The relationships among types of errors observed during probability problem solving were studied. Subjects were 50 graduate students in an introductory probability and statistics course. Errors were classified as text comprehension, conceptual, procedural, and arithmetic. Canonical correlation analysis was conducted on the frequencies of specific types of procedural and conceptual errors. Variables loading on the first canonical pair suggest that general ability is a factor in successful probability problem solving. Additive trees were fit to the correlation matrix for procedural and conceptual errors, and salient clusters identified on the additive tree provided a visual correspondence to the results of the canonical correlation analysis. Additive trees were also used to investigate the structure of the correlation matrix for errors in all four main categories. Results indicate that difficulties in text comprehension and poor arithmetic skills are responsible for a considerable proportion of observed errors in probability problem-solving. Greater emphasis should be placed on these prerequisite skills, as well as on fostering an appreciation of probability concepts themselves. Suggestions for teaching probability concepts are presented, including a model of normative performance that could be taught in an effort to build more efficient schemas for probability problem solving. Three tables and two figures are included. An appendix summarizes error codes. (Contains 23 references.) (Author/SLD)