The success of applications of item response theory (IRT) depends upon the properties of the estimates of model parameters. Many theoretical properties of these estimates have been extensively studied. However, the properties of estimates obtained empirically from real data depend not only on the theoretical results, but also on the data and the estimation procedures used to obtain them. In this paper, the properties of estimates obtained from a commonly-used implementation of the joint maximum likelihood approach (LOGIST) are examined extensively and shown to be, in part, functions of the properties of the test or item set being calibrated. A small study is also made of the properties of estimates obtained from a commonly-used implementation of the marginal maximum likelihood approach (BILOG). Recommendations are made for the improvement of both procedures. (Contains 3 tables, 23 figures, and 29 references.) (Author)