Students’ evaluations of teacher performance (SETs) are increasingly used by universities and colleges for teaching improvement and decision making (e.g., promotion or tenure). However, SETs are highly controversial mainly due to two issues: (1) teachers value various aspects of excellent teaching differently, and, to be fair, (2) SETs should be determined solely by the teacher’s actual performance in the classroom, not by other influences (related to the teacher, the students or the course) which are not under his or her control. To account for these two issues, this paper constructs SETs using a specially tailored version of the popular non-parametric Data Envelopment Analysis (DEA) approach. In particular, in a so-called ‘Benefit of the doubt’ model we account for different values and interpretations that teachers attach to ‘good teaching’. Within this model, we reduce the impact of measurement errors and a-typical observations, and account explicitly for heterogeneous background characteristics arising from teacher, student and course characteristics. To show the potentiality of the method, we examine teacher performance for the Hogeschool Universiteit Brussel (located in Belgium). Our findings suggest that heterogeneous background characteristics play an important role in teacher performance.