In a post-mortem study of item sampling, 1,050 examinees were divided into ten groups 50 times. Each time, their papers were scored on four different sets of item samples from a 150-item test of academic aptitude. These samples were selected using (a) unstratified random sampling and stratification on (b) content, (c) difficulty, and (d) both. There were no systematic relationships between method of sampling and accuracy or stability (defined in terms of the means and variances of the distributions of the estimates) of estimated total-test mean or variance. Implications for both generalizability theory and item sampling methodology are discussed. (Author)