Belief Disagreement and Instability under Stochastic Recall: Quantitative Beliefs

How to cite

Tang, Li; Kuang, Pei; Weber, Michael. 2026. "Belief Disagreement and Instability under Stochastic Recall: Quantitative Beliefs." AEA RCT Registry. AEARCTR-0017805. https://doi.org/10.1257/rct.17805-1.1

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

Li Tang* , Pei Kuang , Michael Weber

* Primary investigator

Registry ID
AEARCTR-0017805
DOI
https://doi.org/10.1257/rct.17805-1.1
First registered
Last updated
Status
In development
Trial dates
2026-02-01 – 2027-01-31
Countries
Private
JEL codes
E31, D84, C83
Registry record
http://www.socialscienceregistry.org/trials/17805

Abstract

We study how imperfect memory shapes belief formation, cross-sectional belief disagreement, and within-individual belief dynamics over time. In a two-part online study, participants are randomly assigned to one of three groups, each of which views the same sequence of numerical data generated from a stationary AR(1) process. After observing the sequence, participants report recall of realized values and estimate the underlying mean of the process, allowing us to isolate the role of imperfect memory in belief formation in a controlled statistical environment. One day later, participants are recontacted to re-elicit beliefs and recall of the previously presented data, enabling us to measure belief disagreement across individuals and belief changes within individuals over time.

Intervention

Participants take part in a two-part online study.

In Part 1, participants are randomly assigned to one of three groups. Participants are randomly assigned at the individual level to one of three experimental conditions with equal probability. Randomization is implemented by the survey software prior to the start of the task and is independent of participant characteristics. The three conditions differ only in the persistence and noisiness of the underlying data-generating process from which participants observe a sequence of 20 numerical realizations.

Group 1 observes a low-persistence, low-noise process and serves as the control condition. Group 2 observes a high-persistence, low-noise process Group 3 observes a high-persistence, high-noise process

Aside from the persistence and noise parameters of the process, all aspects of the experimental protocol—including instructions, number of observations, display format, timing, incentives, and outcome elicitation—are identical across conditions.

Within each group, participants view the same sequence of numerical data generated from a stationary AR(1) process, with observations presented one at a time. After viewing the full sequence, participants complete a set of questions eliciting their recall of realized values and their beliefs about the underlying mean of the process.

In Part 2, conducted one day after Part 1, participants are recontacted and asked a follow-up set of questions eliciting recall of the previously presented data and updated beliefs about the mean of the process.

Experimental design

The study uses a two-part online experimental design. In Part 1, participants are randomly assigned to one of three groups, each of which views the same sequence of numerical data generated from a stationary AR(1) process, with observations presented one at a time. The data-generating process differs across groups. After viewing the data sequence, participants provide assessments related to the process, including estimates of the underlying mean, as well as measures of recall of realized values.

Participants observe 20 numerical realizations from a stationary AR(1) process and displayed sequentially on separate screens. Participants are randomly assigned to one of three groups that vary the persistence and volatility of the underlying process. Group 1 observes a low-persistence, low-noise process and serves as the control condition. Group 2 observes a high-persistence, low-noise process Group 3 observes a high-persistence, high-noise process. In all groups, the long-run mean of the process is fixed at 5 and unknown to participants. After viewing the sequence, participants are directed to some unrelated tasks. Afterward, they report their best estimate of the series average, with accuracy incentives, and then complete a free recall task in which they list any values they remember freely. This design mirrors the theoretical environment in which agents observe a common realized history but form beliefs based on stochastic recall.

Part 2 of the study is conducted approximately one day after Part 1. Participants are recontacted and asked a follow-up set of questions related to the same data sequence, including repeated belief elicitation about the mean of the process and recall of the previously presented values.

The experimental design allows for comparisons of beliefs and recall across participants exposed to the same data sequence and within individuals across the two parts of the study.

Design details

Randomization method
Done within qualtrics using javascript.
Randomization unit
Individual-level.
Clusters
Around 1800 respondents.
Observations
Around 1800 respondents.
Arms
Around 600 respondents per group.
Primary outcomes
The primary outcomes are participants’ beliefs and recalls following exposure to the data sequence. These include beliefs about the underlying statistical properties of the process, in particular participants’ estimates of the mean, as well as recall of realized values. The main outcomes of interest are (i) cross-sectional belief disagreement among participants exposed to the same data sequence and (ii) within-individual belief changes between Part 1 and Part 2.

Cited by

No citation data available for this registration yet.

Keywords