Diagnosis of misconceptions or "bugs" in procedural skills is difficult because of their unstable nature. This study addresses this problem by proposing and evaluating a probability-based approach to the diagnosis of bugs in children's multicolumn subtraction performance using Bayesian networks. This approach assumes a causal network relating hypothesized subtraction bugs to the observed test items. Two research questions are tested within this framework. First, it is investigated whether more reliable assessment of latent subtraction bugs can be achieved by hypothesizing and using subskill nodes in the Bayesian network as causal factors affecting bugs. Second, network performance is evaluated using two types of testing situations, one using binary data (items scored as correct or incorrect) and the other simulating a multiple-choice test format with diagnostic use of specific wrong answers. The resulting four types of Bayesian networks are evaluated for their effectiveness in bug diagnosis. All four networks show good performance, with even the simplest network (bug nodes only, binary data) giving overall bug diagnosis rates of at least 85%. Prediction is best with the most complex network (bug and subskill nodes, diagnostic use of specific wrong answers), for which the correct diagnosis rate reaches 99%. These results suggest that stable and reliable bug diagnosis can be achieved using a Bayesian network framework, but that the stability and effectiveness of diagnosis is increased when the network includes latent subskills in addition to bugs as causal factors, and when specific wrong answers are used for diagnostic purposes. (Contains 4 tables and 3 figures.)