Conventional value-added (VA) models estimate teacher quality as a simple average of the difference between students’ actual and predicted standardized test scores. These models therefore implicitly assume it is just as important to raise test scores of lower-achieving students as it is to raise test scores of higher-achieving students. I consider whether a weighted average of residuals might be more useful. Using data from North Carolina, I find that teacher VA measures become more predictive of teachers’ long-run impacts when the highest-achieving students are weighted more than the median student. These differences in weights may reflect that either (i) small-sample efficiency (some students are more informative about teachers’ true test-score effects than others) or (ii) differences in true effects (e.g. test-score effects for different students might capture different general aspects of teaching). I find empirical evidence supporting both explanations..