Investigating Causal Questions in Human Development using Marginal Structural Models: A Tutorial Introduction to the devMSMs Package in R

Causal inference and questions about timing are central to developmental science. Yet scientists are often limited to observational data, and conceptual, methodological, and computational barriers have prevented uptake of relevant methodological tools from other disciplines that are designed to improve causal inference. This paper provides a practical guide to implementing one powerful causal inference tool, marginal structural models (MSMs), to strengthen our ability to test causal questions about exposures/treatments that unfold over time. We draw on an illustrative example testing the potentially causal roles of dose and timing of exposure to economic strain from infancy through early childhood on children’s behavior problems in early childhood. We first introduce the advantages of MSMs for investigating causal questions over time. Specifically, we review the logic of ‘exposure histories’ and the potential outcomes framework, using sequential randomization as a conceptual tool. We then discuss inverse-probability-of-treatment-weighting (IPTW) for addressing the threat of confounding over time. We then outline how to implement MSMs with observational data, before providing step-by-step guidance for using our novel, open-source R package, devMSMs. In our example, we create longitudinal IPTW balancing weights to attenuate a number of confounder-exposure relations. We then apply these weights to test an outcome model relating economic strain across infancy, toddlerhood, and early childhood to future behavior problems. Finally, we estimate and compare the predicted causal effects of different developmental histories of high economic strain that vary in dose and timing and discuss substantive implications. We close with ongoing considerations.