Identifiability and indistinguishability of linear compartmental models

Structural identifiability is concerned with whether the parameters of a specified model can be identified (uniquely or locally) from a given input-output experiment, assuming perfect data, while regional indistinguishability is concerned with generating all models, within a given class, which are input-output indistinguishable from the specified model over some region of parameter space. The two types of analysis are compared and contrasted for linear compartmental models and it is shown that indistinguishability analysis, using algebra alone, often presents a very difficult problem. This is alleviated through the use of a set of geometrical rules, based on structural controllability and observability, which must be obeyed for all candidate models for indistinguishability. The rules are necessary but not sufficient and it is shown by an example that establishing the regions of positive parameter space over which one model is indistinguishable from another can sometimes prove a lengthy and difficult calculation.