Employment and the risk of domestic violence

UNIVERSITY OF THE BASQUE COUNTRY UPV/EHU MASTER THESIS Employment and the risk of domestic violence Author: David Gastorf Supervisor: Arantza Ugidos Olazabal A thesis submitted in fulfillment of the requirements for the degree of Master in Economics: Empirical Applications and Policies September 25, 2022 i Abstract David Gastorf Employment and the risk of domestic violence In this master’s thesis we study the effect on the risk of intimate partner violence (IPV) for women of the emplyoment status of herself, her partner and income together with a set of exogenous sociodemographic factors. In doing so we account for the possible endogeneity of the employment statuses as well as the incomne. We use the most recent data available which originates from the Violence Agains Women survey (VAW) in Spain from 2019. We apply three different estimation methods to study their differences and in order to be able to compare our results to previous studies which used the same approaches. The estimation methods are a linear univariate probability model which we use in order to examine the results without taking possible endogenity into account, a linear two-stage least squares probability model and a non-linear multivariate probability model. Where the latter two models account for endogeneity. Our main findings with respect to the employment statuses are that only the partners employment status plays a major role on reducing the risk of IPV for the woman and only when the woman is also employed and only on the non-physical IPV type. Furthermore, the lowest risk of non-physical IPV appears when both partners are employed. Additionally we find that especially the education of a woman and her partner plays a major role in reducing the risk for both types of IPV when successfully finished college. Keywords: intimate partner violence, gender based violence, domestic violence, employment, multivariate probit, mvprobit, 2SLS, two-stage leas squares, endogeneity ii Acknowledgements First of all I would like to thank my supervisor Arantza Ugidos Olabazal for her consistent support and guidance during the running of this thesis. I would also like to acknowledge all my other professors of the University of the Basque Country for encouraging me along the whole way and for giving me this research opportunity. iii Contents Abstract i Acknowledgements ii 1 Introduction 1 2 Literature Review 3 3 Data and Descriptive Statistics 10 4 Methodology 18 5 Results 22 6 Conclusions 31 A Appendix 33 Bibliography 37 iv List of Tables 3.1 Household’s income intervals . . . . . . . . . . . . . . . . . . . . . 12 3.2 Ageintervals ............................. 12 3.3 Categories of serious abuse in the Spanish VAW surveys . . . . . . 13 3.4 IPVbyType.............................. 13 3.5 IPVtypebyRegion.......................... 14 3.6 Summary statistics for woman’s, partners’ and household characteristicsbyIPVstatus .......................... 16 3.7 Summary statistics for partners’ employment and IPV status . . . . 17 5.1 Estimates for risk of Physical IPV . . . . . . . . . . . . . . . . . . 26 5.2 Cross-equation correlations coefficients for IPV Physical (from mvprobitestimation)............................. 27 5.3 Estimates for risk of Non-Physical IPV . . . . . . . . . . . . . . . . 28 5.4 Cross-equation correlations coefficients for IPV Non-Physical (from mvprobitestimation) ......................... 29 5.5 Estimated marginal effects of woman and partner employment on IPV on the basis of the 2SLS regression . . . . . . . . . . . . . . . 29 5.6 Estimated marginal effects of household characteristics on IPV on basis of the 2SLS regression . . . . . . . . . . . . . . . . . . . . . 30 A.1 Estimates for risk of Physical IPV - including interaction of woman’s and partner’s education level: college . . . . . . . . . . . . . . . . . 33 A.2 Continuing: Estimates for risk of Physical IPV - including interaction of woman’s and partner’s education level: college . . . . . . . . 34 A.3 Estimates for risk of Non-Physical IPV - including interaction of woman’s and partner’s education level: college . . . . . . . . . . . 35 A.4 Continuing: Estimates for risk of Non-Physical IPV - including interaction of woman’s and partner’s education level: college . . . . . 36 v List of Abbreviations IPV Intimate Partner Violence VAW Violence Agains Women HBS Household Budget Survey HH HouseHold EAPS Economically Active Population Survey 2SLS 2 Stage Least Squares 1 Chapter 1 Introduction Violence against women (VAW) is a product of gender based inequalities and has been associated with different factors such as e.g. employment instability (Showalter, Yoon, and Logan, 2021). Hence, VAW is not only of great political interest but also of economic interest all around the globe. Lots of studies are devoted to examine determinants which affect whether a woman experiences violence conducted by her partner or not and many studies focus on the impact of a woman’s employment status on her risk of experiencing the so called intimate partner violence (IPV) (Allen et al., 2019; Goodey, 2017; Capaldi, 2012). Still the findings in the literature differ to a great extent, they are inconsistent and sometimes even contradictory. This very fact makes it desirable take the studies further, using the most recent data available for Spain, from 2019. In this master’s thesis we want to examine the effect of the woman’s employment status, her partner’s employment status as well as a set of sociodemographic variables on the risk of IPV. We distinguish two types of IPV, such are non-physical IPV and physical IPV. IPV refers to behaviour by an intimate partner or ex-partner that causes physical or non-physical harm - where physical harm also includes sexual abuse and non-physical harm includes controlling behaviour (World Health Organization, 2019). Furthermore we apply three different estimation procedures such are a two-stage least squares linear probability model, a non-linear multivariate probit model and a linear univariate probit model. In applying those we want to study the differences between the methods and, furthermore, we aim at establishing conclusions about the importance of taking the possible endogeneity of the employment statuses of the woman and her partner as well as the income into account. For this purpose there are two very recent studies - beside others - which we mainly follow due to the fact that we carry out the same estimation approaches as them and compare the results between each other. The first to mention is the work from Lenze and Klasen, 2017, who carried out their research with data from Jordan and published their work in 2017. They used a two-stage least square linear probability approach in order to account for the possible endogeneity of woman’s Chapter 1. Introduction 2 employment. After accounting for a possible endogeneity bias, they came to the conclusion that there is no significant evidence for an effect of woman’s labor force participation on domestic violence. The second paper we mainly follow is the study from Alonso-Borrego and Carrasco, 2017. This study is carried out with partly the same data as we use for this very master’s thesis. They use data from the Spanish VAW survey from the years 1999, 2002 and 2006 while we use data from the same survey but from the survey year 2019. They studied the relationship between, not only, the woman’s employment status, but also the partner’s employment status and income on the risk of IPV. They use a multivariate probit model in order to account for the possible endogeneity of both of the employment statuses and income. The master’s thesis is organized as follows. In section 2 we report previews studies and literature regarding the examination of IPV and employment statuses. The underlying data base for our analysis, the sample creation and the corresponding descriptive statistics are described in section 3. The theoretical aspects of the different estimation models used are presented in section 4. In section 5, we discuss the results before we finally present the main conclusions in section 6. 3 Chapter 2 Literature Review In this part of the master’s thesis we firstly discuss the relevant research that took place before the works of Alonso-Borrego and Carrasco, 2017, and Lenze and Klasen, 2017. In the second part of this chapter we will not only look closer into the two main papers mentioned above, but also focus on the research that has come to light afterwards, until the writing of this master’s thesis. Intimate partner violence (IPV) initially has been a topic of criminology and sociology where it is known to mainly serve two purposes: Expressive purposes and instrumental purposes. Expressive here means that some men derive a direct benefit from violence while instrumental refers to the case when the partner is increasing his utility indirectly, by the control of the woman’s behaviour (Alonso-Borrego and Carrasco, 2017). One of the older papers - and one of the first studies to provide negative empirical evidence on the relationship between IPV and the female employment status - is the work of Gelles, 1976. Gelles used the sociological absolute resource theory to explain why wives stay with their partners despite being abused by them. He identifies three factors which mainly influence the actions of abused wives. Firstly, the less violence is exerted and the less frequent it is, the more likely it is that the woman will stay with her husband, secondly; if a woman was struck by her parents when she was a child, it is more likely for her to stay with her abusive husband - the more abuse experienced the more likely it is. Thirdly; the more resources and the more power a wife has in her relationship to her husband, the less likely she is to stay with him. He also concludes that IPV might be a problem of poor households hence identifying households income as possibly influenced by an endogenous bias. At about the same time and further, altruistic models claimed by Becker, 1965 and Becker, 1973 were predominant. Here we find the assumptions to be that in a marriage each person tries to do as well as possible and that the "marriage market" is in equilibrium. With the aid of several additional simplifying assumptions, a number of significant implications about behavior in this market have been derived. He found that the gain to a man and woman from marrying compared to remaining single is 10 Chapter 3 Data and Descriptive Statistics In the first part of this chapter of the master’s thesis we will describe the data we worked with, the data sources and the survey from which most of our data basis originates. The second part of this chapter denotes the descriptive statistics analysis. The main data source of this master’s thesis is the cross-sectional surveys on violence against women (VAW) in Spain from 2019. This survey is not only the most relevant statistical operation carried out in Spain on this type of violence but also the only official statistic to measure the prevalence of violence against women. The survey has been carried out approximately every 4 years since 1999 and is included in the National Statistical Plan Against Domestic Violence. The Government Delegation against Gender Violence has been in charge of preparing the last three editions from 2011 to 2019. In total, the 2019 Macro-survey is the sixth to be carried out in Spain. Its main objective is to find out the percentage of women aged 16 and over residing in Spain who have suffered or currently are suffering some type of violence only because they are women (Ministry of Equality, 2020). This survey has already served several other researchers and authors as the data basis for their work in previous years such as Alonso-Borrego and Carrasco, 2017, Brassiolo, 2016 and Tur-Prats, 2016. The survey were promoted by the First National Action Plan against Domestic Violence established in 1998, which led to subsequent legislative proposals that resulted in the first constitutional law against gender-based violence in 2004 . This law not only provided for harsher penalties for perpetrators, but also funded public assistance services and shelters for battered women, promoted training programmes for health professionals and judges, and public education and media campaigns to raise awareness of violence against women. Following the adoption of the law, the number of complaints increased, as did the number of emergency calls and the number of women contacting the special telephone service for victims of abuse (Alonso-Borrego and Carrasco, 2017). The gender violence surveys are large, nationally and regionally representative samples of women living in Spain, in this case for the year 2019. The surveys were conducted by telephone with women Chapter 3. Data and Descriptive Statistics 11 aged 16 or older1. The second data source that feeds into this work is the household budget survey (HBS) publicly available at the website from the national institute of statistics in Spain (National Institute of Statistics, 2019). In particular we selected data from the third quarter in 2019 since this is the same time period in which the 2019 VAWsurvey was carried out as well. The HBS serves our purpose to receive information on the type and purpose of consumption expenditure, as well as on a number of characteristics relating to the living conditions of households and the exact amount of total household net monthly income. For us, the net incomes are particularly interesting in order to be able to calculate the mean income for Spain in this period and then to be able to create the variable "low income/poor household", as AlosnoBorregeo et al. have also done. Note that although we do the same as Alonso-Borrego and Carrasco, 2017, we have to take the changed intervals into account and adjust the creation of the variables accordingly to the new intervals provided by the VAW survey 2019. While Alonso-Borrego and Carrasco, 2017, defined large municipalities as those with at least 200.000 inhabitants or more, we are facing given intervals (from the changed VAW survey in 2019) in another predefined range (e.g. from 50.001 to 100.0000 inhabitants, from 100.001 to 400.00 inhabitants etc.) and therefore define a large municipality as those who have 100.000 or more inhabitants. Furthermore, note that, Alonso-Borrego et al. have considered households to be a "low income/poor household" if their income is at least one standard deviation less than the national wide average (Alonso-Borrego and Carrasco, 2017). However in our case - due to our data basis and the changed survey, the VAW statistics - we have different predefined income intervals that we use in order to create this variable, see table 3.1. Based on the HBS data, we have calculated the national average of household income, which equals C 2,205.313, and the standard deviation, which equals C 1,386.94. This shows that, according to the definition of Alonso-Borrego et al. households with an income of C 818,373 (2,205.313-1,386.94 = 818.373) or less are classified as poor households. In our case, we set the threshold with the third interval, i.e. we classify households up to an income of 900 C as poor households because we do not have that precise information as Alonso-Borrego et al. had about the households incomes. 1In 2015, the Macro survey questionnaire was significantly modified in relation to previous editions (1999, 2002, 2006, 2011). With this changes, which mainly took as a reference the Guidelines for the Production of Statistics on Violence against Women prepared by the United Nations Statistics Division, the aim was to measure more rigorously the reality of violence against women in Spain. With the 2019 Macro-survey on Violence against Women, this process of improving the quality of the survey has continued (Pérez-Sánchez, Dávila-Cárdenes, and Gómez-Déniz, 2022). One example of a change would be the age of the interviewed women changed from 18 to 16. Chapter 3. Data and Descriptive Statistics 12 TABLE 3.1: Household’s income intervals Interval HH’s monthly net income in C 1 none or less than 300.00 2 from 301.00 to 600.00 3 from 601.00 to 900.00 4 from 901.00 to 1,200.00 5 from 1,201.00 to 1,800.00 6 from 1,801.00 to 2,400.00 7 from 2,401.00 to 3,000.00 8 from 3,001.00 to 4,500.00 9 4,501.00 or more Source: Spanish VAW Survey 2019 The third and final source of data for this thesis is the Economically Active Population Survey (EAPS) - also publicly available over the website form the Spanish National Institute of Statistics - from which we retrieve, after short calculations, the employment and unemployment rates for 2019 by gender, region and age-interval. Here we have to note that, again, we are facing different data than Alonso-Borrego and Carrasco, 2017 and hence we need to use different age-intervals since the unemployment and employment rates are only provided in the intervals displayed in table 3.2. TABLE 3.2: Age intervals Interval interval in ages 1 from 16 to 19 2 from 20 to 24 3 from 25 to 54 4 55 or older Source: Spanish VAW Survey 2019 We restrict our sample to woman older than 242. The difference in sample sizes to the work of Alonso-Borrego and Carrasco, 2017, is mainly due to the fact that we only use data from one survey year, whereas Alonso-Borrego et al. use several survey years as a data basis. Thus our final sample consists of 1,716 observations. 2We cannot collapse the age intervals "from 16 to 19" and "from 20 to 24" into new or other of the given intervals because we are facing predefined intervals for the employment and unemployment rate variables which would not match these adjusted age intervals Chapter 3. Data and Descriptive Statistics 13 We constructed the IPV indicator variables for comparison purposes in the same way as in previous studies (e.g. Alonso-Borrego and Carrasco, 2017, etc.). Table 3.3. displays the list of main behaviours on which the construction of these indicators is based. As can be seen, the IPV indicators take account for either physical or nonphysical abuse. Whereas non-physical abuse refers to abuse not including physical violence, physical abuse refers to the opposite, including sexual violence. TABLE 3.3: Categories of serious abuse in the Spanish VAW surveys Behavior Physical Non-Physical Abuse Abuse Stopped from seeing relatives, friends and neighbors × Prevented from fair share of household money × Insulted or threatened you × Prevented from deciding by yourself × Forced to have sexual intercourse × Deprived of your necessities × Scared you sometimes × Pushed you or hit you × Scorned about your capacity × Criticized for the things you do × Despised for your beliefs × Disregarded for your work × Disrespected in front of your children × Table 3.4 shows the frequencies and percentages of the respondents for 2019 by type of IPV. From our sample, almost half - 47.65 per cent - of the women respondents report having experienced some type of abuse. Furthermore, from those who experienced some type of abuse, almost half of them, precisely 47.06 percent report to have been victimized by non-physical violence which accounts for the largest proportion here, while physical violence accounts for 22.15 percentage points. TABLE 3.4: IPV by Type IPV Type Obs. Frequency Percentage Physical 1,716 264 22.15 Non-Physical 1,716 561 47.06 Any 1,716 568 47.65 No Abuse 1,716 624 52.35 Source: Own calculations from Spanish VAW Survey 2019 In Table 3.5 we report the frequencies of IPV types by region. Here we find that the IPV type "Any" ranges from 1.58 percent in the region La Rioja up to 11.62 Chapter 3. Data and Descriptive Statistics 14 percent in the community of Valencia, which provides evidence for statistically significant differences across regions. The same realization holds for physical IPV, ranging from 0.76 percentage points in Navarra to 12.88 percent in Cataluña. For non-physical IPV which ranges from 1.60 percent in Navarra and La Rioja to 11.41 percent in the community of Valencia. The difference between regions with respect to the variable "No Abuse", also shows a similar extent of fluctuation with a range from 0.96 percent in Melilla up to 10.10 percent in Andalucía. TABLE 3.5: IPV type by Region Region Obs No Abuse Physical Non-Physical Any Andalucía 104 10.10 8.33 7.31 7.22 Aragón 38 3.37 2.65 3.03 2.99 Asturias 48 4.33 2.65 3.74 3.70 Baleares 55 5.77 4.55 3.39 3.35 Canarias 75 6.25 6.44 6.42 6.34 Cantabria 47 4.81 3.03 3.03 2.99 Castilla-La Mancha 57 5.29 4.17 4.28 4.23 Castilla y León 48 4.49 4.17 3.57 3.52 Cataluña 98 5.93 12.88 10.87 10.74 Comunitat Valencia 112 7.37 11.74 11.41 11.62 Extremadura 59 4.65 5.30 5.35 5.28 Galicia 83 5.93 9.47 7.84 8.10 Madrid 116 9.29 8.33 10.16 10.21 Murcia 57 5.45 5.68 3.92 4.05 Navarra 29 3.04 0.76 1.60 1.76 País Vasco 79 6.89 4.17 6.42 6.34 La Rioja 24 2.40 1.14 1.60 1.58 Ceuta 43 3.69 3.41 3.57 3.52 Melilla 20 0.96 1.14 2.50 2.46 Source: Own calculations from Spanish VAW Survey 2019 The main summary statistics by IPV status are shown in Table 3.6. These statistics show that the characteristics of women, partners and households differ according to the presence of abuse, with the strongest differences for physical abuse. Most women who have been abused are found in the 25-54 age segment with huge differences to the older segment. Also the mean difference between non-physical and physical abuse is significant on the 10% level. When it comes to the education level of women, we find different results to previous studies but also have to keep in mind that we do use different, more precise, levels of education. We find that the higher the education level of a woman is, the higher the numbers of not abused women. But, we also find that women have suffered more physical as well as non-physical abuse when they only have completed secondary education in comparison to when they Chapter 3. Data and Descriptive Statistics 15 have finished college or a vocational training. Here we have to be careful because finishing a vocational training might indicate that the woman is employed and the result is more likely due to the employment status than due to the finalization of the vocational training. However, we will gain further particular insight in the part of the model estimation of this very thesis. Important to point out particularly is that the woman’s education level "college" shows significant mean differences even on the 1% level for physical abuse. When it comes to the partners age and education we find similar tendencies. Most women who suffered abuse are found to be in the segment of 25 to 54 years. Interestingly we find more abused women when a woman’s partner has finished secondary education in comparison to when the partner finished college. A woman’s college degree also does seem to show that women in this sample were less abused. Note that the mean differences for the education level "college" are significant at the 1% for all types of abuse. For those woman who are more educated than her partner we find the means to be very close to each other with respect to the IPV type. Furthermore we find the mean differences to be significant only for the physical IPV type. There are very slightly differences in the household size between abused and non-abused women. The martial status differs only to a slightly larger extent when it comes to physical abused women, while the mean differences are significant at all levels for all types of IPV. When it comes to the income of an household we find that in average income households are the most reported cases of non-physical and physical violence with significant mean differences for all types of IPV at the 10% level and in particular, for physical abuse, even on the 1 % level. In table 3.7 we display the pure prevalence of IPV depending on the employment status of the woman or her partner respectively. We find that there are no significant differences in the likelihood of being abused regardless of whether the woman is employed or not. In the work from Alonso-Borrego and Carrasco, 2017, this was a significant difference with woman who are not employed being more likely to be a victim of IPV. Furthermore only small differences appear in the likeliness of being abused when both - the woman and her partner - are employed. Compared to the previous realizations, the occurrence of IPV is the smallest when both partners are not employed (Any violence = 0.44). When we take a more detailed look at the interaction of the employment status of both, we find only small differences for the cases when the woman is not employed and the partner is employed (e.g. physical = 0.23) and when both are not employed (e.g. Physical = 0.20). The evidence for when the woman is employed and the partner is not shows that it is more likely (compared to the previous two cases) for the woman to be abused (e.g. physical = 0.33). Chapter 3. Data and Descriptive Statistics 16 TABLE 3.6: Summary statistics for woman’s, partners’ and household characteristics by IPV status IPV Type No Abuse Physical Non-Phys. Any Mean (SD) Mean (SD) Mean (SD) Mean (SD) Woman age 25-54 years 0.8 (0.40) 0.86* (0.34) 0.84* (0.37) 0.84 (0.37) 55 or more 0.19 (0.39) 0.13* (0.34) 0.15 (0.36) 0.15 (0.36) Woman education Primary or less 0.08 (0.27) 0.12* (0.33) 0.1 (0.3) 0.1 (0.3) Voc. Train. 0.24 (0.43) 0.24 (0.43) 0.24 (0.43) 0.24 (0.43) Secondary 0.34 (0.47) 0.41* (0.49) 0.38 (0.48) 0.37 (0.48) College 0.34 (0.47) 0.22*** (0.42) 0.29* (0.45) 0.29* (0.45) Partner age 25-54 years 0.74 (0.44) 0.81* (0.39) 0.8* (0.40) 0.79* (0.41) 55 or more 0.25 (0.44) 0.19* (0.39) 0.20* (0.40) 0.21* (0.40) Partner educ. Primary or less 0.10 (0.30) 0.14 (0.34) 0.11 (0.32) 0.11 (0.32) Voc. Train. 0.22 (0.42) 0.21 (0.41) 0.24 (0.43) 0.24 (0.42) Secondary 0.37 (0.48) 0.5 (0.50)*** 0.42* (0.49) 0.42* (0.49) College 0.31 (0.46) 0.16 (0.36)*** 0.23*** (0.42) 0.23*** (0.42) Woman more educated 0.31 (0.46) 0.36 (0.48)* 0.34 (0.47) 0.34 (0.47) Household size 3.44 (1.02) 3.48 (1.19) 3.45 (1.12) 3.45 (1.12) Married (y/n) 0.76 (0.43) 0.54 (0.5)*** 0.63 (0.48)*** 0.63 (0.48)*** Household income Below average 0.10 (0.30) 0.13 (0.33) 0.11 (0.31) 0.11 (0.31) Average 0.67 (0.47) 0.76 (0.43)* 0.72* (0.45) 0.72* (0.45) Above average 0.23 (0.42) 0.12 (0.32)*** 0.17* (0.37) 0.17* (0.38) Large Municipality 0.36 (0.48) 0.33 (0.47) 0.36 (0.48) 0.36 (0.48) Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Chapter 3. Data and Descriptive Statistics 17 TABLE 3.7: Summary statistics for partners’ employment and IPV status IPV Type No Abuse Physical Non-Phys. Any Mean (SD) Mean (SD) Mean (SD) Mean (SD) Woman empl. 0.66 (0.2) 0.65 (0.3) 0.65 (0.2) 0.65 (0.2) Partner empl. 0.82 (0.2) 0.79 (0.3) 0.82 (0.2) 0.82 (0.2) Both empl. 0.60 (0.2) 0.55 (0.3) 0.57 (0.2) 0.57 (0.2) Woman not empl., 0.5 (0.5) 0.23 (0.42) 0.5 (0.5) 0.5 (0.5) Partner empl. Woman empl., 0.46 (0.5) 0.33 (0.47) 0.53 (0.5) 0.54 (0.5) Partner not empl. Both not empl. 0.56 (0.5) 0.20 (0.40) 0.43 (0.5) 0.44 (0.5) Source: Own calculations from Spanish VAW Survey 2019 Summarizing the descriptive statistics analysis we find that depending on the occurrence of abuse the sociodemographic attributes differ to quite a big extent. Dramatically we find that almost half of the respondents experienced some kind of physical or non-physical abuse with non-physical abuse being the bigger share. Furthermore we found huge differences on the occurrence of IPV depending on the region were the women live. In the regions of Cataluña, Valencia and Madrid most incidence have been reported and the regions of Navarra and La Rioja reported the least. The most cases of abused women are found in the age group of 24 years old up to 54 years, which applies to both the age of the woman and the age of the partners. The same tendencies apply for the level of education, for the women and also for her partner. When not finished college more abuse is found to be present. In accordance with Alonso-Borrego and Carrasco, 2017, we find that there are only small differences in the presence of abuse whether woman is working or not but less abuse depending on the employment status of her partner, finding less abuse, when the partner is employed. Also we find that when the woman is employed and her partner is not, abuse seems to be more likely to appear and the least appearance of abuse is found in relationships where both partners are employed. 18 Chapter 4 Methodology In this chapter of the master’s thesis, we describe the empirical model underlying the analysis that we used to estimate the risk of the occurrence of potential determinants depending on different variables, taking into account endogeneity and comparing our results with previous studies, in particular with the work of Alonso-Borrego and Carrasco, 2017 and Lenze and Klasen, 2017. To achieve results that are as comparable as possible to those of Alonso-Borrego and Carrasco, 2017 and Lenze and Klasen, 2017, we examine the effect of employment status of females and their partners on the probability of female abuse, in a linear probability model estimating it in a two-stage least squares approach (2SLS) for the comparison with Lenze and Klasen, 2017 and a nonlinear model estimation using a simultaneous multivariate probit approach like Alonso-Borrego and Carrasco, 2017, used. The latter is a nonlinear discrete model where IPV∗ iis the latent process that steers IPV which is identified by the following behavioral model: IPV∗ i=α0+α1fi+α2pi+α3(fi×pi)+X′ iδ+vi≡W′ iβ+vi(1) where Xidenotes a set of exogenous variables. The dummy variables for women’s and her partner’s employment are denoted by fiand pirespectively. The interaction between the female’s and her partner’s employment is therefore characterized by (fi×pi). We find ourselves observing the binary variable, IPVi, that indicates if women iexperiences IPV or not. Taking on the value 1 for the case that she does and 0 otherwise 1. We represent this variable with an indicator function as follows: IPVi=1(IPV∗ i>0) = 1(α0+α1fi+α2pi+α3fi×pi+X′ iδ+vi≥0)(2) 1Also known as dummy endogenous variable model (Amemiya et al., 1985) Chapter 4. Methodology 19 This model (2) becomes a standard probit model if vi|Xi,fi,pi∼N(0,1)which we also estimate for comparison reasons. The woman’s and her partner’s empployment status may not be exogenous and be related to the IPV status through unobserved factors. In our case we account for endogeneity by estimating a multivariate probit model. The first step is to define reduced form equations for female and male employment: fi=1(f∗ i>0) = 1(Z′ 1iλ1+εi1>0)(3) pi=1(p∗ i>0) = 1(Z′ 2iλ1+εi2>0)(4) Here Z1iand Z2iare sets of exogenous variables, that include XiFurthermore we assume that vi,εi1,εi2are jointly normally distributed with zero mean vector and covariance matrix. Ω=   1ρvε1ρvε2 1ρε1ε2 1   (5) For the case that ρvε1=ρvε2=ρε1ε2=0 we are not forced to estimate the equations simultaneously but can indeed obtain consistent parameters by estimating each equation separately (Alonso-Borrego and Carrasco, 2017). We use the most popular simulation method (simulation by maximum likelihood - SML) by Geweke, Hajivassiliou and Keane (GHK) to estimate our multivariate probit model, which is based on the expression of the multivariate normal distribution as the product of sequentially conditioned univariate normal distributions (Börsch-Supan and Hajivassiliou, 1993). Since we face a model in which we have an additional possible endogenous dummy variable (beeing a poor household or not) we add this variable to our reduced form model so that our multivariate probit model takes the correlation between the error terms of the four auxiliar equations (for IPV, woman’s employment status, partner’s employment status and poor houshold) into account. However, since we run the multivariate probit regression first and then realize that we only find evidence for the endogeneity (see Chapter 5) of the parnter’s employment status for IPV physical and only for low income/poor households for the non-physical IPV we will carry out our estimations of the auxiliar equations to create the instruments as well as the final estimation of the in stage 1 predicted values and the outcome IPV iteratively instead of simultaneously, by using a 2-Stage Least Squares (2SLS) model. The corresponding evidence can be found in Tables 5.2 & and 5.42. The p-values of the correlation of the errors are only lower than 0.1 for 2We will take a closer look at this in chapter 5 Chapter 5. Results 26 TABLE 5.1: Estimates for risk of Physical IPV Probit 2SLS mvprobit IPV IPV IPV VARIABLES Physical Physical Physical Woman employed 0.24 0.61 0.09 (0.17) (0.5) (0.26) Partner employed -0.15 0.06 0.13 (0.14) (0.22) (0.20) Women empl. ×Partner empl. -0.30 -0.23 -0.251 (0.19) (0.38) (0.19) Household size 0.005 0.01 -0.0044 (0.04) (0.01) (0.04) Woman Age 25-54 0.13 -0.11 0.13 (0.29) (0.16) (0.29) Woman Age 55+ -0.19 -0.05 -0.08 (0.3) (0.12) (0.31) Woman: Secondary 0.01 -0.01 0.01 (0.08) (0.03) (0.08) Woman: College -0.23* -0.08* -0.23* (0.1) (0.04) (0.1) Partner: Secondary 0.06 0.004 0.06 (0.08) (0.03) (0.08) Partner: College -0.23* -0.06* -0.26* (0.11) (0.03) (0.11) ln(province GDP per capita) 0.03 -0.01 -0.05 (0.19) (0.08) (0.2) Prov. population density -0.00003 -0.00001 -0.00003 (0.00004) (0.00001) (0.00003) Woman more educated 0.07 0.02 0.071 (0.09) (0.03) (0.09) Low income household 0.04 0.55* -0.26 (0.14) (0.32) (0.26) Large Municipality -0.02 -0.01 -0.02 (0.07) (0.02) (0.07) RNorMed 0.05 0.004 0.04 (0.08) (0.03) (0.08) Observations 1,716 1,716 1,716 Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Chapter 5. Results 27 TABLE 5.2: Cross-equation correlations coefficients for IPV Physical (from mvprobit estimation) Woman Partner Low VARIABLES empl. empl income IPV Physical 0.05 -0.19* 0.18 (0.11) (0.10) (0.12) Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Chapter 5. Results 28 TABLE 5.3: Estimates for risk of Non-Physical IPV Probit 2SLS mvprobit IPV IPV IPV VARIABLES Non-Physical Non-Physical Non-Physical Woman employed 0.24 1.25* 0.02 (0.16) (0.57) (0.27) Partner employed 0.04 0.58* 0.26 (0.12) (0.27) (0.2) Women empl. ×Partner empl. -0.33* -1.09* -0.28 (0.17) (0.45) (0.17) Household size -0.01 0.002 -0.02 (0.03) (0.01) (0.03) Woman Age 25-54 0.01 -0.16 0.01 (0.25) (0.18) (0.26) Woman Age 55+ -0.18 -0.03 -0.11 (0.27) (0.14) (0.27) Woman: Secondary 0.06 0.02 0.067 (0.08) (0.03) (0.08) Woman: College -0.18* -0.09* -0.18* (0.09) (0.04) (0.09) Partner: Secondary -0.1 -0.06* -0.1 (0.07 ) (0.03) (0.08) Partner: College -0.14 -0.08* -0.17* (0.1) (0.04) (0.1) ln(province GDP per capita) 0.23 0.03 0.12 (0.17 ) (0.11) (0.18) Prov. population density 0.00006* 0.00002 0.00006* (0.00003) (0.00001) (0.00003) Woman more educated 0.1 0.04 0.09 (0.08) (0.03) (0.08) Low income household 0.08 0.16 -0.58* (0.12) (0.38) (0.26) Large Municipality 0.05 0.02 0.05 (0.07) (0.03) (0.07) RNorMed 0.03 -0.0039 0.02 (0.07) (0.03) (0.07) Observations 1,716 1,716 1,716 Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Chapter 5. Results 29 TABLE 5.4: Cross-equation correlations coefficients for IPV NonPhysical (from mvprobit estimation) Woman Partner Low VARIABLES empl. empl income IPV Non-Physical 0.11 -0.17 0.36* (0.12) (0.11) (0.14) Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 TABLE 5.5: Estimated marginal effects of woman and partner employment on IPV on the basis of the 2SLS regression Effect of woman employment IPV IPV Physical Non-Physical Partner not employed 0.51 0.42* (0.39) (0.64) Partner employed 0.4 0.54* (0.71) (0.84) Effect of partner employment Woman not employed 0.11 0.31* (0.33) (0.27) Woman employed -0.64 -0.27* (0.47) (0.72) Source: Own calculations from Spanish VAW Survey 2019 Marginal effects evaluated for 2019, using the sample mean values at that year. Standard errors in parentheses (using Delta method) *** p<0.01, ** p<0.05, * p<0.1 Chapter 5. Results 30 TABLE 5.6: Estimated marginal effects of household characteristics on IPV on basis of the 2SLS regression IPV IPV Variable Physical Non-Physical Household size 0.00 0.001 (0.00) (0.01) Woman Age 25-54 -0.05 -0.1 (0.08) (0.08) Woman Age 55+ -0.02 -0.002 (0.06) (0.05) Woman: Secondary -0.01 0.01 (0.01) (0.01) Woman: College -0.04* -0.03* (0.002) (0.04) Partner: Secondary 0.01 -0.03* (0.02) (0.01) Partner College -0.02* -0.04* (0.01) (0.002) ln(province GDP per capita) -0.001 0.001 (0.007) (0.01) Prov. population density -0.00 0.00 (0.00) (0.00) Woman more educated 0.04 0.02 (0.01) (0.02) Low income household 0.08* 0.06 (0.03) (0.04) Large Municipality -0.04 0.01 (0.02) (0.01) RNorMed 0.0002 -0.004 (0.001) (0.002) Source: Own calculations from Spanish VAW Survey 2019 Marginal effects evaluated for 2019, using the sample mean values at that year. Standard errors in parentheses (using Delta method) *** p<0.01, ** p<0.05, * p<0.1 31 Chapter 6 Conclusions In this master’s thesis we used data from different sources but mainly from the violence against women (VAW) survey in Spain from 2019 (more details can be found in chapter 3) in order to examine the effect of woman’s and her partner’s employment status on the risk of experiencing IPV. For this purpose we distinguished between two IPV types, physical and non-physical. We took account of the separate effects of the two different employment statuses (for the woman and her partner) as well as of the interaction between both of them. Additionally we conditioned our analysis on income and a set of covariates. We used three different estimation approaches, such as an univariate probit regression, a two-stage least square linear probability regression and a multivariate probit regression. While the univariate probit regression does not take the potential endogeneity of the variables woman’s employment status, partner’s employment status and poor/low income households into account, the other two regression strategies do. In the case of physical IPV the results of the different analysis approaches taking account of endogeneity and the approach which is not taking account of it (the univariate probit analysis) are quite similar. As you can see later, two of the significant variables which lower the risk of experiencing IPV physical can also be found to lower the risk of experiencing non-physical IPV. Those variables are having a college degree and having a partner who finished college. The 2SLS approach provides further evidence such is, that living in a low income/poor household increases the risk of physical IPV. Furthermore the results in this master’s thesis underline the importance of taking account of endogeneity when it comes to non-physical IPV. While in both of the approaches which take account of endogeneity the results show that for IPV non-physical, finishing college and having a partner who finished college significantly decreases the risk, the univarate probit approach provides no evidence that having a partner who finished college reduces the risk of experiencing non-physical IPV. Also taking the results of the marginal effects into account, we find that finishing college and having a partner who finished college underlining the findings from the estimation results and, furthermore, that the employment status of the partner plays Chapter 6. Conclusions 32 a major role only if the woman is employed too and in this case lowering the risk of non-physical IPV. For physical IPV we did not find any evidence with respect to the employment statuses. The main realization therefore are that only the partners employment status plays a major role in reducing the risk of IPV and only when the woman is also employed and only on the non-physical IPV type. This realization differs from those from Alonso-Borrego and Carrasco, 2017 realizations in the way that they found the same influence but also influencing the physical IPV. This results matches with the realizations of Kaukinen, 2004 who found that the risk of abuse rather depends on the employment status of the partner than on the employment status of the woman herself. Our results match the results from Alonso-Borrego and Carrasco, 2017 in the point that the lowest risk of non-physical IPV appears when both partners in a relationship are employed. Our results do not match with the results from Lenze and Klasen, 2017 since we found significant evidence for the partner’s employment status, if the woman also is employed and they did not find any significant evidence after accounting for endogeneity. Furthermore we found that especially the education of a woman and her partner plays a major role in reducing the risk for both types of IPV when successfully finished college. Intriguingly, when comparing these results with the results from Pérez-Sánchez, Dávila-Cárdenes, and Gómez-Déniz, 2022, (see Chapter 2 for more details) we realize that they found a decreasing effect for sexual IPV only when the partner is currently studying and furthermore that the education level of the woman is not relevant to any type of IPV. Reflecting on our work we have to note that our sample is relatively small when comparing the sample size of 1,716 from our work to the sample size of e.g. AlonsoBorrego and Carrasco, 2017, with more than 30.000 observations and we only considered data from one year. Hence, in order to get even more precise results it would be desirable to extent the underlying data set with data from previous years. With special respect to the 2SLS approach one could might obtain better results when finding better instruments. 33 Appendix A Appendix TABLE A.1: Estimates for risk of Physical IPV - including interaction of woman’s and partner’s education level: college Probit 2SLS mvprobit IPV IPV IPV VARIABLES Physical Physical Physical Woman employed 0.24 0.62 0.08 (0.17) (0.5) (0.26) Partner employed -0.15 0.05 0.13 (0.14) (0.22) (0.20) Women empl. ×Partner empl. -0.3 -0.23 -0.25 (0.19) (0.38) (0.19) Household size 0.01 0.01 -0.003 (0.04) (0.014) (0.04) Woman Age 25-54 0.13 -0.11 0.13 (0.29) (0.16) (0.29) Woman Age 55+ -0.19 -0.05 -0.08 (0.3) (0.12) (0.31) Woman: Secondary -0.002 -0.01 0.0004 (0.08) (0.03) (0.08) Woman: College -0.15 -0.06 -0.15 (0.12) (0.05) (0.12) Observations 1,716 1,716 1,716 Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Appendix A. Appendix 34 TABLE A.2: Continuing: Estimates for risk of Physical IPV - including interaction of woman’s and partner’s education level: college Probit 2SLS mvprobit IPV IPV IPV VARIABLES Physical Physical Physical Partner: Secondary 0.07 0.01 0.07 (0.08) (0.03) (0.08) Partner: College -0.15 -0.05 -0.18 (0.13) (0.04) (0.13) Women college ×Partner college -0.22 -0.05 -0.20 (0.19) (0.05) (0.19) ln(province GDP per capita) 0.02 -0.01 -0.06 (0.19) (0.08) (0.2) Prov. population density -0.00003 -0.00001 -0.00003 (0.00003) (0.00001) (0.00003) Woman more educated 0.03 0.01 0.04 (0.1) (0.03) (0.1) Low income household 0.05 0.58* -0.25 (0.14) (0.32) (0.27) Large Municipality -0.018 -0.01 -0.02 (0.07) (0.02) (0.07) RNorMed 0.05 0.003 0.04 (0.08) (0.03) (0.08) Observations 1,716 1,716 1,716 Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Appendix A. Appendix 35 TABLE A.3: Estimates for risk of Non-Physical IPV - including interaction of woman’s and partner’s education level: college Probit 2SLS mvprobit IPV IPV IPV VARIABLES Non-Physical Non-Physical Non-Physical Woman employed 0.24 1.27* 0.01 (0.16) (0.57) (0.27) Partner employed 0.04 0.57* 0.25 (0.12) (0.27) (0.2) Women empl. ×Partner empl. -0.33* -1.09* -0.28 (0.17) (0.46) (0.17) Household size -0.01 0.003 -0.02 (0.03) (0.01) (0.03) Woman Age 25-54 0.01 -0.16 0.01 (0.25) (0.18) (0.26) Woman Age 55+ -0.18 -0.03 -0.11 (0.27) (0.14) (0.27) Woman: Secondary 0.05 0.01 0.05 (0.08) (0.03) (0.08) Woman: College -0.08 -0.06 -0.09 (0.11) (0.05) (0.11) Observations 1,716 1,716 1,716 Source: Own calculations from Spanish VAW Survey 2019 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1