Testing students and using test information to hold schools and, in some cases, teachers accountable for student achievement has arguably been the primary national strategy for school improvement over the past decade and a half. Tests are also used for diagnostic purposes, such as to predict students at-risk of dropping out of high school. But there is policy debate about the efficacy of this usage, in part because of disagreements about whether tests are an important schooling outcome. We use panel data from three states – North Carolina, Massachusetts and Washington State – to investigate how accurate early test scores are in predicting later high school outcomes: 10th grade test achievement, the probability of taking advanced math courses in high school, and graduation. We find 3rd grade tests predict all of these outcomes with a high degree of accuracy and relatively little diminishment from using 8th grade tests. We also find evidence that using a two-stage model estimated on separate cohorts (one predicting 8th grade information using 3rd grade information, and another predicting high school outcomes with 8th grade information) only slightly diminishes forecast accuracy. Finally, the use of machine learning techniques increases accuracy of predictions over widely used linear models, but only marginally.

Working Paper 235-0520 was originally released in May 2020. This is an updated version, released August 2021.