Token Impact: A Two-Condition Giving Experiment

Fundraising advice treats framing as the lever. Make the donor the hero of the story and the giving follows. It is a claim that is easy to state, easy to believe, and rarely tested on money the donor had to work for.

Most giving experiments hand participants an endowment. Money that arrives free is given away more easily than money that was earned, and a framing effect measured on a windfall may not survive contact with a wage. Testing the claim properly needs three things in one place: a task that earns the tokens, an appeal that varies, and a decision recorded to the second.

An application that runs the whole study without me in the room.

On first login a participant is assigned at random to one of two conditions and keeps that assignment for every session afterwards. The heroism condition sees a superhero appeal and the line How many of your tokens will you use to be a hero? The plain condition sees a neutral image and How many of your tokens would you like to contribute? Everything else is identical and sits in the same place on screen.

Consent comes first, on its own page, with a checkbox. Then a tutorial, then the appeal image, then the game: a memory-match grid cut from the appeal image itself. Solving it earns 100 tokens plus one token for every second left on a five-minute clock, so speed pays and the endowment is genuinely unequal.

Then the decision. A slider from zero to everything they earned — not preset percentages, which push people toward round numbers. Below it, a progress bar showing the collective total against a goal of 10,000 tokens.

Every session records the condition, the time taken, the tokens earned, the tokens given and kept, how long the decision itself took, and the device. The researcher’s view holds the per-condition breakdown and exports the lot as CSV. It is the same kind of apparatus as Research Lab, built narrower and for one question.

Sixty-nine participants have been through it across 109 completed sessions: 30 in the heroism condition playing 58 times between them, and 39 in the plain one playing 51.

The heroism group gave away 55.6 per cent of what they earned, on average, against 62.7 per cent for the plain group. The framing that was supposed to raise giving is, in this sample, associated with seven percentage points less of it.

The averages hide the more interesting pattern. The median donation in both conditions is exactly 50 per cent. The gap lives entirely in the tail: 11 of 58 heroism sessions ended in a donation of zero, against 1 of 51 plain sessions. Outright refusal was close to ten times more common under the appeal built to inspire. Heroism participants also took longer to decide — 28.3 seconds against 23.1.

None of this is a result. No hypothesis was registered, no test was run, the participants are a convenience sample, and a difference of this size on a sample of this size is well inside what chance produces. The sessions are not independent either: the heroism group averaged almost two plays each against the plain group’s one and a third, so the arm that gave less is also the arm whose repeat players weigh heaviest in the average. What the application has established is that the apparatus works and the data comes out clean enough to analyse. The finding, if there is one, needs a pre-registered study with a power calculation behind it.

One defect matters more than the headline numbers. The global counter that drove the progress bar had drifted badly out of step with the session records — it was reporting a collective total of about 5,500 tokens when the sessions themselves summed to more than 21,000. Participants saw a goal that looked three-quarters unmet when it had in fact been passed twice over. The counter is now computed from the session records every time it is shown, rather than kept as a running total of its own. The goal itself stays at 10,000, and the application does not tell participants that a real donation follows from reaching it. Tokens here are an experimental currency, not a pledge.

Anything on the screen is part of the treatment. The progress bar was built as decoration, a bit of warmth before the decision. It is nothing of the sort: a running total of what other people have given is social proof, shown at the exact moment of choosing, and a counter that drifts is a stimulus that drifts. I would not have called it a confound while I was building it. It is one.

Freeze the instrument before you collect. The earning task was rebuilt three times — a 3×3 sliding puzzle, then 2×2, then 2×3, then abandoned for memory-match — each time because it felt too hard or too easy. Every one of those changes moves the difficulty, the time taken and therefore the size of the endowment. Tuning a task is development work. Doing it while sessions are arriving quietly turns one experiment into several.

The mean was the wrong statistic. Read the averages and the story is that heroism dampens giving a little. Read the distribution and the story is that heroism splits the room: the same median, far more refusals. Those are different mechanisms and they suggest different follow-ups. The dashboard showed me the mean first because means are what dashboards show.