The number of years to estimate value-added measures (VAM) has received insufficient attention. Researchers often use as many years as available, to increase precision and decrease transitory sorting bias. However, this decision has little theoretical or empirical backing. We develop a theoretical framework and evaluate data from thousands of teachers in three large school districts across several years. For evaluations of past effectiveness for policy purposes, one should almost always use one-year of VAM, while for personnel purposes, use one year of VAM or take a weighted average. Predicting future effectiveness depends on mean reversion or persistence in VAM.