A methodology has been constructed to assess the uncertainty in an output consequence calculated by a large code, due to the uncertainties in input data. A sensitivity analysis was first applied to the code to screen the input variables, leaving only those most affecting the output consequences. The variations of these effective inputs were prescribed by an effective combination of statistical designs, which accounted for the linear, quadratic, and two-factor interaction effects of the inputs on the calculated consequence. A key result of the methodology was the probability density function of the consequence of interest, expressed as a distribution of the Pearson family. The confidence level in calculating a consequence was readily obtained from this distribution function. The methodology was applied to the computer code MELT-IIIA, a major code for the analysis of the hypothetical core disruptive accident in liquidmetal fast breeder reactors, and the confidence level in predicting the time of initial pin failure during a transient overpower accident in the fast test reactor was determined. The sensitivity of this confidence level to the uncertainties of the input data was also shown, thereby establishing the need for well-documented statistical properties of data used in nuclear reactor safety analysis.