We propose a Bayesian hierarchical Age-Period-Cohort model to predict elementary school enrollment in New York City. We demonstrate this model using student enrollment data for grades K-5 in each Census Tract of Brooklyn’s 20th School District over the 2001-02 to 2010-11 school years. Specifically, our model disaggregates enrollment into grade (age), year (period), and cohort effects so that each can be interpreted and extrapolated over the 2011-12 to 2017-18 school years. We find this approach ideal for incorporating spatial information indicative of the socioeconomic forces that determine school enrollment in New York City. This work is the result of a 2016 “Call for Innovation” initiated by the Department of Education, and it won the grand prize for having a lower prediction error than competing teams on a held out test set.

Code and data available at github.com/stan-dev/stancon_talks