We propose a novel approach to examine uncertain sources that cause variations in solder joint lifetimes and to predict statistical distributions of solder joint lifetimes under actual operating conditions. The uncertainty in input variables of a life prediction model is propagated to the output. The variations in the output are statistically compared to experimentally-measured lifetimes with censoring data. A set of input variables that minimize the discrepancy between predicted and experimental results is obtained through the optimization technique. The proposed approach is implemented for chip resistor assemblies, and the fatigue life of solder joints under the actual operating conditions is predicted after calibration.