A fundraising model produced its gains among past donors,
not from rich data about them

A fundraiser for the Catholic Church operating in a German urban area randomly assigned almost 20,000 people to receive a solicitation with a small unconditional gift enclosed, or to a control group. Among the 2,354 who had donated at least once before, machine-learning targeting increased net donations by 13.8% against a benchmark in which everybody received the gift. Among the 17,425 who had been asked before and never given, it did not increase net donations. The study, by Tobias Cagala, Ulrich Glogowsky, Johannes Rincke and Anthony Strittmatter, ran the field experiment in 2014 and repeated it in 2015. Those figures are from the 2014 campaign, at one charity, using one fundraising instrument.¹

The paper is specific about what the model used. It reports that socioeconomic characteristics were "of very limited use" for machine-learning-based optimal targeting, and that its baseline targeting rule did not use them at all. A rule built from donation history combined with geospatial information performed as well as a rule using all the available data. The authors go further, stating that "knowledge about who is in the warm list and publicly available geospatial information is sufficient" for the gains to materialise. Detailed donor histories, on this evidence, were largely dispensable.

The gains occurred among people the organisation recorded as past donors. Set against a live CRM, that can be read as asking whether an organisation can assemble a reliable list of its donors at all. The experiment began with such a list.

NonProfit PRO reported that Make-A-Wish America held donor records scattered across spreadsheets, fundraising databases and volunteer management systems, which prevented staff from understanding a supporter's full relationship with the organisation, and that unified profiles later enabled segmentation by giving history and previous engagements. At the same event, Cherian Koshy said that many organisations do not trust their data, do not have time to use it or do not know how to interpret it. Read alongside the experiment, that fragmentation can present a single supporter as a past donor in one record and a stranger in another.

A foundation weighing a predictive fundraising tool inherits four questions its CRM does not answer. The first is whether it can tell that a donation, an event registration, a newsletter subscription and a volunteer shift belong to the same person. The second is what it is permitted to use of what it holds, and for what. The third is whether it records what happened after each ask, since the study's rule was estimated against donation outcomes and net cost and cannot be built without them. The fourth is who has decided what counts as a best donor before any model ranks one.

The Office of the Australian Information Commissioner addresses how health information may be shared. Its position is that health information may be used for the primary purpose it was collected for, and that another purpose requires consent or a limited permitted circumstance. Its guidance for not-for-profits adds that an organisation should only retain personal information where there is an ongoing need to hold it, and that indefinite retention is unlikely to be compliant. Placed beside the experiment, that guidance can be read as pointing the same way. The regulator expects organisations to hold less, and richer data added almost nothing to the targeting. Blackbaud, a fundraising software vendor, states in its own guidance that not all patient data is suitable for fundraising and lists exclusions including adverse outcomes, deaths and legal cases, opt-outs and sensitive areas such as mental health notes and genetic testing. That guidance is written to a United States regulatory framework and is not Australian regulatory guidance.

Jagsi and colleagues surveyed 1,282 American adults and found that 9.9% considered it acceptable for hospital fundraising staff to check the value of patients' homes or other public information to identify who could make a large donation. The same survey measured consent directly. A physician sharing a patient's name with fundraising staff was acceptable to 47.0% of respondents when the patient had given permission, and to 8.5% when the patient had not. The same act on the same data differed only in whether the patient had agreed. That peer-reviewed American research measures attitudes rather than behaviour.

The study is explicit about its objective. It targeted net donations after solicitation costs rather than gross donations, directing effort at individuals whose expected donations exceeded the cost of asking. One reading of that design is that the objective, not the model, determined which supporters the rule ranked highly.

On the cold list, the study reports, targeting those people did no better than sending them nothing in that campaign. It does not establish that people without a giving history cannot be modelled at all. The study did not test individual-level wealth screening and says nothing about whether it would work. In a health setting the Jagsi survey points to a separate limit, since the public found checking a patient's donation capacity broadly unacceptable.

Where a record does not identify who has already given, a predictive model cannot compute that status from its other fields. It ranks whatever identities the record supplies, not the supporters they stand for.