Some 500 cases of benchmark calculations on criticality problems for homogeneous experimental systems have been made with the KENO-IV Monte Carlo calculation code using the MGCL cross-section data library. The calculation results have been analyzed to classify the experimental systems so as to make the variance of calculated keff bias as small as possible in each classified system. The trends of bias are identified and illustrated to be optimumly expressed by a multiple variable regression equation in terms of several variables, which adequately correlate with the bias value of keff calculated for the experiments. The uncertainty accompanied by bias correction for calculated keff is clearly determined, and the margin set aside for the experimental error is assessed. Finally, the procedure to estimate nuclear criticality safety is proposed.