Post your analysis of three aspects of motivation
American Economic Review 2013, 103(4): 1325–1359 http://dx.doi.org/10.1257/aer.103.4.1325
1325
The Lure of Authority: Motivation and Incentive Effects of Power†
By Ernst Fehr, Holger Herz, and Tom Wilkening*
Authority and power permeate political, social, and economic life, but empirical knowledge about the motivational origins and consequences of authority is limited. We study the motivation and incentive effects of authority experimentally in an authority- delegation game. Individuals often retain authority even when its delegation is in their material interest—suggesting that authority has nonpecuniary consequences for utility. Authority also leads to over- provision of effort by the controlling parties, while a large percentage of subordinates underprovide effort despite pecuniary incentives to the contrary. Authority thus has important motivational consequences that exacerbate the inefficiencies arising from suboptimal delegation choices. (JEL C92, D23, D82)
Authority and power play an important role in human societies. Influential schol- ars from various social science disciplines—such as Marx (1867), Russell (1938), Parsons (1963), Dahl (1957), and Weber (1978)—have contributed to our under- standing of the origins, characteristics, and potential consequences of these forces.
Despite some notable early exceptions (Simon 1951; Zeuthen 1968; Harsanyi 1971; Bowles and Gintis 1988), the study of authority and power has not been a major focus in economics. More recently, however, organizational economists have taken interest in the incentive effects of decision rights by studying situa- tions where one party has the contractual right to make decisions that influence another party’s payoffs and potential choices (Grossman and Hart 1986; Hart and Moore 1990; Aghion and Tirole 1997; Baker, Gibbons, and Murphy 1999;
* Fehr: Department of Economics, Zurich University, Bluemlisalpstrasse 10, CH-8006 Zurich (e-mail: ernst. fehr@econ.uzh.ch); Herz: Department of Economics, Zurich University, Bluemlisalpstrasse 10, CH-8006 Zurich (e-mail: holger.herz@econ.uzh.ch); Wilkening: Department of Economics, The University of Melbourne, Economics and Commerce Building, Victoria, 3010, Australia (e-mail: Tom.Wilkening@unimelb.edu.au). We are grateful to Sam Bowles, Wouter Dessein, Jon Elster, Bob Gibbons, Sergei Izmalkov, Bentley MacLeod, Nikos Nikiforakis, Klaus Schmidt, Chris Smith, seminar participants at the European University Institute in Florence, the London School of Economics, Bocconi, Harvard, MIT, NYU, Zurich, participants at the ESA International Meetings 2008 at Caltech, participants at the 24th annual congress of the European Economic Association, the sec- ond Australian Workshop for Experimental Economics, the second Behavioral Economics Conference at Cornell, and three anonymous referees for helpful suggestions and feedback. We are particularly grateful to a fourth ref- eree whose careful reading of our work greatly improved the article. We acknowledge the generous support of the Swiss National Science Foundation (Grant No. 100014_130127/1 on “The Social Dynamics of Normative Behavior”). The project was conducted while Tom Wilkening was at MIT and received human subjects approval from MIT’s institutional review board (COUHES No. 0703002170). The project also received the approval of the Institutional Review Board at the faculty of economics, business administration, and computer science of the University of Zurich.
† To view additional materials and author disclosure statement(s), visit the article page at http://dx.doi.org/10.1257/aer.103.4.1325.
1326 THE AMERICAN ECONOMIC REVIEW JuNE 2013
Dessein 2002; Aghion, Dewatripont, and Rey 2004). The granting of decision rights can mitigate inefficiencies by shielding the controlling party from potential holdup and expropriation.
There is, however, very little empirical work in economics that examines the behavioral consequences of authority and power or their motivational origins. This paper explores these forces using a laboratory experiment where we study how individuals manage and respond to authority in a hierarchical relationship. We propose a new “authority-delegation game” based on a model developed in Aghion and Tirole (1997). A principal and an agent must select one of a large number of potential projects for implementation. One party, initially the principal, has the right to decide which project to implement. The other party, initially the agent, can only make a project recommendation but lacks direct power to deter- mine the project. We follow Aghion and Tirole (1997) by defining authority as the right to determine the project.
Payoffs to the principal and agent for implementing a project are unknown ex ante, and both parties can provide effort which directly controls the probability with which they will be informed about the value of each project. One of the projects is best for the principal, while a different project is best for the agent. After the parties have exerted effort, four states are possible: both parties are informed, only the principal is informed, only the agent is informed, or neither party is informed. Before the parties provide effort, the principal can delegate authority to the agent and become the sub- ordinate party. Delegation of authority means that the agent becomes the controlling party and has the right to choose the project.
Both the controlling party (i.e., the party with the decision right) and the subordi- nate party have pecuniary incentives to exert effort in this setup because both parties earn only a low outside option in the absence of any information. The subordinate’s incentives are lower, however, because if both parties are informed, the control- ling party may overrule the subordinate’s proposal and choose the project which is best for him or her. Delegation therefore increases the agent’s effort because he can now implement his preferred project in cases where he is informed. However, delegation also reduces the principal’s control over project choice. When the princi- pal’s return from the agent’s preferred project is high, the cost of losing authority is small. A rational principal who maximizes her expected payoff should thus delegate authority in this case. When a principal’s return from the agent’s preferred project is low, however, the cost of losing authority is high, and a rational principal who maximizes expected payoff should retain authority.
Our first main result is that the principals show a proclivity for retaining author- ity in situations in which they could improve their expected income by delegating it, i.e., in situations in which their return from the agent’s preferred project is rela- tively high. However, the principals delegate only in roughly 40 percent of these cases. Pessimistic expectations about the agent’s effort in case of delegation can- not explain this reluctance. On the contrary, the principals have quite reasonable beliefs about the agent’s effort, meaning that it would be profitable to delegate in the clear majority of cases based on these beliefs. Nevertheless, principals prefer retaining authority.
These findings suggest that the principals might view authority not just as an instrument that helps them increase their earnings, but that the allocation of decision
1327fehr et al.: lure of authorityVol. 103 No. 4
rights has nonpecuniary consequences that inhibit the delegation of authority.1 In our experiment, the fact that the principals are willing to sacrifice some of their earn- ings to keep authority suggests a preference for the decision right.
Why are the principals willing to forgo money in order to keep the decision right? Our empirical data indicate that a disutility for being overruled appears to be an important driver behind their reluctance to delegate. A principal is overruled if (i) she delegates authority and (ii) both the principal and the agent are informed about project values so that (iii) the agent can disregard the principal’s information and choose his preferred project. Note that the principal’s pecuniary payoff from the agent’s preferred project is the same regardless of whether the principal is informed about the project values or not. Conditional on effort, an expected utility maximiz- ing principal who is the subordinate of an informed agent should thus be indifferent between the case where she is informed and overruled by the agent and the case where she remains uninformed. It follows that her behavior after being informed, overruled, and receiving the payoff from the agent’s preferred project should be identical to her behavior after being uninformed and receiving the same payoff.
If, however, a principal experiences a nonpecuniary disutility from being over- ruled, her behavior after these two outcomes may differ: the principal may be less willing to delegate in the next period if overruled. This is exactly what we find in our data. Principals who are overruled are significantly less likely to delegate in the next period relative to those who are uninformed, even if it is in their pecuniary interest to delegate. Moreover, we observe significantly higher delegation rates in a con- trol treatment where delegation is profitable and the principal cannot be overruled after delegation (because she always remains uninformed) relative to a treatment in which the profitability of delegation is larger but the principal can be overruled. Thus, a disutility from being overruled appears to be an important nonpecuniary factor behind the reluctance to delegate.
Our second main finding is that the controlling party substantially overprovides effort relative to the Nash equilibrium and relative to his or her best response to the subordinate party’s anticipated effort. This overprovision is persistent, with no con- vergence to the Nash equilibrium over time. Being in the position of the controlling party thus generates additional motivation for effort provision.
Our third main finding is that many subordinates substantially underprovide effort relative to the Nash equilibrium. In fact, a substantial minority of the subordinate parties (between 30 and 50 percent across various treatments) chooses a zero effort level even though—due to the very small cost of low positive effort levels—zero effort is almost never an optimal choice. This result suggests that the lack of author- ity has a demotivating effect on a substantial minority of the subordinate parties.
It turns out that many aspects of our data can be captured by the notion that subjects want to avoid ex post regret about their choices—a hypothesis that was introduced by Loomes and Sugden (1982). Regret aversion is a form of reference dependent utility and is based on the idea that subjects derive disutility from regret
1 Psychologists have postulated a preference for power (Mulder 1975) or a preference for agency, autonomy, and self-determination (Rotter 1966; deCharms 1968; Deci 1981). To our knowledge, however, no evidence yet exists that demonstrates a willingness to pay (i.e., a preference) for power, agency, or autonomy.
1328 THE AMERICAN ECONOMIC REVIEW JuNE 2013
that arises by comparing their actual ex post outcomes with those the subject could have had by choosing a different action.
A distaste for ex post regret can explain our underdelegation result in the follow- ing manner:2 the principals exert strictly less effort throughout the experiment as a subordinate (i.e., after delegation) than they did as the controlling party. Thus, a principal who delegates and is informed in the role of a subordinate would also have been informed as the controlling party.3 As being overruled leads to a project with a lower value to the principal, a principal who delegates and is informed and overruled may regret her delegation choice ex post, because if she had kept her decision right she could have chosen her preferred project. Anticipating such regret, an individual may maintain control in order to mitigate the potential for regret. Moreover, a prin- cipal who has delegated can minimize feelings of regret by minimizing the probabil- ity of being informed, i.e., by choosing a zero effort. Thus, the desire to avoid regret can explain both the propensity for underdelegation and the low effort choices of the subordinates.4 Finally, regret may analogously generate the overprovision of effort by the controlling parties if they experience regret when they remain uninformed because the right to choose the project is of little value in this case. We discuss the evidence in favor of regret aversion in more detail in Section IIIB, Section IIIE, and online Appendix A.
Our article is related to the experimental literature on the consequences of del- egation on punishment choices (Bartling and Fischbacher 2012, Coffman 2011). While these papers study the assignment of punishment in response to the alloca- tion choices of either a principal or a delegate, our article studies the willingness of the principal to delegate and the willingness of a principal and an agent to invest effort in response to the assignment of decision rights. Our results on effort pro- vision are also related to the literature on the hidden costs of control (Fehr and Rockenbach 2003; Fehr and List 2004; Falk and Kosfeld 2006; Charness et al. 2012). While this literature shows that the exercise of control reduces an individu- al’s positive reciprocity towards the principal, our article shows that lack of control has demotivating consequences on subordinates that induce them to act against their material self-interest.
We believe that our results have potentially important implications across many domains. In relation to the property rights literature (Grossman and Hart 1986; Hart and Moore 1990), (re)allocating property rights across firms may be difficult if deci- sion rights have nonpecuniary consequences for utility because organization mem- bers with decision rights may oppose their reallocation, even if they would benefit economically from it. The underdelegation of authority not only reduces the prin- cipals’ earnings, but also causes the agent to lose money in some of our treatments.
2 For further details see online Appendix A where we apply a formal model of regret aversion to our experiment. 3 Effort in the experiment is identical to the probability of becoming informed about the value of each project.
Each party’s effort is a number in the set {0, 5, …, 100 } ; this effort is compared to a random number equally distrib- uted between 1 and 100. If the random number is equal or below the effort of the party, the party was informed about the value of all projects; if the random number is above the effort of the party, the party remained uninformed. Thus, if the principal’s effort as a controlling party is higher than her effort as a subordinate party, then being informed as a subordinate party implies that the principal also would have been informed as a controlling party.
4 Note that regret can affect not only principals, but also agents in the subordinate role. The agents’ efforts might be wasted even if they are informed, since in cases where the principal is also informed, the agent will be overruled. In these cases, agents may regret positive effort levels ex post. Therefore, regret averse agents in the subordinate role may reduce effort relative to an agent that maximizes his or her expected earnings.
1329fehr et al.: lure of authorityVol. 103 No. 4
Thus, the distortion in the allocation of control rights can lead to organizational structures that reduce the value of the organization as a whole. The identification of motivational obstacles to delegation adds an important component to the theoreti- cal work by Baker, Gibbons, and Murphy (1999), Sliwka (2001), and Bester and Krahmer (2008), which predicts limits to delegation in environments with limited commitment, dynamic incentives, or limited liability.
A reluctance to delegate decision rights may also play a role in corporate finance, in the political sphere, and in the design of optimal institutions for regulating rela- tions between firms. Models of empire-building investment (Jensen 1986, Hart and Moore 1995), which have been used extensively in the literature to understand the trade-offs between financial instruments, may, in part, be founded on nonpecuni- ary motives to retain authority. In view of incumbents’ advantages for reelection (Gelman and King 1990), these motives also strengthen the case for term limits because politicians may otherwise try to keep their political power positions beyond what is good for the polity. In addition, they may provide a rationale for models in the spirit of Niskanen (1971) which assume that bureaucrats seek to maximize their discretionary budget.
The motivational consequences of authority for effort provision may be equally important. The motivation enhancing effect for the controlling parties and the detri- mental effect on the motivation of a large minority of the subordinates suggest that the incentive effects of authority are larger than the standard model predicts: a real- location of authority may cause a marked increase in effort by the new controlling party and a large reduction in effort by the party previously in control. The note- worthy gap between the controlling and the subordinate parties’ efforts also implies that when contracts are incomplete, the efficiency losses due to authority are likely to be higher than the standard model predicts. Furthermore, our finding that a lack of authority seems to demotivate only a minority of people suggests that putting the right people into positions that lack authority is important. The development of tools for detecting this type of employee may thus be important in minimizing the cost associated with the (re)allocation of authority.
Despite the systematic deviations from the predictions of the Aghion and Tirole (1997) model, we believe that their model is very useful for the study of authority because the main comparative static predictions of the model are nicely met, and the precise numerical predictions of the model enabled us to detect the motivational forces we described above. The model is thus incomplete in terms of the underlying motivational forces, but the (incomplete) model is remarkably robust in terms of the comparative static predictions. It remains to be seen whether this robustness is a general feature of the broader organizational economics literature where commu- nication (Dessein 2002, Rantakari 2008), monetary incentives (Athey and Roberts 2001), and dynamic learning (Aghion, Dewatripont, and Rey 2004) are possible. However, even if the robustness of the comparative static predictions of the Aghion and Tirole model extends to the broader organizational economics literature, we believe that this literature should take the behavioral forces observed in our paper into account because—as we show here—they may have important consequences.
The remainder of the paper is structured as follows. We present a simplified ver- sion of the model of Aghion and Tirole (1997) in Section I and derive its theo- retical predictions. Section II details our experimental design and hypotheses.
1330 THE AMERICAN ECONOMIC REVIEW JuNE 2013
Section III reports the main results of our experiment and is separated into three parts. Section IIIA summarizes the data and provides an overview of the major results. Section IIIB explores possible reasons why principals might choose to keep control rights. The third part, consisting of Sections IIIC–IIIE, examines the rea- sons for the controlling parties’ overprovision of effort and why subordinate par- ties might want to underprovide effort relative to the risk neutral Nash equilibrium. Section IV concludes.
I. Theoretical Motivations
The basis of our experimental design is a model of authority developed in Aghion and Tirole (1997). We consider a world in which a principal (she) and an agent (he) are organized in a hierarchical structure and must decide to implement one or zero projects out of a set of n ≥ 3 potential projects. With each project k ∈ {1, … , n }, there is an associated noncontractible gain of P k for the principal and a private ben- efit A k for the agent. If no project is implemented, the profit and private value are both equal to a known outside value of P 0 and A 0 , respectively.
For ease of exposition, we index the principal’s preferred project by 1 and the agent’s preferred project by 2. The principal’s preferred project yields known profit P 1 to the principal and A 1 to the agent where P 1 > P 0 and A 1 > A 0 . Likewise, the agent’s preferred project yields known benefit P 2 to the principal and A 2 to the agent with A 2 > A 0 and P 2 > P 0 . As their name suggests, the principal’s preferred proj- ect yields a strictly higher value to the principal than the agent’s preferred project ( P 1 > P 2 ). Likewise, the agent’s preferred project yields strictly higher value to the agent than the principal’s preferred project ( A 2 > A 1 ).
While the potential values of projects are known, all projects look identical ex ante and information must be collected in order to differentiate between them. The principal and agent acquire information in a binary form. At private cost g A (e ), the agent learns his payoffs to all candidate projects with probability e. With prob- ability 1 − e, the agent learns nothing and cannot differentiate between the projects. Similarly, at private cost g P (E ), the principal becomes perfectly informed about the payoffs of all projects with probability E and learns nothing with probability 1 − E. Effort choices are made simultaneously and privately. We concentrate on the case where g A (e ) and g P (E ) are quadratic, g A ′ (0 ) = g P ′ (0 ) = 0, P 1 − g P ′ (1 ) < 0, and A 2 − g A ′ (1 ) < 0. These assumptions ensure that the reaction functions are linear and that a unique interior solution exists for both authority allocations.5
We consider a four-stage game which relates decision rights, incentive conflict, and effort. In the first stage, the principal decides whether to keep decision rights or to delegate them to the agent. In the second stage, both parties privately and simul- taneously gather information about the n projects’ payoffs. In the third stage, the subordinate recommends a project to the controlling party. Finally, the controlling party implements a project or the outside option on the basis of his information and the information communicated by the subordinate.
5 In the experiment, we also use a discrete effort space to reduce cognitive burden. All predictions in Section IIB are relative to the restricted effort space.
1331fehr et al.: lure of authorityVol. 103 No. 4
We assume that the principal and agent are risk neutral. For a given effort level and implemented project k, the principal’s utility is P k − g P (E ). The agent’s utility is A k − g A (e ). As outcomes and effort choices are noncontractible, performance- or outcome-contingent payments are ruled out and the introduction of wages is neces- sary only to satisfy the agent’s participation constraint, which, to avoid further nota- tion, we assume to be satisfied.
Information in the model is soft so that information passed between parties cannot be verified. As such, if one party is informed and the other party is uninformed, the informed party can limit the amount of information given to the other party. As there is always an incentive conflict between the parties and outcomes are noncontract- ible, there is always an incentive to restrict information to the preferred project of the informed individual. It follows that communication between parties is reduced to a recommendation for a single project choice.
A. Analysis and Theoretical Implications
We denote the party that has authority as the controlling party, while the party without authority is called the subordinate. For each party, the expected value for selecting a project at random is less than their respective outside option. Thus, under the assumption of risk neutrality or risk aversion, the subordinate prefers to recom- mend the outside option rather than a random project. Similarly, an uninformed controlling party never chooses unilaterally to undertake a project other than the outside option.
Given that A 2 > A 1 > A 0 , P 1 > P 2 > P 0 , and information is soft, the subordinate always has an incentive to recommend his or her preferred project to the control- ling party. The controlling party has an incentive to follow this recommendation if uninformed and to overrule the project and implement his or her preferred project if informed. It follows that if the principal keeps control, the utilities of a risk- neutral principal and agent are
(1) E V P = E P 1 + (1 − E )e P 2 + P 0 − g P (E),
(2) E V A = E A 1 + (1 − E )e A 2 + A 0 − g A (e),
where
(3) P i = P i − P 0 , for i ∈ {1, 2},
(4) A i = A i − A 0 , for i ∈ {1, 2}.
If the agent receives control, the utility of the principal and agent are
(5) E V P d = (1 − e )E P 1 + e P 2 + P 0 − g P (E ),
(6) E V A d = (1 − e )E A 1 + e A 2 + A 0 − g A (e),
where the superscript d denotes the expected payoffs in the delegation case.
1332 THE AMERICAN ECONOMIC REVIEW JuNE 2013
From equations (1) and (2), the reaction functions if the principal keeps control are the solutions to the following first-order conditions:
(7) P 1 − e P 2 = g P ′ (E),
(8) (1 − E ) A 2 = g A ′ (e).
Equation (7) describes the principal’s reaction function which we denote by r P (e ). Equation (8) describes the agent’s reaction function denoted by r A (E ). Note that both r P (e ) and r A (E ) are downward sloping in (E, e )-space, implying that the princi- pal’s and agent’s effort level are strategic substitutes. Thus, an increase in the agent’s effort induces the principal to reduce her effort and vice versa. By the additional assumptions placed on g P (E ) and g A (e ) above, the reaction functions are also linear, and there exists an interior intersection of reaction functions, ( e NE , E NE ), which con- stitutes the Nash equilibrium of this subgame.
If the agent receives control, the reaction curves of the principal and agent are the solutions to:
(9) (1 − e ) P 1 = g P ′ (E),
(10) A 2 − E A 1 = g A ′ (e),
and denoted by r P d ( e d ) and r A d ( E d ). As in the case when the principal keeps control, the reaction functions are downward sloping in ( E d , e d ) space. Our uniqueness criteria assumed above again ensure the existence of an interior intersection
of reaction functions, ( e d NE , E d
NE ), which constitutes the Nash equilibrium of this subgame.
A careful examination of the reaction functions if the principal keeps control and if the agent receives control reveals that the principal decreases her effort when giv- ing up control, while the agent increases his effort. Delegation thus has two effects on the principal’s payoff: (i) a cost saving effect since delegation reduces the equi- librium effort of the principal and increases the agent’s equilibrium effort, and (ii) a project selection effect which decreases the probability that the principal’s preferred project is undertaken. As these effects are, in general, of opposite sign, the overall incentive for delegation depends on the specifics of the cost function and the degree of interest alignment. In our experiment, we chose cost functions and parameters such that the magnitude of P 2 determines whether delegation or retention is optimal for the principal. Full details of the experimental design and its param- eterizations are discussed in more detail in the next section.
II. The Experiment
A. The Authority Game
At the center of our experimental design is a computerized authority-delegation game with the following features. In each of ten periods, a principal is matched with an agent and shown a set of 36 cards on her computer screen representing potential
1333fehr et al.: lure of authorityVol. 103 No. 4
projects.6 One of these cards has a small positive payoff for both players and is placed face up representing the outside option. The remaining 35 cards are shuffled face down so that the location of each project is unknown. One of these cards is red and represents the principal’s preferred project. Following the theory section, we refer to this card as Project 1. A second card is blue and represents the agent’s pre- ferred project. We refer to this card as Project 2. The remaining 33 cards are white and result in zero payoff for both parties. The task of each principal-agent pair is to select a card which will be used for payment. The payoffs ensure that individuals prefer to implement the outside option relative to picking a card at random.
Play of the game is done in six stages which are illustrated in Figure 1 and dis- cussed here. Initially principals are given the decision right which corresponds to being able to select a card at the end of the game. In the first stage of the game, each principal is asked whether he wishes to keep this right or to transfer the right to the agent. Giving the right to the agent is binding and irreversible.
In the second stage, subjects choose their effort levels simultaneously and in private.7 Both subjects select their effort in increments of five from {0, 5, … , 95, 100 }. This effort corresponds to the probability that the subject learns the location of all projects. Effort has an associated cost generated via a quadratic cost function which is constant across treatments and player types:
(11) g P (E ) = 25 ( E _ 100 ) 2 , g A (e ) = 25 ( e _ 100 )
2 .
Subjects are presented information on the cost of effort in a table where each pos- sible effort and its associated cost is displayed. In all but one session, agents’ effort levels are recorded via the strategy method where an effort level is elicited both for the case where principals keep decision rights and the case where these rights are
6 Subjects are randomly assigned the role of a principal or of an agent and remain in this role throughout the experiment. In the instructions, they were referred to as participant A and participant B.
7 In the experiment we refer to effort as “search intensity.”
Stage 1 Stage 2 Stage 3 Stage 4 Stage 5
Delegation i i
Effort i i
Beliefs
Principal
decision decision
Formal choice of
Agent
Effort decision
Delegation revealed
subordinate controlling party
(strategy) (strategy)
Project valuations determined/not
determined
Project valuations determined/not
determined
Signal of
Beliefs
Stage 6
(strategy)
Figure 1. Experimental Procedures in the Authority Game
1334 THE AMERICAN ECONOMIC REVIEW JuNE 2013
delegated.8 Thus, agents choose their effort levels before they know whether the principal has delegated authority to them.
In the third stage, we elicit beliefs of both subjects. Principals and agents are asked their beliefs about the effort of the other party both in the case where decision rights are kept and where they are delegated. For principals this is done in two steps. Beliefs are first elicited for the chosen authority assignment followed by beliefs for the counterfactual. For agents, beliefs for both potential authority assignments are elicited simultaneously. To prevent hedging, no incentives are used in the elicitation of beliefs. In the fourth stage, agents are informed about whether principals kept or transferred decision rights. Then, given a subject’s effort for the principal’s assign- ment of authority, a random process determines whether that subject learns the pay- offs of all projects or whether he stays uninformed. The effort of the other subject is not revealed, nor is information indicating the success or failure of the other subject’s effort. All information gained at this stage is private.
In the fifth stage, the subordinate is given the ability to recommend a project to the controlling party. This is accomplished by visibly marking a single project on the computer screen, which can include the outside option. The recommendation is shown to the controlling party, but the payoffs associated with the recommended project are kept hidden in the case where the controlling party remains uninformed.
In the final stage, after seeing the recommendation of the subordinate and the information from his own effort, the controlling party selects a project. Payment for the round is based on the selected project and the costs of effort of each subject.
B. Experimental Design and Hypotheses
The experimental design involves four treatments implemented in a between-sub- jects design. Treatments vary in the amount that principals and agents are paid for the selection of the project preferred by the other party ( P 2 and A 1 ). By changing the payoff given to the other party, the level of incentive conflict in the environment is changed, which, as indicated by the first-order conditions in Section IA, leads to differences in predicted delegation and effort levels.
Table 1 summarizes the value of projects across the four treatments. In each treat- ment, each party earns 40 points for the selection of his preferred project and a smaller amount for the other party’s preferred project. Treatments are divided into two groups—symmetric and asymmetric—where symmetry refers to the relative values of P 2 and A 1 . In the symmetric treatments (LOW and HIGH) the payoffs from the other party’s preferred project are the same for the principal and agent. In the low alignment treatment (LOW), the payoffs from the other party’s preferred project are small (20) leading to a high degree of incentive conflict. In the high alignment treatment (HIGH), the payoffs from the other party’s preferred project are
8 We test whether the strategy method influences our results by comparing the results of three sessions of the HIGH treatment where the strategy method was run (N = 70) to the session of the HIGH treatment where a stan- dard elicitation method was used (N = 30). We find no difference across treatments. The p-values of a Kolmogorov- Smirnov test, which tests whether the distribution of agent effort is identical in treatments with and without strategy method, are 0.79 for effort with decision rights and 0.81 for effort without decision rights. Delegation frequencies differ by 1.6 percent. This difference is also not significant ( p = 0.71 in a Fisher’s exact test). The data from the treatments using the strategy method are therefore pooled with the data from treatments using the standard elicita- tion method in the subsequent analysis.
1335fehr et al.: lure of authorityVol. 103 No. 4
large (35) leading to less incentive conflict. In the asymmetric treatments (PLOW and PHIGH), the payoffs from the other party’s preferred project are large for one of the two parties (35) and small for the other (20). As a naming convention, we use PHIGH to denote the case where the principal’s value is high under the agent’s preferred project. The PLOW treatment is the case where the principal’s value is low under the agent’s preferred project.
Table 2 shows the predicted Nash equilibrium effort levels and expected profits for each treatment under the case where authority is kept and transferred. As in the model developed in Section I, E represents the effort level of the principal, while e represents the effort level of the agent. As can be seen in Table 2, the LOW treat- ment has a high degree of incentive conflict, and authority should be kept by the principal, because the principal’s expected profit if she keeps control, E V P , is 20.1, while the expected payoff if she delegates control, E V P d , is only 17.3. In the HIGH treatment, incentive conflict is reduced, and the principal should delegate authority (E V P = 23.3 versus E V P d = 24.0)9.
In the asymmetric treatments the rewards to delegation are either exacerbated or further diminished relative to the symmetric treatments. Of the four treatments, principals are predicted to have the highest expected value from delegation in the PHIGH treatment (E V P d = 25.6) and the lowest expected value from delegation in the PLOW treatment (E V P d = 17.2).
In addition to the delegation predictions, the different interest alignments also lead to different predictions with regard to equilibrium effort levels. All point pre- dictions are given in Table 2. Note that the delegation decisions predicted by the
9 As will be shown below, while the difference in theoretical expected value is small, the empirical difference turns out to be large.
Table 1—Overview of Project Payoffs
Project 1 Project 2 Outside option
Other projectsPrincipal Agent Principal Agent
PLOW 40 35 20 40 10 0 LOW 40 20 20 40 10 0 HIGH 40 35 35 40 10 0 PHIGH 40 20 35 40 10 0
Table 2—Predicted Effort Levels and Expected Profits
Principal has control Agent has control
Delegation? E NE e NE E V P E V A E d NE e d
NE E V P d E V A d
PLOW 55 25 20.1 25.6 35 45 17.2 23.3 No LOW 55 25 20.1 17.3 25 55 17.3 20.1 No HIGH 45 35 23.3 24.0 35 45 24.0 23.3 Yes PHIGH 45 35 23.3 17.2 25 55 25.6 20.1 Yes
Notes: E NE and E d NE denote Nash equilibrium predictions for the principal depending on the control allocation.
e NE and e d NE denote Nash equilibrium predictions for the agent depending on the control allocation. E V P , and
E V P d denote expected equilibrium profits for the principal depending on the control allocation. E V A and E V A
d denote
expected equilibrium profits for the agent depending on the control allocation.
1336 THE AMERICAN ECONOMIC REVIEW JuNE 2013
Nash equilibrium are always in the set of welfare maximizing delegation choices. In the PLOW treatment, aggregate expected earnings E V p + E V A are highest if the principal keeps authority, while in the PHIGH treatment aggregate expected earn- ings are highest if the principal delegates authority. In the symmetric treatments, LOW and HIGH, the delegation decision has no effect on the overall welfare if subjects choose Nash equilibrium effort levels.
In the experiment described above, the delegation decision of the principal and the joint effort decisions of the principal and agent generate one of many potential compound lotteries. If an individual’s preferences are reference dependent (as, e.g., stipulated in regret theory), the individual’s preferred action profile may depart from the equilibrium action profile which assumes players are expected value maximiz- ers. In order to control for such heterogeneity in preferences, we ran a lottery task and used choices from this lottery task as a proxy for the degree to which a subject’s preferences exhibit reference dependence. In the lottery task, each subject is pre- sented with the opportunity to participate in six different lotteries with payoffs in Swiss francs, each having the following form:
Win CHF 6 with probability 1 _
2 , lose CHF X with probability
1 _
2 .
If subjects reject the lottery they receive CHF 0.
The six lotteries varied in the amount X, that could be lost, where X took on the values X ∈ {2, 3, 4, 5, 6, 7 }. One of the six gambles was randomly selected and paid. As these lotteries are binary, any reference dependent utility function with a reference point between the lowest and the highest outcome can lead to a rejection of gambles with X ≤ 6. In particular, individuals who are regret averse and compare their outcome to the action which is optimal ex post will reject actuarially fair gam- bles. Thus, the amount X at which a subject starts rejecting the lottery can therefore be taken as an indicator of the degree to which a subject’s preference exhibits refer- ence dependence, such as, e.g., regret aversion. For example, a subject who rejects all lotteries with a potential loss of X > 3 is classified as exhibiting more regret aversion than a subject who rejects only all lotteries with a potential loss of X > 5.10
In principle, the rejection of actuarially fair gambles in the lottery choice task may also reflect a subject’s loss or risk aversion. However, in Section IIIE and the online Appendices D and E we show that loss and risk aversion have little explana- tory power with regard to effort choices, while regret aversion can rationalize both the controlling parties’ overprovision of effort and the preference for extremely low (i.e., zero) effort levels among the subordinates.11
10 One hundred forty-three out of 150 subjects who participated in the lottery task and played the authority- delegation game in one of our main treatments have a unique switching point. We use the accepted gamble with the largest potential loss as the independent variable when using the lottery task and do not exclude subjects in the analysis. Excluding subjects with multiple switching points does not significantly alter any of our results.
11 Moreover, it has been shown by Rabin (2000) that the rejection of lotteries for X ≤ 6 cannot be reconciled with the assumption that utility is a (reference-independent) strictly concave function of total wealth. Risk aversion based on concave utility of wealth at such low-stake levels would imply unreasonable levels of risk aversion at higher stakes, which makes risk averse behavior in this task incompatible with expected utility theory.
1337fehr et al.: lure of authorityVol. 103 No. 4
Thus, if regret aversion is a motive behind subjects’ rejection of lotteries in the lottery task, then the propensity to reject lotteries should also be a predictor of subjects’ effort level as a controlling party and the tendency to provide zero effort as a subordinate party. Likewise, if regret aversion is a motive behind both the propensity to reject lotteries and the reluctance to delegate, then we should observe a correlation between the two phenomena. The lottery task may thus provide further evidence regarding the motivational forces behind effort and del- egation choices.
C. Procedures
Typically, between 20 and 30 subjects participated in each experimental session which consisted of three parts.12 In part one, subjects played seven periods of a single-player version of the authority game. This single-player game is identical to the authority game except that there is no second party. Subjects choose an effort and receive information probabilistically based on their effort. Each individual must then select a project based solely on his own information. The selected project does not affect the payoff of a second party, nor does a second party recommend a project. This single-player variant gives subjects a chance to get familiar with the effort cost schedule and the computer program.
In part two, the subjects are divided into matching groups of ten subjects consist- ing of five principals and five agents. Subjects play ten periods of the main authority game in one of the four treatments. Subjects are informed that in a new period they would be matched with another randomly chosen partner.
In part three, subjects are asked to take a short questionnaire in which demo- graphic information is recorded. Instructions for the experiment include a control quiz and a verbal summary of the authority game.
Our subject pool consisted primarily of students at Zurich University and the Federal Institute of Technology in Zurich.13 The first series of experiments took place in May and June 2007 with a second series of experiments conducted in May and October 2008. Further control experiments were conducted in May 2009 and April 2011.14 In total, 504 subjects participated in the experiment, divided into 17 sessions. Experiments were computerized using the software z-tree (Fischbacher 2007). Payment was given for each period of the main authority game, for the last five periods of the single-player game, and for one randomly chosen gamble from the lottery task. On average, an experimental session of the main treatments lasted one hour with an average payment of 33.5 CHF ($35.00).15
12 More details on individual sessions is provided in online Appendix B. 13 Subjects were drawn from a database of volunteers using the Online Recruitment System for Economic
Experiments (Greiner 2004). 14 See also online Appendix B. 15 Some of the control treatments had 25 or 50 periods and therefore lasted longer. Additional information on
these treatments is given in Section IIIB.
1338 THE AMERICAN ECONOMIC REVIEW JuNE 2013
III. Experimental Results
A. The Main Facts
Our experimental design generates predictions with regard to delegation, effort, and project choices. With regard to project choices and project recommendations the theory does very well:
RESULT 1: Controlling parties who are informed about the project valuations almost always choose their preferred projects, implying that they overrule the sub- ordinates’ recommendations. Informed subordinates almost always recommend their preferred project, and uninformed controlling parties almost always imple- ment this recommendation.
Result 1 is supported by the following numbers. Principals (Agents) in the role of controlling parties who were informed implemented their preferred proj- ect in 100 percent (97.3 percent) of the cases. Principals (agents) in the role of the informed subordinate party recommended their preferred project in 92.6 percent (92.5 percent) of the cases. Finally, principals (agents) in the role of the uninformed controlling party followed the subordinate party’s recommendation in 94.1 percent (96.5 percent) of the cases. If the subordinate parties were not informed they typically recommended the outside option (principals: 95.3 percent; agents: 97.0 percent).
Result 1 indicates that the controlling parties used the decision right in their favor. As predicted by theory, this generates a disincentive for subordinates’ effort provi- sion, but it also makes it reasonable for the principals to delegate authority if their payoff loss at the agents’ preferred project is low. Therefore, we next turn to the principals’ delegation choices. Recall that in case of Nash equilibrium effort choices by the principal and agent, the principal has an incentive to delegate authority in the HIGH and PHIGH treatments and to keep authority in the LOW and PLOW treat- ments. Empirically, we find in our experiment:
RESULT 2: (i ) When the principals’ interests are misaligned with the agent (LOW and PLOW) such that the principals are predicted to keep authority, delegation decisions are close to the equilibrium predictions. (ii ) When the principals’ inter- ests are strongly aligned with the agent (HIGH and PHIGH treatments) such that principals should delegate, we observe strong underdelegation of authority relative to the equilibrium predictions.
Figure 2 shows the frequency of delegation for each treatment graphically. As can be seen on the left-hand side of the figure, delegation rates in the PLOW and LOW treatments are 16.3 percent and 13.9 percent. While these levels are above the pre- dicted level of zero, deviations from the prediction appear to be due to infrequent experimentation rather than heterogeneity in delegation strategies. There is little per- sistence in the strategy of delegation, with 67.4 percent of individuals who delegated authority in one period switching to keeping authority in the next. The frequency of delegation for most individuals is also low, with 39.4 percent of individuals choosing to never delegate and 89.4 percent of individuals delegating in three periods or less.
1339fehr et al.: lure of authorityVol. 103 No. 4
Average delegation rates in the HIGH and PHIGH treatment are 35.5 percent and 42.7 percent, far below the predicted rate of 100 percent. These low delegation rates are also rather stable over time. In the HIGH treatment the overall delegation rate is 33.5 percent in the first five periods and 37.5 percent in periods six through ten. In the PHIGH treatment the overall delegation rate is 36.7 percent in the first five peri- ods and stabilizes around 48.7 percent from period six onwards.16
In contrast to the LOW and PLOW treatments, the underdelegation of authority in the HIGH and PHIGH treatments appears to be due to heterogeneity in delegation strategies across individuals. Less than 20 percent of individuals delegate eight or more times in the experiment, and individuals who delegate in one period are more likely to delegate in the next period, suggesting some persistence in the delegation strategy. However, even in the PHIGH treatment in which delegation incentives are highest according to the Nash prediction, 40 percent of individuals have a delega- tion frequency of zero, suggesting that underdelegation is rather pervasive.
One possible reason for the observed underdelegation might be that actual effort pro- vision if the principal keeps control compared to the case in which the agent receives control makes it more profitable to keep authority. Table 3, which shows the realized profits of principals who kept and delegated authority, shows that this is not the case. In the HIGH and PHIGH treatments, realized profits for the principal are lower than predicted if she keeps control and higher than predicted if the agent receives control. Principals who delegate have on average 30.4 percent greater earnings in the HIGH treatment and 44.5 percent greater earnings in the PHIGH treatment.
16 While the difference in delegation rates between the first and second half of the experiment is insignificant for the HIGH treatment, the difference of 12 percentage points in the PHIGH treatment is significant ( p < 0.01 in a probit regression of delegation on a dummy for periods 6–10). We also ran a 50 period session to test for potential long-run learning effects. While delegation increased over the first 20 periods, under-delegation was still pervasive. See Section IIIB for details.
Predicted frequency: 0% Predicted frequency: 100% 100
80
60
20
40
D e
le g
a tio
n f
re q
u e
n cy
( p
e rc
e n
t)
PLOW LOW HIGH PHIGH
16.3% 13.9% 35.5% 42.7% 0
Figure 2. Delegation Frequencies by Alignment
1340 THE AMERICAN ECONOMIC REVIEW JuNE 2013
The second main hypothesis of the experiment is about effort provision. In theory, an incentive conflict leads the controlling party to put in more effort than would be optimal in the case of contractible effort and causes the subordinate to put in less. Relative to this Nash equilibrium benchmark, we observe
RESULT 3: Controlling parties provide an excess of effort relative to the Nash equi- librium. Subordinates underprovide effort relative to the Nash equilibrium.
Figure 3 plots the average deviation of effort levels from the predicted equilibrium values by the principal and agent with both means and 95 percent confidence inter- vals calculated from individual average efforts. It can be seen that, when authority is kept, the principal overprovides and the agent underprovides relative to the pre- diction. This phenomenon is reversed, again in all treatments, when authority is delegated, and these deviations are significant for the majority of treatments.17 In the low treatment, for example, the principal overprovides effort by roughly ten units relative to the prediction if she keeps control, while the agent underprovides effort by about ten units. This deviation pattern is reversed when the agent is the controlling party.
The deviations in effort levels from the equilibrium prediction cause inefficiencies that are reflected in the low actual payoff levels of the principals and the agents rela- tive to the predicted payoff levels. Table 3 shows that the principals earn less than predicted in five out of eight cases. In particular, if control is kept, which occurs most frequently in all treatments, the principals always earn less than predicted. For the agents the income loss relative to the prediction is even more extreme (see Table 4): in all eight cases they earn on average less than predicted.
The combined effect of underdelegation and deviations in effort provision has particularly strong pecuniary consequences in the PHIGH treatment. Unlike the other treatments, delegation in this treatment lead both the principal and agent to be better off in expectation relative to held control. Principals who delegate receive 45 percent more income compared to those who hold on to decision rights. Likewise agents who are left as the subordinate receive 13.8 percent less profit than those who
17 We report results from a nonparametric Wilcoxon Signed-Rank test in Table C1 of online Appendix C.
Table 3—Realized Profits and Predicted Equilibrium Profits for Principals
Principal has control Agent has control
Actual Predicted Observations Actual Predicted Observations
PLOW 18.3* 20.1 251 17.6 17.2 49 LOW 19.0 20.1 310 15.0** 17.3 50 HIGH 19.1*** 23.3 316 24.9 24.0 174 PHIGH 18.4*** 23.3 172 26.6 25.6 128
Notes: Significance levels calculated by regressing earnings on a constant and testing whether the constant is equal to the prediction. Errors clustered at the individual level.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1341fehr et al.: lure of authorityVol. 103 No. 4
are delegated to. Taken together, in PHIGH the welfare loss of keeping authority amounts to 30 percent in terms of expected income.18 The loss in aggregate payoff due to the deviations from the Nash equilibrium are not restricted to PHIGH, how- ever. Table C2 in online Appendix C shows that both the principals and the agents earn less than in the Nash equilibrium in each of the four treatments. We summarize these findings in the following result:
RESULT 4: In each treatment, the deviation in effort provision and delegation leads to monetary losses for both parties relative to the Nash equilibrium. Monetary losses are most acute in the PHIGH treatment where delegation would lead to higher aver- age earnings for both parties.
18 The implicit assumption in this calculation is that if the principals who kept control rights were to delegate instead, they would exert effort similar to those who delegated. If the principal were to exert less effort, the overall monetary loss would be slightly smaller. In the extreme case where we assume principals exert zero effort in the counterfactual, the monetary loss would amount to 27 percent.
Observations under principal control
ObidACl
20
Observations under agent control
95% confidence interval
PHIGH
10LOW
HIGH
PHIGH
PLOW
10 20–10–20
LOW
HIGH PLOW
–10
–20
̃ e = e − e NE
̃ E = E − E NE
Figure 3. Deviations from Equilibrium Effort Predictions
Notes: The vertical axis shows the difference ( ̃ E ) between the principal’s observed effort (E) and the Nash equilib- rium effort ( E NE ). The horizontal axis shows the difference ( ̃ e ) between the agent’s observed effort (e) and the Nash equilibrium effort ( e NE ). Ninety-five percent confidence intervals and mean effort calculated at the individual level.
1342 THE AMERICAN ECONOMIC REVIEW JuNE 2013
B. Exploring the Principals’ Reluctance to Delegate
Pecuniary Motivations.—A natural initial hypothesis for the observed under- delegation of authority in the HIGH and PHIGH treatments is that individuals believe that they are monetarily better off retaining authority. To see whether this hypothesis has merit, we consider the following counterfactual: suppose that a prin- cipal who did not delegate would elect to delegate instead. Given her beliefs about the agent’s actions if she keeps control and if she delegates control, what would be her gain or loss in expected earnings?
As the effort of the principal was elicited only in the case of her chosen authority allocation, a comparison of the principal’s expected earnings for the cases of delega- tion and nondelegation requires assumptions about her effort in the counterfactual authority allocation. As we have the principals’ beliefs about the agents’ effort from both the delegation case and nondelegation case, a natural approach is to use the principal’s best reply effort as a proxy for effort. If, for example, the principal kept authority we can compute the principal’s best reply effort for the case in which the principal had delegated authority. Using this effort proxy and the principal’s belief about the agent’s effort enables us to compute the principal’s expected profit for the counterfactual case of delegation.19
As a comparison value, we next compute the expected profits of the principal for the case of retained authority, taking the principal’s actual effort and his beliefs about the agent’s effort into account.20 Subtracting the expected profit from retained authority from the expected profit from delegation yields our first measure for the expected gains from delegation.
Figure 4 shows the cumulative density function of the gains from delegating under the assumption that the principal would have played a best reply in case he had
19 Under the assumption that the principal best replies to his beliefs, the expected earnings for the counterfactual case of delegation is given by
E V P d ( E d = r P d ( e d ), e d ) = e d P 2 + (1 − e d ) r P d ( e d ) P 1 + P 0 − g P ( r P d ( e d )), where e d is the principal’s belief about the agent’s effort under delegation, P 0 is the principal’s payout under the out- side option, P 2 is the principal’s payment under the agent’s preferred project net of P 0 , P 1 is the principal’s payment under the principal’s preferred project net of P 0 , and r P d ( e d ) is the best response function constructed in equation (9).
20 This comparison value is given by
E V P (E, e ) = E P 1 + (1 − E ) e P 2 + P 0 − g P (E).
Table 4—Realized Profits and Predicted Equilibrium Profits for Agents
Principal has control Agent has control
Actual Predicted Observations Actual Predicted Observations
PLOW 23.0*** 25.6 251 18.8** 23.3 49 LOW 16.1** 17.3 310 17.9 20.1 50 HIGH 21.0*** 24.0 316 20.1*** 23.3 174 PHIGH 15.9** 17.2 172 18.1** 20.1 128
Notes: Significance levels calculated by regressing earnings on a constant and testing whether the constant is equal to the prediction. Errors clustered at the individual level.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1343fehr et al.: lure of authorityVol. 103 No. 4
delegated. As can be seen in this graph by looking at the mass to the right of the zero line, 68 percent of observations in the HIGH treatment and 92 percent of observa- tions from the PHIGH treatment are from individuals who would have been better off if they had delegated. The retention of authority in the PHIGH treatment is espe- cially noteworthy since both the principal and the agent would be made better off through delegation. Thus, in this treatment, the underdelegation is suboptimal not only from the principals’ perspective, but also from the perspective of the organiza- tion as a whole.
One might worry that using the best response to beliefs as a proxy for effort might overstate the expected return to delegation. Perhaps some individuals may not per- fectly best respond to their beliefs.
As a conservative secondary measure for the expected gains from delegation, we next consider the case where the principal provides zero effort after delegation. This criterion is selected for three reasons. First, an individual who puts in zero effort has no potential losses and minimal exposure to risk. Relative to the actual strategies typically employed by principals, the zero effort criterion should thus be an attrac- tive strategy for principals who are extremely risk or loss averse. Second, besides very high effort choices which are observed very infrequently, zero effort minimizes the expected value of delegation, giving us the lowest reasonable expected value of delegation. Finally, zero effort is, in fact, the modal strategy taken after delegation, suggesting it is a relevant benchmark for analysis.
In Figure 5 we depict the cumulative density function for the expected gains from delegation under the assumption that the principal would have chosen zero effort if he had delegated. We find that 46.8 percent of observations in the HIGH treatment and 75 percent of observations in the PHIGH treatment are from individ- uals who would have been better off in case of delegation. This result is remarkable
Figure 4. Cumulative Density Function of Expected Gain from Delegation if Principal Best Replies to Beliefs
1No delegation optimal
0.6
0.7
0.8
0.9
0.3
0.4
0.5
32.0%
0
0.1
0.2
–15 –10 –5 0 5 10 15 20 25
HIGH
PHIGH
7.6%
Delegation optimal
Expected gain/loss from switching from kept authority to delegating authority and best responding to beliefs
1344 THE AMERICAN ECONOMIC REVIEW JuNE 2013
because even if we assume that principals choose highly suboptimal effort levels after delegation, it would have often been better for them (given their beliefs) to delegate authority.
One might also be concerned that the underdelegation effect is due to having a limited number of periods in which to learn the optimal delegation strategy. Perhaps underdelegation is a consequence of incomplete learning in our ten- period experiments.
In order to study this hypothesis, we ran an additional treatment (PHIGH50) in which we increased the number of periods to 50 and increased the equilibrium returns to delegation from 10 percent (in the PHIGH treatment) to 17 percent. As with the original PHIGH treatment, we use an asymmetric design in which the agent’s pay- ment for his preferred project is much higher than his payment under the principal’s preferred project, and where both the principal and agent would benefit highly from delegation at the equilibrium effort levels.21 To further facilitate learning we also simplified the design by making the subordinate’s recommendation automatic, i.e., the controlling party knew with certainty that an informed subordinate would rec- ommend his preferred project, while an uninformed subordinate would recommend the outside option.
Figure 6 shows the time path of delegation decisions in this treatment. As can be seen, while there is an increase in delegation in the first 25 periods, delegation
21 In the PHIGH50 treatment, the payment for the principal and agent under the principal’s preferred project were 45 and 20, respectively. Under the agent’s preferred project, the payments were 40 (for the principal) and 45 (for the agent). Equilibrium payoffs were 31 points for the principal and 23.1 points for the agent if the agent had control, and 26.5 points for the principal and 18.1 points for the agent if the principal had control. We randomly paid 20 of the 50 periods at the end of the experiment. 64 subjects participated in two sessions of this treatment, and the average session time of this experiment was 2.5 hours with an average payment of 72 CHF ($75).
Figure 5. Cumulative Density Function of Expected Gain from Delegation if Principal Chooses Zero Effort after Delegation
No delegation optimal
0.9
1
0.6
0.7
0.8
53.2%
0.3
0.4
0.5
25.0%
Delegation optimal
0
0.1
0.2
–25 –20 –15 –10 –5 0 5 10 15 20 25
HIGH
PHIGH
Expected gain/loss from switching from kept authority to delegating authority and exerting zero effort
1345fehr et al.: lure of authorityVol. 103 No. 4
rates remain fairly constant in the remaining 25 periods. Of the 32 principals in the treatment, 43.8 percent never switched their delegation decision after period 25. Subjects are also on average fairly persistent in their delegation choice. On the basis of a median split, we find that those who predominantly keep authority del- egate in only 11 percent of the cases in the second half of this treatment. Those who predominantly delegate authority keep it in only 12.5 percent of the cases in the second half of this treatment. Overall, the delegation rate is 8 percent higher in the second half of this treatment compared to the second half of the PHIGH treatment, but this is what one would expect because of the increased incentives to delegate in the PHIGH50 treatment. The overall delegation rates in PHIGH50 are, however, still well below the level predicted by theory, with an average delegation rate of 56.5 percent in the last 25 periods of the experiment. As in the PHIGH treatment, this large amount of underdelegation occurred despite the fact that both the princi- pals and the agents were substantially better off in cases where decision rights were delegated relative to cases where decision rights were held.
Nonpecuniary Motivations.—The results above suggest that many principals had strong pecuniary incentives to delegate. They further suggest that a large portion of underdelegation is not due to incomplete learning. Why then do we observe this strong reluctance to delegate?
One nonpecuniary force behind the principals’ choices appears to be a disutility from being overruled. Recall that the principal’s return from the agent’s preferred project is the same regardless of whether the principal is informed or uninformed. Thus, for a principal who has delegated and faces an agent who selects his pre- ferred project, the pecuniary value of being informed and uninformed is the same. Conditional on effort, an expected utility maximizing principal in the subordinate
1
0.6
0.8
F re
q u e n cy
o f d e le
g a tio
n
0.4
48%
Average delegation
All periods:
0
0.2
Period
Last 25 periods: 56%
0 5 10 15 20 25 30 35 40 45 50
Figure 6. Average Delegation Frequencies of Principals in the PHIGH50 Treatment
Notes: Period averages shown by the diamonds, quadratic time trend shown by the solid line. In total, 32 principals participated in these 50-period experiments.
1346 THE AMERICAN ECONOMIC REVIEW JuNE 2013
position should be indifferent between the case where she is informed and overruled by the agent and where she remains uninformed. Thus, her response to being over- ruled and receiving the payoff from the agent’s preferred project should be identical to her response to being uninformed and receiving the payoff of the agent’s pre- ferred project. However, if a principal experiences a nonpecuniary disutility from being overruled, her response to these two outcomes may differ: In the next period, the principal may correctly anticipate the potential disutility from being overruled and, therefore, she may be less willing to delegate.
To examine spillovers across periods, we take the principals who delegate in period t − 1 and regress the probability of delegating in period t on the information the principal and agent received in the previous period. To account for potential differences in effort and beliefs, we also condition on the effort of the principal in the previous period and the beliefs the principal has about the agent’s action in the current period.
Table 5 reports the marginal effects of this regression for various subsets of our data. The omitted category in all regressions is the case in which both the prin- cipal and the agent are uninformed. As can be seen in column 1, the principal is more likely to delegate in period t if either she was informed in t − 1 or the agent was informed in t − 1. However, when both the principal and agent were informed in t − 1 and the principal was overruled by the agent, her delegation probability falls to the same level that also prevails if both were uninformed in the previous period, i.e., when the outside option was implemented. This suggests that the prin- cipal reacts negatively to being overruled—a behavior that cannot be explained if the principal maximizes the expected utility from her monetary payoffs only. The behavioral response to overruling is robust to specifications which include period
Table 5—Delegation Conditional on Previous Experience
(1) (2) (3)
Principal informed in t − 1 0.170 0.194 * 0.191 (0.120) (0.110) (0.129)
Agent informed in t − 1 0.208 *** 0.189 *** 0.122 ** (0.062) (0.061) (0.059)
Overruled in t − 1 –0.371 ** –0.410 *** –0.323 ** (0.153) (0.149) (0.159)
Effort in t − 1 –0.004 ** –0.005 –0.002 (0.002) (0.003) (0.002)
Treatment controls? Yes Yes Yes Belief and time controls? No Yes Yes
Pseudo R 2 0.170 0.249 0.199 Observations 360 360 271
Notes: Marginal effects from a probit regression are reported in the table. Standard errors in parentheses, clus- tered by individual. Sample is restricted to principals who delegated in t − 1. The omitted category is the case in which both the principal and agent are uninformed. Regressions (1) and (2) contain the data from all treatments. Regression (3) contains only data from the HIGH and PHIGH treatments, where delegation is predicted. Belief con- trols are beliefs of principals under both authority allocations. Time controls are period fixed effects.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1347fehr et al.: lure of authorityVol. 103 No. 4
dummies and belief data (column 2), and specifications which use only the HIGH and PHIGH treatments where delegation is expected to take place (column 3).
The importance of the overruling effect is also corroborated by data from questions in the ex post survey administered after the PHIGH50 treatments in April 2011. We asked participants in the role of the principal to evaluate two delegation scenarios in both of which they delegated and the agent’s project is implemented, and which differ only in whether they were informed or not. On a seven-point likert scale, the scenario in which the principal was uninformed had an average evaluation of 5.97, while the scenario in which the principal was informed and thus overruled had a lower average evaluation of 4.59. This difference is highly significant both in a paired t-test ( p < 0.01) and in a nonparametric sign-rank test ( p < 0.01).22 Hence, even though the monetary outcome is identical across both scenarios, prin- cipals on average assigned significantly lower value to the scenario in which they were overruled.
We performed another test of the hypothesis that the nonpecuniary disutility of being overruled causes a reluctance to delegate by conducting two additional 25-period treatments which had nearly identical equilibrium returns to delegation but varied in the extent to which overruling was possible. The first of these treatments restricted the subordinate’s effort to zero and did not allow him to make recommen- dations. This treatment was symmetric with a payment of 40 if the own preferred project was chosen and 30 if the other party’s preferred project was chosen. We refer to this treatment as HIGH NOREC. There are relatively high (15.8 percent) equilibrium returns to delegation since the commitment not to exert effort as the subordinate increases the effort exerted by the controlling party. We compare these data to a 25-period version of the PHIGH50 treatment (PHIGH25) which had simi- lar equilibrium returns to delegation.23
Recall that the difference between the HIGH NOREC and the PHIGH25 treat- ment is that in the former the principals cannot be overruled. Therefore, if the principals derive disutility from being overruled they should be more willing to delegate in HIGH NOREC. We find indeed that the average delegation rate of the HIGH NOREC treatment was higher (67.1 percent) than in the PHIGH25 treat- ment (41.3 percent)—a difference that is statistically significant ( p = 0.011).24 The higher delegation rate in HIGH NOREC occurred despite the fact that the empirical return on delegation is 21 percentage points higher in the PHIGH25 treatment than in HIGH NOREC. For this reason, the higher delegation rates in HIGH NOREC provide additional support for the hypothesis that being overruled is associated with disutility.
What is the source of this disutility? A plausible answer to this question is pro- vided by the notion of regret aversion (which we formalize in online Appendix A). To examine the role of regret aversion in our setting it is useful to recall that the princi- pals chose considerably higher effort in the role of the controlling party compared to
22 The two survey questions were randomly ordered and asked on separate screens of the exit survey. 23 The 25-period experiments lasted on average 1.75 hours, 32 subjects participated in the PHIGH25 treatment,
and 28 subjects participated in the HIGH NOREC treatment. Pooling earnings across the treatments, subjects earned 44 CHF ($47) on average.
24 This p-value is calculated using a probit regression controlling for period fixed effects. Standard errors are clustered by the individual.
1348 THE AMERICAN ECONOMIC REVIEW JuNE 2013
when they were the subordinate party. Therefore, if the principals in the subordinate role became informed about which project was best for them they knew that they would also have been informed if they had kept the decision right. In other words, they could have chosen their best project if they had kept their decision right. It is thus quite plausible that the principals regretted their delegation decision when they were informed and the agent’s preferred project was implemented. If this regret is psychologically aversive the delegation option becomes less valuable. This account of underdelegation in terms of regret aversion can explain the following four facts: (i) the underdelegation of authority in the HIGH and the PHIGH treatment; (ii) the sharp reduction in delegation rates after being overruled; (iii) the subjects’ prefer- ence (in the postexperimental survey) for being not informed about their best project when the agent’s preferred project is implemented anyway; and (iv) the much higher delegation rate when the experimental design rules out ex post regret by preventing the principal from providing effort after delegation.
Our explanation of underdelegation in terms of regret aversion would receive further support if an individual measure of regret aversion would predict individ- ual differences in the reluctance to delegate. Subjects’ behavior in the lottery task described in Section IIB may be interpreted as such an individual difference mea- sure. If the subjects’ reluctance to accept the lotteries is at least partly driven by subjects’ regret aversion, the propensity to reject lotteries should also predict the reluctance to delegate (in the treatments where overruling can occur). This is, in fact, what we observe. In online Appendix C (Table C3) we report probit regressions of principals’ delegation choices with treatment conditions, period fixed effects, prin- cipals’ beliefs about agents’ effort and the number of rejected lotteries as explana- tory variables. Individuals in the HIGH and PHIGH treatments are 12 percent less likely to delegate for each gamble in the lottery task that they decline ( p = 0.021). If we combine the observations in the HIGH and PHIGH treatment and split the sample at median lottery acceptance, the difference in delegation frequency between the group which accepts more gambles and the group which accepts fewer gambles is 20 percent. These findings further support the view that regret aversion appears to be an important motive behind the reluctance to delegate authority. The next subsec- tions will show that the same motive may also partly explain the controlling and the subordinate parties’ effort choices.
C. Exploring the Controlling Parties’ Overprovision of Effort
We saw in Figure 3 that the provision of effort by the controlling party exceeds the Nash equilibrium prediction across all treatments, while the effort of the subordinate is below the Nash prediction. These deviations are persistent, with no convergence to the Nash equilibrium over time.25
Persistent deviations from the Nash equilibrium might be due to one of two sources. First, for a given belief about the other party’s effort, an individual may respond to those beliefs differently than the best reply. For example, if the controlling party
25 Looking at agents’ effort in the first five periods and the last five periods of the four main treatments, average effort declines by 2.4 points as a subordinate ( p = 0.014 if effort is regressed on a dummy for periods 6–10) and decreases by 0.2 points as controlling party ( p = 0.76).
1349fehr et al.: lure of authorityVol. 103 No. 4
systematically overprovides effort relative to the best reply, his or her effort is likely to be higher than the Nash equilibrium effort. Likewise, if the subordinate party underprovides effort relative to the best reply, then the effort is likely to be lower than the Nash equilibrium effort.
Second, beliefs about the other party’s effort provision may deviate from those predicted in the Nash equilibrium. Because of strategic substitutability, a controlling party whose beliefs about subordinate effort are below those predicted by the Nash equilibrium will increase his or her effort relative to the Nash equilibrium. Likewise, a subordinate party whose beliefs are above the Nash equilibrium will decrease effort in substitution. In this section we examine both the best reply channel and the belief channel as potential sources of the controlling parties’ overprovision of effort.
We first look at systematic deviations from the best response function by con- structing the theoretical best response for the controlling party if control is kept and if control is delegated under the assumption of risk neutrality:
(12) r P ( e ) = 100 P 1 − e P 2
_ 50
, r A d ( E d ) =
100 A 2 − E d A 1 __
50 .
By comparing these best responses with the actual response of controlling parties to their beliefs, we can examine systematic deviations from the best response function.
Figure 7 shows this comparison pooled for all treatments. The dashed 45° line represents those cases where the actual response to beliefs coincides with the best response to these beliefs. Points above the 45° line represent observations in which the controlling party overprovides relative to the best response, while points below the 45° line represent an underprovision of effort.
The solid line in Figure 7 shows the empirical relationship between the actual response to beliefs about the subordinates’ effort and the best response. The positive slope of this line indicates that the best response has some (qualitative) predictive power. However, the overwhelming feature in the data is the systematic over- provision of effort by the controlling party relative to the best response. Counting all observations strictly above the 45° line, 66 percent of observations for principals and 77 percent of observations for agents provide more effort than is predicted by a best response to beliefs. The magnitude of this overprovision is typically large, with 48 percent of observations 15 points or more above prediction.
While Figure 7 shows the combined data across all treatments, the pattern doesn’t vary qualitatively across treatments. Table 6 shows the average effort of the control- ling party and the corresponding average of the best response to beliefs. As can be seen, effort provision of the controlling party is above the average best response prediction in all treatments and for both authority allocations, and in seven of these eight cases the difference is significant.26
26 Technically, the effort provisions of individuals within a matching group may be correlated due to shared histories. However, as the information concerning the actions of others in the matching group is limited, we expect the effect of heterogeneous learning to be limited. As an additional control, we ran matching group-clustered ver- sions of each sign-rank and rank-sum test included in this article to check whether matching group-level effects are driving our results. As expected, the p-values of these tests are slightly higher but have a minor effect on the significance levels reported throughout the paper. See Datta and Satten (2008) and Datta and Satten (2005) for details of the two tests.
1350 THE AMERICAN ECONOMIC REVIEW JuNE 2013
Based on this data, we conclude
RESULT 5: Controlling parties overprovide effort relative to their best response to beliefs about the subordinate’s effort.
Result 5 suggests that having authority appears to have a motivational effect on the effort provision of the controlling parties. We next turn to beliefs. Since the efforts of the two parties are strategic substitutes, deviations from the Nash equi- librium prediction may partially be explained by pessimistic beliefs of controlling parties.
Table 7 compares actual beliefs to the Nash equilibrium beliefs for all treatments and both authority allocations. As can be seen by comparing the first two columns, the principal’s belief about agent effort if control is kept is comparable to the Nash equilibrium prediction. In fact, in three out of four cases (i.e., in PLOW, LOW,
Principal as controlling party Agent as controlling party
66% of observations 77% of observations
Predicted
Observations: Fitted
Best response to beliefsBest response to beliefs
100
80
60
40
20
0 0 20 40 60 80 100 0 20 40 60 80 100
1
2–4
5–7
8–15
16–31
32–63
64+
A g e
n t’s
a ct
u a
l r e
sp o
n se
t o
b e
lie fs
100
80
60
40
20
0
P ri
n ci
p a
l’s a
ct u
a l r
e sp
o n
se t o
b e
lie fs
Predicted
Observations: Fitted
1
2–4
5–7
8–15
16–31
32–63
64+
E above r P ( ˆ e )
r P ( ˆ e )
above e d
r A d ( ˆ
E d )
r A d ( ˆ
E d )
Figure 7. Controlling Party: Actual Effort versus Best Response to Beliefs (Combined data from all main treatments)
Table 6—Comparison of Effort Provision of the Controlling Party to the Best Response to Beliefs
Principal has control Agent has control
Actual Best response Actual Best response effort effort effort effort
PLOW 55.7 53.9 68.1*** 49.1 LOW 66.1*** 54.5 68.3*** 55.8 HIGH 48.2*** 42.1 58.7*** 45.3 PHIGH 58.2** 45.9 65.1** 56.2
Note: Significance levels calculated using a signed-rank test with beliefs and effort averaged by individual prior to estimation.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1351fehr et al.: lure of authorityVol. 103 No. 4
and HIGH) the principals’ effort expectation is above e NE , but the deviation is not significant. Thus, pessimistic beliefs of the principal cannot contribute to the over- provision of effort in these cases. The situation is somewhat different if authority was delegated. Here, the controlling party (the agents) expected in all four treat- ments that the subordinate party will underprovide effort relative to the Nash equi- librium. Thus, beliefs of the agents do account for some of the overprovision of effort relative to the Nash equilibrium prediction.27
D. Exploring the Subordinate Parties’ Underprovision of Effort
We next examine possible reasons for deviations from the Nash equilibrium on the part of the subordinates. We will again examine the role of beliefs as well as the role of systematic deviations from the best response function as potential sources of the observed deviation of effort relative to the Nash equilibrium.
Figure 8 shows the relationship between the theoretical best response and the actual response function for the subordinates. As before, the 45° line represents the predicted best response function of the subordinate in response to beliefs about the effort of the controlling party, while the solid line shows the empirical best response behavior from a simple linear regression. Points above the 45° line represent observa- tions in which the subordinate overprovides effort relative to the best response, while points below the 45° line represent an underprovision of effort.
As can be seen in the left-hand panel of the figure, the actual response function is positively sloped but relatively flat, suggesting a relatively weak effort response to beliefs. Unlike the controlling parties’ efforts, which were clustered above the best response correspondence, effort provisions by subordinates are heterogeneous: 52 percent of individual choices are below the best response to beliefs for agents, while 56 percent of individual choices are below the best response to beliefs for principals. In addition, a large number of individual choices are considerably below the best response.
27 Table C4 in online Appendix C shows that beliefs of controlling parties about subordinate effort do, on aver- age, exceed actual subordinate effort.
Table 7—Comparison of Actual Beliefs of the Controlling Party to the Nash Prediction
Principal has control Agent has control
Nash Actual Nash Actual prediction belief prediction belief
PLOW 25 30.4 35*** 21.8 LOW 25 27.5 25* 20.9 HIGH 35 35.8 35* 29.4 PHIGH 35* 28.2 25** 19.0
Note: Significance levels calculated using a signed-rank test with beliefs and effort averaged by individual prior to estimation.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1352 THE AMERICAN ECONOMIC REVIEW JuNE 2013
A particularly salient fact in Figure 8 is that a large number of subordinates put in zero effort, i.e., lack of control appears to have a strong demotivational effect for a large minority. Recall that the cost of effort is convex with the cost of increasing effort from zero to five equaling g P (5 ) − g P (0 ) = 0.06 points. Since incremental effort is nearly costless, zero effort is predicted only in cases where the subordinate believes in an effort of 100 by the controlling party, which almost never occurred.
The heterogeneous behavior of subordinates also appears to be a robust phenom- enon at the treatment level. Table 8 reports the average effort of the subordinate, the average theoretical best response to beliefs, and the proportion of individuals who provide zero effort for each treatment and for principals and agents separately. As can be seen by comparing the first two columns of each treatment, there is little dif- ference between the actual effort and the theoretical best response to beliefs at the mean. The similarity in these two averages reflects the heterogeneous nature of sub- ordinate effort provision where both under- and overprovision of effort is observed.
Looking at the third column of each row, it is apparent that zero effort is a modal strategy for the subordinates in all treatments. Zero effort is observed at least 25 per- cent of the time in all eight cases, and in three of these cases, zero effort accounts for roughly 50 percent of the observations.28 The high frequency of observed zeroes
28 One possible explanation for zero effort is that individuals who exert zero effort simply don’t understand the environment. In studying the effort that agents exert as the controlling party, however, this explanation is unlikely. Remember that we collect effort choices of agents in both roles, the subordinate and the controlling party. We can therefore directly compare whether those subjects who exert zero effort as subordinates are different from those who do not when in the controlling party role. A regression of controlling party effort on a dummy that takes on the value 1 if subordinate effort is zero, controlling for treatment and clustering standard errors at the individual level, reveals that those agents who chose zero effort as subordinates on average exert three units of additional effort (this difference is not significant ( p = 0.35)). In fact, as will be discussed later in Section IIIE, a positive difference is to be expected if regret aversion directly affects effort choices. This suggests that a lack of understanding is not the driver of zero effort choices.
ac
Agent as subordinate
Best response to beliefs
100
80
60
40
20
0 0 20 40 60 80 100
Predicted
Observations:
Fitted
1
2–4
5–7
8–15
16–31
32–63
64+
A g
e n
t’s a
ct u
a l r
e sp
o n
se t o
b e
lie fs
52% of observations
below
56% of observations below
Principal as subordinate
0 20 40 60 80 100
P ri
n ci
p a
l’s a
ct u
a l r
e sp
o n
se t o
b e
lie fs
100
80
60
40
20
0
Best response to beliefs
Predicted
Observations:
Fitted
1
2–4
5–7
8–15
16–31
32–63
64+
r A ( ˆ E )
r A ( ˆ E )
e r P d ( ˆ e d )
r P d ( ˆ e d )
E d
Figure 8. Subordinates: Actual Effort versus Best Response to Beliefs (Combined data from all main treatments)
1353fehr et al.: lure of authorityVol. 103 No. 4
is due in large part to a subset of individuals who always exert zero effort in the subordinate role.29
Based on these observations, we conclude
RESULT 6: The subordinates’ effort behavior is heterogeneous. While on average effort provision is close to the theoretical best response, there is a large group of subordinates who provide zero effort, far below the optimal best response. For this group, authority appears to have a strong demotivational effect. In addition, there is a smaller group of subordinates who systematically overprovide effort.
Turning to beliefs, Table 9 shows the beliefs of the subordinate compared to the Nash equilibrium. As can be seen, agents and principals have optimistic beliefs relative to the Nash equilibrium. As optimistic beliefs are expected to lead to a decrease in effort, beliefs may be contributing to the underprovision of effort by the agent and the principals. However, as we noted in Figure 8, the actual response function is much flatter than would be predicted by the best response. Whereas theory would predict that effort increases by six points when beliefs fall by ten points, the actual response to beliefs is significantly smaller. For agents, a ten-point reduction in beliefs about the controlling parties’ effort leads to only a 1.6-point increase in effort.30
E. The Motivational and Demotivational Forces of Authority: Nonpecuniary Explanations
Thus far we have seen that for both the controlling party and the subordinate, deviations from best response behavior account for much of the observed departure from the Nash predictions. A significant proportion of controlling parties provide
29 Between 18 and 33 percent of agents exert zero effort in nine or ten periods. Between 8 and 50 percent of principals who delegated at least once exerted zero effort after each delegation.
30 For principals, effort increases by 3.8 points when beliefs fall by ten points. Both coefficients are significantly smaller than the theoretically expected six-point increase ( p < 0.01) in a simple regression of effort on beliefs. Table C4 in online Appendix C reports beliefs about controlling party effort in comparison to actual controlling party effort choices.
Table 8—Comparison of Effort of Subordinates to Their Best Response to Beliefs
Agents in subordinate role Principals in subordinate role
Actual Best response Percent Actual Best response Percent effort to beliefs zero effort to beliefs zero
PLOW 22.8 21.1 39.0 16.5 24.2 36.7 LOW 14.3** 19.8 49.4 16.2 18.9 54.0 HIGH 26.5 24.6 28.5 19.6 26.3 36.8 PHIGH 17.3 18.4 50.3 20.7 22.6 36.7
Note: Significance levels calculated using a signed-rank test with beliefs and effort averaged by individual prior to estimation.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1354 THE AMERICAN ECONOMIC REVIEW JuNE 2013
effort which exceeds the best response function, leading to effort levels higher than predicted. Similarly, a significant proportion of subordinates provide zero effort despite the almost negligible cost of providing incremental effort. Having estab- lished these observational facts, the question remains which behavioral force shows promise in rationalizing our data.
As we saw in the delegation section, many aspects of our data support the inter- pretation that regret aversion is an important force behind the underdelegation of authority. Might the same behavioral force also have promise to rationalize the observed deviations in effort choices?
To begin our analysis, we take the same overruling effect which was found to be of importance in delegation and ask to what extent it can explain the effort pat- terns of a subordinate. Recall that as a subordinate there is the potential of being overruled. If such overruling generates nonpecuniary losses—as predicted by regret aversion—a subordinate who anticipates this disutility may reduce effort, thereby reducing the probability with which overruling occurs (see online Appendix A for more details). For cases where this nonpecuniary loss is particularly strong, effort provision in the subordinate role may be driven to zero.
The explanation that zero effort is a response to anticipatory regret would receive support if our individual measure of regret aversion would correlate with individual differences in zero effort. Our interpretation is supported in precisely this way: in a probit regression, the probability that a subordinate exerts zero effort increases by 5.9 percent per additional gamble rejected ( p = 0.030).31
Regret in the domain of effort choices may also result in an overprovision of effort by the controlling party if there is a nonpecuniary disutility for being unsuccessful in implementing one’s own preferred project. Filiz-Ozbay and Ozbay (2007) have shown that individuals who lose in first price sealed bid auctions to bids which are below their true valuation behave as if they experience “loser regret,” i.e., these individuals behave as if they suffer a disutility from losing when they could have
31 A dummy variable is created that takes on the value 1 in case of zero effort provision in the role of the subordi- nate. Data from the four main treatments is included in this regression. The probit regression also contains controls for the treatment, being in the role of the agent, interactions of role and treatment, controls for beliefs, and period dummies. Standard errors are clustered at the individual level.
Table 9—Comparison of Actual Beliefs of Subordinates to the Nash Prediction
Agents in subordinate role Principals in subordinate role
Nash Actual Nash Actual prediction belief prediction belief
PLOW 55*** 64.8 45*** 59.8 LOW 55*** 66.9 55*** 68.4 HIGH 45*** 59.0 45*** 56.2 PHIGH 45** 69.3 55*** 62.3
Note: Significance levels calculated using a signed-rank test with beliefs and effort averaged by individual prior to estimation.
*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
1355fehr et al.: lure of authorityVol. 103 No. 4
won, and earned profit, by making a higher bid. In our setting, the controlling party may, hence, regret his or her effort choice in cases where it remains uninformed and thus could have improved the project selection by putting in more effort. An individual in the role of the controlling party who anticipates such regret optimally raises his or her effort (see online Appendix A for more details). Again, we can test this conjecture by using regression analysis and looking at the correlation between overprovision of effort and our individual measure of regret aversion. In a probit regression, we find that the probability of overprovision increases by 4.5 percent per additional gamble rejected ( p = 0.058).32 Thus, taken together, regret aversion may be a driving force behind all three major experimental patterns—the underdelega- tion of authority, the high frequency of zero effort choices among subordinates, and the overprovision of effort by the controlling parties.33
While regret aversion shows promise in rationalizing our data, it is reasonable to ask whether alternative hypotheses show similar promise. Could it be, for example, that the agents’ effort choices are influenced by reciprocity or other forms of social pref- erences? Or, may risk or loss aversion generate a similar pattern to regret aversion?
A common reason for deviations from standard economic predictions is the exis- tence of social preferences. In our setting, if agents view the delegation of authority as a kind act they may overprovide effort because of reciprocal motivations. Likewise, if they view a lack of delegation as an unkind act they may underprovide effort relative to their best response. Thus, positive and negative reciprocity may, in principle, explain the agents’ effort pattern. We tested for the impact of reciprocity motives by conduct- ing an additional treatment in the HIGH condition in which the delegation decision was decided exogenously by the computer. In this HIGH RAND treatment, a virtual coin is flipped each period which determines whether control rights are kept by the principal or whether the principal is forced to delegate them. Since the agents know that the principals are forced to make a choice it is impossible to attribute kind or unkind intentions to the principal. If positive or negative reciprocity play a role, the agents’ effort choices in the HIGH RAND condition will deviate from their choices in the HIGH condition. However, neither as a controlling party (Kolmogorov Smirnov test, p = 0.20) nor in the position of the subordinate party (Kolmogorov Smirnov test, p = 0.99) do the agents’ effort choices differ in the two conditions, implying that reciprocity is unlikely to explain their effort pattern.
In all of our treatments, the controlling party overprovides effort relative to her best response which directly increases the expected earnings of the subordinate. Thus, altruism on the part of the controlling parties could explain this pattern of effort. To control for this possibility we implemented an additional control treatment with the following features. Only one of the two subjects was given the ability to
32 A dummy variable is created that takes on the value 1 if a subject overprovided effort in the role of the control- ling party on average. Data from the four main treatments are included in this regression. The regression also con- tains controls for the treatment, being in the role of the agent, and interactions of role and treatment. Each individual in our dataset who chose effort in the controlling party role at least once is an observation.
33 As was stated in result 6, we also observe a minority of subordinates who actually overprovide effort relative to their best response. In online Appendix A we show that heterogeneity in subordinate effort is also consistent with a model with both loser regret and overrule regret. Recall that the role of controlling party is influenced only by loser regret, and therefore individuals are predicted to exert effort above the Nash equilibrium in this role. As the subordinate, an individual is exposed both to the potential of being overruled and to the potential of loser regret. These forces go in opposite directions, and the response in effort will therefore depend on an individual’s inclination toward both types of regret.
1356 THE AMERICAN ECONOMIC REVIEW JuNE 2013
provide effort and to choose the project, but both parties were paid based on the controlling party’s project choice. Thus, in this treatment the passive party never receives the decision right and never makes an effort choice but only collects her payoffs. We compare this treatment with the single player game (described at the beginning of Section IIC) which is identical to the above control treatment except that no passive recipient exists. Thus in the additional control treatment social pref- erences can affect the active subject’s effort, while in the single player game social preferences cannot play a role. It turns out that the effort choices of the active party and the single player are indistinguishable (Kolmogorov Smirnov test, p = 0.44), indicating that social preferences do not significantly affect effort.
Another potential reason for deviations from theoretical predictions is that the assumption of expected value maximization may be violated due to loss aversion. In the online Appendix D, we show that loss aversion with a refer- ence point at the outside option cannot explain the overprovision of effort by the controlling parties. The intuitive reason for this claim is as follows. For loss averse individuals, an increase in effort above the risk neutral optimum increases the magnitude of a potential loss which reduces utility. This follows from the fact that an increase in effort causes a sure increase in costs, but as long as the possibility of success is below 1 the controlling parties’ ex post payoff from unsuccessful search may not cover the effort cost. Thus, for rea- sonable amounts of loss aversion, optimal effort is decreasing in an individual’s degree of loss aversion. For unrealistically extreme levels of loss aversion, an individual may prefer to guarantee a payoff rather than playing any lottery. For controlling parties with such extreme levels of loss aversion, providing maxi- mal effort (which guarantees a payoff of 15) may be preferable to providing low effort and hoping for success by the subordinate. In these cases, loss aver- sion would predict a maximal effort level of 100.
Looking at both cases in combination, loss aversion cannot explain effort levels which are above the best response function but below an effort level of 100. As these are the observations which need to be rationalized in order to explain the overprovi- sion of effort by the controlling parties, loss aversion cannot explain our effort results.
In online Appendix E, we show that similar to loss aversion, neither risk aversion nor risk loving preferences can account for overprovision of effort by the control- ling party. Moreover, neither risk nor loss aversion can explain the subordinates’ choice of zero effort levels, because effort costs are negligible at low effort levels. For example, assuming risk aversion, beliefs about controlling party effort need to be extremely high to rationalize a subordinate’s effort choice of zero. Using a CRRA utility specification of the following form, U(x) = x
1−σ _
1 − σ , an effort of 0 is predicted only if the belief in controlling party effort is 100 (up to σ = 8). Hence, only for counterfactually high beliefs (or unrealistically extreme levels of risk aversion) is zero effort expected to occur.
IV. Conclusion
Authority and power permeate political, social, and economic interactions. It is therefore important to understand their motivation and incentive effects. In this
1357fehr et al.: lure of authorityVol. 103 No. 4
article we tackle this question by using a novel experimental design. We find a strong behavioral bias among principals to retain authority against their pecuniary interests and often to the disadvantage of both the principal and the agent. We demonstrate that underdelegation cannot be attributed to principals’ beliefs nor incomplete learn- ing, and that the individual and aggregate income losses of this delegation bias are substantial. Our results suggest that authority has nonpecuniary consequences that inhibit the reallocation of authority.
Our results also show that authority has effects on the motivation to provide effort that are not captured by the theoretical model. The fundamental trade- off between incentives and control, as modeled by Aghion and Tirole (1997), indeed exists; relative to the first-best the subordinate provides too little effort, and the controlling party provides too much. Further, the comparative statics between treatments are well met. However, the inefficiency generated by the incentive conflict is much greater than predicted by theory. The controlling parties provide significantly more effort, and the subordinate parties provide significantly less relative to the Nash equilibrium prediction. For controlling parties and a large fraction of subordinates, this is also true relative to the best response to beliefs.
A deeper look at our data suggests that a distaste for being overruled is a sub- stantial determinant of the desire to retain control. Principals who are overruled after delegation and earn the return from the agent’s preferred project are less likely to delegate in the future relative to those who are uninformed and earn the very same return. This difference in delegation behavior—driven by informa- tional states and not by pecuniary payoffs—suggests that individuals are suffer- ing a nonpecuniary disutility from being overruled. One potential source of such nonpecuniary disutility is regret, a theory which can also help to explain the high frequency of zero effort among subordinates and the overprovision of effort by the controlling parties.
Given the importance of authority and power in the functioning of economic and political organizations we believe that the motivational biases revealed by our data should receive more attention. In addition, further explorations into the motives behind delegation and control are suggested by our data. Although a significant part of underdelegation can be explained by principals’ regret aversion we also observed a nonnegligible underdelegation in the treatment HIGH NOREC, where the princi- pals’ recommendation could not be overruled after delegation. This raises the ques- tion whether some subjects intrinsically prefer to be the controlling rather than the subordinate party. In Bartling, Fehr, and Herz (2013) it is shown that this is indeed the case. Moreover, it is well possible that delegation of authority is affected by the mechanism by which authority has been initially assigned. For example, if the pos- session of authority is the result of prior superior performance, the principals might even be more reluctant to delegate compared to the random assignment of authority. This additional underdelegation may result from the perception of superior com- petence, from overconfidence, or from the status gains associated with a position of authority. In addition, research on the cultural determinants of the sources of underdelegation and the motivational effects of decision rights may be interesting because societies seem to be quite heterogeneous with regard to how they view and legitimize hierarchical structures.
1358 THE AMERICAN ECONOMIC REVIEW JuNE 2013
REFERENCES
Aghion, Philippe, Mathias Dewatripont, and Patrick Rey. 2004. “Transferable Control.” Journal of the European Economic Association 2 (1): 115–38.
Aghion, Philippe, and Jean Tirole. 1997. “Formal and Real Authority in Organizations.” Journal of Political Economy 105 (1): 1–29.
Athey, Susan, and John Roberts. 2001. “Organizational Design: Decision Rights and Incentive Con- tracts.” American Economic Review 91 (2): 200–205.
Baker, George, Robert Gibbons, and Kevin J. Murphy. 1999. “Informal Authority in Organizations.” Journal of Law, Economics, and Organization 15 (1): 56–73.
Bartling, Björn, Ernst Fehr, and Holger Herz. 2013. “The Intrinsic Value of Decision Rights.” Uni- versity of Zurich Department of Economics Working Paper 120.
Bartling, Björn, and Urs Fischbacher. 2012. “Shifting the Blame: On Delegation and Responsibility.” Review of Economic Studies 79 (1): 67–87.
Bester, Helmut, and Daniel Krahmer. 2008. “Delegation and Incentives.” RAND Journal of Econom- ics 39 (3): 664–82.
Bowles, Samuel, and Herbert Gintis. 1988. “Contested Exchange: Political Economy and Modern Economic Theory.” American Economic Review 78 (2): 145–50.
Charness, Gary, Ramón Cobo-Reyes, Natalia Jiménez, Juan A. Lacomba, and Francisco Lagos. 2012. “The Hidden Advantage of Delegation: Pareto Improvements in a Gift Exchange Game.” American Economic Review 102 (5): 2358–79.
Coffman, Lucas C. 2011. “Intermediation Reduces Punishment (and Reward).” American Economic Journal: Microeconomics 3 (4): 77–106.
Dahl, Robert. 1957. “The Concept of Power.” Behavioral Science 2 (3): 201–15. Datta, Somnath, and Glen A. Satten. 2005. “Rank-Sum Tests for Clustered Data.” Journal of the Amer-
ican Statistical Association 100 (471): 908–15. Datta, Somnath, and Glen A. Satten. 2008. “A Signed-Rank Test for Clustered Data.” Biometrics
64 (2): 501–507. deCharms, Richard. 1968. Personal Causation: The Internal Affective Determinants of Behavior. New
York: Academic Press. Deci, Edward L. 1981. The Psychology of Self-Determination. Lexington, MA: Lexington Books. Dessein, Wouter. 2002. “Authority and Communication in Organizations.” Review of Economic Stud-
ies 69 (4): 811–38. Falk, Armin, and Michael Kosfeld. 2006. “The Hidden Costs of Control.” American Economic Review
96 (5): 1611–30. Fehr, Ernst, Holger Herz, and Tom Wilkening. 2013. “The Lure of Authority: Motivation and Incentive
Effects of Power: Dataset.” American Economic Review. http://dx.doi.org/10.1257/aer.103.4.1325. Fehr, Ernst, and John A. List. 2004. “The Hidden Costs and Returns of Incentives: Trust and Trustwor-
thiness among CEOs.” Journal of the European Economic Association 2 (5): 743–71. Fehr, Ernst, and Bettina Rockenbach. 2003. “Detrimental Effects of Sanctions on Human Altruism.”
Nature 422 (3): 137–40. Filiz-Ozbay, Emel, and Erkut Y. Ozbay. 2007. “Auctions with Anticipated Regret: Theory and Experi-
ment.” American Economic Review 97 (4): 1407–18. Fischbacher, Urs. 2007. “z-Tree: Zurich Toolbox for Ready-Made Economic Experiments.” Experi-
mental Economics 10 (2): 171–78. Gelman, Andrew, and Gary King. 1990. “Estimating Incumbency Advantage without Bias.” American
Journal of Political Science 34 (4): 1142–64. Greiner, Ben. 2004. “The Online Recruitment System ORSEE 2.0—A Guide for the Organization
of Experiments in Economics.” University of Cologne Department of Economics Working Paper Series in Economics 10.
Grossman, Sanford J., and Oliver D. Hart. 1986. “The Costs and Benefits of Ownership: A Theory of Vertical and Lateral Integration.” Journal of Political Economy 94 (4): 691–719.
Harsanyi, John C. 1971. “The Dimension and Measurement of Social Power.” In Power in Economics, edited by K. W. Rothschild, 77–96. Harmondsworth, UK: Penguin Books.
Hart, Oliver, and John Moore. 1990. “Property Rights and the Nature of the Firm.” Journal of Politi- cal Economy 98 (6): 1119–58.
Hart, Oliver, and John Moore. 1995. “Debt and Seniority: An Analysis of the Role of Hard Claims in Constraining Management.” American Economic Review 85 (3): 567–85.
Jensen, Michael C. 1986. “Agency Costs of Free Cash Flow, Corporate Finance, and Takeovers.” Amer- ican Economic Review 76 (2): 323–29.
1359fehr et al.: lure of authorityVol. 103 No. 4
Loomes, Graham, and Robert Sugden. 1982. “Regret Theory: An Alternative Theory of Rational Choice under Uncertainty.” Economic Journal 92 (368): 805–24.
Marx, Karl. 1867. Das Kapital. Hamburg: Verlag von Otto Meisner. Mulder, Mauk. 1975. The Daily Power Game. Leiden: Martinus Nijhoff. Niskanen, William A. 1971. Bureaucracy and Representative Government. Chicago: Aldine-Atherton. Parsons, Talcott. 1963. “On the Concept of Political Power.” Proceedings of the American Philosophi-
cal Society 107 (3): 232–62. Rabin, Matthew. 2000. “Risk Aversion and Expected-Utility Theory: A Calibration Theorem.” Econo-
metrica 68 (5): 1281–92. Rantakari, Heikki. 2008. “Governing Adaptation.” Review of Economic Studies 75 (4): 1257–85. Rotter, Julian B. 1966. “Generalized Expectancies for Internal versus External Control of Reinforce-
ment.” Psychological Monographs: General and Applied 80 (1): 1–28. Russell, Bertrand. 1938. Power: A New Social Analysis. London: George Allen and Unwin. Simon, Herbert A. 1951. “A Formal Theory of the Employment Relationship.” Econometrica 19 (3):
293–305. Sliwka, Dirk. 2001. “On the Costs and Benefits of Delegation in Organizations.” Journal of Institu-
tional and Theoretical Economics 157 (4): 568–90. Weber, Max. 1978. Economy and Society. Translated by Guenther Roth and Claus Wittich. Los Ange-
les: University of California Press. Zeuthen, Frederik. (1930) 1968. “Economic Welfare.” In Problems of Monopoly and Economic Wel-
fare, 104–51. London: Routledge and Kegan Pail Ltd.
Copyright of American Economic Review is the property of American Economic Association and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
- The Lure of Authority: Motivation and Incentive Effects of Power
- I. Theoretical Motivations
- A. Analysis and Theoretical Implications
- II. The Experiment
- A. The Authority Game
- B. Experimental Design and Hypotheses
- C. Procedures
- III. Experimental Results
- A. The Main Facts
- B. Exploring the Principals’ Reluctance to Delegate
- C. Exploring the Controlling Parties’ Overprovision of Effort
- D. Exploring the Subordinate Parties’ Underprovision of Effort
- E. The Motivational and Demotivational Forces of Authority: Nonpecuniary Explanations
- IV. Conclusion
- REFERENCES