TL;DR
This paper applies machine learning to identify factors associated with child marriage in India, moving beyond traditional regression to uncover novel insights. By combining advanced ML techniques with an iterative thematic analysis, the researchers highlight the under-recognized role of poor integration with health systems and inadequate nutrition as key correlates, suggesting new avenues for intervention.
What Problem Does This Paper Address?
Child marriage remains a significant global issue, particularly in South Asia, with profound negative impacts on the health and human rights of girls and their children. While global efforts have led to a decline, the pace is insufficient to meet Sustainable Development Goals. Existing research primarily relies on traditional demographic and epidemiological analyses, which may overlook complex or less obvious correlations. The authors aim to leverage machine learning to identify novel factors associated with child marriage, thereby expanding the targets for intervention and accelerating its elimination. They also seek to validate machine learning's utility in this domain and develop a more robust methodology for hypothesis generation.
Methods
The study employed a multi-stage machine learning approach using data from India's National Family Health Survey (NFHS-4).
- Data Pre-processing: Domain experts reviewed and cleaned the data, eliminating irrelevant variables and categorizing others. Categorical variables were one-hot encoded, and continuous variables were normalized, resulting in over 6500 features per individual.
- Feature Selection (Stage 1): A Least Absolute Shrinkage and Selection Operator (LASSO) or L-1 regularized logistic regression model was used for initial feature selection. This technique helps reduce dimensionality and identify the most relevant features by shrinking less important coefficients to zero.
- Predictive Modeling (Stage 2):
- L2 Regularized Logistic Regression (Ridge): This model was applied to the selected features from LASSO. Ridge regression shrinks coefficients towards zero but doesn't force them to be exactly zero, useful for identifying variables with potential relationships.
- Neural Network Model: A feed-forward neural network with four hidden layers was also employed. Neural networks are powerful for identifying non-linear patterns and feature combinations.
- Iterative Categorization and Theme Generation: This is the paper's novel contribution. Building on the results from the ridge regression model, the researchers used an iterative process inspired by grounded theory. Two domain experts qualitatively coded the identified features into themes. This process involved multiple iterations where themes contributing the most variance were removed, allowing for the emergence of new, less obvious themes. This iterative approach aimed to uncover hidden patterns and generate hypotheses.
Datasets
The study utilized data from the National Family Health Survey (NFHS-4), a nationally representative Indian household survey conducted from 2015 to 2016. The analysis focused on women aged 20-24 years who were ever-married (N = 78,542), as this age group allows for the analysis of child marriage as a more recently practiced phenomenon.
Key Insights and Ideas
- Machine Learning Outperforms Traditional Methods: Regularized regression (LASSO, Ridge) and neural networks demonstrated superior accuracy and lower error rates compared to traditional logistic regression, highlighting their potential for complex social science data.
- Beyond Socioeconomic Factors: While traditional correlates like poverty, geography, and fertility remain important, the study identified non-utilization of health system benefits (specifically related to nutrition services like Anganwadi/ICDS centers) and inadequate diet as significant, previously under-emphasized factors associated with child marriage.
- Iterative Thematic Analysis is Crucial: The novel iterative theme generation approach, combined with expert qualitative coding, was instrumental in uncovering these less obvious but important correlates, which might have been masked by more dominant factors in standard analyses.
- Health System Integration is Key: The findings suggest that girls and women affected by child marriage may have poorer integration with essential public health and nutrition services, potentially explaining observed adverse maternal and child health outcomes.
Key Results
- Model Performance:
- Traditional Logistic Regression: ROC AUC of 79%, BER of 49%.
- Ridge Logistic Regression: ROC AUC of 91%, BER of 16%.
- Neural Network Model: ROC AUC of 90%, BER of 34%.
- Identified Themes: The iterative process generated nine explanatory themes, including sexual history, marriage, reproductive history, geography, socio-economic status, family planning, receipt of benefits from health systems, exposure to media, and diet.
- Novel Correlates: Key emergent themes included "receipt of benefits from health systems" (specifically non-receipt of Anganwadi/ICDS services) and "diet" (e.g., never eating eggs, occasional fruit intake), which were not as prominent in prior research.
Limitations and Future Work
- Data Comprehensiveness: The survey data, while extensive, may not have captured all potentially relevant demographic or socioeconomic features, potentially limiting the identification of all possible themes.
- Measurement Error: Reliance on self-reported data is subject to recall and social desirability biases.
- Cross-Sectional Data: The study design precludes causal inference; identified "predictors" are correlates.
- "Black Box" Nature of Neural Networks: While powerful, neural networks can make it difficult to understand the precise reasoning behind individual predictions. The paper suggests including approaches like ridge regression alongside neural networks to provide more interpretability.
- Future Work: The authors suggest further research into barriers to public health program use among adolescent girls and young women. They also emphasize the need for targeted public health and nutrition interventions, especially in areas with high rates of child marriage, and highlight the ongoing importance of child marriage prevention efforts, particularly in light of potential pandemic-related increases. They also suggest exploring other machine learning algorithms and potential interactions between variables.
Summary updated