Experimental Design
We propose an online experiment to study correlation neglect bias on social media, under the context of civil context exam in China. We will explore how reposts from civil servants affect college students’ career choices about being a civil servant. We will recruit Chinese college students as our study participants.
We will recruit participants from Wechat groups and pre-screen college students based on the following criterion:
i. In the last two years in college or in graduate schools.
ii. Spend at least 30 minutes on Microblog every day.
iii. Hesitant about signing up for civil service exams
After getting consent from those participants who meet the criterion above, the experiment will be conducted in the following steps:
1. We will send a baseline survey via email. The baseline survey includes information about demographics, including gender, age, school, major etc. We will then ask about their baseline beliefs on civil service jobs as well as willingness to become civil servants.
2. We will randomly assign participants into two groups and incentivize them to special-follow different Microblog accounts operated by us at the end of the baseline survey (special-following makes sure that our posts appear at the top of followers’ feeds). Those in the Treatment group will follow accounts A, B, C and Z while those in the Control group will follow accounts C, D, E and Z. Accounts A, B, C and Z will post (and repost) things relevant to civil service jobs and Accounts D and E will post irrelevant things to make sure that participants in both groups follow the same amount of information. Participants will be incentivized to upload a screenshot in order to prove that they special follow us. We will also incentivize them to “like” our posts.
3. For the next three weeks, we will post and repost things with our six accounts (all posts will be based on actual posts on Sina Microblog, there won’t be any deception). For accounts A, B and C, B will repost A’s post and C will repost B’s post so that we can get the reposting chains: B//@A and C//@B//@A. (An example of what participants in each group will see can be found in “example_posts.doc”)
Those in the control group will see:
- C // @B // @ A: negative info about civil service jobs.
- Z: positive info about civil service jobs.
- D: irrelevant posts (for example about pests).
- E: irrelevant posts (for example about pests).
Those in the treatment group will see:
- C // @B // @ A: negative info about civil service jobs.
- B // @ A: negative info about civil service jobs.
- A: negative info about civil service jobs.
- Z: positive info about civil service jobs
(Note: The content of the three posts “C // @B // @ A”, “B // @ A” and A will be the same.)
The key point here is that both groups receive the exact same amount of information on civil service jobs. However, for those in the treatment group, due to repeated exposure to the same information sources, if participants are subject to correlation neglect bias, they will overreact to A’s post be more likely to overweight negative information about civil service jobs.
4. During these three weeks, we will keep track of the number of likes and who “liked” it for each post to get a proxy for attention for each post.
5. After following these accounts for three weeks, we will elicit their posteriors about civil service jobs, and willingness to become a civil servant. We will also ask them whether they want to sign up for an email list sending them reminders about civil service exams and track the click-through rate for a video about civil service exams. This endline survey will also be distributed via email.
(New online experiment, May 2024) To further shed light on the mechanism, we implement an online experiment in a more controlled setting. The core design, utilizing reposting chains to construct redundant information, is similar to the field experiment above, except that (1) we present screenshots of Microblog posts, (2) we vary the salience of correlated posts to explore the role of distraction and (3) we use positive information about civil service jobs to construct redundant posts.
In the online experiment, we included two sets of microblog posts that involved civil servant salaries and presented them to participants.
- C // @B // @ A: positive info about civil servant salary.
- @B // @ A: positive info about civil servant salary.
- A: positive info about civil servant salary.
- Z: negative info about civil servant salary.
We randomly assign participants into one of the three groups: Control, Treat-Salient and Treat-Distraction. In Treat-Salient, screenshots from C, B, A are displayed in consecutive order, together with two irrelevant posts and post Z. In Treat-Distraction, screenshots from C, B and A are split apart by irrelevant posts. Finally, in Control group, participants see screenshots from C and Z, as well as irrelevant posts.
The main outcome variable in this online experiment is participants' beliefs about civil servant salary.