Approximating Trends to Recover Dynamics: BurstiVAR for Vector Autoregressive Modeling Without Trend Specification

Google map

Intensive longitudinal data often contain two co-occurring components: gradual mean-level changes over time and short-term dynamic processes such as autoregressive and cross-regressive effects. When researchers attempt to estimate both simultaneously on the same time scale, trends and dynamics become entangled, and most existing methods require specifying a particular functional form for the trend in order to disentangle them. This talk introduces BurstiVAR, an approach that sidesteps this identification challenge by partitioning the time series into discrete segments (bursts), freely estimating person-specific intercepts for each burst to absorb mean-level changes, and modeling within-burst observations as deviations from these intercepts according to a VAR process. Because no functional form is imposed on how the intercepts change across bursts, the approach can theoretically accommodate arbitrary mean-level trajectories while preserving unbiased recovery of the within-burst VAR parameters. I present results from two Monte Carlo simulation studies evaluating BurstiVAR across varying burst lengths and sample sizes, and comparing it to a parametric variant (GoBurstiVAR) that constrains intercepts to follow a Gompertz growth function. Results show that BurstiVAR recovers VAR parameters with acceptable bias and appropriate coverage even with as few as two bursts and three time points per burst, given a sufficient sample size. The method remains robust even when the true trend is a continuous nonlinear function that BurstiVAR does not explicitly model. I illustrate BurstiVAR with data from the Student Engagement, Learning, and Flourishing (SELF) project, a 21-day daily diary study of college students' belonging uncertainty and alcohol use (N = 621), in which pronounced weekday–weekend drinking cycles are exploited to define five alternating high- and low-consumption bursts within continuous diary data.