Social Media Toxicity and Mental Health

Information on the intervention is hidden until the end of the trial.

Experimental intervention is introduced through a browser extension that all participants have to install.

In the treatment groups, during the intervention period all text content (posts, comments) exceeding a set threshold level of toxicity is hidden. There are two treatment groups. In the first one, Perspective API will be used to assign a toxicity score between 0 and 1 to each post and comment on Facebook, X (formerly Twitter), and Reddit. Each piece of content with a score exceeding 0.33 will be hidden. The second treatment group will additionally rely on the Alienation classifier, newly developed by Jigsaw. This classifier also assigns toxicity scores between 0 and 1. Each piece of content with an Alienation score exceeding 0.7 or with a Perspective score exceeding 0.33 will be hidden.

In the control group, participants will either experience no hiding (half of the control participants) or experience random hiding of social media content (the other half of control participants). The probability of randomly hiding content will be platform-specific and informed by the proportion of content hidden by the two toxicity classifiers used for the treatment groups.

Additionally, between the midline and the endline surveys, we are planning to cross-randomize the respondents in the control into two groups that will either experience an intervention that instrumentates user engagement (e.g., changing latency of the page/feed) or not. The goal is to quantify the direct effect of toxicity hiding on outcome variables as well as establish what part of the treatment effect operates through varying user engagement, which in turn affects the outcomes.

2024-07-10

2025-03-31

1. Mental Health of the respondents measured with two validated scales: PHQ8 (depression) and GAD7 (anxiety).

Heterogeneity:
We will look at heterogeneity with respect to: baseline level of mental health, political affiliation, minority status, and interaction of minority status with baseline level of mental health. Details are provided below.

Re 1: For each scale, we will sum the points for each question. Moreover, we will create a composite index which is the average between the PHQ8 score and the GAD7 score. We will z-score the index. We provide the list of PHQ8 and GAD7 from our surveys for completeness.

PHQ8: Over the last 2 weeks, how often have you been bothered by any of the following?
- Little interest or pleasure in doing things.
- Feeling down, depressed, or hopeless.
- Trouble falling or staying asleep, or sleeping too much.
- Feeling tired or having little energy.
- Poor appetite or overeating.
- Feeling bad about yourself — or that you are a failure or have let yourself or your family down.
- Trouble concentrating on things, such as reading the newspaper or watching television.
- Moving or speaking so slowly that other people could have noticed. Or the opposite – being so fidgety or restless that you have been moving around a lot more than usual.

For each statement, participants select one of four options: not at all, several days, more than half of the days, nearly every day. These will be coded as 1, 2, 3, and 4 respectively when the outcomes are computed (as explained above).

GAD7: Over the last 2 weeks, how often have you been bothered by any of the following?
- Feeling nervous, anxious, or on edge.
- Not being able to stop or control worrying.
- Worrying too much about different things.
- Trouble relaxing.
- Being so restless that it is hard to sit still.
- Becoming easily annoyed or irritable.
- Feeling afraid, as if something awful might happen.

For each statement, participants select one of four options: not at all, several days, more than half of the days, nearly every day. These will be coded as 1, 2, 3, and 4 respectively when the outcomes are computed (as explained above).

Heterogeneity:
- baseline level of mental health (the respondents will be split into two groups: above and below median score of the composite index of mental health),
- political affiliation (Democrats and Independents leaning towards Democrats vs. Independents leaning towards Republicans and Republicans),
- minority status (female respondents, members of LGBTQ+ community, ethnic minorities),
- interaction of minority status and baseline level of mental health (above/below median baseline MH score x having/having minority status).

1.Beliefs about likelihood of discrimination

2. Moral superiority

3. Beliefs about altruism of others

4. Willingness to follow a bot/page/group offering particular content (midline)

5. Willingness to accept to temporarily deactivate their accounts on social media platforms after the endline

Heterogeneity:
We will look at the same angles of heterogeneity as for the primary outcomes.

Re 1: For each of the following statements, we measure agreement on a scale from 0 to 100 regarding the prevalence of various forms of discrimination.

How prevalent do you think are the following situations?
1) Women facing sexual harassment (ranging from inappropriate jokes to unwanted advances) in the workplace.
2) Members of ethnic minorities and LGBTQ+ individuals experiencing hate crime (ranging from abusive language to assault).
3) Despite having equal qualifications, women are not getting hired for a managerial role.
4) Despite having equal qualifications, straight white men are not getting hired because they do not add to the diversity of the workplace.
5) Supervisors in a workplace being excessively critical and demeaning towards employees.
6) Older individuals are denied equal opportunities because of their age.

To compute the outcome, we will look at the mean of standardized scores for each statement. Separately, we will consider the mean of standardized scores but only for scenarios that describe discrimination against a group of which the respondent is a member (for “old people” we will use a cutoff of 50 years old).

Re 2: First, we ask participants to decide whether each of the following statements applies to them (“yes” or “no”).
1) I sometimes feel resentful when I don’t get my way.
2) I’m always willing to admit it when I make a mistake.
3) There have been times when I was quite jealous of the good fortune of others.
4) I am always courteous, even to people who are disagreeable.
5) There have been occasions when I took advantage of someone.

To compute an index of absolute moral superiority, we first standardize the proportion of “yes” answers for each statement. Then, the scores for statements that indicate lower moral superiority (1, 3, 5) are multiplied by −1. Lastly, we compute the mean of the sign-adjusted standardized scores.

Separately, we look at relative moral superiority. Participants are asked the following question:

Considering the statements listed above, over the last 2 weeks, do you think that you represented more or fewer positive qualities than an average person?

They can indicate agreement from 0 to 100 using a slider.

Re 3: In a different survey (with a representative sample of U.S. adults), we offered respondents who use social media a surprise bonus payment of $1 at the very end.

They could anonymously split the bonus between themselves and Save the Children USA – a charity offering aid and support to children with economic and emotional needs.

We ask “On average, what proportion of the bonus payment do you think the respondents who use social media donated to the charity?” ($0-$1 slider with $0.01 increments).

Re 4: We provide an example of the outcome for a bot on X (formerly Twitter). The decision made by participants is whether or not they want to follow the bot.

The bot will be sharing content with some toxicity. We signal toxicity by showing some screenshots of what the first posts will be, but we will not mention it directly.

Overall, the bot is designed to post tweets that have been carefully curated to drive high engagement. The bot tweets five times a day, and each tweet is crafted to spark discussions, generate likes, and be shared widely.

These tweets might include a mix of the following (we will provide screenshots of toxic tweets as an example):

- Thought-provoking questions that encourage people to share their opinions.
- Interesting facts or statistics that prompt conversations.
- Humorous content that is likely to be retweeted and liked.
- Trending topics or current events that are relevant and timely.

Re 5: We will use the same approach as Butera et al. (2022) to measure WTP/WTA to deactivate their social media accounts for 4 weeks after. We will inform participants that a random subset of respondents will see their choices implemented. The number of implemented choices will depend on the funding availability.

Information on the experimental design is hidden until the end of the trial.

The primary purpose of the experiment is to measure the effect of an intervention that reduces exposure to toxic social media content on mental health of users, and to understand the mechanisms linking toxicity and mental health. This study is conducted simultaneously with an RCT registered under the following title “Social Media Toxicity and Political Attitudes”. The trial was registered by Mateusz Stalinski (University of Warwick).

We recruit participants to install a browser extension called "Social Media Research", which has the ability to hide toxic content (posts, comments, replies) on Twitter, Facebook, and Reddit. The extension was specifically designed for Google Chrome browser. The extension may work on other Chromium browsers, but this functionality was not tested. Recruitment is conducted both via social media ads and using survey companies (such as CloudResearch). We will recruit in two waves, the first one in the summer of 2024 (approximately 1,000 people) and the second one in the fall (starting in mid-November). We will include wave fixed effects in our analysis.

Users are randomized into one of 3 groups. In the treatment groups, during the intervention period all content exceeding a set threshold level of toxicity is hidden (please see Intervention section for details, including a description on how we assign toxicity scores). There will be two treatment groups that will use different combinations of classifiers to flag toxic content (see Intervention section). In the control group, half of the respondents will experience random hiding of content. The other half of the control group will not experience any hiding. We will use the following probabilities of assignment to different treatment groups:
- Treatment 1 (first classifier): 30% chance,
- Treatment 2 (second classifier): 30% chance,
- Control: 40% chance (which means an overall 20% chance of random hiding control and a 20% chance of pure control).

After installation and completion of a baseline survey, the user enters an intervention period of 6 weeks, where the extension hides toxic/random content in the groups where this is required. After 6 weeks, we will survey the respondents again in a midline survey phase. After that, the respondents will keep using the extension for 4 weeks before an endline survey. During that time (after the midline survey), the respondents in the control group will be cross-randomized in two groups where they will either experience an intervention that instrumentates user engagement (e.g., changing latency of the page/feed) or not. The goal is to quantify the direct effect of toxicity hiding on outcome variables as well as establish what part of the treatment effect operates through varying user engagement, which in turn affects the outcomes.

The primary outcome and secondary outcomes 1-3 will be measured in all three surveys (baseline, midline, endline). In order to test whether hiding toxicity has an impact on these outcomes, we will use a canonical difference-in-difference specification with outcomes measured in two periods (baseline survey and midline survey) and two treatment groups (toxicity hiding with both sets of classifiers pooled vs. controls pooled). We consider this approach our main specification. This approach takes into account the high overlap in the type of content flagged as toxic by the two sets of classifiers that we employ. We will separately look at whether the additional usage of Alienation classifier led to heterogeneous treatment effects. For secondary outcomes 4-5, which are only measured in either midline or endline surveys, we will rely on between subjects mean comparisons between treatment groups (for the main specification, we will compare toxicity hiding with both sets of classifiers pooled vs. controls pooled).

Before fielding the endline survey of the study starting the week of 21st October, we want to amend the pre-registration by adding a few questions to the endline survey to understand additional mechanisms. These are the additional questions:

a) In the past weeks, did social media make you feel you were missing out on important information, experiences and/or discussions? (Likert scale 5 points)

(b) UCLA 3-item (we will index the responses):
- First, how often do you feel that you lack companionship: Hardly ever, some of the time, or often?
- How often do you feel left out: Hardly ever, some of the time, or often?
- How often do you feel isolated from others? (Is it hardly ever, some of the time, or often?)

(c) In the past four weeks, how often have you felt angry? (Likert scale 5 points)

We will compare the additional outcomes across experimental conditions using the same specifications described above.

After the user installs the browser extension (and agrees to the data collection), the extension assigns the user one of the experimental groups. The following example visualizes our approach well. The extension generates a random number between 0 and 1. If the number exceeds 0.4, the user is assigned one of the treatments: if it is between 0.4 and 0.7 to the treatment that uses Perspective API as classifier, if it is above 0.7 to the treatment that additionally uses Alienation as classifier. If it is below 0.4, the user is assigned to the control group: if it is between 0 and 0.2 it is assigned to the control group with no hiding, while if it is between 0.2 and 0.4 then it is assigned to the control group with random hiding.

Between the midline and the endline, we are planning to cross-randomize the respondents in both control groups into two groups that will either experience an engagement intervention or not. Randomization will be performed by the extension.

Individual

No

N/A

The number of observations depends on how successful we will be in promoting the browser extension (it will be promoted through advertisement on social media in addition to recruitment via survey companies). We estimate to be able to recruit approximately 2,500 users.

We will randomly assign individuals to treatment groups with the following probabilities: 40% to the control group (equal split between pure control with no hiding and control with random hiding), 30% to a treatment group that uses Perspective API classifier, and 30% to a treatment group that additionally uses Alienation classifier.

We use the software GPower to derive the optimal sample size for the experiment. The family test is a one-sided t test and the statistical test is a linear bivariate regression with two groups and differences between slopes (diff-in-diff). We assume that parameters alpha and 1-beta are equal to 0.05 and 0.8, we fix that the allocation ratio of the respondents of the treatment group relative to the control group is 1.5, we assume a standard deviation for the outcomes being the same across groups and equal to 1, and we target a minimum detectable effect of 0.1SD. This implies a total sample of 2,578 respondents (divided in 1,547 in the treatment group and 1,031 in the control group). The minimum detectable effect of our study is reasonable given the results of similar experiments. For example, Braghieri et al. (2022) reports the effect of Facebook adoption on an index of mental health of college students (based on the NCHA survey) to be around 0.085 standard deviation units. Similarly, in a Facebook deactivation experiment, Allcott et al. (2020) report an effect on mental health of approximately 0.09 standard deviation units. Furthermore, it is likely that our intervention, hiding toxic content on three social media platforms, might be stronger than the interventions cited above. First, we hide toxicity on three key social media platforms: X, Facebook, Reddit. Second, we remove arguably “the worst” part of social media (toxic content), while keeping many benefits (such as interacting with friends). Lastly, it is important to point out that some interventions (such as providing access to mental health support apps) are very successful in improving an index of anxiety, depression, and stress even when an intervention lasts only several weeks: Shreekumar and Vautrey (2024) reports effects 0.38 standard deviations at two weeks and 0.46 SDs at four weeks, with persistent effects three months later.