The method of paired comparisons is one of the most common experiment techniques to quantify image quality. The common data analysis is the Thurstone's traditional procedure that assumes normal distribution and equal variance of observer judgments across the test samples. When the distributions of observer judgments are apparently unequal, less restrictive method should be used. However, most existing statistical procedures for the purpose are limited in practice to the data with complete pairwised comparisons and with no unanimous judgments. This paper describes a new procedure that can overcome the limitations. It utilizes linear regression and iterative error reduction techniques to directly estimate the standard deviations of observer judgments and the scale values of test samples without approximation.