The federal Race to the Top competition provided significant impetus for states to require evidence of student learning in teacher evaluations. A growing number of states are considering or have adopted “value-added” models as a part of their teacher evaluation systems. Such models typically link students to their teachers in the spring semester when statewide tests are administered and estimate a teacher’s performance based on her students’ learning between the test date in the previous school year and the test date in the current year. However, between those two time points, a significant portion of student learning is beyond the control of teachers of the spring semester. Due to data limitations in many states, the effect of all these student learning opportunities cannot be distinguished from, and hence is often mistakenly attributed to the value-added of teachers in the spring classrooms. This study investigates how teacher evaluations are affected by such misattribution.

An important source of this misattribution is instruction by multiple teachers between two consecutive tests, either concurrently or successively. Administrative data in most states provide very limited, if any, information about these student experiences as classrooms are typically observed once during the school year. One of the few exceptions is Florida, which conducts two enrollment surveys and thus provides detailed information about within-semester student movement across schools (e.g. schools attended, entry and withdrawal dates) as well as classrooms in fall and spring semesters (e.g. time spent with each teacher in a given week). Using these data, we compare how commonly used value-added models behave under two informational settings: (1) the limited information setting where each teacher is fully responsible for the students in her classroom at the time of testing; (2) the full information setting where each teacher is responsible for all students she has instructed between two consecutive tests, but at varying rates based on the student’s dosage rate (i.e. percentage of time spent with the teacher). Comparisons between these two settings reveal that the additional information make a significant difference in terms of the estimated teacher value-added. Misattribution results in considerable bias for both reading and math teachers, leading to many teachers being mistakenly labeled as ‘effective’ or ‘ineffective’.   

So far, the fairness of teacher value-added models has typically been assessed based on their capability to adequately isolate the contribution of non-teacher factors on student outcomes. In this study, on the other hand, we investigate how fairly current value added specifications allocate reward/blame for the performance of students among teachers. In the midst of the current move towards teacher-level accountability and as many states design their teacher evaluation systems as mandated by RTTT, these findings provide valuable and timely information for policy-makers and the wider education policy community.