Why labour market experiments?Advantages of laboratory experimentsObjections to laboratory experimentsAdding realismThe future of experimentationReferencesWhy labour market experiments?Armin Falk*, Ernst FehrInstitute for Empirical Research in Economics, University of Zurich, Blu¨mlisalpstrasse 10,CH-8006 Zurich, SwitzerlandAbstractOver the last decades, there has been a steady increase in the use of experimental methods ineconomics. We discuss the advantages of experiments for labour economics in this paper. Control isthe most important asset behind running experiments; no other empirical method allows a similarlytight control as do experiments. Moreover, experiments produce replicable evidence and permit theimplementation of truly exogenous ceteris paribus changes. We also discuss frequent objections toexperiments, such as a potential subject pool bias, the stake levels used in experiments, the numberof observations as well as internal and external validity. We argue that although these objections areimportant, careful experimentation can circumvent them. While we think that lab and fieldexperiments offer a very valuable tool, they should not be viewed as substitutes but as complementsto more traditional methods of empirical economic analysis.D 2003 Elsevier Science B.V. All rights reserved.Keywords: Experiments; Labour economics1. Advantages of laboratory experimentsScientific progress relies crucially on the testing of theories. The researcher hasdifferent data sources available for performing such testing. Theses sources can be roughlyclassified along two dimensions (Friedman and Sunder, 1994). We can distinguishhappenstance from experimental data, and field data from laboratory data. Happenstancedata is the by-product of uncontrolled, naturally occurring economic activity. In contrast,experimental data is created explicitly for scientific purposes under controlled conditions.Field data is data from natural environments while laboratory data comes from labenvironments. The two distinctions allow for four combinations, and all of these are used0927-5371/03/$ - see front matter D 2003 Elsevier Science B.V. All rights reserved.doi:10.1016/S0927-5371(03)00050-2* Corresponding author. Tel.: +41-1-6343722; fax: +41-1-6344907.E-mail addresses: [email protected] (A. Falk), [email protected] (E. Fehr).www.elsevier.com/locate/econbaseLabour Economics 10 (2003) 399 – 406in economics. In labour economics, the data most commonly used is field happenstancedata such as unemployment rates or data on wages, education, or income.Given the richness of happenstance field data, why should one bother creating own dataand performing experiments? Performing an experiment is, after all, time-consuming andcostly, as the participants of an experiment are paid based on their decisions. Let usillustrate the potential advantages of lab experiments with the help of a concrete example,the testing of tournament theory. Tournament incentives mean that workers compete for aprize, just as in a sport tournament. The worker with the higher performance gets the prize,e.g., a higher salary or a promoti on. According to tournament theory, the workers’equilibrium efforts should be chosen in such a way that marginal effort costs equalmarginal gains. The latter depend on the level of the prize and inversely on the importanceof chance for getting the prize (Lazear and Rosen, 1981). Given the equilibrium effortchoices, the calculation of the optimal prize is straightforward.A direct empirical test of this theory requires that the researcher knows the number ofworkers who compete for the prize, the effort cost functions of the workers, the exact levelof the prize, and the production function including the nature of the error term. Knowledgeof the payoff function of the firm and the participation constraints of the workers is furthernecessary for the determination of the optimal prize level. All of these features can beimplemented in a lab experiment and are therefore known by the researcher. Consequently,it is possible to derive a precise prediction and to test this prediction by observing effortand prize choices of subjects who participate in the experiment. With happenstance fielddata, conducting such a direct test seems impossible. The researcher never knows any ofthe ingredients mentioned above with a sufficient level of confidence when using fielddata. Moreover, he cannot be sure that the environmen t in which the interaction takes placeis similar to that assumed by the theory. For example, it is well known in theory that if theworkers can engage in sabotage activities, both the optimal level of effort as well as of theprize are different from a situation where sabotage is not part of the workers’ strategy set.By the same token, the researcher who uses field data knows only little about whether theinteraction between workers and the firm is one-shot or repeated. In repeated games,however, theoretical predictions are usually quite different from those derived in one-shotsituations. Other factors may be important too. How intensely do workers know eachother? Is a lot of communication and peer pressure going on, making collusion morelikely? Or is the interaction rather anonymous? These and many other details of theenvironment affect behaviour in the field in an uncontrolled manner. In a lab experiment,on the other hand, these environmental factors can be controlled and systematicallystudied. It is easy to study a treatment where subjects interact only once in an experimentand to contr ast this treatment with a condition where workers interact repeatedly. Like-wise, it is possible to study anonymous vs. face-to-face interactions. The study of theimpact of sabotage activities on effort choices in a controlled manner is also possible.The existence of superior controls is, of course, not restricted to the study of tourna-ments. In the debate about inter-industry wage differentials, for instance, it has beennotoriously difficult to judge whether the observed differentials reflect true rents orwhether they are simply the result of unobserved heterogeneity (Gibbons and Katz, 1992).Likewise, it seems alm ost impo ssible on t he bas is of field data to ju dge whe therunemployment is involuntar y or voluntary, i.e., whether unemployed workers would inA. Falk, E. Fehr / Labour Economics 10 (2003) 399–406400fact be willing to work for less than the going wage . In contrast, it is possible to controlworkers’ outside