Optimal significance levels of prior tests in the presence of multicollinearity