New Model Enables Person-Specific Lag Detection in Longitudinal Data

A new statistical model, DTVEM-RE (Differential Time-Varying Effect Model with Random Effects), has been introduced to address a limitation in existing methods for analyzing intensive longitudinal data. While the original DTVEM is effective for identifying optimal time lags, it assumes a uniform lag structure across all individuals. DTVEM-RE overcomes this by allowing each person to have their own unique lag coefficients, aligning with the modern clinical research premise that individual differences are crucial. The model comes in two versions: a discrete-time hierarchical Bayesian Vector Autoregression (VAR) implemented in Stan, which pools information across individuals to provide calibrated uncertainty estimates, and a continuous-time per-person Ornstein-Uhlenbeck model in ctsem, designed to handle unevenly spaced data directly. Extensive simulations demonstrated that the Bayesian version accurately recovers between-person spread with low bias and high coverage. When applied to a real-world EMA dataset, DTVEM-RE showed that person-specific lag-1 effects varied significantly across mood items, and it provided the best one-step-ahead predictions among several discrete-time methods. Furthermore, a multi-lag version revealed that the lag at which individuals differ most can vary by item, a nuance missed by single-lag methods. The two versions of DTVEM-RE showed strong agreement on person-specific lag-1 estimates, confirming its robustness. This marks the first person-specific implementation of DTVEM-style lag detection.