TL;DR
This paper applies machine learning, specifically iterative thematic analysis (ITA), to a large Indian survey dataset to identify factors associated with non-marital sexual violence (NMSV). It reveals that past exposure to violence, geography, and sexual/reproductive health knowledge are key predictors, and importantly, that economic agency and mobility are associated with higher risk, challenging assumptions about vulnerability.
What Problem Does This Paper Address?
Non-marital sexual violence (NMSV) is a significant public health and human rights issue in India, disproportionately affecting adolescents and young adults. Despite its prevalence, NMSV is under-reported due to stigma, making it difficult to understand its contributing factors through traditional research methods. Existing studies often narrow down variables to avoid multiple comparisons, potentially missing crucial insights from large, complex datasets. This paper aims to leverage machine learning to explore a broader range of potential correlates of NMSV, providing a more nuanced understanding of this under-researched problem.
Methods
The study employs a combination of machine learning techniques, primarily supervised learning models, to analyze a large dataset. The core methodology involves:
- Regularized Logistic Regression (Lasso and Ridge): Lasso (L-1 regularization) is used for variable selection and shrinkage, forcing some coefficients to zero to identify the most relevant predictors. Ridge (L-2 regularization) then shrinks the remaining coefficients, helping to identify variables with potential relationships to the outcome.
- Iterative Thematic Analysis (ITA): This novel approach, developed by the research team, combines the statistical outputs of the regularized models with qualitative coding by domain experts. Variables are grouped into themes, and this process is iterated to refine understanding and identify new themes.
- Neural Networks: Feed-forward neural networks are used to capture non-linear relationships among predictors, offering an alternative perspective to the linear models.
The analysis was conducted in two main parts: first, on the complete sample of women aged 15-49, and second, on a subsample of never-married adolescent girls aged 15-19, a group particularly vulnerable to NMSV.
Datasets
The primary dataset used is the National Family Health Survey (NFHS-4), a nationally representative, household-based survey conducted in India during 2015-16. This dataset is extensive, containing over 4000 variables. For this study, a subsample of women who responded to questions about experiences of violence was used. A further subsample of never-married adolescent girls (15-19 years) was also analyzed. No new benchmarks were introduced; the study leverages existing, publicly available data.
Key Insights and Ideas
- Machine Learning as a Tool for Under-Reported Phenomena: The study demonstrates the effectiveness of machine learning, particularly with ITA, in uncovering factors associated with issues like NMSV that suffer from low reporting rates.
- Challenging Vulnerability Assumptions: Contrary to the expectation that the most socially vulnerable girls are at higher risk, findings suggest that economic agency (income generation, control over income) and mobility are associated with increased odds of experiencing NMSV. This implies that factors beyond traditional markers of vulnerability play a significant role.
- The Interplay of Violence and Knowledge: Past exposure to violence is a strong predictor, but the study also highlights the importance of sexual and reproductive health (SRH) knowledge and service utilization as key correlates. Low SRH knowledge and service access are linked to NMSV.
- Adolescent Girls' Complex Risk Profile: For adolescent girls, attitudes accepting of marital violence, combined with factors like poor HIV knowledge and higher fertility preferences, emerge as significant correlates, suggesting a complex interplay of social norms and health literacy.
Key Results
- Overall Sample (15-49 years): The most predictive themes for NMSV included experiences of/exposure to violence, geography, economic circumstances/engagement/empowerment, and health care access/use. Novel themes identified included knowledge of and access to SRH services and sexual behavior.
- Adolescent Girls (15-19 years): In addition to violence exposure, attitudes toward marital violence, freedom of movement/mobility, nutrition, access to information/media, tobacco/alcohol consumption, menstrual health, and social positioning were identified as correlates.
- Model Performance: The iterative models showed good reliability, with AUC values ranging from 73% to 80% for the complete sample and 77% to 86% for the adolescent subsample. The neural network models also performed well, with AUC values around 79% for the complete sample and 74% for the adolescent sample.
Limitations and Future Work
- Self-Reported Data and Under-Reporting: The study acknowledges that data relies on self-report, and NMSV is likely under-reported due to stigma. This can lead to reporting biases and potentially underestimate the true prevalence and correlates.
- Cross-Sectional Data: The data is cross-sectional, meaning causality cannot be definitively established. For example, it's unclear if economic agency is a cause or a consequence of NMSV.
- Limited NMSV Focus in Survey: The NFHS-4 survey, while comprehensive, is not specifically designed to assess NMSV, potentially leading to inadequate correlates and measurement error.
- Future Work: The authors suggest the need for machine learning analysis on longitudinal data to explore causality. They also emphasize the importance of interventions that address sexual and reproductive health, gender equity, and safety, particularly for adolescents, and call for further research into why women with economic agency might be at increased risk.
Summary updated