Transcript Slide 1
Modeling and Money: The Two DO Mix TAIR February 1, 2006 Baylor University Located in Waco Texas Affiliated with Baptist General Convention of Texas Bachelors/Masters/Doctoral degrees Seminary Mdiv and Dmin Law Fall Enrollment approximately 14,000 Nuggets “If you’ve got terabytes of data, and you’re relying on data mining to find interesting things in there for you, you’ve lost before you’ve even begun.” — Herb Edelstein Predictive Modeling at BU Enrollment Management Inquiry to Net Deposit Accept to Enroll – Applications of model • Moving from one stage to another • Classification of students-new freshmen, new transfers, graduate, etc. • Texas and non-Texas students Enrollment Management Stages Inquiry Applied Accepted Deposit Net Deposit Enroll Retention Graduation Student Retention – Applications -- • Fall to Spring Retention • Fall to Fall Retention • Enroll to Graduation Donor Management – Annual Gift – Major Donor – Planned Gift – Retention/Upgrade – New Donors Business Questions How can we identify potential major donors? How can we predict propensity of a donor to make an annual gift? How can we identify potential planned giving donors? How can we identify current donors that can move to next level of giving? How can we identify non-donor constituents with characteristics of a donor? How can we predict expected value of a gift? Required Expertise – Domain – Data – Analytical Methods Project Team Representatives from University Development Representatives from Institutional Research SAS Consultants Process/Steps Explore Development data Build datasets for descriptive models Validate datasets Create profiles for analysis Build datasets for predictive modeling/mining Mine the data Create predictive models Apply the models Test the models Data Exploration New database for IR –Learn and learn more! –Edit reports and data cleansing Profiles Donor Non-donor Alumnae donor Hispanic donor African-American donor – More data cleansing! Indicator Score Creation of indicator variables with yes/no (1/0) values For Single households -- 18 indicators For Two-person households -- 25 indicators (7 indicators could be duplicated) Indicator Variables DOB_50_ind – over 50 years of age? Married-Widowed_ind - married or widowed? Children_ind – any info on children? Alumni_ind – an alumni? Contact_ind – any contact info for donor? Executive_ind – executive job code? Leader_ind –Baylor relationship? gift count – has donor made 15 gifts over lifetime? gift_5k – total cum gifts >= $5,000? gift_25k – total cum gifts >= $25,000? gift_100k – total cum gifts >= $100,000? year5_ind – has donor made $250 gift in EACH of last 5 years? year2_ind – has donor made ANY gift in EITHER of last 2 years? Rating_ind – does donor have Echelon rating? Athletic_gift_ind – has donor made gift to Athletic Department? Alumn_assoc_ind – has donor made gift to Alumni Association? Spouse_alum_ind - is spouse coded an alum? Indicator Score Distribution Average Cumulative Gift Donor Household Profile 64,000+ Households 72% One donor in household 50% Alumni 60% Males 57% Married 19% indicate Baptist religion 58% indicate Texas residences Non-Donor Household Profile 77,000+ Households Most data fields have a large percent of missing values Donor Model for 2004 Use donors for previous 10 years Create target variable Identify predictor variables Build model Apply to 2005 donors Categories of Predictors Biographical/demographic - 20 Contact information - 12 Degree data – 9 Activities - 15 Gift information - 31 External rating information - 5 Research data - 4 Building Model Target variable – gift in 2004 – 1 for household with 2004 donation – 0 for household with no donation in 2004 Predictors constructed from donors in 1994-2003 time period Tools -- SAS Enterprise Miner – Used to build, validate, and score ROC Model Comparison Lift Chart Distribution of Scores Model Comparisons ROC curves and Lift charts indicate all models are performing well Misclassification rates for the models are all close to 16% Very little difference between average profit for the models Logistic regression was chosen as the model to employ Model Application Analyze 2004 donors at the end of June 2005 Determine those who have not made a donation Use probability scores to target those most likely to make a gift Future Work Application of general model – Annual gifts – Major gifts – Planned gifts Non-donor model Gift amount model Life time value model Thanks! Questions or Comments