Can conversations about minorities reduce discrimination? Evidence from anti-transgender discrimination in Chennai

Discriminatory behaviour is known to adversely impact equity and efficiency in a range of economic and social domains. In the context of discrimination against transgender people in urban Chennai, this article examines whether discrimination may be reduced by peer-to-peer communication among majority-group members about a minority. It finds that such communication has significant positive effects, driven by group members persuading one another to be more pro-transgender.

Discrimination is costly as it causes distortions in many economic domains – from school admissions and hiring decisions, to the ability to rent a home (Hjort 2014, Hsieh et al. 2019, Rao 2019, Lowe 2021, Christensen and Timmins 2023). It is normally considered to be quite hard to change, stemming from deep-seated prejudice or stereotypical negative beliefs about minorities. But sometimes societies do shift towards less discriminatory preferences: people have become more accepting of same-sex marriage, inter-ethnic marriage (Ford 2014), or women’s rights (Chingwete et al. 2014) over the course of less than a generation. These changes often coincide with changes in the way people communicate with each other about a minority, suggesting that there might be a causal relationship – that is, the way people talk about a minority might drive rapid reductions in discrimination.

In new research (Webb 2024), I examine whether ‘horizontal communication’ – peer-to-peer conversations among majority-group members about a minority – can reduce discrimination. Why might such communication help? Perhaps through virtue signalling: if people do not want to appear to be discriminatory, they might communicate in a pro-minority manner in group settings and persuade others to discriminate less. Another channel focuses on who communicates in a group: if those who are more supportive of a minority are particularly vocal, they may persuade others to discriminate less.

I ran a field experiment in 2023 in urban Chennai in Tamil Nadu involving a sample of 3,397 individuals, examining discrimination against one of the most visible LGBTQ+ (lesbian, gay, bisexual, transgender and queer) groups in India: a community of transgender women called thirunangai. Their social identity has been historically marginalised in the country, and they are both visually recognisable and vulnerable to extensive economic discrimination and violence. There are at least 490,000 thirunangai in India, and they are stereotypically associated with begging and sex work, making it particularly difficult for them to find other forms of work – as indicated by very low rates of formal employment among this group. They also face other forms of discrimination, such as rejection by family, difficulty in obtaining housing, exclusion from medical services, and police harassment. Despite this discrimination, there is some evidence of nascent social change towards greater acceptance (Thomas 2017), making this an apt setting in which to study the effects of horizontal communication on discrimination.

Measuring how communication affects discrimination

In my experiment, I measure discrimination against transgender workers by offering participants a free grocery delivery and asking them to make a series of binary choices over which worker carries out the delivery, after seeing photos of the workers, along with the items they receive, which are randomly varied across choices. All of the participants are non-transgender and were recruited using door-to-door canvassing in Chennai. Using the random variation in the items on offer across choices, I can evaluate the trade-off participants make between workers and item value. Following economists’ standard definition of discrimination, I define it as the difference in the probability of choosing a transgender worker compared to a non-transgender worker, holding fixed the other attributes of the worker and items.

Then, to measure the effects of horizontal communication between participants, I randomise whether participants are earlier involved in a discussion with two of their neighbours (see Figure 1). In this discussion, participants are also shown a series of hiring options and asked to make collective choices over which options they prefer. Some of the options include transgender workers, and participants discuss whether they want to hire transgender workers. Crucially, the only communication about transgender people in this discussion comes from the participants themselves, rather than from the discussion facilitator. Moreover, the purpose of the experiment was hidden from participants to avoid response bias: only 8% of respondents guessed that the study was about discrimination, and these respondents do not drive the results.

I also compare this to the effects of another form of ‘top-down communication’ from a societal authority – specifically, communication about the legal rights of transgender individuals in India. To measure this, I show some participants a video at the start of the survey that informs them about an Indian Supreme Court ruling that affirmed the fundamental rights of transgender persons.

Figure 1. Experimental design

The effect of communication

I first document that participants are highly discriminatory when making choices in private, on average. In the control group, which does not engage in any communication, participants are 19 percentage points (32%) less likely to hire transgender workers than non-transgender workers (Figure 2, left panel). Their choices imply that they are willing to sacrifice grocery items worth 1.9 times the median daily food expenditure to avoid interacting with a transgender worker.

Figure 2. Effects of communication on probability of hiring a transgender worker

Note: (i) Thick bars are 90% confidence intervals, thin bars are 95% confidence intervals. A confidence interval is a way of expressing uncertainty about estimated effects – specifically, it means that if the study was repeated over and over with new samples, 95% of the time the calculated confidence interval would contain the true effect. (ii) “Obs” indicates the number of observations.

However, horizontal communication between these discriminatory participants substantially reduces discrimination. The effects of the group discussion on discrimination are stark: even when people are making private, individual choices after the discussion has ended, they are 17 percentage points (42%) more likely to select a transgender worker than the control group (who were not involved in a discussion) (Figure 1, right panel). The level implies that there is no significant anti-transgender discrimination, on average, in the discussion arm.

The discussion’s effects are substantially larger than a the top-down communication about the legal rights granted to transgender people by the Supreme Court. This communication does reduce discrimination, but by only 59% as much as the discussion. The discussion’s effects are also partially persistent: when I re-survey participants around one month later, discussion participants are still 5 percentage points more likely to select transgender workers in a series of hypothetical hiring choices. While the effects are persistent, they are also substantially attenuated, implying that this light-touch 10-minute discussion is not sufficient to generate large changes in behaviour over the medium run. Such changes would likely require more intensive, repeated interventions.

Mechanisms for reducing discrimination

It is noteworthy that generating communication between privately discriminatory individuals can sharply reduce discrimination post-communication. How does this occur? My additional results suggest that two channels – namely, correcting a misperceived norm and virtue signalling – are not sufficient for explaining the effects. Instead, the results seem to be driven by a third channel – persuasion. The three possible channels are discussed below:

i) Correcting a misperceived norm: One plausible explanation would be that participants initially overestimate how discriminatory their neighbours are (as in Bursztyn et al. 2020). Then, when they communicate, this misperception is corrected, so everyone feels more comfortable selecting a transgender worker. But this pattern is not enough to explain the large effects of the intervention. Although people in the control group do overestimate how discriminatory their neighbours are (by 5 percentage points), the reduction in predicted discrimination generated by the discussion is far larger (24 percentage points) than the initial misperception. This suggests that a simple correction of a misperception could only account for around 21% of the effect of the intervention.

ii) Virtue signalling: The results also do not seem to be explained by a simple story of virtue signalling. If participants have social image concerns and do not want to appear discriminatory in a group setting, they may act more favourably towards transgender persons in a group setting, in turn persuading others to be more pro-transgender. I test this channel using an additional ‘public’ variation in which participants do not discuss with each other, but instead make individual hiring choices that they know will be revealed to other members of their group. This has no effect on discrimination, suggesting that virtue signalling alone cannot drive pro-transgender behaviour in the absence of horizontal communication.

iii) Persuasion: By contrast, persuasion does seem to explain the effects of the discussion. To show that persuasion occurs, I add a variation in which one participant silently listens to two other people who have a discussion. The effect on these ‘listeners’ is just as large as on people who actively participate in a discussion. The effects therefore seem to be driven by persuasion among participants, rather than participants persuading themselves in the process of producing arguments about who to hire.

But why do participants persuade each other to be more pro-transgender, rather than more anti-transgender? I suggest that this is because pro-transgender participants are more vocal in the discussions. They are more likely to speak first and dominate a discussion, specifically when faced with a choice involving a transgender worker. On net, therefore, participants are persuaded to discriminate less. They also use highly pro-social arguments: participants often explicitly say that it is important to help transgender people, to give them an opportunity to work, or to not discriminate.

Leveraging the insights

Under the right conditions, horizontal communication between discriminatory individuals can substantially reduce discrimination. I show this can occur when there are persuasive arguments against discriminating, and a subset of people who are willing to speak out in favour of a minority even within a discriminatory group. While I focus on a specific setting – anti-transgender discrimination in India – further research can examine how such results apply in other contexts.

Nevertheless, the study provides a proof-of-concept that could be applied to policies. First, we could design and evaluate policies that create discussions at scale to change anti-minority attitudes. Previous work shows that it may be possible to influence discriminatory attitudes by leading discussion-based interventions in schools (Dhar et al. 2022), or by door-to-door canvassing (Kalla and Broockman 2020). My results suggest that discrimination can be reduced without even having to lead a discussion; instead, just creating a scenario where minorities are naturally discussed might be sufficient.

Second, we could encourage group-based decision-making, rather than individual decision-making, in high-stakes environments where discrimination takes place. For example, processes for hiring, college admissions, and housing allocation may avoid the substantial costs of discrimination if they are based on committees’ rather than individuals’ decisions.

I4I is now on Substack. Please click here (@Ideas for India) to subscribe to our channel for quick updates on our content.