Decision trees that are limited to testing a single variable at a node are potentially much larger than trees that allow testing multiple variables at a node. This limitation reduces the ability to express concepts succinctly, which renders many classes of concepts difficult or impossible to express. This paper presents the PT2 algorithm, which searches for a multivariate split at each node. Because a univariate test is a special case of a multivariate test, the expressive power of such decision trees is strictly increased. The algorithm is incremental, handles ordered and unordered variables, and estimates missing values.