It is well-known from experience that low order models perform well in change detection problems even if the true system is quite complicated. By computing the asymptotic power function, it is here shown that common hypothesis tests proposed in literature attain their maximum power for a low order model under certain conditions on how much each parameter change.