The leadership of Dickinson College understands the importance of predictive analytics to help shape their class. But they also know that models can get stale, and will improve only when they are transparent and institutional departments collaborate during the model building and enrollment process.
In this web seminar, the vice president for enrollment, marketing and communications and the senior research analyst at Dickinson discussed how they use predictive analytics to build the best possible enrollment models, as well as the importance of creating a feedback loop between the enrollment and institutional research departments, and some of the key lessons they have learned during the process.
Vice President for Enrollment, Marketing and Communications
Senior Research Analyst
Senior Data Analyst
Rapid Insight Inc.
Jon MacMillan: Rapid Insight is a predictive analytics and data blending software company. We were founded in 2002 and our goal has always been to make data analysis and predictive modeling easy for any customer regardless of ability and background. We started out in higher education, and that’s sort of been our main focus. We have over 200 institutions nationwide that run the gamut from graduate schools to law schools, community colleges and two- and four-year public and private colleges and universities.
When we talk about predictive modeling in higher education, there are two main focuses: admit to enroll and retention. When you bring a tool like Rapid Insight in-house, you have an automated predictive modeling solution. We could be looking at enrollment, admit to enroll, or a summer melt model of early success indicators. Any of the questions that you want to ask can be asked and answered with our tools.
We give you the tools to integrate data from any source, then an automated process to clean up that data, perform those statistical functions and find the key indicators of whichever outcome you’re investigating, and automatically build predictive models.
We can output results back to database tables, extracts, reports and dashboards; or if you’re using third-party applications like Tableau, we can output directly there as well. We give you the ability to look at this data at any level. For example, we can look at the student level and see each student’s enrollment likelihood. If we’re interested in overall college or major enrollment, we can see that as well. We’re also often asked to produce institutional-level reports, such as: What is our overall enrollment going to be and how does that affect our financial aid outlay?
We offer the tools to be able to investigate at any level.
Stefanie Niles: Dickinson College is a selective private liberal arts college. We admit less than half of those who apply—we enroll between 2,300 and 2,400 undergraduate students. Our incoming class usually ranges between 610 and 640 students, and we retain those at a rate of 90 to 92 percent. That’s been our average for the last two or three years. Our four-year graduation rate is 80 to 81 percent.
Korey Paul: We’ve constructed enrollment prediction models with Rapid Insight since 2008. The software is incredibly intuitive, and the support from Rapid Insight has been invaluable.
Having an external vendor for modeling is valuable for verification and validation, but doing internal modeling allows us a greater breadth and accuracy, and gives us a hands-on daily view of the data. We’re able to cater the models based on internal knowledge. External services, in my experience, tend to have canned models, which are variables that tend to be related to enrollment anywhere and that are not specific to an institution. We’ve seen lower accuracy with those external models. Historically our internal models have allowed us to predict within plus or minus 10 students.
I use Veera for the data construction, putting together the previous years of data, then I build the model in Analytics, and then I bring that back into Veera for daily scoring and updating. You can build the model and do the scoring in Analytics without Veera, but it’s a lot smoother to get the data together in a workable format in Veera.
When we’re constructing a model, we have to figure out how many years we want to put into the model. Typically, we build out three-year models, sometimes four-year. You also need to consider what your key drivers are within the model. If you’re seeing variables that tend to be related, such as campus visits or multiple campus visits or multiple forms of contact, then you can expect a higher potential for yield. But then you also have to manage those expectations and realize that models are fairly accurate only if students behave the same as they have in the past. And the reality is, we don’t know. So we’re always considering new possible approaches to building out new models.
Stefanie Niles: That’s is something that we talk about each and every year in great depth. What have we seen that worked? What are aspects of our process that we need to think about differently? What do we need to keep the same? Because by being consistent we have a greater likelihood of understanding how students will perform. That collaborative nature is definitely a significant part of the way we operate.
Keeping strong lines of communication open is also important. One specific example from this year is a particular student we talked about in a committee-type setting. Everything about that student predicted that the individual would enroll and would be a solid contributor to the class. But the staff member who was reviewing the application—who had spent many years as a high school counselor—could just tell that something was off the mark. Going back to that student’s high school counselor, we found that student had some academic issues that led us to believe they would not be a good fit. Again, it wasn’t evident to us in the model, but that’s where maintaining those strong lines of communication became valuable so that both sides understood what the data was telling us.
Korey Paul: You have to keep your assumptions in check about what you think is happening with the incoming class and just trust your model. Trust that it is looking at bits of the variables in connection to all of the other variables. Then continue to find new approaches of figuring out what’s happening with the actual behavior of the students.
To watch this web seminar in its entirety, visit www.universitybusiness.com/ws071317