An alternative method is shown to consist of two parts: the test syste
m itself; and a prediction model for converting in vitro endpoints int
o predictions of in vivo toxicity. For the alternative method to be re
levant and reliable, it is important that its prediction model compone
nt is of high predictive power and is sufficiently robust against sour
ces of data variability. In other words, the prediction model must be
subjected to criticism, leading successful models to the state of conf
irmation. It is shown that there are certain circumstances in which a
new prediction model may be introduced without the necessity to genera
te new test system data.