Erasmus MC's Deep Generalised Mixed Models merge mixed effects with deep learning

Longitudinal studies like experience sampling (ESM) collect intensive repeated measurements from participants, but they face a dual challenge: traditional statistical models struggle with high-dimensional data, while standard machine learning methods introduce bias when data are missing-not-at-random. To address this, Nina van Gerwen and colleagues from Erasmus MC introduced the Deep Generalised Mixed Model (DGMM), a novel neural network architecture that extends mixed effects models into the deep learning domain. The model flexibly estimates both mean and correlation structures through fixed and random effects, and training relies on a combination of variational auto-encoders and a Bayesian data augmentation algorithm.

DGMM was motivated by the GrowIt! app, which tracked daily emotions in adolescents during the COVID-19 pandemic—a dataset where dropout correlated with negative emotional states, making missingness informative. The authors validated DGMM on simulations and real data, showing it can accommodate generic outcome distributions and provide valid inference under missing-at-random mechanisms. However, performance was suboptimal due to model instability, indicating the architecture needs further refinement before it can be reliably deployed. As an arXiv preprint, the work is not yet peer-reviewed, but it offers a promising direction for blending statistical rigor with deep learning flexibility in hierarchical health data.