Dynamic time warp versus vector autoregression models for network analyses of psychological processes

Psychopathological disorders are increasingly conceptualized as complex dynamic systems, which can be represented as networks of interconnected symptoms. These dynamic networks are often constructed using Multilevel Vector AutoRegression (mlVAR) models. However, psychological processes frequently violate these assumptions. An alternative approach for examining temporal relationships between variables is Dynamic Time Warping (DTW). This paper evaluates the potential applications, advantages, and disadvantages of DTW and mlVAR. As part of the Netherlands Study of Depression and Anxiety, an Ecological Momentary Assessment module was administered five times daily for 2 weeks, to 376 participants (Mean age 49.3 years, 64.4% women). We created item networks based on 20 of the mood and physical condition items from this module using the mlVAR and DTW techniques, and repeated these analyses using simulated data to explore violations of mlVAR assumptions, including various lagged relationships and the presence of collider variables. Analysis of simulated datasets revealed that mlVAR networks were more susceptible to spurious connections, while DTW produces more reliable networks under these conditions. While mlVAR better reveals causal relationships when assumptions are met, DTW provides a robust method for examining co-occurrence, synchrony, and the directionality of lagged connections in real-world psychological data.