1 Introduction

Studies of labor market discrimination typically refer to preference-based and statistical discrimination. When discrimination is preference-based, it is rooted in preferences and based on group dislike, by animus or prejudice, involving discrimination against individuals, who would be effectively willing to pay the price of their choices (Becker, Reference Becker1957). By contrast, statistical discrimination relies on accurate or biased beliefs about the abilities or skills of the group an individual belongs to (Arrow, Reference Arrow2015; Phelps, Reference Phelps1972; Bordalo et al., Reference Bordalo, Coffman, Gennaioli and Shleifer2016).Footnote ¹ A lot of work has been done on this, both theoretically and empirically, to untangle the sources of discrimination.

To mitigate statistical discrimination, one might suggest providing ample specific information about individual candidates in order to avoid relying on average group information for important decisions such as hiring, promotions, and so on Guryan and Charles (Reference Guryan and Charles2013). By contrast, where preference-based discrimination is concerned, decreasing subjectivity in judgment by providing information on the quality of a specific individual has no effect because it influences beliefs but not tastes (Bohren et al., Reference Bohren, Imas and Rosenberg2019). Given that individuals engaging in preference-based discrimination seem willing to bear the costs involved in not selecting people from the group they dislike, increasing such costs could induce employers to forego their own preferences when making choices.

This paper replicates the carefully designed experiment by Coffman et al. (Reference Coffman, Exley and Niederle2021) who separated taste-based and statistical discrimination by comparing the choices of employers acting in one of two hiring environments (treatments), both characterized by the same ability distributions of workers in tasks on which men are found to outperform women on average, but only one allowing for gender-specific considerations. Moreover, we used this setting to vary the amount of employers’ financial compensation (either low or high) in a within-subject design to study whether increasing employers’ monetary costs eliminates preference-based gender discrimination in hiring.

Coffman et al. (Reference Coffman, Exley and Niederle2021) show that employers choosing between two workers with identical résumés are less willing to hire a worker from a group that performs worse on average, regardless of whether this group is defined by gender or by a non-stereotypical characteristic. Thus, discrimination against women is not specific to gender but statistical as it depends on beliefs about average group differences. However, the authors show that tastes do have a part to play: when the gender dimension is salient, females (males) are significantly more (less) likely to be hired than equally able males (females) when labeled as such than when labeled by means of their month of birth. Thus, while discrimination against women is statistical, preference-based considerations produce another type of discrimination that works in favor of women, disadvantaging equally able men (belonging to the better performing group). This sort of over-compensating behavior, inducing discrimination in favor of the discriminated group, has also been documented in other settings, for instance with immigrants (Alesina et al., Reference Alesina, Carlana, La Ferrara and Pinotti2018). Such preference-based discrimination in favor of females disappears when image concerns are reduced through the introduction of something that might excuse the hiring of a member of the non-preferred group, such as an apparent risk associated with hiring a female worker.

Our purpose in this paper is twofold. On the one hand, we replicated Coffman et al. (Reference Coffman, Exley and Niederle2021)’s experiment. We therefore selected a set of female workers born in even months and male workers born in odd months and studied how employers’ hiring decisions change when the labels used to describe the same workers are framed in terms of gender or month of birth. On the other hand, we used the same design to study whether increasing the monetary cost of relying on tastes or other non-monetary motives makes individuals less likely to do so, thus eliminating or at least reducing gender discrimination.

We found that when employers’ monetary incentives to employ more productive workers are as they were in Coffman et al. (Reference Coffman, Exley and Niederle2021), there is preference-based discrimination in favor of females whereby, controlling for beliefs about the average ability gap between the two groups, females are hired more often when defined as such than when they are labeled with their birth month. What we show that is new—thanks to our experimental design—is that preference-based discrimination, in terms of greater likelihood of hiring a female intentionally when labeled female instead of even-month worker, significantly reduces and is no longer present when employer’s monetary incentives to employ more productive workers are higher. Results point in the same direction for the probability of hiring a male worker intentionally (i.e., the reduction in this probability for a male versus an odd-month worker is lower when monetary incentives to employ more productive workers are higher), but the effect is smaller and less precisely estimated. We also find that the decision to leave it to chance is not significantly different between the gender and month-of-birth treatment with low incentives. Yet, when monetary incentives to employ more productive workers are higher, employers in the gender treatment opt to leave the decision to chance significantly more frequently than in the month-of-birth treatment. Thus, our results suggest that increasing employers’ financial compensation reduces the likelihood of them relying on preferences, especially in terms of hiring a female intentionally.

Moreover, our study shows that monetary incentives to employ more productive workers are a powerful tool to attenuate other broader motives in hiring decisions. Indeed, when we look at in-group preferences, namely at whether sharing the same identifier with an employee increases the likelihood of being hired (Tajfel et al., Reference Tajfel, Billig, Bundy and Flament1971; Chen and Li, Reference Chen and Li2009; Chen and Chen, Reference Chen and Chen2011; Chen et al., Reference Chen, Li, Liu and Shih2014), we found that, when monetary incentives to employ more productive workers are low, in-group preferences are significantly stronger in the gender treatment than in the birth-month treatment. By contrast, such reliance on in-group status in the gender treatment is significantly reduced and no longer present with higher monetary incentives. In contrast with Coffman et al. (Reference Coffman, Exley and Niederle2021), we found no evidence of in-group preferences in the birth-month treatment: employers born in even months do not hire even-month employees more than odd-month employers do.

The main takeaway for policymakers and organizations interested in limiting discrimination within the workplace is that the remuneration of individuals making employment decisions (human resource offices or paid hiring committees) should be contingent on employees’ performance. Further research should investigate whether monetary incentives to employ more productive workers also mitigate other (clearly identified) gender or race preference-based discrimination.

Concerning this last point, this article contributes to a recent current in economic literature which has found mixed results regarding the effect of financial incentives on various sorts of discrimination. Anukriti (Reference Anukriti2018), for instance, studied an Indian policy that offers financial incentives to families designed to lower the fertility rate and the gender ratio at birth (increasing daughters over sons). She found that financial incentives have a modest effect on reducing the fertility but do increase the sex ratio at first birth. In two randomized vignette studies, Underhill (Reference Underhill2019) varied the race of a hypothetical patient in need of a kidney transplant (black or white) and the financial incentives involved, showing that incentives encourage donation but introduce a significant bias favoring white patients. However, this “crowding in” of racial bias seems to be limited to medium-sized incentives, while large incentives may be so attractive as to override any bias effect. Hortaçsu et al. (Reference Hortaçsu, Hwang and Mathur2019) have estimated the effect of a policy in India that offers monetary incentives for inter-caste marriages. They found that a 10,000-rupee increase in the incentive raises exogamy between Scheduled Caste (historically disadvantaged caste) men and non-Scheduled Caste women by 4 percent in rural India.

The remainder of the paper is organized as follows: Section 2 describes our experimental design and the data. Section 3 presents the results and Sect. 4 concludes the paper.

2 Experimental design and data

Our hiring experiment follows Coffman et al. (Reference Coffman, Exley and Niederle2021)’s as regards treatments and employment settings but varies in the monetary incentives given to participants and the set of decisions they are required to make.

We conducted two preliminary studies to collect performance information for subjects to be used as the available “workers” for hire in the main hiring experiment. The first study asked just a few questions, including month of birth, separately to a female and a male pool, each made up of 150 people. Using the data from the first preliminary study, we then focused on two groups of workers: male workers born in odd months (79 subjects) and female workers born in even months (69 subjects). In a second preliminary study, these two groups of workers were invited to complete one easy and one hard three-minute math quiz, each consisting of 10 multiple-choice questions.Footnote ² Workers received 10 cents for each question answered correctly in a randomly selected quiz.Footnote ³ On average, males answered 6.47 questions in the easy quiz correctly (ranging from 3 to 10) and females 5.67 (ranging from 0 to 9).

In the hiring experiment, participants were “employers” asked to make incentivized hiring decisions on the available workers.Footnote ⁴ At the time of the decision, employers received information about the easy-quiz performances of the applicants but not about their hard-quiz performance, which would determine their earnings in the experiment. On top of the fixed show-up fee, employers earned a bonus that depended on the hard-quiz performance of the selected worker in a randomly selected decision.

The treatment is the presence of a gendered label, instead of the neutral label of the control group, used to describe the same workers, thus keeping their performance constant. Thus, depending on treatment, the same available workers were labeled with their own gender (males and females) or with their own month of birth (categorized as even or odd months). We collected data on a total of 210 employers, 100 for the gender treatment and 110 for the month-of-birth treatment.

The experiment was divided into two parts. At the end, subjects completed a short questionnaire.

In the first part, we elicited participants’ prior beliefs about the average ability gap (male/female gap and odd/even month gap)Footnote ⁵. Then, employers were asked to make nine hiring decisions between groups of workers (i.e., female-even-month vs male-odd-month) and information on worker performance was provided in the form of a bar chart comparing the distribution of the two groups.Footnote ⁶ In each decision, employers could choose whether to hire a randomly selected worker from the female-even-month group, the male-odd-month group or leave the hiring decision to chance (in this case the computer randomly determined which group to hire from). The option to leave the hiring decision to chance was introduced by Coffman et al. (Reference Coffman, Exley and Niederle2021) with the aim of allowing for expressions of indifference; the availability of this option should increase the likelihood that choosing one of the other two groups reflects a strict preference. If one of the decisions from this part was selected for payment, the hired worker received an additional 25 cents as bonus payment and the employer received 10 cents for each question answered correctly—on the hard quiz—by the worker hired in the randomly selected decision. Finally, we elicited posterior beliefs, using the same questions asked to elicit prior beliefs, in order to see how a given set of beliefs affected behavior in the gender and month-of-birth treatments.

Therefore, in both treatments, employers received accurate information about the distribution of easy-quiz performance across worker groups (female and male workers in the gender treatment; odd- and even-month workers in the month-of-birth treatment) prior to making their hiring decisions and forming their beliefs on the group average performance difference.Footnote ⁷ The information set did not change since the set of workers was the same. However, employers in the gender treatment associated the performance information with male and female workers. By contrast, employers in the month-of-birth treatment, were never told workers’ gender because the same workers were presented to them as two groups of workers, one born in odd months and the other in even months.

In the second part of the experiment, participants took two sets of nine hiring decisions between two specific workers, each belonging to a different group (male or female in the gender treatment; even- or odd-month in the birth-month treatment). For each hiring decision, employers knew the exact easy quiz performance of both applicants rather than knowing only the distributions of the two applicant groups as in Part 1, a design feature intended as a means of countering potential statistical discrimination. The displayed performances for the female-even-month and the male-odd-month workers for each decision were: 4-4, 5-4, 6-4, 7-4, 8-4, 6-6, 7-6, 8-6, 8-8.Footnote ⁸ Thus, individual performance was the same for both workers in three decisions and female-even-month workers did strictly better in the other six decisions. Considering only decisions in which the female-even-month workers weakly outperformed the male-odd-month workers enabled all decisions not to hire female workers to be classified as discrimination against women. Employers could choose whether to hire the male-odd-month worker, the female-even-month worker or leave the hiring decision to chance. If, for a given decision, employers chose to leave the decision to chance, the computer randomly determined which of the two workers (in that decision) to hire. If one of the decisions from this part was selected for payment, the worker hired received 25 cents.

Employers’ payment depended on the hard quiz performance of the worker they decided to hire. In the first set of nine hiring decisions in Part 2 of the experiment, the employer earned 10 cents for each correct answer by the hired worker in the hard quiz. This monetary incentive to employ more productive workers was also used in Coffman et al. (Reference Coffman, Exley and Niederle2021). In the second set of nine hiring decisions, the monetary incentives to employ more productive workers were higher: the employer earned 50 cents for each hard quiz question answered correctly by the hired worker. Thus, while the structure of the experiment up to this point replicated that of Coffman et al. (Reference Coffman, Exley and Niederle2021), this set of decisions was an addition aimed at studying the role played by monetary incentives.

Table 1 shows the descriptive statistics for our variables separately by treatment and, for the dependent variables, by the magnitude of the monetary incentives. As in Coffman et al. (Reference Coffman, Exley and Niederle2021), we focused our attention on the three decisions in which employers chose between workers with the same easy quiz performance.

Table 1

Descriptive statistics

	Mean	Std. dev	Min	Max
	(1)	(2)	(3)	(4)
PANEL A: gender
Low incentive
Female-even-month and chance	0.478	0.275	0	1
Female-even-month	0.13	0.337	0	1
Chance	0.697	0.46	0	1
Male-odd-month	0.173	0.379	0	1
High incentive
Female-even-month and chance	0.453	0.264	0	1
Female-even-month	0.097	0.296	0	1
Chance	0.713	0.453	0	1
Male-odd-month	0.19	0.393	0	1
Controls
Posterior (easy gap)	2.69	2.86	− 5	10
Posterior (hard-easy gap)	− 0.19	2.261	− 12	8
No. of obs	300
PANEL B: month of birth
Low incentive
Female-even-month and chance	0.332	0.286	0	1
Female-even-month	0.052	0.221	0	1
Chance	0.561	0.497	0	1
Male-odd-month	0.388	0.488	0	1
High incentive
Female-even-month and chance	0.344	0.295	0	1
Female-even-month	0.067	0.250	0	1
Chance	0.555	0.498	0	1
Male-odd-month	0.379	0.486	0	1
Controls
Posterior (easy gap)	3.827	2.434	− 3	10
Posterior (hard-easy gap)	− 0.282	1.671	− 6	6
No. of obs	330

The dependent variables are Female-even-month and chance taking the value of 1 if a female-even-month worker is hired intentionally, 0.5 if chance is chosen and 0 if male-odd-month is chosen, Female-even-month taking the value of 1 if a female-even-month worker is hired intentionally and 0 otherwise, Chance taking the value of 1 if chance is chosen and 0 otherwise, Male-odd-month taking the value of 1 if a male-odd-month worker is hired intentionally and 0 otherwise

When making each choice, employers could decide whether to select a female-even-month worker, a male-odd-month worker or leave the hiring decision to chance. We classified employers’ decisions across these three options using four methods. The first looks at the choice of a female-even-month worker in terms of the outcome of the choice (i.e., not only strict preference for female-even-month workers but also the likelihood that these are randomly hired via the chance option). The corresponding dependent variable is Female-even-month and chance, taking the value of 1 if a female-even-month worker was hired intentionally, 0.5 if chance was chosen and 0 if male-odd-month was chosen, and it was the main dependent variable used in Coffman et al. (Reference Coffman, Exley and Niederle2021). The other three dependent variables, on the other hand, consider employers’ preferences and are represented by strict preference for one of the two groups and indifference. Specifically, these are: Female-even-month, taking the value of 1 if a female-even-month worker was hired intentionally and 0 otherwise; Chance, taking the value of 1 if chance was chosen and 0 otherwise; Male-odd-month, taking the value of 1 if a male-odd-month worker was hired intentionally and 0 otherwise. The same classification was also used by Coffman et al. (Reference Coffman, Exley and Niederle2021) in the online appendix.

We saw that on average female workers were hired, intentionally and by chance, in 47.8% of the decisions when there were low monetary incentives to employ more productive workers and in 45.3% of the decisions when incentives were higher; both figures are significantly below the 50% benchmark (p value = 0.087 and 0.001, respectively).Footnote ⁹ However, this discrimination is not specific to gender. Indeed, even-month workers were hired in just 33% of low incentive decisions and 34% of higher incentive decisions; both figures are significantly below the 50% benchmark (p value = 0.000). Thus, as in Coffman et al. (Reference Coffman, Exley and Niederle2021), we found evidence of statistical discrimination because employers in both treatments preferred to hire workers associated with the higher-performing group.

On closer examination, we can see that in the gender treatment employers left the hiring decision to chance in about 70% of choices, while females were hired intentionally only with a probability of 13% with low monetary incentives to employ more productive workers and 9.7% with higher incentives. Males were hired intentionally in about 17% and 19% of decisions, respectively. In the month-of-birth treatment (Panel B) the probability of hiring an even-month worker intentionally was considerably lower (5.2% and 6.7% with low and high incentives, respectively), the probability of leaving the outcome to chance was about 56% and the probability of selecting an odd-month worker intentionally was higher (38.8% and 37.9%, respectively).

As in Coffman et al. (Reference Coffman, Exley and Niederle2021), we included employers’ posterior beliefs as controls—after observing the distribution of abilities in the easy quiz for the two groups of workers—of the average performance gap in the easy quiz, Posterior (easy gap), which on average favors male workers (2.69 in the gender treatment and 3.8 in the month-of-birth treatment) and a difference between the posterior beliefs of the average performance gap in the hard quiz and in the easy quiz, Posterior (hard-easy gap), which is on average close to zero ( $-$ 0.19 and $-$ 0.28 in the gender and birth-month treatment, respectively). Finally, as in Coffman et al. (Reference Coffman, Exley and Niederle2021), we found that employers initially believed that male workers outperform female workers by a significantly larger margin than odd-month workers outperform even-month workers0 (p value < 0.01 in both quizzes). After taking the decisions in Part 1, posterior beliefs in both treatments point to a greater advantage of male-odd-month workers and the reported performance gap is slightly greater for employers in the month-of-birth treatment. By including posterior beliefs in our controls, we estimated gender-driven discrimination on top of stated beliefs.