Cross-Symptoms Interactions Fail to Predict Future Symptom States Across Four Ecological Momentary Assessment Datasets

Symptom networks have emerged as a popular framework for characterizing the interactions among psychiatric symptoms. Estimated from longitudinal ecological momentary assessment (EMA) recordings, they promise individualized insights into psychopathology and its treatment. Yet whether these networks can be robustly estimated from the short, noisy, and irregularly sampled time series typical of EMA studies remains an open question. We address this through the lens of predictive validity: do estimated cross-symptom connections improve forecasts of future symptom states? In a systematic analysis spanning four public EMA datasets covering different symptoms and sampling schemes, we find that they largely do not. In 273 of 313 participants, symptom networks are outperformed by networks that only use every individual symptom's past to predict its future, indicating that estimated cross-symptom connections tend to worsen out-of-sample predictions rather than improve them. This pattern holds across estimation strategies from unregularized Ordinary Least Squares (OLS) to multilevel hierarchical modeling and nonlinear state-space models, suggesting that cross-symptom connections are difficult to identify given the characteristics of real EMA data. Simulation studies support this interpretation, showing that even when ground-truth data is generated by a network model mirroring realistic EMA conditions, cross-symptom structure is often not recovered with sufficient precision to yield a predictive advantage. These findings raise questions about the interpretive framework built on idiographic symptom networks, and suggest that progress may require either measurement designs that make cross-symptom influence more identifiable, or models with less ambitious inferential targets than the recovery of directed cross-symptom effects.