Computational Justice: Simulating Structural Bias and Interventions

Ida Momennejad, Stacey Sinclair, Mina CikarabioRxiv 2019 · DOI

TL;DR

This paper uses agent-based simulations to quantify the costs of structural gender bias (unequal representation) in institutions and to evaluate the effectiveness of interventions. The simulations reveal that unequal gender ratios significantly increase costs for women and institutions, and even a sudden shift to equal representation doesn't fully resolve the issue without increased interpersonal bias confrontation.

What Problem Does This Paper Address?

Gender inequality is a persistent issue in high-prestige professions, stemming from both structural biases (like underrepresentation) and interpersonal biases (like sexism). While interventions exist, their effectiveness, especially for structural biases, is hard to measure due to logistical challenges and the long time scales involved. This paper aims to bridge this gap by using computational simulations to estimate the costs of structural bias and to compare the long-term impact of different interventions, particularly focusing on how to effectively reduce these costs and promote equality.

Methods

The researchers employed an agent-based simulation model to explore the dynamics of gender bias in simulated institutions. Key aspects of their methodology include:

  • Agent-Based Modeling: Each simulation involved 100 agents representing individuals within an institution. These agents interacted in simulated meetings over 1000 rounds.
  • Parameterization: Crucially, the model's parameters (probabilities of bias, objection, and learning rates) were derived from empirical studies on human behavior related to gender bias and confrontation. This grounding in real-world data aims to make the simulation results more relevant.
  • Isolation of Structural Bias: To specifically understand the impact of structural bias, the simulations were initialized with equal probabilities of interpersonal bias for male and female agents. This allowed the researchers to attribute observed disparities solely to the unequal gender ratios.
  • Intervention Simulation: The model allowed for the simulation of two main types of interventions:
  • Interpersonal Bias Confrontation: Agents could object to sexist comments, with different costs associated with being a target or an ally.
  • Structural Equality Policy: The simulation tested the effect of enforcing a 50/50 gender ratio.
  • Learning Mechanism: Agents learned from interactions, with bias probabilities increasing when sexism occurred without objection and decreasing when objections were made.

Datasets

The paper does not use traditional datasets for training or evaluation. Instead, it relies on a comprehensive literature review of empirical studies on gender bias, sexism, and bias confrontation to derive the parameters for its agent-based simulation. This approach allows them to create a simulated environment that reflects observed human behaviors and their consequences.

Key Insights and Ideas

  • Structural Bias is Costly: Unequal gender ratios are not just a symptom of inequality but actively contribute to higher costs for individuals (especially women) and the institution as a whole. This highlights the systemic nature of the problem.
  • Interpersonal Confrontation Matters: While structural changes are important, the active confrontation of bias by both targets and allies plays a significant role in mitigating the negative impacts of structural bias.
  • Equality Isn't Instantaneous: Simply enforcing a 50/50 gender ratio in an institution with a history of structural disparity is insufficient to eliminate bias. The ingrained probabilities of sexism and the learned behaviors persist.
  • Combined Approach is Key: The most effective way to combat persistent gender bias appears to be a combination of structural interventions (like equality policies) and robust interpersonal bias confrontation strategies.

Key Results

  • Increased Costs with Skewed Ratios: Institutions with lower proportions of female agents experienced significantly higher costs of bias for female agents and the institution overall.
  • Bias Propagation: In environments with unequal gender ratios, sexism probabilities increased over time, particularly among the majority group.
  • Limited Impact of Equality Policy Alone: Enforcing a 50/50 gender ratio in a previously skewed institution did not immediately equalize sexist encounters or costs. The gap between male and female agents widened over time in these simulations.
  • Effectiveness of Increased Objection: The gender gap in sexist comments disappeared only when objection probabilities were significantly increased, with 20% more agents (10% targets and 10% allies) objecting with certainty.

Limitations and Future Work

The authors acknowledge several limitations:

  • Binary Gender Model: The simulations focused on a binary gender system, not accounting for the complexities of intersectionality (e.g., race, ethnicity, sexual orientation).
  • No Institutional Exit: The model did not include agents leaving institutions due to the cumulative costs of bias, which is a significant factor in real-world "leaky pipelines."
  • Assumed Equal Probabilities: For simplicity, they assumed equal probabilities of displaying bias and equal costs for men and women in certain scenarios, where empirical data was lacking.
  • Simplified Learning: The learning mechanisms were based on specific studies and could be more nuanced.

Future work suggested includes:

  • Intersectionality: Incorporating multiple social group memberships for agents.
  • Institutional Exit: Modeling the phenomenon of individuals leaving organizations.
  • Norms and Social Influence: Exploring the impact of social norms and network structures on bias.
  • Data Collection: Identifying specific areas where empirical data is needed to refine simulation parameters and improve predictive accuracy.

Summary updated

Abstract

Abstract Gender inequality has been documented across a variety of high-prestige professions. Both structural bias (e.g., lack of proportionate representation) and interpersonal bias (e.g., sexism, discrimination) generate costs to underrepresented minorities. How can we estimate these costs and what interventions are most effective for reducing them? We used agent-based simulations, removing gender differences in interpersonal bias to isolate and quantify the impact and costs of structural bias (unequal gender ratios) on individuals and institutions. We compared the long-term impact of bias-confrontation strategies. Unequal gender ratios led to higher costs for female agents and institutions and increased sexism among male agents. Confronting interpersonal bias by targets and allies attenuated the impact of structural bias. However, bias persisted even after a structural intervention to suddenly make previously unequal institutions equal (50% women) unless the probability of interpersonal bias-confrontation was further increased among targets and allies. This computational approach allows for comparison of various policies to attenuate structural equality, and informs the design of new experiments to estimate parameters for more accurate predictions.
BibTeX
@misc{momennejad2019computational,
  author = {Ida Momennejad and Stacey Sinclair and Mina Cikara},
  title = {{Computational Justice: Simulating Structural Bias and Interventions}},
  year = {2019},
  doi = {10.1101/776211},
  howpublished = {bioRxiv},
}
Figure