Ensaios em econometria, estatística e aprendizado de máquina aplicados à violência doméstica

Abstract

This thesis presents three essays on the theme of domestic violence, using data from the Survey on Socioeconomic Conditions and Domestic and Family Violence against Women (PCSVDFMulher). The first chapter seeks to analyze the occurrence of loss of sampling units, that is, sampling attrition, in PCSVDF-Mulher. During the first two waves of the research, which took place between 2016 and 2017, a percentage of attrition of 55.13% was found. To better understand the phenomenon, machine learning techniques and a discrete choice model were used, capable of identifying variables and their importance in the occurrence of friction. The results showed that characteristics of the woman’s housing (type of home - owned or not - and whether she intends to move) have an influence on her participation in another wave, in addition to the factor of having children. Characteristics of the census sector in which the participant resides and the relationship between interviewee and interviewer are also explored. Knowing the existence of sample attrition, it is possible that it will be necessary to replace some women during the study longitudinal. The second chapter then aims to verify whether the replacement protocol adopted by PCSVDF-Mulher was complied with by the contracted company and whether the sample of women who are classified as replacement has the same profile as those who stopped participating in the research (attrited), considering that it is necessary to maintain the statistical representativeness of the sample. For this, spatial and statistical techniques are used with the aim of verifying whether the company complied with the protocol in order to replace the sample within the sector census or in closer sectors and to compare replacement and attrited samples. The findings show that the company presented an optimizing behavior in relation to spatial replacement. In census tracts that recorded a deficit of households from one wave to the next, they are spatially closer to census tracts that had a surplus of households, with evidence that the replacements were spatially dependent. Furthermore, most of the variables compared between the samples, rubbed and replacement, had statistically equal distribution, thus concluding that the replacement was carried out satisfactorily. Finally, the third chapter analyzes another problem in longitudinal research, which is the logical inconsistency in the response, which can be explained by several areas, whether psychological, cognitive, types of importance of events, among others. In particular, on the topic of domestic violence, inconsistent reports of intimate partner violence (IPV) potentially lead to inaccurate estimates of the prevalence of such behaviors. For inconsistency regarding general violence, in PCSVDF-Mulher, between waves 1 and 2, a rate of 18.08% was found, and, when disaggregated by type of violence, emotional, physical and sexual inconsistency reached percentages equal to 17.17%, 12.57% and 6.54%, respectively, reflecting a different profile. The results found indicate that knowledge variables, whether regarding the woman’s education or knowledge about the Maria da Penha Law, have negative relationship with inconsistent reporting. On the other hand, if the woman has children or has been a recent victim of violence, it has a positive relationship with inconsistency.