Using Causal Inference to Investigate Contraceptive Discontinuation in Sub-Saharan Africa

Victor Akinwande, Megan A. MacGregor, Celia Cintas, Ehud Karavani, Dennis Wei, Kush R. Varshney

Abstract: Contraceptive discontinuation—stopping a method while still wishing to avoid pregnancy—poses a significant challenge to reproductive health in developing regions. While Randomized Controlled Trials (RCTs) are ideal for establishing causality, they are often unfeasible at scale. This study utilizes machine learning-based causal inference on observational data from the Demographic and Health Surveys (DHS) across seven Sub-Saharan African countries. By employing the OverRule algorithm to identify valid sub-populations and propensity score weighting to adjust for socio-economic confounders, we investigate the discontinuation risks associated with injectable contraceptives. Our findings reveal that while injectables exhibit higher discontinuation rates due to health concerns, these risks are significantly influenced by socio-economic factors. Causal adjustment resulted in risk ratios between 1.08 and 1.69, providing more accurate insights for targeted policy interventions.

Introduction

Understanding the drivers of contraceptive discontinuation is crucial for improving reproductive health outcomes. Associative studies often fail to distinguish between the effects of the contraceptive method itself and confounding socio-economic factors such as wealth, education, and geographic location. For instance, if specific demographics are more likely to choose injectables and also face greater barriers to healthcare, simple correlations may misattribute discontinuation to medical side effects. This research applies a potential outcomes framework to observational data to isolate the causal impact of injectable contraceptives on discontinuation due to health concerns.

Methodology and Data

The study analyzed the DHS Contraceptive Calendar (2016–2018) for Burundi, Ethiopia, Liberia, Nigeria, Sierra Leone, Uganda, and Zambia, encompassing 20,703 individuals. To address selection bias, the researchers implemented:

  • Propensity Score Weighting: Inverse Probability Weighting (IPW) and Overlap Weighting (OW) were used to balance 23 socio-economic variables between treatment (injectables) and control (other modern methods) groups.
  • OverRule Algorithm: This method identifies sub-populations where the "overlap" condition for causal inference is met, generating interpretable Boolean rules for demographic targeting.
  • Calibrated Logistic Regression: Used to ensure robust propensity score estimation.

Key Results and Insights

The analysis identified a "Popularity-Risk Paradox": injectables are the most prevalent modern contraceptive yet show the highest discontinuation rates for health reasons. However, the study found that the "unadjusted" risk of discontinuation was consistently higher than the "causal" risk calculated after socio-economic adjustment. This suggests that a portion of reported health concerns may stem from external stressors or lack of follow-up care rather than the method itself.

Risk ratios for full and overlap samples across seven Sub-Saharan African countries.

As shown in the figure above, the risk ratios vary significantly by country. In Nigeria, Sierra Leone, Uganda, and Burundi, the increased risk was statistically significant. The OverRule algorithm confirmed that these causal estimates are generalizable to 91%–95% of the studied populations.

Detailed risk ratio metrics for Marginal, IPW, and OW models across all studied countries.

The table above provides the precise Average Treatment Effect (ATE) risk ratios. For example, in Nigeria, the IPW-adjusted risk ratio of 1.67 indicates that injectable users are 67% more likely to discontinue due to health concerns than users of other modern methods, a figure notably lower than the unadjusted marginal risk of 2.01.

Discussion and Limitations

The application of the OverRule algorithm allows for actionable interpretability, enabling policymakers to identify specific demographic "bins" (e.g., based on age and education) most at risk. While the study is subject to limitations such as recall bias and the assumption of no unmeasured confounders, it demonstrates the power of causal machine learning in refining public health policy. Future research using more frequent survey data could provide even more timely guidance for reproductive health programs.

Summary updated

BibTeX
@inproceedings{akinwande2024using,
  author = {Victor Akinwande and Megan A. MacGregor and Celia Cintas and Ehud Karavani and Dennis Wei and Kush R. Varshney},
  title = {{Using Causal Inference to Investigate Contraceptive Discontinuation in Sub-Saharan Africa}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=Whd4u9BVvL},
}
Figure