The following study was conducted in conjunction with a client of American Philanthropic and all relevant organization and donor information has been anonymized.
This study was born out of an attempt to better understand an organization’s active donor base in order to identify donors that were most likely to increase their giving—and whether, therefore, predictive modeling could be used to make fundraising efforts more effective in identifying donors likely to upgrade.
Not only do nonprofits want to retain their donor base, they want to upgrade their donors where possible. Upgrading a donor often requires investing in a personal relationship with them, but this takes time. After 30 years of operation, one human services nonprofit has reached over 20,000 donors annually. At this scale, they have so many donors that it's impossible to reach each one personally. A good problem to have!
To address this problem, they sought to prioritize their donor list with the use of a “predictive model.” The goal of the model was to predict which donors who gave less than $1,000 in 2019 would upgrade to the $1,000 level in 2020. Multiple variables regarding donor giving history and wealth indicators were assessed using a random forest algorithm, and then a predictive model was built based on those findings. The model was applied to the subset of the 2019 donor file that gave under $1000, and each was assigned a score from 0-10, with 10 being the most likely to upgrade.
During CY2020, the gift officers were made aware of these scores, and each individual gift officer chose to rely on these scores to differing degrees. Overall, CY2020 carried on without any major concerted efforts to target donors based on these scores. At the conclusion of 2020, the upgrade scores were compared to the 2020 giving behavior of the scored donors to test for accuracy.
The results showed that the donors who upgraded to the $1,000 level in 2020 had a higher median upgrade score than all other donors, as was expected. In the future, gift officers can use the model to making a more concerted effort to prioritize donor communications based on the modeling scores in CY2021 is expected to result in an even greater showing of donor upgrades in CY 2021.
Donor retention is an uphill battle for most nonprofits, and many may not even have time to think about what comes next. But once a nonprofit gets into a successful annual rhythm, it is important to avoid treading water and to identify ways to improve fundraising efforts. With a solid base of reliable donors in the file, the next step is to encourage loyal donors to increase their giving. This effort, however, often requires a major commitment from the nonprofit: while some donors may offer an especially generous gift unprompted, most donors will increase their giving after growing a personal relationship with someone on staff.
This relationship-building takes time, and no development department has unlimited time. This means that tough decisions must be made to prioritize time with certain donors over others, which makes us wonder which donors should be prioritized?
How should it be decided which donors get assigned to gift officers or receive other special attention, especially if they have not yet reached a set “major giving” level? Is there a way to know which donors have more potential to give more or more frequently than others?
Creating a predictive model is one way to determine which donors have the most potential to become major donors. For the duration of this paper, major donors will be defined as those who give $1,000 or more within a given calendar year.
A human services nonprofit with over 20,000 donors annually sought to use a predictive model to answer this question—that is, to identify which donors should be prioritized for more personal attention from gift officers. The predictive model they used read in several variables from each donor’s giving history, and then generated a score for each donor on a scale from 0-10. Only donors below the $1,000 giving level were scored, and those with higher scores were predicted by the model as being more likely to upgrade to major donor status than lower-scored donors. In this way, an entire donor file was ordered according to “likelihood to upgrade.”
The goal of the model was to predict which donors who gave less than $1,000 in 2019 would upgrade to the major donor level in 2020. All individual donors were exported from the database, along with several other pieces of data.
Data pulled from the database for the model included:
Once this data was obtained from the database, a new variable, the “Donor Value Score,” was calculated. To get this score, a proprietary formula factors in various data points—such as First Gift Date, Last Gift Date, Largest Gift Amount, and Estimated Capacity—giving each of them a different weight in the formula in order to determine which donors are ready to upgrade.
The organization’s donors were sorted into four groups: new donors in 2019, non-new donors who had given below $1,000 in both 2018 and 2019, non-new donors who gave below $1,000 in 2018 and $1,000 or more in 2019, and donors who gave $1,000 or more in 2018. These basic groupings were used to build the model donor population. Using a random forest algorithm, the second and third groups were contrasted, and the model was built off these differences.
This model was then applied to the predictive model population of all donors who gave less than $1,000 in 2019. Each donor in this population received a score from 0-10, with 10 being most likely to upgrade to the major donor level in 2020.
Finally, the scores were shared with the development staff and entered into the organization’s database for easy reference. It was up to each person’s discretion how they chose to employ the scores in their strategy for the year. There was no unified effort to use the scores to prioritize certain donors over others in 2020.
Once 2020 had drawn to a close and all 2020 gifts were properly attributed, the total giving in 2020 was tallied for each donor. The donors that had been assigned 2020 upgrade propensity scores were then pulled from the database, along with the new total 2020 gift amount. The donors were split into four groups based on their total 2019 and 2020 gift amounts.
Those who gave at least $1,000 in 2020 were considered the “Success” group because that is what the model aimed to predict. The remainder were split among whether they gave less in 2020 than in 2019 (“Downgrade”); whether they gave more in 2020 than in 2019, though not quite $1,000 (“Not Quite”); and those who gave the same amount in both years (“Same”).
The Success group had a significantly higher median score in the predictive model than the other three groups, whose distributions were all about equivalent.
The main finding that the group of upgrading donors had higher upgrade propensity scores indicates that this predictive model was accurate in identifying donors that were more likely to upgrade. That the other groups had similar score distributions also indicates that the model worked well to answer this particular question and that it did not seek to simply answer which donors were likely to give more in 2020 than in 2019. In fact, the score distribution of the three non-Success groups also reflected the broader overall distribution of all scores.
Another important element of this finding is that the development team at this organization did not make any major efforts to upgrade the donors with higher scores, and yet the donors that upgraded were scored more highly by this model than those who did not. This might indicate that this model highlighted donors who were likely to upgrade, even without much prompting from development staff.
A secondary analysis of this data required obtaining other new information from the database, including whether an ask date field had been filled in for each of these donors in 2020. This time, donors who were directly asked by a staff member of the organization to make a gift in 2020 were compared to donors who were not directly asked to make a gift in 2020. Except at the very highest upgrade propensity scores, the donors who were asked to make a gift ended up upgrading to the major donor level 2 to 3 times as often as donors who were not asked to make a gift.
In future years, these scores can be used when writing donors plans, so that highly scored donors can be nudged or given the opportunity to become major donors, and lower-scored donors with other strong characteristics of major donors can be assigned a new gift officer to encourage them to become major donors. This model, then, can provide fundraising leaders the information they need to invest time and resources effectively.
It is important to note that there were very low-scoring donors who upgraded to the major donor level. That is okay and not indicative of a poor model. Most donors were assigned relatively low scores. A low score did not necessarily indicate that the low-scoring donor would not upgrade and the high-scoring donor would. Instead, a more correct interpretation is that a donor with a higher score was more likely to upgrade than a donor with a lower score.
In this case, the predictive model scored donors who were more likely to upgrade higher than those less likely. It can be worth repeating this model every year, incorporating more variables as their collection becomes more reliable, to inform strategy and to increase number of upgrades each year. Using this predictive model can help fundraisers and their supervisors invest their time in the most efficient way possible.