Conditions when statistical tests for model discrimination have high power. Some examples from pharmacokinetics,...

The situation where data pairs xi, yi are actually generated by a true model, f(delta, x), but erroneously fitted by a deficient model, g(phi, x), is explored. A function, Q(delta), is described which is the average squared distance between f(delta, x) and the best-fit false model, g(phi, x). For a range of x covering 5-95% 'saturation' of f(delta, x), Q(delta) is calculated numerically for sums of exponentials, binding functions and rational functions. In each case, the region of delta when the second model in the series can be reliably differentiated from the first by statistical tests is described.