Personalization with Self-Selection: Experimental Evidence on Social Information and Charitable Giving (Wave 1)

How to cite

Sasaki, Shusaku; Kitano, Shodai; Inoue, Kosuke. 2026. "Personalization with Self-Selection: Experimental Evidence on Social Information and Charitable Giving (Wave 1)." AEA RCT Registry. AEARCTR-0018087. https://doi.org/10.1257/rct.18087-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Shusaku Sasaki* , Shodai Kitano , Kosuke Inoue

* Primary investigator

Registry ID
AEARCTR-0018087
DOI
https://doi.org/10.1257/rct.18087-1.0
First registered
Last updated
Status
On going
Trial dates
2026-03-13 – 2029-03-31
Countries
Japan ()
Registry record
http://www.socialscienceregistry.org/trials/18087

Abstract

Policy targeting with machine learning can personalize interventions at the individual level. Yet when individuals are allowed to accept or decline the assigned treatment, noncompliance arises. The welfare properties of targeting policies under such noncompliance are not well understood. This study empirically evaluates how such self-selection affects social welfare in the context of policy targeting. We apply the framework for targeting under noncompliance proposed by Athey et al. (2025) to a charitable giving setting and evaluate the welfare implications through a two-wave experimental design.

This pre-registration covers Wave 1 of the two-wave experiment. Wave 1 collects data for constructing the machine learning models used for targeting in Wave 2. We conduct a randomized controlled trial (RCT) in a donation experiment to measure how presenting social comparison information and offering participants the option to view it affect charitable giving behavior. Using the Wave 1 data, we will also estimate heterogeneous treatment effects and conduct simulation-based welfare evaluation of policy targeting.

*Athey, S., K. Inoue, and Y. Tsugawa. 2025. "Targeted Treatment Assignment Using Data from Randomized Experiments with Noncompliance." AEA Papers and Proceedings 115: 209–14.

Intervention

Participants are randomly assigned to one of three groups. The initial endowment is 1,000 JPY for all groups.

1. Control Group:No information about others' donation behavior (hereafter, social comparison information) is presented.
2. Forced Disclosure Group:Social comparison information is presented prior to the donation decision. The information consists of the donation rate and the average donation amount, obtained from a pilot study.
3. Self-Selection Group:Participants are given the option to view social comparison information and choose whether to view it before making their donation decision. If they choose to view the information, the same content as in the Forced Disclosure Group is presented. If they choose not to view it, the same donation decision screen as in the Control Group is displayed. The default option is set to “not view.”

Experimental design

In Wave 1, we collect data for constructing the machine learning models used for targeting in Wave 2 and estimate heterogeneity in the effects of presenting social comparison information and granting the right to view it. We administer two online surveys to panelists of an online survey company.

1. Screening Survey (scheduled to begin on March 13, 2026)
The screening survey measures covariates used for constructing the targeting models and estimating treatment effect heterogeneity. The covariates include demographic variables (age, gender, prefecture of residence, education, marital status, household members, occupation, and income), individual characteristics (psychological traits such as reciprocity and social comparison orientation, Big Five, self-image, risk preferences, and loss aversion), donation-related questions (hypothetical donation behavior, beliefs about others' donation amounts, interest in charitable giving, total donations in 2025, and organizations donated to in 2025), and questions on usage of generative AI services. We include two large language model (LLM) bot detection questions to address the concern that respondents may use generative AI to answer surveys.

2. Main Experiment (scheduled to begin on March 19, 2026)
The main experiment implements a randomized controlled trial on charitable giving. Participants are first informed that they have a chance of receiving an additional payment of 1,000 JPY through a lottery (1 in 10 odds). They then report what they would spend the 1,000 JPY on if they were to use it for themselves. After that, they receive an explanation of the donation experiment and complete a comprehension check to confirm their understanding of the incentive mechanism. They then read a description of the donation recipient, the Japan Committee for UNICEF. Participants are randomly assigned to one of the three groups described in the Intervention section. Under their respective treatment conditions, they decide, in increments of 1 JPY, how much to donate if they receive the 1,000 JPY. At the end of the experiment, participants are asked whether they used generative AI to answer the survey. They are assured that their response will not affect their payment.

Design details

Randomization method
Stratified randomization by age, gender, residence in a major city, and intended donation amount measured in the screening survey
Randomization unit
Individual
Clusters
N/A
Observations
9,000 individuals
Arms
3,000 individuals per each group
Primary outcomes
- Donation amount (JPY) - Whether the participant made a donation or not
Secondary outcomes
- Donation amount conditional on donors - Whether the participant chose to view the social comparison information or not (Self-Selection Group only)

Cited by

No citation data available for this registration yet.

Keywords