Why Isn’t Scholarship (always) Predictive?

Lately I’ve had a few conversations with college enrollment staff who are concerned that scholarship dollars are not entering their models as major predictors of whether or not a student will enroll. Worse, some colleges are finding scholarship is negatively correlated with enrollment – that is, the higher the scholarship offered, the less likely a student is to enroll! Understandably, this kind of result can be disconcerting – even unbelievable. It goes against the very reasonable assumption that the more money you offer someone, the more they’ll like you.

The underlying reason behind this modeling behavior is actually pretty reasonable, once we take a breath and think about it. Typically, academic scholarships are based upon academic performance. Put another way, you could probably predict the amount of scholarship a student will receive based on their GPA, SATs, and other characteristics. Depending on the way that scholarships are awarded, these variables act as proxies for the scholarship amount. So, if you include those more predictive proxy variables in a modeling dataset, they often displace scholarships in the model. Another common issue is that while scholarship might enter a model, it might seem to have a negative effect on enrollment – so the more money offered to a student, the less likely they might be to enroll. The reasoning for this behavior is that schools generally tend to award higher scholarships to those students who perform very well academically.  Students who perform well typically have more options available to them, making them less likely to enroll at any particular school than a student who did not perform as well. Does this mean more scholarship money drives students away?  Definitely not- but the way scholarships are awarded tends to reflect the financial aid packaging techniques of the college awarding the scholarship.

While raw scholarship amount may not be the best predictors, there are a few tricks to try if you’d like to test whether scholarship is predictive at all. One idea would be to create flags for the various scholarship types that can be awarded. Prospective student would be assigned a “1” if they received the scholarship and a “0” otherwise. Another would be to flag whether a student received any scholarships. Creating new versions of variables in your dataset gives the model more information to work with and can increase the likelihood of a particular variable entering a model – if indeed it is adding value to the model’s predictions.

Decentralize analytics.
Harness the power of many.

Create and share reports and datasets across the enterprise, and put analytical power in the hands of everyone. Veera creates a truly data-driven culture. Try it for yourself today.



Decentralize analytics. Harness the power of many.