Machine Learning Meets Tradition to Reveal Who Is Missing Out on Family Planning in Bangladesh

Millions of women worldwide want to delay their next pregnancy or stop having children altogether, yet they are not using any form of contraception. Public health researchers call this the unmet need for family planning, and it remains one of the most stubborn obstacles to safer motherhood, healthier families, and women’s autonomy. A new analysis of national survey data from Bangladesh, published in PLOS Global Public Health, takes an unusually rigorous look at who falls into this gap and why, combining the statistical tools that have guided family planning programs for decades with machine learning methods that can detect patterns traditional models tend to overlook.

The study, led by Muhammad Khairul Alam and Rumana Rois, draws on the Bangladesh Demographic and Health Survey of 2022, a nationally representative household survey that collects detailed information on fertility, contraceptive use, and household characteristics. Rather than treating unmet need as a single undifferentiated category, the researchers separated it into its two distinct components: the need for spacing births, which applies to women who want to wait at least two years before their next pregnancy, and the need for limiting family size, which applies to women who want no more children. This distinction matters because the two groups face different barriers, require different counseling approaches, and respond to different kinds of contraceptive services.

The analytical strategy unfolded in two stages. First, the team built a conventional multivariable multinomial logistic regression model, the workhorse method for estimating how demographic and socioeconomic factors shift the odds of belonging to one category rather than another. This model produced adjusted odds ratios, the familiar risk profiles that program planners can act on directly. Second, the researchers trained two ensemble machine learning classifiers, Random Forest and XGBoost, on the same data to see whether these flexible algorithms could capture non-linear relationships and complex interactions between variables that a regression equation would smooth over or miss entirely.

Working with survey data of this kind presents a well-known technical challenge: severe class imbalance. Most women in the survey either use contraception or have no need for it, so the categories representing unmet need contain relatively few observations. Machine learning classifiers trained on imbalanced data tend to simply predict the majority class, achieving deceptively high accuracy while failing to identify the minority cases that matter most. To counter this, the researchers applied a synthetic hybrid resampling technique based on SMOTE, the Synthetic Minority Over-sampling Technique, which generates synthetic examples of the underrepresented classes to balance the training data and give the algorithms a fair chance to learn the rarer patterns.

Feature selection posed another methodological question: with dozens of potential predictors, which ones genuinely carry signal? The team used the Boruta algorithm, a wrapper method built on Random Forests that compares the importance of real variables against shuffled, randomized versions of themselves. Only features that consistently outperform their random shadows are retained, providing a conservative filter against noise. To interpret what the trained models had actually learned, the researchers turned to Shapley Additive Explanations, or SHAP, a technique borrowed from cooperative game theory that assigns each feature a contribution value for every individual prediction, making it possible to visualize both global patterns and local interactions across the entire dataset.

The bivariate analysis revealed significant associations between unmet need and several key characteristics: maternal age, parity, household wealth, urban or rural residence, and the woman’s decision-making power within the household, all at the conventional threshold of statistical significance. The multivariable regression then sharpened these findings into striking effect sizes. Women with four or more children faced dramatically elevated odds of unmet need, with an adjusted odds ratio of 20.60 for spacing and 2.24 for limiting. In other words, high-parity mothers who still wanted to space a future birth were more than twenty times as likely to be without contraception as their lower-parity counterparts, a signal that current family size alone does not capture the full complexity of reproductive intentions.

Wealth emerged as a protective factor, particularly for limiting needs. Women in the richest household quintile had an adjusted odds ratio of 0.48 for unmet need for limiting, meaning their odds were roughly half those of the poorest women. This gradient points to the role of economic access: wealthier households can more easily reach clinics, pay for methods, absorb the indirect costs of seeking care, and negotiate the logistics of obtaining supplies. The SHAP analysis added exploratory depth, suggesting non-linear interactions between residence and factors such as parity and education, hinting that the barriers facing rural women may operate differently depending on how many children they have and how much schooling they received.

Yet the study delivers an honest verdict on the machine learning component. The predictive accuracy of the Random Forest and XGBoost models was only moderate, with macro-averaged areas under the receiver operating characteristic curve comparable to those achieved by traditional regression. The sophisticated algorithms did not outperform the simpler statistical model in ways that would justify relying on them for population-level inference. The authors therefore conclude that conclusions about the population should be drawn primarily from the multinomial logistic regression, while the SHAP-based insights into interactions remain exploratory rather than definitive. This is a valuable caution for a field where machine learning is sometimes presented as a wholesale replacement for classical epidemiological methods.

The cross-sectional design of the survey imposes further limits. A single snapshot in time cannot establish causation, and unmet need for spacing is often a fluid, transient state shaped by relationships, recent births, and shifting intentions. A woman classified as having an unmet need today may adopt a method next month or decide she wants another child soon. The researchers argue that this fluidity means programs should not treat the unmet need gap as a fixed deficit to be closed by a single intervention, but as a dynamic condition that requires continuous, responsive contact with women across their reproductive lives.

The practical implications are concrete. The authors recommend moving away from broad media campaigns toward localized, face-to-face counseling that addresses the specific circumstances of each woman. They call for ensuring a stable supply of spacing methods in rural areas, where stockouts and distance can quietly push women into the unmet need category. Perhaps most strikingly, they emphasize involving husbands and mothers-in-law in family planning conversations, reflecting evidence that reproductive decisions in Bangladesh are frequently negotiated within the household rather than made by women alone. Turning contraception from a private, sometimes contested choice into a shared household conversation, the study suggests, may be the key to closing a gap that decades of national programs have not yet eliminated.

Subject of Research: Unmet need for spacing and limiting family planning in Bangladesh, analyzed with regression and machine learning on national survey data

Article Title: Factors associated with the unmet need for spacing and limiting family planning: Evidence from a national survey

Article References: Alam, M. K., & Rois, R. (2026). Factors associated with the unmet need for spacing and limiting family planning: Evidence from a national survey. PLOS Global Public Health, 6(10), e0007412. https://doi.org/10.1371/journal.pgph.0007412

Image Credits: AI Generated

DOI: 10.1371/journal.pgph.0007412

Keywords: family planning, unmet need, Bangladesh, Demographic and Health Survey, contraception, machine learning, Random Forest, XGBoost, SHAP, multinomial logistic regression, SMOTE, reproductive health