Using Predictive Modeling to Target Student Recruitment: Results of a Replication By Gillian Butler Student Affairs Research & Information University of California, Davis
Replication of Controlled Experiments at SUNY Stony Brook Thomas, Reznik & Dawes
Similarities Public, highly selective (63% vs. 56%) Multi-campus system One application, low marginal cost to apply to additional campuses ($40 vs. $30) Similar yield (approximately 30%)
Fence – Sitters vs. Hot Prospects
Goal: To identify which students are most responsive to targeted recruitment efforts Construct a predictive model to estimate individual admits’ probability of enrolling Design a controlled recruitment experiment Assess the results of the experiment
Logistic Regression Model
UC Davis Model Correct predictions = 72%
Predicted and Actual Enrollment by Probability Range, Fall 2000
Distribution of Admits by Probability Band
Mean Combined SAT1 Score by Probability Band UCD Fall 2000 Freshman Social Science Admits
SUNY at Stony Brook 1998 Treatment: Additional invite to visit Two additional mailings to parents Expedited financial aid packaging Contact in financial aid marathons (p) of enrollment 30% – 90% Control group: n=819Experimental Group: n=326
SUNY at Stony Brook 1999 Treatment: Three additional mailings One-half contacted by telephone (p) of enrollment20% – 100% Control group: n=2442Experimental Group: n=700
UC Davis 2001 Treatment: Personalized invitation to lunch w/faculty from major interest area Recruitment CD Note: Social Science Admits Only (p) of enrollment 0% - 60% Control group: n = 3383Experimental group: n = 400
What’s wrong? Theory is incorrect Model was misspecified Size of experimental group was inadequate Level of treatment was insufficient Type of treatment was ineffective