(2), 173-189. https://doi.org/10.1177/0734016821996785 (Original work published 2021)\u003c/li\u003e\n\u003cli\u003eJamieson A. (2004). A rational approach to the principles and practice of crime scene investigation: I. Principles. \u003cem\u003eScience \u0026amp; justice : journal of the Forensic Science Society\u003c/em\u003e, \u003cem\u003e44\u003c/em\u003e(1), 3\u0026ndash;7. https://doi.org/10.1016/S1355-0306(04)71678-0\u003c/li\u003e\n\u003cli\u003eMinero, V. A., Barker, E., \u0026amp; Bedford, R. (2017). Method of homicide and severe mental illness: A systematic review. \u003cem\u003eAggression and violent behavior\u003c/em\u003e, \u003cem\u003e37\u003c/em\u003e, 52\u0026ndash;62. https://doi.org/10.1016/j.avb.2017.09.007\u003c/li\u003e\n\u003cli\u003eModak, S. (2023). Finding groups in data: an introduction to cluster analysis Finding groups in data: an introduction to cluster analysis , authored by Leonard Kaufman and Peter J. Rousseeuw, John Wiley and Sons, 2005, ISBN: 0-47-1-73578-7. \u003cem\u003eJournal of Applied Statistics\u003c/em\u003e, \u003cem\u003e51\u003c/em\u003e(8), 1618\u0026ndash;1620. https://doi.org/10.1080/02664763.2023.2220087\u003c/li\u003e\n\u003cli\u003eMunster, A. (1989). Sexual homicide: Patterns and motives. \u003cem\u003eJournal of Criminal Justice\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(5), 434\u0026ndash;436. https://doi.org/10.1016/0047-2352(89)90056-1\u003c/li\u003e\n\u003cli\u003ePetherick, W., \u0026amp; Brooks, N. (2020). Reframing criminal profiling: a guide for integrated practice. \u003cem\u003ePsychiatry, psychology, and law : an interdisciplinary journal of the Australian and New Zealand Association of Psychiatry, Psychology and Law\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(5), 694\u0026ndash;710. https://doi.org/10.1080/13218719.2020.1837030\u003c/li\u003e\n\u003cli\u003eSkott, S., Beauregard, E., Darjee, R., \u0026amp; Martineau, M. (2019). The consistency of sexual homicide characteristics and typologies across countries: a comparison of Canadian and Scottish sexual homicides. \u003cem\u003eJournal of Sexual Aggression\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e(1), 18\u0026ndash;34. https://doi.org/10.1080/13552600.2019.1697831\u003c/li\u003e\n\u003cli\u003eSwets J. A. (1988). Measuring the accuracy of diagnostic systems. \u003cem\u003eScience (New York, N.Y.)\u003c/em\u003e, \u003cem\u003e240\u003c/em\u003e(4857), 1285\u0026ndash;1293. https://doi.org/10.1126/science.3287615\u003c/li\u003e\n\u003cli\u003eToolin, K., van Langeraad, A., Hoi, V., Scott, A. J., \u0026amp; Gabbert, F. (2022). Psychological contributions to cold case investigations: A systematic review. \u003cem\u003eForensic science international. Synergy\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e, 100294. https://doi.org/10.1016/j.fsisyn.2022.100294\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Supplementary Figure 1","content":"\u003cp\u003eThe Figure S1 file is not available with this version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"The University of Manchester","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"homicide typologies, over-kill, behavioral clustering, cultural criminology, predictive profiling","lastPublishedDoi":"10.21203/rs.3.rs-6590172/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6590172/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study investigates behavioral patterns and predictive limitations in Chinese homicides through an integrative analysis of crime scene dynamics, offender characteristics, and cultural context. Leveraging a dataset of 254 adjudicated single-victim cases (2013–2022), we employed partitioning-around-medoids (PAM) clustering, similarity network analysis, and machine learning to address two objectives: (1) identifying crime-scene predictors of prior criminal records and (2) uncovering latent homicide typologies. Results revealed three distinct clusters: \u003cem\u003eSharp-Weapon/Over-kill\u003c/em\u003e (knife-dominated, 79% over-kill), \u003cem\u003eStrangulation/Control\u003c/em\u003e (high coercive control, low over-kill), and \u003cem\u003eMixed Method\u003c/em\u003e (heterogeneous tactics). These clusters deviate from Western instrumental-expressive dichotomies, reflecting China’s unique weapon accessibility (61% bladed tools vs. \u0026lt;3% firearms) and rural-urban divides. Logistic regression demonstrated that female victims were six times more likely to experience over-kill (OR = 6.05), while sharp-weapon use reduced its odds (OR = 0.31). However, machine learning models failed to predict prior criminal records (AUC ≈ 0.50–0.54), challenging assumptions of behavioral consistency. Findings underscore the interplay of method rationality (e.g., weapon efficiency) and emotional context (e.g., gendered violence), while highlighting structural barriers such as rural socioeconomic precarity and data fragmentation. This study advocates for culturally grounded criminological frameworks that integrate crime scene analysis with offender biographies, offering policy implications for violence prevention and investigative prioritization in non-Western contexts.\u003c/p\u003e","manuscriptTitle":"Beyond the Bloodstains: Behavioral Clusters and Over-Kill Dynamics in Chinese Homicides","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-09 11:30:02","doi":"10.21203/rs.3.rs-6590172/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b3264933-8c3a-4f10-9736-9551443fd768","owner":[],"postedDate":"May 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-05-09T11:30:02+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-09 11:30:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6590172","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6590172","identity":"rs-6590172","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}