The increased detection of ductal carcinoma in situ (DCIS) by mammographic screening, the greater use of breast-conserving surgery, and the recognition that certain histological subtypes are associated with a greater risk of local recurrence has led to the formulation of several new classifications of DCIS in recent years. There are, however, no data concerning the degree of consistency with which these schemes can be applied by reasonable numbers of pathologists. Thirty-three cases of DCIS were thus examined by a working group of 23 European pathologists who categorized them using five recently published classifications: (1) that of the European Pathologists' Working Group based on differentiation (a combination of nuclear grade and cell polarization) with categories of poorly, intermediately, and well differentiated; (2) one based entirely on nuclear grade with categories of high, intermediate, and low, currently in use in the UK national and EC-funded breast screening programs; (3) the same classification in which only two categories, high nuclear grade and other, were used; (4) the Van Nuys system in which lesions are divided into high grade, non-high grade with necrosis and non-high grade without necrosis; and (5) a two-category classification based entirely on the presence or absence of comedo necrosis. Of the three systems with three categories, Van Nuys gave the highest overall kappa statistic of 0.42. Others gave similar values of 0.37 and 0.35 showing that assessing cell polarization in addition to nuclear grade neither improves nor worsens consistency. In all three systems, the middle category was associated with the lowest value for kappa. Of the two systems with two categories, that based on nuclear grade gave the highest overall kappa of 0.46 and that based on comedo necrosis the lowest of 0.34. The most robust histological features were thus high- and low-grade nuclei and necrosis as long as the latter did not involve the recognition of a comedo growth pattern. These values probably represent the maximum achievable, at least by reasonable numbers of pathologists in everyday practice. They are better than those previously reported for classification based entirely on architecture, but further improvement is needed.