Weight, Balance, Repeat: Marginal Structural Models for Improving Causal Inference With Time-Varying Treatments

• In Event: 3-238 - New (or Under Appreciated) Statistical Methods for Probing Causal Hypotheses in Developmental Research

Integrative Statement

Substantive Context: Families struggling with economic adversity tend to face a litany of external stressors that can undermine the predictability and organization of children’s everyday experiences—ranging from the quality of their physical environments to their interactions with meaningful adults (Dearing, Berry, & Zalow, 2006). Theory and a growing body of empirical evidence suggests that such experiences may calibrate children’s emerging physiological stress systems toward states of pronounced psychophysiological vigilance (Blair & Raver, 2012; Del Guidice et al., 2011). Although contextually adaptive in the short-term, when chronic, such vigilance can lead to longer-term wear and tear—or allostatic load—that can compromise cross-system functioning and physical and psychological well-being (Juster et al., 2010).

Methodological Context: Observational studies of children in context provide some support for these ideas. However, the nature of these data typically makes it difficult to infer causal relations. For obvious reasons, experiential risk and parenting quality are not assigned randomly. Moreover, unlike most randomized treatments, children move in and out of these experiences over time—acquiring different ‘treatment histories,’ with respect to dose and/or developmental timing. This is a non-trivial complexity. To make causal inferences with respect to any single ‘dose” of treatment or the accumulation and/or the timing of repeated doses, individuals would need repeated randomization to treatment/ control over time. Suffice it to say, this is rare. Further, traditional regression-based approaches aimed at adjusting time-varying confounds often introduce more bias than they eliminate (e.g., blocked effects, collider bias). These issues have been discussed widely in the econometrics and epidemiology literatures (e.g., Robins, Hernan, & Brumback, 2000); however, they are less well known amongst developmentalists.

Aims: In this talk, we leverage longitudinal data concerning early environmental risk, parenting quality, and children’ stress physiology to introduce one approach from epidemiology——marginal structural models (MSMs; Robins, et al., 2000)—as a way to strengthen casual inferences with time-varying treatments. Specifically, we use longitudinal data from the Family Life Project—a study of 1,292 families in rural poverty—to test the extent to which the respective dosage and timing effects of household chaos and maternal sensitivity on stress physiology in infancy and toddlerhood are evident, after using MSMs to balance observed time-varying confounds across treatment histories. Using this worked example, we describe how MSMs serve as a longitudinal extension of more common inversed propensity-score weighting (IPWT) approaches for balancing observed confounds across treatment levels. We highlight multiple contemporary methods for establishing and checking balance with time-varying treatments and confounds using readily implemented algorithms (e.g., covariate balancing propensity score [e.g., see Figure 1; Imai & Ratkovic, 2012]; machine learning [Karim et al., 2017]). We also discuss practical ways of integrating these weights into one’s substantive models, culminating in evidence of dosage and timing effects of household chaos and maternal sensitivity on children’s cortisol and salivary alpha amylase levels across infancy and toddlerhood.

Authors

• Dan Berry, University of Minnesota
  Presenting Author

• Clancy Blair, New York University
  Non-Presenting Author

• W. Roger Mills-Koonce, University of North Carolina at Chapel Hill
  Non-Presenting Author