Parameter estimation and uncertainty analysis theory provide a series of guiding principles which can be used to assess the costs and benefits of model simplification in decision making. It can be shown that if a modelling prediction is data-informed, partially data-informed or not data-informed, then the approach to modelling can be very different.
We present an empirical demonstration of these guiding principles for the most challenging context of model simplification, e.g. partially data-informed. Successful strategies for partially data-informed problems must avoid the perils of model complexity which include long run times and numerical instability. These strategies must also navigate the perils of model simplification, i.e. errors in uncertainty estimates and predictive bias, so that such models retain the benefits of complexity, namely the ability to quantify uncertainty. This requires that prediction-specific complexity is retained, while those parts of a model that are of secondary importance to management-critical predictions can be simplified.
The empirical demonstration explores the costs and benefits of adopting the concept of “steady state” as our simplification strategy in the context of predicting the increase in the duration of low stream flows in response to pumping of a nearby well.