On Balance: Using Social Media and Machine Learning for Violence Prevention

The revelations of the Facebook Files have again focused public attention on the spread of misinformation, hate speech and incitement to violence on Facebook. Youths across the globe use Facebook's Instagram and similar platforms to share violent videos. The question inevitably arises if authorities could use these platforms for violence prevention.

Bystander programs represent an established prevention measure. They motivate people to intervene in violent situations and teach the skills for safe and effective intervention. Studies suggest that bystander programs could reduce violent victimization and perpetration. However, face-to-face programs are cost-intensive and difficult to scale. Online programs face the challenge of reaching enough relevant participants. Social media can help on both counts. They allow reaching large audiences at relatively low cost and precisely addressing individuals at risk, i.e. micro-targeting.

 

With our recent paper, we explored ways to implement bystander programs in social media and assess their economic efficiency (Ebers & Thomsen, 2021). In this blog post, we briefly describe how our conceptual framework could be applied to a hypothetical bystander program for illustration. Estimating program costs is straightforward as they simply include all expenses on implementing the program. Estimating benefits requires more effort and makes up the core of our work. Within our framework, benefits arise from the costs of crime avoided and are defined as follows:

  BENEFITS = PARTICIPANTS × BYSTANDER × INTERVENTION ×
(1 - REPLACE) × DAMAGE
(1)


PARTICIPANTS
is the estimated number of program participants. The more participate, the more will have the opportunity to prevent a crime. Our paper explains how to use social media analytics tools and metrics to estimate this parameter. However, it emphasizes that such metrics do not give perfectly accurate information due to cookie deletion, multi-machine browsing, and non-human traffic, for example.

The BYSTANDER term captures the share of future bystanders among participants, who are the only ones who get to intervene. We can maximize the share of future bystanders by micro-targeting individuals at risk of encountering a violent situation. In the paper, we explain how to use social media data and machine learning (ML) for this purpose.

More specifically, we could train a model on social media data; extract the relevant risk factors and micro-target similar individuals. Subsequently, we could estimate the BYSTANDER term based on the precision of our trained model and the overall crime rate, which is a good proxy for the share of individuals at risk within the population.

The INTERVENTION term captures the causal effect on the willingness-to-intervene. It thus captures the share of bystanders who will intervene just because of program participation. Our paper describes how to conduct an RCT on Instagram in order to identify this effect.

Finally, a certain share of bystanders will be injured instead of the victim. We defined this as REPLACE. Logically, (1 - REPLACE) is the share of bystanders who intervene without being injured. Only in these cases, the costs of crime are avoided yielding a social benefit. The product of the four factors described above yields the total number of crimes prevented. Multiplied by the unit costs per crime, DAMAGE, this gives us the program benefit.

To illustrate the application of our framework, consider the following example. Assume a hypothetical bystander program reaching 10,000 participants, which identifies individuals at risk with 69% precision (cf. Hassanpour et al., 2019). Together with a 40% crime rate (DOJ & FBI, 2021), we arrive at a share of 0.40/(1-0.69(1-0.40)) ≈ 68% future bystanders.

In line with meta-analytic evidence (Kettrey & Marx, 2020), assume the program could achieve a 22% intervention effect. Finally, as participants will acquire the skills for safe intervention, we assume little replacement. To keep our estimate conservative, nevertheless, we choose 20% replacement following the Pareto principle.

Based on these assumptions, our hypothetical bystander program would prevent 10,000 × 0.68 × 0.22 × (1 - 0.20) ≈ 1,196 crimes. Multiplied by estimated unit costs of $91,110 (Miller et al., 2021), we would arrive at a benefit of $91,110 × 1,193 = $108,694,230. As long as costs are below this threshold, the program is efficient. Given the relatively low costs of implementation and scaling of social media interventions, they represent a suitable means for crime prevention, likely to break even. With our paper, we aimed to make new assessment methods accessible to a broad readership and promote evidence-based policy-making in this area.

References
Ebers, A., & Thomsen, S. L. (2021). Benefit–Cost Analysis of Social Media Facilitated Bystander Programs. Journal of Benefit-Cost Analysis, 12(2), 367–393.

Hassanpour, S., Tomita, N., DeLise, T., Crosier, B., & Marsch, L. A. (2019). Identifying substance use risk based on deep neural networks and Instagram social media data. Neuropsychopharmacology, 44(3), 487.

Miller, T. R., Cohen, M. A., Swedler, D. I., Ali, B., & Hendrie, D. V. (2021). Incidence and Costs of Personal and Property Crimes in the USA, 2017. Journal of Benefit-Cost Analysis, 12(1), 24–54. https://doi.org/DOI: 10.1017/bca.2020.36.

United States Department of Justice [DOJ], & Federal Bureau of Investigation [FBI]. (2021). Crime in the United States, 2020. https://crime-data-explorer.app.cloud.gov/pages/explorer/crime/crime-trend.

Share this post:

Comments on "On Balance: Using Social Media and Machine Learning for Violence Prevention"

Comments 0-5 of 0

Please login to comment