Educational robotics is widely used to develop computational thinking, but the instructional practices that account for its effects remain poorly specified. This comparative study of robotics programmes in the United Arab Emirates and Japan assesses computational thinking gains among 620 middle-school students enrolled in one-semester robotics courses in six schools, three in each country. Computational thinking was measured with a validated test administered at the beginning and end of the semester, and instructional practices were recorded through classroom observation and teacher logs, yielding measures of the intensity of structured pair programming, reflective journal use, teacher-led direct instruction and open-ended project time for each class. Students in both countries showed substantial and statistically significant gains, with an average effect size of 0.68 standard deviations, and the gains did not differ significantly between countries after controlling for prior achievement. Hierarchical regression revealed that structured pair programming and reflective journals explain most of the variance in outcomes across both contexts, jointly accounting for 41 per cent of the variance in gain scores at the class level, whereas direct instruction and unstructured project time contributed little. Both practices were effective in each setting, with no significant country interaction. The paper argues that the benefits of robotics education depend on specific collaborative and metacognitive practices rather than on the technology itself and discusses implications for curriculum design and teacher professional development in both regions.