Is Your Fundraising Mail Guided by a Model No One Can Explain?

In one major annual fundraising campaign, donations fell significantly. The offer, creative, and timing had not changed enough to explain the decline. The campaign drew prospects from more than 75 rented donor and response files, many of which had performed predictably for years.

No one could identify the cause.

It took several weeks of analysis to discover that one important control list had changed the offer used to generate its names. The list owner had previously attracted people through an offer requiring a meaningful expression of interest. It replaced that offer with a contest.

The names were still classified as responders. But entering a contest required less interest and commitment than responding to the earlier offer. The label on the file had not changed. The behavior behind it had.

That one change weakened the annual campaign in two ways. The organization wasted acquisition dollars mailing a control list whose past performance could no longer be trusted. It also lost a major source of prospective donors from its available universe.

The lesson extends far beyond one list:

Past response is predictive only while the method that produced the names continues to generate consistent results.

Today, large fundraisers increasingly rely on propensity-modeled cooperative audiences and traditional rented donor files. That evolution has expanded the available universe, improved segmentation, and made multichannel targeting more practical. But it has not removed the risk revealed by this campaign. It has moved the risk deeper into the data and the model.

Management may now approve millions of acquisition contacts without knowing whose behavior created the prediction, what outcome the model was designed to estimate, or whether its inputs still represent the current fundraising environment.

Fundraising Acquisition Lists Have Changed

For decades, large direct-mail fundraising plans were built primarily from rented response files. These files contained people who had donated, purchased, subscribed, joined, inquired, or taken another measurable action.

That history mattered. Someone who had already responded to another organization's offer was generally more likely to respond again than someone selected solely on the basis of age, income, geography, or other demographic characteristics.

But traditional list plans were difficult to manage at scale. An annual program mailing millions of names might require dozens—or even hundreds—of separate files. Each could carry its own prices, selection charges, one-time-use terms, reuse rules, offer approvals, creative restrictions, drop-date requirements, and competitive exclusions.

Public list-rental pages still show this structure. The Upper Room Masterfile, for example, publishes a base price of $105 per thousand, with additional charges for recency, geography, and other selections. Reuse requires separate approval. The Direct Marketing Fundraisers Association also describes list rental as the one-time use of names and explains that list owners may restrict recent donors, competing organizations, sweepstakes offers, and other users. Published rates may differ from the negotiated prices paid by large mailers, but these public terms illustrate the cost and operational limits of building acquisition volume one rented file at a time.

Fundraising cooperatives changed that structure. Participating nonprofits contribute donor transactions and related promotional data to a much larger shared environment. The provider combines that history with other demographic, behavioral, and commercial information, then develops predictive models to rank prospective donors for a particular organization or appeal.

This is not a simple contest between “response data” and “compiled data.” Modern fundraising cooperatives are hybrids. Actual donation behavior supplies much of the foundation. Modeling, identity resolution, and additional attributes determine which prospects appear most likely to respond, give more, or develop greater long-term value.

Data Axle reports that its DonorBase cooperative contains more than 83 million donors and 1.2 billion individual donations. Moore describes donor cooperatives as combining giving and promotional histories with added transactional, demographic, attitudinal, and behavioral data. These are provider descriptions, not independent performance guarantees. They do show how far fundraising acquisition has moved beyond the rental of individual donor lists.

Traditional rented donor files provide traceable response evidence; donor cooperatives add scale and ranked modeling. Both must prove performance through controlled donor economics.

The Model Can Be Right About the Wrong Result

A propensity model is designed to estimate an outcome. That sounds obvious, but the word “propensity” is often treated as if it defines the outcome by itself.

Propensity to do what?

To open an envelope? Visit a page? Enter a contest? Make a first donation? Give above a certain amount? Renew? Become a profitable long-term donor?

Those are not interchangeable results.

During an earlier engagement involving multivariate testing, I repeatedly asked what data had been used in the propensity model. Was it built from customers, qualified leads, or all responders? If leads were used, how recent were they? Which offers generated them? Were cancellations, weak responders, low-value customers, or other exclusions removed? What population represented the nonresponders? When was the model last updated?

Neither my boss nor the head of statistics could answer those questions.

That did not prove the statistical work was defective. It revealed something more fundamental: the people applying the model could not explain what it meant. They could not determine whether it supported the client’s current business objective, whether its exclusions made sense, or whether market changes had weakened its predictive value.

A model can be statistically sound yet strategically wrong if it was directed to the wrong people, the wrong outcome, or the wrong period.

If a fundraising model is trained to maximize initial response rates, it may identify many low-dollar donors who never give again. If it is trained on all donors without distinguishing recent from stale behavior, it may preserve relationships that no longer predict current giving. If it includes contest entrants, unqualified leads, or records generated by an outdated offer, it may identify more people who resemble a population the organization should no longer pursue.

The model does not correct a weak strategic definition. It executes it at scale.

Marketing Must Curate the Inputs

Most of the major weaknesses attributed to propensity modeling are not inherent weaknesses in the method. They are failures of discipline before and after the statistical work.

Marketing must define the business result before the model is built. It must also understand enough about the input population to determine whether that population represents the result it wants repeated.

At a minimum, management should require clear answers to these questions:

• What fundraising result is the model designed to estimate: initial response, gift amount, renewal, net revenue, long-term donor value, or another outcome?

• Was it built from first-time gifts, repeat gifts, total revenue, donor value, or another result?

• Which campaigns, offers, channels, and time periods produced the input population?

• Who was excluded, and why?

• Were refunds, cancellations, low-value donors, unusual events, or unrepresentative campaigns removed?

• What non-donor comparison population was used?

• Which outside attributes were added, and are they current, lawful, and relevant?

• How often are the input data refreshed and the model recalibrated?

• How has each ranked segment performed against actual donations, net revenue, renewal, and subsequent donor value?

The answers do not require executives to become statisticians. They require the organization to retain control of the business question.

This is especially important when conditions change. Giving behavior during a national crisis, a political event, a natural disaster, or a heavily promoted premium campaign may not represent normal donor behavior. A model trained during an unusual period may remain statistically coherent while becoming less useful as the environment changes.

Good input curation is therefore not a one-time technical step. It is an ongoing marketing responsibility. The seed population, exclusions, variables, and outcome must be reviewed as the offer, donor base, competition, and economy evolve.

This Is a Targeting World, Not a Digital World

The growth of digital media did not make prospect data less important. It made dependable targeting more valuable.

A strong fundraising cooperative can create a ranked audience that supports direct mail, email, and digital activation. Data Axle, for example, says its cooperative audiences can be activated through mail and digital channels and that some email addresses are linked directly to donation history. That is different from purchasing a generic digital audience based only on platform behavior or broad inferred interests.

Digital platforms can find audiences without a donor file. But the platform generally controls identity matching, audience definitions, delivery, and reporting. A named or identity-resolved prospect universe provides the fundraiser with a more consistent starting point for deciding whom to reach and how to contact those prospects across channels.

This does not mean every channel can be used without restriction. Postal-list rights do not automatically provide permission for email, telephone, text, or platform matching. Each channel introduces its own licensing, privacy, consent, and suppression rules.

Mail and digital do not compete for the right to define the prospect. Better targeting increases the value of both.

The audience should come first. Channels should be assigned according to the evidence, economics, and rights available—not because the organization has declared itself “digital-first” or “mail-first.”

Testing Must Decide How Deeply to Roll Out

A modeled cooperative universe need not be accepted or rejected as a single block. It can be ranked from the highest to the lowest estimated propensity, then divided into quintiles, deciles, or smaller scoring bands.

That gives marketing a disciplined way to determine rollout depth. The strongest segment may justify a larger rollout quantity and coordinated digital support. The next segment may support mail alone. Lower-ranked segments may be tested at smaller quantities or withheld entirely.

But the stopping point should not be determined solely by response rate.

Fundraising needs to track the progression from prospect source to response, completed donation, average gift, net revenue, subsequent gifts, and long-term donor value. A segment with a high response rate may attract low-dollar donors who do not renew. Another segment may respond less often but produce greater revenue and stronger retention. The modeling plan should decide which of these results matters before the model is built—and testing should verify all of them after deployment.

The practical governing measure is the allowable acquisition cost: the highest cost the organization can accept for the kind of donor it is acquiring, based on its financial objective and recovery period.

Marketing should recalculate the cumulative, blended cost per donor each time another quintile is added. A fourth quintile may exceed the allowable on its own yet remain acceptable if the combined results of the first four remain within the allowable. Rollout should stop only when adding the next quintile causes the cumulative cost per donor for the entire selected universe to exceed the allowable—unless leadership has deliberately approved a longer payback or a different objective.

Traditional rented files can be tested against the modeled cooperative universe—and often should be. Some rented donor lists may retain a strong cause affinity, gift pattern, or recent response advantage that the broader model does not capture. They may also add valuable names outside the cooperative’s best segments.

The fair test must control the offer, creative, mailing window, geography, merge-purge treatment, suppression rules, source coding, and response period. Overlap must also be handled carefully. If the cooperative receives favorable merge priority, it may claim obvious responders that also appear on rented lists, making the comparison look stronger than it is.

The question is not whether modeling has made rented donor files obsolete. It is which source—or combination—produces the required donors within the allowable cost.

Management Must Control the Prediction

Propensity modeling has made fundraising acquisition more scalable, more segmented, and more useful across channels. It has also made weak assumptions harder to see.

With traditional rented files, management could trace a list to a particular organization or offer, select donors by gift amount or recency, and evaluate that source separately. A cooperative model combines donation histories from many organizations with far more transactions and outside attributes. That broader evidence may improve the estimate, but it also makes the basis of any individual prospect score harder for management to see and question.

Management does not need every coefficient, variable weight, or proprietary formula. It does need to know what behavior created the model, what fundraising result the model is intended to estimate, what data and exclusions shaped it, and how accurately controlled market results have supported that estimate.

That is why transparency about model construction is not an optional technical courtesy. It is part of acquisition governance.

Before approving a major campaign, leadership should be able to explain:

1. The fundraising outcome the model is designed to estimate—and how that estimate has been validated.

2. The donor behavior and time period used to build the model.

3. The exclusions and unusual conditions removed from the input.

4. The performance of each ranked segment against actual donor economics.

5. The deepest rollout point at which the cumulative results of all selected segments remain within the allowable acquisition cost.

The model may be proprietary. Its strategic purpose cannot be.

Before trusting a propensity model, can management explain whose behavior created it, what result it predicts, who was excluded, and when it was last validated against actual economic outcomes?

 If the answer is no, the organization does not control its targeting strategy. It is simply financing a prediction it does not understand.

References

• Data Axle Nonprofit: DonorBase — Provider description of a large fundraising cooperative built from donation activity and additional demographic, behavioral, and commercial data, with mail and digital activation.     

• Moore: What Is a Donor Co-op and Why Should You Use One? — A useful explanation of cooperative inputs, modeling, testing, data hygiene, frequent updates, and the importance of nonprofit-specific donor behavior. Performance statements are vendor claims.

• Data Axle Nonprofit and TrueSense acquisition case study — A provider case study describing year-long, head-to-head testing of nonprofit cooperatives across 39 campaigns. It does not compare cooperatives directly with traditional rented donor files.

• Statlistics: Upper Room Masterfile — Public response-file listing illustrating base pricing, selection charges, approval requirements, and separately approved reuse.

• Direct Marketing Fundraisers Association: List Strategy for Acquisition — Industry presentation explaining one-time list rental, reuse, exchanges, donor-recency selections, list-owner restrictions, and acquisition planning.