Local responses to gender-based violence, with femicide as its most extreme form, remain uneven across Europe. This column combines machine-learning forecasting with policy evaluation to provide new evidence on where femicides occur in Italy and whether current interventions are working. Contrary to the common narrative, femicide risk is not highest in marginalised or rural regions, revealing a misalignment between the geography of risk and deployment of public support. Policymakers must use predictive analytics to better target resources, ensuring services are strategically placed, well-funded, and integrated across institutions.
Gender-based violence is a systemic problem, with femicide representing its most extreme form. The urgency of this issue has recently permeated high-level policy debates, culminating in the European Council’s adoption of Directive 2024/1385 in May 2024. This directive mandates that Member States criminalise various forms of violence against women and strengthen protection mechanisms by 2027. However, while legislative frameworks are evolving, the practical implementation of local policies remains a complex challenge, due in part to a limited comprehension of the phenomenon. In this article, we aim to shed light on this topic from a territorial perspective by addressing the following questions: Is it possible to identify the areas most exposed to risk of gender-based violence? How should resources to address violence be allocated across territories? Do specialised support services effectively reduce violence, or do they primarily encourage reporting?
Despite the prevalence of gender-based violence, socio-economic research has frequently treated femicides as idiosyncratic events driven by individual or family-specific risk factors, partially neglecting the territorial dimension. Existing literature has extensively explored specific determinants, such as the backlash against female political empowerment (Frisancho et al. 2022), the role of divorce laws (Garcia-Ramos 2021), or the impact of compulsory schooling (Akyol and Kırdar 2022). Yet, systematic territorial patterns of femicides have been largely overlooked in the literature.
Without understanding the geography of violence, policymakers risk misallocating economic resources. Most importantly, we often lack systematic and informative data on gender-based violence with adequate spatial disaggregation and temporal coverage. Without such data, a correct understanding of the phenomenon and its determinants remains impossible.
In a recent paper (Cerqua et al. 2026), we address this gap by constructing a novel, granular dataset of femicides in Italy spanning 2006–2022 and evaluating the efficacy of local anti-violence centres. By combining machine-learning forecasting with causal policy evaluation, we provide new evidence on where violence occurs and whether current interventions are working.
Public discourse often frames femicide as an unpredictable crime, complicating the design of preventive measures. However, our analysis suggests that while individual cases may contain unpredictable (or difficult to measure) elements, the aggregate risk of femicide follows detectable socio-economic and geographic patterns.
Using a set of machine learning algorithms trained on over 100 territorial variables, ranging from labour market conditions to social capital indicators, we developed a forecasting model to predict femicide risk at the local labour-market level (a statistical unit between province and city). Crucially, our risk analysis challenges some traditional narratives regarding violence against women.
We find that femicide risk is not necessarily highest in marginalised, rural, or ‘backward’ areas. Even after accounting for population size, femicides are less likely to occur in rural areas where female emancipation levels are generally lower. This is true even after accounting for the mechanical correlation with demographic predictors and population-related differences.
This finding aligns with the ‘backlash hypothesis’ proposed in recent economics literature (Bergvall 2024, Bulte and Lensink 2019, Daniele 2023, Poblete-Cazenave and Martínez 2025), which posits that violence may increase as men react negatively to women’s economic or political empowerment. Consequently, narratives emphasising the clustering of gender-based violence in fragile territories and policies targeting only socioeconomically deprived areas may miss high-risk territories where changing gender norms are generating friction.
If we can predict which areas are at high risk, are we adequately protecting women in those areas? To answer this, we mapped the coverage of the Italian network of anti-violence centres, which are specialised facilities offering legal, psychological, and logistical support to victims of violence and represent one of the main public tools to combat gender-based violence.
As illustrated in Figure 1, there is a misalignment between the geography of risk and the deployment of public support. While coverage of anti-violence centres has expanded significantly over the last two decades, our analysis reveals that high-risk areas identified by our algorithms do not fully overlap with the presence of anti-violence centres.
Figure 1 Coverage of anti-violence centres and machine-learning-based femicide risk
This implies that the current ‘bottom-up’ approach to establishing these centres – often relying on local volunteer initiatives and sporadic funding – may be leaving the most vulnerable women without adequate protection. This serves as a proof of concept that predictive analytics can, and should, complement local knowledge to improve the targeting of public funds.
The second part of our study evaluates the causal impact of opening a new anti-violence centre. Using a staggered non-parametric difference-in-differences design, we compared changes in gender-based violence outcomes in provinces that opened an anti-violence centre against those that did not, ensuring that the control group followed similar pre-opening trends.
First, we find that the opening of a new anti-violence centre does not lead to a statistically significant reduction in femicides. This null result persists even when we restrict the analysis to intimate partner femicides or focus on high-risk areas as identified in the machine-learning analysis.
While this finding is sobering, it is perhaps unsurprising given the nature of femicides as rare, extreme events often serving as the culmination of long-term abuse. Furthermore, as noted by García-Hombrados et al. (2024), in the context of specialised courts in Spain, standard institutional interventions often face limitations in preventing the most lethal forms of violence.
However, when we broaden the scope to other forms of gender-based violence, the picture changes. We observe a substantial and statistically significant reduction in reported cases of sexual violence, a decline of approximately 20% following the opening of an anti-violence centre.
Figure 2 The effect of anti-violence centre openings on sexual violence and femicides
Interpreting crime data requires caution, particularly regarding gender-based violence, which is notoriously under-reported. An intervention like an anti-violence centre could theoretically increase reported crimes by encouraging victims to come forward (a ‘reporting effect’), or decrease them by preventing violence (a ‘reduction effect’).
Our findings on sexual violence suggest that the ‘reduction effect’ might dominate. While anti-violence centres undoubtedly support victims in navigating the legal system, they also act as hubs for community awareness, prevention training in schools, and cultural change. The significant drop in sexual violence suggests that these broader preventive activities may be deterring assaults.
Conversely, for crimes such as stalking and abuse within the family, we find no significant net change in reported cases. This stability likely reflects the offsetting nature of the two mechanisms: anti-violence centres may be preventing some abuse while simultaneously empowering other victims to report crimes that would have otherwise remained hidden. This interpretation aligns with qualitative evidence from anti-violence centre operators, who note that many women seek help to escape violence but stop short of filing formal police reports due to fear of secondary victimisation by the justice system (Demurtas et al. 2019).
The fight against gender-based violence is moving to the core of the political agenda, yet our research highlights that good intentions must be matched by data-driven strategies.
First, the misalignment between risk and resources suggests that predictive analyses should play a central role in policy targeting. The establishment of support centres should not rely solely on local initiative but should be prioritised in territories identified as high-risk through objective data analysis.
Second, the disparity in outcomes, reducing sexual violence but failing to curb femicides, indicates that anti-violence centres are necessary but not sufficient. Local support structures effectively address sexual violence and support victims of abuse, yet they struggle to prevent femicides. This points to the need for distinct strategies. Combating femicide likely requires earlier intervention points, such as stricter enforcement of restraining orders, better risk assessment by law enforcement (as suggested by Grogger et al. 2021), and broader cultural interventions that address the root causes of toxic masculinity and possessiveness.
Finally, the recently adopted EU Directive emphasises the need for specialised support services. Our findings confirm that these services work, particularly in reducing sexual violence. However, for them to reach their full potential, they must be adequately funded, strategically located, and integrated into a holistic network that includes law enforcement, the judiciary, and educational institutions. Only by treating femicide not as a tragic anomaly, but as a partially preventable systemic failure, can we hope to bring the numbers down.
Source : VOXeu
Experts say $200bln bond-buying effort unlikely to significantly lower housing costs. There's scant evidence so…
That has helped at least to put a floor under euro zone bond prices. Euro…
Bank profitability will remain strong this year despite lower interest rates, says S&P. Saudi banks…
The 2026 review of the EU ETS must be anchored in facts and focus on…
Federal Open Market Committee statements typically sound unanimous, but the Committee’s internal debates rarely are.…
Calls for a digital euro increasingly invoke monetary sovereignty, often on the grounds that Europe…