Testing Narrative Models in the Field: Domestic Violence, Donations, and Cross-Cultural Persuasion

Participants are randomly assigned to one of three video messages about the Solidarity Bus, an NGO initiative in Kyrgyzstan that supports women facing domestic violence through legal assistance and activities aimed at changing gender norms. The videos are approximately four minutes long and are matched in narrator, structure, tone, graphics, music, length, and underlying factual information. All videos describe the NGO’s work since 2016, present information on reported domestic-violence cases in Kyrgyzstan, show cross-country evidence on the association between gender norms and domestic violence, and end with the same donation appeal.
The intervention varies only the interpretive frame used to present the same underlying facts. The Baseline video provides a descriptive informational message about the NGO’s legal-support activities and gender-norm programming, without drawing strong causal conclusions from the evidence. The Easy-Fix video emphasizes the gender-norm component of the NGO’s work and frames changing attitudes toward women’s roles as a relatively simple, scalable, and effective way to reduce domestic violence. The Overfitting video emphasizes the legal-support component and interprets temporal patterns in reported domestic-violence cases as suggestive evidence that expansions and reforms of the NGO’s legal-aid activities have contributed to reductions in reported violence.
In the Kyrgyzstan field experiment, adult women in rural villages first complete a baseline survey and make an incentivized baseline donation decision. They then view one of the three randomly assigned videos and complete a post-video survey including a second incentivized donation decision, questions on their interpretation of the NGO’s work, and a measure of willingness to share a donation appeal. After the individual part of the study, participants are reassigned to small deliberation groups with randomly varied treatment composition. Each participant proposes a group donation before and after a structured discussion, and discussions are recorded for later coding of the arguments used.
In the U.S. online experiment, adult women recruited through Prolific are randomly assigned to view one of the same three video messages. After watching the assigned video, participants make an incentivized donation decision from a bonus payment and answer questions about their interpretation of the NGO’s work. The U.S. study uses the same treatment arms as the Kyrgyzstan field experiment but does not include the group-deliberation component.

2026-09-01

2026-11-30

The primary outcomes are measured after exposure to the assigned video treatment.
In both the Kyrgyzstan field experiment and the U.S. online experiment, the main primary outcomes are:
Individual donation amount. In Kyrgyzstan, this is the amount, from 0 to 400 KGS, that the participant chooses to donate to the Solidarity Bus in the post-video incentivized donation decision. In the U.S. online experiment, this is the amount, from £0 to £3, that the participant chooses to donate from her bonus payment.
Treatment-congruent narrative endorsement. This is an indicator equal to one if the participant selects the interpretation of the Solidarity Bus that corresponds to the narrative emphasized in her assigned treatment, and zero otherwise. For the Overfitting treatment, the congruent interpretation is that the NGO has shown that legal support can reduce violence. For the Easy-Fix treatment, the congruent interpretation is that the NGO has found a relatively quick way to reduce violence by changing gender norms. For the Baseline treatment, the congruent interpretation is that the NGO is making efforts to reduce violence but that results are still uncertain.
In the Kyrgyzstan field experiment, we additionally define the following primary endpoint:
Verified WhatsApp sharing of the donation appeal. This is an indicator equal to one if the participant actually posts the donation appeal as a WhatsApp status and the posting is verified by a facilitator, and zero otherwise.
These are 3 main outcomes for Kyrgyzstan experiment.

For the Kyrgyzstan group-deliberation phase, the primary endpoints are:
Individual proposed group donation before and after discussion. Each participant proposes how much of the group endowment should be donated to the Solidarity Bus before the structured group discussion and again after the discussion.
Narrative content of group discussions. Video-recorded group discussions are transcribed and coded using a pre-specified codebook. The main group-level endpoints are the share of coded statements aligned with each narrative category—Overfitting, Easy-Fix, or Baseline/uncertainty—and indicators for whether a given narrative category is dominant in the discussion.

Most primary outcomes are directly observed rather than index-based.
The individual donation outcome is the amount donated in the post-video incentivized donation decision. In Kyrgyzstan, the variable ranges from 0 to 400 KGS and equals the number of KGS the participant chooses to donate to the Solidarity Bus after watching the assigned video. In the U.S. online experiment, the variable ranges from £0 to £3 and equals the amount donated from the participant’s bonus payment.
Treatment-congruent narrative endorsement is a constructed binary variable. It equals one if the participant selects the interpretation of the Solidarity Bus that corresponds to the narrative emphasized in her assigned treatment, and zero otherwise. For participants assigned to the Overfitting treatment, the variable equals one if they select the statement that the Solidarity Bus has shown that legal support can reduce violence. For participants assigned to the Easy-Fix treatment, it equals one if they select the statement that the Solidarity Bus has found a relatively quick way to reduce violence by changing gender norms. For participants assigned to the Baseline treatment, it equals one if they select the statement that the Solidarity Bus is making efforts to reduce violence but that results are still uncertain.
The WhatsApp diffusion outcome in Kyrgyzstan is a constructed binary variable. It equals one if the participant posts the donation appeal as a WhatsApp status and the posting is verified by a facilitator during the session window. It equals zero otherwise. Participants who express willingness to post but do not have a verified post are coded as zero for this outcome.
The proposed group donation outcomes in Kyrgyzstan are directly observed individual-level variables. Before and after the structured group discussion, each participant states how much of the group endowment she proposes to donate to the Solidarity Bus. The group endowment equals 400 KGS multiplied by the number of participants in the discussion group. We will analyze both the pre-discussion proposal and the post-discussion proposal, with the post-discussion proposal serving as the main group-deliberation endpoint.
The discussion-content outcomes are constructed from transcribed and coded group discussions. The unit of coding is a statement or argument segment, defined as a contiguous stretch of speech expressing one distinct reason, claim, or objection. Each coded statement is classified as Overfitting-aligned, Easy-Fix-aligned, Baseline/uncertainty-aligned, or Other/uncodable, using a pre-specified codebook. At the group level, we construct the share of coded statements in each narrative category. We also construct indicators for whether Overfitting-aligned statements or Easy-Fix-aligned statements form the largest narrative category in the group discussion. Coders will be blind to treatment status, and discussions will be double-coded independently.

The study consists of two parallel randomized experiments evaluating how alternative video messages about an NGO initiative against domestic violence affect support for the organization.

The first experiment is a field experiment with adult women in rural Kyrgyzstan. Community meetings are organized in participating villages. After consent, participants complete a short baseline survey and make an incentivized baseline donation decision. Participants are then individually randomized within village to one of three video messages about the Solidarity Bus, an NGO program that provides legal assistance to women facing domestic violence and conducts activities aimed at changing gender norms. The three videos contain the same underlying factual information and the same donation appeal, but differ in how the evidence is interpreted. After watching the assigned video, participants complete a post-video survey and make a second incentivized donation decision.

The Kyrgyzstan field experiment also includes a group-deliberation component. After the individual post-video survey, participants are reassigned to small discussion groups with randomly varied composition based on the video messages previously viewed by group members. Each participant proposes how much of a group endowment should be donated to the NGO before discussion and again after discussion. Groups then participate in a structured discussion about whether and why to donate. Discussions are recorded and later coded using a pre-specified codebook.

The second experiment is an online experiment with adult women residing in the United States and recruited through Prolific. Participants are individually randomized to one of the same three video messages. After viewing the assigned video, they make an incentivized donation decision from a bonus payment and answer a short survey about their interpretation of the NGO’s work. The U.S. experiment uses the same treatment arms as the Kyrgyzstan individual-level experiment but does not include the group-deliberation component.

The design allows us to compare the effect of alternative interpretations of the same factual information on individual giving, narrative endorsement, diffusion behavior, and, in Kyrgyzstan, collective discussion and proposed group donations.

Not available

In the Kyrgyzstan field experiment, treatment assignment is randomized at the individual level within each village. Upon arrival at the community meeting, eligible consenting participants draw numbered tokens under researcher supervision. The token determines assignment to one of the three video treatments with equal probability: Baseline, Easy-Fix, or Overfitting. Randomization therefore occurs publicly on site within each village.

For the group-deliberation phase in Kyrgyzstan, participants are reassigned after the individual post-treatment survey to discussion groups using a pre-specified randomized assignment procedure. Group composition is randomly varied so that some groups are mostly composed of participants previously assigned to Baseline, some mostly of participants previously assigned to Easy-Fix, and some mostly of participants previously assigned to Overfitting.

In the U.S. online experiment, participants recruited through Prolific are individually randomized by the survey software to one of the three video treatments with equal probability.

individual

Yes

The Kyrgyzstan field experiment will be conducted in approximately 150 villages. Treatment is randomized at the individual level within villages, but standard errors will be clustered at the village level.
The U.S. online experiment has no geographic or institutional clusters; treatment is randomized at the individual participant level.

The planned sample consists of 3,060 adult women in the Kyrgyzstan field experiment and 3,000 adult women in the U.S. online experiment, for a total planned sample of 6,060 individual participants. In Kyrgyzstan, the 3,060 participants will be recruited across approximately 150 rural villages. The group-deliberation phase is expected to generate approximately 300 discussion groups. In the U.S. online experiment, the 3,000 participants will be recruited through Prolific and randomized at the individual level.

In the Kyrgyzstan field experiment, the planned sample is 3,060 adult women randomized with equal probability across three treatment arms:

Baseline: approximately 1,020 participants
Easy-Fix: approximately 1,020 participants
Overfitting: approximately 1,020 participants

These participants will be recruited across approximately 150 rural villages. Treatment is randomized at the individual level within villages.

In the U.S. online experiment, the planned sample is 3,000 adult women randomized with equal probability across the same three treatment arms:

Baseline: approximately 1,000 participants
Easy-Fix: approximately 1,000 participants
Overfitting: approximately 1,000 participants

Overall planned sample by treatment arm across both experiments:

Baseline: approximately 2,020 participants
Easy-Fix: approximately 2,020 participants
Overfitting: approximately 2,020 participants

Using two-sided tests with α = 0.05 and 80% power, and focusing on pairwise comparisons between each narrative treatment and the Baseline, our planned samples provide power to detect effects of approximately 0.124 standard deviations in phase 1 outcomes in Kyrgyzstan (N = 3,060, or 1,020 per arm) and 0.125 standard deviations in the U.S. (N = 3,000, or 1,000 per arm) for standardized continuous outcomes. These calculations assume equal allocation across the three treatment arms and do not account for multiple-hypothesis testing adjustments, which would reduce effective power, nor for the inclusion of baseline outcome controls in the regression specification, which would increase precision under the assumption of positive outcome persistence. In Kyrgyzstan, treatment is randomized at the individual level within villages; we therefore plan to include village fixed effects and cluster standard errors at the village level, which may further affect precision depending on intra-village correlation in outcomes. For the reduced sample used in the analysis of group discussions, effective power is lower due to the smaller number of observations. Since arguments and collective decisions are observed at the group level rather than the individual level, the relevant sample size is determined by the number of groups rather than individuals. With 300 groups in total, the minimum detectable effect is approximately 0.40 standard deviations.