The Wisconsin Dropout Early Warning System
I co-developed the award-winning Wisconsin Dropout Early Warning System (DEWS) during my time working at the Wisconsin Department of Public Instruction. DEWS is a machine-learning application built on the state longitudinal data system at DPI. It uses hundreds of thousands of records that are mandated to be submitted to the department for various accountability and reporting purposes, analyzes them, and return to schools a prediction of on-time high school graduation for their students in grades 5-9.
While I led the work on designing the algorithm and software, this work was only possible because of co-developing DEWS with educators across the state. Through this co-development process, where I would present the current state of DEWS, educators would share their values, concerns, and insights. Together with a fantastic working group back at DPI we would identify how to incorporate this information into the statistical model or the communication materials around the statistical model. The end result is an ethically designed prediction system that provides schools with a timely list of students who should be considered for additional support and attention. You can learn all about DEWS, and more importantly, the communication and guidance around its use, at the DEWS homepage.
DEWS is a sophisticated enterprise scale machine-learning system built in R. It iterates through more than 50 potential algorithms for each grade-level of students, selects a suite of high-performing, but dissimilar, algorithms for final training, and then blends those algorithms together to maximize its strength. Models are re-trained twice a year, and each grade-level model is built on a training set of hundreds of thousands of records.
But, more important than the technical achievement, DEWS is widely used by schools across Wisconsin. And DEWS is still in use, and has been extended to include college-readiness indicators as well. This success is due to the co-development model, which created ownership among users of the outcome of the system, and created enthusiasm for its use which has sustained it.
DEWS was the first state-wide machine learning early warning system implemented. A number of elements of its development have been covered and discussed in-depth from the technical pieces of building an application on a state longitudinal data system to the critical human element of appropriately communicating predictions to educators in a clear and ethical manner. See the links below to learn more:
Media Coverage of DEWS
- https://academiccommons.columbia.edu/doi/10.7916/D86W9N4XOf Needles and Haystacks (main DEWS Paper)
- Bowers, Sprott, and Taff metaanalysis of EWI performance
- DEWS Announcement blog post
- DEWS comparison blog post
Civilytics has advised a range of educational organizations on the development, refinement, and support of early warning systems including state education agencies, school districts, and vendors. If your organization is interested in implementing an early warning system (you can!) and would like to discuss, please get in touch.