Background: Household shocks are hypothesized to impair child linear growth, but conventional regression adjustment is biased when earlier shocks change later wealth, food security, and growth, creating treatment confounder feedback. Standard covariate adjusted models can condition away part of the causal pathway or leave residual confounding.
Methods: We analysed Young Lives cohort data from Ethiopia, India, Peru, and Vietnam across five survey rounds spanning 2002 to 2016, retaining a balanced panel of 7,126 children with complete treatment histories and non missing final height for age z score (HAZ), 88.5% of the 8,049 younger cohort children interviewed at baseline. Cumulative shock exposure over childhood was treated as a longitudinal exposure process. Effects on final HAZ were estimated using cross fitted g methods, including a marginal structural model with inverse probability of treatment weighting, g computation, and longitudinal targeted maximum likelihood estimation, alongside a conventional same learner covariate adjusted benchmark. Attrition was addressed with inverse probability of censoring weighting, and estimator behaviour was first validated in semi synthetic data with a known oracle effect.
Results: In validation data, the marginal structural model and targeted maximum likelihood estimator recovered the true effect while covariate adjustment and g computation were attenuated toward the null. In the real data, each additional shock year lowered final HAZ by 0.107 SD (95% CI -0.121 to -0.093), essentially unchanged after re-weighting the full cohort for censoring, while the conventional adjusted benchmark was close to null (-0.001, 95% CI -0.011 to 0.006). The central finding was substantial cross country heterogeneity: the per shock effect ranged from -0.128 SD in Ethiopia and -0.089 SD in India to -0.063 SD in Peru and -0.029 SD in Vietnam, whose interval included the null. The Vietnam-Ethiopia difference was 0.098 SD (95% CI 0.038 to 0.162, P < 0.001), formally rejecting a common per shock effect. Transportability analysis showed this heterogeneity was not explained by measured covariate composition, indicating structural rather than compositional effect modification. Mediation analysis suggested measured household wealth carried almost none of the effect, and E values of 1.40 and 1.37 indicated only moderate robustness to unmeasured confounding.
Conclusions: A single pooled shock effect is not transportable across settings: the same measured household shock produces significantly different consequences for child growth in different countries, for structural reasons unrelated to observed household composition. Shock vulnerability, and the evidence used to target interventions against it, should be treated as setting specific rather than generalised from one population to another.
