Periodic Reporting for period 3 - DYNNET (Opinion Dynamics)
Reporting period: 2021-12-01 to 2023-05-31
Our project studies opinion dynamics in social networks and groups with a particular focus on social identity. One subproject focuses on opinion dynamics in committees and the role of committee deliberation for gender biases. Gender biases have been documented in areas including hiring, promotion or performance evaluations. Many of these decisions are made by committees. Yet little is known about the role of committee deliberation. We experimentally investigate whether committee deliberation contributes to gender biases. There is substantial evidence of gender bias with open committee deliberation. In this case 60 percent of ratings received by men are revised upwards after deliberation compared to only 25 percent of ratings received by women. As a consequence women are ranked on average three positions lower after deliberation. We explore several mechanisms and test two interventions for open deliberation. Randomizing the order of speaking does not reduce gender bias, but an information intervention where raters are informed of gender bias in prior sessions does. These are important findings for anyone interested in reducing gender bias in committee decisions.
Another strand of research focuses on the role of beliefs and opinions in labour markets. It is well known that in the absence of complete information on worker characteristics, employers might systematically discriminate against workers from given identity groups. This is often referred to as "Statistical Discrimination" by Economists. Such statistical discrimination is typically modelled assuming that employers are Bayesian. However, a large literature shows that most people fail to update rationally. We use a model and two experiments to show that if employers are naive, in the sense of signal neglect, workers from disadvantaged groups will be discriminated against more often than when employers are rational. Employers overdiscriminate disadvantaged groups especially when signals are very informative. These results have important implications for policy and highlight the promise of affirmative action policies.
Another strand of research focuses on the role of beliefs and opinions in labour markets. It is well known that in the absence of complete information on worker characteristics, employers might systematically discriminate against workers from given identity groups. This is often referred to as "Statistical Discrimination" by Economists. Such statistical discrimination is typically modelled assuming that employers are Bayesian. However, a large literature shows that most people fail to update rationally. We use a model and two experiments to show that if employers are naive, in the sense of signal neglect, workers from disadvantaged groups will be discriminated against more often than when employers are rational. Employers overdiscriminate disadvantaged groups especially when signals are very informative. These results have important implications for policy and highlight the promise of affirmative action policies.
We found substantial evidence of gender bias with open committee deliberation and tested several mechanisms and test two interventions for open deliberation. Our main experimental treatments involve open committee deliberation. We find significant gender biases under open committee deliberation. After deliberation 60 percent of ratings received by men are revised upwards compared to only 25 percent of ratings received by women. As a consequence women are ranked on average three positions lower after deliberation. We tested two further interventions both designed to reduce gender bias in the presence of open deliberation. The first intervention randomized the order of speaking in the committee. This intervention was unsuccessful and in fact produced weakly larger gender biases compared to our baseline open deliberation treatment. The second intervention we tested is an information intervention, where participants are made aware of gender bias in previous sessions prior to entering their ratings. Similar interventions have sometimes been shown to be successful in non-committee decision-making. We also find that this intervention is successful. There is no gender bias in this treatment. These results carry potentially actionable policy consequences. Our interventions have shown that care must be taken when designing rules for committee deliberation. Changes designed to reduce bias, such as randomizing the order of speaking in a committee, can have unintended consequences and in our case led to very strong gender bias. On the other hand our information intervention was successful and did not lead to gender bias (neither against men nor women). We also did not find evidence that this intervention would lead to greater polarization of opinions.
In a subproject on statistical discrimination we use theoretical modelling and two experiments to study the implications of such failures of Bayesian rationality on discrimination in the labour market. Our research shows not only that wrong beliefs matter, but moreover that the source of these wrong beliefs is crucial for our understanding of discrimination and of workers' human capital investments. We show that naive employers discriminate against disadvantaged groups much more often than bayesian employers. Such irrational discrimination makes workers from the disadvantaged group less willing to pursue education, further exacerbating the initial inequalities. Compared to the Bayesian benchmark, we observe excess discrimination especially when signals are highly informative. An example where we encounter this pattern is women and computer science. There are many fewer women studying computer science than men and---in line with this pattern---women are perceived as ``on average worse'' in coding than men. However, conditional on having programming knowledge, there is suggestive evidence that women are better coders than men. These are the situations where Bayesians and non-Bayesians will make different decisions. In particular, since conservatives neglect the education signal, they will be less likely to hire the disadvantaged group than Bayesians. We refer to this type of discrimination as irrational statistical discrimination. We then design two experiments that allow us to test the intuitions developed in the theory. We find substantial evidence of conservatism, with a larger share of decisions being consistent with conservatism than with Bayesian reasoning. As a result, the disadvantaged group is hired around 52 percent less frequently compared to what we would expect if all employers were Bayesian. We also find that---conditional on their real productivity---workers from the disadvantaged group seek education much less frequently than others. This finding is important not only because market exit further exacerbates imbalances between the two groups and implies a substantial welfare loss due to the loss of high-skilled individuals in the work-force, but also because the fact that many high quality disadvantaged workers exit the market early makes it harder for employers to learn. We do indeed not find any evidence that the quality of employers' decisions improves over time.
In a subproject on statistical discrimination we use theoretical modelling and two experiments to study the implications of such failures of Bayesian rationality on discrimination in the labour market. Our research shows not only that wrong beliefs matter, but moreover that the source of these wrong beliefs is crucial for our understanding of discrimination and of workers' human capital investments. We show that naive employers discriminate against disadvantaged groups much more often than bayesian employers. Such irrational discrimination makes workers from the disadvantaged group less willing to pursue education, further exacerbating the initial inequalities. Compared to the Bayesian benchmark, we observe excess discrimination especially when signals are highly informative. An example where we encounter this pattern is women and computer science. There are many fewer women studying computer science than men and---in line with this pattern---women are perceived as ``on average worse'' in coding than men. However, conditional on having programming knowledge, there is suggestive evidence that women are better coders than men. These are the situations where Bayesians and non-Bayesians will make different decisions. In particular, since conservatives neglect the education signal, they will be less likely to hire the disadvantaged group than Bayesians. We refer to this type of discrimination as irrational statistical discrimination. We then design two experiments that allow us to test the intuitions developed in the theory. We find substantial evidence of conservatism, with a larger share of decisions being consistent with conservatism than with Bayesian reasoning. As a result, the disadvantaged group is hired around 52 percent less frequently compared to what we would expect if all employers were Bayesian. We also find that---conditional on their real productivity---workers from the disadvantaged group seek education much less frequently than others. This finding is important not only because market exit further exacerbates imbalances between the two groups and implies a substantial welfare loss due to the loss of high-skilled individuals in the work-force, but also because the fact that many high quality disadvantaged workers exit the market early makes it harder for employers to learn. We do indeed not find any evidence that the quality of employers' decisions improves over time.
Due to lab closures caused by Covid-19 epidemic, we still have to finalize data collection on some subprojects. By the end of the project we expect results on how network structure shapes opinion dynamics. We also hope to gain insight into the question when and whether social identity is a barrier to communication and when “opinion bubbles” are to be expected. Social identity can potentially affect social learning in different ways. If people disregard information from other identity groups they might not be able to sufficiently aggregate disperse information. However, in some circumstances, discrimination might be beneficial to social learning. This can be the case whenever social identity is a good proxy for expertise.