Abstracts of Distinction: Applying Time-Varying Effect Models (TVEM) to Prevention and Developmental Science

Introduction: The purpose of this paper is to demonstrate the utility of Time-Varying Effect Models (TVEM) (Li et al., 2015). Traditional longitudinal models (multi-level models, growth models) produce a single mean estimate of the effect of a time-varying covariate on an outcome. TVEM estimates a smoothed nonparametric regression of the outcome on time, as well as the smoothed association between a covariate and outcomes across time.

Rationale: We demonstrate several potential applications of TVEM for prevention and developmental science: (1) Mapping normative trajectories of developmental constructs, (2) Identifying constructs’ susceptibility to time-varying ecological conditions, (3) Examining prevalence rates of behaviors or ecological conditions across historical time, and (4) Expanding traditional event history analyses by linking event timing to outcomes with greater time specificity.

Method: We conducted TVEM analyses with data from the National Longitudinal Study of Youth 79 Child and Young Adults cohort. As part of a cohort-sequential design, adult participants and their children were assessed beginning at the first child’s birth; we utilized data from biennial data collection from 1988-2012 and coded child age to the nearest month, resulting in coverage of ages birth to 14 years in nearly continuous time. Analyses address the four applications described above: (1) Levels of maternal cognitive stimulation and emotional support across early childhood, (2) Their association with each other, and the time-varying association between income poverty and each construct, (3) The historical prevalence of spanking children under three and the association between child behavior problems and spanking from 1988-2012, and (4) The prevalence of bearing a second child within two years of the first, based on maternal age at first birth. The optimal form of each coefficient function was selected by P-spline smoothing; we interpreted intercept and coefficient curves with 95% confidence intervals.

Results and Implications: The demonstration indicated that TVEM provides additional, complimentary information to traditional longitudinal modeling approaches by identifying critical time periods during which developmental constructs (i.e. parenting, spanking) are most prevalent and most associated with other time-varying behaviors. Importantly limitations include an inability to draw conclusions about growth, given that the data reflect estimates similar to cross-section data, and an inability to model mediation or causal effects. TVEM represents one analytic tool for addressing questions and generating future hypotheses about sensitive periods for intervention in both age and historical time that are uniquely applicable to prevention and developmental science.