Predicting Contraceptive Discontinuation in Nigeria: A Machine Learning Analysis of Family-Planning Method Adherence Using NDHS 2023–24 Calendar Data
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Contraceptive discontinuation — stopping a method while still at risk of an unintended pregnancy — is a major but under-examined driver of unmet need in Nigeria, where modern contraceptive prevalence remains far below the country’s FP2030 target. Distinguishing discontinuation that a health system could plausibly prevent from discontinuation that reflects a woman’s achieved fertility intentions is essential for targeting intervention, but this distinction has not been studied systematically, or compared algorithmically, in the Nigerian literature.
Methods
We used the reproductive calendar from the 2023–24 Nigeria Demographic and Health Survey to create 14,178 contraceptive use episodes for 7,723 women and identified two outcomes: Gap A, twelve-month discontinuation (n = 11,624), and Gap B, among women who discontinued within twelve months, method-related versus intentional discontinuation (n = 3,701). We evaluated four algorithms (survey-weighted logistic regression, elastic net, Random Forest, and XGBoost) on the same held-out test set (80/20) and reported the same area under the receiver operating characteristic curve (AUC).
Results
Discontinuation after 12 months was 32.3 percent (survey-weighted), a value that is very similar to the Nigeria-published life-table value of 37.2 percent. The best-performing algorithm in both cases, Gap A (AUC = 0.699) and Gap B (AUC = 0.664), was Random Forest, while the worst was XGBoost (AUC = 0.621) in Gap B, which struggled even with the basic logistic baseline. The most common predictor across all four algorithms and both outcomes was contraceptive method type, with long-acting reversible contraceptive (LARC) much less likely to be discontinued than short-acting methods. Among individuals who discontinued within 12 months, reasons were almost evenly split between method-related (49.7 percent) and intentional (49.2 percent).
Conclusions
Factors (side effects, cost, access, and efficacy dissatisfaction) account for approximately half of Nigeria’s contraceptive discontinuation problem and could be reduced with counseling, task shifting to pharmacists, and ongoing investments to increase the availability of longer-acting methods (LARCs) in Nigeria. Algorithm performance is outcome-dependent, not fixed, and it is better to compare algorithms by outcome than to default to a particular algorithm in DHS-based machine learning research.
