Models of Causal Belief Systems and Misperceptions: Experimental Tests

Participants complete an online experiment with repeated computer-generated decision environments involving three binary variables: an action, an intermediate variable or signal, and a payoff-relevant outcome. Across treatments, the environments vary the causal and informational structure of these variables. Participants make incentivized choices and report beliefs about causal relationships among the variables. See the attached document for rationale and details.

2026-05-28

2026-09-01

Primary outcomes are causal belief systems and action frequencies. Belief outcomes include elicited causal-effect beliefs, including beliefs that fix the third variable interventionally, and the graphs participants draw to represent causal influences. Action outcomes are P(A=1) in reverse-causality and two-variable tasks, P(A=1 | X=0) and P(A=1 | X=1) in confounded-choice tasks, choices in the last 30 of 80 rounds, and corresponding high-stakes WTP/WTA choices.

Action frequencies summarize how often participants choose the relevant action in each environment, including state-contingent action rates when X is observed before choice. Belief outcomes measure whether participants perceive causal effects that are absent in the true DGP and whether their reported belief systems satisfy the restrictions implied by single-DAG or DAG-mixture interpretations of their observed histories.

Participants are assigned to online decision environments designed to study causal belief formation and repeated choice. The design includes reverse-causality environments, confounded-choice environments, and two-variable calibration environments. The main comparisons examine how beliefs and choices vary with the environment's causal structure, noise, display order, and signal informativeness. See the attached document for rationale and details.

The main treatment comparisons concern the effect of increasing noise in the DGP on the action frequency and WTP/WTA within each decision environment. Each environment features a high-noise condition and a low-noise condition.

Not available

Done by the survey script

Individual

Yes

There are 11 within-subject treatment comparisons. For 7 of them, we will enlist 120 subjects each, for the remaining four we will enlist 60 subjects each.

1080 individuals We include a subject in the data analysis only if all of the following conditions are met: 1. the subject's prolific authenticity checks has “high” for both fields, LLM and bots. Subjects with missing fields are excluded, too. 2. the subject does _not_ give a correct answer to the question "How many elements does the largest sporadic simple group have?” 3. the subject has a reCaptcha score of at least 0.7 (based on https://web.archive.org/web/20260316085956/https://researcher-help.prolific.com/en/articles/445222-understanding-the-limitations-of-recaptcha-bot-detection-in-research) 4. the subject does not miss the “Monday” attention check (which terminates the survey) 5. the subject passes the technical check in which a sequence of numbers is displayed in video format The target sample size refers to subjects who satisfy all of the above conditions

1080 individuals

aea_attachment_analysis_plan.pdf