AUDIENCE & BUSINESS
HGO · Audience-to-Donor Strategy
Which Audience Members Could Become Donors?
Working with Houston Grand Opera, we combined attendance, donation, and survey records to study first and repeat giving and inform who to engage and when.
View presentation posterCustomer Records
Before preprocessing
Data Sources
Survey · Ticketing · Donations · Marketing
Features
Background · Attendance · Engagement
After preprocessing, we built task-specific samples for first and repeat donations, combining customer background, attendance, and time-to-event records.
Audience Geography · Analysis & Visualization
Mapping ZIP codes turns many separate customer records into a view of geographic reach. The audience extends across the US, so a Houston-only profile would miss part of it. This view informs regional grouping; attendance and donation records are then needed to evaluate geographic differences.
Why Model When Someone Gives?
Outreach takes time. A new attendee and a long-standing audience member may need different engagement schedules. Survival analysis keeps the waiting time and can use records where no donation has been observed yet. Cox offers interpretable feature associations; Random Survival Forest tests more complex combinations of features.
| Task and Model | Training C-index | Evaluation C-index |
|---|---|---|
| First donation · Cox, full feature set | 0.731 | 0.717 |
| First donation · Random Survival Forest | 0.800 | 0.733 |
| Second donation · Cox, before feature selection | 0.603 | 0.581 |
The study used an 80/20 training–evaluation split. These are exploratory results: the evaluation set also informed feature and parameter choices, so it was not an untouched final validation set.
Reading the Results
C-index: Who Donates First?
If A donates before B and the model puts A first, that pair is correctly ordered. A score of 0.733 is roughly 73 correct out of 100 comparable pairs; 0.717 is about 72, and 0.581 about 58. A score of 0.5 is near random ordering—not a customer’s donation probability.
Cox: Which Features Relate to Earlier Giving?
The full first-donation model gives annual attendance a hazard ratio (HR) of 1.11. For 4 versus 3 annual visits, holding other features equal and before either person donates, the instantaneous donation rate is about 11% higher—not an 11-percentage-point probability increase.
Kaplan–Meier: Who Is Still Waiting?
It estimates a waiting curve from observed records. If 10 people are followed for a full year and 3 donate, the curve ends at 70%. Someone observed for only six months still contributes those six months; they are not treated as someone who will never donate.
Think of the curve as “how many are still waiting to give?” A height of 70% at year 2 is like expecting about 70 out of 100 similar customers still to be waiting, with about 30 having given. A faster drop means an earlier expected donation.
Audience Profile from Analysis and Model Evaluation
A Profile Summarized by the Project
- Age55+
- EducationBachelor’s or higher
- Household Income$125k+ / year
- LocationHouston
- RelationshipSubscriber
- EngagementHigh satisfaction & recommendation intent
A profile drawn from audience analysis and model findings to understand differences and discuss outreach.
What Do These Features Help Us Ask?
- Depth of engagement
Read attendance alongside subscription status to distinguish occasional visits from sustained involvement.
- Quality of the relationship
Use satisfaction and recommendation intent as separate features: enjoying an experience and recommending it are different signals.
- Timing of engagement
Study the wait for first and repeat donations to discuss when to build a relationship and when to follow up.
Engagement features combine attendance records with survey responses: annual visits, satisfaction, and recommendation intent are linked to each customer and analyzed alongside donation timing. The NPS-style recommendation response is an individual input, adding relationship quality to the frequency of attendance.
EXPLORE THE OUTPUT
Interactive Demo
Set a prospective donor’s profile, then compare one change at a time. The examples explain how customer relationships and the time window affect the simulated output.
University–industry collaboration: source data and models are private. This demo visualizes the project’s output format using illustrative rules and example values.
Fields reflect the project’s research and audience profile. Age increases in small steps in this demo only; the Cox results did not establish a positive age effect. Region bands are illustrative: ≤50, >50–150, and >150 km.
How the Time Window Changes the View
More time creates more opportunity for a first donation. The example keeps this profile unchanged.
Compare One Change
Why Does the Number of Years Matter?
Five illustrative propensity levels: very low <15%, low 15–<30%, medium 30–<50%, high 50–<70%, very high ≥70%. Comparisons explain the simulation’s rules, not measured causal effects.
