Introduction

Domestic violence is a prevalent and costly societal issue. The lifetime prevalence of domestic violence for aged 15–49 was about 27% worldwide (World Health Organization, 2018). It accounts for 21% of all violent crime and can happen to any race, age, gender and affects people of different socioeconomic statuses and education levels (Office for National Statistics, 2018). Domestic violence could generate substantial costs, including lost productivity, medical costs from the health system, and costs of children witnessing and living with violence (World Health Organization, 2021). The total cost of domestic violence was estimated to be $5.8 billion in the United States of America (Max et al., 2004).

Identifying risk factors for domestic violence is an essential first step so to screen potential victims early for prevention to mitigate negative costs and consequences. Past research has identified risk factors associated with domestic violence (e.g., World Health Organization, 2010b), and using machine learning techniques to model and forecast domestic violence (Rodriguez-Rodriguez et al., 2020). Nonetheless, in practice, the difficulty of identifying victims was one of the barriers for health professionals to screen for domestic violence (Duchesne et al., 2021; Miller & Jaye, 2007).

To fill the gap, the present study aimed to identify pertinent risk factors that can predict domestic violence, particularly spousal violence victimization and understanding the complex association across various risk factors, using population-level administrative data and machine learning and network approaches. The study could contribute to the existing literature related to domestic violence issues. Additionally, it could also advance our understanding of opportunities and strategies for upstream intervention and inform the community and the professionals for more comprehensive screening and assessment for domestic violence.

Existing Literature on Risk Factors of Domestic Violence Victimization

The nature of domestic violence could be explained by the ecological framework (Heise, 1998), which suggested that violence could be a result of the complex interplay of multiple factors at the individual, relationship, community, and societal levels. World Health Organization (2010b) applied the ecological framework and identified key risk factors for experiencing intimate partner and sexual violence by women, including young age, low education, and exposure to child maltreatment.

Individual Factors

Individual factors refer to biological factors and personal history that could be associated with an increased likelihood of being a victim of domestic violence. Younger age could be an individual risk factor, especially in females. Past research suggested that marriage at a young age could be linked with early childbearing and reduced opportunities for further education or employment for women, leading to power disparities in the families and subsequently increasing the risk of domestic violence victimization (Hong Le et al., 2014; Speizer & Pearson, 2011).

The association between education level and domestic violence has been well documented. Women with lower levels of education were about two to five times as likely to experience intimate partner violence as compared to women with higher education levels (Ackerson et al., 2008; Boy & Kulczycki, 2008; Martin et al., 2007). Lower educational attainment may reduce a person’s access to resources and increase the acceptance of violence (World Health Organization, 2010b).

A history of childhood maltreatment could also be an individual risk factor for domestic violence. As suggested by the social learning theory (Bandura & Walters, 1977), parents are children’s role models in shaping their behaviors, and thus children’s exposure to abuse or parental violence might result in developing norms that it is appropriate to use violence under certain circumstances. Past studies using survey data have also provided empirical evidence to support the association between childhood maltreatment and domestic violence victimization in different countries and regions (Bensley et al., 2003; Heyman & Slep, 2002).

A history of being a perpetrator could also be a risk factor for domestic violence victimization. Existing literature revealed that domestic violence could be bi-directional (Kwong et al., 1999), as victims might view perpetrators experiencing positive consequences and subsequently adopt similar strategies in their relationship (Palmetto et al., 2013). It could also be attributed to the victim’s desire of seeking vengeance or retaliation (Walker, 2006). Such a cycle of violence indicated that perpetrators might become victims, put differently, prior history as perpetrators might be associated with subsequent victimization of domestic violence.

Another personal history factor associated with domestic violence victimization could be a history of mental health conditions. Individuals with severe mental health conditions are highly vulnerable and may have impaired functioning (Habtamu et al., 2018). These individuals may also face stigma, discrimination, and social exclusion which led to decreased self-confidence, and less hope for the future (World Health Organization, 2010a). Rossa-Roccor, Schmid, and Steinert (2020) have shown that among persons with severe mental illness conditions, psychotic disorder was associated with lifetime violent and non-violent victimization. Other studies have shown that depression and anxiety were associated with a higher risk of exposure to partner violence (Tolman & Rosen, 2016; Trevillion et al., 2012).

In addition, criminal history could also be associated with domestic violence victimization. Prior research has also shown that the risk of domestic violence victimization would be higher among incarcerated individuals than the general population (Chesney-Lind & Shelden, 2013). This could be explained by the assortative mating perspective and lifestyle role, that is, the criminogenic contexts and the relationships with criminally involved partners would lead to increased risks of being exposed to domestic violence (Carbone-Lopez & Kruttschnitt, 2010).

Relationship Factors

Relationship factors associated with domestic violence refer to factors related to an individual’s closest social circle, including peers, intimate partners, and family members. The number of children in the family could be viewed as a relationship factor for domestic violence. An increasing number of children might lead to economic burden and stress in the family, which could result in a higher likelihood of domestic violence (Gokler et al., 2014).

Divorce could be another relationship factor, though the research on the association between divorce and domestic violence has also yielded inconsistent findings (Brownridge et al., 2008; Vatnar & Bjørkly, 2011). The exposure reduction thesis suggested that removing a victim from an abusive marriage or violent relationship could reduce the physical contact between spouses and thus decrease the likelihood of domestic violence (Dugan et al., 2003). On the other hand, the retaliation thesis suggested that violence might continue between ex-spouses after a marriage is dissolved as men could be more violent after separation or divorce (Campbell, 1992).

Community Factors

Community factors refer to the community contexts where social relationships are embedded, including schools, workplaces, and neighborhoods. Poverty or low socioeconomic status could be viewed as a community factor, and its association with domestic violence victimization has been well documented in the existing literature (e.g., Reichel 2017; Semahegn & Mengistie, 2015). Based on a comprehensive review of domestic violence literature, Wilt and Olson (1996) suggested that domestic violence was more common among people with lower socioeconomic status, measured by income, occupation, etc.

Societal Factors

Societal factors refer to the macro-level factors that may influence the rates of domestic violence. Such factors may include cultural and social norms that support males’ superiority and dominance over females or the acceptance of the use of violence to resolve conflicts (World Health Organization, 2010b). Past studies have adopted the ecological framework and examined the role of societal factors in a specific country or region (e.g., Jewkes et al., 2002; Tekkas Kerman & Betrus, 2020). However, limited research has applied the ecological framework and examined the risk factors of domestic violence in the Asian context, where people may have different cultural and social norms (Ang & Stratton, 2018).

The Present Research

This paper aimed to examine risk factors associated with spousal violence victimization in Singapore, a multicultural city-state located in Southeast Asia. The prevalence of domestic violence in Singapore was 11% based on self-reported data, which was lower than the global prevalence rate (World Health Organization, 2018). However, there has been an increase in the number of inquiries and new cases handled by community centers in recent years. Therefore, a study of risk factors associated with spousal violence in Singapore is timely. Furthermore, Singapore reported a lower proportion of people disagreeing with traditional gender beliefs as compared to Western countries, but a higher proportion compared to many other Asian societies (Mathews et al., 2021). In this sense, Singapore would be an ideal test ground to complete our understanding about the societal factors (e.g., gender norms) associated with spousal violence.

As spousal violence usually occurs behind closed doors and detection could be challenging, this study used the administrative records of individuals’ contact with the Personal Protection Orders (PPO) system as a proxy. In Singapore, PPO is granted where the Court is satisfied that family violence has been committed or is likely to be committed and that it is necessary for the protection of the family member (Family Justice Courts Singapore, 2021)Footnote 1. Although individuals who apply for PPO may not necessarily be spousal violence victims, this study aimed to identify upstream intervention opportunities, and thus it took a broader perspective to capture anyone who had applied for PPO.

Using administrative data covering various life domains, this study aimed to identify pertinent risk factors that can predict the likelihood of applying for PPO, and further explored the association of various risk factors with PPO application. Specifically, this paper aimed to address the following two research questions (RQ):

RQ1

What are the pertinent risk factors predicting the likelihood of applying for PPO?

RQ2

What are the interrelationships between the different life adversities and disadvantages that surround PPO application?

To address the RQs, analyses were conducted based on population-level administrative data from two full birth cohorts in Singapore. Past research examining risk factors associated with domestic violence was based on self-reported survey data. Little research has utilized population-level data to predict the risk of domestic violence victimization or examine the complex relationships across various risk factors. The use of population-level data could minimize selection bias and include difficult-to-reach populations. Furthermore, the birth cohorts research design allows us to control for common macro life experiences, that is, individuals included in our study would have experienced similar social, political, and cultural developments. Taken together, this study could contribute to our understanding of spousal violence and facilitate early identification and upstream intervention relating to spousal violence.

Data and Method

Method

Data. This study is part of the Intergenerational Transmission of Criminality and Other Social Disadvantages (INTRACS) research program, which consisted of linked administrative data of multiple birth cohorts in Singapore (see Ting et al., 2022, for more information). The study sample comprised of 51,853 ever-married individuals from the 1980 and 1985 birth cohorts. Individuals with missing educational attainment data (n = 14) were removed from the analysis, which reduced the population size to 51,839. Administrative records of PPO were obtained to indicate whether individuals appeared in the system as applicants who filed a PPO by 30 June 2017, which indicated the allegation of victimization of spousal violence.

A total of 51,839 married individuals were included in the analysis (Table 1). Of these, 51.6% were female, and the majority had degree level and above educational attainment (41.9%). Among all ever-married individuals, 830 were PPO applicants (1.6%). The average age of PPO applicants was 28.4 years (SD = 4.4) when they had their first contact with the PPO system. Note that the age range was up to 37.2 years by the end of the study period. Regarding gender distribution, majority of the PPO applicants were females (n = 754, 90.8%).

Table 1 Counts and percentages of event happening before Personal Protection Order (PPO) filing among PPO applicants and non-applicants
Full size table

When filing a PPO, the applicant would need to provide information about the latest and past incidents of family violence including time, place, details, type of violence, and injuries, and indicate the reasons for filing the PPO. Among the 830 PPO applications in our study, approximately 73.4% were due to placing or attempting to place one in fear of hurt, causing hurt, and wrongfully confining or restraining. The remaining 26.6% resulted from continual harassment.

Variables. Twenty variables were included in the analyses. The outcome variable was a binary variable indicating whether an individual was an application of a PPO (i.e., 1 = Yes, 0 = No). Information was collected from official records in the PPO filing system.

The independent variables of the study captured individual, relationship and community factors. Individual factors include gender, highest qualification attained, housing status, criminal history, and mental illness diagnosis. Relationship factors include age of first marriage, parenthood, history of divorce, number of children, and any prior history of family violence as a PPO applicant or PPO respondent. Finally, the community factor includes receiving social assistance from the government.

Some of the independent variables were static, whereas others measuring the occurrence of certain life events were dynamic in nature. For dynamic variables, given that the administrative data collected in this study captured the dates of the life events in these domains, this study could identify whether the life event happened before or after filing PPO for applicants. Given that this study aimed to identify pertinent factors predicting the likelihood of PPO application, life events after the PPO application would not be applicable. Therefore, for the dynamic variables examined in this study, different cut-off dates were applied to PPO applicants and non-applicants. For PPO applicants, these variables only captured the events happened before the first contact with the PPO system as applicant. For non-applicants, it captured all the events that happened as of the time of data collection.

In this study, the demographic variable controlled for was gender. Data were collected from official records and coded as a binary variable (i.e., 1 = Female, 0 = Male).

Highest qualification attained (i.e., HQA) was captured by this study to measure socioeconomic status. Data on HQA were obtained from the administrative records updated by Singapore residents at age 30 years old. The variable was coded as an ordinal variable with four levels, with “1” indicating primary level and below, “2” secondary level, “3” post-secondary or diploma level, and “4” degree level and above. Furthermore, information about housing status (i.e., whether an individual was staying in a public rental flat at birth and at age 30 years old was also obtained to measure socioeconomic status. Another variable relating to socioeconomic status was whether an individual received short-to-medium term assistance as of the cut-off date. These variables were coded as a binary variable (i.e., 1 = Yes, 0 = No).

For variables in the domain of marriage and parenthood, information of the age of first marriage, first parenthood was extracted from local administrative records and coded as binary variables to indicate whether an individual was married or had his or her first child before 21 years old, respectively. Similarly, data on divorce was collected from official marriage records to indicate whether the individual was divorced or separated as of the cut-off date. The number of children that individuals had as of the cut-off date was also collected and coded as a continuous variable.

Prior history of family violence included childhood maltreatment (i.e., whether a person had contact with the child protection and welfare system in Singapore) and prior history as PPO respondents as of the cut-off date. These variables were coded as binary variables based on administrative records (i.e., 1 = Yes, 0 = No).

To measure criminal history, official information of criminal offenses in Singapore were used to indicate whether an individual had any contact with the criminal justice system by the cut-off date. Particularly, offenses were categorized into four types in this study—violent offense, drug offense, sexual offense, and other types of offense. Correspondingly, four binary variables were generated.

Whether an individual had been diagnosed with any mental illness based on administrative health records before the cut-off date was also captured. Four binary variables were generated to capture four types of mental illness— mood disorders, anxiety disorders, schizophrenia and other psychotic disorders, and other mental health conditions.

Plan of Analysis

Machine learning approach. First, the machine learning (ML) approach was adopted to address RQ1, that is, to identify the pertinent predictors and develop algorithms that could accurately predict PPO application. ML is a subfield of artificial intelligence (AI) that involves training a computer software or model to recognize patterns and relationships in data and using it to predict or make judgments based on new data (Nichols et al., 2019; Sarker, 2021). Analyses were performed using Python (version 3.6.5), Python package numpy 1.16.5, pandas 0.25.1, scikit-learn 0.21.3, and xgboost 0.90. The dataset was split into a training set which was comprised of 70% of the overall population (n = 36,287) and a testing set which was comprised of 30% of the population (n = 15,552). The training data were used to develop the models to predict PPO application, using four ML algorithms: logistic regression (LR), classification and regression trees (CART), random forest (RF) and extreme gradient boosting (XGB). LR is a machine learning technique that uses one or more independent variables to predict the likelihood of a binary (‘yes’ or ‘no’) or categorical result. In a clinical setting, for example, it can be used to predict clinical outcomes in terms of mortality. A CART algorithm is a top-down flowchart-like structure that splits data recursively into smaller groupings based on predictor variable values. After the outcomes have been trained into a decision tree, new data and information may be loaded into the tree, and the node decisions can be followed to forecast the likely clinical outcome. RF is an ensemble method for constructing several decision trees from distinct subsets of training data. Following training, all trees in the forest are run in parallel on new data entries, and the model’s final forecast is determined by the forest’s majority prediction opinion. XGB is a boosting ensemble method that constructs several models (i.e., decision trees), and each model learns from prior models by correcting prediction mistakes (see Pettit et al., 2021, and James et al., 2013 for more information about the different algorithms).

When developing the models, cost-sensitive learning (BrownLee, 2020; He & Ma, 2013) was adopted given that the dataset was highly imbalanced, that is, the prevalence of PPO applicants in the two birth cohorts was approximately 1.6%. The cost-sensitive learning algorithm could introduce different misclassification costs into the learning process. Given that spousal violence could have serious negative consequences, missing an actual PPO applicant would be more problematic as compared to categorizing a non-applicant as applicant and following up with more detailed assessment. Therefore, higher cost was assigned for the misclassification of PPO applicants when developing the ML models using the training data.

The models were then applied to the testing dataset to evaluate the performance of the selected model on unseen data. To evaluate model performance, overall accuracy, recall, precision, specificity and F2 scoreFootnote 2 were reported. The model with the best performance was selected and five-folded cross-validated grid search was carried out to further determine the optimal values of hyper-parameters and meanwhile reduce overfitting. The estimated coefficients of the final model were used to identify pertinent risk factors which could predict the likelihood of applying for PPO.

Table 2 Model performance evaluation metrics
Full size table

Network approach. Next, network approach (NA) was adopted to examine the inter-relationship across all the risk factors (Borsboom & Cramer, 2013) and answer RQ2. Each variable was represented as ‘node’(circles) and the statistical association between two variables, after controlling for other nodes in the network, was represented as an ‘edge’ (lines). All network construction, visualization, and analyses were performed using R statistical software (version 3.6.3), RStudio (version 1.3.1056) (RStudio Team, 2018), R package qgraph 1.6.9, and mgm 1.2–12 (Epskamp et al., 2012; Haslbeck & Waldorp, 2020). In this analysis, we estimated a pairwise mixed graphical model (MGM), which allowed estimation of undirected edges for both continuous and binary variables. To control spurious connections among the variables, we employed the “least absolute shrinkage and selection operator” (LASSO) to correct for false positive by reducing the overall strength of parameter estimates, retaining more solid edges, and shrinking small values edges to zero. For this purpose, we relied on the LASSO regularization with the Extended Bayesian Information Criterion (EBIC) and selected using a high EBIC of 0.5 as recommended by Epskamp, Borsboom, and Fried (2018) to yield adequate network structures. Each edge was estimated twice (once for each node an edge is connected to). We adopted the AND-rule, indicating that an edge was included in our network if it was included in both performed regression models (refer to Haslbeck & Waldorp 2020, p. 8–10 on estimating mixed graphical models).

We assessed the importance of each factor that surrounded PPO application by computing the centrality indices of the overall network structure and assessed the robustness of the network (Epskamp et al., 2018). Node strength provides information on the importance of a node based on the number of connections it has with another node in the network structure. Closeness examines how close a node by considering the indirect connections from that node to all other nodes in the network. Betweenness indicates how important a node is in the average pathway between other pairs of nodes. Robustness analysis was conducted using R package bootnet 1.4.3 to ensure the stability and accuracy of the results (Epskamp et al., 2018) (see Appendix A. Accuracy and stability checks).

Results

Pertinent Risk Factors Predicting the Likelihood of Applying for PPO

The performance of the four models using LR, CART, RF and XGB was reported in Table 2. In general, the LR and XGB models had superior performance in recall and F2 score whereas the CART and RF models had higher specificity and overall accuracy. The precision rates were generally low across the four models. However, given that PPO applicants were rare in the population (i.e., 1.6%), the model performance was acceptable (Berk et al., 2016). As the objective of the study was to identify the true PPO applicants in the population, recall was prioritized, and thus we favored LR and XGB models over the other two models. Furthermore, to build an interpretable model and uncover the direction of the relationships between the likelihood of applying for PPO and other variables, we selected the LR model.

Subsequently, cross-validated grid search was executed on the LR model. The recall of the final model was 83.1%, that is, out of the 249 actual PPO applicants in the testing dataset, 207 were correctly identified. On the other hand, out of the 3,869 cases categorized as PPO applicants by our model, 207 were actual PPO applicants (i.e., precision was 5.4%). The F2 score was 21.3%. The confusion matrix based on the final model was reported in Table 3.

Table 3 Confusion matrix of the final model based on testing data
Full size table

The estimated coefficients of the final model were reported in Fig. 1. First, there was a significant gender difference in the likelihood of applying for PPO. Females were more likely to apply for a PPO as compared to males. With gender accounted for, the most pertinent risk factor predicting the likelihood of being PPO applicant was prior history as PPO respondent, based on the magnitude of estimated coefficients in the final model. Other pertinent risk factors were marriage before 21 years, staying in a public rental flat at 30 years, parenthood before 21 years, childhood maltreatment, anxiety disorders, other mental disorders, drug offenses and other offenses. In addition, the results showed that HQA had a negative coefficient, meaning that higher education level was associated with reduced likelihood of being PPO applicant on average. Put differently, lower educational attainment was found to be a risk factor associated with higher likelihood of being PPO applicants. Note that in the final model, Divorce was negatively associated with the likelihood of being PPO applicant. One plausible reason could be the PPO was only applicable to married people during the study period.

Fig. 1
figure 1

Estimated coefficients based on cost-sensitive logistic regression model

* p < 0.05; ** p < 0.01; *** p < 0.001.

Full size image

Interrelationships Between Different Risk Factors Surrounding PPO Application

Figure 2 presents the undirected network depicting the interrelationship of the factors associated with the likelihood of applying for PPO. Results showed similar findings to the LR model. We identified a positive relationship between PPO applicants and lower HQAFootnote 3 after controlling for all other variables in the network, indicating that individuals with low educational attainment had a higher likelihood of applying for PPO. The network also depicted additional findings: females, those who had PPO made against them previously, marriage before age 21 years, staying in public rental flat at age 30 years, parenthood before age 21 years, and childhood maltreatment had direct association with PPO application, whereas criminal history and mental illness diagnosis were linked to PPO application through the above-mentioned factors (i.e., indirect association).

Figure 3 illustrates the magnitude of the node strength, closeness, betweenness for each risk factor. HQA was the one of the most central nodes (other than Gender – a covariate) across all centrality indices in the entire network, with a relatively high node strength and closeness, highlighting the importance of education level in influencing other risk factors (i.e., life adversities and disadvantages) in the network.

Fig. 3
figure 3

Computed centrality indices for each life adversity and disadvantage that surround personal protection order (PPO) application. Centrality indices are shown as standardized z-scores

Full size image

Discussion

In summary, this study used population-level administrative data to examine the pertinent risk factors which could predict the likelihood of being PPO applicant in Singapore and further explored the associations of different risk factors. The pertinent risk factors identified included covered critical life events in various domains, including educational attainment, staying in a public rental flat, early marriage and parenthood, experience of childhood maltreatment, prior history of being respondent to PPO, criminal history, and mental health concerns. Findings were mostly consistent with international literature related to domestic violence victimization (Bensley et al., 2003; Carbone-Lopez & Kruttschnitt, 2010; Flury et al., 2010; Speizer & Pearson, 2011).

The study further revealed a direct relationship between specific risk factors with PPO application, such as those from low socioeconomic status, had early marriage and parenthood, history of being a PPO respondent, and childhood maltreatment. In contrast, criminal history and mental illness diagnosis were linked indirectly to PPO application through the abovementioned factors. This indicated that criminal history and mental illness diagnosis may not directly lead to a higher risk of experiencing spousal violence, but they can be associated with poorer socioeconomic status background or other adverse events, which in turn leads to spousal violence victimization.

Moreover, lower educational attainment was identified as one of the risk factors with the highest centrality that was linked to many other life adversities and disadvantages such as criminal offenses, marriage before 21 years old, receiving social assistance and staying in a public rental flat at age 30 years old. This finding aligned with the past literature where individuals with higher educational attainment were associated with better well-being (Tran et al., 2021) and life events (Halim et al., 2018).

Implications

The findings of this study could contribute to existing literature related to spousal violence and help policymakers to identify opportunities and strategies for upstream intervention. The pertinent risk factors identified in the study could provide useful information for more targeted community outreach to tackle the domestic violence issue.

For instance, the top-line findings of the study were used to inform the development of the recommendations by Singapore’s multi-stakeholder Taskforce on Family Violence (FVTF)Footnote 4. Specifically, early marriage and young parenthood were found to be pertinent risk factors for filing a PPO, which indicated that providing additional support and proactive outreach to this group would be beneficial. This finding informed the FVTF’s recommendation to encourage young couples to attend marriage preparation programs. Additionally, our findings suggested that staying in a public rental flat could be linked with increased likelihood of being PPO applicants. This informed the FVTF’s recommendation to tap on potential touchpoints such as the Community Link programme (for families with children that are living in public rental flats) for upstream preventive efforts, such as regular engagement and check-ins with families that may have elevated risk of family violence. All the recommendations from the FVTF were accepted in principle by the Government, which has started implementation efforts.

Furthermore, the pertinent risk factor identified in this study could be used as a checklist for first responders (e.g., police or social workers) in suspected family violence incidents to identify at-risk individuals and families. Given that most factors included in the checklist were coded as binary variables to indicate the presence of various life experiences, it would be relatively straightforward for social workers to obtain the necessary information from the individuals involved. In this sense, the checklist could assist with front-line triaging and augment professional judgement when tackling spousal violence issues. Additionally, as risk factors for spousal violence may co-occur and cumulate, the factors identified in the study could be used to form a cumulative risk score to identify the turning point where the risk of domestic violence victimization increases. This could further facilitate the early identification of people and families at higher risk of experiencing spousal violence in the community.

From an operational perspective, there are several numbers of ethical concerns that must be considered in spousal violence intervention which includes: confidentiality of victims be maintained, autonomy where victims of the violence having the right to make their own decisions about their safety and well-being, and non-judgmental approach. It is crucial to note that the objective of the checklist or a screener was not to determine whether an individual was a victim of spousal violence, but to be used as a preliminary screener in the community. However, such information should be further supplemented by dynamic risk and protective factors that could be assessed by the attending professional. At-risk individuals flagged by the checklist could be followed up with additional home visits and attention to further understand the family circumstances and assessed by more comprehensive and structured risk assessment measures. Nonetheless, this approach is not meant to stigmatize family or individual of certain characteristics but to take steps to minimize the risk of harm to all parties involved (including other family members such as children).

Limitations and Future Research

There were several limitations in study. First, by the end of the study period, individuals in the 1980 and 1985 birth cohorts were at age 37 or 32 years respectively. In this sense, the findings of this paper might not be generalizable to victims who experienced spousal violence in their later life stage. It would be rewarding for future research to track spousal violence as well as other life events across lifetime to examine if the risk factors associated with spousal violence would be different across life stages.

Second, this research used official records of life events to capture risk factors of spousal violence. However, there could be other personal or relational factors linked to spousal violence that were not captured by the current administrative data. Additionally, the use of administrative data allowed us to capture spousal violence cases among married people, but not other types of domestic violence. Future research could collect more data, such as attitudes towards male privilege, family functioning and dating violence, through qualitative studies as well as questionnaire surveys, to draw a more complete picture of risk factors leading to domestic violence. Future research could also investigate risk factors for diverse types of spousal violence, such as sexual abuse and emotional abuse.

Third, this research identified risk factors associated with spousal violence victimization, instead of establishing a causal relationship. Further efforts on understanding how and why the pertinent risk factors lead to domestic violence would be crucial.

Fourth, the follow-up period used to examine the dynamic variables in this study differed between individuals who applied for PPO and those who did not. This resulted in a longer tracking duration for individuals who did not apply for PPO. As a result, the presence of dynamic risk factors in PPO applicants could be under-reported.

Last, this study examined the pertinent risk factors associated with spousal violence victimization in the Singaporean context. Interpretation should be made with caution when generalizing the findings to other countries or regions. Future research could apply the model in the current study to samples with different cultural backgrounds or social norms to further understand the risk factors of spousal violence around the globe.

Conclusion

As a concluding remark, this paper identified risk factors associated with the likelihood of being PPO applicants, using population-level administrative records in Singapore. Findings could contribute to existing literature related to spousal violence and help policymakers to identify opportunities and strategies for upstream intervention.