Recently, the Brookings Institution and New America authored white papers criticizing the use of predictive analytics in enrollment management. The overarching concern of these think tanks is that predictive modeling and optimization algorithms may inadvertently introduce or reproduce bias against historically disadvantaged groups during the college admissions and financial aid awarding processes.
Due to the opportunity for bias to arise when using these technologies, the authors’ recommendations range from the rational (e.g., employ humans to oversee the use of algorithmic outputs), to the irrational (e.g., ban predictive analytics and optimization algorithms from admissions and awarding processes).
These authors recommend just about everything except an actual methodological solution. In what follows, I make the argument that these authors’ concerns are important, but their recommendations are myopic and castigate the very tools that help higher education institutions increase opportunity, access and affordability for historically under-served groups. Predictive models and optimization algorithms, when used with intention, can mitigate bias and the reproduction of historical disadvantage in admissions and financial aid awarding processes.
A Simple Between- and Within-Group Approach
Where does the opportunity for bias arise when admissions offices use predictive analytics? Perhaps counterintuitively, the opportunity for bias arises when admissions offices don’t consider advantaged/disadvantaged group membership when they utilize enrollment or awarding predictions.
Enrollment predictions themselves are outcomes of data generated by historical human processes, and to the extent that those human processes were biased in some fashion, predictions based on those processes will be similarly biased. The way to mitigate the reproduction of bias is not to try to eliminate bias in enrollment predictions themselves, but instead, to utilize enrollment predictions within pre-defined groups, with the intention of mitigating bias between groups. Identifying between-group differences can help reveal the presence of bias for disadvantaged groups. Within-group differences can be used to optimize processes while controlling for the influence of group membership.
A simple between- and within-group approach to using the outputs of predictive models and constraining optimization algorithms can mitigate bias and the reproduction of historical disadvantage in admissions and financial aid awarding processes.
For example, say an admissions office used group comparisons to identify declining enrollments in first-generation students over time. The university wants to optimize internal processes to improve their first-generation yield rate. Admitted students can be grouped into first-generation and not-first generation categories, and then enrollment predictions within the first-generation group can be used to optimize the allocation of scarce resources to encourage more first-generation students to enroll. First-generation students with higher enrollment likelihoods can be prioritized by college counselors for outreach, and relevant support materials can be sent to those students at higher rates.
In this example, the knowledge of between-group differences (i.e., bias) is used to intentionally redistribute recruiting resources within a group to mitigate representational disadvantage. Between-group differences are useful for determining whether bias arises, but only the within-group variation is used to optimize internal processes.
Bias Constraint in Optimization Algorithms
Mathematical optimization procedures can also be used to intentionally reduce bias and increase affordability during financial aid awarding. Since it is only the rare and typically highly selective school that can afford to meet the full financial need of each student, most colleges determine a percentage of financial need that each student will be awarded in institutional aid. Awarding procedures can reduce bias and increase access and affordability for under-represented groups when group membership is considered during the optimization phase.
First, subgroups that measure advantage and disadvantage are defined. Typically, this is accomplished by determining the financial need of each student and then grouping students into statistically equivalent ranges of need. To further categorize students for awarding, schools typically use a “grid” or “matrix” that organizes admitted students into financial need and academic performance categories.
An optimization algorithm can be executed across the cells in this grid to test how different award amounts are likely to impact the incoming class size and composition. The algorithm can identify the optimal combination of awards across categories that best balances the total number of likely enrollments with the amount of financial aid offered and the amount of net tuition revenue likely received, all while enforcing strict bias constraints.
With properly defined constraints, this approach ensures that students with greater need are always awarded aid amounts that meet a higher percentage of their need, when compared to students with lesser need.
When projected enrollments are not likely to meet the yield goal, some admissions offices establish scholarships specifically to improve enrollment yield. Admissions and financial aid offices can offer an unoptimized, constant scholarship award increase to every admitted student.
Alternatively, they can vary the scholarship award amount at the student-level, within disadvantaged groups, with help from an optimization algorithm. Such an optimization algorithm can offer more aid when a student is likely to be more sensitive to receiving those dollars, but only for students who share group membership.
To increase affordability and access for historically disadvantaged groups, schools can define the groups that are likely to be under-served in this admissions cycle and provide optimized scholarships for individuals within those disadvantaged groups. This controls for group membership and ensures that students in a disadvantaged group aren’t competing with more advantaged students for the same scholarship dollars.
Increasing Access and Opportunity
Higher education institutions have the responsibility to increase life chances and social mobility for not only the privileged, but also for the disadvantaged. Most schools do not have enormous endowments and cash reserves that can help them weather years of low enrollments. As demographic and economic constraints reduce the population of enrolling students, schools face increasing economic pressure.
Admissions offices do their best to perform both the public good of increasing access and opportunity, while also meeting their own financial budget requirements and enrollment needs. To assist admissions offices in this effort, the solution is not to tell them they can’t use the analytic tools that are available to every other industry. The predictive models, optimization algorithms, and strategies for mitigating bias to help higher education institutions be successful already exist.
Capture Higher Ed helps admissions and financial aid offices use predictive analytics and optimization algorithms to increase opportunity, access, and affordability for historically underserved groups, while meeting their enrollment and revenue targets.
By Pete Barwis, Ph.D., Principal Data Scientist, Capture Higher Ed