COVID-19 lockdown and domestic violence: Evidence from internet-search behavior in 11 countries

Abstract

We study the impact of the COVID-19 pandemic on domestic violence in 11 countries with different ex-ante incidence of domestic violence (DV) and lockdown intensity. We use a novel measure of DV incidents that allows us to make cross-country comparisons: a Google search intensity index of DV-related topics. Our difference-in-difference estimates show an increase in DV search intensity after lockdown (30%), with larger effects as more people stayed at home (measured with Google Mobility Data). The peak of the increase in DV appears, on average, 5 weeks after the introduction of the lockdown. While we observe that the positive impact on DV is a widespread phenomenon, the effect in developed countries is more than twice as strong as in Latin American countries. We show that the difference in impact correlates with the intensity of compliance with stay-at-home measures in the two groups.

1. Introduction

Violence against women is a serious health concern all around the world. About 1 in 3 (30%) of women worldwide have experienced some form of physical and/or sexual violence by their intimate partner in their lifetime (World Health Organization, 2013). The situation may have worsened during the COVID-19 pandemic, due to the increase in unemployment and because the stay-at-home orders forced victims to stuck at home with abusers and decrease their possibility of escaping from a violent situation (Aizer, 2010, Anderberg, Rainer, Wadsworth, Wilson, 2016, Bhalotra, Britto, Pinotti, Sampaio).

However, there are important data challenges for the analysis of the patterns of domestic violence (DV hereafter) around the world. First, there is a lack of comparable data to make cross-country comparisons (Jayachandran, 2015). Second, because of its private nature, most of the cases of domestic violence remain hidden and are not usually reported, neither in police reports nor in surveys (Aizer, 2010).1 Finally, when data exists, there is a significant delay between the occurrence of the offences and the availability of the data for researchers, making impossible any analysis of the effect of the COVID-19 pandemic on DV.

In this paper we analyze the impact of the COVID-19 pandemic on domestic violence in several countries, using a novel indicator of its incidence based on Google searches of DV related topics. This indicator overcomes the issues listed before, as it comes from an almost real time high-frequency data (daily) available for many countries. Furthermore, Google searchers “express interests not easily elicited by other means” (Stephens-Davidowitz, 2014), which can help to avoid the under-reporting issues explained before. Our sample includes the United States and —according to their GDP— the five largest Latin American countries (Argentina, Brazil, Chile, Colombia and Mexico) and the 5 largest European countries (France, Germany, Italy, Spain and United Kingdom). All these countries were significantly affected by the COVID-19 in terms of deaths and economic impacts, and faced different degrees of lockdowns.2

Fig. 1 shows the evolution of our measure of DV-incidence in 2020 (bold line) and previous four years (grey lines). We observe a large increase in DV search intensity after the lockdown that coincides with an increase in the time people stay-at-home —measured with Google Mobility Data (blue data). Our event-study calculations based on this data show that the positive effect of the lockdowns on our measure of domestic violence remains statistically significant for 10 weeks after lockdown, with a peak on week 5. This result is further reflected in our difference-in-difference regressions which show an increase in search intensity of 30% relative to the week prior to lockdown. When we analyze how residential mobility impacts our index of DV we find a 20% increase in search intensity for every one standard deviation increase in residential mobility.

Fig. 1.

Trends in online searches and residential mobility.

Notes: The figure above plots the average number of weekly search intensity of domestic violence related topics across 11 countries by week since lockdown in 2020 (bold black line) and previous four years (grey lines), and the residential mobility index in 2020 (blue line). The blue curve reports Google residential mobility data. The vertical red dashed line corresponds to the week of the introduction of lockdown type 2 —as defined in Hale et al. (2020)— for each country. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Although we observe an increase in DV in every country, the effect of the lockdown in developed countries is more than twice as strong as in Latin American countries. This difference seems to be related to a differential in compliance to the stay-at-home measures. While as a response to the lockdowns residential mobility goes up by one-third of a standard deviation in developed countries in our sample, it only increases by less than one-fifth of a standard deviation in Latin America, on average. Although there are differences in the intensity of the actual lockdowns the effect of a one standard deviation change in mobility on the DV indicator is similar between the two groups of countries, and only statistically different from zero in Latin America.

In the final section of the paper we provide evidence that our search-intensity index is a good measure to monitor the evolution of DV incidence. Using data for Spain, we observe that search intensity related to DV topic and calls to the DV helpline present a similar behavior both before and after lockdown.

This paper is related to a large emerging literature analyzing the effects of the COVID-19 pandemic on domestic violence. The majority of these papers focus in one particular country or city and use administrative data from calls-for-service or crime/police reports.3 The literature shows an increase in the rate of domestic violence calls for service during the lockdowns for a diverse set of countries.4 However, some of the studies that find an increase in the number of calls to DV-services also show a decrease in DV crime reports rates during the pandemic (Miller, Segal, Spencer, 2020, Bullinger, Carr, Packham, 2020, Silverio-Murillo, Balmori de la Miyar, Hoehn-Velasco).5 We contribute to this literature with a novel measure of DV incidence that complements the existing one by allowing to monitor cross-country DV-incidence in real-time and with less issues of under-reporting. In a concurrent paper, Anderberg et al. (2020) propose an algorithm for measuring variation in DV incidence based on internet search activity and test it using data for two cities, London and Los Angeles, during the COVID-19 pandemic. Whereas the focus of Anderberg et al. (2020) is methodological, ours is to exploit Google search data to estimate the impacts of the lockdowns on DV for a large set of countries, in order to have a broader picture of the phenomenon and to understand the role of lockdown compliance in the heterogeneous effect between countries.

2. Data

We analyze data for eleven countries: United States and —according to their GDP— the five largest Latin American countries (Argentina, Brazil, Chile, Colombia and Mexico) and the five largest European countries (France, Germany, Italy, Spain and United Kingdom). All these countries were significantly affected by the COVID-19 in terms of deaths and economic impacts, and faced different degrees of lockdown. Important also for the kind of data we exploit, these are countries with high internet penetration (Internet-World-Stats, 2019).

Our main outcome variable is a Google trends index of search intensity for terms related to domestic violence. The data comes from Google Trends, that is a publicly available data with a weekly frequency, representing the search behavior of Internet users. The Google trends’ domestic violence search intensity index is calculated, for each country, as the fraction all Google searches devoted to the topic “Domestic Violence” or terms related to the domestic violence hotline in each country (e.g. ”domestic violence hotline”, ”linea mujer”, etc.).6 Fig. 2 shows for each of the eleven countries the evolution of the average search intensity of domestic violence related topics by month since lockdown in 2020 (bold black line) and previous four years (grey lines).

Fig. 2.

By country: Trends in 2020 vs previous years.

Notes: The figures above plot the average number of weekly search intensity of domestic violence related topics for 11 countries by month of year in 2020 (bold black line) and previous four years (grey lines). The blue curve reports Google residential mobility data, the green line reports workplace mobility data and the brown line reports retail mobility. The vertical red dashed line corresponds to the month of the introduction of lockdown type 2 —as defined in Hale et al. (2020)— for each country. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

The dates of the countries’ introduction of stay-at-home orders comes from Hale et al. (2020). For each country this date indicates the moment of the first introduction of a lockdown type 2. This type of lockdown requires not leaving house with exceptions for daily exercise, grocery shopping, and ‘essential’ trips. Table 4 of the Appendix lists the date of the lockdown for each country in our sample.

Finally, we use Google mobility data, which provides information about how the length of stay at different places change in comparison to a baseline. The baseline is the median value, for the corresponding day of the week, during the five-week period 3 Jan 6 Feb 2020. This data also has a weekly frequency. In this paper we mainly focus on the mobility trend data for places of residence, which gives us a measure of the intensity of the actual lockdown in each country. Fig. 2 displays, for the 11 countries in our sample, the Google mobility index for residential places (blue lines), workplace mobility (green line) and retail mobility (brown line).

Putting all these data together, Fig. 2 suggests a correlation between the introduction of the lockdowns, a subsequent drop in mobility, and an increase in the search intensity of DV-related topics some weeks after. This correlation is present everywhere and does not seem to be explained by seasonality, as such increase in DV searches was not observed the years before during the same calendar months (compare bold black lines and grey lines). In the next section we run an event-study and a difference-in-difference model with this data in order to estimate the magnitude of the impact of the lockdown on our index of DV incidence.

3. Lockdown impacts on domestic violence: event study approach

We estimate the impact of COVID-19 related lockdowns on search intensity of domestic violence related topics using both a week-by-week event study specification and a two-period before-after specification. The event study specification is

$$Log\left(SearchIntensit{y}_{ctwy}\right)=\sum _{\tau =-10}^{30}{\beta }_{\tau }\mathbb{1}{\left(Week\tau \right)}_t\times Year{2020}_y+{\lambda }_{cy}+{\varphi }_{cw}+{ϵ}_{ctwy},$$ (1)

where the outcome is the google trends’ search intensity index in country $c$ in week $t$, week-of-year $w$, and year $y$. The indicator function $\mathbb{1}{\left(Week\tau \right)}_t$ takes a value of one if week $t$ falls $\tau$ weeks before or after the week prior to lockdown -our reference week.7 The sample is restricted to weeks -10 through 30 from lockdown week. $Year{2020}_y$ is an indicator for weeks in 2020. The $\beta$ coefficients track weekly changes in the search intensity during 2020 relative to the previous five years. We include country-by-year (${\lambda }_{cy}$), and country-by-week (${\varphi }_{cw}$) fixed effects to allow for country-specific trends in search intensity across years and season. This means that our estimates are obtained using within-country variation of weekly searches in 2020 relative to the previous five years. Because we use data for 11 countries, we report wild bootstrapped confidence intervals and p-values to account for clustering at the country-level.8

Fig. 3 presents the event study coefficients using as outcome the log of search intensity. There is a clear break in searches starting the week of the lockdown. There is a raise in search intensity that peaks at about seven weeks into the lockdown. The effect remains statistically different from zero for 10 weeks after lockdown (keep in mind the average lockdown length is 121 days or 17 weeks). The point estimates during the first 10 weeks of lockdown indicate an increase in search intensity ranging from 25% to 85% relative to the week prior to lockdown. Results using levels instead of logs are virtually the same (see Fig. A.1 in Appendix).

Fig. 3.

Event study - All countries.

Notes: The figure above shows event study coefficients from Eq. (1) where the outcome is the (log of) search intensity at the country-by-week level. Country-by-year and country-by-week fixed effects are included. The vertical lines for each coefficient show 95% confidence intervals, cluster corrected at the country level using the wild bootstrap. The omitted week is the week before the lockdown took place in each country. We use the type 2 lockdown as defined in Hale et al. (2020).

Fig. A.1.

Event study - All countries.

Notes: The figure above shows event study coefficients from Eq. (1) where the outcome is the search intensity at the country-by-week level. Country-by-year and country-by-month fixed effects are included. The vertical lines for each coefficient show 95% confidence intervals, cluster corrected at the country level using the wild bootstrap. The omitted week is the week before the lockdown took place in each country. We use the type 2 lockdown as defined in Hale et al. (2020).

The event study results provide evidence that trends in 2020 were similar to those of the previous five years in the pre-lockdown weeks. There was a marked divergence of trends coinciding with the time each lockdown was imposed and mobility patterns changed towards more time at home.

As a placebo check, we repeat the event study analysis using 2019 (period with no lockdown measures) as treatment year and 2016/2018 as controls. In other words, we assign the lockdown to take place one year (52 weeks) before it actually happened in each country. As expected, Fig. A.2 in Appendix shows that there is no statistically significant change in search behavior after the imputed week of lockdown 2019.

Fig. A.2.

Event study - Placebo test using 2019 as treatment year.

Notes: The figure above shows event study coefficients from Eq. (1) a placebo test where we use 2019 as treatment year instead of 2020. Country-by-year and country-by-month fixed effects are included. The vertical lines for each coefficient show 95% confidence intervals, cluster corrected at the country level using the wild bootstrap. The omitted week is the week before the lockdown took place in each country in 2020 but assigned to 2019 instead. We use the type 2 lockdown as defined in Hale et al. (2020). The sample contains search data from January 2016 through November 2019.

4. Difference-in-difference model

4.1. Lockdown – extensive margin

To quantify average effects, we estimate a difference-in-differences model comparing search intensity in 2020 (treated) and the previous five years (controls), between periods with and without stay-at-home orders (lockdown). We estimate the following equation:

$$Log\left(SearchIntensit{y}_{ctwy}\right)=\beta Lockdow{n}_t+{\lambda }_{cy}+{\varphi }_{cw}+{ϵ}_{ctwy},$$ (2)

where $Lockdow{n}_t$ is an indicator that equals one if the week $t$ is in year 2020 and during the lockdown period. The coefficient of interest is $\beta$, which can be interpreted as the overall lockdown increase in search intensity compared to those same weeks in prior years. We include the same set of rich fixed effects as in Eq. (1). The identifying assumption is a parallel trends assumption. We must assume that, if the lockdown and social distancing had not occurred, the –country-specific– weekly DV search intensity would have followed the same trend after the lockdown-week in 2020 as it did after that same week in the prior years (2016 to 2019). Because we already observed in the event study analysis that the effect is stronger during the first months after lockdown starts, we include one specification in which we only use data up to twelve weeks from the moment stay-at-home orders are introduced.

Table 1 presents difference-in-differences results. In column (1) we include the simple correlation between the log of search intensity and our indicator for the lockdown period. The partial correlation estimate suggests there was, on average, a 35% raise in search intensity related to domestic violence in each country every week during lockdown relative to other periods. Once we add fixed effects for country-by-year and country-by-week (Column 2) the effect is slightly smaller (30%) but still statistically significant at the one percent level.

Table 1.

Changes in (log of) search intensity by lockdown.

	(1)	(2)	(3)	(4)
Lockdown	0.349	0.299	0.393	-0.030
	(0.000)	(0.000)	(0.000)	(0.315)
	[0.221, 0.467]	[0.194, 0.417]	[0.276, 0.509]	[-0.0881, 0.0301]
$N$	2746	2746	2517	2176
r2	0.473	0.726	0.730	0.741
Mean dep.	39.29	39.29	38.91	37.58
Country FE	Yes	Yes	Yes	Yes
Country x Year FE		Yes	Yes	Yes
Country x Week FE		Yes	Yes	Yes
First 12 weeks			Yes	Yes
Placebo (2019)				Yes

Notes: Observation at the country-by-week level for 11 countries, from January 2016 to November 2020. The outcome is the log of google search intensity related to the ‘domestic violence’ topic. Column (1) includes fixed effects by country, while column (2) adds fixed effects for country-by-year and country-by-week. Column (3) uses only observations up to twelve weeks from the introduction of stay-at-home orders. Column (4) is a placebo test using 2019 as the treatment year. 95% confidence intervals from wild bootstrapped standard errors corrected for clustering at the country-level are reported in brackets, with the associated p-value in parentheses.

In column 3 we repeat the exercise using data only for the first twelve weeks from lockdown. In this specification, search intensity went up by an average 39% during lockdown. Results using levels instead of logs are virtually the same (see Table A.2 in Appendix). Finally, in column 4 we perform a placebo test assigning to each country the beginning of the lockdown in 2020 to the same week in 2019. As we can see, the point estimate is not statistically nor economically significant.

Table A.2.

Changes in search intensity by lockdown.

	(1)	(2)	(3)	(4)
Lockdown	17.517	14.087	19.912	-0.849
	(0.000)	(0.000)	(0.000)	(0.499)
	[12.67, 22.55]	[9.932, 18.94]	[14.81, 25.11]	[-3.496, 1.841]
$N$	2783	2783	2554	2211
r2	0.463	0.709	0.718	0.723
Mean dep.	38.77	38.77	38.35	36.98
Country FE	Yes	Yes	Yes	Yes
Country x Year FE		Yes	Yes	Yes
Country x Week FE		Yes	Yes	Yes
First 12 weeks			Yes	Yes
Placebo (2019)				Yes

Notes: Observation at the country-by-week level for 11 countries, from January 2016 to November 2020. The outcome is the google search intensity related to the ‘domestic violence’ topic. Column (1) includes fixed effects by country, while column (2) adds fixed effects for country-by-year and country-by-week. Column (3) uses only observations up to twelve weeks from the introduction of stay-at-home orders. Column (4) is a placebo test using 2019 as the treatment year. 95% confidence intervals from wild bootstrapped standard errors corrected for clustering at the country-level are reported in brackets, with the associated p-value in parentheses.

As a robustness check, we re-estimate Eq. (2) excluding one country at a time, to make sure no one country is driving the results. Fig. 4 shows that these estimates are very stable across regressions.

Fig. 4.

Effect of lockdown - Leave out one country at a time.

Notes: The figure above shows coefficient estimates for $Lockdown$ from the difference-in-differences model in Eq. (2) where the outcome is the (log of) search intensity at the country-by-week level. Country-by-year and country-by-week fixed effects are included. Each regression excludes one country. Country-by-year and country-by-week of year fixed effects are included. Wild bootstrapped standard errors are corrected for clustering at the country-level.

4.2. Mobility – intensive margin

We repeat our analysis using as treatment the google’s residential mobility measure instead of lockdown. We will test two models, one with the continuous measure and a second one using a dummy treatment equal to one for weeks when the residential mobility is above a one standard deviation. Different from Eq. (2), here the only controls are country and week fixed effects, since the mobility data is only available since mid-February 2020.

Table 2 presents difference-in-differences results using the continuous and discrete mobility measures in Panels A and B, respectively. Each column corresponds to a different specification, and the final column uses only observations up to twelve weeks after lockdown. We observe a close to 20% increase in search intensity for every one standard deviation increase in residential mobility. A similar pattern is observed when we use an indicator variable for mobility above one standard deviation instead of a continuous one -although coefficient estimates are less stale. As before, we re-run our estimates excluding one country at a time. Fig. 5 shows that these estimates are very stable across regressions.

Table 2.

Changes in (log of) search intensity by mobility.

	(1)	(2)	(3)	(4)
Panel (A)
Residential Mob.	0.238	0.183	0.204	0.148
	(0.002)	(0.000)	(0.008)	(0.033)
	[0.158, 0.307]	[0.133, 0.233]	[0.110, 0.352]	[0.0156, 0.266]
Panel (B)
Mob $>1$ sd	0.522	0.335	0.254	0.206
	(0.002)	(0.000)	(0.006)	(0.182)
	[0.284, 0.745]	[0.197, 0.475]	[0.0885, 0.390]	[-0.125, 0.362]
Observations	428	428	428	199
r2	0.227	0.645	0.703	0.663
Mean dep.	48.22	48.22	48.22	53.67
Country FE		Yes	Yes	Yes
Week FE			Yes	Yes
First 12 weeks				Yes

Notes: Observation at the country-by-week level for 11 countries, from February to November 2020. The outcome is the log of google search intensity related to the ‘domestic violence’ topic. Each column corresponds to a different specification, until column (4) which only includes observations in the first six months of the calendar year. 95% confidence intervals from wild bootstrapped standard errors corrected for clustering at the country-level are reported in brackets, with the associated p-value in parentheses.

Fig. 5.

Effect of mobility - Leave out one country at a time.

Notes: The figure above shows coefficient estimates for $ResidentialMobility$ from the difference-in-differences model in Eq. (2) where the outcome is the (log of) search intensity at the country-by-week level. Country and week-of-the-year fixed effects are included. Each regression excludes one country. Country-by-year and country-by-week of year fixed effects are included. Wild bootstrapped standard errors are corrected for clustering at the country-level.

One may wonder whether the raise in DV related google searches is due to worse economic conditions or because the stay-at-home orders forced victims to be locked up in the same house with potential abusers. Unfortunately, we lack a proper experiment where some people suffer from one of these factors but not the other. Still, we can look at whether the timing and correlations of either factor is closer and stronger to the raise in searches. Fig. A.3 in the Appendix shows averages of searches, mobility and unemployment rate for the eight OECD countries in our sample. It shows that the peak of both unemployment and mobility take place in the first month after lockdown starts, as well as the peak of searches. However, while both mobility and searches start decreasing immediately and go back to previous levels by month five after lockdown, unemployment remains above 30% of previous levels even after seven months of lockdown. Finally, we also perform a ‘horse race’ between both mobility and unemployment in a regression setup. Table A.3 shows the results of regressing search intensity on residential mobility (column 1), unemployment rate (column 2), and both (column 3). As we can see, the coefficient on unemployment is one-third the size of that of mobility and not statistically significant. Although not conclusive, this is suggestive evidence that the effect is mainly due to the physical lockdown.9

Fig. A.3.

Trends in online searches, residential mobility and unemployment.

Notes: The figure above plots the average number of monthly search intensity of domestic violence related topics across 8 OECD countries by month since lockdown in 2020 (bold black line) and previous four years (grey lines). The blue curve reports Google residential mobility data, while the green line corresponds to the yearly-change in unemployment rate. The vertical red dashed line corresponds to the month of the introduction of lockdown type 2 —as defined in Hale et al. (2020)— for each country. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table A.3.

Changes in (log of) search intensity by lockdown.

	(1)	(2)	(3)
Residential mob.	0.267		0.250
	(0.000)		(0.008)
	[0.211, 0.299]		[0.186, 0.301]
Unemployment rate		0.121	0.076
		(0.352)	(0.500)
		[-1.050, 1.169]	[-0.678, 0.959]
Observations	80	80	80
r2	0.330	0.122	0.355
Mean dep.	44.13	44.13	44.13

Notes: Observation at the country-by-month level for the 8 OECD countries in our sample, from February to November 2020. The outcome is the log of google search intensity related to the ‘domestic violence’ topic. Column (1) repeats the main model using residential mobility. Column (2) uses unemployment rate instead. Column (3) includes both regressors. All models include additive fixed effects for country and month. 95% confidence intervals from wild bootstrapped standard errors corrected for clustering at the country-level are reported in brackets, with the associated p-value in parentheses.

4.3. Heterogeneity by lockdown intensity

As we see in Fig. 2, there is heterogeneity in the effect of lockdown on mobility across our sample of countries. We repeat the analysis separating the sample in two groups: developed (Europe and the US) and developing countries (Latin America). These two groups are substantially different in the level of income of their inhabitants and the government capacity to alleviate income losses due to lockdown measures. Table 3 presents the results. Columns one and four show the effect of lockdown on search intensity for Europe and America respectively. The effect of the lockdown in richer countries is more than twice as strong as in the Latin American countries.

Table 3.

Heterogenous effects of lockdowns.

	Europe+US			Latin America
	(1)	(2)	(3)	(4)	(5)	(6)
	Search Intensity	Resid. Mobility	Search Intensity	Search Intensity	Resid. Mobility	Search Intensity
Lockdown	0.411	0.310		0.191	0.183
	(0.000)	(0.125)		(0.000)	(0.563)
	[0.221, 0.637]	[-0.0614, 0.519]		[0.0955, 0.262]	[-0.362, 0.613]
Res. Mob.			0.133			0.082
			(0.125)			(0.000)
			[-0.0504, 0.270]			[0.0132, 0.446]
$N$	1515	234	234	1231	195	194
Mean dep.	31.89	38.04	38.04	48.41	60.18	60.49
Country FE	Yes	Yes	Yes	Yes	Yes	Yes
Country+Week FE	Yes	Yes	Yes	Yes	Yes	Yes
Country x Year FE	Yes			Yes
Country x Week FE	Yes			Yes

Notes: Observation at the country-by-week level for 11 countries. The first three columns use European countries plus the US, while the last two use Latin-American countries only. 95% confidence intervals from wild bootstrapped standard errors corrected for clustering at the country-level are reported in brackets, with the associated p-value in parentheses.

This could be the result of a higher response of the violence perpetrators to the stay-at-home measures in developed countries, a differential in compliance to these measures across groups, or a combination of both.10 In this second scenario, we would observe a lower response on mobility to lockdown. This is what columns two and five show. While residential mobility goes up by one-third of a standard deviation in Europe, it only increases by less than one-fifth of a standard deviation in Latin America, on average. Finally, columns three and six show that the effect of a one standard deviation change in mobility is similar between the two groups of countries, and only statistically significant for the Latin American countries.

5. Online-search intensity vs helpline calls

Finally, we provide evidence that during lockdown both search intensity related to DV topic and calls to a DV helpline observed an expected increase with respect to previous years. We use data for Spain, for which we have monthly data on number of calls to the DV helpline (016) from 2015 up to the first semester of 2020.

The analysis follows the spirit of Eq. (1), but at the monthly level and using only one country. Controls now include month and year fixed effects. Fig. 6 presents the event study coefficients using as outcome the log of search intensity and the log of DV calls in panels (a) and (b), respectively.

Fig. 6.

Event study analysis of DV online-search and calls.

Notes: The figures above show event study coefficients from Eq. (1) for the case of Spain. In 6 a the outcome is the (log of) search intensity, while in 6 b the outcome is the (log of) DV calls, both at the monthly level. Year and month fixed effects are included. The vertical lines for each coefficient show 95% confidence intervals. The lockdown in Spain started in March, hence the omitted month is February, and is denoted by the shaded area. We use the type 2 lockdown as defined in Hale et al. (2020). The sample contains data from January 2016 through November 2020.

The two figures exhibit a remarkably similar pattern, with no difference with previous years in January, and a bell-shaped effect starting March and ending around June. However, although the peak for the effect of DV calls is 50%, it reaches about 80% for google searches.11 Another difference is that the effect on search intensity drops and is not statistically significant starting June, while it remains significant for DV calls. The higher peak in searches can be explained by a selection of women who, after searching, actually call the number. Meanwhile, the shorter duration of the effect on search intensity may be due to the innate nature of the search, which is most likely related to look for information about how to make a complain. Once the woman obtains this information, she does not need to google it again in future cases. However, DV call would still show if she calls again.

Although this analysis uses data for only one country in our sample, the strikingly similar behavior and timing of the two series reassures us that search intensity related to DV topics are a good tracker of DV cases during the lockdown.

6. Conclusions

Domestic violence is a global public health problem with large social and economic costs. DV incidence can be further exacerbated during times of crisis, high unemployment and social-stress (like the COVID-19 pandemic). However, most countries lack the necessary information to implement rapid public polices to contain the situation. Timely administrative data from police reports and DV service calls, when available, suffer from substantial under-reporting.12 On the other hand, survey data collecting information on DV incidence are rarely speedily available and are costly. This paper proposes the use of already publicly available data on google searches to track the incidence of DV. This data presents important advantages: it is free, timely, available at daily frequency, and allows for comparisons across different geographical areas. Furthermore, Google-trends data provides information about DV episodes that are not usually reported in surveys or administrative records.

Using a panel of eleven countries and five years of google search data on DV related topics, we find an average increase in searches of about 31% after stay-at-home orders are put in place. Furthermore, using data on DV service calls for Spain, we find a similar although slightly smaller effect of lockdown. Finally, we show evidence that the effect is stronger in countries with a stricter lockdown. We remain agnostic as to whether these results driven by economic activity, time of exposure to potential perpetrator, or other factors.

This online search index has, at least, one important limitation: it requires that victims have access to internet. This means that our findings, if anything, are a lower bound, since previous work has shown that DV is more prevalent among poor and low-educated women (Aizer, 2010), and we expect them to also be more affected by the lockdown.

All in all, we believe that online searches can be an extra source of information for governments and complement existing data -specially in places where other sources of information are not available. As the COVID-19 pandemic is still hitting hard in many countries and mobility restrictions are increasing in several nations, our results call for more policies to particularly protect domestic violence victims, reinforcing the communication channels to seek help and accelerating processes to release victims.

Appendix A

 

Table A.1.

Dates of lockdown for countries in our sample.

	Lockdown
Country	Start	End	Length (in days)
Argentina	19/03/2020		242
Brazil	05/05/2020		195
Chile	25/03/2020		236
Colombia	25/03/2020	31/08/2020	159
France	17/03/2020	10/05/2020	54
Germany	21/03/2020	05/05/2020	45
Italy	23/02/2020	03/05/2020	70
Mexico	30/03/2020	10/09/2020	164
Spain	14/03/2020	26/05/2020	73
United Kingdom	22/03/2020	12/05/2020	51
United States	15/03/2020	19/07/2020	126

Notes: Country dates for mandated Stay-at-Home orders obtained from Hale et al. (2020), using lockdown type 2 definition.

Supplementary material

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.euroecorev.2021.103775

Appendix B. Supplementary materials

Supplementary Data S1

Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/