MMA: Model Misspecification Assessment

Predictivity of simulations using mechanistic models is hampered by different classes of uncertainties:

  • uncertainties in data
  • uncertainties in model parameter estimation
  • uncertainties in model structure

Assessment and management of uncertainties in data and their effect on model predictivity is thoroughly understood and sufficient methods for handling the respective simulation errors are established in the theory of inverse problems. However, the effects of insufficient parameter estimation and erroneous model structures on predictivity are much less understood. They do not affect the standard error norms on the data used for model fitting but cause higher order error signals on the data. As they have the potential to cause significantly erroneous predictions in extrapolations, uncertainty quantification of the parametric and structural errors is essential to achieve reliability in extension of the model range, such as in translational learning or prediction of behavior in special populations.

Uncertainties in model parameter estimation are typically due to missing of the global optimum of the error functional in the parameter fitting process. This may be caused by an ill-posed parameter identification problem or an insufficient greedy parameter identification strategy.

Uncertainties in model structure are either due to the closure problem excluding “false irrelevant” processes from the model in order to reduce the model complexity or they are based on erroneous understanding of the “relevant” process structure controlling the system.