Intimate Partner Homicide Among Women of Childbearing Age: Identifying Multilevel Risk Factors with Machine Learning

Abstract

Intimate partner homicide (IPH) remains a major yet understudied cause of maternal mortality among U.S. women of childbearing age (WCBA). We leveraged the National Violent Death Reporting System (NVDRS) and county-level Maternal Vulnerability Index (MVI) data from 2018–2022 to train three machine learning models—logistic regression, random forest, and XGBoost—to classify whether homicides were IPH. Among 11,498 homicides involving WCBA, 33% were IPH. XGBoost achieved the best performance (F1-score = 0.83, AUPRC = 0.87), prompting further examination of key predictors via model explainability. Results indicated that acute interpersonal conflicts (e.g., arguments, jealousy), prior IPV victimization, and structural vulnerabilities (e.g., reproductive healthcare access, physical environment) were influential predictors of IPH. By illustrating the interplay of individual, interpersonal, and broader community-level risk factors, our study shows how machine learning can inform multilevel strategies to prevent IPH and improve maternal health.

Introduction

The United States (U.S.) is facing a maternal mortality crisis, with the highest maternal mortality rates among high-income nations.¹ While leading causes include mental health conditions (e.g., suicide), hemorrhage, and cardiac and coronary conditions,² intimate partner homicide (IPH)—homicide perpetrated by a current or former spouse or dating partner³—is a significant, yet understudied contributor.^4,5 In 2021, the U.S. Federal Bureau of Investigation estimated that 4,970 women were murdered;⁶ alarmingly, one-third to one-half of female homicide victims in the U.S. are killed by their intimate partner.^6,7 Women of childbearing age (WCBA), typically women aged 15-49, are at the highest risk of intimate partner violence (IPV),⁸ with two-thirds of cases that occur before pregnancy persisting into pregnancy.⁹

Multilevel determinants of health and their complex interactions shape women’s risk of IPH. Guided by the socioecological model,^10,11 research has identified various risk and protective factors for IPV across four levels: individual (e.g., history of violent victimization and substance use), relationship (e.g., social isolation and family conflict), community (e.g., neighborhood poverty and high unemployment rates), and societal (e.g., gender inequality and weak health, educational, economic, and social policies/laws).^12–14 Societal factors shape the distribution and effects of community, relationship, and individual factors on IPV through women’s socioeconomic positions, which determine their access to power, prestige, and resources.¹⁵ As a result, marginalized women, including those who are Black or rural residents, experience disproportionately higher risks of IPH and suffer from severe physical and psychosocial consequences of IPV.^16–20 This heightened vulnerability stems from systemic disadvantages, such as unemployment, poverty, residential segregation, unstable housing, and limited IPV-related services, healthcare, and substance use treatment.^21–23 To prevent maternal mortality caused by IPV, it is essential to identify and address multilevel determinants of health specific to IPH among a broader population of WCBA, particularly when IPV mostly begins before pregnancy and persists or escalates to IPH during the perinatal period.⁹

However, significant methodological limitations in the current literature hinder the identification of multilevel determinants of IPH among WCBA. First, prior studies on IPH have primarily focused on individual and relationship factors as underlying mechanisms, leading to a lack of empirical evidence on the role of community and societal factors in IPH.^17,20 This gap is partly due to the absence of comprehensive data that would allow researchers to examine the association between multilevel determinants of health and IPH. Although the National Violent Death Reporting System (NVDRS) is the most widely used dataset for investigating patterns and contributors to IPH, its collected information is also limited to individual (e.g., victim and perpetrator characteristics) and relationship factors (e.g., victim-perpetrator relationship, circumstances surrounding the death). While some studies have incorporated community factors, these have often been derived from victims’ perceptions rather than objective measures.¹⁷ Even when objective measures were used, findings on the relationship between community disadvantage and IPV remained inconclusive.^22,23 This underscores the need to employ more objective measures of community and societal factors17 that accurately capture how broader systemic disadvantages contribute to increased IPH risk.²²

Second, the complex interplay between multilevel determinants of health and IPH, likely characterized by non-linear associations, necessitates more flexible and sophisticated analytical approaches beyond traditional linear models to disentangle the impact of each level of determinants. Previously, community and societal factors, when analyzed alongside individual, more proximal factors, often failed to demonstrate statistical significance,²⁴ which has led researchers, practitioners, and policymakers to focus primarily on individual factors, overlooking community and societal factors as root causes of public health issues, including IPH. To address this limitation, we propose leveraging machine learning, which is well-suited for handling such complexities and identifying important predictors while taking interactions between factors into account. To our knowledge, no study has applied machine learning to ascertain the underlying mechanisms of IPH among WCBA. Therefore, this study aims to predict IPH and identify its key predictors among WCBA in the U.S. by leveraging machine learning and NVDRS data linked with the Maternal Vulnerability Index (MVI) at the county level, incorporating variables across individual, relationship, community, and societal levels.

Methods

Study Design and Data Sources: This study used a retrospective, observational design analyzing secondary data to identify the risk factors driving IPH prediction. The primary dataset was the NVDRS, a state-based surveillance system maintained by the Centers for Disease Control and Prevention (CDC). State VDRS entities systematically aggregate data from death certificates, coroner/medical examiner reports, and law enforcement reports to provide incident-level details about violent deaths. Our analytic sample included NVDRS-recorded homicides involving female victims aged 15-49 years (WCBA) due to their heightened vulnerability to IPV.⁸ We restricted our sample to homicides occurring between 2018 and 2022, leveraging the improved geographic coverage and larger sample size and representativeness following NVDRS’s nationwide expansion in 2018.

Given the critical role of community-level and societal determinants in shaping IPV risk and the limited representation of these factors in the NVDRS data, we further enriched the dataset by linking it to the county-level Maternal Vulnerability Index (MVI) developed by Surgo Ventures.²⁵ The MVI is a composite measure of multidimensional determinants of maternal health risk with six subdomains: reproductive healthcare, physical health, mental health and substance use, general healthcare access, socioeconomic factors, and physical environment (Table 1). We merged MVI data using the county FIPS code of the injury location recorded in NVDRS to reflect the immediate community context and structural vulnerabilities at the time of the homicide, which may differ from the victim’s residence or place of death.

Table 1.

Subdomains and indicators of the Maternal Vulnerability Index.

Reproductive healthcare	Family planning needs, access to abortions, public reproductive health funding, and reproductive health system capacity
Physical health	Hypertension, diabetes, obesity, prevalence of sexually transmitted infections, HIV prevalence, and self-rated health
Mental health and substance abuse	Stress, mental health status, accessibility, and substance abuse
General healthcare	Affordability (postpartum extension status, income eligibility limit for pregnant individuals, and Medicaid expansion status), accessibility (distance to the nearest hospital), care seeking behavior, and quality
Socioeconomic determinants	Educational attainment, minority status, poverty, food insecurity, and social capital
Physical environment	Housing, violent crime rate, transportation, and air pollution

Features and Preprocessing: Our binary outcome of IPH was manually coded by trained state VDRS abstractors using official incident reports to identify whether a homicide was perpetrated by an intimate partner.

We derived structured NVDRS features from categorical variables, which we standardized and converted using one-hot encoding to create binary features. These included:

• Incident characteristics, specifically the year, state and region where the incident occurred, type (single homicide, multiple homicides, or homicide[s] followed by suicide), and the primary weapon implicated in the incident.

• Victim demographics, such as age (continuous feature), race and ethnicity, educational attainment, military status, pregnancy or postpartum status, sexual orientation, and transgender identity.

• Homicide circumstances, including the victim’s history of mental health disorders and treatment, substance use disorders and treatment, IPV perpetration, IPV victimization, and child abuse. These measures also included events that precipitated the homicide: stalking, a verbal argument, a physical fight, family relationship issues, other relationship problems, and jealousy between partners. Lastly, one feature indicated whether post-mortem toxicology screening provided evidence that the victim had consumed at least one substance prior to the homicide.

The MVI features included six subdomain scores, each ranging from 0 to 100, with higher values indicating greater vulnerability. Table 1 outlines the indicators that informed the calculation of each subdomain score. We also included an overall MVI score representing a percentile ranking of county-level vulnerability across the U.S. to capture cumulative vulnerability across all subdomains. Table 2 displays all features and descriptive characteristics of the analytic sample.

Table 2.

Sample characteristics (n=11,498)

IPV: intimate partner violence, NH: Non-Hispanic, GED: Generational Educational Development

¹According to the state where the homicide occurred.

²Jealousy over a current or former intimate partner’s relationship or suspected relationship with another person.

Structured NVDRS variables underwent cleaning, including removal of irrelevant variables and handling of missing data by listwise deletion. Lastly, we merged NVDRS incident records with county-level MVI data using the county FIPS code of the injury location.

Model Development: We selected logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost) classifiers for their robust performance, interpretability, and widespread use and accessibility in public health predictive modeling tasks. We did not include more advanced classifiers, such as deep learning models, due to their complexity and computational resource requirements, making them less feasible for practitioners and organizations working in IPV prevention with constrained resources or technical capacity. We trained each classifier using 10-fold cross-validation (CV) to optimize hyperparameters and reduce overfitting. We optimized hyperparameters through nested CV and grid search, training multiple models on each parameter set, and selecting those that yielded the highest training set F1 score:

• LR utilized an elastic-net regularization approach, tuning both alpha (elastic-net mixing parameter) and lambda (regularization strength)

• RF models underwent hyperparameter tuning for the number of trees (ntree) and the number of variables available for splitting (mtry).

• XGBoost employed a similar approach, tuning tree complexity (max depth), learning rate (eta), and the regularization parameter (gamma).

We trained final models for each classifier with optimal hyperparameters determined through inner CV folds and evaluated on independent test folds generated through outer CV partitions.

Model Evaluation: We evaluated model performance using multiple metrics calculated on independent test sets generated by the outer CV folds. Metrics included the accuracy, precision, recall, F1-score, and area under the precision-recall curve (AUPRC). In selecting the best-performing model, we prioritized average F1-score, recall, and AUPRC across folds. These metrics reflect our primary goal of balancing precision-recall and robustness in imbalanced settings and minimizing false negatives when predicting IPH.

For the best-performing classifier, we assessed feature importance using explainability metrics—calculated using the relevant package’s built-in importance function—that rank variables according to their contribution to predictive performance. These findings allowed the identification of key predictors of IPH, thus providing actionable insights for targeted multilevel intervention efforts.

Statistical Software and Packages: We conducted all analyses in R version 4.4.2 using the glmnet, randomForest, and xgboost packages for model development; caret for cross-validation and tuning; and model-specific packages and pROC and PRROC for performance evaluation.

Ethical Considerations: This study was exempt from the Institutional Review Board review as it used de-identified data. To protect the confidentiality of victims, we restricted access to the NVDRS data to core research team members and reported all linked results using the MVI at the county level rather than more granular geographic units. We thus minimized any risk of re-identification and aimed to uphold ethical standards in the use of sensitive mortality data.

Results

After removing 449 cases due to missing MVI data, the final analytic sample included 11,498 female homicide victims aged 15-49 years, of whom 3,808 (33.1%) were victims of IPH. All homicide victims had a mean age of 31.2 years (SD = 9.2), and almost half (45.3%) were Black women. Most were single or never married (66.7%), and over half (58.0%) had consumed at least one substance prior to death. Firearms were the most common weapon used (72.4%). Verbal arguments (27.6%) and physical fights (8.1%) were the most frequently documented precipitating events, with a smaller proportion involving jealousy (3.3%) or conflicts with family (3.9%) or friends or associates (3.7%). Importantly, only 4.0% of victims were noted as having previously experienced IPV. The highest average MVI scores, representing greater vulnerability, were observed for physical environment (mean = 73.4, SD = 22.2) and socioeconomic determinants (mean = 60.9, SD = 22.7). Table 2 presents additional sample characteristics.

Table 3 displays model performance results. Across the models evaluated, XGBoost demonstrated the best overall performance, achieving a mean (SD) F1-score of 0.83 (0.01), precision of 0.79 (0.01), recall of 0.89 (0.00), and AUPRC of 0.87 (0.01). The best-performing hyperparameters for XGBoost were a maximum depth of 9, learning rate (eta) of 0.3, and gamma of 0. LR showed a slightly improved F1-score (0.84 [0.01]) and recall (0.91 [0.01]) but demonstrated lower AURPC (0.86 [0.01]) and lower precision (0.78 [0.01]) than XGBoost, suggesting relatively reduced precision-recall performance and a higher prevalence of false positives. RF had the lowest performance on the prioritized metrics, with an F1-score of 0.58 (0.02) and recall of 0.49 (0.02).

Table 3.

Comparison of predictive model performance¹ for identifying intimate partner homicide

Model	Accuracy	Precision	Recall	F1-Score	AUPRC
Logistic Regression2	0.77(0.02)	0.78(0.01)	0.91(0.01)	0.84(0.01)	0.86(0.01)
Random Forest3	0.76(0.01)	0.70(0.03)	0.49(0.02)	0.58(0.02)	0.68(0.02)
XGBoost	0.76(0.01)	0.79(0.01)	0.89(0.00)	0.83(0.01)	0.87(0.01)

AUPRC: area under the precision-recall curve, XGBoost: eXtreme Gradient Boosting

¹Performance metrics are reported as means with standard deviations across all outer cross-validation folds.

²Best hyperparameters: alpha = 0.5, lambda = 0.002

³Best hyperparameters: mtry = 8, ntree = 300

⁴Best hyperparameters: max depth = 9, eta = 0.3, gamma = 0

Feature importance analysis from the XGBoost model indicated that the most influential variables included acute circumstances preceding the homicide, such as verbal arguments (gain=0.17), previous IPV victimization (gain=0.07), and demographic characteristics including single status (gain=0.07) and age (gain=0.05) (Figure 1). Notably, all MVI features ranked among the top 15 most important variables, with physical environment emerging as the most influential MVI subdomain (gain = 0.06). The most important contributing features remained the same across racial and ethnic groups (non-Hispanic White, non-Hispanic Black, and Hispanic), suggesting consistent predictors of IPH irrespective of race/ethnicity (Figure S1).

Figure 1.

Top 15 features by gain in predicting intimate partner homicide

Discussion

Among more than 11,000 homicides involving U.S. WCBA between 2018 and 2022, approximately one in three (33.1%) were perpetrated by intimate partners. All victims had a mean age of 31 years, and a plurality (45.3%) were Black, reflecting a profound loss of life during young adulthood among women already burdened by systemic inequities, with consequences reverberating across families, social networks, and entire communities.^16,21 Importantly, only 4.0% of all victims had a documented history of prior IPV victimization, likely underestimating the true prevalence of abuse given persistent underreporting.^26,27 Most incidents were preceded by acute conflicts (i.e., verbal arguments, physical fights, jealousy), and firearms were the most common weapon used, reinforcing their central role in IPV lethality.²⁸ Together, these patterns reveal how women, often without formal indicators of IPV risk, are killed during escalating and potentially preventable conflicts with intimate partners.

Our machine learning modeling further reinforced and extended these insights. We found that XGBoost achieved the best performance in classifying IPH. The model’s F1-score of 0.83 reflects a high overall ability to balance precision and recall. While its recall (0.89) was slightly lower than that of LR (0.91), XGBoost demonstrated better precision (0.79). This improvement in precision is critical, as it minimizes false positives that could otherwise divert limited resources and dilute the focus on truly high-risk cases. Additionally, XGBoost’s AUPRC of 0.87 underscores a stronger ability to discriminate the relatively rare positive class within a highly imbalanced dataset. Although LR showed comparable performance, it assumes linear relationships and lacks the capacity to capture the complex, non-linear interactions likely present among multilevel risk factors. In contrast, XGBoost’s ability to model such relationships more effectively, combined with its robust performance under class imbalance, makes it a more flexible and reliable tool for identifying IPH and informing targeted intervention strategies.

The XGBoost model identified a coherent set of predictors that align with existing knowledge of IPV dynamics while also emphasizing the influence of broader multilevel factors. Acute interpersonal conflicts, including verbal arguments, previous IPV victimization, and jealousy, emerged as some of the most influential predictors, consistent with evidence that spontaneous or escalating disputes are common triggers for IPH.^29,30 Jealousy, often rooted in accusations of infidelity or coercive control, is a well-established precipitant of lethal IPV.³¹ These findings reinforce the importance of recognizing such relationship-level warning signs, which are often incorporated into risk assessments such as the Danger Assessment.³² Victim age and marital status also contributed substantially to prediction, with younger women who had never been married more likely to be victims of IPH. This is consistent with epidemiological patterns showing that girls and women in late adolescence through their 30s experience the highest rate of IPV victimization and lethality.^8,15 Our findings support targeted prevention and screening efforts in settings that serve younger women—such as schools and universities, reproductive health clinics, and prenatal care—where early identification of controlling behaviors and relationship conflict could be lifesaving. Recent calls to expand IPV screening in obstetric and gynecological care to reach WCBA are well-aligned with this evidence.^33,34

Most significantly, all six MVI subdomains appeared among the top 15 predictors, with county-level vulnerability in the physical environment and reproductive health access emerging as particularly influential. The physical environment domain encompasses features such as housing quality, crowding, pollution, and transportation, all of which reflect the safety and stability of the surroundings in which women live. One plausible explanation is that environmental stressors may intensify relationship conflict or constrain women’s ability to leave abusive relationships—for instance, limited housing options may leave victims with few viable alternatives.³⁵ Moreover, physical environment vulnerability may be indicative of disadvantaged areas, where access to critical support services (e.g., legal aid, shelters) that have been associated with reduced IPV risk^36,37 is often restricted. In particular, prior studies highlight rural women’s heightened vulnerability due to physical isolation, stigma, and limited services.36 The prominence of reproductive health access as a predictor similarly points to the role of structural factors: counties with poor access to reproductive health services might indicate broader healthcare shortages and social inequities that leave women with fewer touchpoints for IPV intervention.³⁸ Overall, the inclusion of MVI-derived features allowed us to capture community-level and societal determinants in line with a socioecological understanding of IPV. Our results thereby extend previous work by empirically demonstrating that the context of disadvantage, beyond individual and relationship factors, significantly shapes the risk of IPH.

The low proportion of victims in our sample with documented prior IPV victimization, despite its strong predictive value for IPH, likely reflects limitations in data collection rather than the true absence of prior abuse. Research on intimate partner femicide consistently shows that most victims experience previous violence by the same partner, but such histories are often underreported or undocumented.^26,27 In NVDRS, abstractors derive the “prior IPV” variable from investigative reports, which may overlook abuse known only to the victim or close contacts. Underreporting presents a significant challenge for prevention, as victims are not likely to seek help before IPV escalates and prior incidents may go unrecognized by authorities. These findings underscore the importance of addressing community-level and societal determinants as root causes in efforts to prevent IPH, both among women with known IPV histories and those without.

Our analysis has several limitations that warrant careful consideration. First, we restricted our analysis to traditional machine learning classifiers—LR, RF, and XGBoost—due to their interpretability and accessibility for public health practitioners. Further research is warranted to explore the use of more complex deep learning models that could potentially offer performance gains by capturing intricate patterns. However, these models typically demand significant computational resources and often lack transparency. Advancements in interpretability techniques, such as Grad-CAM and other model explanation methods, may help bridge this gap.⁴¹ Second, coding of many NVDRS features is based on abstractors’ subjective interpretations of law enforcement and medical examiner reports, and potential misclassifications could have introduced bias. We also relied exclusively on structured NVDRS variables and county-level measures, excluding incident narratives available in the NVDRS data that may contain valuable contextual information. While natural language processing of these narratives could enhance predictive performance, their use poses risks of data leakage, particularly when narratives include direct references to IPV or intimate partners, and narratives vary widely in completeness and quality across jurisdictions. Finally, this retrospective analysis does not evaluate the model’s prospective utility for real-time identification of IPH risk. Despite these limitations, this study offers several important contributions. We leveraged a large, recent, multi-state dataset and enhanced it with a novel linkage to the MVI, allowing us to incorporate objective, multidimensional measures of community and societal disadvantage, addressing a key gap in prior IPH research. Our application of cross-validated machine learning methods also enabled the identification of non-linear relationships and complex interactions among multilevel predictors, and our use of interpretable models allowed us to extract actionable insights.

These findings carry important implications for public health research, practice, and policy. The prominence of acute relationship-level triggers, such as arguments and jealousy, alongside community-level and societal predictors underscores the need for multilevel prevention strategies. Routine IPV screening in healthcare settings, particularly during the pregnancy and postpartum period, should extend beyond physical abuse to include indicators of controlling or volatile relationship dynamics. Prior IPV victimization, when known, should prompt immediate intervention, including tailored safety planning. At the community level, improving housing stability, transportation access, and reproductive healthcare in vulnerable areas could reduce IPV lethality by expanding women’s options for support and safety. Our results also highlight the need for policy action to reduce access to firearms among individuals with histories of IPV, through better enforcement of existing laws and expansion of extreme risk protection orders. Beyond direct prevention, this work demonstrates how integrated data systems, such as linking NVDRS to indices like the MVI, can support data-driven responses by identifying high-risk contexts and guiding resource allocation. Predictive models like ours may eventually aid in flagging potential IPV-related cases in real time, enabling more timely service engagement. Future research should build on this foundation by refining models with additional data sources and incorporating qualitative insights from IPV survivors. As machine learning becomes more widely used in violence prevention, maintaining ethical oversight and centering survivor autonomy will be essential. Ultimately, our findings demonstrate that IPH may be a predictable outcome of identifiable patterns and structural conditions. Intervening early through screening, community investment, and coordinated policy is essential to disrupting preventable patterns of IPV and safeguarding the lives of WCBA.

Conclusion

IPH remains a leading cause of maternal mortality in the U.S. Using machine learning on multilevel homicide data obtained from NVDRS and MVI, this study identified individual, relationship, community, and societal risk factors contributing to IPH risk among WCBA. These include acute interpersonal conflicts (e.g., verbal arguments, jealousy) and structural vulnerabilities such as poor physical environments and limited reproductive healthcare access. These findings reinforce the need for prevention strategies that address the determinants of IPV lethality beyond those at the individual level. Future research should incorporate IPV survivor perspectives to refine predictive models and evaluate models prospectively for real-time risk identification, while prioritizing survivor autonomy and health equity.

Supplementary Material

Figure S1.

Top 15 features by gain in predicting intimate partner homicide, by race/ethnicity