Almost every charity’s pool of donors includes plenty of people who have both the means and the inclination to make a far bigger gift than they ever did in the past. The trick, of course, is to figure out just which people will make the leap.
To that end, Memorial Sloan-Kettering Cancer Center, in New York, has become one of a small but growing number of institutions to embrace a technique known as predictive modeling to help it set priorities and decide which donors deserve the most attention. While the approach requires fairly sophisticated statistical software and a staff member or a consultant who knows how to use it, fund raisers say predictive modeling can be an option for even relatively small organizations in need of a way to sort through records of previous donors.
The software works by figuring out a combination of variables that, taken together, predict which donors are most likely to make a large gift based on the characteristics of people who previously made big gifts.
Memorial Sloan-Kettering got started two years ago by taking the information it had about donors who had made gifts of $50,000 or more and feeding the data through sophisticated statistical software commonly used by scholars and other researchers to create a model—essentially an equation. The center could then use that information to rate how much people who had made small gifts resembled donors who had made large gifts.
“Instead of us determining what a good prospect is, modeling allows behavior to show what makes a good prospect,” says Kate Chamberlin, campaign strategic research director at Memorial Sloan-Kettering.
While the hospital’s fund raisers still also rely on intuition and indicators such as largest previous gift to identify people likely to make big donations, they say predictive modeling has shown strong promise: It has already yielded its first $1-million gift.
The donor of that gift is a good example of the not-so-obvious prospect Memorial Sloan-Kettering hopes to uncover among its 1.5 million contributors. Before making the $1-million gift, he had given a total of $2,861 to the hospital. His largest single donation had been $300.
“We never would have found him without this process,” says Ms. Chamberlin.
Many Possibilities
Among the variables identified by the software as important predictors: the size of the first gift, an executive job title, and a residence in New York or Washington.
In addition, the software gave more weight to donors who asked that correspondence from the hospital go to his or her business address.
So far more than 244 people identified by the software model have been studied by staff members who conduct research on prospective donors and then assigned to a fund raiser charged with building a closer relationship with those donors. Nearly half are considered potential donors of $50,000 or more, with 23 classified as potential donors of $1-million or more.
The most difficult part of predictive modeling to explain to fund raisers at the medical center has been that there is no single model that predicts who is most likely to make a big gift, says Ms. Chamberlin. The software can build an infinite number of models that can be useful in predicting a behavior.
To determine whether she likes a particular model, Ms. Chamberlin looks carefully at the variables on which it is based. While the variables offer useful predictions only when taken in combination, individual variables can point to possible flaws in a model.
As an example, Ms. Chamberlin says that she would question the predictive power of a model that relied heavily on the notes field of the donor’s record being filled in. That fund raisers wrote in the notes field shows that the organization was already in relatively close contact with the donor, she says, which would make her wonder how well that particular model would identify new prospects.
“I don’t want to keep finding the same people over and over again,” she says.
As new gifts of $50,000 or more have come in since Memorial Sloan-Kettering built its initial model, fund raisers have run those donors’ records through the model to see how they would score. That provided a way for the hospital to test the model with real data.
Of the 164 people who made such gifts, three quarters scored in the top 10 percent of scores for everyone in the database—and 63 percent were in the top 2.5 percent.
Memorial Sloan-Kettering rebuilds from scratch the major-gift model it uses to identify prospects each year to make sure it reflects any changes in the types of donors who are making gifts to the hospital.
Ms. Chamberlin says that it’s possible to “overfit a model” by creating an equation that is so specific that it pertains only to the people who have already made a large gift rather than leaving room for variations that help identify future donors.
“A good but slightly flawed model is actually better,” she says.
Gaining in Popularity
Nonprofit organizations’ use of statistical analysis in their fund raising has grown significantly in recent years, most often to identify the donors most likely to respond to requests for large gifts or to direct-mail appeals, says Joshua Birkholz, a fund-raising consultant at Bentz Whaley Flessner, in Minneapolis, and author of Fundraising Analytics: Using Data to Guide Strategy.
He says Prospect-dmm, an e-mail discussion list run by Rob Scott, the head of development services at the Massachusetts Institute of Technology, on data mining and modeling in fund raising, has seen its subscriber list grow from about 40 people when the list got started in 2005 to nearly 600 today.
Currently large universities and medical centers are “leading the charge” on the use of modeling to identify potential major donors, says Mr. Birkholz, who helped Memorial Sloan-Kettering develop its program.
But, he says, efforts in recent years to make statistical software easier to use have made modeling more accessible than many smaller institutions realize.
“If you were able to learn, say, LexisNexis for Development Professionals as a prospect researcher, or even how to do a query on your database, you could easily learn how to build a predictive model,” he says.
Basic statistical software starts at less than $3,000, according to Mr. Birkholz, while more advanced software is closer to $10,000.
Proponents of predictive modeling stress that they are not trying to replace the traditional ways that charities identify potential donors—such as buying data about donors’ assets to determine who is wealthy, or getting referrals from other top donors—but rather to supplement them.
Ms. Chamberlin, of Memorial Sloan-Kettering, says fund raisers at the medical center have become believers in the approach. “That an experienced fund raiser will choose to put these people into their portfolio is one of the clearest indications of the model’s success,” she says.
Other Uses
Both Ms. Chamberlin and Mr. Birkholz say predictive modeling could be used in ways that go well beyond direct mail and major gifts.
For example, Ms. Chamberlin says, charities might be able to use modeling to predict which donors are most likely to stop giving, and then try to prevent that from happening rather than trying to win back people who already have stopped making gifts.
Cellphone companies are already experimenting with that concept, using data about customers who have canceled their plans to predict which of their current customers are most likely to leave—and then the companies reach out to those customers to try to keep them, she says.
Mr. Birkholz thinks the technology might be able to help identify the most successful approaches for securing large gifts, not just the people who are most likely to make them.
Nonprofit groups have started to record more information about the relationship-building process, he says, things like how often a fund raiser meets with the prospective donor, where they have met, and whether the donor has met the organization’s chief executive or taken a tour of the campus.
With enough information, modeling could identify which steps are most important in winning big gifts from different types of donors, says Mr. Birkholz.
“It’s cool stuff, and it’s doable,” he says. “If the data’s there, it’s doable.”