Generally, proportional hazard and location-scale model classes are used in modeling lifetime data. Both these model classes are derived based on the constant variance assumption of lifetime distribution. Lifetime improvement experiments in quality engineering always aim to maximize mean lifetime of a product. The most important problem in an industrial process is to find the operating condition that achieves the target value of the mean lifetime of a process characteristic, and simultaneously minimizes the process variability. This article examines some real lifetime data. It is identified that variance of lifetime distributions may not always be constant. Consequently, classical lifetime model classes such as proportional hazard and location-scale may mislead the true model of a lifetime data set. This article examines this point with some real examples. These examples are analyzed based on generalized and joint generalized linear models. Present analyses give completely different results from the earlier studies, and in the process, it is observed that log-normal distribution (with constant and non-constant variance) fits well these three data sets.