Although automobile crash test data have a comparatively large variation because of the complexity of the tests, only a limited number of crash tests are usually conducted due to monetary and time limitations. Thus, it is necessary to control input variables that cause the variation in test data to obtain consistent crash test results and to correctly assess the safety performance of an automobile under development. In this study, a MADYMO model was validated deterministically to yield the head, chest, pelvis deceleration pulses of anthropomorphic test devices and the belt load pulses similar to those from actual tests, and it was also validated stochastically to yield means and standard deviations of the head and chest injury numbers, i.e., $HIC_{15}$ and 3 msec clip similar to those from actual tests. A stochastic analysis was conducted using the validated MADYMO model to calculate the sensitivity of the standard deviations of the injury numbers to the standard deviations of influential input variables to determine the most influential input variable that makes the largest contribution to the variation in the injury numbers. Moreover, the Taguchi approach was used to determine the optimal values of the influential input variables to improve safety performance.