Identity in journalism: Evidence from News Reporting of Violence Against Women

This project aims to investigate whether gender differences in reporting about violence against women also translate into changes in readers' preferences and behavior, by examining readers' attitudes, policy preferences, and engagement as outcomes using an online experiment with Italian respondents. The intervention is an information treatment that entails showing respondents a snippet from a newspaper article about the murder of a woman at the hands of her partner, in which I randomize the presence of different narratives coming from real newspaper articles, that mimic female-like and male-like writing.

The sample will be randomly split into four groups, with each group comprising 25% of the respondents. The intervention is an information treatment that entails showing respondents a snippet from a newspaper article about the murder of a woman at the hands of her partner, in which I randomize the presence of different narratives coming from real newspaper articles. Each of the four groups will be randomly assigned to one of four different treatment conditions.

The first is a Control where respondents will see one of the two baseline articles about two different crimes. The remaining groups are treatment groups and will be shown the same baseline article that respondents in the Control see, with an additional sentence including a different narrative. Treatment EFP provides an episodic framing to the article by including perpetrator-centered narratives, namely a sentence focusing on the perpetrator's point of view and motive for the crime. Treatment EFV consists again in giving an episodic framing to the article, but instead includes a victim-centered narrative. For this treatment arm, I will employ three distinct narrative framings for each crime. Finally, Treatment TF will include in the article some information about the phenomenon of violence against women in Italy, as well as information on help resources, giving a thematic framing to the article.

2026-02-11

2026-03-31

Empathy towards the victim and perpetrator, preference for government intervention (measured through petition signing), perceived prevalence of help-seeking behavior and crime incidence

There are three sets of primary outcomes. The first set aims at measuring attitudes towards the perpetrator and the victim of the crime:
- Empathy: I use an empathy thermometer to capture empathy towards the victim and the perpetrator on a scale from 0 (no empathy) to 100 (max empathy)

The second set of outcomes focuses on preferences for government intervention related to these crimes
- Petition signing: a binary outcome that measures respondents' willingness to sign a petition supporting government intervention to counter violence against women. To protect respondents' anonymity while still making it realistic, I inform participants that I will communicate the share of respondents who signed the petition to the Parliamentary Commission on Femicide and any form of Gender-Based Violence.

The last set of outcomes aims at capturing perceptions about the incidence of gender-based violence-related crimes and help-seeking behavior.
- Perceived prevalence of help-seeking behavior: I will ask the prevalence of calls to the dedicated helpline number and reports to the police about domestic violence and stalking for each 10000 women in the province of the respondents. I will provide respondents with monetary incentives to answer the question correctly. I am interested in how much their perception deviates from the true value. Therefore, I will calculate the difference between their perception and the true value and standardize it to have mean of 0 and standard deviation of 1.
- Perceived crime incidence: I will ask respondents for their best guess about the prevalence of domestic violence and stalking for each 10000 women in the province where they live. I cannot provide any monetary incentive for this question since there is no data on real crime incidence. Consequently, in this case, I cannot compute deviations from the true value, so I will simply standardize the outcome.

Fairness of compensation and punishment, engagement

The first part of secondary outcomes still focuses on the perceptions of the crime and will allow me to provide additional insights and contribute to the interpretation of the effect of the treatments on the primary outcomes.

- Fairness of Compensation: I ask respondents whether they think the compensation by the Ministry of Justice was too low or too high on a scale from 0 to 10, which I will then standardize to have mean of 0 and standard deviation of 1.
- Fairness of punishment: perceived fairness of the sentence length given to the perpetrator on a scale from 0 to 10, which I will then standardize to have mean of 0 and standard deviation of 1.

The last set of outcomes is motivated by previous conversations with journalists, who have precise priors about which reporting practices lead to higher engagement. Therefore, this part aims at understanding the levels of engagement with the newspaper articles and the newspaper itself. These outcomes will not be part of the main hypothesis testing, as they do not address the primary research question.

- Interactions on social media: I will ask respondents how they would react to the article if they saw it on social media (like, dislike, share, report, or nothing)
- Preference for news outlet:
1. I ask them if they would use the newspaper as a source of information for crime articles (binary outcome)
2. I allow them to interact with the newspaper as they would on a news website. Specifically, I include links to other articles and monitor whether they click on them.
3. I will show respondents all the variation of the article and ask them which they would prefer to read in a newspaper.

The experiment will be run entirely online and administered through the survey platforms Prolific and Bilendi on a sample of Italian respondents, February 11 and March 31, 2026. The recruitment will remain open until either the target number of participants is reached or the end date. The final sample will be randomly split into four groups, with each group comprising 25% of the respondents and seeing a different version of the crime article. These articles are designed starting from real newspaper articles about murders of women, where the names have been changed to protect the identity of the victims.

The experiment will be run entirely online and administered through the survey platforms Prolific and Bilendi. I will recruit 1500 to 2000 respondents on these survey platforms, and through the platform's filtering systems, I will only recruit those who are based in Italy, proficient in Italian (the survey language), at least 18 years old, and have a balanced gender distribution. Moreover, to ensure that respondents in the sample devote sufficient attention to the survey content, I will include two instructions manipulation checks as described in Stantcheva (2023). Respondents who fail both attention checks will be redirected out of the survey, and their submission will be returned. All of these sample restrictions will be implemented before randomizing the respondents into the treatments. Finally, to ensure that respondents are not using ChatBots to complete the survey, I will include an instruction manipulation check where the instructions will be given in a video following Celebi et al. (2025).

I will recruit respondents between February 11 and March 31, 2026. The recruitment will remain open until either the target number of participants is reached or the end date.

At the beginning of the survey, I will collect basic demographic information about respondents, including gender, age, number of children and their sex, and province of residence. Moreover, I will also collect information about political preferences, their preferred medium for news information, and agreement with statements on gender equality from the World Value Survey, as a proxy for gender norms. Next, I will randomize respondents into four different groups with 25% of respondents per group, each seeing a different version of the news article. Finally, I collect the primary and secondary outcomes.

Randomization done by Qualtrics.

Individual

No

1500 to 2000 respondents

1500 to 2000 respondents

375 to 500 respondents

- empathy: 3 percentage points, corresponding to 0.2 S.D. and 43% of the baseline - petition signing: 10 percentage points, corresponding to 0.21 S.D. and 15% of the baseline - Preceived help seeking: 435 calls out fo 10000 women, corresponding to 0.206 S.D. and 22% of the baseline