Causal effects of Antenatal Care (ANC) on child malnutrition: a machine learning approach in Ethiopia and Rwanda Authors: Alehegn Moges Tessema, Temesgen Zewotir, Richard Kabanda Published: 2025 ABSTRACT Malnutrition among children under five remains a critical public health challenge in Ethiopia and Rwanda, with nearly half of the children in this study population affected by at least one form of undernutrition. This study provides a robust causal analysis of the heterogeneous effects of Antenatal Care (ANC) on child malnutrition, as measured by the Composite Index of Anthropometric Failure (CIAF). Leveraging data from 33,737 mother-child pairs from the Demographic and Health Surveys (DHS) in Ethiopia (N = 22,668) and Rwanda (N = 11,069) across three waves (2005–2015), we employ a machine learning-based Causal Forest model. This approach was chosen specifically to overcome the limitations of traditional regression methods, allowing for the estimation of how the impact of different levels of ANC attendance varies across diverse sociodemographic and health-related subgroups. The results reveal a clear and powerful dose-response relationship. While a single ANC visit has a negligible effect, attending 2–3 visits is associated with a modest 3.5% point reduction in the risk of malnutrition. The strongest impact is seen with the completion of 4 or more visits, which is associated with an average reduction of 5.7% points (ATE: -0.057). Crucially, this average effect masks profound and policy-relevant heterogeneity. The benefits of ANC are massively amplified for the most vulnerable populations; for children in the poorer wealth quintile, 4 + ANC visits are associated with a massive 17.1% point reduction in malnutrition risk—an effect nearly three times the population average. The intervention is also particularly impactful for older mothers (aged 35–49), where it is associated with an 11.9% point risk reduction. Furthermore, the benefits of ANC are amplified by a healthier environment; for children in households with an improved water source, the associated risk reduction is a substantial 7.8% points, demonstrating a powerful synergistic effect between clinical care and public health infrastructure. The application of the Causal Forest model represents a significant advancement, moving beyond a single average effect to uncover this critical heterogeneity and identify for whom, and under what conditions, ANC is most effective. These findings provide robust causal evidence for a necessary shift away from a one-size-fits-all public health strategy. The results strongly advocate for stratified and multi-sectoral interventions. Policy and resources should be intensely focused on ensuring the most vulnerable populations—particularly the poorest households and older mothers—complete the full ANC schedule, as this is where the public health return on investment is highest. Additionally, the synergistic effect with WASH highlights the need to integrate clinical maternal health programs with investments in community-level water and sanitation infrastructure. By tailoring ANC programs and combining them with broader public health improvements, governments can develop more holistic and effective strategies to accelerate progress against child malnutrition in Ethiopia, Rwanda, and comparable high-burden contexts. Introduction Malnutrition persists as a significant public health issue in Ethiopia and Rwanda, especially among children under five, contributing to high rates of illness, mortality, and impaired development [1]. Among the various strategies to combat this challenge, Antenatal Care (ANC) is promoted as a cornerstone of maternal and child health, yet its direct causal impact on child nutritional outcomes remains insufficiently explored [2]. In response to this high burden, both nations have implemented ambitious national policies. Ethiopia’s National Nutrition Program (NNP II) and its successor, the National Food and Nutrition Policy, set out multi-sectoral strategies [3, 4], while Rwanda’s National Food and Nutrition Policy and National Strategy for Transformation (NST1) similarly prioritize a comprehensive approach [5, 6]. However, despite these ambitious frameworks, malnutrition levels remain persistently high, with stunting affecting approximately 37% of children in Ethiopia and 33% in Rwanda [7, 8]. This persistence suggests critical gaps between policy design and on-the-ground implementation. Key challenges common to both nations include inconsistent service quality, weak integration across sectors like health and agriculture, and socioeconomic disparities that limit the ability of the most vulnerable households to benefit from existing programs [9, 10]. These implementation gaps create an urgent need for more targeted and effective interventions that can translate national policy into tangible improvements in child health. To build a robust model capable of exploring these challenges, this study’s pooled analysis of Ethiopia and Rwanda is a deliberate methodological choice. First, both nations face a severe and persistent burden of child malnutrition, making them critical settings where evidence-based interventions are urgently needed. Second, both countries have built their primary healthcare strategies on a similar foundational platform: a formalized cadre of frontline health workers—Health Extension Workers in Ethiopia and Community Health Workers in Rwanda—acting as the essential bridge to the formal health system [11, 12]. This shared reliance on a structured community platform makes a pooled analysis of ANC’s impact both valid and powerful. Building on this, our approach leverages the combined dataset to enhance statistical power while including a country variable to statistically control for fixed effects like governance and health financing [13, 14]. A substantial body of research has explored the determinants of child malnutrition, providing a crucial foundation for public health policy. However, a critical review of this literature reveals two significant limitations that hinder the design of effective and equitable interventions. The first major limitation is that the vast majority of existing research, particularly that using large-scale observational datasets, establishes correlation rather than causation. While influential studies have documented strong associations between interventions like ANC and better child nutrition [15, 16], their methodological limitations make it difficult to prove causality due to the persistent challenge of selection bias [17]. This creates a serious challenge for policymakers, as acting on correlational evidence carries the risk of investing in an intervention that is merely a marker of healthier households, rather than the true cause of the improved outcome [18]. Beyond the challenge of establishing causality, a second limitation emerges even within the smaller set of studies that do employ more rigorous methods. These studies often focus on estimating a single, overarching Average Treatment Effect (ATE)—the impact of an intervention averaged across the entire study population [19]. Yet, as there is a growing consensus in the field of impact evaluation, treatment effects are rarely constant across different populations and contexts [20, 21]. This phenomenon, known as treatment effect heterogeneity, creates a crucial policy blind spot. A policy based on a single average effect may fail to target resources where they are needed most, leading to inefficient programs and the perpetuation of health inequities. To address these critical research gaps, this study applies the Causal Forest model—a machine learning approach for estimating treatment heterogeneity—allowing for a robust estimation of how ANC impacts malnutrition across various subgroups. While this study primarily focuses on the causal impact of ANC, it also considers related factors such as maternal education and access to clean water to understand how they interact with and modify ANC’s effects. By identifying for whom and under what conditions ANC is most impactful, this study aims to contribute context-specific evidence to inform the design of more effective and equitable health policies, helping guide policymakers as they refine national strategies to accelerate progress against persistent child malnutrition. Methods The data sources and preparation This study utilized data from three comparable waves of the Demographic and Health Surveys (DHS) in Ethiopia and Rwanda (2005, 2010, 2015). Our selection criteria prioritized methodological consistency, using full DHS waves that contained all necessary variables for both countries across similar time periods. This led to the exclusion of the most recent survey round (2019–2020) from our pooled analysis. Specifically, Ethiopia’s 2019 Mini DHS was unsuitable as it lacked the anthropometric measurements required to construct our outcome variable. To maintain a balanced and comparable dataset across both nations, the corresponding 2019-20 DHS from Rwanda was therefore also excluded. The study’s objective was to estimate the causal effect of multi-level ANC attendance (0, 1, 2–3, and 4 + visits) on child malnutrition, as measured by the Composite Index of Anthropometric Failure (CIAF)—a composite indicator of stunting, wasting, or underweight. The data for this study were extracted from the Kids’ Record (KR) and Individual Recode (IR) files of the Demographic and Health Surveys (DHS) datasets. These files provide comprehensive information on household demographics, maternal and child health indicators, socioeconomic characteristics, and access to essential services such as healthcare, water, and sanitation. Based on established literature and their theoretical role as potential confounders, key independent variables were selected, including the wealth index, maternal education, and various health and environmental factors. Detailed definitions for each variable included in the analysis are provided in Table 1 below.Table 1Descriptions of the response variables and the respective covariates included in the modelVariable NameCategoriesData TypeDescriptionComposite Index of Anthropometric Failure (CIAF)0 = No1 = YesBinary(Outcome) The Composite Index of Anthropometric Failure. A binary indicator where 1 indicates the child has any form of malnutrition (stunting, wasting, or underweight).Number of ANC attendance0 = None1 = 1 visit2 = 2–3 visits3 = 4 + visitsOrdinal(Treatment) The number of Antenatal Care visits attended by the mother.Country1 = Ethiopia2 = RwandaCategoricalThe country where the survey data was collected.Type of place of residence1 = Urban2 = RuralBinaryThe household’s place of residence.Highest educational level of woman0 = No education1 = Primary3 = Secondary4 = HigherOrdinalThe highest level of formal education attained by the mother.Sex of child1 = Male2 = FemaleBinaryThe biological sex of the child.Received Vitamin A supplement0 = No1 = YesBinaryWhether the child received a Vitamin A dose in the first two months after delivery.Child’s age in months (grouped)1 = 0–5 months2 = 6–11 months3 = 12–17 months4 = 18–23 months5 = 24–35 months6 = 36–47 months7 = 48–59 monthsOrdinalThe age of the child in months, grouped into seven standard categories (e.g., 0–5, 6–11, etc.).Wealth index1 = Poorest2 = Poorer3 = Middle4 = Richer5 = RichestOrdinalA composite measure of household economic status, ranked into quintiles.Survey Year (DHS Wave)1 = 20052 = 20103 = 2015CategoricalThe year of the Demographic and Health Survey (DHS) wave.Access to improved drinking water0 = No1 = YesBinaryWhether the household has access to an improved source of drinking water.Access to improved toilet facility0 = No1 = YesBinaryWhether the household has access to an improved, non-shared toilet facility.Iron supplementation during pregnancy0 = No1 = YesBinaryWhether the mother took at least one iron tablet or any syrup during her last pregnancy.Delivery in a health facility0 = No1 = YesBinaryWhether the child was delivered in a health facility versus a non-facility setting.Firstborn0 = No1 = YesBinaryWhether the child is the mother’s first live birth.Birth Type (Single vs. Twin)1 = Single2 = TwinsBinaryWhether the child was from a single or multiple (e.g., twins) birth.Child was ever breastfed0 = No1 = YesBinaryWhether the child was ever breastfedMaternal age (grouped)1 = 15–242 = 25–343 = 35–49OrdinalThe mother’s age in years, grouped into three categories.Size of child at birth1 = Smaller than average2 = Average3 = Above averageOrdinalThe mother’s subjective assessment of the child’s size at birth.Maternal Body Mass Index (BMI)1 = Underweight2 = Normal3 = Overweight/ObeseOrdinalThe mother’s Body Mass Index (BMI), grouped into three standard categories.Household size (grouped)1 = < 42 = 5–93 = 10+OrdinalThe total number of members in the household, grouped into three categories.Number of children under five in the household1 = 0–12 = 23 = 3 or moreOrdinalThe number of children under the age of five currently in the household. Table 1Descriptions of the response variables and the respective covariates included in the modelVariable NameCategoriesData TypeDescriptionComposite Index of Anthropometric Failure (CIAF)0 = No1 = YesBinary(Outcome) The Composite Index of Anthropometric Failure. A binary indicator where 1 indicates the child has any form of malnutrition (stunting, wasting, or underweight).Number of ANC attendance0 = None1 = 1 visit2 = 2–3 visits3 = 4 + visitsOrdinal(Treatment) The number of Antenatal Care visits attended by the mother.Country1 = Ethiopia2 = RwandaCategoricalThe country where the survey data was collected.Type of place of residence1 = Urban2 = RuralBinaryThe household’s place of residence.Highest educational level of woman0 = No education1 = Primary3 = Secondary4 = HigherOrdinalThe highest level of formal education attained by the mother.Sex of child1 = Male2 = FemaleBinaryThe biological sex of the child.Received Vitamin A supplement0 = No1 = YesBinaryWhether the child received a Vitamin A dose in the first two months after delivery.Child’s age in months (grouped)1 = 0–5 months2 = 6–11 months3 = 12–17 months4 = 18–23 months5 = 24–35 months6 = 36–47 months7 = 48–59 monthsOrdinalThe age of the child in months, grouped into seven standard categories (e.g., 0–5, 6–11, etc.).Wealth index1 = Poorest2 = Poorer3 = Middle4 = Richer5 = RichestOrdinalA composite measure of household economic status, ranked into quintiles.Survey Year (DHS Wave)1 = 20052 = 20103 = 2015CategoricalThe year of the Demographic and Health Survey (DHS) wave.Access to improved drinking water0 = No1 = YesBinaryWhether the household has access to an improved source of drinking water.Access to improved toilet facility0 = No1 = YesBinaryWhether the household has access to an improved, non-shared toilet facility.Iron supplementation during pregnancy0 = No1 = YesBinaryWhether the mother took at least one iron tablet or any syrup during her last pregnancy.Delivery in a health facility0 = No1 = YesBinaryWhether the child was delivered in a health facility versus a non-facility setting.Firstborn0 = No1 = YesBinaryWhether the child is the mother’s first live birth.Birth Type (Single vs. Twin)1 = Single2 = TwinsBinaryWhether the child was from a single or multiple (e.g., twins) birth.Child was ever breastfed0 = No1 = YesBinaryWhether the child was ever breastfedMaternal age (grouped)1 = 15–242 = 25–343 = 35–49OrdinalThe mother’s age in years, grouped into three categories.Size of child at birth1 = Smaller than average2 = Average3 = Above averageOrdinalThe mother’s subjective assessment of the child’s size at birth.Maternal Body Mass Index (BMI)1 = Underweight2 = Normal3 = Overweight/ObeseOrdinalThe mother’s Body Mass Index (BMI), grouped into three standard categories.Household size (grouped)1 = < 42 = 5–93 = 10+OrdinalThe total number of members in the household, grouped into three categories.Number of children under five in the household1 = 0–12 = 23 = 3 or moreOrdinalThe number of children under the age of five currently in the household. Descriptions of the response variables and the respective covariates included in the model 0 = No 1 = Yes 0 = None 1 = 1 visit 2 = 2–3 visits 3 = 4 + visits 1 = Ethiopia 2 = Rwanda 1 = Urban 2 = Rural 0 = No education 1 = Primary 3 = Secondary 4 = Higher 1 = Male 2 = Female 0 = No 1 = Yes 1 = 0–5 months 2 = 6–11 months 3 = 12–17 months 4 = 18–23 months 5 = 24–35 months 6 = 36–47 months 7 = 48–59 months 1 = Poorest 2 = Poorer 3 = Middle 4 = Richer 5 = Richest 1 = 2005 2 = 2010 3 = 2015 0 = No 1 = Yes 0 = No 1 = Yes 0 = No 1 = Yes 0 = No 1 = Yes 0 = No 1 = Yes 1 = Single 2 = Twins 0 = No 1 = Yes 1 = 15–24 2 = 25–34 3 = 35–49 1 = Smaller than average 2 = Average 3 = Above average 1 = Underweight 2 = Normal 3 = Overweight/Obese 1 = < 4 2 = 5–9 3 = 10+ 1 = 0–1 2 = 2 3 = 3 or more Analytical framework and causal inference methodology The study employs a multi-stage analytical framework designed to move beyond correlation and rigorously estimate the causal impact of Antenatal Care (ANC) on child malnutrition. The framework is built around the Causal Forest Double Machine Learning (DML) model, a state-of-the-art method chosen for its ability to flexibly control for a large number of potential confounders while simultaneously estimating heterogeneous treatment effects. The entire analysis was conducted using Python with the econml and scikit-learn libraries. The analytical framework, outlined in Fig. 1 follows process moves systematically from data preparation to final policy interpretation, ensuring robust and actionable findings. It employs a four-stage procedure described as follows.Fig. 1Analytical Framework Analytical Framework Stage 1: data balancing and confounder control The analysis begins by addressing potential biases in the dataset. Confounders—variables associated with both receiving ANC and child nutritional outcomes—were identified based on established evidence from maternal and child health literature. These include maternal education, household wealth, residence (urban/rural), and other key maternal health practices. To mitigate the risk of model bias from an imbalanced distribution of the ‘CIAF’ outcome, the Synthetic Minority Over-sampling Technique (SMOTE) was applied. This step creates a balanced dataset by generating synthetic samples for the minority class, ensuring the subsequent models learn the characteristics of all outcome groups effectively. After balancing, the Average Treatment Effect (ATE) is estimated, which quantifies the overall impact of ANC on child malnutrition across the entire population. Stage 2: causal forest model for heterogeneous effect estimation To gain deeper, policy-relevant insights, this study employs a Causal Forest Double Machine Learning (DML) model, an ensemble-based machine learning method that excels at estimating heterogeneous treatment effects. This model was chosen over traditional regression methods for its ability to handle high-dimensional confounders and automatically uncover complex, non-linear interactions without requiring pre-specification. This approach provides a more robust and detailed understanding of how the effects of ANC vary across the population. The model was implemented using the CausalForestDML function from the Python econml package. The DML framework operates by first “residualizing” the outcome (Y) and the treatment (T) to isolate the variation that is not explained by the confounders (W). The core of the model can be conceptually represented by the following steps: Model nuisance functions Two machine learning models (Random Forest Regressors) are trained to predict the outcome and the treatment based on the confounders. An outcome model estimates the expected outcome given the confounders: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\text{E}\left[\text{Y}\:|\:\text{W}\:=\:\text{w}\right]$$\end{document} An outcome model estimates the expected outcome given the confounders: An outcome model estimates the expected outcome given the confounders: A treatment model estimates the probability of receiving treatment given the confounders: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\text{E}\left[\text{T}\:|\:\text{W}\:=\:\text{w}\right]$$\end{document} A treatment model estimates the probability of receiving treatment given the confounders: A treatment model estimates the probability of receiving treatment given the confounders: Estimate causal effect The model then estimates the causal effect (τ) by examining the relationship between the residuals (the portion of the outcome and treatment unexplained by the confounders). 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:=Effect(Y\:-\:E[Y\:\left|\:W\right],\:T\:-\:E\left[T\:\right|\:W\left]\right)$$\end{document} A primary output of the Causal Forest is the Conditional Average Treatment Effect (CATE), denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:\left(w\right)\:=\:E\left[Y\right(1)\:-\:Y(0\left)\:\right|\:W\:=\:w$$\end{document}. The CATE analysis, conducted using the fitted causal forest model, estimates how ANC effectiveness varies across specific sociodemographic subgroups. This study examines CATEs for groups defined by residence, maternal education, household wealth, country, breastfeeding practices, access to clean water, sanitation, and vitamin A supplementation, thereby identifying which populations benefit most or least from the standard ANC intervention. Stage 3: model validation and robustness checks The reliability of the findings was verified using two key validation techniques to ensure the model’s stability and robustness. 5-Fold Cross-Validation: The stability of the ATE was assessed by training the model on four parts of the data and testing on a fifth, repeating this process five times. The final reported ATE is the average across all folds, ensuring the result is not an artifact of a specific data split. Analysis of Treatment Effect Distribution: The consistency of individual causal estimates was confirmed by analysing a histogram of the treatment effects. A distribution with low variance indicates that the model’s estimates are stable and reliable across the population. Stage 4: interpretation and policy implications The final stage translates the statistical results (ATE and CATEs) into actionable policy insights. By identifying specific subgroups where ANC is more or less effective—for instance, comparing rural vs. urban populations or differing wealth quintiles—the analysis provides clear evidence for policymakers. This allows for the strategic design of targeted interventions and the efficient allocation of resources to improve access to and quality of child nutrition services in underserved communities, ultimately leading to more equitable and impactful health outcomes. Results Characteristics of the study participants and prevalence of the Composite Index of Anthropometric Failure (CIAF) The study sample comprises 33,737 children under five and their mothers, drawn from three Demographic and Health Survey (DHS) waves (2005, 2010, and 2015) in Ethiopia (N = 22,668) and Rwanda (N = 11,069). While the two countries started from different socioeconomic and health system baselines, both demonstrated significant progress in reducing child malnutrition over the decade. The overall prevalence of malnutrition, measured by the Composite Index of Anthropometric Failure (CIAF), was high in both nations, affecting 49.2% of children in Ethiopia and 47.2% in Rwanda. This overall prevalence marks a substantial improvement over time; in Ethiopia, the CIAF rate fell from 59.0% in 2005 to 43.1% in 2015, while Rwanda saw a notable decline from 54.8% to 40.2% during the same period (Table 2).Table 2Socio-Demographic and health characteristics of mothers and children under five in Ethiopia and rwanda: DHS data from 2005, 2010, and 2015VariableEthiopiaRwanda2005 (N = 4122)2010 (N = 9752)2015 (N = 8794)Total (N = 22668)Number of Women2005 (N = 3653)2010 (N = 3937)2015 (N = 3479)Total (N = 11069)Number of WomenComposite Index of Anthropometric FailureNo4149.456.9*50.811,50745.253.559.8*52.85839Yes59.0*50.643.149.211,16154.8*46.540.247.25230Type of place of residenceUrban13.116.118.5*16.5373219.913.521.8*18.22015Rural86.9*83.981.583.518,93680.1*86.5*78.281.89054Highest educational level of womanNo education76.5*70.364.16915,64526.5*18.913.919.82193Primary16.624.9*25.5*23.6535664.371.9*72.2*69.57691Secondary6.0*3.36.7*5.111588.78.311.3*9.41036Higher0.81.53.7*2.25090.512.6*1.3149Sex of childMale50.950.951.351.111,57549.150.351.350.25559Female49.149.148.748.911,09350.949.748.749.85510Received Vitamin A supplement14.412.244.5*25.1569422.738.9*37.5*33.13667Child age group0–5 months9.511.2*11.1*10.8245511.6*8.89.61011056–11 months9.81010.410.1229810.610.211.710.8119612–17 months11.9*10.111.3*10.9246911.3*9.210.910.4115518–23 months8.37.98.38.118429.6109.79.8108424–35 months19.119.319.219.2436222.5*21.420.921.6239236–47 months21.3*21.2*19.120.4461916.920.320.9*19.4214548–59 months20.120.320.620.4462317.520.0*16.2181992Wealth indexPoorest26.63137.0*32.5737419.621.524.3*21.72405Poorer18.418.6*17.118407220.421.8*20.8212329Middle18.2*16.61415.935982020.1*1919.72184Richer17.5*16.2*12.415339521.0*19.316.919.12114Richest19.217.619.5*18.7422918.917.41918.42037Access to Improved Drinking Water58.152.162.2*57.1129453672.0*73.0*60.46688Access to Improved Toilet Facility9.915.9*16.9*15.2345296.9*73.370.780.38884Took Iron g PregnancyTablets Durin11.221.616.9*28.7650730.974.879.9*61.96854Delivery yin a Health Facilit9.212.835.3*20.9473229.374.392.0*657198Cornbhild is First16.918.519.6*18.6422016.724.827.9*23.12555Type of Birth (Single vs. Twin)Single98.6*97.797.797.822,18097.797.697.897.710,814Twins1.42.3*2.3*2.24882.32.42.22.3255Child was Ever Breastfed66.668.167.567.61531865.872.877.3*71.97959Number of ANC attendanceNone70.1*5836.151.711,7165.0*1.71.12.62871 visit3.64.65.1*4.6105113.1*4.63.77.17902–3 visits11.218.425.1*19.7445668.8*59.753.460.767194 or more visits15.21933.7*24544513.13441.9*29.63273Mothers’ age: 3-year groups15–2423.723.523.623.6534217.818.118.518.1200725–3450.252.352.45211,7795253.956.9*54.2600235–4926.1*24.22424.5554730.3*27.924.527.63060Size of child at birthSmaller than average30.6*28.530.5*29.6672038.445.9*38.541.14549Average41.639.942.4*41.2933750.4*38.846.845.24998Above average27.831.6*27.129.2661111.215.3*14.7*13.81522Maternal Body Mass Index (BMI)Underweight12.913.514.5*13.831184.5*2.82.13.2349Normal Weight66.9*75.37.3973.21659762.663.961.662.76943 Overweight/Obese20.2*11.211.713.0295333.033.336.3*34.13777Number of household members<423.324.627.1*25.4574827.936.640.2*34.938625–970.1*67.866.767.815,36667.5*59.957.461.7682510+6.67.6*6.26.915544.5*3.42.33.5382Number of Children Under Five0–132.132.434.9*33.3755427.636.944.9*36.44024249.7*47.9*4547.110,67948.6*49.9*43.947.652693 or more18.219.720.0*19.6443523.8*13.211.2161776Note: Values in the same row with * are significantly different at p < 0.05 in the two-sided test of equality for column proportions Table 2Socio-Demographic and health characteristics of mothers and children under five in Ethiopia and rwanda: DHS data from 2005, 2010, and 2015VariableEthiopiaRwanda2005 (N = 4122)2010 (N = 9752)2015 (N = 8794)Total (N = 22668)Number of Women2005 (N = 3653)2010 (N = 3937)2015 (N = 3479)Total (N = 11069)Number of WomenComposite Index of Anthropometric FailureNo4149.456.9*50.811,50745.253.559.8*52.85839Yes59.0*50.643.149.211,16154.8*46.540.247.25230Type of place of residenceUrban13.116.118.5*16.5373219.913.521.8*18.22015Rural86.9*83.981.583.518,93680.1*86.5*78.281.89054Highest educational level of womanNo education76.5*70.364.16915,64526.5*18.913.919.82193Primary16.624.9*25.5*23.6535664.371.9*72.2*69.57691Secondary6.0*3.36.7*5.111588.78.311.3*9.41036Higher0.81.53.7*2.25090.512.6*1.3149Sex of childMale50.950.951.351.111,57549.150.351.350.25559Female49.149.148.748.911,09350.949.748.749.85510Received Vitamin A supplement14.412.244.5*25.1569422.738.9*37.5*33.13667Child age group0–5 months9.511.2*11.1*10.8245511.6*8.89.61011056–11 months9.81010.410.1229810.610.211.710.8119612–17 months11.9*10.111.3*10.9246911.3*9.210.910.4115518–23 months8.37.98.38.118429.6109.79.8108424–35 months19.119.319.219.2436222.5*21.420.921.6239236–47 months21.3*21.2*19.120.4461916.920.320.9*19.4214548–59 months20.120.320.620.4462317.520.0*16.2181992Wealth indexPoorest26.63137.0*32.5737419.621.524.3*21.72405Poorer18.418.6*17.118407220.421.8*20.8212329Middle18.2*16.61415.935982020.1*1919.72184Richer17.5*16.2*12.415339521.0*19.316.919.12114Richest19.217.619.5*18.7422918.917.41918.42037Access to Improved Drinking Water58.152.162.2*57.1129453672.0*73.0*60.46688Access to Improved Toilet Facility9.915.9*16.9*15.2345296.9*73.370.780.38884Took Iron g PregnancyTablets Durin11.221.616.9*28.7650730.974.879.9*61.96854Delivery yin a Health Facilit9.212.835.3*20.9473229.374.392.0*657198Cornbhild is First16.918.519.6*18.6422016.724.827.9*23.12555Type of Birth (Single vs. Twin)Single98.6*97.797.797.822,18097.797.697.897.710,814Twins1.42.3*2.3*2.24882.32.42.22.3255Child was Ever Breastfed66.668.167.567.61531865.872.877.3*71.97959Number of ANC attendanceNone70.1*5836.151.711,7165.0*1.71.12.62871 visit3.64.65.1*4.6105113.1*4.63.77.17902–3 visits11.218.425.1*19.7445668.8*59.753.460.767194 or more visits15.21933.7*24544513.13441.9*29.63273Mothers’ age: 3-year groups15–2423.723.523.623.6534217.818.118.518.1200725–3450.252.352.45211,7795253.956.9*54.2600235–4926.1*24.22424.5554730.3*27.924.527.63060Size of child at birthSmaller than average30.6*28.530.5*29.6672038.445.9*38.541.14549Average41.639.942.4*41.2933750.4*38.846.845.24998Above average27.831.6*27.129.2661111.215.3*14.7*13.81522Maternal Body Mass Index (BMI)Underweight12.913.514.5*13.831184.5*2.82.13.2349Normal Weight66.9*75.37.3973.21659762.663.961.662.76943 Overweight/Obese20.2*11.211.713.0295333.033.336.3*34.13777Number of household members<423.324.627.1*25.4574827.936.640.2*34.938625–970.1*67.866.767.815,36667.5*59.957.461.7682510+6.67.6*6.26.915544.5*3.42.33.5382Number of Children Under Five0–132.132.434.9*33.3755427.636.944.9*36.44024249.7*47.9*4547.110,67948.6*49.9*43.947.652693 or more18.219.720.0*19.6443523.8*13.211.2161776Note: Values in the same row with * are significantly different at p < 0.05 in the two-sided test of equality for column proportions Socio-Demographic and health characteristics of mothers and children under five in Ethiopia and rwanda: DHS data from 2005, 2010, and 2015 Note: Values in the same row with * are significantly different at p < 0.05 in the two-sided test of equality for column proportions The study population in both nations was predominantly rural, accounting for 83.5% of participants in Ethiopia and 81.8% in Rwanda. However, stark contrasts emerged in key socioeconomic indicators. A large majority of mothers in Ethiopia (69.0%) had no formal education, whereas this figure was far lower in Rwanda (19.8%), where nearly 70% had completed primary education. This disparity was also evident in sanitation, with 80.3% of Rwandan households having access to an improved toilet facility, compared to only 15.2% in Ethiopia. Furthermore, over a third of households in Ethiopia (32.5%) fell into the poorest wealth quintile. The most significant differences between the two countries were observed in the utilization of maternal and child health services. The gap in Antenatal Care (ANC) coverage was particularly wide. In Ethiopia, the proportion of mothers with no ANC visits, while improving, remained high with an overall rate of 51.7%. In stark contrast, Rwanda maintained exceptionally high coverage, with only 2.6% of mothers reporting no ANC visits. This trend was mirrored in facility-based deliveries. Rwanda saw a dramatic increase in facility deliveries from 29.3% in 2005 to 92.0% in 2015. While Ethiopia also showed progress, its rate only increased from 9.2% to 35.3% over the same period. Similarly, the uptake of iron tablets during pregnancy and Vitamin A supplementation for children was substantially higher in Rwanda. Other notable characteristics provide additional context. The practice of ever breastfeeding was widespread and comparably high in both countries (67.6% in Ethiopia and 71.9% in Rwanda). There were, however, differences in maternal nutritional status, with mothers in Ethiopia more likely to have a normal BMI (73.2%) and mothers in Rwanda having a higher prevalence of being overweight (34.1%). In summary, while both countries achieved significant reductions in child malnutrition, their distinct profiles regarding maternal education and access to health services underscore the importance of controlling for these contextual factors in the subsequent causal analysis. A bivariate analysis was conducted to examine the association between various socio-demographic, health, and environmental factors and child malnutrition, as measured by the Composite Index of Anthropometric Failure (CIAF). As detailed in the Table 3, all examined factors were found to have a highly statistically significant association with CIAF (p ≤ 0.001).Table 3Prevalence of CIAF across socio-demographic and environmental factorsComposite Index of Anthropometric FailureNumber of womenChi-squareSig.Nourished (%)CIAF (%)Composite Index of Anthropometric Failure (CIAF)No50.849.222,66811.760.001*Yes52.847.211,069Type of place of residenceUrban67.532.55747720.180.000*Rural48.151.927,990Highest maternal educational levelNo education465417,838860.790.000*Primary53.846.213,047Secondary71.328.72194Higher83.916.1658Sex of childMale49.150.917,13472.400.000*Female53.846.216,603Child’s age in months (grouped)0–5 months70.829.235601219.540.000*6–11 months66.733.3349412–17 months52.847.2362418–23 months44.555.5292624–35 months42.557.5675436–47 months45.954.1676448–59 months49.950.16615Wealth indexPoorest44.155.997791038.410.000*Poorer45.954.16401Middle49.550.55782Richer53.346.75509Richest68.531.56266Survey Year (DHS Wave)200543577775418.900.000*201050.649.413,689201557.742.312,273Access to improved drinking water54.945.119,633Access to improved toilet facility57.742.312,336Iron supplementation during pregnancy55.844.213,361Delivery in a health facility61.338.711,930Firstborn56.643.46775Birth Type (Single vs. Twin)Single51.748.332,99461.920.000*Twins37.162.9743Child was Ever Breastfed53.346.723,277Number of Antenatal Care (ANC) VisitsNone44.555.512,003575.180.000*1 visit46.253.818412–3 visits52.147.911,1754 or more visits61.138.98718Maternal Age (grouped)15–2452.947.1734932.060.000*25–34524817,78135–4948.851.28607Size of child at birthSmaller than average56.743.311,269348.720.000*Average51.948.114,335Above average43.256.88133Maternal Body Mass Index (BMI)Underweight22.777.334671625.390.000*Normal Weight57.642.423,540Overweight/Obese44.755.36730Household Size (grouped)<453.846.2961034.580.000*5–950.349.722,19110+52.347.71936Number of Children Under Five in the household0–1544611,57871.610.000*2495115,9483 or more52.747.36211 Table 3Prevalence of CIAF across socio-demographic and environmental factorsComposite Index of Anthropometric FailureNumber of womenChi-squareSig.Nourished (%)CIAF (%)Composite Index of Anthropometric Failure (CIAF)No50.849.222,66811.760.001*Yes52.847.211,069Type of place of residenceUrban67.532.55747720.180.000*Rural48.151.927,990Highest maternal educational levelNo education465417,838860.790.000*Primary53.846.213,047Secondary71.328.72194Higher83.916.1658Sex of childMale49.150.917,13472.400.000*Female53.846.216,603Child’s age in months (grouped)0–5 months70.829.235601219.540.000*6–11 months66.733.3349412–17 months52.847.2362418–23 months44.555.5292624–35 months42.557.5675436–47 months45.954.1676448–59 months49.950.16615Wealth indexPoorest44.155.997791038.410.000*Poorer45.954.16401Middle49.550.55782Richer53.346.75509Richest68.531.56266Survey Year (DHS Wave)200543577775418.900.000*201050.649.413,689201557.742.312,273Access to improved drinking water54.945.119,633Access to improved toilet facility57.742.312,336Iron supplementation during pregnancy55.844.213,361Delivery in a health facility61.338.711,930Firstborn56.643.46775Birth Type (Single vs. Twin)Single51.748.332,99461.920.000*Twins37.162.9743Child was Ever Breastfed53.346.723,277Number of Antenatal Care (ANC) VisitsNone44.555.512,003575.180.000*1 visit46.253.818412–3 visits52.147.911,1754 or more visits61.138.98718Maternal Age (grouped)15–2452.947.1734932.060.000*25–34524817,78135–4948.851.28607Size of child at birthSmaller than average56.743.311,269348.720.000*Average51.948.114,335Above average43.256.88133Maternal Body Mass Index (BMI)Underweight22.777.334671625.390.000*Normal Weight57.642.423,540Overweight/Obese44.755.36730Household Size (grouped)<453.846.2961034.580.000*5–950.349.722,19110+52.347.71936Number of Children Under Five in the household0–1544611,57871.610.000*2495115,9483 or more52.747.36211 Prevalence of CIAF across socio-demographic and environmental factors The data presented in Table 3 shows strong geographic and socioeconomic disparities. Children in rural areas had a significantly higher prevalence of malnutrition (51.9%) compared to their urban counterparts (32.5%) (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}= 720.18, p < 0.001). A steep educational gradient was observed; malnutrition rates were highest among children of mothers with no formal education (54.0%) and progressively decreased to just 16.1% for those with higher education (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}= 860.79, p < 0.001). Likewise, household wealth was a powerful predictor, with CIAF affecting 55.9% of children in the poorest households compared to 31.5% in the richest (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}= 1038.41, p < 0.001). Child-level characteristics also revealed significant patterns. Male children were slightly more likely to experience malnutrition (50.9%) than female children (46.2%) (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}=72.40, p < 0.001). The risk of malnutrition varied dramatically with age, being lowest for infants aged 0–5 months (29.2%) and peaking among toddlers aged 24–35 months (57.5%) (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}=1219.54, p < 0.001). Children from multiple births were far more vulnerable, with a CIAF rate of 62.9% for twins compared to 48.3% for single births (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}=61.92, p < 0.001). Access to healthcare services, particularly antenatal care (ANC), was strongly associated with a reduced risk of child malnutrition. Children of mothers with no ANC had the highest rates of malnutrition (55.5%), which steadily decreased to 38.9% for those who completed four or more visits (χ^2 = 575.18, p < 0.001). Furthermore, other health and environmental factors, including access to improved water, sanitation, iron supplementation, and breastfeeding, were also significantly associated with lower malnutrition rates. Maternal nutritional status, measured by BMI, showed a complex relationship with child outcomes. The risk was extremely high for children of mothers with moderate thinness (92.0%) and also elevated for mothers classified as severely thin (62.2%) or obese (64.5%), while being lowest for those with a normal BMI (42.4%) (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}=2040.00, p < 0.001). A slight age trend was also observed for mothers, with the prevalence of CIAF increasing from 47.1% in the youngest maternal age group to 51.2% in the oldest (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\chi\:}^{2}$$\end{document}=32.06, p < 0.001). These associations informed the selection of features for further multivariate and machine learning analysis, ensuring the inclusion of statistically and contextually relevant variables in estimating causal effects. Estimation of treatment effects This section presents the findings from the Causal Forest model, which estimated the association between different levels of Antenatal Care (ANC) attendance and the risk of the Composite Index of Anthropometric Failure (CIAF). The analysis compares outcomes for children whose mothers had 1 visit, 2–3 visits, and 4 or more visits against the reference group of None. A negative treatment effect indicates that ANC attendance is associated with a reduced risk of child malnutrition. Overall effect of ANC on child malnutrition: average treatment effect The analysis of the Average Treatment Effect (ATE) reveals a powerful and statistically significant dose-response relationship, showing that increased engagement with ANC leads to a substantially greater reduction in child malnutrition. The analysis of the Average Treatment Effect (ATE), presented in Table 4, reveals a clear and statistically significant dose-response relationship, where the causal impact of Antenatal Care (ANC) on reducing child malnutrition grows substantially with increased maternal engagement. The benefit begins with a 9.5% point reduction in malnutrition risk for a single ANC visit (ATE = −0.095, 95% CI [−0.103, −0.087], p < 0.001). This impact becomes more pronounced for mothers attending 2–3 visits, which corresponds to a 12.1% point reduction (ATE = −0.1209, 95% CI [−0.129, −0.113], p < 0.001). The effect culminates with the most comprehensive level of care, as four or more visits are associated with a 14.5% point reduction in the risk of malnutrition (ATE = −0.145, 95% CI [−0.153, −0.137], p < 0.001). This clear dose-response pattern is a critical finding, suggesting that policy efforts should focus not only on initial contact but also on retaining mothers in the healthcare system to ensure they complete the full, most effective course of antenatal care.Table 4Overall effect of ANC on child malnutrition: average treatment effect (ATE)1 ANC Visit vs. No ANC2–3 ANC Visits vs. No ANC4 + ANC Visits vs. No ANCEstimate95% CIp-valueEstimate95% CIp-valueEstimate95% CIp-valueAverage Treatment Effect (ATE)−0.095[−0.103, −0.087]< 0.001−0.1209[−0.129, −0.113]< 0.001−0.145[−0.153, −0.137]< 0.001 Table 4Overall effect of ANC on child malnutrition: average treatment effect (ATE)1 ANC Visit vs. No ANC2–3 ANC Visits vs. No ANC4 + ANC Visits vs. No ANCEstimate95% CIp-valueEstimate95% CIp-valueEstimate95% CIp-valueAverage Treatment Effect (ATE)−0.095[−0.103, −0.087]< 0.001−0.1209[−0.129, −0.113]< 0.001−0.145[−0.153, −0.137]< 0.001 Overall effect of ANC on child malnutrition: average treatment effect (ATE) Subgroup-specific effects: Conditional Average Treatment Effect (CATE) To explore how ANC effectiveness varies across different population groups, we estimate the conditional average treatment effect (CATE) across key sociodemographic and contextual factors. These factors include residence (urban vs. rural), maternal education, wealth status, country-specific differences, and temporal trends. Additionally, we assess the influence of complementary health and environmental factors such as breastfeeding, access to improved water, sanitation, and vitamin A supplementation. The results presented in Table 5 provide a detailed and insightful analysis of how antenatal care (ANC) influences child malnutrition across various population groups. By employing Conditional Average Treatment Effect (CATE) analysis, the study moves beyond average effects to uncover how the impact of ANC varies based on factors such as residence (urban vs. rural), education level, wealth status, country, and other health and environmental conditions. These findings are particularly valuable for policymakers and healthcare providers, as they highlight disparities in ANC effectiveness and identify population groups that may require more targeted interventions. The CATE not only quantifies the extent to which ANC reduces malnutrition but also sheds light on the reasons behind these variations, offering critical insights for designing equitable and effective health strategies.Table 5Subgroup-Specific Effects: Conditional Average Treatment Effect (CATE)Subgroup CategoryCharacteristicCATE for 1 ANC Visit vs. No ANCCATE for 2–3 ANC Visits vs. No ANCCATE for 4 + ANC Visits vs. No ANCEstimate95% Confidence Intervalp-valueEstimate95% Confidence Intervalp-valueEstimate95% Confidence Intervalp-valueCountry (Ref: Ethiopia)Rwanda−0.0395[−0.050, −0.029]< 0.001−0.0573[−0.068, −0.047]< 0.001−0.0757[−0.086, −0.065]< 0.001Type of place of residence (Ref: Urban)Rural−0.1091[−0.118, −0.100]< 0.001−0.1365[−0.146, −0.127]< 0.001−0.1583[−0.167, −0.149]< 0.001Highest maternal educational level (Ref: No Education)Primary−0.0639[−0.074, −0.054]< 0.001−0.0805[−0.090, −0.071]< 0.001−0.0998[−0.110, −0.090]< 0.001Secondary−0.0462[−0.063, −0.029]< 0.001−0.0714[−0.088, −0.055]< 0.001−0.1074[−0.124, −0.091]< 0.001Higher−0.0408[−0.075, −0.006]0.02−0.0347[−0.070, 0.000]0.052−0.0669[−0.102, −0.032]< 0.001Sex of Child (Ref: Male)Female−0.0966[−0.107, −0.086]< 0.001−0.1237[−0.135, −0.113]< 0.001−0.1432[−0.154, −0.132]< 0.001Wealth Index (Ref: Poorest)Poorer−0.1596[−0.180, −0.139]< 0.001−0.1731[−0.194, −0.153]< 0.001−0.2021[−0.223, −0.181]< 0.001Middle−0.034[−0.052, −0.016]< 0.001−0.0725[−0.091, −0.054]< 0.001−0.0892[−0.108, −0.071]< 0.001Richer−0.0579[−0.075, −0.040]< 0.001−0.0756[−0.093, −0.058]< 0.001−0.0911[−0.109, −0.074]< 0.001Richest−0.0889[−0.100, −0.078]< 0.001−0.1025[−0.114, −0.091]< 0.001−0.1415[−0.152, −0.131]< 0.001Survey Year (DHS Wave) (Ref: 2005)2010−0.1349[−0.147, −0.122]< 0.001−0.1673[−0.180, −0.155]< 0.001−0.1923[−0.205, −0.180]< 0.0012015−0.0323[−0.045, −0.020]< 0.001−0.0419[−0.055, −0.029]< 0.001−0.068[−0.081, −0.055]< 0.001Maternal age (grouped) (Ref: 15–24)25–34−0.0851[−0.096, −0.074]< 0.001−0.1136[−0.125, −0.102]< 0.001−0.1362[−0.147, −0.125]< 0.00135–49−0.1439[−0.160, −0.128]< 0.001−0.1655[−0.181, −0.150]< 0.001−0.193[−0.209, −0.177]< 0.001Maternal Body Mass Index (BMI) (Ref: Underweight)Normal−0.0902[−0.100, −0.080]< 0.001−0.1173[−0.127, −0.107]< 0.001−0.1368[−0.147, −0.127]< 0.001Overweight−0.1443[−0.157, −0.132]< 0.001−0.1774[−0.190, −0.165]< 0.001−0.2136[−0.226, −0.201]< 0.001Child’s age in months (grouped) (Ref: 0–5 mos)6–11 mos0.0344[0.016, 0.053]< 0.001−0.003[−0.021, 0.015]0.749−0.0395[−0.058, −0.021]< 0.00112–17 mos0.049[0.029, 0.069]< 0.0010.0125[−0.008, 0.033]0.224−0.0254[−0.046, −0.005]0.01518–23 mos0.0244[0.001, 0.047]0.0390.0064[−0.017, 0.030]0.589−0.0376[−0.061, −0.014]0.00224–35 mos−0.0775[−0.094, −0.061]< 0.001−0.0887[−0.105, −0.072]< 0.001−0.098[−0.115, −0.081]< 0.00136–47 mos−0.1236[−0.143, −0.104]< 0.001−0.1468[−0.167, −0.127]< 0.001−0.1873[−0.207, −0.167]< 0.00148–59 mos−0.3435[−0.363, −0.324]< 0.001−0.3591[−0.379, −0.339]< 0.001−0.3545[−0.374, −0.335]< 0.001Size of child at birth (Ref: Above average)Average−0.054[−0.066, −0.042]< 0.001−0.0825[−0.095, −0.070]< 0.001−0.109[−0.121, −0.097]< 0.001Smaller−0.1096[−0.123, −0.096]< 0.001−0.1355[−0.149, −0.122]< 0.001−0.1539[−0.168, −0.140]< 0.001Firstborn (Ref: Not Firstborn)Yes−0.0531[−0.066, −0.041]< 0.001−0.0669[−0.079, −0.054]< 0.001−0.0847[−0.097, −0.072]< 0.001Household size (grouped)(Ref: <4 members)5–9 members−0.098[−0.108, −0.087]< 0.001−0.1278[−0.138, −0.117]< 0.001−0.1532[−0.164, −0.143]< 0.00110 + members−0.1736[−0.200, −0.148]< 0.001−0.1854[−0.212, −0.159]< 0.001−0.2014[−0.228, −0.175]< 0.001Number of Children Under Five in the household(Ref: 1)2 children−0.1036[−0.116, −0.091]< 0.001−0.1304[−0.143, −0.118]< 0.001−0.1521[−0.164, −0.140]< 0.0013 + children−0.1455[−0.163, −0.128]< 0.001−0.174[−0.191, −0.157]< 0.001−0.1949[−0.212, −0.177]< 0.001Received Vitamin A supplement (Ref: No)Yes−0.0162[−0.029, −0.004]0.012−0.0227[−0.035, −0.010]< 0.001−0.0533[−0.066, −0.041]< 0.001Access to improved drinking water (Ref: No)Yes−0.0823[−0.091, −0.073]< 0.001−0.0937[−0.103, −0.084]< 0.001−0.1206[−0.130, −0.111]< 0.001Access to improved toilet facility (Ref: No)Yes−0.0475[−0.057, −0.038]< 0.001−0.0613[−0.071, −0.052]< 0.001−0.0837[−0.093, −0.074]< 0.001Iron supplementation during pregnancy (Ref: No)Yes−0.0047[−0.015, 0.005]0.3550[−0.010, 0.010]0.995−0.0221[−0.032, −0.012]< 0.001Delivery in a health facility (Ref: No)Yes0.0132[0.004, 0.023]0.0060.004[−0.006, 0.014]0.407−0.0251[−0.035, −0.016]< 0.001Breastfeeding (Ref: No)Yes−0.0216[−0.030, −0.013]< 0.001−0.0516[−0.060, −0.043]< 0.001−0.0825[−0.091, −0.074]< 0.001Birth Type (Ref: Single)Twin−0.0857[−0.138, −0.034]0.001−0.1056[−0.158, −0.053]< 0.001−0.1207[−0.174, −0.068]< 0.001 Table 5Subgroup-Specific Effects: Conditional Average Treatment Effect (CATE)Subgroup CategoryCharacteristicCATE for 1 ANC Visit vs. No ANCCATE for 2–3 ANC Visits vs. No ANCCATE for 4 + ANC Visits vs. No ANCEstimate95% Confidence Intervalp-valueEstimate95% Confidence Intervalp-valueEstimate95% Confidence Intervalp-valueCountry (Ref: Ethiopia)Rwanda−0.0395[−0.050, −0.029]< 0.001−0.0573[−0.068, −0.047]< 0.001−0.0757[−0.086, −0.065]< 0.001Type of place of residence (Ref: Urban)Rural−0.1091[−0.118, −0.100]< 0.001−0.1365[−0.146, −0.127]< 0.001−0.1583[−0.167, −0.149]< 0.001Highest maternal educational level (Ref: No Education)Primary−0.0639[−0.074, −0.054]< 0.001−0.0805[−0.090, −0.071]< 0.001−0.0998[−0.110, −0.090]< 0.001Secondary−0.0462[−0.063, −0.029]< 0.001−0.0714[−0.088, −0.055]< 0.001−0.1074[−0.124, −0.091]< 0.001Higher−0.0408[−0.075, −0.006]0.02−0.0347[−0.070, 0.000]0.052−0.0669[−0.102, −0.032]< 0.001Sex of Child (Ref: Male)Female−0.0966[−0.107, −0.086]< 0.001−0.1237[−0.135, −0.113]< 0.001−0.1432[−0.154, −0.132]< 0.001Wealth Index (Ref: Poorest)Poorer−0.1596[−0.180, −0.139]< 0.001−0.1731[−0.194, −0.153]< 0.001−0.2021[−0.223, −0.181]< 0.001Middle−0.034[−0.052, −0.016]< 0.001−0.0725[−0.091, −0.054]< 0.001−0.0892[−0.108, −0.071]< 0.001Richer−0.0579[−0.075, −0.040]< 0.001−0.0756[−0.093, −0.058]< 0.001−0.0911[−0.109, −0.074]< 0.001Richest−0.0889[−0.100, −0.078]< 0.001−0.1025[−0.114, −0.091]< 0.001−0.1415[−0.152, −0.131]< 0.001Survey Year (DHS Wave) (Ref: 2005)2010−0.1349[−0.147, −0.122]< 0.001−0.1673[−0.180, −0.155]< 0.001−0.1923[−0.205, −0.180]< 0.0012015−0.0323[−0.045, −0.020]< 0.001−0.0419[−0.055, −0.029]< 0.001−0.068[−0.081, −0.055]< 0.001Maternal age (grouped) (Ref: 15–24)25–34−0.0851[−0.096, −0.074]< 0.001−0.1136[−0.125, −0.102]< 0.001−0.1362[−0.147, −0.125]< 0.00135–49−0.1439[−0.160, −0.128]< 0.001−0.1655[−0.181, −0.150]< 0.001−0.193[−0.209, −0.177]< 0.001Maternal Body Mass Index (BMI) (Ref: Underweight)Normal−0.0902[−0.100, −0.080]< 0.001−0.1173[−0.127, −0.107]< 0.001−0.1368[−0.147, −0.127]< 0.001Overweight−0.1443[−0.157, −0.132]< 0.001−0.1774[−0.190, −0.165]< 0.001−0.2136[−0.226, −0.201]< 0.001Child’s age in months (grouped) (Ref: 0–5 mos)6–11 mos0.0344[0.016, 0.053]< 0.001−0.003[−0.021, 0.015]0.749−0.0395[−0.058, −0.021]< 0.00112–17 mos0.049[0.029, 0.069]< 0.0010.0125[−0.008, 0.033]0.224−0.0254[−0.046, −0.005]0.01518–23 mos0.0244[0.001, 0.047]0.0390.0064[−0.017, 0.030]0.589−0.0376[−0.061, −0.014]0.00224–35 mos−0.0775[−0.094, −0.061]< 0.001−0.0887[−0.105, −0.072]< 0.001−0.098[−0.115, −0.081]< 0.00136–47 mos−0.1236[−0.143, −0.104]< 0.001−0.1468[−0.167, −0.127]< 0.001−0.1873[−0.207, −0.167]< 0.00148–59 mos−0.3435[−0.363, −0.324]< 0.001−0.3591[−0.379, −0.339]< 0.001−0.3545[−0.374, −0.335]< 0.001Size of child at birth (Ref: Above average)Average−0.054[−0.066, −0.042]< 0.001−0.0825[−0.095, −0.070]< 0.001−0.109[−0.121, −0.097]< 0.001Smaller−0.1096[−0.123, −0.096]< 0.001−0.1355[−0.149, −0.122]< 0.001−0.1539[−0.168, −0.140]< 0.001Firstborn (Ref: Not Firstborn)Yes−0.0531[−0.066, −0.041]< 0.001−0.0669[−0.079, −0.054]< 0.001−0.0847[−0.097, −0.072]< 0.001Household size (grouped)(Ref: <4 members)5–9 members−0.098[−0.108, −0.087]< 0.001−0.1278[−0.138, −0.117]< 0.001−0.1532[−0.164, −0.143]< 0.00110 + members−0.1736[−0.200, −0.148]< 0.001−0.1854[−0.212, −0.159]< 0.001−0.2014[−0.228, −0.175]< 0.001Number of Children Under Five in the household(Ref: 1)2 children−0.1036[−0.116, −0.091]< 0.001−0.1304[−0.143, −0.118]< 0.001−0.1521[−0.164, −0.140]< 0.0013 + children−0.1455[−0.163, −0.128]< 0.001−0.174[−0.191, −0.157]< 0.001−0.1949[−0.212, −0.177]< 0.001Received Vitamin A supplement (Ref: No)Yes−0.0162[−0.029, −0.004]0.012−0.0227[−0.035, −0.010]< 0.001−0.0533[−0.066, −0.041]< 0.001Access to improved drinking water (Ref: No)Yes−0.0823[−0.091, −0.073]< 0.001−0.0937[−0.103, −0.084]< 0.001−0.1206[−0.130, −0.111]< 0.001Access to improved toilet facility (Ref: No)Yes−0.0475[−0.057, −0.038]< 0.001−0.0613[−0.071, −0.052]< 0.001−0.0837[−0.093, −0.074]< 0.001Iron supplementation during pregnancy (Ref: No)Yes−0.0047[−0.015, 0.005]0.3550[−0.010, 0.010]0.995−0.0221[−0.032, −0.012]< 0.001Delivery in a health facility (Ref: No)Yes0.0132[0.004, 0.023]0.0060.004[−0.006, 0.014]0.407−0.0251[−0.035, −0.016]< 0.001Breastfeeding (Ref: No)Yes−0.0216[−0.030, −0.013]< 0.001−0.0516[−0.060, −0.043]< 0.001−0.0825[−0.091, −0.074]< 0.001Birth Type (Ref: Single)Twin−0.0857[−0.138, −0.034]0.001−0.1056[−0.158, −0.053]< 0.001−0.1207[−0.174, −0.068]< 0.001 Subgroup-Specific Effects: Conditional Average Treatment Effect (CATE) Maternal age (grouped) (Ref: 15–24) Maternal Body Mass Index (BMI) (Ref: Underweight) Child’s age in months (grouped) (Ref: 0–5 mos) Size of child at birth (Ref: Above average) Household size (grouped) (Ref: <4 members) Number of Children Under Five in the household (Ref: 1) The results presented in Table 5 provide a detailed and insightful analysis of how antenatal care (ANC) influences child malnutrition across various population groups. By employing Conditional Average Treatment Effect (CATE) analysis, the study moves beyond average effects to uncover how the impact of ANC varies based on factors such as residence (urban vs. rural), education level, wealth status, country, and other health and environmental conditions. These findings are particularly valuable for policymakers and healthcare providers, as they highlight disparities in ANC effectiveness and identify population groups that may require more targeted interventions. The CATE not only quantifies the extent to which ANC reduces malnutrition but also sheds light on the reasons behind these variations, offering critical insights for designing equitable and effective health strategies. Model validation and robustness checks Evaluating the performance of the Causal Forest model is crucial to ensure the reliability and robustness of the findings. This section presents two key assessments of the model’s performance: the stability of the estimates through cross-validation and the distribution of the individual treatment effects. These metrics provide insights into the model’s ability to consistently and accurately estimate causal effects across the dataset. Cross-Validated results Cross-validation is a widely used technique for evaluating the generalizability of a model’s findings. In this analysis, a 5-fold cross-validation procedure was implemented to assess the reliability of the Causal Forest model. The dataset was divided into three subsets, where the model was iteratively trained on two subsets and tested on the remaining one. The Average Treatment Effect (ATE) was calculated for each fold, and the cross-validated ATE was obtained by averaging these values. The cross-validated ATE for attending 4 or more ANC visits was estimated at −0.0736, indicating a 7.4% point reduction in malnutrition risk. This result is more conservative than the overall ATE of −0.1450 calculated on the full dataset, but it is a critically important finding. It confirms that even under a more rigorous validation process, the direction and high statistical significance of the effect (p < 0.001) are maintained. The consistency of this strong, negative effect across the cross-validation folds highlights the model’s robustness and generalizability. While the magnitude of the point estimate is lower, the unwavering finding of a significant reduction in malnutrition risk suggests the model is not overfitting to the training data. This reliability ensures the estimated causal effects are credible and suitable for informing programmatic decision-making. Distribution of treatment effects The distribution of individual treatment effects was analyzed to evaluate how the estimated causal effects vary across individuals in the dataset. Histograms were generated for each treatment level (1, 2–3, and 4 + ANC visits versus no ANC) to visualize the spread of effects, providing a clear picture of their variability and consistency, as shown in Fig. 2. This analysis is crucial for understanding the model’s ability to generate reliable and stable estimates of the causal association between Antenatal Care (ANC) and child malnutrition.Fig. 2Distribution of Treatment Effects Distribution of Treatment Effects The distributions of the individual treatment effects for each level of ANC revealed approximately symmetric, sharply peaked, bell-shaped distributions, each centered around its respective Average Treatment Effect (ATE). The visual placement of these peaks confirms the overall dose-response relationship: the distribution for 1 ANC visit is centered almost perfectly on the “No Effect” line (zero); the peak for 2–3 ANC visits shifts visibly to the left into negative territory; and the peak for 4 or more visits is centered even further to the left, representing the strongest protective effect. The clear and increasingly negative central tendency of these distributions indicates that the model consistently captures the average causal association and that its estimates for the ATE are stable and reliable. While the central peaks of the distributions are notably narrow, indicating a high degree of consistency in the estimated effect for the majority of the population, the presence of tails extending in both directions is still direct visual evidence of treatment effect heterogeneity. This variability shows that while the average association is a reduced risk of malnutrition (for 2 + visits), some individuals experience a much stronger risk reduction while others experience little to no benefit. This finding is crucial as it validates the importance of the CATE analysis, which is designed to explore the specific characteristics of individuals who fall into the different parts of these distributions. The combined results of these metrics—the cross-validated ATE and the analysis of the treatment effect distributions—confirm the suitability of the Causal Forest model for this study. The model’s demonstrated ability to produce consistent and stable average estimates, while also capturing significant individual-level heterogeneity, ensures its reliability for policy and programmatic decision-making. These findings provide confidence that the model accurately captures the causal relationships necessary to draw actionable insights for targeted interventions. Interpretation of the results The results from the Causal Forest model, confirmed as robust and reliable, provide a powerful understanding of the causal impact of Antenatal Care (ANC) on child malnutrition. This interpretation will focus on the effects for 4 + ANC visits, which was identified as the most effective intervention tier with an Average Treatment Effect (ATE) corresponding to a 14.5% point reduction in the risk of child malnutrition (ATE = −0.1450, 95% CI [−0.153, −0.137], p < 0.001). The Conditional Average Treatment Effect (CATE) analysis reveals where and for whom this impact is significantly amplified or diminished. All findings discussed are statistically significant at the p < 0.001 level unless otherwise noted. Contextual disparities in ANC effectiveness The analysis reveals that the causal impact of ANC is profoundly modified by a mother’s environment and socioeconomic status, highlighting how these contextual factors dictate the effectiveness of this health intervention. Residence and wealth: a lifeline for the rural poor The impact of ANC is substantially greater in rural areas, which saw a 15.8% point reduction in the CIAF score (CATE = −0.1583, 95% CI [−0.167, −0.149]). This finding is magnified when considering economic status. The analysis reveals a profound effect for the poorer wealth quintile, where 4 + ANC visits are associated with a massive 20.2% point reduction in malnutrition risk (CATE = −0.2021, 95% CI [−0.223, −0.181]). This effect, nearly 40% stronger than the population average, indicates that for households facing resource constraints and with fewer alternative sources of health information, ANC is not just a routine service but a transformative lifeline for child health. Maternal education: enhancing the ability to act on advice The effectiveness of ANC generally increases with maternal education. The benefit grows from the reference group (no education) to mothers with primary education (−10.0% points, 95% CI [−0.110, −0.090]) and secondary education (−10.7% points, 95% CI [−0.124, −0.091]). This suggests that education enhances a mother’s “absorptive capacity”—her ability to understand, retain, and effectively act upon the complex health and nutritional advice provided during ANC visits. The greater impact among educated mothers may also reflect differences in baseline nutritional status and health-seeking behavior, as well as an increased ability to act on medical advice. Country and temporal trends: health system scale-up, marginal returns, and quality dilution The analysis reveals intriguing variations in ANC’s impact both between countries and over time, potentially reflecting the challenges of scaling up public health interventions. Antenatal care was associated with a larger reduction in child malnutrition in Ethiopia (the reference country) compared to Rwanda, where it corresponded to a 7.6% point reduction in the CIAF score (CATE = −0.0757, 95% CI [−0.086, −0.065]). One explanation for this difference may be related to the scale of program implementation. With a much larger proportion of women in Rwanda reporting ANC attendance compared to their Ethiopian counterparts, Rwanda’s health system may face challenges related to the dilution of service quality as it serves a near-universal population, leading to a smaller marginal impact for each individual. This same principle may explain the temporal trend observed in the data. The effect was most profound in the 2010 survey wave, which saw an exceptional 19.2% point reduction in the CIAF score (CATE = −0.1923, 95% CI [−0.205, −0.180]). The subsequent decline in effect size by 2015 could reflect the inherent difficulties of maintaining service quality during a massive scale-up. The 2010 period may represent a “peak effectiveness” phase when health systems had matured significantly during the Millennium Development Goal era but had not yet faced the full strain of reaching the most remote, “last-mile” populations with complex needs. This suggests that as ANC coverage expands, ensuring consistent quality of care becomes the paramount challenge for achieving the best possible health outcomes. Impact of ANC dosage and synergies with other factors The Importance of consistent care: a dose-response effect The analysis of the overall impact of ANC shows a clear dose-response effect. A single visit provides a significant benefit (−9.5% points, 95% CI [−0.103, −0.087]), but the full, substantial benefit is only realized after consistent engagement, with four or more visits yielding the largest impact (−14.5% points, 95% CI [−0.153, −0.137]). This implies that policy efforts must focus not just on initial contact but on implementing strategies to ensure mothers complete the full recommended schedule to achieve the greatest possible reduction in child malnutrition. WASH: creating an enabling environment for nutrition The effectiveness of ANC’s nutritional guidance is fundamentally dependent on a clean and safe environment. The impact of 4 + ANC visits is substantially amplified for households with access to an improved water source, causing a 12.1% point reduction in risk (CATE = −0.1206, 95% CI [−0.130, −0.111]). Access to an improved toilet facility also strengthened the effect (−8.4% points, 95% CI [−0.093, −0.074]). This is because clean water and sanitation are prerequisites for good health; they prevent the diarrheal diseases and infections that impair nutrient absorption and can completely undermine the benefits of a better diet, highlighting the need for multi-sectoral collaboration. Maternal experience and child vulnerability: a detailed picture The results paint a detailed picture of how maternal experience and child vulnerability interact with ANC’s effectiveness. Guidance for the Inexperienced: Attending 4 + ANC visits provides a strong, significant benefit for firstborn children, reducing malnutrition risk by 8.5% points (CATE = −0.0847, 95% CI [−0.097, −0.072]). This shows that ANC serves as a crucial source of formal guidance and support, helping inexperienced mothers navigate the challenges of their first child.The Synergy with Experience: While beneficial for first-time mothers, the causal impact of ANC is supercharged when it combines with experience. The effect is most profound for older mothers aged 35–49, who saw a massive 19.3% point reduction in risk (CATE = −0.1930, 95% CI [−0.209, −0.177]). This suggests that ANC is most effective when it synergizes with a mother’s existing practical knowledge.Protecting the Vulnerable at Birth: ANC demonstrates its critical role in mitigating early-life disadvantage. The effect is significantly stronger for children from single births (−12.1% points, 95% CI [−0.174, −0.068]) compared to twins. Most importantly, the intervention is exceptionally protective for children born smaller than average, for whom it is associated with a remarkable 15.4% point reduction in malnutrition (CATE = −0.1539, 95% CI [−0.168, −0.140]), showcasing its power as a tool to help the most biologically vulnerable infants catch up. Guidance for the Inexperienced: Attending 4 + ANC visits provides a strong, significant benefit for firstborn children, reducing malnutrition risk by 8.5% points (CATE = −0.0847, 95% CI [−0.097, −0.072]). This shows that ANC serves as a crucial source of formal guidance and support, helping inexperienced mothers navigate the challenges of their first child. Guidance for the Inexperienced: Attending 4 + ANC visits provides a strong, significant benefit for firstborn children, reducing malnutrition risk by 8.5% points (CATE = −0.0847, 95% CI [−0.097, −0.072]). This shows that ANC serves as a crucial source of formal guidance and support, helping inexperienced mothers navigate the challenges of their first child. The Synergy with Experience: While beneficial for first-time mothers, the causal impact of ANC is supercharged when it combines with experience. The effect is most profound for older mothers aged 35–49, who saw a massive 19.3% point reduction in risk (CATE = −0.1930, 95% CI [−0.209, −0.177]). This suggests that ANC is most effective when it synergizes with a mother’s existing practical knowledge. The Synergy with Experience: While beneficial for first-time mothers, the causal impact of ANC is supercharged when it combines with experience. The effect is most profound for older mothers aged 35–49, who saw a massive 19.3% point reduction in risk (CATE = −0.1930, 95% CI [−0.209, −0.177]). This suggests that ANC is most effective when it synergizes with a mother’s existing practical knowledge. Protecting the Vulnerable at Birth: ANC demonstrates its critical role in mitigating early-life disadvantage. The effect is significantly stronger for children from single births (−12.1% points, 95% CI [−0.174, −0.068]) compared to twins. Most importantly, the intervention is exceptionally protective for children born smaller than average, for whom it is associated with a remarkable 15.4% point reduction in malnutrition (CATE = −0.1539, 95% CI [−0.168, −0.140]), showcasing its power as a tool to help the most biologically vulnerable infants catch up. Protecting the Vulnerable at Birth: ANC demonstrates its critical role in mitigating early-life disadvantage. The effect is significantly stronger for children from single births (−12.1% points, 95% CI [−0.174, −0.068]) compared to twins. Most importantly, the intervention is exceptionally protective for children born smaller than average, for whom it is associated with a remarkable 15.4% point reduction in malnutrition (CATE = −0.1539, 95% CI [−0.168, −0.140]), showcasing its power as a tool to help the most biologically vulnerable infants catch up. Discussion This study builds on existing literature by applying a machine learning-based causal inference approach—specifically, the Causal Forest model—to identify heterogeneity in ANC’s effect on child malnutrition, measured via the Composite Index of Anthropometric Failure (CIAF), across multiple Sub-Saharan African countries. The findings not only confirm the overall efficacy of ANC, as established in the literature, but, more importantly, reveal a complex tapestry of effect heterogeneity that challenges conventional assumptions and provides a clear evidence base for more targeted, equitable, and efficient maternal health policies. A central finding of this analysis is the amplified effectiveness of ANC among the most vulnerable populations, specifically those in rural areas and in the poorer wealth quintiles. Our model estimates that the malnutrition reduction for the poorer quintile is a massive 20.2% points (CATE = −0.2021, 95% CI [−0.223, −0.181], p < 0.001). This finding presents a compelling counter-narrative to research that often focuses on the numerous barriers to quality care in such settings [22]. It strongly supports the hypothesis that where alternative health resources are scarce, formal interventions like ANC serve as a critical, high-impact lifeline. While urban mothers may have multiple avenues for health information, consistent with findings that child malnutrition is generally lower in urban areas [23], our results suggest that for a rural mother in poverty, the guidance from ANC is exceptionally valuable. This underscores a key policy directive: maximizing public health returns requires prioritizing investment in reaching the rural poor, potentially through community-based models driven by Community Health Workers (CHWs), which have proven effective in both Ethiopia and Rwanda [11, 12]. The interplay between maternal education and experience revealed an insightful pattern. The finding that ANC’s impact generally increases with education aligns with established literature showing that educated mothers are better able to understand and act upon health advice [15, 16]. However, the synergy between ANC and maternal experience is particularly profound. The causal impact is strongest among older, more experienced mothers (aged 35–49), who saw a 19.3% point reduction in risk (CATE = −0.1930, 95% CI [−0.209, −0.177], p < 0.001). Importantly, ANC also provides a strong benefit for firstborn children (−8.5% points, CATE = −0.0847, 95% CI [−0.097, −0.072], p < 0.001), indicating its role as crucial formal guidance for the inexperienced. However, the significantly larger effect for older mothers suggests that ANC’s impact is supercharged when it synergizes with existing practical knowledge, a critical period for child development [1]. The heterogeneity of ANC’s impact across countries and over time offers a data-driven narrative on the complexities of health system scale-up. Specifically, ANC was associated with a larger reduction in child malnutrition in Ethiopia than in Rwanda (CATE = −0.0757, 95% CI [−0.086, −0.065], p < 0.001). This disparity can be interpreted through the complementary lenses of diminishing marginal returns and service quality dilution. The principle of diminishing marginal returns suggests that as intervention coverage approaches universality, as is the case in Rwanda, the average benefit per individual may decrease because the program has already reached many of the highest-risk individuals. In contrast, Ethiopia’s lower-coverage context means the measured impact is likely greater because the program is still primarily serving a population with a higher initial unmet need. This concept is linked to service quality; the logistical strain of maintaining high-quality care at a near-universal scale can contribute to a smaller average effect size. This same framework helps explain the temporal pattern, where the effect of ANC peaked in the 2010 survey wave with a 19.2% point reduction in risk (CATE = −0.1923, 95% CI [−0.205, −0.180], p < 0.001) before declining. This suggests a “peak effectiveness” phase, followed by the challenges of reaching “last-mile” populations and maintaining quality during a massive expansion, a difficulty noted in Ethiopia’s federal system [9]. This pattern underscores that after achieving coverage targets, the subsequent policy frontier must be a relentless focus on quality assurance and continuous training for health workers [24]. Finally, this study provides strong causal evidence for the necessity of a multi-sectoral approach to nutrition. The finding that ANC’s effectiveness is significantly amplified by access to an improved water source, which strengthens the malnutrition reduction to 12.1% points (CATE = −0.1206, 95% CI [−0.130, −0.111], p < 0.001), is critical. It confirms that the benefits of health and nutrition counseling can be severely undermined by an unsafe environment that fosters infectious disease, a cornerstone of maternal and child health research [25]. This aligns perfectly with nutrition-sensitive frameworks that call for integrating health interventions with improvements in WASH [26]. Therefore, our results provide a strong argument for policymakers to break down silos and co-locate investments to achieve a synergistic and lasting impact on child well-being. Conclusion This study employed a machine learning-based causal inference framework to move beyond simple correlations and rigorously estimate the causal impact of Antenatal Care (ANC) on child malnutrition. The central finding is that while ANC is beneficial on average, its true value is revealed through its profound heterogeneity. The evidence clearly shows that the intervention’s impact is dramatically amplified for the most vulnerable populations, while a “ceiling effect” is observed among more advantaged groups, rendering a “one-size-fits-all” policy approach suboptimal. Despite these robust findings, the study’s reliance on cross-sectional data limits the ability to infer long-term causal impacts. The causal estimates, while rigorously controlled for observed confounders, cannot account for unobserved factors that may change over time. Future research should therefore leverage longitudinal designs to track developmental trajectories and incorporate qualitative methods to uncover the specific mechanisms driving the observed heterogeneity—for instance, to understand why ANC is so much more effective for experienced mothers. The clear policy imperative arising from this research is the need to shift from a universal to a stratified and targeted model of care. This requires practical action, such as expanding rural healthcare infrastructure and addressing economic barriers through pro-poor policies. Furthermore, achieving the full potential of ANC requires moving beyond the health sector. Much of the counseling mothers receive during ANC—on diet, hygiene, and safe food preparation—can only be effective if it is supported by the availability of necessary resources and infrastructure. Our findings provide a strong evidence base for integrating ANC programs with multi-sectoral initiatives, particularly in Water, Sanitation, and Hygiene (WASH) and social protection. Programs such as Rwanda’s Fortified Blended Food (FBF) initiative offer a valuable model for this integrated approach, ensuring mothers have both the knowledge from ANC and the means to act on it. The application of machine learning techniques, particularly the Causal Forest model, represents a significant methodological advancement, allowing for the precise estimation of treatment effects across diverse subpopulations. Ultimately, this research demonstrates that by tailoring ANC programs to local contexts and leveraging machine learning tools for precision targeting, governments can significantly accelerate progress toward eliminating child malnutrition.