Reliable detection of longitudinal change in computational models

Computational models are increasingly used to infer latent cognitive change across development and treatment, yet the reliability of longitudinal change detection remains poorly understood. Here, we present a simulation-based framework for evaluating whether latent change can be detected reliably before data collection. Using reinforcement learning as a testbed, we compared two approaches to change detection: fit-based model comparison and parameter-based thresholding of estimated change parameters. Reliable detection depends not only on effect size and sample size, but also on likelihood geometry, model parsimony, and task design. Fit-based detection provided a more practical basis for validation, whereas parameter-based detection was limited by threshold ambiguity and parameter trade-offs. In regimes with poor detectability, simulation-based task optimisation improved sensitivity without increasing sample size, a finding of particular relevance for studies involving hard-to-recruit populations. Our framework provides a practical route for validating longitudinal change models before their use in developmental or clinical intervention research.