Statistical discrimination in labor markets: an experimental analysis.

• Author: Dickinson, David L.

1. Introduction

   When membership in a particular group conveys valuable information about an individual's skills, productivity, or other characteristics, a nonprejudiced agent may still find it rational to statistically discriminate. Examples of statistical discrimination include wage or hiring decisions in labor markets, racial profiling in law enforcement, determinants of loan approval rates, voting the party ticket in elections, or differential premiums for insurance, among others. In some settings statistical discrimination is legal and acceptable (for example, insurance rates); whereas, in others it is controversial and/or illegal (for example, racial profiling and employment discrimination). Existing research has focused on first-moment statistical discrimination: that is, discriminatory wage offers to females or lower loan approval rates for minority applicants are based on average productivity and default rates, respectively. Agents attribute average group characteristics to each individual from that group when it is costly to gather information.

   In this article we explore the possibility that statistical discrimination extends beyond differential treatment based on average group characteristics. Specifically, discrimination may also exist if agents base decisions on productivity distribution risk (or default rates, accident rates, etc.). Using labor markets as an example, risk-averse employers may make lower wage offers to females if their productivity variance is believed to be higher, even though average productivity may be identical to that of males. If such variance-based statistical discrimination is empirically documented, then existing measures of statistical discrimination are biased, and measures of prejudiced-based discrimination may be overstated. Some have found field evidence of statistical discrimination based on higher-order moments of a distribution (Ayers and Siegelman 1995; Goldberg 1996; List 2004), but we also recognize that statistical variance may not be the only behaviorally important measure of distributional risk. Thus, we contribute additionally by examining the support of the productivity distribution (Tversky and Kahneman 1973; Curley and Yates 1985; Griffin and Tversky 1992; Babcock et al. 1995) and the possibility for loss (Kahneman and Tversky 1979) as two other potentially important cognitive assessments of risk. In short, some discrimination labeled as personal prejudice or taste based may really be just a different form of statistical discrimination than what is typically examined.

   We report results from a controlled laboratory experiment in which subjects are engaged as employers and workers in a laboratory double-auction labor market. We choose a labor market context for our experiments for several reasons. First, we believed that because payoffs to employer-subjects are determined by an outcome variable minus a contract payment to another subject, the labor market context would aid subjects in understanding payoffs in our experiment. Second, statistical discrimination is very relevant to labor markets, highlighted by the many existing empirical studies of statistical discrimination that examine labor markets. Finally, the use of labor market context in a competitive double-auction market environment is a logical context with precedence in the experimental economics literature (Fehr et al. 1998). That said, the insights we gain from our data extend to other contexts, and the implication of our results is that statistical discrimination may be more pervasive than previously thought. Our results show that subjectemployers make significantly lower wage offers when the probability of loss is greater, and this measure of risk mattered more than the statistical variance or support distribution.

   Statistical theories of discrimination have been advanced by Arrow (1972), Phelps (1972), Aigner and Cain (1977), and Lundberg and Startz (1983). Some studies base statistical discrimination on noisier productivity signals for certain worker groups, while others base it on imperfect or incomplete information. (1) Most researchers advance theories that depend on differences in average productivity characteristics; although, others note that statistical discrimination need not be based on differences in average productivity (e.g., Aigner and Cain 1977; Curley and Yates 1985). For risk-averse individuals, it seems clear that a less risky outcome distribution would be preferred to a more risky distribution; although, "risk" may be defined in ways other than just a statistical variance, as we note. Empirical evidence alluding to statistical discrimination can be found in a variety of settings; although, it is often difficult to identify taste-based versus statistical discrimination (see discussion in Arrow 1998). Probably the only easily observable forms of statistical discrimination are the legal forms, such as those found in the insurance industry. In labor markets there is some direct evidence from employer interviews that race is used as a proxy in employment decisions (Wilson 1996). Neumark (1999) uses field data to uncover discrimination not based on productivity characteristics, but Altonji and Pierret (2001) find little evidence for statistical discrimination based on race. (2)

   Given identification and causation issues inherent in field data examinations of discrimination, some have used controlled experiments to study statistically based discrimination resulting from imperfect information (Anderson and Haupert 1999), asymmetric information (Davis 1987), or ethnic stereotypes (Fershtman and Gneezy 2001). We employ a full information environment to examine higher-order statistical discrimination and to explore which of several risk measures is more behaviorally important. Our design is such that causation can go in only one direction (that is, exogenous wage distributions imply that wage contracts cannot affect future worker productivity), and the market institution for determining wage contracts is one that produces strong convergence to the competitive equilibrium prediction. Nevertheless, we find evidence for statistical discrimination based on one important measure of worker risk.

2. Experimental Design

   We implement a two-sided auction market design to simulate a labor market. Specifically both employers (buyers) and workers (sellers) negotiate in an open-pit fashion, with no central auctioneer. Workers are more plentiful than employers, and so there is an equilibrium level of "unemployment" in this design. Both supply and demand for labor are induced upon the experimental subjects using standard experimental techniques, discussed in Smith (1982). (3)

   The baseline design we use is simple in that it generates clear equilibrium predictions. Specifically the demand side of the experimental market consists of five employers, each capable of hiring one unit of labor in each experimental market round. The productivity of a unit of labor in the baseline (treatment 1) is certain and fixed at three units of output (each unit of output sells for SI experimental), and so the demand for labor is perfectly elastic at $3.00 up to five units of labor. The supply side of the market consists of 10 workers, each with a reservation wage of S0.40, and each is able to sell at most one unit of labor services in each experimental market round. As such, the supply curve is perfectly elastic at SO.40 up until 10 units of labor. The predicted market wage is S0.40, and the predicted market quantity of labor traded is five units. We used the labels "worker," "employer," and "wages" to facilitate the subjects' understanding of the connection between productivity and final payoff, but it was clear to all subjects that no labor task would be completed in the experiment. In this way we maintain strict control over productivity in the experiment. Figure 1 shows the experimental design graphically.

   The baseline experimental design is quite similar to that used in Smith (1965), though Smith does not use a labor market context. That is, at the predicted equilibrium the entire market surplus is allocated to one side of the market (the buyers of labor). In our design the employers are not given information on worker reservation wages, and workers are not informed as to the value (to employers) of a unit of output. Payoff information is therefore private to each subject, as in Smith (1965), who shows that, even when market surplus at equilibrium is designed to be extremely imbalanced, this trading institution produces strong convergence of equilibrium prices to the competitive equilibrium prediction. Any evidence of statistical discrimination in the uncertain productivity treatments would then be significant, given the strong competitive tendencies inherent in our baseline design.

   [FIGURE 1 OMITTED]

   The stochastic or uncertain productivity treatments are labeled treatments 2, 3, and 4. The difference across these uncertain productivity treatments lies in the particular (known) productivity distribution for the labor pool. After hiring a unit of labor in an uncertain productivity treatment the employer discovers the realized productivity of that unit of labor by means of an ex post random draw. Specifically, in treatment 2, productivity of the labor pool is either one, two, three, four, or five units of output with probability 10%, 10%, 60%, 10%, and 10%, respectively. Productivity is determined by a random draw from a Bingo cage, and an independent draw is conducted for each employer who hires a unit of labor. Though wage contracts are made with a specific experimental subject in any given trading round, it is made clear that productivity draws are independent of the actual worker-subject (that is, you cannot contract in the next round with John Doe to ensure productivity of five just because it happened to turn out that way in the current or past rounds when contracting with John Doe). The...