4.1. Data collection and sample

As noted, the PANGEA data was gathered for inter-disciplinary, mixed-method research. Quantitative data was collected through a large cross-sectional survey with retrospective questions targeting the female adult population domiciled in the region of Tuscany and living with their partners during lockdowns.

To complement the quantitative findings and enrich their interpretation, qualitative data was collected by re-interviewing (willing) interviewees having reported abuse during lockdowns as well as by extending the survey to abused women who had sought shelter in anti-violence centres in Tuscany and had not been part of the survey first round. The findings from qualitative analysis are discussed at length elsewhere [50] but are selectively recalled here where relevant.

The data collection took place in three rounds: from August 17^th to September 18^th, 2021 (3000 interviews); from November 22nd to December 17^th, 2021 (743 interviews); and from November 5^th, 2022 to May 30^th, 2023 (35 interviews in anti-violence centres). In the first round of data collection, the interviewers administered the questionnaire by Computer-Assisted Telephone Interview (CATI). In the second round, a combination of CATI and Computer-Assisted Mobile Interview (CAMI) was used to reduce the bias of under-representation in the sample of those not appearing in telephone directories and those who do not have landline phones (typically young women). The third round oversampled actual victims by interviewing 35 women who had suffered from intimate partner violence during lockdowns and had sought shelter in anti-violence centres in Tuscany. These interviews were based on a semi-structured questionnaire that added open-ended questions to the closed questions of the CATI questionnaire. The oversampling technique is recommended when the group of interest tends to elude standard sample surveys on the general population [62].

The PANGEA survey questionnaire included 45 closed-ended questions (see S1 Table for list). The entire project, including the questionnaire, was approved by the Comitato per l’Etica Clinica dell’Azienda Ospedaliero – Universitaria Senese (COMEC) of the University of Siena. Before the interview started, respondents were informed that their data was strictly confidential, and that privacy and anonymity were guaranteed. All interviews were conducted in Italian by selected and trained interviewers by Numeria Srl, the company that carried out the survey. Interviews began only after obtaining informed verbal consent from each participant. This consent was documented through authorised audio recordings made and stored by the interviewers. All the data was anonymised by Numeria, so the authors did not have access to any information that could identify the women interviewed during or after data collection. The terms of the PANGEA project granted access to the anonymised data to the authors and other members of the project with no expiration. Therefore, we accessed the data for this research multiple times from September 15^th, 2021, until March 31st, 2026.

We developed the questions using the classification and language of the National Statistical Institute ‘Women’s Safety Survey’ [56,63], the European Fundamental Rights Agency Survey on women’s well-being and safety [64], and the EU survey on gender-based violence against women and other forms of inter-personal violence (EU-GBV) [16]. Accordingly, we classified violence against women into three types – psychological, physical, and sexual – with an added subdivision between COVID-related psychological violence and general psychological violence. Table S1 in S1 Appendix lists the individual acts of violence in each type. To address the missing data from variables with no response, we used stochastic imputation methods. Table S3 in S1 Appendix provides further details about the imputation.

The sampling strategy hypothesized a 10% IPV prevalence, a figure broadly in line with previous estimates (see previous section). The quantitative survey contacted 35709 women, 3778 of whom agreed to be interviewed. The sample from the first two rounds of 2021 (3743 observations) underwent filtering and qualitative analysis of the interviews which led to the validation of 3599 observations. After adding the 35 observations from the CAVs, the final available sample comprised 3634 women between 18 and 75 years of age. Although the response rate was modest, representativeness was ensured through careful stratified sampling and post-survey weighting. Specifically, sampling was stratified by province and implicitly by municipality size to ensure geographical representativeness. Subsequently, we weighted all 3599 observations from the first two rounds according to the official age distribution of women in Tuscany across five age groups (18–34, 35–44, 45–54, 55–64, 65–75) and the proportional female population across the ten Tuscan provinces. Finally, since the additional observations in the third round had an equal probability of experiencing violence as in the prior rounds, we redistributed the weights of those who experienced violence proportionally, to ensure that the relationship between violence and non-violence remained unchanged. Therefore, the sample is representative of the population in terms of age distribution among women and their geographical distribution across provinces.

For the purposes of this paper, we restricted the sample to working-age women (18–64) who are not pensioners, as our research focus is specifically on intimate partner violence during working age. This restriction resulted in a working sample comprising 2061 observations.

4.2. Measurement and overview

IPV measures and findings. As anticipated, a new feature of this study is the way IPV outcomes are measured. We used two measures, the conventional prevalence of violence whereby women are considered abused if they suffer a single act of violence, irrespective of intensity and severity, and a fuzzy index which also accounts for severity and intensity. This index was derived by [19] and exploits fuzzy logic under the following premises. First, intensity is gauged by the frequency with which each abuse occurs. Second, severity is understood as socially perceived severity and is proxied by the inverse probability of occurrence of a certain abusive act, based on the assumption that the effort societies put into fighting adverse outcomes is broadly commensurate with the perceived severity of these outcomes. This assumption is discussed at length by the proponents of the index who also validated the resulting severity scale by comparing it with existing scales, usually much more costly to compute and less fine grained. Third, acts of violence are ranked by severity only if they pertain to the same type of violence (e.g., psychological or physical or sexual), since ranking across types is problematic. Fourth, since certain acts of violence tend to occur together, the severity scale based on inverse prevalence is corrected for correlation between acts of violence in order to counter problems of measurement and avoid redundancy. Under these premises, fuzzy logic can be used to combine prevalence, intensity, and severity into type-specific indexes that are computed for each individual and can be aggregated across individuals. Further aggregation of type-specific indexes into an overall index can be controversial because of the difficulty of comparing severity across violence types. The most transparent option, and the one we adopt in this paper, is to take a simple average, which is equivalent to abstaining from ranking violence types in terms of severity, i.e., physical violence is not viewed as more severe, per se, than other types as it all depends on the specific abuse that is perpetrated.

[19] formalized the construction of the index, and applied it to the first Europe-wide survey of violence against women carried out by the Fundamental Rights Agency (FRA) [64]. A Stata code to actually implement the index is available on GitHub [65].

The fuzzy index we present here uses the severity scale obtained by [19]. Given sufficient cultural homogeneity between the population of Tuscany and that of the European Union as a whole, using a scale which is based on a very large sample has clear advantages. However, since differences between the FRA questionnaire and our own required some amendments, we report the revised scale in S1 Appendix. Moreover, the present index proxies intensity with multiplicity of abuses rather than frequency, i.e., with the number of different abuses of a given type the woman reported. This is because the PANGEA questionnaire traded off investigation of mere frequency with that of perceived, overall change in IPV (see Section 6).

Table 1 summarizes the evidence from our sample on the distribution of IPV by types in terms of prevalence and fuzzy index values. Average prevalence refers to the whole population while average fuzzy index values are also reported for the subgroup of abused women. Of the 2061 women in our working sample, approximately 150 experienced abuse at least once from their partner during lockdowns. After applying sampling weights (to correct for oversampling and to account for the age and geographical distribution of women, as noted), the overall prevalence is 5.7%, which corresponds to approximately 60,245 women in the population (considering the total population of women between 18 and 64 minus the recipients of pension: data available at Istat [66]). Psychological violence was the most widespread as it affected nearly all self-reported victims (5.4% prevalence). Physical violence affected less than half the victims (2.1% prevalence) while sexual violence was reported by 1.3% of women. Lockdown conditions made it easier for the partner to exercise old and new forms of control, like preventing the woman from working, contacting friends and family and even contacting doctors, clinics and hospitals, and more generally from leaving home (as far as allowed by confinement measures). Nearly 40% of the victims reported suffering at least one such form of control.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Prevalence and Fuzzy index of IPV during lockdown, by type.

https://doi.org/10.1371/journal.pone.0349889.t001

The ranking among IPV types stays the same when the fuzzy index is computed for the entire female population of working age women while reversing when computed only for women having experienced abuse. In this latter case, the index is not influenced by prevalence and gauges severity and intensity of abuses: its value reaches 0.47 for sexual violence (out of 1) followed by physical (0.36) and psychological violence (0.26). The overall fuzzy index value for abused women (0.18) is lower than any type-specific value since not all victims suffered from all types of violence.

How do the above figures compare with those from other studies on IPV during lockdown? Concerning prevalence, there is unfortunately no exhaustive answer. According to the 2014 ISTAT estimates for Tuscany we reported earlier [59], 12 months prevalence before the pandemic was lower than our estimates during lockdowns for sexual and physical violence but higher for psychological violence. However, in view of the numerous differences between our questionnaire and that used by ISTAT in 2014 (e.g., age groups or inclusion of non-partners among potential perpetrators) a before – after COVID comparison is problematic, except perhaps for the indication that physical and sexual violence are likely to have increased during the pandemic.

As concerns international comparisons, only a minority of studies based on telephone surveys can be taken as reference, especially those for European countries. [67] conducted a telephone survey in January 2021 on a representative sample of 1541 Portuguese women to investigate episodes of violence since March 2020 and found that 4.9% of respondents experienced violence for the first time in that period. This is a lower figure than ours, but the difference is easily explained by the exclusion of women with a previous history of violence as well as by a much shorter questionnaire in the Portuguese survey. A higher figure (9.4%) than ours is instead reported by [68] based on a large, nationally representative telephone survey carried out in Germany between February and March 2021 to interrogate women about their experience in the previous 12 months. Overall, however, the above differences are still small with respect to those resulting from online surveys. For example, an online survey for Portugal [69] found a 13.7% prevalence of domestic violence, mostly psychological, between April and October 2020. Back to Germany, a large online survey found an overall prevalence of 8% but for a much shorter reference period (April to May 2020) [38]. One factor that may have amplified differences between online and telephone surveys is that recent experiences of violence tend to be associated with a higher reporting rate [70] and online surveys were often conducted during or soon after the worst lockdown episodes. Moreover, because online surveys are self-administered, they often provide a sense of anonymity that encourages more confident disclosure compared to telephone surveys where the presence of an interviewer can trigger social desirability bias [71]. Additionally, online surveys typically have lower rates of item nonresponse than telephone surveys [72].

On balance, we cannot rule out underestimation of psychological violence by our survey. While there is still insufficient knowledge about the influence that different methods of data collection may have on IPV recall bias [73], our qualitative evidence suggests that psychological abuse may have been comparatively underreported in the PANGEA survey. The qualitative investigation we conducted by re-interviewing women with experience of abuse during the lockdown revealed a tendency for some of them to minimize and even disavow ‘lighter’ forms of abuse that they had previously acknowledged [50]. This suggests that recollection of psychological abuse months after it occurred may be less accurate than that of physical or sexual abuse.

Victim’s profile. The main demographic and socioeconomic characteristics of the women in our sample are set out in detail in S1 Appendix (Table S2) distinguishing between those who experienced at least one instance of abuse (V) and those who did not (NV). The former were, on average, younger (37.7% less than 45 years old compared with 50.4% for non-abused women), disproportionately of foreign citizenship at birth (18.3% against 4.4%), almost equally educated (i.e., the share with tertiary education is practically the same for the two groups), less likely to have no child living in the household (48.9% against 58.2%), significantly more likely to be unemployed or out of the labour force (39.7% against 23.2%) as well as more likely to live in a household that experienced an economic shock due to job loss for at least one partner during lockdowns (46.6% against 31.0%). Additionally, 28.26% of the victims had lost their job against 21.67% of the non-victims. On average, the partners of abused women were slightly less educated, more likely to be unemployed or non-active and to have lost their jobs during the lockdown.

Typical household characteristics also differed between the two groups of women. Abused women more often reported that it was somewhat difficult or very difficult for the household to make ends meet from one paycheck to another (48.9% against 25.5% for non-abused women). Furthermore, where they lived was more crowded and less likely to have outdoor space. The dummy indicator we constructed to capture crowded living space and lack of privacy, takes value 1 if the household (i) has no outdoor space and (ii) accommodates more than one person per room (excluding bathroom and storage rooms and counting children as half persons). On average, abused women scored higher than non-abused women on this indicator (0.17 against 0.12).

On the whole, this evidence yields a profile for abused women which is broadly familiar in the literature and which we used to define and select the variables for our estimations.

5.1. Instruments’ validity and strength

Given heterogeneous development patterns at the territorial level, and the related occupational segregation, the two instruments we chose reflect local employment opportunities for women in the long and medium run. As such, they are likely to affect a woman’s decision to work, whereas the reverse effect is negligible if only for statistical reasons. Earlier empirical literature has used proxies of area-specific employment opportunities as instruments for women’s labour market outcomes [76,77,83,84]. In principle, local-level indicators like the average employment rate in a given area may also be affected by the kind of gender norms that influence violence in the area. This may seem to contradict the exclusion restriction assumption that a valid instrument requires. However, we are comparing average employment rates within a geographically small and relatively culturally homogeneous area – Tuscany – whereby the extent to which women work in, say, municipality A compared to municipality B is unlikely to reflect significant differences in gender norms between these municipalities.

The results of first stage estimation (Equation 4) are reported in Table S5 of S2 Appendix and provide evidence of the relevance and strength of our combination of instruments. The female employment rate has a negative and significant association with non-employed women, and a negative and non-significant relation with the variation in female employment rates. Thus, both instrumental variables exhibit the expected signs, though only the first shows strong predictive power. Regarding the exclusion restrictions, neither instrument is significantly correlated with our dependent variables (see Table S6 in S2 Appendix). In the same Appendix we further document the strength of our joint instrument set (Table S7) [85].

6.1. Main results

Did IPV worsen? In our questionnaire, psychological violence included items like forbidding the woman to go to the doctor or hospital, which we termed COVID-specific and listed separately (see Table S1). For each non-COVID-specific item, victims were asked whether they had been abused ‘more’ or ‘less’ compared to before the pandemic, or no change had occurred. The ‘more than before’ category includes three possibilities: abuse occurring for the first time, occurring more frequently or occurring more forcefully. Only a small minority of abused women answered ‘less than before’, the percentage ranging from a tiny 3.1% for sexual abuse to 7.3% for psychological abuse. Depending on the type of violence, slightly more or slightly less than half abused women indicated no change (Table 2). Hence, nearly half the women reporting sexual abuse and 37–38% of those reporting psychological or physical violence answered ‘more than before’. This is clear, if subjective, evidence of heightened violence.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Percentage distribution of perceived change in IPV experience by violence type.

https://doi.org/10.1371/journal.pone.0349889.t002

The role of shocks. Table 3 answers our second and third research questions. The table sets out the estimated, average marginal effects of the COVID shocks as well as of the instrumented control variable (women’s employment status). It also reports estimates of the coefficient of generalized residuals in the CF model. From a methodological perspective, the salient finding is that the coefficients of generalized residuals lack significance, which indicates that endogeneity has been controlled for. An additional indication in the same direction is that the CF estimates for the most significant COVID shock (presence of children and job loss) are close to those from probit/fractional regressions (across IPV measures and types). In fact, instrumentation affects the results mainly by reducing the statistical significance of the instrumented variable. This last finding is not new in the literature [77,84,86], and since our interest in the actual role of women’s employment is primarily instrumental, we refer readers to the vast literature on the topic.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Main estimates.

https://doi.org/10.1371/journal.pone.0349889.t003

In reading the table recall that, unlike the CF model, the probit and the fractional models do not control for endogeneity and are used when the dependent variable is measured by prevalence and by the fuzzy index, respectively. From a substantive perspective, the findings reveal a hierarchy of the COVID-specific effects. The presence of children under 18 capturing parental overburden yields the most consistently significant and among the largest effects across different models, measurement and types of IPV. For example, with reference to the CF model estimates, the increase in prevalence amounts to 2.5 pp. for all types of abuse, i.e., a 44% rise with respect to the average predicted value (column 2a, Table 3). Psychological violence is the strongest driver in this case (+2.6 pp.) though the difference with respect to sexual and/or physical violence is small. The results for the fuzzy measure of violence tell a similar story: the estimated increase in the IPV index for all types of violence amounts to 0.014 index points in the CF model, corresponding to a 127% rise with respect to the average predicted value. In this case too, the increase is primarily driven by psychological violence.

The results are mixed in the case of job loss, and they vary with IPV measurement. The estimated effect is almost as strong as for parental overburden when IPV is measured by prevalence: + 2.8 pp. for violence of all types corresponding to a 49% rise. In this case too, the aggregate result is driven by psychological IPV (+2.5 pp.) while the effect for physical/sexual IPV is lower and lacks significance (+1.0 pp). In comparison, significance drops below conventional levels across models and violence types when IPV is measured by the fuzzy index. Finally, confinement in crowded space apparently had no significant effect across models, measures and violence types, although the sign is as expected.

Evidence from the qualitative in-depth interviews of the PANGEA project [50] may throw light on the above results. Some of the victims taking part in qualitative interviews linked their experience of violence during lockdowns with the conflict arising from the need to constantly renegotiate a childcare workload that had become more onerous and complex during confinement. An additional source of conflict they highlighted was the fact that children were perceived by some partners as vehicles of contagion [50].

In contrast, most victims viewed confinement in spaces lacking privacy as an impediment to organizing support from outside rather than a factor conducive to violence. Controlling behaviour was widespread during confinement and these women felt that partners could check on them everywhere, even in the bathroom, no matter how large the living space might have been.

All this may offer some explanation of why parental overburden and confinement in crowded space lacking privacy appear to have had the strongest and the weakest impacts, respectively. In between stands job loss for either partner, which may be viewed as epitomizing COVID-specific economic shocks. An important distinction that would have been lost had we only used prevalence to measure IPV, is that job loss appears to have played a significant role only for less severe IPV abuses. This is indicated by the fact that (i) the increase in overall prevalence was driven by psychological rather than physical or sexual violence, and (ii) when severity of IPV is accounted for (by our fuzzy index) the association with job loss consistently loses significance.

Concerning controls (other than her employment status), and focusing again on the CF estimates, the results for foreign status, young age and education are worth mentioning. The marginal effect of being of foreign citizenship, typically of migrant origin from outside but also from within the EU, is consistently significant across all estimates and of an order of magnitude larger than for any other variable. For example, the average, predicted likelihood of reporting any kind of violence nearly triples for foreign citizens. In comparison, estimated marginal effects are lower for being young (less than 31 years) and living with a low educated partner, although both these variables emerge as significant risk factors across estimates. In all these cases measurement matters since both the level of significance and the size of the marginal effects tend to be higher when severity is accounted for, i.e., for fuzzy outcomes rather than prevalence. Interestingly, her education seems to have been a protective factor since marginal effects for this variable carry a negative sign, although exhibiting weaker significance than for partner’s education in several instances.

Size of municipality, partner’s employment status, and alcohol consumption in the health district of residence rarely or never exhibit statistical significance. In the case of alcohol consumption this may seem surprising given clear indications to the contrary in the literature, but we cannot rule out that the proxy we used for individual alcohol abuse is poor. Also, confinement need not have increased alcohol consumption for all users: as some of the women participating in the qualitative survey pointed out, certain users may have found it more difficult to secure supply during lockdown [50]. Finally, the findings are mixed for the making-ends-meet variable which is discussed in greater depth in the following paragraphs.

6.2. Gender norms as a moderation factor

Was the effect of parental overburden moderated by gender-related attitudes? There is a fair deal of consensus in the literature that authoritarian parenting and resistance to sharing childcare are salient indicators of traditional gender attitudes, especially on the part of the male partner. Adherence to these attitudes may thus have affected the impact of parental overburden on IPV during the pandemic. The PANGEA survey did not directly investigate partners’ views; hence we proxied gender-related attitudes with two of our most impactful control variables, namely male partner’s education and woman’s citizenship. Low education has long been associated with traditional gender values [87], and its moderating role in the occurrence of violence is especially worth investigating for the male partner. Concerning foreign status, our survey only recorded the woman’s citizenship, and we know that the majority of abused women in the survey originated from countries that rank low on international indexes of gender equality (like Honduras, Peru, Afghanistan, Morocco among others) or from former socialist countries where women’s employment tends to be accepted but support for women’s equality remains comparatively low [88]. This evidence is compatible with the possibility that, in comparison to local women, women of foreign citizenship felt or were actually less empowered to diffuse tensions over parenting.

The estimations in Table 3 were thus repeated after adding two terms, respectively the interaction of parental overburden with the woman’s foreign status and with the partner’s education. The first notable result is that our main estimates (the average marginal effects of the COVID shocks reported by Table 3) hardly change with the introduction of the interaction terms. Moreover, both these terms attain statistical significance across specifications for fuzzy IPV outcomes but hardly ever for prevalence IPV, echoing our previous findings on how measurement influences outcomes (Section 6.1). Finally, we gauged the strength of the interaction from the differential marginal effect of parental overburden for women of foreign citizenship compared to local women and for women with low-to-middle educated partner compared to women with a highly educated partner. Both these differential effects are significant at conventional levels for fuzzy IPV outcomes, not for prevalence IPV; also, the differential effect is larger for foreign citizenship (Table 4).

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Moderation of parental overburden effects by foreign status and partner’s education.

https://doi.org/10.1371/journal.pone.0349889.t004

A plausible indication we would draw from the above findings is that adherence to traditional gender norms on the part of one or both partners may have exacerbated the effect of parental overburden on the intensity and severity of violence rather than on prevalence. However, our use of proxy variables to obtain these results warrants caution in interpreting them.

6.3. Sensitivity and robustness

We carried out a number of tests in order to assess the robustness and sensitivity of our key findings. The first robustness test involved repeating the estimation after dropping the making-ends-meet variable (Table S8, S3 Appendix). This variable gauges the prevalent condition of the household in late 2021, when the economy (GDP) had almost recovered to pre-pandemic levels; as such, it need not have been influenced by job loss during lockdowns. However, it probably captured ‘usual’ economic conditions for that household, since poverty tends to persist in time. Removing this variable serves two purposes, namely verifying that the effects of interest are not affected by endogeneity, and that our third COVID-specific shock, crowding and lack of privacy, does not act as a proxy for poor economic conditions. The results are reassuring since the effects of our ‘shocks’ maintain the same level of statistical significance and very similar values after removing this variable.

Redistributing sampling weights in the third round of our data collection in order to preserve abuse prevalence from previous rounds could have distorted the results. In our second robustness test (Table S9, S3 Appendix) we therefore repeated all estimations after dropping all the observations from the third round. In comparison with our main estimates (Table 3), all marginal effects of parental overburden on prevalence IPV get smaller in size and lack significance across models and violence types in the reduced sample, whereas job loss effects gain both in size and significance across the board. However, results are very different for the marginal effects on fuzzy IPV outcomes. In this case job loss effects lose significance across the board compared to the main estimates (although most gain in magnitude) while those for parental overburden reduce in size but retain significance for aggregate (all types) as well as psychological IPV. Confinement in crowded space continues to exhibit nonsignificant effects.

Arguably, these results are consistent with our main findings in one important respect, namely that measurement by prevalence may lead to overstating the role of factors involved in widespread types of violence while understating the role of factors influencing intensity and/or severity of violence. We therefore see no strong reason to question the hierarchy of effects that can be inferred from the main estimates.

As noted earlier (Sections 3 and 4) we found a lower, overall IPV prevalence figure than expected when designing the PANGEA survey. The small number of IPV cases on which our results rest may thus raise concerns about the statistical power of our analysis. While ex-post power analysis is not advisable, comparing Minimum Detectable Effects (MDEs henceforth) with actual estimates reveals possible power deficiencies [89–91]. We therefore compared all our average estimated marginal effects (AMEs) for the three covariates of interest with the corresponding MDEs computed at 0.80 power and three different significance levels (0.01, 0.05, and 0.10: Table S10, S3 Appendix) The comparison suggests that, at conventional levels (0.80 for power, 0.05 for significance) our estimates of the impact of parental overburden on all types of fuzzy IPV outcomes and on sexual and physical IPV prevalence are adequately powered. In contrast, the estimated AMEs of job loss or confinement in crowded space appear to be underpowered across models and specifications (although it is worth recalling that some exhibit conventional statistical significance, specifically the AMEs of job loss on prevalence IPV). We cannot therefore rule out that small sample size contributed to lack of significance for the effects of confinement in crowded space or to the patchy significance pattern of the job loss effects.

Finally, we carried out sensitivity tests by probing the assumptions about errors in the CF model (see Table S11 and Table S12). The results are reassuring, but we relegate the evidence and the related discussion to S3 Appendix for reasons of length.