This kind of study is difficult to refute, given the strong internal validity (random assignment of caste to exams) and external validity (actual grading versus a laboratory setting).
The study involved an exam competition in which teachers graded exams that included a cover sheet listing randomly assigned characteristics (caste, gender, and age) about the students. The researchers chose India because of its entrenched caste system and its practice of recording students’ caste and religious background.[…]
The study showed that some teachers graded the exams they believed to be of upper caste higher by 3 percent to 9 percent higher. Interestingly, low-caste teachers drove the discrimination against lower-caste students. No discrimination was found among higher-caste teachers, and overall there was no evidence of discrimination on the basis of gender or age.
The study also showed that papers graded earlier in the process received lower scores for students from minority groups, with this effect fading as grading continued. This suggests, the researchers say, that as teachers became more confident in their understanding of a test, they were less likely to rely on student social status to determine grades, and that improving teachers’ skills with testing instruments may reduce discrimination.