Multiple-choice-questions are common in medical examinations, but guessing biases assessment results. Confidence-based-testing (CBT) integrates indicated confidence levels. It has been suggested that correctness of and confidence in an answer together indicate knowledge levels thus determining the quality of a resulting decision. We used a CBT approach to investigate whether decision quality improves during undergraduate medical education. 3rd- and 5th-year students attended formative multiple-choice exams on pharmacological issues. Students were asked to indicate their confidence in a given answer. Correctness of answers was scored binary (1-correct; 0-wrong) and confidence levels were transformed to an ordinal scale (guess: 0; rather unsure: 0.33; rather sure: 0.66; very sure: 1). 5th-year students gave more correct answers (73 ± 16 vs. 49 ± 13%, p < 0.05) and were on average more confident regarding the correctness of their answers (0.61 ± 0.18 vs. 0.46 ± 0.13, p < 0.05). Correlation of these parameters was stronger for 5th-year students (r = 0.81 vs. r = 0.52), but agreement of confidence and correctness ("centration") was lower. By combining the Bland-and-Altman approach with categories of decision-quality we found that 5th-year students were more likely to be "well-informed" (41 vs. 5%), while more 3rd-students were "uninformed" (24 vs. 76%). Despite a good correlation of exam results and confidence in given answers increased knowledge might be accompanied by a more critical view at the own abilities. Combining the statistical Bland-and-Altman analysis with a theoretical approach to decision-quality, more advanced students are expected to apply correct beliefs, while their younger fellows are rather at risk to hesitate or to act amiss.