If teacher value-added estimates (VAEs) are to be used as indicators of individual teacher performance in teacher evaluation and accountability systems, it is important to understand how much VAEs are affected by the data and model specifications used to estimate them. In this study we explored the effects of three conditions on the stability of VAEs and evaluated their relative impact. We varied how we accounted for differences among students in their prior learning; whether we estimated VAEs from single or multiple cohorts of students; and the number of students contributing to the VAE for each teacher. Using data from one of the largest school districts in the nation, we created a single, complete data set and used it to estimate several sets of VAEs for each of the 3,651 5th grade mathematics teachers. We found that approximately two thirds of teachers' were stable in that they remained in the same performance group across all conditions. We also found that differences in number of students used for VAEs accounted for up to one third of teacher reclassifications into different performance groups; single versus multiple cohort models accounted for about one fifth; and different methods for controlling for student prior learning accounted for about one sixth (16%) of teacher reclassifications. We relate our findings to characteristics of our data and discuss implications for educational policy. (Contains 7 tables and 5 figures.)