Selective Information Processing: A Field Experiment on De-Biasing

In collaboration with the Bavarian State Ministry of the Interior, we will develop an online platform designed to educate users against disinformation. Participants are either recruited among those who directly visit the Ministry's online platform or via Bilendi. They will be randomly assigned to a control or treatment condition. The treatment group is exposed to content on the online platform that comprises two educational interactive components, each targeting one of the cognitive biases under study. The first component uses a simulated social media feed to explain correlation neglect. The second component employs a short video narrative to illustrate the neglect of unobserved information. Both components conclude with a concise explanation of the respective bias. See PAP for details.

2025-11-05

2026-12-31

1. Measures of correlation neglect and neglect of unobserved information bias
2. Intended Information consumption
3. (Social) media behavior and attitudes

The study is a field experiment conducted in collaboration with the Bavarian State Ministry of the Interior, using two samples of participants: (i) the “Ministry sample” consisting of voluntary platform visitors, and (ii) the “Bilendi sample” recruited online to be representative of the German population in terms of age, gender, and region. Each sample is randomly assigned to a treatment group, which engages with interactive educational elements on an online platform, or a control group, which completes the study survey without prior exposure to these elements.

The treatment group interacts with two educational modules illustrating the mechanisms behind correlation neglect and neglect of unobserved information, followed by concise explanations of each bias. After completing the modules (or directly after joining for control groups), participants complete abstract, incentivized tasks measuring the two cognitive biases and a comprehensive survey covering: (i) participants’ performance in practical tasks that operationalize correlation neglect and neglect of unobserved information in the news domain, (ii) participants’ perceived breadth of their news consumption, their intention to read cross-cutting news with contrasting opinions, and their plans to diversify news sources, (iii) trust in media, perceptions of social media risks and benefits, attitudes toward a hypothetical social media ban, and self-reported ability to detect fake news, and (iv) demographic and cognitive background characteristics.

The design allows for causal inference on the effects of the educational intervention within each sample and assessment of short-term impacts on cognitive biases, intentions to diversify news consumption, and (social) media attitudes and behavior. Data collection and experimental procedures are implemented via oTree, using unique codes to link platform interactions with survey responses. See PAP for details.

Not available

Fair coin flip done by computer to assign to treatment vs. control group

individual

No

Ministry sample: maximum duration of 3 months starting from the platform launch, or until a maximum sample size of 2,000 participants is reached;
Bilendi sample: 2,000 observations

Ministry sample: maximum duration of 3 months starting from the platform launch, or until a maximum sample size of 2,000 participants is reached; Bilendi sample: 2,000 observations

Ministry sample: maximum duration of 3 months starting from the platform launch, or until a maximum sample size of 1,000 participants per condition (control and treatment group) is reached;
Bilendi sample: 1,000 observations per condition (control and treatment group)

For both samples, we aim to recruit 1,000 participants per condition. This sample size allows us to identify a minimum effect size of Cohen's d = 0.125 for both our main outcome measures, at the 5% significance level, with 80% statistical power. We are also interested in checking for treatment effects in substrata (e.g., split by age, gender, or region). Our power analysis suggests that if we want to consider, for example, a quarter of the observations, this should yield a minimal detectable standardized effect of d = 0.251.