Is sharing caring? Disentangling prosociality from virtue signalling

Interventions manipulating the action set and the observability in a 2x3 design.

2026-03-30

2026-05-31

Amount donated (GBP). Donation amount chosen from {0, £2.50, £5.00, £7.50, £10.00}.
Decision to donate (binary). Indicator equal to 1 if donation amount > 0, else 0.

Donation amount is recorded as the participant’s selected transfer to the chosen charity (in GBP). Primary analyses use (i) the unconditional amount, including zeros, and (ii) a two-part decomposition into extensive and intensive margins as pre-specified.

Donate indicator equals 1 if the selected donation is strictly positive, 0 otherwise.

Message sent (binary). In the Message condition, indicator equal to 1 if a supportive message is selected, 0 otherwise.
Message content choice. Categorical indicator for which pre-written message is selected (within cause).
Beliefs about peer donations. For Gen 0, this is belief about the amound donated from Gen 1 participants who observe thir decision. For Gen 1, this is belief about other Gen 1 participants. Both beliefs are incentivized with a small bonus for accuracy.
Mechanism measures (post-experiment). Agreement scales on motives related to influencing others, concern for generous image, and learning from predecessors.
Attitudinal covariates and demographics. Attitudes on charity trust and giving, frequency and typical annual giving, and standard demographics.

We run an online randomized experiment with a factorial structure at the charity decision screen. Participants are randomly assigned to one of six conditions formed by crossing:
1. Action set: Baseline (donation only) vs Message (donation plus the option to select a supportive message)
2. Observability: Private vs Observed Before vs Observed After.

In Observed-Before, participants may observe a predecessor’s decision before choosing, and their own decision may be shown to later participants before those participants choose, enabling analysis of direct and indirect social transmission across generations. In Observed-After, information (if any) is revealed only after the participant decides, separating pure observability (image) from strategic influence. In Private, decisions are not shown to anyone.

Primary outcomes are donation participation (donate indicator) and donation amount.

Not available

Computer-generated randomization implemented within the online survey (simple random assignment with fixed assignment probabilities across treatment cells). Allocation is performed automatically at the individual’s entry to the survey.

Primary randomization is at the individual participant level into the 2 × 3 design cells. In conditions involving observability, participants are linked into chains for information display (generational structure). Participants in later generations are hence randomly allocated a specific generational sequence. For later analysis standard errors are clustered at the chain level in analyses that use these links.

No

8,000 participants organised in 2,000 chains. Each chain has 1 Gen 0 participant and 3 Gen 1 participants.

8,000 donation decisions, and 4,000 message decisions.

Each treatment arm has 2,000 participants organised in 500 chains. Each chain has 1 Gen 0 participant and 3 Gen 1 participants.

The main hypotheses we are planning to test are whether the availability of the message option crowds out donations. We are planning to test this primary hypotheses by non-parametric and parametric tests comparing across the action space arm (either within a single observability cell or aggregating across them). Secondary hypotheses are (i) about the moderating effect of observability and (ii) about the direct effect of observed actions on later generations. Given our budget, the target sample is approximately 8,000 participants, i.e. 2,000 per treatment arm. Power calculations focus on pairwise “corner” comparisons between two Gen 0 treatment cells (500 vs 500) using two-sided tests at α=0.05. Based on expected behaviour in our setting (from past literature), we assume an unconditional mean donation of £3.50, a donation rate of 70%, and a standard deviation of donations not exceeding £2.00 on the discrete £0–£10 donation grid. Under these assumptions, we have approximately 80% power to detect a 10% change in mean donations (a drop from £3.50 to £3.15) in any single cell-to-cell comparison. For the extensive margin, with a baseline donation rate of 70%, we have approximately 80% power to detect a 7.8% relative change in donation participation (0.70 vs 0.62). Pooled main-effect comparisons (e.g., Message vs No-message across observability) are substantially higher powered for the same effect sizes.