Peer Instruction (PI), a student-centred teaching method, engages students during class through structured, frequent questioning, facilitated by classroom response systems. The central feature of PI is the ConcepTest, a question designed to help resolve student misconceptions around the subject content. Within our coordination chemistry PI session, we provide students two opportunities to answer each question -- once after a round of individual reflection, and then again after a round of augmented reality (AR)-supported peer discussion. The second round provides students with the opportunity to "switch" their original response to a different answer. The percentage of right answers typically increase after peer discussion: most students who answer incorrectly in the individual round switch to the correct answer after the peer discussion. For the six questions posed, we analysed students' discussions, in addition to their interactions with our AR tool. Furthermore, we analyse students' self-efficacy, and how this, in addition to factors such as ConcepTest difficulty influence response switching. For this study, we found that students are more likely to switch their responses for more difficult questions, as measured using the approach of Item Response Theory. Students with high pre-session self-efficacy switched from right-to-wrong (p < 0.05) and wrong-to-different wrong less often, and switched from wrong-to-right more often than students with low self-efficacy. Students with a low assessment of their problem solving and science communication abilities were significantly more likely to switch their responses from right to wrong than students with a high assessment of those abilities. Analysis of dialogues revealed evidence of the activation of knowledge elements and control structures.