To prevent the spread of Covid19, many governments have been taking strict measures such as closing borders, imposing nationwide lockdowns and setting up quarantine facilities. While these measures may ensure that social distancing is followed seriously, they may have indirect effects on the economy and adverse effects on the well-being of people, especially the vulnerable population.
To help governments make data-driven policy decisions to effectively deal with pandemics like Covid19, Omdena provided an enabling platform to Artificial Intelligence (AI) experts, data scientists, and domain experts so that they could study the effects of Covid19 policy measures on the vulnerable population. This article describes the results of one of the many facets of this challenge, which focused on the impact of Covid19 on domestic violence.
The goal of this task was to get a better grip on domestic violence and gauge the scale of the problem. To this end, different data sources were used – including news articles, policy data, mobility trends, and domestic violence search rates. The results indicate that the problem of domestic violence could be much bigger than indicated by some of the key figures mentioned in the news. Further, restrictions on movement and strict enforcement of lockdowns may have further amplified the issue. It can be said that domestic violence is a shadow pandemic and it is integral to understand the gravity of the problem, and ensure redressal and support to survivors and vulnerable populations.
Domestic violence — a growing shadow pandemic
The UN Women recently labeled the increase of violence against women as ‘a growing shadow pandemic’. As a consequence of Covid19 policy measures, many victims find themselves in proximity to their abusers due to lockdown measures. The world is witnessing a sharp rise in the number of helpline calls and domestic violence reports, as illustrated in the infographic in Fig. 1. This highlights the pressing need to reflect upon the pre-existing and growing incidence of domestic violence and sensitizing organisations and communities at the grassroots level to provide help and support.
The shadow pandemic’s size — news coverage
The news is replete with reports and cases of domestic violence and its surge during the pandemic. In the beginning of March, the increase in the domestic violence in China received coverage in the news. In the Hubei Province the number of reported cases had tripled in February, compared to the same period last year. Weeks later, similar articles appeared from all over the world.
To get a first grip on gravity and spread of this shadow pandemic, a dataset of about 80,000 Covid19-related news articles was used. This dataset was created using GDELT to query relevant articles and news-please to extract contents. The said dataset has been used for different analyses in the Omdena AI pandemic challenge. To identify the news articles related to domestic violence, the corpus was filtered based on domestic violence-related keywords. In total 1,500 articles were linked to both Covid19 and domestic violence — revealing a connection.
Covid19 and domestic violence related articles
To assess the relevance of the subset of domestic violence-related news articles, topic modelling was performed to discover the abstract topics in the news articles. Latent Dirichlet Allocation (LDA) topic modelling was performed, using gensim. Three topics were modeled, and one of these clearly illustrates that the considered subset covers domestic violence. The word-cloud of this topic is shown in Fig. 2.
Absolute increase in domestic violence related articles
The number of news articles related to both Covid19 and domestic violence started to increase a couple of weeks after the first lockdown measures were implemented in Europe (end of February). This trend is illustrated in Fig. 3.
The increasing trend in domestic violence-related articles could be explained by an overall increase in Covid19-related articles. To study whether the topic of domestic violence has become more dominant in the discussion, the ratio of domestic violence-related articles versus the total number of Covid19 related articles is illustrated in Fig. 4. An increasing trend can be observed, indicating that the issue of domestic violence has become more dominant post the onset of the pandemic.
The shadow pandemic’s size — search rates
The data mentioned in the news is typically in summary form, similar to the key figures shown in the Infographic of UN Women. To get a more detailed grip on the extent and size of the shadow pandemic, different datasets were used:
OXFORD COVID-19 Government Response Tracker (OxCGRT), covers the policy measures taken in 152 countries (accessed on May 8, 2020).
Google COVID-19 Community Mobility Reports, indicates the percentual changesin mobility patterns in 132 countries (accessed on May 8, 2020). The data is relative (_rel) to the mobility patterns between January 7 and February 7, 2020. To limit stochasticity, a moving average (_ma) filter of 7 days (1 week) was applied.
Google Trends data, indicates the search trend of a certain topicover time (accessed on May 8, 2020). To get the percentual change (_rel) in search rates, this date is made relative to a baseline period as well (January 3 – February 13). To remove stochasticity a moving average filter (_ma) of 14 days (2 weeks) was applied to the Google Trends data.
The analysis focuses on countries that are present in all three datasets, and that have sufficient Google Trends data available. The condition of having data available for at least 50% of the considered time period (January 3 – May 8) was imposed. This ensured that the analysis was expansive and included a total number of 53 countries.
The search trend data is considered to be relevant for studying the scale of the problem in situations where one is in search of help, has access to the internet, and has a certain level of trust in societal organizations to be able to offer help. Evidently, the last two conditions are not met in different countries to the desired level across the world. This is, amongst others, reflected in the Human Development Data – for example, the % of the (female) population that has access to the internet. Hence, the results should be considered with these conditions, caveats, and nuances in mind.
Further, the use of search rates has a clear advantage. The victim’s quest for help and receiving help is expected to consist of several steps; and more courage is required for every succeeding step that needs to be taken. The most basic step might be to browse the web for ways to deal with and seek help for domestic violence. Hence, search rate data might reflect the scale of the real problem more accurately than the number of domestic violence reports, because the search rate is probably the first step a victim might take in seeking assistance.
Correlation between policy measures, mobility and domestic violence search rates
The first step in the analysis is to study correlations between the different features in the dataset. To reveal whether there is a mutual relationship or a connection between the variables, a correlation plot is used. The correlation plot for France is shown in Fig. 5. A highly negative correlation (-0.95) between workplace mobility and domestic violence search rates can be observed. And, as expected, workplace mobility highly correlates with the workplace closing policy measure that was implemented by the government (-0.98). These correlations indicate that with the closing of workplaces workplace mobility decreases, and with the decrease in mobility domestic violence increases.
In Fig. 6, the trend of workplace mobility and domestic violence search rates is visualized over time. The negative correlation between both variables is illustrated by the decrease in workplace mobility, while at the same time there is an increase in domestic violence search rates. Compared to the baseline, search rates almost doubled (100% increase). This indicates that the incidence of searching for information related to domestic violence increased with the decline in workplace mobility and as people found themselves stuck at home.
Regression models to quantify the effect of mobility on domestic violence search rates
Regression analysis was used to assess the size and significance of the relationship between workplace mobility and domestic violence search rates. The outcome of this analysis is a regression model that indicates the impact of workplace mobility on domestic violence search rates. Regression analysis is a predictive modelling technique used for forecasting, time series modelling, and finding causal effect relationship between variables.
The linear line in Fig. 7 is the illustration of the output of the regression model for the case study of France. The relationship between mobility and domestic violence is significant, and the slope indicates that with every 1% decrease in mobility, domestic violence search rates increase by 1.4%.
The results of the models for the countries in the top 10 and bottom 10 are listed. In the top 10 countries, decreasing mobility correlates with a steep increase in domestic violence search rates. In the bottom 10 countries, the opposite trend is observed: mobility and domestic violence both decrease at the same time. To further study and explain the results of the different models, the individual plots for the first six in the categories of the top 10 and bottom 10 countries are shown in the next section.
Countries illustrating a strong relationship between a decrease in mobility and an increase in domestic violence
The individual figures for the first six among the top 10 countries are shown in Fig. 8. These countries have a strong relationship between mobility decrease and domestic violence increase.
With the exception of Japan, the peak in search rates has doubled or even tripled in each of the illustrated countries.
Although the coefficient in Japanis relatively high, the peak in search rate is ‘just’ 60%. This is due to a relatively limited decrease in mobility, likely due to less strict lockdown measures in this country.
Vietnamstands out with a peak in domestic violence search rates that increased by more than triple of the baseline. The issue of domestic violence in light of social distancing in Vietnamis stressed in this article as well, stating that the number of people who are in need of shelter has doubled compared to 2018 and 2019.
The figures for Germany, France, Belgium, and South Africa, clearly illustrate the increasing trend in domestic violence search rate as mobility drops.
Countries not illustrating a relationship between a decrease in mobility and increase in domestic violence
The individual figures for the final six countries among the bottom 10 countries are displayed in Fig. 9 and show a positive relationship between mobility and domestic violence.
First of all, the plot for Australia stands out, which witnessed a high increase in domestic violence towards end of February. The sudden rise in domestic violence in Australia is assumed to be a consequence of the bushfires which occurred around this time. This relationship is also expressed in this article: ‘the bushfires’ hidden aftermath: Surging risk of domestic abuse’.
In South Korea, lockdown measures could be considered to be more targeted instead of strict blanket measures, and this could explain the unique trend displayed for this country as compared to the others.
For the Philippines, Thailand, El Salvador,and Jamaica, the simultaneous drop in domestic violence search rates and mobility is visible. This does not mean that there have been less domestic violence incidents. There can be various other factors influencing the observed search rate trends. For example, the peaks in search rates in these countries towards the late February / beginning of March could be explained by the (media) attention given to domestic violence in light of International Women’s Day on March 8. there was a large turnout for the different marches that were held that day, both in Asia and Latin America.
Discussion — action is needed to mitigate the increase in domestic violence
This article studies the impact of the Covid19 global pandemic on domestic violence. The increase in domestic violence can be viewed as the ‘growing shadow pandemic’. This is stressed by the news as well – there is an increasing trend in the number of articles that cover the issue. Some of these articles give insight into the gravity and scale of the ‘growing shadow pandemic’ in summary form. For example, the Infographic of UN Women, shown at the beginning of this article, mentions that in France, Argentina, Cyprus and Singapore domestic violence emergency calls and reports have increased by more than 30%.
The results indicate that the problem of domestic violence could be much bigger than indicated by some of the key figures in the news
The analysis of Google mobility and search rate trends shows that the effect of lockdown measures on domestic violence, such as the closing of workplaces, can be much higher than 30%. In countries where the inverse relationship between the decrease in mobility and increase in domestic violence is strongest, search rates have doubled, and some more than tripled. A search query could be considered the most accessible step in seeking out help. This could explain why the results in this article indicate that the problem of domestic violence could be much bigger than the previously mentioned key figures.
It is important to note that there are many other factors that can influence the search rates results. The extent to which the search rates may accurately reflect the growing scale of the problem of domestic violence also depends on the situation the countries are in. As stated before, a victim is only expected to perform a search query if she or he has access to the internet and a certain level of trust in societal organizations to be able to offer help. These assumptions could explain that a strong relationship is found in many European countries in this study.
The aim of this work is to help build awareness on the issue of domestic violence. Although some countries have adopted steps to mitigate the problem, the results clearly indicate that the issue persist. In this light, the UN recently published a brief with ‘recommendations to be considered by all sectors of society, from governments to international organizations and to civil society organizations in order to prevent and respond to violence against women and girls, at the onset, during, and after the public health crisis with examples of actions already taken’.
Credits — Omdena AI pandemic challenge
The work presented in this article is part of a dedicated task of
the Omdena AI pandemic
challenge. This work would not have been possible without the help of
all team members. Special thanks to Albina Latifi for all efforts
related to the news articles analysis and topic modeling.
Elke Klaassen is an experienced data scientist in the energy domain. She has a multidisciplinary background, with a BSc degree in Innovation Sciences. During her MSc and PhD studies she specialized in the energy transition, and the integration of renewable generation into the electricity grid. She likes to consider the broader context and consider problems from a system perspective. And, by doing so, make project results understandable by using data visualization techniques to explain (complex) models and their outcomes.
 GDELT is a Global Database of Events, Language, and Tone (GDELT), and news-please is news crawler that can extract information from news websites.