TL;DR
This paper uses machine learning on a massive Indian health survey dataset to uncover surprising drivers of low Intrauterine Device (IUD) usage. The findings highlight that shared family planning goals between couples, particularly involving male engagement, are the strongest predictors, suggesting a shift from individual-focused interventions to couple-centric approaches.
What Problem Does This Paper Address?
Despite being a safe, effective, and low-cost method for delaying or spacing pregnancies, Intrauterine Device (IUD) uptake in India remains remarkably low. While efforts have been made to increase access to modern contraceptives, India continues to heavily rely on female sterilization, often used only after desired family size is achieved. Existing research on IUD use determinants is limited, often relying on smaller samples or focusing primarily on supply-side factors. This paper aims to address this gap by using advanced machine learning techniques on a large, nationally representative dataset to explore a broad range of potential demand-side factors influencing IUD use, particularly in a low-prevalence context.
Methods
The researchers employed a combination of machine learning (ML) techniques and iterative thematic analysis (ITA) to explore the determinants of IUD use.
- Machine Learning Models:
- Lasso (L1 Regularized Logistic Regression): Used for initial feature selection to reduce the dimensionality of the dataset by shrinking less relevant coefficients to zero.
- Ridge (L2 Regularized Logistic Regression): Applied after Lasso to further refine the model, reduce multicollinearity, and mitigate overfitting with the reduced set of predictors.
- Neural Network (NN) Model: A feed-forward neural network with four hidden layers was used to capture potential non-linear relationships among predictors and to validate findings from the regularized models.
- Iterative Thematic Analysis (ITA): This qualitative component was crucial for interpreting the ML model outputs. Content experts thematically coded variables with significant coefficients from the ML models. This process was repeated iteratively: after identifying themes, the most influential variables within those themes were removed, and the ML models were re-run. This iterative process helped to generate coherent themes and identify key constructs associated with IUD use.
- Model Performance: Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC) and the Balanced Error Rate (BER). K-fold cross-validation was used to estimate hyperparameters for the regularized regression models.
Datasets
The study utilized data from the 4th National Family Health Survey (NFHS-4), conducted in India from 2015 to 2016. This survey is a nationally representative dataset that collected information on population, health, and nutrition. The analytic sample was limited to 499,627 currently married women aged 15-49 years due to the very low prevalence of IUD use among unmarried women.
Key Insights and Ideas
- Couple Dynamics Trump Individual Factors: The most striking finding is that shared family planning goals between husbands and wives were the strongest predictor of IUD use. This suggests that interventions need to move beyond focusing solely on women and actively engage men in family planning decisions.
- Male Engagement is Key: The emphasis on shared goals points to the critical role of male support and communication in a couple's decision-making process regarding contraception.
- IUDs for Spacing, Not Just Limiting: The correlation between the desire for "no more children" and IUD use, while seemingly counterintuitive given sterilization's role in limiting births, suggests a potential misunderstanding or lack of awareness about IUDs as an effective method for birth spacing.
- Integrated Healthcare Matters: Access to a broader range of maternal and child health (MCH) services, alongside quality family planning counseling, emerged as significant predictors, highlighting the importance of a holistic approach to reproductive healthcare.
- ML for Low-Prevalence Phenomena: The study successfully demonstrates the utility of ML techniques, particularly when combined with ITA, for uncovering determinants of behaviors with very low prevalence, which can be challenging for traditional statistical methods.
Key Results
- Strongest Predictor: Couples' shared family planning goals were the most significant determinant of IUD use.
- Other Key Predictors: Following shared goals, other important factors included:
- Receipt of family planning (FP) services.
- Desire for no more children.
- Higher wealth and education levels.
- Receipt of maternal and child health (MCH) services.
- Model Performance:
- The AUC for the first round of ridge regression was 88% with a BER of 22%.
- The AUC for the second round was 87% with a BER of 23%.
- The Neural Network model achieved a very high AUC of 99% and a low BER of 2%.
- Prevalence: Approximately 1.5% of currently married women aged 15-49 years reported using IUDs.
Limitations and Future Work
- Cross-Sectional Data: The study relies on cross-sectional data, which limits the ability to establish causality. Social desirability and recall biases are also potential concerns.
- Limited Variable Scope: The analysis is constrained by the variables available in the NFHS-4 dataset. Potentially relevant factors, such as awareness of IUD side effects or fear of insertion, were not directly captured.
- ML Model Limitations: The ML models used do not account for confounding and are exploratory rather than hypothesis-driven. Other ML approaches, like tree-based methods, might offer different insights.
- Need for Qualitative Research: While the study provides broad, generalizable findings, a deeper, nuanced understanding of why and how women engage with IUDs would benefit from qualitative research.
Future Work Suggestions: The authors suggest that future research should:
- Focus on couple-centric interventions that actively engage men in family planning.
- Emphasize IUDs as an effective option for birth spacing in FP counseling.
- Explore the impact of integrated reproductive and maternal care services.
- Conduct mixed-methods research, combining quantitative findings with qualitative insights.
Summary updated