Profit, morality and discrimination

ABSTRACT

Using an original vignette survey, we study the normative acceptability of the trade-off between immoral profit (discrimination) and costly morality (non-discrimination). We test the causal influence of three factors: i) the origin of discrimination, ii) the steepness of the morality/profit trade-off and iii) anti-discriminatory moral injunctions. Contrasting with past experimental and attitudinal studies, we find that a significant minority of respondents believe that labour market discrimination is acceptable when morality results in profit loss. We also find that the three tested factors have significant effects on normative opinions. Respondents are more likely to choose profit over morality when discrimination is taste-based than when it is caused by imperfect information. Discrimination’s acceptability rises with the cost of non discrimination. Anti-discriminatory moral injunctions sharply reduce the acceptability of profitable discrimination.

KEYWORDS:

• Discrimination
• profit/morality trade-off
• moral suasion
• factorial survey
• JEL classification: J23,J7,J78,C9

Highlights

• vignette-based survey experiments reveal causal effects on normative preferences

• normative opinions on discrimination are affected by profit considerations

• customer taste discrimination is more acceptable than screening or statistical discrimination

• moral suasion effects have an impact on normative choices between profit and morality

Acknowledgments

Warm thanks to Sophie Harnay and Régis Lanneau for their help during the administration of survey. Thanks to the participants of the Nouméa workshop on discrimination, Frédéric Chantreuil, Isabelle Lebon and Dominique Meurs. Thank you also to the two EconomiX referees, and this journal’s anonymous referee.

Disclosure statement

We declare that this manuscript is original, has not been published before and is not currently being considered for publication elsewhere. We know of no conflict of interest associated with this publication.

This work has been funded by the 2015 JCJC ANR Grant of the French Agence Nationale de la Recherche.

Notes

¹ In this paper, we narrow the focus on labour market discrimination, although similar mechanisms are at play on other markets, such as the housing of the credit market.

² Recently, Barr, Lane, and Nosenzo (2018) found that 1 out of 6 participants in a lab experiment discriminated against individuals who did not belong to their group when the groups were nationality-based. This rate was 3 out of 10 when the groups were artificial (randomly drawn).

³ For recent methodological surveys on vignette-based factorial surveys , see Taylor (2006), Wallander (2009), Atzmüller and Steiner (2010), and Evans et al. (2015).

⁴ Recent papers feature the topic of paid participation in ‘repugnant transactions’ such as clinical trials, (Ambuehl et al., 2015; Ambuehl and Ockenfels, 2017) and of the overconfidence of of health workers and the quality of provided healthcare (Kovacs et al. 2020).

⁵ Although some papers on discrimination have chosen the opposite protocol. For example, in Kübler, Schmid, and Stüber (2018) and Terum, Torsvik, and Øverbye (2018), managers were shown a set of experimentally manipulated resumes and were asked to tell which person, in the real world, they would have chosen to hire.

⁶ Implicit association tests developed in social psychology show that such discriminatory preferences can be unconscious (Greenwald et al., 1998), and that discriminatory behaviours may be influenced by contextual effects (Devine, 1989; Bertrand, Chugh, and Mullainathan 2005).

⁷ Paris-Descartes, Paris-Nanterre and Nouméa.

⁸ This is standard practice in normative questionnaire-experiments (see Gaertner and Schokkaert 2011). Since the point is to collect opinions on moral issues, incentivizing respondents would likely create a huge bias the results, a fortiori if one of the alternative principles tested is profit-based.

⁹ See Appendix B for the full text of the vignette.

¹⁰ Having three equal-sized ethnic groups allows to neutralize the employer’s potential in-group favouritism: having three groups means that the employer can belong to a race not represented among the applicants, and have no intrinsic reason to prefer one race of applicants over the other. Equal-sized groups also mean that no race is demographically dominant, so we are able to leave out minority/majority and dominated/dominant issues.

¹¹ The setting on a faraway planet was driven by the repeated success of science fiction literature, TV and movie franchises within our respondents’ age demographics.

¹² Encompassing all the channels through which moral injunctions can transit (for example through vote-issued laws or generally accepted social norms) is beyond the scope of this paper. Using a universally respected leader, we bypass the need to specify a social choice procedure. Moreover it allows us to skip the issue of whether the scenario’s protagonists got to vote for and/or agree with the law, and to establish that all of them (employers, applicants from the discriminated group and others) abide by the leader’s decisions. Using a leader is also a handy way to avoid dealing with the plausibility of anti-discriminatory general social norms when our scenarios depict situations where a race is discriminated against.

¹³ The question was not how many applicants the respondents would themselves hire if they were in the employer’s place. Normative survey experiments provide information on preferences, and are not designed to ask about hypothetical behaviours. To avoid any confusion in the respondent’s mind, we stated that ‘your opinion [as a respondent] on what is just will have no effect on the employer’s actual hiring decision’.

¹⁴ We ran robustness checks based on alternative categorizations. First, we changed the threshold for the ‘no support of discrimination’ category, classifying respondents as not supporting discrimination when they answered that 5 members of the target group or more should be hired (instead of ‘4 or more’). Second, we split the ‘no support of discrimination’ category in two to be able to differentiate between the respondents who supported a strict egalitarian choice (exactly 5 Tenkas should be hired) from those who supported a positive discrimination choice (6 and more members of the discriminated group should be hired). We found only small differences across these specifications, with no modification of the findings presented in this paper. The results using this alternatives specifications are available upon request. The paper’s results are robust to alternative definitions of the strong, weak and ‘no discrimination’ categories (evidence available upon request). The specification presented here (no discrimination = 4 Tenkas or more) is one with the better Akaike Information Criterion score.

¹⁵ In this paper, we use either the Kruskal-Wallis or the Jonckheere-Terpstra test to test the difference in the acceptability of discrimination. Note that several tests are available to test the relation between two qualitative variables or to test if a nominal outcome differs between k-groups: Chi-square test (Pearson, 1900), Kruskal-Wallis test (1952) and Jonckheere-Terpstra test (1952, 1954). The chi-square test is recommended when both the outcome and the k-populations are non-ranked. The Kruskal–Wallis test is more powerful (higher probability that the test will reject the H0 when the H1 is true) when the outcome is ranked but not the populations. The Jonckheere-Terpstra test is more powerful when both the outcome and the populations are ordered. This a priori ordering of the k populations is due to the intensity of treatment (for more information, see Agresti, Mehta, and Patel 1990).

¹⁶ Jonckheere-Terspstra test rejects the null hypothesis of no dependence between the variables with an error of 1.6%. We find no significant difference between the median profit loss (50%) and either of the two extreme losses (25% or 75%).

¹⁷ The p-value of the Jonckheere-Terspstra test is 0.35 if discrimination is justified by screening issues and 0.23 if it is justified by statistical discrimination.

¹⁸ We thank the anonymous referee to outline this point.

¹⁹ We use the command predictnl in Stata to implement this method.

²⁰ Due to sample size issues, interaction effects between individual characteristics could not be tested; nor was it possible to replicate the results on gender- or minority-based sub-samples.

Additional information

Funding

This work was supported by the Agence Nationale de la Recherche [JCJC2015].