When Brenda Ulin started developing new ways to use institutional data in the 1990s, she and her colleagues knew they were doing meaningful work.
But they didn’t expect that one day they’d be helping to predict who would enroll—and stay enrolled—at the UI.
Working with biostatisticians at the College of Public Health and recruitment and retention specialists in the Office of Enrollment Management and University College, they’re using predictive models to forecast which undergraduates are likely to apply, accept admission, enroll, return year over year, and ultimately graduate.
“We’ve mapped more than 400 data attributes to behavior that cover areas such as demographics, academic performance, behavioral characteristics, and engagement,” says Ulin, group leader of the university’s Business Intelligence Shared Service Center.
The team has developed models that support planning and outreach across the student life cycle. The prospect index is a numeric score that rates an individual’s probability of applying to Iowa, while the admit model predicts whether an applicant will enroll.
Once a student enrolls, retention and persistence models predict her or his likelihood to return. Additional models can project cumulative GPA, and a years-to-graduation model is in the development queue.
Turning to faculty experts
The initiative has roots in OneIT’s Business Intelligence Project, which included a strategic needs assessment and the development of admissions data dashboards based on predictive models developed in 2015 by Enrollment Management analysts and Public Health biostatisticians. In 2017, new funding supported expansion to add a focus on student success and retention.
Today, the team is developing models that assess the impact of financial aid awards on application and enrollment. They’re also helping admissions staff chart whether outreach to younger high school students results in earlier college decisions.
It’s rare for a university to develop its own enrollment-prediction models in-house—many schools look to outside vendors for help.
“At meetings, we frequently have peers ask, ‘How did you get there?’” Ulin says. “Working with our functional experts and our biostatisticians is a key factor. We have amazing knowledge, creativity, and opportunities for mutually beneficial partnerships.”
Faculty members Grant Brown and Knute Carter with the College of Public Health’s Center for Public Health Statistics lead development of the project’s predictive models, working with graduate students Uche Nwoke and Josh Tomiyama (and, previously, Daren Kuwaye and Biyue Dai).
The project is one of several where the center is helping university units address business needs. It offers graduate students hands-on experience with advanced analytical techniques in a production environment.
“We use the university’s high-performance computing system to help us tune our models,” says Brown, an assistant professor of biostatistics. “Machine-learning methods depend on a number of algorithm parameters and selecting which configurations to use in the final models is very computationally intensive. The HPC system lets us experiment with hundreds or thousands of potential configurations in a way that otherwise wouldn’t be feasible.”
Making data-driven decisions
In addition to making individual predictions, Brown and colleagues are exploring new ways to produce aggregate forecasts. “Our individual-level models use hundreds of variables to rank students with respect to outcomes,” Brown says. “Aggregate models can complement these predictions with less sensitive but more stable estimates of quantities like total retained students.”
Such predictions might help academic units gauge resource needs the way individual models help Enrollment Management teams make data-driven decisions about whom to target.
“We use predictive modeling to determine where we spend recruitment time, money, and resources,” says Mike Hovland, director of enrollment management data. “Counselors use data to decide which high schools to visit or which students to contact individually with email or postcards. Event staff use modeling to determine who received registration priority, and financial aid staff use it to manage scholarship budgets.”
Predictive models also help staff and faculty identify enrolled students who might be at risk of dropping out. They’re integrated into Excelling@Iowa, a UI-built system used by 500-plus users in more than 30 departments to track and support student success.
“Excelling@Iowa leverages predictive models along with responses to a transitional survey taken by first-year, transfer, and some returning students,” says Danielle Martinez, associate director of Academic Support and Retention for University College. “It helps us prioritize outreach and have more timely and meaningful conversations with students.”
In turn, Hovland, Martinez, and their colleagues provide invaluable perspective on factors that influence student decisions. “They identify process and policy changes—the university’s decision to join the Common Application, for example, or changes in scholarship award schedules—that might affect how our data skews,” Ulin says.
She and fellow data scientists couldn’t see the future when they started building the UI data warehouse. But their work from decades ago turns out to be critical today.
“Our statisticians tune their models based on historic activity, using data snapshots that go back years. Without all that data we couldn’t have developed to this degree,” Ulin says.
“Index strategies designed for aggregate-level planning needs in conjunction with individual predictions will shift the way campus planners approach strategic and operational decision making,” she adds. “We’re excited to be part of the movement toward a data-informed culture on campus.”