Domestic violence in Nepal: Insights from machine learning-based prediction

Abstract

Introduction

Conducting surveys on domestic violence across diverse countries, particularly in lower-middle-income nations like Nepal, poses significant challenges in understanding and addressing the multifaceted dynamics involved in domestic violence research. However, integrating machine learning can help uncover patterns and predictive factors. Therefore, this study aimed to evaluate and compare machine-learning models to identify population-level risk patterns of domestic violence associated with male demographic characteristics using nationally representative data from Nepal.

Methodology

We utilized nationally representative data from the Nepal Demographic and Health Surveys (DHS) conducted in 2016 and 2022. A total of 7,813 observations were analyzed. The outcome variable captured whether women reported experiencing any form of physical or sexual violence. Data preprocessing and analysis were conducted using Stata and Python, with machine learning models implemented through the PyCaret framework. Multiple algorithms were evaluated based on performance metrics including accuracy, precision, recall, F1-score, and AUC.

Result

Significant demographic shifts were observed between 2016 and 2022, including an increase in husbands with only primary education (from 23.2% to 42.52%) and rising rates of alcohol consumption. Among all models tested, LDA achieved the highest accuracy (74.61%) and F1-score (0.6924), while CatBoost and AdaBoost also showed competitive performance.

Conclusion

This study demonstrates the potential of machine learning models in predicting DV risk using male demographic profiles. While acknowledging that findings derived from Nepal-specific data may not be directly generalizable to other sociocultural settings, the findings highlight critical socio-economic determinants such as education, wealth, and substance use and support the use of predictive modeling as a complementary tool for early identification and targeted intervention.

Keywords

domestic violence gender-based patterns gender dynamics intimate partner violence machine learning male dominance

Introduction

Domestic violence (DV) is currently a serious global problem, taking the form of both physical or sexual abuse within intimate relationships among couples.¹ This arises from complex interactions between cultural, psychological, and individual factors.¹ It is noticeable in societies that gender roles, especially for males, have more dominant power, and this uneven power between men and women can create cases where males try to control their partners by being abusive.²

These dynamics are well explained by established theoretical frameworks. Gender power theory suggests that unequal power relations within intimate partnerships enable control and coercion, increasing the risk of violence.³ Resource theory posits that economic stress, unemployment, or perceived loss of provider status among men may trigger violence as a compensatory mechanism.^4,5 Additionally, the ecological framework of violence emphasizes that DV emerges from the interaction of individual characteristics, household dynamics, and broader social and cultural norms rather than isolated personal traits.^6,7

As DV rises, these issues are becoming more visible worldwide now. Different theories of violence mainly focus on women’s rights. Studies on cyber-attacks against women and research into privacy matters are all helping to put DV in a worldwide context.⁸ Some studies showed that cultures where men are dominant lead to more DV against women.^9–11 This can be done by restricting what women can do or wear in public, controlling women’s private lives and sex, and limiting their jobs and free time. This may turn to threats of assault, sharing private information, and even circulating altered photos with nudes. Additionally, a lack of privacy at home can be connected to DV, to.¹²

It is seen that the prevalence of DV within families shot up during the COVID lockdowns.¹³ A study by WHO shows big differences from country to country: 15% in Japan and over 70% in rural Ethiopia, where DV became prevalent.¹⁴ Another study highlights that men’s frustration over not providing enough for their families, along with cultural attitudes, contributes to DV.¹⁵ When men exerted dominance in decision-making, it correlated with increased rates of violence.¹⁶ Several studies examined the factors contributing to violence, recognizing that male demographics are the key element among many interconnected ones.^17,18 Male demographics within a relationship can serve as an indicator of instances of DV.^16,19 From a theoretical perspective, these male demographic characteristics can be interpreted as observable indicators of the broader relational and structural mechanisms described by gender power theory, resource theory, and the ecological framework. For instance, differences in age or education between partners may reflect shifts in household power dynamics, while alcohol consumption and economic conditions may represent behavioral or socioeconomic stressors that influence conflict within relationships.²⁰ Examining these demographic factors will provide a practical way to operationalize theoretical concepts within large-scale survey data. In this context, machine learning methods offer an advantage because they allow multiple interacting demographic characteristics to be analyzed simultaneously and can detect complex patterns that may not be easily captured using traditional regression approaches. Different demographic factors, including age, wealth, education, and occupation, may play crucial roles in shaping the risk of DV.²¹ Studies showed that younger husbands feel stressed when starting their careers and taking family responsibilities. On the other hand, older males experience frustration and a sense of losing control, leading to aggressive behavior at home, but sometimes they control themselves, not becoming violent, which younger people can’t control easily.^22,23 Some other factors, like spousal educational disparity, also influence violence. Wives with higher education than their husbands were more likely to ever experience DV as compared to equally educated couples.²⁴

In low- and lower-middle-income countries, cultural norms, economic insecurity, and social stratification often intensify these mechanisms.²⁵ Nepal represents a particularly relevant context due to its pronounced ethnic, caste, and socioeconomic diversity. While the country comprises over 60 ethnic groups with distinct languages and social structures,^26,27 many communities continue to operate within patriarchal family systems where male authority over household decision-making is normalized. These intersecting cultural and socioeconomic factors may influence both the prevalence of DV and the way demographic risk patterns manifest. In Nepal, DV targets women, manifesting in various forms shaped by entrenched cultural practices rooted in patriarchal norms.²⁸ These patterns, ingrained over generations, have perpetuated a social condition where women often find themselves subordinated to men.²⁸ However, certain men have shown hesitation in accepting violence within families and are increasingly aware of changing societal expectations related to gender roles, suggesting potential shifts in DV dynamics. So, it is noticeable that males contain many contributing factors that trigger DV. This raises a question: if male factors are contributing to DV, is there any way we can identify patterns of risk among female partners by examining male demographic characteristics? Also, conducting large-scale surveys across multiple time points is not easy, especially in lower-middle-income countries like Nepal. To address this challenge, advanced machine-learning (ML) techniques have increasingly been applied in social science research to explore complex social phenomena.^29,30

Previous studies applying machine learning to domestic violence have primarily focused on victim reports, service utilization, or incident-level data.^31,32 Fewer studies have examined the predictive relevance of male demographic characteristics, particularly in low-resource and culturally diverse settings. This highlights a methodological and contextual gap that the present study seeks to address. In the machine learning approach, several models can be useful, including random forests, decision trees, neural networks, etc., and they have been employed to predict different events in multiple domains.³³ Still, there are some challenges, which include difficulties in adopting ML algorithms, issues with sourcing data, and the time-consuming nature of preparing the data.³⁴ To address these challenges, it is crucial to develop and evaluate ML algorithms using proper data to maintain the accuracy of a model.^29,35 Thus, this focus on considerations in machine learning applications is particularly relevant to our study approach. Therefore, rather than aiming to establish causal explanations or develop individual-level prediction tools, this study aims to evaluate and compare machine-learning algorithms to identify population-level risk patterns associated with domestic violence using male demographic characteristics in Nepal. By applying multiple classification models to nationally representative DHS data, the study explores how demographic factors may contribute to identifying broader patterns of domestic violence risk that can inform public health surveillance and prevention strategies.

Methodology

Data collection

Our study used the Nepal Demographic and Health Survey (NDHS) dataset. Nepal is classified as a middle-income country by the World Bank.³⁶ To gain insights, we delved into the DHS program’s data, which meticulously documents instances of DV across different regions of the country. Our study focused on a crucial target group: women who have either experienced violence or have not been abused by their husbands. By analyzing the data collected from Nepal’s 2016³⁷ and 2022³⁸ DHS datasets, we set out on a mission to reveal the findings that we aimed for.

Study design & population

The main outcome variable for this study is individuals who have ever experienced any violence, including physical or sexual violence. Since we used secondary data and relied on the Demographic and Health Survey (DHS) dataset, the data collection was done by DHS, which is globally recognized for its representative data collection on various health indicators in developing countries.³⁹ The DHS employs a two-stage sampling design based on enumeration areas (EAs) and conducts data collection through both in-person and virtual interviews. To ensure privacy and data security, DHS field staff receive standardized training on privacy principles, regulatory compliance, incident handling, and participants’ rights. During interviews, respondents are informed about the purpose of data collection and the authority under which it is conducted.

Inclusion and exclusion criteria were defined in accordance with the study objectives. Data from all regions and provinces of Nepal were included to ensure national representativeness and minimize geographic bias. The analysis was restricted to the most recent standardized DHS datasets from Nepal (2016 and 2022) to maintain consistency and comparability. The study population was derived from the DHS Individual Recode (IR) files, yielding a total of 7,783 records after extraction of variables related to experiences of physical or sexual violence.

Outcome measures

The outcome variable for this study was the experience of DV, derived from the standardized domestic violence module of the Nepal Demographic and Health Surveys. Three measures were constructed from the survey items. The first measure, physical violence, indicates whether respondents reported experiencing acts such as being pushed, slapped, punched, kicked, or otherwise physically harmed by a current or former partner. The second measure, sexual violence, captures experiences of forced sexual intercourse or other unwanted sexual acts by a partner. Each measure was coded as a binary variable, where 0 indicates no reported experience and 1 indicates reported experience. To support the machine learning analysis, a composite outcome variable termed “experienced physical or sexual violence” was created. This variable was coded as 1 if a respondent reported experiencing either physical violence, sexual violence, or both, and 0 if neither form of violence was reported. In this study, the outcome represents the lifetime experience of physical or sexual violence, consistent with the Nepal DHS domestic violence module. Physical and sexual violence were combined to capture overall exposure to severe forms of partner violence and to ensure sufficient event prevalence for stable model training. Previous DHS-based studies commonly adopt this composite measure because these forms of violence frequently co-occur and are measured consistently across survey rounds.

Emotional violence was not included in the composite outcome because its measurement relies on a different set of survey questions and reporting structures within the DHS framework. For example, emotional violence in the Nepal DHS is typically assessed through items such as whether the partner insults the respondent, humiliates her in front of others, or threatens harm, which capture psychological aggression rather than direct physical acts. These questions differ conceptually and behaviorally from those used to measure physical or sexual violence (e.g., being pushed, slapped, kicked, or forced into sexual acts). Because emotional violence reflects a broader range of subjective experiences and may have different reporting patterns across survey rounds, including it in the composite outcome could introduce heterogeneity and reduce comparability between the 2016 and 2022 Nepal DHS datasets used in this study.

Independent variable

We have carefully defined several key dependent variables, including the “Husband age” variable, which was kept as a continuous variable. The “Wealth Index” variable represents different wealth levels among participants, was categorical in nature, and labeled as Poorest, Poorer, Middle, Richer, and Richest. “Husband Education Level” categorizes individuals into tiers like “No Education,” “Primary Education,” “Secondary Education,” and “Higher Education,” providing insights into educational backgrounds. The “Husband working status” variable identifies participants as either employed (“Yes”) or not (“No”), offering valuable information about their employment status. The “Husband Alcohol Consumption” variable was labeled as a binary option, Yes or No, adding to the behavioral context of the husband. Lastly, the “Spousal Education Difference” variable categorizes participants into groups based on the difference in education levels between spouses, which are categorized as husbands better educated, Wives better educated, both equally educated, and neither educated. These offer insights into educational dynamics within couples. Since we used the DHS dataset, we also kept the weighted value, stratification, and primary sampling unit as continuous values so the machine could train itself to make accurate responses while predicting.

Model selection, data training, and analysis

To construct a theoretically grounded and mathematically robust machine learning framework for predicting health-related outcomes in the Nepalese context, we employed an advanced supervised learning pipeline implemented through PyCaret (version 3). This framework facilitated a fully modular architecture encompassing data preprocessing, algorithm benchmarking, hyperparameter tuning, and model interpretation. The predictive task was defined as a binary classification problem where the outcome variable $y \in {0, 1}$ , with 0 denoting didn’t experience DV and 1 indicating experienced DV. The input feature matrix was denoted as $X = {x_{1}, x_{2}, \dots, x_{n}} \in R^{m \times n}$ , where $m$ is the number of samples and $n$ the number of features.

Data preprocessing commenced in Stata version 18, where rigorous inclusion and exclusion criteria were applied to isolate theoretically salient features. After preliminary extraction, we constructed a pairwise correlation matrix to detect multicollinearity among independent variables. As visualized in Figure 1, none of the feature pairs exceeded the critical collinearity threshold ( $∣ corr (x_{i}, x_{j}) ∣ < 0.8$ validating their concurrent inclusion in multivariate analysis and upholding assumptions of feature independence.

Figure 1.

Correlation matrix of key predictors. Caption: The correlation matrix illustrates relationships among predictor variables and confirms the absence of strong multicollinearity, supporting their simultaneous inclusion in the machine learning models.

After confirming the integrity of the variable set, we implemented a comprehensive cleaning protocol that resolved missingness, standardized categorical responses, and harmonized structural definitions across survey years. The curated datasets were exported as CSV files and processed within a Python 3.11.3 environment, using Jupyter Notebook version 5.5.0 for reproducible analysis. Categorical features were encoded via a dual-function mapping:

ϕ (x) = {\begin{cases} OneHot (x), & if x is nominal \\ OrdinalEncoder (x), & if x is ordinal \end{cases}

where

ϕ (x)

denotes the feature transformation applied to the raw categorical variable

x

. The transformations were orchestrated using scikit-learn’s ColumnTransformer, preserving structure across feature types while preparing the design matrix for learning.

We initialized PyCaret’s classification environment using the cleaned 2016 dataset (m=3,664), allowing internal stratified k-fold cross-validation (default $k = 10$ ), automatic scaling, missing value imputation, and feature-type inference. Using the compare models() function, we empirically evaluated a library of classifiers, including tree-based, ensemble, and linear models under identical data conditions. Model performance was assessed using a range of classification metrics derived from the confusion matrix:

• Accuracy: $\frac{T P + T N}{T P + T N + F P + F N}$

• Precision: $\frac{T P}{T P + F P}$

• Recall (Sensitivity): $\frac{T P}{T P + F N}$

• F1-Score: $2 \cdot \frac{Precision \cdot Recall}{Precision + Recall}$

• AUC: Area under the ROC curve generated from the true positive rate (TPR) and false positive rate (FPR)

Here, $T P, T N, F P,$ and $F N$ represent the elements of the binary confusion matrix. These metrics enabled a multidimensional view of model performance, particularly important under class imbalance. To optimize model generalization, we applied tune_model() using randomized search cross-validation. The hyperparameter search aimed to minimize empirical risk via:

θ^{*} = \arg \underset{θ \in Θ}{\min E_{(x, y) \sim D_{v a l}} [L (y, f_{θ} (x))]}

where

L

denotes the binary cross-entropy loss function,

f_{θ}

is the classifier parameterized by

θ

, and

D_{v a l}

The validation distribution under stratified folds. This aligns with minimizing generalization error, a cornerstone of statistical learning theory.

The finalized models were trained on the complete 2016 dataset and deployed on the 2022 dataset ( $m = 4, 149$ ), where the target variable was masked during prediction. Predictions were generated via PyCaret’s predict_model() and subsequently evaluated against the true labels using the aforementioned metrics. Diagnostic visualizations, including confusion matrices and ROC curves, were generated for interpretability and threshold sensitivity assessment.

To enhance transparency and domain alignment, we performed post hoc feature importance analysis using PyCaret’s model interpretation tools. Let the global importance score for the feature $x_{j}$ be $γ_{j}$ ; the features were ranked such that:

γ_{j} = E_{f} [\frac{\partial \hat{y}}{\partial x_{j}}]

where

\hat{y}

denotes the predicted probability. Features with higher

γ_{j}

values exerted a stronger marginal influence on the model’s output. This analytical layer was instrumental in translating algorithmic predictions into actionable insights, allowing for the identification of leverage points in policy design or health interventions.

Results

Table 1 presents the demographic characteristics of the study population in Nepal for 2016 and 2022. The mean age of husbands increased modestly from 35.8 ± 9.45 years in 2016 to 36.3 ± 9.39 years in 2022. The proportion of households in the poorest wealth quintile rose from 22.43% to 29.31%, while the proportion in the richest quintile declined from 15.91% to 12.85%. Husbands’ workforce participation remained consistently high, increasing slightly from 96.67% in 2016 to 97.9% in 2022. Educational attainment showed improvement over time, with a decrease in the proportion of husbands with no formal education (15.72% to 13.79%) and a substantial increase in primary-level education (23.2% to 42.52%). Alcohol consumption among husbands increased from 46.64% in 2016 to 51.7% in 2022. Regarding spousal educational differences, the proportion of couples in which husbands were more educated than their wives declined from 55.19% to 53.56%, while cases where wives were more educated increased from 15.47% to 21.64%.

Table 1.

Demographic characteristics of the individuals.

Characteristics	Observation	Mean (SD)	Observation	Mean (SD)
Husband Age	3664	35.8±9.45	4149	36.3±9.39

	Freq.	Percentage %	Freq.	Percentage %
Wealth Index
Poorest	822	22.43	1216	29.31
Poorer	781	21.32	848	20.44
Middle	754	20.58	816	19.67
Richer	724	19.76	736	17.74
Richest	583	15.91	533	12.85
Husband Occupation
Working	3542	96.67	4062	97.9
Not working	122	3.33	87	2.1
Husband Education
No education	576	15.72	572	13.79
Primary	850	23.2	1764	42.52
Secondary	1616	44.1	1562	37.65
Higher	622	16.98	251	6.05
Husband Consumes Alcohol
No	1955	53.36	2004	48.3
Yes	1709	46.64	2145	51.7
Spousal Educational Difference
Husband better educated	2022	55.19	2222	53.56
Wife better educated	567	15.47	898	21.64
Both equally educated	571	15.58	576	13.88
Neither educated	504	13.76	453	10.92

Feature importance analysis (Figure 2) showed that the husband’s age was the most influential predictor of domestic violence experience, with an importance score of 0.584998. Household wealth index (0.123602) and alcohol consumption (0.117512) were the next most influential predictors, followed by husband’s education status (0.081528) and spousal education differences (0.078606). Husband’s working status contributed relatively little to predictive performance (0.013755). These rankings reflect statistical associations within the predictive models and are not indicative of causal relationships.

Figure 2.

Bar Diagram Showing the Importance of the Machine Learning Model. Caption: Higher importance scores indicate a stronger contribution of the corresponding predictor to the classification performance of the models.

Model performance was assessed using multiple classification algorithms implemented in the PyCaret framework (Table 2). Linear Discriminant Analysis (LDA) showed the highest overall accuracy (0.7461) and a strong AUC (0.6815), along with the highest F1-score (0.6924) and short training time (0.053 s). Logistic Regression demonstrated comparable performance, with an accuracy of 0.7453, AUC of 0.6821, precision of 0.7089, and F1-score of 0.6847, while maintaining low computational cost (0.672 s). Among ensemble methods, CatBoost achieved similar accuracy (0.7414) and a higher F1-score (0.6970) and MCC (0.2026), although with longer training time (1.105 s). Moreover, Ada showed an accuracy of 0.7414 with AUC 0.6705, with training time 0.1060. Based on overall performance across evaluation metrics, LDA, Logistic Regression, CatBoost, and Ada were selected for subsequent analyses.

Table 2.

Performance metrics of the models.

	Model	Accuracy	AUC	Recall	Prec.	F1	Kappa	MCC	TT (Sec)
lda	Linear Discriminant Analysis	0.7461	0.6815	0.7461	0.7092	0.6924	0.1585	0.1980	0.0530
lr	Logistic Regression	0.7453	0.6821	0.7453	0.7089	0.6847	0.1378	0.1842	0.6720
catboost	CatBoost Classifier	0.7414	0.6740	0.7414	0.7057	0.6970	0.1710	0.2026	1.1050
ada	Ada Boost Classifier	0.7402	0.6705	0.7402	0.6987	0.6817	0.1285	0.1687	0.1060
ridge	Ridge Classifier	0.7414	0.6815	0.7414	0.7063	0.6675	0.0943	0.1482	0.0520
svm	SVM - Linear Kernel	0.7395	0.6467	0.7395	0.5933	0.6403	0.0350	0.0510	0.0580
dummy	Dummy Classifier	0.7367	0.5000	0.7367	0.5428	0.6251	0.0000	0.0000	0.0520
gbc	Gradient Boosting Classifier	0.7356	0.6728	0.7356	0.6889	0.6778	0.1181	0.1522	0.1400
lightgbm	Light Gradient Boosting Machine	0.7246	0.6629	0.7246	0.6856	0.6893	0.1532	0.1701	14.2960
xgboost	Extreme Gradient Boosting	0.7129	0.6489	0.7129	0.6765	0.6846	0.1461	0.1555	0.0820
knn	K Neighbors Classifier	0.7090	0.6224	0.7090	0.6711	0.6811	0.1363	0.1439	0.0590
nb	Naive Bayes	0.6876	0.6577	0.6876	0.6927	0.6892	0.2058	0.2068	0.0590
rf	Random Forest Classifier	0.6810	0.6202	0.6810	0.6614	0.6695	0.1248	0.1260	0.2120
et	Extra Trees Classifier	0.6794	0.5872	0.6794	0.6631	0.6697	0.1292	0.1304	0.1960
dt	Decision Tree Classifier	0.6673	0.5641	0.6673	0.6554	0.6605	0.1104	0.1110	0.0540
qda	Quadratic Discriminant Analysis	0.5620	0.5671	0.5620	0.6346	0.5540	0.0565	0.0593	0.0590

Diagnostic performance of selected models

Confusion matrix analysis demonstrated consistently high specificity across all selected models (Figure 3). Linear Discriminant Analysis (LDA) and Logistic Regression (LR) correctly classified a large number of non–domestic violence cases, with true negative counts of 2,917 and 2,927, respectively. Sensitivity was comparatively lower across models, reflecting class imbalance in the outcome variable. Among the evaluated approaches, CatBoost identified the highest number of true positive cases (n = 163), followed by AdaBoost (n = 123), LDA (n = 135), and LR (n = 116). Ensemble-based methods. Therefore, demonstrated greater capacity to detect positive cases, while linear classifiers showed more conservative classification with fewer false positives.

Figure 3.

Confusion matrices of top classification models for predicting experience of domestic violence. Caption: The confusion matrices show the distribution of true positives, false positives, true negatives, and false negatives, illustrating the effect of class imbalance on model sensitivity.

Receiver operating characteristic (ROC) analysis demonstrated moderate discrimination across the evaluated models (Figure 4). The area under the curve (AUC) values ranged from 0.6705 to 0.6821. Logistic Regression achieved the highest AUC (0.6821), followed closely by Linear Discriminant Analysis (0.6815), while CatBoost (0.6740) and AdaBoost (0.6705) showed slightly lower but comparable performance.

Figure 4.

Combined receiver operating characteristic (ROC) curves for top models. Caption: The ROC curves demonstrate the trade-off between sensitivity and specificity across classification thresholds.

Precision–recall curves are presented in Supplementary Figure S1, showing comparable performance across models, with AdaBoost achieving the highest average precision (AP = 0.20) followed by Linear Discriminant Analysis, Logistic Regression, and CatBoost (each AP = 0.19). Calibration curves are shown in Supplementary Figure S2, indicating more stable mid-range probability estimates for AdaBoost, generally consistent calibration for linear models, and greater divergence for CatBoost at extreme probability values.

Discussion

In this study, importance rankings indicate the relative contribution of variables to classification performance rather than causal effects. Such measures are sensitive to correlations among predictors, unmeasured confounding, and the underlying data utilization process. In addition, social and behavioral variables captured in cross-sectional survey data may reflect complex, intertwined mechanisms that cannot be disentangled through predictive modeling alone. Therefore, while feature importance can help identify population-level risk patterns and guide hypothesis generation, it does not establish causal pathways linking male demographic characteristics to domestic violence. Therefore, feature importance results should be interpreted cautiously. In this study, husband’s age emerged as the most influential predictor of domestic violence, with a high importance score (0.584998), consistent with evidence that women whose husbands are more than ten years older face higher odds of experiencing violence in Nepal.⁴⁰ This finding highlights the role of age asymmetry within marital relationships in shaping power dynamics associated with domestic violence. Household wealth index also contributed to the prediction (importance score = 0.123602), aligning with international evidence showing heterogeneous associations between wealth and domestic violence across settings, including declining prevalence among wealthier groups in India and increased risk within higher wealth quintiles in Mozambique.⁴¹

In the case of ‘alcohol consumption,’ it emerged with a significant importance score of 0.117512 in predicting DV. This score indicates a notable connection between substance use, specifically alcohol consumption, and a higher prevalence of DV. A study showed that women whose husbands consume alcohol are at a greater risk of experiencing various forms of violence, including physical, psychological, or sexual abuse. This correlation is particularly pronounced during sensitive periods such as pregnancy and the postpartum phase,⁴² with ‘education status’ and ‘spousal education differences’ scoring 0.081528 and 0.078606, respectively. This aligns with research suggesting that women with higher education than their husbands are more susceptible to DV. This phenomenon is likely due to these women challenging entrenched gender norms, making their husbands feel threatened, resentful, or jealous.⁴³ Such findings highlight the complex interplay between educational disparities and DV, emphasizing the need to consider educational backgrounds in DV risk assessment and understanding. Lastly, ‘women’s working status’ emerged with an importance score of 0.013755, suggesting a potential impact on predicting domestic DV. This finding is consistent with other research that highlights the complex interplay between women’s employment and DV risk. Women’s employment may sometimes provoke feelings of resentment, jealousy, or insecurity in their partners, potentially leading to DV as a means of asserting authority or control.⁴⁴

Overall, the machine learning models demonstrated moderate but meaningful predictive performance in identifying domestic violence risk using nationally representative DHS data from Nepal (2016–2022). Linear models, particularly Linear Discriminant Analysis and Logistic Regression, showed consistent performance across key evaluation metrics while maintaining computational efficiency, whereas ensemble models such as CatBoost and AdaBoost offered comparable performance with advantages in handling nonlinear relationships and class imbalance. The relatively strong performance of Linear Discriminant Analysis may be related to the structured and low-dimensional nature of the demographic variables used in this study, where linear classifiers can generate stable decision boundaries and generalize effectively compared with more complex ensemble methods. Model development included hyperparameter tuning through randomized search cross-validation within the PyCaret framework, evaluation under class imbalance using multiple performance metrics rather than resampling techniques, and post hoc feature importance estimation using PyCaret’s model interpretation tools to assess the relative contribution of predictors to classification performance. The moderate predictive performance observed in this study likely reflects the complex and multidimensional nature of DV. Violence within intimate relationships is influenced by interpersonal dynamics, cultural norms, psychological factors, and community-level conditions that are not fully captured in demographic survey data. While the DHS provides valuable population-level information, it includes relatively limited indicators of behavioral or relational dynamics that may influence DV. As a result, demographic characteristics can provide meaningful signals of DV risk but may not fully capture the broader social and relational processes underlying violence within intimate relationships.

While this study focuses on male demographic characteristics as indicators associated with domestic violence risk, these findings should not be interpreted as attributing individual blame or reinforcing stigmatizing narratives. Domestic violence is driven by complex interactions between institutional, cultural, economic, and structural factors rather than inherent traits of individuals. The purpose of this analysis is to support population-level understanding and prevention efforts, not to justify profiling or punitive actions. Careful and ethical interpretation of predictive models is therefore essential to ensure that they inform supportive, preventive, and system-level interventions. These findings underscore the potential utility of predictive modeling in public health surveillance and targeted interventions, particularly in low-resource settings like Nepal, where early identification of DV risk is crucial. Yet, model outputs should be interpreted with caution and used to complement rather than replace context-aware, qualitative assessments.

From a public policy perspective, predictive models of this nature have the potential to inform population-level domestic violence prevention strategies rather than individual-level decision-making. Risk pattern information derived from routinely collected survey or administrative data can support policymakers in identifying demographic groups or geographic areas where enhanced social support services, awareness initiatives, or alcohol harm-reduction programs may be most needed. In resource-constrained settings, such models may further assist public health agencies and non-governmental organizations in prioritizing outreach, counseling, and community-based prevention efforts.^45,46 Importantly, the application of predictive modeling should be embedded within existing social protection and gender-based violence response frameworks and applied as a complementary tool alongside qualitative assessment, professional judgment, and context-specific expertise.

Limitation

This study has several limitations that should be considered when interpreting the findings. The analysis relies on self-reported data from the Nepal Demographic and Health Surveys, which are subject to reporting biases, including social desirability bias, fear of disclosure, and recall inaccuracies. These factors may influence the reporting of domestic violence, particularly given the sensitive nature of DV in sociocultural contexts where disclosure may be constrained. As with most large-scale household surveys, variations in response accuracy and willingness to report sensitive experiences cannot be fully eliminated.

The use of secondary survey data also restricted the selection of explanatory variables to those available within the DHS framework. As a result, several potentially important contextual and cultural determinants, such as family environment, gender norms, attitudes toward violence, and community-level social dynamics, could not be directly incorporated into the models. While male demographic characteristics may serve as indirect indicators of broader structural and socioeconomic conditions, they do not fully capture the complex relational and cultural mechanisms underlying domestic violence. This limitation reflects data availability rather than conceptual omission.

Although a range of machine learning algorithms were evaluated using the PyCaret framework, including linear, ensemble, and tree-based models, more complex approaches such as deep neural networks were not applied. Neural network models are typically better suited to high-dimensional or image-based data and often require large feature-rich datasets to achieve reliable performance. Given the structured, tabular nature of the DHS variables and the importance of interpretability for public health relevance, neural network approaches were not considered appropriate for the current analysis. This represents a methodological trade-off between model complexity and transparency. The demographic shifts observed between 2016 and 2022, including changes in education levels, wealth distribution, and alcohol consumption, may reflect broader social, economic, and cultural transformations occurring in Nepal during this period. However, the cross-sectional design of the DHS limits the ability to directly assess how evolving gender norms, family structures, or institutional changes influenced domestic violence dynamics over time. Longitudinal data or mixed-methods approaches would be better suited to examine these contextual processes in greater depth. Finally, the findings are specific to the Nepalese context and should be interpreted accordingly. Cultural norms, gender relations, and institutional structures vary widely across countries, particularly between different income settings. Consequently, the predictive patterns identified in this study may not be directly generalizable to other sociocultural contexts. The results should therefore be understood as population-level insights relevant to Nepal rather than universally applicable models of domestic violence risk.

Conclusion

This study demonstrates the potential of machine learning models in predicting domestic violence (DV) risk using nationally representative data from the Nepal Demographic and Health Surveys (2016 and 2022). By applying a suite of algorithms through the PyCaret framework, we identified key predictors, including husband’s age, wealth index, alcohol consumption, and spousal educational differences, that significantly contribute to DV risk. Among the models tested, Linear Discriminant Analysis (LDA) emerged as the most effective overall, offering a balanced performance across accuracy, precision, recall, and F1-score. Ensemble models such as CatBoost and AdaBoost also showed competitive results, particularly in handling class imbalance and capturing feature interactions. However, the overall discriminatory power of all models remained moderate, highlighting the complexity of DV and the limitations of relying solely on structured survey data. Despite these challenges, the findings offer valuable insights for early risk identification and targeted interventions, particularly in low-resource settings. As machine learning continues to evolve, integrating richer behavioral, contextual, and longitudinal data may enhance predictive accuracy and support more effective public health strategies aimed at preventing domestic violence.

Supplemental Material

Supplemental Material - Domestic violence in Nepal: Insights from machine learning-based prediction

Supplemental Material for Domestic violence in Nepal: Insights from machine learning-based prediction by MD Nahid Hassan Nishan, M Z E M Naser Uddin Ahmed in Journal of Public Health Research.

Supplemental Material

Supplemental Material - Domestic violence in Nepal: Insights from machine learning-based prediction

Supplemental Material for Domestic violence in Nepal: Insights from machine learning-based prediction by MD Nahid Hassan Nishan, M Z E M Naser Uddin Ahmed in Journal of Public Health Research.

Footnotes

Acknowledgments

We are grateful to the Demographics and Health Survey (DHS) for providing us with access to their valuable survey data. This data has played a valuable role in our research, providing us with their datasets.

ORCID iD

MD Nahid Hassan Nishan

Ethical considerations

This study used publicly available secondary data from the Nepal Demographic and Health Surveys (2016 and 2022). Because the analysis relied exclusively on anonymized datasets with no direct participant involvement, additional ethical approval was not required. Ethical clearance and informed consent were obtained by the DHS program during the original data collection. Details of the DHS privacy and ethical procedures are available at: .

Author contributions

Conceptualization: MD Nahid Hassan Nishan. Data curation: MD Nahid Hassan Nishan. Formal analysis: MD Nahid Hassan Nishan, M Z E M Naser Uddin Ahmed. Methodology: MD Nahid Hassan Nishan, M Z E M Naser Uddin Ahmed. Writing- original draft: All authors. Writing-review & editing: All authors

Funding

The authors received no financial support for the research, authorship, and/or publication of this article: This study did not receive any funds from the public or any donor agency.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Supplemental material

Supplemental material for this article is available online.

Appendix

References

Rakovec-Felser

. Domestic Violence and Abuse in Intimate Relationship from Public Health Perspective. Health Psychol Res 2014; 2: 1821. https://doi.org/10.4081/hpr.2014.1821

Mshweshwe

. Understanding domestic violence: masculinity, culture, traditions. Heliyon 2020; 6: e05334. https://doi.org/10.1016/j.heliyon.2020.e05334

Kearns

D’Inverno

Reidy

. The Association Between Gender Inequality and Sexual Violence in the U.S. American Journal of Preventive Medicine 2020; 58: 12–20. https://doi.org/10.1016/j.amepre.2019.08.035

Meyer

Hardt

Brambilla

, et al. Explaining intimate partner violence through economic theories: A systematic review and narrative synthesis. Aggression and Violent Behavior 2024; 77: 101929. https://doi.org/10.1016/j.avb.2024.101929

Liming

. Examining the Differing Effects of Economic Hardship and Poor Maternal Wellbeing on Cumulative Exposure to Adverse Childhood Experiences. Journ Child Adol Trauma 2019; 12: 307–321. https://doi.org/10.1007/s40653-018-0230-z

Heise

. Violence Against Women: An Integrated, Ecological Framework. Violence Against Women 1998; 4: 262–290. https://doi.org/10.1177/1077801298004003002

Shorey

Chua

CMS

Chan

, et al. Women living with domestic violence: Ecological framework-guided qualitative systematic review. Aggression and Violent Behavior 2023; 71: 101835. https://doi.org/10.1016/j.avb.2023.101835

Bailey

Flynn

Henry

(eds). The Emerald international handbook of technology-facilitated violence and abuse. 1st ed. Emerald Publishing, 2021.

Al-Natour

Gillespie

. Men’s Perceptions of Violence against Women: Improving Cultural Literacy for Combating Violence within Family. Clin Nurs Res 2022; 31: 251–260. https://doi.org/10.1177/10547738211027223

10.

Lomazzi

. The Cultural Roots of Violence against Women: Individual and Institutional Gender Norms in 12 Countries. Social Sciences 2023; 12: 117. https://doi.org/10.3390/socsci12030117

11.

Fleming

Gruskin

Rojo

, et al. Men’s violence against women and men are inter-related: Recommendations for simultaneous intervention. Soc Sci Med 2015; 146: 249–256. https://doi.org/10.1016/j.socscimed.2015.10.021

12.

kiener-manu

katharina

. Cybercrime Module 12 Key Issues: Gender-Based Interpersonal Cybercrime. https://www.unodc.org/ (accessed 21 January 2024).

13.

Kourti

Stavridou

Panagouli

, et al. Domestic Violence During the COVID-19 Pandemic: A Systematic Review. Trauma, Violence, & Abuse 2023; 24: 719–745. https://doi.org/10.1177/15248380211038690

14.

García-Moreno

Jansen

Ellsberg

, et al.WHO Multi Country Study on Women’s Health and Domestic Violence Against Women. Initial results on prevalence, health outcomes and women’s responses, 2005.

15.

Pun

Tjomsland

Infanti

, et al. Violence exists to show manhood’: Nepali men’s views on domestic violence - a qualitative study. Glob Health Action 2020; 13: 1788260. https://doi.org/10.1080/16549716.2020.1788260

16.

Conroy

. Gender, Power, and Intimate Partner Violence: A Study on Couples From Rural Malawi. J Interpers Violence 2014; 29: 866–888. https://doi.org/10.1177/0886260513505907

17.

McCarthy

Mehta

Haberland

. Gender, power, and violence: A systematic review of measures and their association with male perpetration of IPV. PLoS ONE 2018; 13: e0207091. https://doi.org/10.1371/journal.pone.0207091

18.

Sikweyiya

Stern

Hanass-Hancock

, et al. Intersections between disability, masculinities, and violence: experiences and insights from men with physical disabilities from three African countries. BMC Public Health 2022; 22: 705. https://doi.org/10.1186/s12889-022-13137-5

19.

Douglas

Hines

. The Helpseeking Experiences of Men Who Sustain Intimate Partner Violence: An Overlooked Population and Implications for Practice. J Fam Violence 2011; 26: 473–485. https://doi.org/10.1007/s10896-011-9382-4

20.

Amir-ud-Din

Idrees

Farooqui

, et al. Exploring spousal disparities: Age, earnings, and education as predictors of intimate partner violence in 29 developing countries. Womens Health (Lond Engl) 2024; 20: 17455057241310289. https://doi.org/10.1177/17455057241310289

21.

Abdi

Mahmoodi

Afsahi

, et al. Social determinants of domestic violence against suburban women in developing countries: a systematic review. Obstet Gynecol Sci 2021; 64: 131–142. https://doi.org/10.5468/ogs.20211

22.

Wisse

Rus

Keller

, et al. “Fear of losing power corrupts those who wield it”: the combined effects of leader fear of losing power and competitive climate on leader self-serving behavior. European Journal of Work and Organizational Psychology 2019; 28: 1–14. https://doi.org/10.1080/1359432x.2019.1635584

23.

Chen

CKM

, et al. Stress in Balancing Work and Family among Working Parents in Hong Kong. IJERPH 2022; 19: 5589. https://doi.org/10.3390/ijerph19095589

24.

Rapp

Zoch

Khan

MMH

, et al. Association between gap in spousal education and domestic violence in India and Bangladesh. BMC Public Health 2012; 12: 467. https://doi.org/10.1186/1471-2458-12-467

25.

Sardinha

Nájera Catalán

. Attitudes towards domestic violence in 49 low- and middle-income countries: A gendered analysis of prevalence and country-level correlates. PLoS One 2018; 13: e0206101. https://doi.org/10.1371/journal.pone.0206101

26.

Pradhan

. The Ain of 1854 and after: Legal pluralism, models of society and ethnicity in Nepal. Epub ahead of print 2015. https://doi.org/10.13140/RG.2.1.2039.7529

27.

Pariyar

Lovett

. Dalit identity in urban Pokhara, Nepal. Geoforum 2016; 75: 134–147. https://doi.org/10.1016/j.geoforum.2016.07.006

28.

Sapkota

Simkhada

Newton

, et al. Domestic Violence Against Women in Nepal: A Systematic Review of Risk Factors. Trauma, Violence, & Abuse 2024; 25: 2703–2720. https://doi.org/10.1177/15248380231222230

29.

Hui

Constantino

Lee

. Harnessing Machine Learning in Tackling Domestic Violence-An Integrative Review. Int J Environ Res Public Health 2023; 20: 4984. https://doi.org/10.3390/ijerph20064984

30.

Tucci

Burns

. What is Machine Learning and How Does It Work? In-Depth Guide. Enterprise AI. https://www.techtarget.com/searchenterpriseai/definition/machine-learning-ML (accessed 21 January 2024).

31.

Hui

Constantino

Lee

. Harnessing Machine Learning in Tackling Domestic Violence—An Integrative Review. IJERPH 2023; 20: 4984. https://doi.org/10.3390/ijerph20064984

32.

Hadjimatheou

Quiroz Flores

Weir

, et al. Using unsupervised machine learning to find profiles of domestic abuse perpetrators. Policing: A Journal of Policy and Practice 2024; 18: paae092. https://doi.org/10.1093/police/paae092

33.

Parmigiani

Barchielli

Casale

, et al. The impact of machine learning in predicting risk of violence: A systematic review. Front Psychiatry 2022; 13: 1015914. https://doi.org/10.3389/fpsyt.2022.1015914

34.

Onyema

Almuzaini

Onu

, et al. Prospects and Challenges of Using Machine Learning for Academic Forecasting. Comput Intell Neurosci 2022; 2022: 5624475. https://doi.org/10.1155/2022/5624475

35.

Kausik

Rashid

Baki

, et al. Machine learning algorithms for manufacturing quality assurance: A systematic review of performance metrics and applications. Array 2025; 26: 100393. https://doi.org/10.1016/j.array.2025.100393

36.

World Bank Country and Lending Groups . World Bank Data Help Desk. https://datahelpdesk.worldbank.org/knowledgebase/articles/906519-world-bank-country-and-lending-groups (accessed 5 January 2024).

37.

Ministry of Health - MOH/Nepal, New ERA/Nepal, ICF . Nepal Demographic and Health Survey 2016. MOH/Nepal, New ERA, and ICF, 2017. https://dhsprogram.com/pubs/pdf/FR336/FR336.pdf

38.

Ministry of Health and Population [Nepal], ICF . Nepal Demographic and Health Survey 2022. Ministry of Health and Population [Nepal], 2023. https://www.dhsprogram.com/pubs/pdf/FR379/FR379.pdf

39.

The DHS Program - Demographic and Health Survey (DHS). https://dhsprogram.com/Methodology/Survey-Types/DHS.cfm (accessed 21 January 2024).

40.

Rahman

Jamil

Nahar

, et al. Factors that provide protection against intimate partner physical violence among married adolescents in Bangladesh. Front Public Health 2023; 11: 1125056. https://doi.org/10.3389/fpubh.2023.1125056

41.

Kebede

Van Harmelen

A-L

Roman-Urrestarazu

. Wealth Inequality and Intimate Partner Violence: An Individual and Ecological Level Analysis Across 20 Countries. J Interpers Violence 2022; 37: NP15568–NP15593. https://doi.org/10.1177/08862605211016337

42.

Bhatta

Assanangkornchai

Rajbhandari

. Does husband’s alcohol consumption increase the risk of domestic violence during the pregnancy and postpartum periods in Nepalese women? BMC Public Health 2021; 21: 5. https://doi.org/10.1186/s12889-020-10021-y

43.

Masaiti

Mapoma

Sikwibele

, et al. The Relationship between Spousal Violence and Levels of Education. IJAHE 2022; 9: 1–20. https://doi.org/10.6017/ijahe.v9i1.15229

44.

Dahal

Joshi

Swahnberg

. A qualitative study on gender inequality and gender-based violence in Nepal. BMC Public Health 2022; 22: 2005. https://doi.org/10.1186/s12889-022-14389-x

45.

Kelly-Weeder

Phillips

Rounseville

. Effectiveness of public health programs for decreasing alcohol consumption. PI 2011; 2011: 29.

46.

Saavedra

Christopher

Illei

, et al. Advancing research on strategies to reduce drug use and overdose-related harms: a community informed approach to establishing common data elements. Harm Reduct J 2025; 22: 169. https://doi.org/10.1186/s12954-025-01301-0

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.10 MB

0.04 MB