The willingness to condemn workplace sexual harassment: An experimental investigation

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

Participants are shown a series of video vignettes describing workplace interactions between a man and a woman. Across participants, we vary which videos to show, and within participants, we vary the order of the videos. In total, we sample from a pool of 2,352 videos, and each participant views 8 videos. Each video starts with a conversation between a man and a woman, and then the man takes an action that might be classified as sexual harassment.

The videos vary by the severity and context of the action. In terms of severity of sexual harassment, we have eight categories:
1) asking out,
2) inappropriate compliment,
3) inappropriate comment about women in general,
4) workplace sexism,
5) showing explicit images on the phone,
6) putting hands on her shoulders,
7) rubbing her thigh,
8) exposing himself.

In terms of the context, we vary across videos:
1) the man’s seniority (boss, work colleague),
2) the location of the interaction (workplace, a nearby coffee shop),
3) repetition (whether or not the man repeats the behavior),
4) the conversation topic (work, personal life),
5) whether the woman gives emotional encouragement, if personal life is discussed,
6) the way the woman is dressed (demure, “sexy”),
7) the ethnicity of the man using both ethnic names and visual cues.

We hold the name and ethnicity of the woman constant across all videos. Our pool of videos consists of all combinations of the above actions and contextual factors, with the following exceptions: the ‘workplace sexism’ action is only shown if the conversation is about work, and the ‘explicit images’ action is only shown if the conversation is about personal life.

As a supplementary intervention, we vary the context of an additional video that is shown to some participants. This video, which is drawn from the same pool of videos, always describes actions likely to be perceived as severe sexual harassment. We randomize some information regarding the context of the interaction such as seniority, location, or the man’s ethnicity. This additional randomization serves to test how contextual information affects participants’ tolerance for leniency on the part of someone else towards the action. We describe how this part of our design is implemented in EXPERIMENTAL DESIGN below.

2025-12-16

2026-04-30

1. Workplace consequences for the man

2. Monetary sacrifice to punish the man

3. Index of willingness to condemn the man’s behaviour

We will analyze heterogeneity of the treatment effects by the following characteristics of the respondent: (a) gender, (b) age, (c) political affiliation, (d) an index of traditional gender norms, (e) an index of masculinity, and (f) full time employment.

Re Outcome 1: We use the following question to elicit this outcome.

Imagine you are a colleague of the same seniority as the man, and you witnessed the above interaction. What consequences should the man face? Pick an option that best represents your view.
- No consequences
- Be required to attend mediation with the woman
- Receive a formal warning and be required to attend additional training on workplace conduct
- Be fired

Moreover, at the end of the survey, we ask participants to rank these consequences from least severe (1) to most severe (4). For each video-person pair, we will code the severity of the consequence that the participant selected according to their own ranking.

Re Outcome 2: Each participant is asked if they would be willing to give up their $1 bonus payment to prevent a person who admitted to or was accused of behaving in a way similar to what is described in the video from getting a $1 bonus. The choice is implemented for one of the videos in the survey, the one which describes behavior that someone else who took our survey either admitted to committing or was accused of. Participants do not know which video describes a real person. They make a hypothetical decision for each of the videos, and one of them (related to a real person) is actually implemented.

Re Outcome 3: We will compute an inverse-covariance-weighted index of willingness to condemn based on three outcomes: workplace consequences for the man (primary outcome 1), monetary sacrifice to punish the man (primary outcome 2), and general assessment of the man’s behavior (secondary outcome 1, defined below).

To construct the index we will follow the steps described in the paper “Multiple inference and gender differences in the effects of early intervention: A reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects.” by Anderson (2008), Journal of the American Statistical Association.

Re Heterogeneity:

(a) For heterogeneity by gender, we will use the following question.

What is your gender?
- Female
- Male
- Non-binary
- Other (please describe if you wish)
- I'd prefer not to answer

We will create an indicator variable for whether a participant selected “Male”. We will report heterogeneity of the treatment effects with respect to that indicator variable.

(b) For heterogeneity by age, we will compare the effects for individuals with above/below median age.

(c) For heterogeneity by political affiliation, we will use the following question.

Which of the following best describes your political view?
- Republican Party
- Democratic Party
- Independent
- I'd prefer not to answer

(d) We will compare the effects for individuals with above vs. below median values of an inverse-covariance-weighted index (Anderson, 2008) of traditional gender norms. The following questions are used to construct the index.

How do you feel about the following statements? Do you agree or disagree with them? [Answers are on a five point scale: Strongly disagree, Disagree, Neither agree nor disagree, Agree, Strongly agree]
- All in all, family life suffers when the woman has a full-time job
- A husband's job is to earn money, a wife's job is to look after the home and family
- Both the husband and wife should contribute to the household income

(e) We will compare the effects for individuals with above vs. below median values of an inverse-covariance-weighted index (Anderson, 2008) of masculinity. The following questions are used to construct the index.

[Answers are on a five point scale: Strongly disagree, Disagree, Neither agree nor disagree, Agree, Strongly agree]
- Winning is the most important thing
- Sometimes violent action is necessary
- It bothers me when I have to ask for help
- I love it when men are in charge of women
- It is important to me that people think I am heterosexual

This analysis will be performed for male respondents only.

(f) We will use answers to the Prolific pre-screener question on whether an individual is in full time employment or not.

1. General assessment of the man’s behavior

2. Beliefs about others’ decisions on consequences for the man

3. Workplace support for the woman

4. Punishment for leniency (dictator game split)

We will analyze heterogeneity of the treatment effects by the same characteristics as for the primary outcomes.

Re Outcome 1: We use the following question to elicit this outcome.

On a scale from 0 to 10, how badly has the man behaved? [slider, with 0 = He did nothing wrong and 10 = The worst behavior you can imagine]

Re Outcome 2: We use the following question to elicit this outcome.

What percentage of participants who watched this video do you believe selected each of the following options as consequences for the man? If your estimate for a category is within +/- 2 percent of the actual answer, you will earn an additional $0.20 for that category.
- No consequences
- Be required to attend mediation with the woman
- Receive a formal warning and be required to attend additional training on workplace conduct
- Be fired

Similarly to primary outcome 1, we will also compute a measure where for each video-person pair, we will apply weights to the predicted shares that correspond to the participant’s own ranking of severity of the consequences

Re Outcome 3: We use the following question to elicit this outcome.

What do you think should happen to the woman? Pick an option that best represents your view.
-The woman should be moved to another team (or workplace)
- The woman should be paid a reasonable sum to sign an NDA (Non-Disclosure Agreement) and leave the firm
- The woman should be offered support, paid leave and compensation
- The woman should be told to conduct herself more professionally in the future
- No action needs to be taken
- The woman should receive acknowledgement, an apology and assurance that the occurrence will not be repeated

Moreover, at the end of the survey, we ask participants to rank these consequences from the least supportive (1) to the most supportive (6). For each video-person pair, we will code the level of support that the participant selected according to their own ranking.

Re Outcome 4: This outcome is elicited for individuals who receive the supplementary intervention as described in EXPERIMENTAL DESIGN. We use the following question to elicit the outcome.

Use the form below to split the money between yourself and the person you were matched with in any way you want. Put the number of dollars you wish to keep in the box labeled "For you". Put the dollars you wish to go to the other person in the box labeled "For the other person". Note that the two amounts must add up to $1.

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

We recruit participants on Prolific. We restrict the pool of eligible participants to individuals on desktop devices who have access to an audio device. Participants need to pass an audio test in order to progress to the substantive part of the survey.

Participants are instructed that they will watch multiple videos depicting interactions between men and women in a workplace. They are informed that each video features different people (even if the names are the same). We ask respondents to imagine themselves as a colleague witnessing the scenarios depicted in the videos. Importantly, we tell participants that one of the videos in the survey showcases a behavior that a real person on Prolific says they did, or were accused of doing (names have been changed).

Participants then proceed to watch 8 videos randomly drawn (without replacement) from a pool of 2,352 videos. Each video starts with a conversation between a man and a woman, and then the man takes an action that might be classified as sexual harassment. The videos vary by the severity and context of the action. For example, the set of severities includes behaviors such as asking out, inappropriate compliment, workplace sexist, or rubbing her thigh. The context includes multiple dimensions of variations, such as the location of the interaction (workplace, a nearby coffee shop) and the man’s seniority (boss, work colleague). See the INTERVENTION section for a comprehensive description of attributes that we vary.

Randomization is performed at attribute level. Before we show each video, we separately randomize the severity, location, man’s seniority, etc., from the set of possible values that these variables can take. Based on the outcome of these randomizations, we show the unique video from the pool of 2,352 videos that matches the randomly selected attributes.

After each video, we collect a number of outcomes, including different measures of willingness to condemn the perpetrator (man). See PRIMARY OUTCOMES and SECONDARY OUTCOMES for a full list.

We evaluate treatment effects using participant-by-video-level regressions, which allows us to utilize both within-subjects and between-subjects variation. Specifically, we regress each outcome variable on a set of indicator variables for each attribute. For attributes with more than 2 possible values, such as severity, we will include indicator variables for each level of the attribute (except for the reference category). The regressions will be estimated with individual fixed effects. Standard errors will be clustered at the individual level.

Separately, we will consider a fully-interacted regression, in which we will include all possible interactions of the indicator variables. We can estimate such a regression as the sample size is calibrated so that each of the 2,352 videos (with its unique set of attributes) will be watched, in expectation, by 20 individuals.

We will ask participants several attention questions in between videos. These questions require recalling basic information from the last video and aim to test whether the participants genuinely watched it. Moreover, at the end of the survey, we ask participants whether they experienced any technical issues with the videos (such as long loading time, not playing smoothly, parts of the videos not displayed properly). As a part of the analysis, we will report results based on the subsample of observations with high attention and no technical issues.

After the 8 treatment videos, participants will watch a 9th video. This video always comes from the pool of videos for which we have a participant who either behaved or was accused of behaving in a way described in the video (some of these participants were identified in the pilot survey). This ensures that the way in which we elicit the bonus sacrifice outcome is not deceptive (see PRIMARY OUTCOMES for more discussion). We elicit the usual outcomes for the 9th video, but we will not include them in the regression analysis.

Lastly, a subsample of participants will watch an additional video describing an action likely to be perceived as severe sexual harassment. For this video, we implement our supplementary intervention (randomization of the context of this video). We randomize information such as seniority, location, or the man’s ethnicity. Participants who watch this video will receive an additional $1 bonus payment. We also match each participant to another person who took the study and saw the same video. The participant receives a few pieces of information about the other person, including that they recommended no consequences for the man (to ensure we are truthful, we randomly select the participant’s match from the pool of people who actually answered this way). Then, the participant plays a dictator game with the matched person, splitting the $1 bonus payment between themselves and the matched person. This measures tolerance for leniency as a function of the interaction context.

We can only implement this additional video and the supplementary intervention with a subsample because we must first gather enough data in the experiment to have a pool of “lenient” participants with all of the characteristics of interest. Once this data has been gathered, we will then implement the additional video and supplementary intervention with all subsequent participants.

Qualtrics randomization + JavaScript randomization in Qualtrics (for video attributes)

Video attribute level

No

N/A

As explained in EXPERIMENTAL DESIGN, the sample size is calibrated such that each of the 2,352 videos (with its unique set of attributes) will be watched, in expectation, by 20 individuals. Given that each person watched 8 videos, this requires recruiting 5,880 people. The total number of observations (video-by-individual pairs) is 47,040.

We have 5,880 people overall who view 8 videos each. Our treatments (attributes of the interaction) are randomized at the video level, independently across people, such that each combination of attributes is equally likely to be picked. Therefore, the number of people exposed to a treatment – for instance, the number of people who are shown at least one video with a given severity – is a random variable.