Investigating parameter identifiability and "sloppiness" in a dynamical model of reading

Theories of many cognitive processes can be expressed as dynamical process models. In order to test the hypotheses that the models implement, we must calibrate them to experimental data by fitting free parameters. In this work, we study a version of the SWIFT model of eye-movement control during reading (Engbert et al., 2005, Psych. Rev.; Engbert & Rabe, 2023, under review) to illustrate two related issues that can arise in models with multiple free parameters: parameter identifiability and sloppiness. The parameters of a model are identifiable for a given data set when it is possible to find a finite confidence interval for the parameter (Raue et al., 2009, Bioninformatics). When a parameter is non-identifiable, parameter fitting can be difficult and misleading, even if the fitted model's output looks reasonable. Sloppiness arises when there are large differences in how sensitive the model's output is to changes in different parameters (Brown & Sethna, 2003, Phys. Rev. E). Sloppiness can also lead to difficulty in model calibration and make interpreting model output challenging, as an analysis of sloppiness often reveals that there are combinations of parameters that vary systematically together with no change in the model's predictions. To our knowledge, parameter identifiability and sloppiness have received little attention in cognitive science, even though the structure of many models is susceptible to these problems. In this talk, we will discuss methods for identifying and addressing parameter non-identifiability and model sloppiness, which can lead to simpler models and more informative fits to experimental data.