Intimate partner violence and women’s health: the private and social burden

ABSTRACT

We assess the impact of intimate partner violence (IPV) against women on their health and healthcare use, finding that IPV substantially worsens health, and increases the hazards of hospitalization, emergency care, and sedative consumption. We exploit two compatible samples from the 2011 Spanish Violence Against Women and National Health surveys. We estimate the effect of IPV on health using a bivariate model that exploits woman’s awareness of IPV among female acquaintances to account for IPV endogeneity, and the effect of IPV on healthcare through a two-stage procedure. Hence, IPV originates high private cost, but also public, by draining healthcare resources.

KEYWORDS:

• Partner abuse
• health
• gender inequality
• medical care costs
• ordered response
• endogeneity

JEL CLASSIFICATION:

• J12
• J16
• I14
• D19
• C24
• C25
• C26
• C35
• C36

Acknowledgment

We thank Maria Errea, Ernesto Villanueva and the audiences at the Workshop on Economic Analysis of Intimate Partner Violence, Royal Holloway London 2017, XXXIX Jornadas AES 2019, and the Economics Seminar at UPNA for their comments. We are also grateful to Margarita Bris, Javier Macas, Amparo Reyes and Juan M. Rodriguez-Poo for their help in providing key data. Research funding from the Ministry of Science, Innovation and Universities, Grant Nos. ECO2015-65204-P, RTI2018-095231-B-I00, MDM2014-0431, and Comunidad de Madrid, Grant No. MadEco-CM (S2015/HUM-3444), is gratefully acknowledged.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ Declaration on the Elimination of Violence against Women, UN General Assembly resolution A/RES/48/104, December, 1993.

² An exception is Agüero (2017), who studies the effect that violence against women has on the health outcomes of their children using a partial identification method to account for possible bias due to omitted variables.

³ A different strand of the literature evaluates the causal effect of certain policy changes on women’s behaviour related to IPV (see Rice and Vall Castelló 2018).

⁴ According to the Special Eurobarometer from the European Commission (2021), $44\mathrm{\%}$ respondents disagree with the idea that acts of gender-based violence should be considered criminal actions.

⁵ The 2014 and 2019 VAWS also report information on women’s health status, but they lack information about woman’s awareness of IPV among her acquaintances.

⁶ We have restrict the sample to women cohabiting with a partner in the last 12 months because cohabiting is a feature of the partner definition in the survey, so that relevant partner information exploited in the analysis is ony available for cohabiting partners.Of course, partner violence can importantly arise from non-cohabiting partners too. But in addition to the aforementioned data limitations, the underlying behavioural models characterizing cohabiting and non-cohabiting partners might strongly differ. Focusing on cohabiting couples is quite standard in this literature (see e.g. Bobonis, González-Brenes, and Castro 2013).

With regard to our age selection criterion, on the one hand labour market situations alternative to inactivity are not relevant for most women over 65, and a large fraction of women over 65 do not currently have a partner. On the other hand, women of at least 25 years old are chosen so as to ensure that they have completed their education.

⁷ Physical and non physical abuse are not mutually exclusive.

⁸ The Spanish VAWS only considers current abuse, which corresponds to any of the situations involving abuse in the last 12 months at the time of the survey. However, past experiences of abuse that ended at least 12 months before the time of the survey are not recorded.

⁹ The remaining motives for an overnight stay at hospital were surgery, diagnostic testing, and hospital treatment.

¹⁰ Cutler and Richardson (1998) also use univariate ordered response models to examine the relationship between different types of disease and self-reported health status. Other examples are Theodossiou (1998) or Chaloupka and Wechsler (1997).

¹¹ Because $IPV$ follows a probit model, it is tempting to try to mimic 2SLS where $IPV$ is replaced with its probit fitted values. But this approach does not produce consistent estimators because the conditional expectations operator cannot be carried through nonlinear functions. Forbidden regressions produce consistent estimates only under very restrictive assumptions, which rarely hold in practice (see for instance Wooldridge 2010, Chapter 15).

¹² Regarding physical assault or abuse, the 2006 NHS poses the generic question “In the last year, have you experienced some type or assault or abuse?”, while the question for non physical abuse is “In the last year, have you experienced discrimination, been impeded from doing something, offended or disregarded because of your gender, origin, education, social class, sex orientation or beliefs?”. Unlike the VAWS, these questions do not fulfill the gold standards methods, so we should expect a much lower level of disclosure, as shown by the low number of affirmative answers.

¹³ Looking for other sources of exogenous variation, among the candidates as instrumental variables for IPV we can mention province-level information on IPV related variables, such as the number of deaths due to IPV or the number of IPV court complaints. However, given that our main model already accounts for geographical effects using the information on the province of residence, these type of instruments are not relevant, as any difference at the province level is already controlled for geographical dummies.

¹⁴ As we do not have any measure of household income, we would expect variables for the woman and her partner’s education to capture partly both individual and household socioeconomic status.

¹⁵ Recall that the IPV measure used in the estimations considers whether or not the woman has experienced some episode of serious abuse in the last 12 months at the time of the survey, irrespective of when such situation started. In order to assess the sensitivity of the results to the definition of this variable, we have considered an alternative measure, which takes on value one if the woman has experienced some episode of serious abuse in the last 12 months, provided that such situation started more than one year ago, and zero otherwise. With this measure, we aim at the effect of lengthier situations of IPV. Qualitatively, the results are similar with both IPV measures. However, the absolute values of the estimated coefficient and the marginal effects of IPV are smaller in magnitude, and are estimated with lower precision than with our original measure. The fact that, under this alternative measure, IPV is set to zero for women reporting IPV that started less than one year ago, is likely to be behind this loss of precision.

¹⁶ Other papers using this approach to check the sample compatibility are Lusardi (1996), Jappelli, Pischke, and Souleles (1998), or Battistin, Miniaci, and Weber (2003), among others.

¹⁷ These estimates are available upon request.

¹⁸ Another way to interpret the marginal effect of $\widehat{H{S}^{\ast }}$ on the probability of healthcare use, is to rescale $\widehat{H{S}^{\ast }}$ into an equivalent normalized measure that takes on values within the $[0,1]$ interval, so that we can interpret the marginal effect in terms of percentage point changes in the continuous health index (see for instance, van Doorslaer and Jones 2003). By rescaling $\widehat{H{S}^{\ast }}$ using the following transformation.${\widehat{H{S}^{\ast }}}^R=\frac{\widehat{HS\ast }-min\left(\widehat{HS\ast }\right)}{\left[max\left(\widehat{HS\ast }\right)-min\left(\widehat{HS\ast }\right)\right]}$

we obtain that 1 pp increase in the normalized health index decreases the probability of hospitalization, emergency care use, and sedative consumption by $0.16$, $0.42$, and $0.50$ percentage points, respectively.

¹⁹ In 2010 the figures are similar to those for 2011: the total expenditure was 99,899 and the expenditure per capita was 2149 euros.

²⁰ Other air pollutants, such as greenhouse gases and nitrogen and sulphur oxides, nitrogen carbons and hydrocarbons, have been disregarded. Their direct consequences on human health are generally less harmful than those caused by the pollutants considered. Furthermore, in some cases, such as greenhouse gases and nitrogen and sulphur oxides, industrial facilities are not usually the major source of release, as motor vehicles play a major role at a very local level. Unfortunately, information about pollutant emissions of these substances from non industrial sources is not available at the postcode or even at the municipality level.

²¹ There are also differences in terms of the classification of the municipalities, which are the ultimate sampling units. In particular, although the number of strata is the same in both cases, there are differences in their size.

Additional information

Funding

The work was supported by the Consejería de Educación e Investigación [MadEco-CM (S2015/HUM-3444)]; Ministerio de Educación y Cultura [RTI2018-095231-B-I00].