Interpersonal Communication and Maternal Behavioral Practice: A Case Study of Explanatory Causal Machine Learning in Nepal

BACKGROUND: There is increasing interest not only in whether a program produced impacts but also the mechanisms through which the impacts were produced. Studies of nutrition interventions have frequently highlighted the role of nutrition education interventions, but few studies have investigated the paths through which those interventions achieve impact. OBJECTIVES: We assessed the total effect of maternal exposure to interpersonal communication with a frontline worker on behavioral practice and the degree to which that effect operated through 2 key pathways: improved awareness and knowledge. We also conducted a comparative case study examining the differences between machine learning (ML) and structural equation model (SEM) approaches. METHODS: We used data from the endline cross-sectional survey of the quasi-experimental Suaahara study of 2040 mothers with a child less than 2 years of age. We estimated the total effect of maternal exposure to interpersonal communication with a frontline worker on her nutrition-related practices and the indirect effects through improvements in awareness and knowledge using ML and conventional SEMs. Models were adjusted for household, child, and maternal factors as well as household location. RESULTS: Maternal exposure to interpersonal communication with a frontline worker translated into practicing more of the ideal, promoted nutrition-related practices. Both methods suggested a similar total effect-ML: 0.58 (95% CI: 0.37, 0.78) and SEM: 0.60 (95% CI: 0.49, 0.72). ML models, however, attributed a much larger proportion of that total effect to indirect effects operating through awareness and knowledge (ML: 0.24; 95% CI: 0.15, 0.33; SEMs: 0.16; 95% CI: 0.09, 0.21). CONCLUSIONS: Interpersonal communication improved nutrition-related practices, in part due to increasing one's awareness and knowledge. The disparities in results between ML and SEMs suggested that the production of effects is a more complex, nonlinear or interactive function of awareness, knowledge, treatment and covariates than what is captured by simple linear models.