This paper uses Multiple Indicator Cluster Surveys data from the Republic of Congo and São Tomé and Príncipe to study the relationships between child stature, mother's years of education, and indicators of early childhood development. The relationships are contrasted between two empirical approaches: the conventional approach whereby control variables are selected in an ad-hoc manner and the double machine-learning approach that employs data-driven methods to select controls from a much wider set of variables. Overall, the findings based on the preferable double machine-learning approach differ across the two countries depending on the measures of early childhood development and child stature (height-for-age Z-score and stunting) used in the analysis. Double machine-learning estimates for the Republic of Congo suggest that height-for-age Z-score and stunting have a direct causal effect on early childhood development. In contrast, for São Tomé and Príncipe, no relationship is found. Thus, country-specific policy advice based on the relationships observed from data in other countries may be quite risky, if not misleading. Double machine-learning provides a practical and feasible approach to reducing threats to internal validity to derive robust inferences based on observational data for evidence-based policy advice.