Inappropriate treatment of missing data may introduce bias into the value-added estimation. We consider a commonly used value-added model (VAM), which includes the past student test score as a covariate. We formulate a joint model of student achievement and missing data, in which the probability of observing a test score depends on observing the test score in the previous grade. We develop a correction procedure that removes the bias in VAM estimates due to data missing not at random and use it to estimate the value-added of elementary school teachers in North Carolina. We find that errors in VAM and missing data equations are uncorrelated, and corrected estimates of teacher productivity are very similar to their ordinary least squares counterparts. Thus, missingness does not cause bias in our data. We also perform simulations to identify scenarios where a correction may be necessary.