Using Predictive Modeling To Manage and Shape Your Enrollments Kevin Crockett President and CEO February 21, 2008 According to the 2008 Institutional Fact Finders submitted in preparation for this conference… • 14% of institutional respondents reported using predictive modeling in their marketing and recruitment programs • 36% reported that they systematically contact inquiries to code their level of interest • 29% reported that they use data analysis to predict dropout proneness What is predictive modeling and how can it support your enrollment management efforts Resource scarcity requires enrollment managers to effectively understand and manage student propensity to enroll/re-enroll Means of qualifying student interest in and commitment to your institution • • • • • • Research/data analysis Tracking student contacts/behavior Telecommunications Personal contact Reply mechanisms in all correspondence Predictive modeling Predictive modeling (pri*dik*tiv mod*el*ing) Statistical analysis of past student behavior to simulate future results Why is funnel qualification important? • Focuses scarce time and resources on those students with the greatest propensity to enroll/reenroll • Facilitates better relationship-building • Enables university staff and advocates to follow-up with students that are genuinely interested in your school • Provides cost-savings by not communicating equally with every student • Enables greater personalization with students • Increases the precision of enrollment forecasting Nationally…enrollment funnel dynamics are changing Source: Noel-Levitz 2006 Admissions Funnel Report Predictive modeling has become more important as the distinction between stages has become blurred The ultimate goal is to build a critical mass of “good fit” students throughout the enrollment funnel How are predictive models built and how well do they work? Models can be built from each stage of the enrollment funnel but they should ultimately predict enrollment or re-enrollment Pre-prospect model Prospect model Inquiry model Applicant/admit model Retention/progression models Modeling converts each trait or behavior into a statistical value Relative Strength of Model Variables Initial Source Code (27.7%) 4.2% First Major as Inquiry (23.4%) 6.0% 27.7% 7.4% 10.1% Enrollment Planning Service Code (8.7%) Categorized Days as Inquiry (12.6%) Email Indicator (10.1%) Categorized Income (7.4%) 12.6% 8.7% 23.4% SCF Code (6%) Prob. of "Mainstream Families" Group (4.2%) Sample inquiry model Sample admitted student model Relative Strength of Model Variables 3.4% Enrollment Planning Service Code (20.9%) Campus Visit Flag (24.3%) 4.3% 4.0% 20.9% 9.1% Categorized No. of Days as Admit (22.4%) SAT Composite Score (11.5%) Primary Academic Interest (9.1%) 11.5% 24.3% 22.4% Binned Distance from Campus (4%) Multiple Self-Initiated Contacts Flag (4.3%) Prob. of "Settled In" Cluster (3.4%) The “Hold” and “Main” Files Models should be built using one half of your historical file so that they can be tested against the other half of your file This ensures that you understand the performance of your model before you ever use it to prioritize your follow-up with prospective students Sample model performance chart Distribution of Model Scores 45% 40% 35% 30% 25% 20% 15% 10% 5% 0% Enrolled Not Enrolled 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Model Score 60% of non-enrollers scored <.30 while less than 4% of enrollers had these scores A model’s output ENROLLED 1 ENROLLED Kate Black .99 Highly Likely Mike Miller .85 Highly Likely Dave Hamilton .72 Likely Jerrica Zwick .68 Likely Angie Mabeus .46 Somewhat Likely Audrey Keppler .41 Somewhat Likely Brian Schuler .21 Less Likely Jordan Clouser .17 Less Likely NOT ENROLLED 0 NOT ENROLLED Sample predictive model performance Model Score Inqs Apps Conv. % Enrolled Yield 0-.20 3,913 21 .5% 4 .1% .20-.29 9,349 87 .9% 12 .1% .30-.39 13,772 107 .8% 14 .1% .40-.49 14,602 172 1.2% 40 .3% .50-.59 10,369 242 2.3% 56 .5% .60-.69 9,085 337 3.7% 66 .7% .70-.79 5,870 512 8.7% 139 2.4% .80-.89 5,305 1,006 19.0% 297 5.6% .90-1.0 8,792 4,965 56.5% 2,289 26.0% 81,057 7,449 9.2% 2,917 3.6% Total At .90 or greater, 11% of the inquiry pool produced 67% of the applications and 78% of the enrolled students. Fall 2007 average client model performance Score Range Inquiry Applicant Admit Gross Deposit Applicant / Inquiry Admit/ Inquiry Gross Deposit/ Inquiry Applicant Lift Admit Lift Gross Deposit Lift 0.00-0.09 3452 101 43 10 2.9% 1.2% 0.3% 0.21 0.13 0.09 0.10-0.19 25455 710 304 79 2.8% 1.2% 0.3% 0.20 0.12 0.10 0.20-0.29 101900 3801 2202 466 3.7% 2.2% 0.5% 0.27 0.22 0.15 0.30-0.39 205783 11685 7770 1782 5.7% 3.8% 0.9% 0.42 0.38 0.28 0.40-0.49 216739 18109 12482 3090 8.4% 5.8% 1.4% 0.61 0.58 0.46 0.50-0.59 153786 19813 14017 3891 12.9% 9.1% 2.5% 0.94 0.92 0.81 0.60-0.69 119424 21496 15641 4593 18.0% 13.1% 3.8% 1.32 1.32 1.24 0.70-0.79 86453 22442 16463 5327 26.0% 19.0% 6.2% 1.90 1.92 1.98 0.80-0.89 58264 22035 16804 5927 37.8% 28.8% 10.2% 2.77 2.91 3.27 0.90-1.00 30170 16582 13422 5997 55.0% 44.5% 19.9% 4.02 4.49 6.39 1001426 136774 99148 31162 13.7% 9.9% 3.1% 1.00 1.00 1.00 Total/Average 83% of the deposited students came from the highest scoring 45% of the inquiry pool. 7% of the deposited students came from the lowest scoring 34% of the inquiry pool Applying predictive modeling technology to your marketing and recruitment program Increase the size of your inquiry pool through more effective mining of your prospect pool (pre-prospect and prospect models) Assign communication channels based on propensity to enroll Strategically created groups Web site E-mail E-newsletters/ communications Direct mail Student calls Professional staff Alumni Faculty Lowest interest Most interested Shape enrollment through targeted communication campaigns Focus admissions travel Applying predictive modeling to your retention efforts We have found that blending a predictive model with data gleaned from a motivation/attitudinal survey produces a powerful data combination ) 8% 26. ( ed a inn ari t-B e V M ) el 5% eed od 12. ( al N %) i M c f (13 an o ned n i d n i F e t -B nn gth ) PA c en -Bi 2% lG en ion Per o r t t o u 12. ( h b i S c r t n me hS Co Hig nc o ) mily ld I a o 1% F h al 10. ( us e o Tot H ed 8% ge inn 26. era it-B ) m Av Ad ) .7% as 2% (10 s g (8. y a a l s F D u r mp ute Ca mm m o Co ) fr 6% ce (6. tan e Dis d o yC 5% unt 12. Co s ble la Re % 6.6 % 8.2 7% 10. e tiv 0% 13. 1% 10. 2% 12. The predictive model provides OBSERVED risk factors While the motivation survey produces ACKOWLEDGED risk factors Risk categories can be used to design both programmatic and student-specific interventions It is critical in this approach that you blend the observed and acknowledged risk factors to create an agenda for action Implementation of this combined approach improved retention rates across entry terms and campuses for this institution Campus Campus 1 Campus 2 Campus 3 Fall 04 Fall 05 Change Spring 05 Spring 06 Change Summer 05 Summer 06 Change 65.5 71.0 5.5 64.7 67.6 2.9 76.0 82.1 6.1 65.1 61.1 -4.0 64.6 68.9 4.3 77.4 72.0 -5.4 61.5 58.9 -2.6 48.3 58.2 9.9 67.1 58.1 -9.0 69.4 74.3 4.9 49.7 57.2 7.5 64.6 12.2 12.8 12.4 72.0 82.2 68.2 71.3 3.1 60.5 76.2 11.9 56.6 67.8 74.2 87.1 3.8 45.5 62.0 11.2 16.5 10.2 15.7 12.9 73.7 78.3 4.6 Campus 4 67.5 66.5 -1.0 48.0 60.2 Campus 5 56.1 58.4 2.3 46.2 59.0 Campus 6 60.2 79.7 19.5 52.2 Campus 7 63.7 72.8 9.1 Campus 8 68.3 80.2 Campus 9 54.9 58.7 Some concluding thoughts Apply modeling to the regions of your funnel that hold the greatest promise for improving your enrollment management outcomes Pre-prospect model Prospect model Inquiry model Applicant/admit model Retention/progression models Identify a resource to develop your institutionspecific models and score your current files Establish project goals and aggressively measure your results…remember the goal is to beat the model! Use the modeling process to improve data collection and data management protocols on your campus…. …while most schools have reasonably good data on student characteristics, the weakness tends to be in tracking student behavior Observations and questions