A least squares approach to estimating the reliability of a measure consisting of more than three content homogeneous or congeneric parts is proposed. The advantages of this method over a more indirect approach in which certain parts of a measure are combined to use Kristof's or Feldt's coefficients are examined. One hundred four-part tests were created by randomly assigning certain items from a 48-item mathematics achievement test to each of the four parts. Data on 400 examinees who took the test were divided into five subsets. Seven estimates of reliability were computed for 100 random splits for each subset of data. Some of the main results included: (1) there is a consistent order in the magnitudes of the means of the coefficients, regardless of sample size; (2) values of the average of the Kristof coefficients are greater than unity when sample size is small; (3) the proposed coefficient has smaller standard deviation and range than the Kristof coefficient; and (4) the Feldt coefficient tends to have smaller standard deviation and range than the proposed coefficient, and depends on the assumption that the error variance of a part is proportional to its length. (Author/RL)