You need to sign in or sign up before continuing.
Back to History Current Version

Bystander and victim support

Last registered on January 22, 2026

Pre-Trial

Trial Information

General Information

Title
Bystander and victim support
RCT ID
AEARCTR-0017729
Initial registration date
January 19, 2026

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
January 22, 2026, 1:47 PM EST

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Primary Investigator

Affiliation
University of Barcelona

Other Primary Investigator(s)

PI Affiliation
LSE
PI Affiliation
Bocconi University
PI Affiliation
LSE

Additional Trial Information

Status
On going
Start date
2026-01-15
End date
2026-02-28
Secondary IDs
Prior work
This trial does not extend or rely on any prior RCTs.
Abstract
This study examines how beliefs about workplace sexual harassment shape workers’ preferences for supporting victims. Using a survey experiment among the Spanish active population, we test whether correcting misperceptions about (i) the prevalence of sexual harassment, (ii) the relative costs faced by victims versus perpetrators, and (iii) knowledge about how bystanders can intervene affects individuals’ willingness to help victims.
We hypothesize that stronger adherence to sexual harassment myths is associated with lower support for victims, and that providing factual information or guidance on bystander intervention can increase support, although responses may differ across individuals. We also examine whether these interventions generate heterogeneous responses, including potential negative or backlash reactions among specific subgroups. In addition, the study documents the prevalence of these beliefs in the Spanish workforce.
External Link(s)

Registration Citation

Citation
Coly, Caroline et al. 2026. "Bystander and victim support." AEA RCT Registry. January 22. https://doi.org/10.1257/rct.17729-1.0
Experimental Details

Interventions

Intervention(s)
Intervention Start Date
2026-01-15
Intervention End Date
2026-01-31

Primary Outcomes

Primary Outcomes (end points)
Our main dependent variables can be divided into four groups:
1) Belief updating: here, we test whether our three information treatments lead to different updating of beliefs around sexual harassment and discrimination issues.
- Continuous variable between 0 and 100, which represents the answer of the respondent to the question: “Out of every 100 working mothers in Spain, how many do you think have suffered gender discrimination at work?”.
- Dummy variable coded to 1 if the respondent answers “the victim” to the question “When a victim of sexual harassment reports their abuser, who do you think is usually more adversely affected in terms of their job? “.
- Dummy variable coded to 1 if the respondent answers “Agree” or “Strongly agree” to the question “If I witness sexual harassment, I know what actions to take.”
2) Real-stake outcomes in willingness to help victims:
- willingness to donate to sexual harassment charity: dummy variable coded to 1 if the respondent agrees to donate to the charity helping victims of sexual harassment (AVA).
- Amount donated to the sexual harassment-related charity: coded between 0 and 100 euros. Respondents can select any amount between 0 and 780 korus (where 780 kors =180 euros) and we will use the conversion rate of 7.8 korus = 1 euro. Respondents who chose another charity to donate or refused to donate will be coded as 0. We will also do a robustness check where the outcome variable will be the amount given, conditional on having said that the respondent wants to give to AVA. In this case, the amount donated will be coded between 1 and 100 euros for those who selected to give to AVA and actually gave, and missing otherwise.
- Dummy coded 1 if the respondent agrees to sign a petition to fight sexual harassment in firms.
- conjoint experiment: respondents are given a choice: “On the next page, we invite you to put yourself in the shoes of an employee in charge of a mission (you can think of a mission related to your own job or imagine a person in charge of organizing an annual event within the company). Here are the profiles of the two volunteer colleagues. Choose who you prefer to work with.” The respondents are then asked to choose 4 times between two profiles. The possible collaborators differ in four characteristics: gender (Male/Female/No information), experience (No experience, 2 years of experience, 4 years of experience); Behavior (No information/Other colleagues have mentioned that this person has behaved inappropriately/Other colleagues have mentioned that this person has reported the inappropriate behavior of another employee); and Professionalism (No information/Is late or fails to complete some tasks/Complies with their tasks on time). The reference category for each characteristic will always be the “No information” or “No experience” one. We create a dummy variable (𝑌𝑖𝑡𝑗) for whether individual i selected a specific colleague j in table t. The profile chosen will be coded as one and the profile not chosen coded as zero. All profiles will be pooled together initially. As we might expect different reactions for male and female respondents depending on whether the potential harasser or victim is a male or a female, we will also analyze separately for them: (1) restricting the sample to choices between two male profiles; (2) restricting the sample to choices between two female profiles.
Additionally, we will also measure the willingness to trade experience to avoid a potential victim or harasser in a similar way as in Folke and Rickne’s (2022) willingness to pay.

3) Policy preferences:
Dummy coded to 1 if respondent answered “Somewhat stricter” or “Much stricter” to the question “To what extent do you think prison sentences for those who commit sexual harassment at work should be less strict, stricter, or remain as they are?”

Dummy coded to 1 if respondent answered “Somewhat stricter” or “Much stricter” to the question “To what extent do you think penalties for employers who fail to act on credible allegations of sexual harassment in the workplace should be less strict or more strict?”

Index comprised of the z-score of the sum of the two above-listed dummy variables
As a robustness check, for these two policy preferences questions, respondents who select “I have no opinion” will be coded as missing and we will replicate the analysis.

4) Open-ended questions: We will analyze open-ended survey responses using large language model (LLM)-based text classification via the Anthropic API (Claude Sonnet 4). We have two open-ended questions:
1. “What is your opinion on the prevalence of sexual harassment in the workplace in Spain?”
2. “What factors could influence someone like you to decide whether or not to act in a situation of sexual harassment? can list as many as you consider relevant”
For the first question, we will prompt the LLM to systematically code multiple dimensions: (1) whether the respondent comments on the magnitude of the phenomenon (too large/not significant) (2) whether the respondent comments on what they’ve learned from the treatment, (3) other dimensions mentioned as possible mechanism to explain treatment effects.
For the second question, we will: (1) prompt the LLM to create a classification of possible factors mentioned, and (2) ask LLM to systematically code whether the proposed factors are 1) individual, 2) firm-based, or/and 3) policy-based.

The coding prompt will be pre-specified and tested on a subset of responses before full implementation to ensure consistency and accuracy. The LLM will return structured outputs (JSON format) that will be merged with the survey dataset using unique respondent identifiers, producing quantifiable variables suitable for statistical analysis alongside closed-ended survey responses. All coded variables will be exported to Stata format for integration with the main analysis dataset.
Primary Outcomes (explanation)

Secondary Outcomes

Secondary Outcomes (end points)
Data collection by the survey provider began on January 15th, 2026, and is expected to conclude by the end of January or early February. We have not yet received access to the data, which will be made available once the data collection is complete.

We plan to test heterogeneities by:
- age groups (18-34, 35-44, 45-65),
- empathy (dummy coded 1 for respondents who answered “disagree” or “totally disagree”),
- right-leaning political views as proxied by a dummy coded 1 if the respondent answers “Totally agree” or “agree” to the statement “Unemployed individuals should accept any available job or lose their unemployment benefits.” Respondents who said “don’t know” will be coded as missing.
- Has already received a training on sexual harassment prevention or gender equality at work (dummy coded to 1 if received any of these two trainings). A robustness check will focus on training specific to sexual harassment with a dummy coded 1 if received the training on sexual harassment.
- By confidence in priors about the prevalence of sexual harassment, measured on a five-point scale from 1: “Very unsure” to 5: “Very sure”.
- sexist views: dummy coded to 1 if answered “Totally agree” or “Agree” to any of the two statements measuring sexism. “Don’t know” will be coded as missing.
- By prior inability to recognize an inappropriate gender-based situation: dummy coded to 1 if the respondent answered to the vignette one “humor” or “neither”, and 0 if answered “sexism” or “sexual harassment”. To further explore heterogeneity, we will also construct separate dummy variables for each response category, provided that at least 10% of respondents select that category, to ensure sufficient statistical power.
- Measures of prior beliefs on prevalence of sexual harassment: (1) a dummy coded to 1 if answered 27 or less, 0 if answered 28 or more; (2) one dummy coded to 1 if answered 22 or less, one dummy coded to 1 if answered between 23 and 33, one dummy coded to 1 if answered 34 or more; (3) one dummy coded to 1 if strictly below the median, 0 if equal or above the median answer; (4) one continuous variable between 0 and 100.
- Measures of prior beliefs on consequences for victims: (1) a dummy coded 1 if the respondent selects “Neither of the previous ones”, 0 otherwise; (2) an index ranging from 0 to 4 constructed as the sum of binary indicators for whether the respondent selected each of the following items: anxiety, depression, trouble sleeping, chronic stress. Each item is coded as 1 if selected, 0 otherwise. “Don’t know” and “Neither of the previous ones” are coded as zero.
- Measures of prior beliefs on costs for harassers: (1) dummy coded to 1 if answered “Moderately often”, “Very often” or “Extremely often”, 0 otherwise; (2) dummy coded to 1 if answered “Very often” or “Extremely often”, 0 otherwise; (3) variable coded on a 1–5 scale, with higher values indicating a belief that sexual harassers are punished more frequently.
- Index on prior misbeliefs about sexual harassment: dummy coded 1 if measure (1) of prior beliefs on prevalence of sexual harassment is equal to 1 and measure (1) of prior beliefs on consequences for victims is equal to 1 and measure (2) of prior beliefs on costs for harassers is equal to 1; 0 otherwise.
- prior beliefs on respondent’s self-rated knowledge of good practices in case of sexual harassment: dummy coded 1 if answered “Good” or “Very good”.
- Actual prior knowledge of good practices for firms: dummy coded to 1 if respondent responded correctly to all 5 items; continuous variable from 1 to 5 for each correct answer.
- prior beliefs about the prevalence of false accusations: (1) dummy coded to 1 if answered “Moderately common”, “Very common” or “Extremely common”. “Don’t know” are coded as missing; (2) dummy coded to 1 if answered “Very common” or “Extremely common”. “Don’t know” are coded as missing.
Since only half of the respondents were presented with this question, we will also test for potential priming effects of seeing this question on the treatment effects by verifying, for each treatment, if effects are similar for respondents who saw that question and those who did not see it.
- working in a male-dominated firm: dummy coded to 1 if answered “Exclusively or almost exclusively men” or “Majority of men”. “Prefer not to answer” will be coded as missing.
- experience of sexual harassment (or discrimination) as a victim [or a witness]: dummies coded to 1 if experienced [or witnessed] any sexual harassment situation (or gender discrimination)

We will further analyze petition support using the same linear model as the main specification with an ordinal outcome variable coded as 0 if the respondent refuses to sign, 1 if they sign anonymously, and 2 if they sign with their name. As a robustness check, we will also estimate an ordered logit model. To isolate public support for the petition, we will additionally estimate our linear model using as outcome a dummy variable equal to 1 if the respondent signs with their name and 0 otherwise.

If applicable, we want to test for potential substitution and moral licensing effects. We distinguish moral licensing from substitution by separately analyzing (i) the likelihood and amount of donating overall and (ii) the allocation of donations across organizations. So, we will analyze also the following outcomes:
- Any donation: dummy variable coded to 1 if the respondent agrees to donate to any of the three charities offered, 0 if he doesn’t wish to give.
- Total donation: continuous variable between 0 and 100 euros representing the amount given to any of the 3 charities. People who refused to give will be coded as 0.
- Other donation: dummy coded to 1 if the respondents agree to donate to any of the 2 charities non-related to sexual harassment (ONCE or Cruz Roja), 0 if they chose AVA or not to donate.
- Amount other: continuous variable between 0 and 100 representing the amount given to any of the 2 charities non-related to sexual harassment (ONCE or Cruz Roja).
If relevant, we will also test for a potential backlash effect with a dummy coded 1 if the respondent answered “Agree” or “Strongly agree” to the question “Some people believe that sexual harassment and discrimination against women are discussed too often in the workplace. To what extent do you agree or disagree with this statement?”. Similarly, we might also test for zero-sum thinking with a dummy coded 1 if the respondent answered “Agree” or “Strongly agree” to the question “Some people believe that when women advance in the labor market, it is often at the expense of men. To what extent do you agree or disagree with this statement?”

To test mechanisms, we also code a dummy equal to 1 if respondent answered “Agree” or “Strongly agree” to the statement “Donations to specialized organizations are an effective way to help victims of sexual harassment at work.” We also code a dummy equal to 1 if respondents answered “Quite useful” or “Very useful” to the question about the usefulness of the training, and a dummy equal to 1 if they answered “Quite applicable” or “Very applicable” to the question about the applicability of the training.

For the conjoint experiment, to test the robustness of the results, we will also reproduce the analysis restricting the sample to the first choice of profiles offered to respondents.

To explore policy preferences further, we construct a continuous variable for each question capturing respondents’ preferred strictness of sanctions related to workplace sexual harassment. The variable will take values from 1 to 5, with higher values representing preference for stricter sanctions. To deal with the issue of multiple hypotheses testing, we will also construct an index defined as Sanction Strictness Indexi=1/2 * (Ziprison+Ziemployer) where the variables are respectively the z-scores for the prison policy question and the employer policy question.
We will address multiple hypotheses testing in two ways: (1) the use of indices as described in the different sections and (2) accounting for the False Discovery Rate using the “sharpened q-value approach” (Benjamini et al. 2006; Anderson 2008) as a robustness check, over the 3 main families of outcomes defined in section 3: belief updating, real-stake outcomes and policy preferences.
As the bystander training treatment is a bit longer than the two others, we will also test whether the time spent on the 3 filler questions after the treatments is not significantly higher than for the other treatment groups. We will also use the filler questions to analyze whether respondents in male-dominated firms are working in less supportive working environments.
Secondary Outcomes (explanation)

Experimental Design

Experimental Design
We first collect baseline demographic and employment characteristics about respondents. They are then asked about any previous experience of different discrimination or harassment items, asked to recognize in a vignette whether a situation is sexism, sexual harassment, humor or neither. We also measure gender norms and empathy. Then, we measure respondents' prior beliefs about different sexual harassment facts to measure adherence to sexual harassment myths. Following that, respondents are randomly assigned to one of the different treatment or control groups.The following block presents some filler questions about workplace climate. Finally, we measure our different outcome and mechanism variables.
Experimental Design Details
Not available
Randomization Method
Randomization done by the survey provider
Randomization Unit
The randomization unit is the individual.
Was the treatment clustered?
No

Experiment Characteristics

Sample size: planned number of clusters
N/A
Sample size: planned number of observations
We aim for 7,000 overall respondents. Any deviation from this figure is due to the survey provider.
Sample size (or number of clusters) by treatment arms
Based on the target of 7000 respondents, we aim for:
1) Active control group: 11.1% of sample
2) Passive control group: 11.1% of sample
3) Prevalence treatment: 22.2% of sample
4) Costs treatment: 22.2% of sample
5) Bystander training treatment: 33.3% of sample
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
IRB

Institutional Review Boards (IRBs)

IRB Name
Ethics Committee - Department of Social Policy - London School of Economics
IRB Approval Date
2026-01-07
IRB Approval Number
673190