Rank as an Inherent incentive: Experimental Evidence from a Cognitively Challenging Task

Tony So

University of Auckland

Abstract

We study the impact of relative performance feedback as an inherent incentive mechanism to enhance productivity in a cognitively challenging task. In each of multiple rounds subjects are shown two cue values, Cue A and Cue B, and asked to predict the value of a third variable X, which is a function of the two cue values. We use forecast errors, the absolute difference between the predicted value of X and the actual value of X, as the metric for performance. Our treatments include: (1) piece rates, where subjects are paid on the basis of only their own absolute errors; (2) piece-rate-win-lose, where subjects are paired and paid a piece rate that depends on their own absolute errors only, but informed about whether they did better or worse than their partners; (3) a two-person winner-take-all-tournament where subjects are paired and the one with the smaller error earns a positive payoff while the other earns nothing. We find that average forecast errors are smaller in the piece-rate-win-lose treatment, compared to the piece-rate and the winner-take-all-tournament treatments, with no difference between the last two.

JEL Codes: C91; J24, J33, J39

Keywords: Experiments; Payment schemes; Tournaments; Piece-rates; Productivity; Learning

1

Preface

This paper is based upon my recently submitted PhD thesis at the University of

Auckland titled: Essays in Behavioural Labour Economics, supervised by Professor Ananish

Chaudhuri and Dr Ryan Greenaway-McGrevy. The thesis analyses how different pay

schemes and the provision of relative performance feedback affects worker performance.

This paper focuses on the notion of tournament decomposition – separating the effects

of competition for rank and competition for payoffs, both of which may play a part in

motivating performance under tournaments. As will be discussed in greater detail in this

paper, competition for rank refers to people’s desire to compete in order to achieve

favourable comparisons with others, while competition for payoffs refers to the desire to

compete for the monetary prizes which are associated with how people rank within the

tournament.

2

1. Introduction

A fundamental question in labour economics relates to how workers are paid and how

productivity is affected by the choice of payment schemes. There is a vast literature looking

at the motivational aspects of various pay schemes. See Prendergast (1999) for a review.

Conventional wisdom in economics suggests that incentive schemes that pay on the basis of

performance, such as output-dependent piece rates, are effective in inducing better

performance (for example, see Lazear (2000)).

Piece rates, however, are inapplicable in situations where output cannot be easily

observed or measured. In these instances, an employer could implement tournament pay

schemes as they do not rely on absolute performance. Rank-order tournaments pay workers a

fixed amount depending on how their performance ranks amongst others, with the ‘prize’

monotonically increasing with rank such that those with a higher rank receive higher pay.

Theoretical analyses of tournaments (Lazear & Rosen, 1981; Green & Stokey, 1983;

Nalebuff & Stiglitz, 1983) show that they are effective in eliciting output at a level analogous

to piece rates. This result is borne out by results in a laboratory experiment by Bull, Schotter,

and Weigelt (1987), where they show that, on average, numerical effort choices made under

tournaments are statistically not different than those under piece rates, though variance of

effort choices under tournaments are larger. There is empirical evidence suggesting that the

rank-dependent incentives of tournaments positively influence behaviour in various field

settings such as corporate promotions and sporting competitions (for example see Eriksson,

1999; Ehrenberg & Bognanno, 1990; Taylor & Trogdon, 2002).

However, while prior work in the area has compared tournaments with piece rates,

one issue has not received enough attention. In moving from piece-rates to tournaments, the

nature of the competition and the inherent incentives change with two distinct components to

that change. The first component – and the more obvious one – arises from the fact that in a

tournament players are competing for a higher payoff, which we will refer to as competing

for payoffs; in a piece rate, payoffs depend on only on one’s own performance while in a

rank-order tournament it depends on one’s rank. If the tournament happens to be of a winner-

take-all type, then coming second implies zero monetary payoff.

But there is a second component to this change, since, in a tournament, agents must

outperform their peers in order to attain a higher rank. While a higher rank may correspond to

3

a higher tangible reward (such as promotion tournaments), agents may simply be motivated

by the higher rank itself, in the sense that they derive pleasure or pain from the act of winning

or losing respectively (as in a friendly game of tennis, squash or chess).1 We will refer to this

loosely as competing for rank. There is ample evidence that information about one’s relative

rank, vis-à-vis one’s peers, has a positive impact on performance, even when that higher rank

does not translate into higher monetary payoffs. We provide a brief overview of this literature

in Section 2 below.

In this paper we study the distinction between the notions of competing for rank and

competing for payoffs and their impact on performance in a multiple cue probabilistic

learning (MCPL) task introduced by Brown (1995, 1998). This is a cognitively challenging

task where in each of multiple rounds subjects are shown two cue values and asked to predict

the value of an unknown variable, which is a function of those two cue values. The cue

values shown to subjects change from one round to the next but the underlying function does

not. Subjects do not know what the underlying functional form is but they do know that this

function remains unchanged in every round. The goal for the subjects is to make accurate

predictions on the basis of the cue values shown to them in each round, where accuracy is

measured by the forecast error – the absolute deviation of their predicted value from the

actual value of the variable in that round. Henceforth, we will refer to forecast errors, rather

loosely, as “productivity”, in the sense that smaller errors represent higher productivity and

vice versa.

We implement three different treatments: (1) piece rates, where subjects are paid on

the basis of their own forecast errors; (2) piece-rate-win-lose, where subjects are paired and

paid a piece rate that depends on their own absolute errors only, but informed about whether

they did better or worse than their partners; (3) a two-person winner-take-all-tournament

where subjects are paired and the one with the smaller error earns a positive payoff while the

other earns nothing. Our primary aim is to understand which one of these treatments leads to

the lowest forecast errors.

We explain our design in greater detail in Section 3 below. In summary, we will show

the following. Average productivity is highest (and forecast errors smallest) under the “piece-

rate win-lose” treatment where subjects get paid a piece rate but are also provided with rank

information even though the latter does not impact their earnings. Average productivity in

1 The idea behind rank competition in our study is similar to what Kräkel (2008) describes as “emotions”, where positive and negative emotions are derived from winning and losing respectively.

4

this treatment is higher (and average errors smaller) compared to that under a piece rate and a

winner take-all tournament. Echoing the results of Bull et al. (1987), average productivity is

not different under piece rates or tournaments. Our results suggest that, in the MCPL task that

we implement, competing for rank provides better incentives than competing for payoffs.

We proceed as follows. In Section 2 we provide a brief overview of the related

literature that looks at the impact of rank information on productivity. In Section 3 we explain

our experimental design and present several hypotheses. In Section 4 we discuss our findings

and finally in Section 5 we make some concluding comments.

2. Impact of Rank Information

As noted above, one thing that changes in going from a piece rate scheme to a

tournament is that subjects now get to see how they are performing vis-à-vis their peers.

There is now a large literature which suggests that information about rank – even if that rank

does not affect monetary payoffs – has implications for worker productivity. Mas and Moretti

(2009) show that supermarket checkout operators who are paid fixed wages became more

productive when, with a change in shifts, the average productivity of their peers increase.

This effect was found to be attributable to workers who were directly in the line of sight of

co-workers, implying that agents’ productivity improved when they were being observed by

others. In other words, peer pressure is found to have a motivating effect.

Productivity spill-overs and the incentive to free-ride off the effort of fellow workers

are key features of the Mas and Moretti (2009) study. Falk and Ichino (2006) look into peer

effects in the absence of productivity spill-overs. They analyse levels and variance of

performance when students were paid a flat wage to stuff envelopes. In the control treatment,

subjects work alone while in an experimental treatment they work on their own but in the

presence of a second subject located in the same room. In the latter setup, subjects are able to

observe each other’s output. The authors find that performance (in terms of number of

enveloped stuffed) was higher in the experimental treatment where subjects work in pairs.

Furthermore, variation of output within pairs is lower than between pairs, indicating that peer

effects mutually motivated each individual to perform at a level similar to her partner.

Blanes i Vidal and Nossol (2011) assess the effect of relative performance feedback

on the performance of German warehouse workers, who are paid a piece rate on top of a base

5

salary.2 Here the workers were notified two months in advance that they would be receiving

additional rank information in their payslips. The revelation of rank information was found to

have a positive effect on productivity. Moreover, a distinct positive effect on productivity

exists even prior to the relative rank information being revealed as long as subjects are aware

that it is impending. In other words, both the anticipation and revelation of relative feedback

has a motivating effect on performance.

In Kuhnen and Tymula (2012) participants solve multiplication problems over a

number of timed rounds and are paid a fixed salary for their participation. In one treatment

participants receive relative performance feedback while in a second they receive such

feedback with 0.5 probability while in a third treatment no feedback is provided. Players in

both certain feedback and probabilistic feedback performed better than those who did not

receive feedback while there are no differences in performance in the first two treatments. It

appears that while feedback matters even the likelihood of receiving feedback can serve as a

motivating force.

Cadsby, Engle-Warnick, Fang and Song (2010) run experiments with Chinese

students and Chinese factory workers, where participants are asked to add five 2-digit

numbers over multiple rounds. The study has a 2 x 2 design which varies payment type –

rank-order tournament or fixed salary – and the nature of feedback after each round – public

or private. Students paid under a tournament scheme perform better when the rank

information is public while the nature of feedback had no effect for factory workers.

Two papers that fail to find a positive impact of peer pressure are Eriksson et al.

(2009) and Bellamare et al. (2010). Eriksson et al.’s subjects undertake a number adding task

like Cadsby et al. while the feedback is either continuous (i.e., up to date feedback is

available at every point in time) or discrete (provided once at the half way mark). There is a

third treatment with no feedback. Eriksson et al. (2009) found no difference in the number of

correct answers between participants who were provided either discrete or continuous

feedback compared to participants who do not receive relative feedback at all. Eriksson et al.

suggest that performance was high in the no feedback treatment and hence there was not

much scope for upward movement in the feedback treatments which may explain the lack of

an effect.

2 In the literature it is usually assumed that payment schemes that combine fixed salaries with piece rates have the same incentive properties as that of the latter.

6

In Bellemare et al. (2010) subjects undertake a data-entry task and are paid either a

fixed wage or a piece rate. Bellemare et al. (2010) report an inverse U-shaped impact of peer

pressure for men under fixed wages in that extremely low and extremely high levels of peer

pressure affected performance adversely. But, by and large, the provision of relative

performance feedback had little impact on performance. However, as opposed to other

studies where relative information is provided in real time with all subjects performing the

task concurrently, Bellemare et al. provide relative feedback by comparing the behaviour of a

current participant with that of someone who took part in the experiment on a previous

occasion. It is possible that this design feature made the relative performance feedback less

salient for subjects.

Two further papers find mixed evidence regarding the effect of providing relative

feedback. Both Hannan, Krishnan and Newman (2008) and Azmat and Iriberri (2016) find

that the provision of relative performance feedback has different effects depending on how

people are paid. In Hannan et al. (2008) subjects participate in a stylised profit maximisation

problem where they took the role of a manager and selected the output level in order to

maximise profit under uncertainty. The manager was paid according to either a tournament

bonus or a piece rate commission on profit earned. For each of these two pay schemes,

subjects in different treatments either received feedback about their rank or not. The

feedback, when provided, could be either coarse or fine. The former notifies the subject

whether the profit points earned is above or below the median, while the latter is more precise

indicating their performance decile. Hannan et al. find that the firm’s profits are higher under

piece rates when relative feedback is provided than when it is not. On the other hand, the

provision of feedback has no effect in tournaments when relative feedback is coarse and

actually worsens earnings when fine feedback is provided.

In an arithmetic task, Azmat and Iriberri (2016) look at the effect of relative feedback

on piece rates and flat salaries. Relative feedback improves performance only under piece

rates, while having no effect whatsoever under salaries. Finally, Azmat and Iriberri (2010)

and Tran and Zeckhauser (2012) use natural data to examine the effect of relative

performance feedback. Azmat and Iriberri (2010) use data from 1986-94 for Spanish high

school students to understand whether providing relative rank information leads to improved

student achievement. In the academic year 1990-91, due to exogenous changes, student report

cards provided information about the average class grade alongside their own grade. This

resulted in students attaining higher grades that year compared to previous and subsequent

7

years where no such relative feedback was provided. Similarly, Tran and Zeckhauser (2012)

found that Vietnamese English-language students who were notified of how they were ranked

within their class, performed better than the control group who were not provided such

information. It made no difference whether feedback was made publicly to all members of

the class or privately to individual students.

3. Experimental Design

3.1. Task

Our experiment is based on a multiple cue probabilistic learning (MCPL) task, where

in each one of 20 rounds (t) participants are required to predict the value of a variable Xt

based on the observation of two numerical “cues” provided to them3. This task is

representative of real life tasks. For example, the variable Xt can be thought of as the

underlying price of a stock; the cues as variables that affect the value of a stock; and the task

at hand as one of forecasting stock prices. The actual stock value is determined by the

underlying equation:

where Xt is the actual stock value participants are required to predict, Cue At and Cue

Bt are the values of the two numerical cues provided to the participant, and εt is a random

variable with uniform distribution drawn from the set [-5, 5] in each round t. Participants do

not know about the error term, the exact relationship between the cue values and the value of

X, or even whether the relationship is linear or non-linear. They do know, however, that

while the cue values change from one round to the next, the underlying relationship does not

change.

We implement two variants of the task. In the “Single Cue” task, Cue A is fixed at

the value of 150 for each of the 20 rounds, while Cue B changes each round. This is designed

to be less difficult than the “Dual Cue” task, where both cue values change in each round.

For both tasks, the sequence in which the cue values appear from one round to the next, is

identical across treatments. Table 1 shows the cue values and corresponding stock prices for

each round.

3 MCPL tasks are commonly used in psychology to study learning (see Balzer, Doherty & O’Connor, 1989 for a review). In economics, besides Brown (1995, 1998), this task has been used by Vandegrift and Brown (2003), Vandegrift, Yavas, and Brown (2007) as well.

8

Our metric of performance is the absolute forecast error, the absolute distance

between the predicted value (XtP) and the actual value Xt

*, (|XtP- Xt

*|). Forecast errors measure

the accuracy of the predicted value, so smaller errors imply more accurate forecasts and

therefore better performance (and higher productivity). Absolute errors are an appropriate

measure of performance as they reflect the cognitively challenging nature of the MCPL task.4

Table 1: Actual Cues and Stock Prices

RoundSingle Cue Task Dual Cue Task

Cue A Cue B Stock Price Cue A Cue B Stock

Price1 150 201 192 105 37 692 150 263 243 242 96 1513 150 88 117 443 159 2564 150 248 232 1 339 2455 150 201 200 41 146 1246 150 196 194 155 32 807 150 353 305 20 288 2238 150 173 173 104 422 3359 150 270 248 102 107 11210 150 243 222 296 188 23111 150 60 102 413 266 32112 150 320 274 165 412 35313 150 340 289 172 167 17414 150 361 311 359 262 29815 150 321 285 271 418 38516 150 361 309 227 31 9817 150 148 155 381 435 42618 150 309 275 262 339 32319 150 135 145 316 92 16420 150 142 156 196 285 269

3.2. Treatments

We report on data from three treatments for the purposes of this paper, which is part

of a larger study involving other treatments and analyses. These three treatments are: (1)

Piece Rate (PR), (2) Piece Rate Win-Lose (PRWL); and (3) two-person Winner-Takes-All

Tournament (WTAT). The treatment intervention comes into play from round 6.

4 As will be clear from Table 1, the MCPL task is difficult enough even when a single cue is changing. When both cues change the degree of difficulty increases. Subjects figure out relatively quickly that the actual stock price is a convex combination of the two cue values and therefore get reasonably close to the actual value when the cue values are close but the errors get large when the cue values are far apart.

9

In the first five rounds of each of these treatments, subjects are paid piece rates

according to their individual performance. At this point, subjects do not know which

treatment they are assigned to, but do know that treatment interventions might or might not

take place prior to the sixth round. Since these rounds are identical in each of the treatments,

we are able to use performance across these rounds to proxy for subjects’ underlying ability

pertaining to the task. It is important to control for ability given that our MCPL task is

relatively difficult and given the possibility of significant heterogeneity in ability levels

across subjects.

The piece rate that takes place across rounds 1 to 5 in each of the treatments pays

participants NZ $1 minus their forecast error for the round. For example, if the absolute error

is 20 then the payment for that round is NZ $0.80. If the forecast error is greater than 100,

then earnings for that round is set to zero. Here subjects aim to minimise their forecast error

in each round, which in turn will lead to higher payoff.

From round 6 onwards, treatment interventions take place. In the PR treatment,

subjects continue to earn piece rates on their individual performance across rounds 6 to 20,

and subjects are instructed accordingly. Figure 1A presents a screenshot to show what the

subjects get to see at the end of a round. This is the information that a subject will be looking

at prior to the beginning of round 10.

The PRWL treatment adds rank competition on top of the PR treatment. Here, starting

with round 6, subjects are paired, with random re-matching of pairs between rounds. They are

paid according to their own forecast errors in each round, i.e. the payment scheme is the same

piece rate. Furthermore, from round 6 onwards, in each round the subjects are also told

whether they have “Won” or “Lost” depending on whether a particular subject’s error was

smaller than or larger than her pair member respectively. However, whether a subject won or

lost a particular round has no bearing on her earnings for that round since each subject

continues to get paid on the basis of her own forecast errors. The rank information is simply

designed to capture positive or negative emotions from winning or losing respectively. Figure

1B shows a screenshot of this treatment.

10

Figure 1A: Screenshot for PR treatment

10 Tso1

1 296 188

Round

1

Cue A

105

Cue B

37

Your Forecast

75

Actual Price

69

Forecasting Error

6

Earnings this round

$0.94

Round number Player name

Player ID Cue A

2 242 96 201 151 50 $0.50 Cue B 3 443 159 174 256 82 $0.18 4 1 339 268 245 23 $0.77 Enter Forecast 5 41 146 113 124 11 $0.89 6 155 32 116 80 36 $0.64 SUBMIT RESET 7 20 288 241 223 18 $0.82 8 104 422 315 335 20 $0.80 9 102 107 108 112 4 $0.96

Figure 1B: Screenshot for PRWL treatment

10 Tso1

1 296 188

Round Cue A Cue B Your Actual Forecasting Earnings WIN Round number

1

105

37

Forecast

75

Price

69

Error

6

this round

$0.94

or LOSE

---

Player name Player ID

Cue A 2 242 96 201 151 50 $0.50 --- Cue B 3 443 159 174 256 82 $0.18 --- 4 1 339 268 245 23 $0.77 --- Enter Forecast 5 41 146 113 124 11 $0.89 --- 6 155 32 116 80 36 $0.64 WIN SUBMIT RESET 7 20 288 241 223 18 $0.82 WIN 8 104 422 315 335 20 $0.80 LOSE 9 102 107 108 112 4 $0.96 WIN

Figure 1C: Screenshot for WTAT treatment

11

10 Tso1

1 296 188

Round Cue A Cue B Your Actual Forecasting Earnings WIN Round number

1

105

37

Forecast

75

Price

69

Error

6

this round

0.94

or LOSE

---

Player name Player ID

Cue A 2 242 96 201 151 50 0.5 --- Cue B 3 443 159 174 256 82 0.18 --- 4 1 339 268 245 23 0.77 --- Enter Forecast 5 41 146 113 124 11 0.89 --- 6 155 32 116 80 36 $1 WIN SUBMIT RESET 7 20 288 241 223 18 $1 WIN 8 104 422 315 335 20 $0 LOSE 9 102 107 108 112 4 $1 WIN

The WTAT treatment also starts in round 6, following five rounds of piece rate

payments. As in the PRWL treatment, from round 6 onwards, we form subjects into pairs

(with random re-matching from one round to the next), except here we implement a winner-

takes-all tournament scheme, where in each round the subject with the smaller absolute error

wins NZ $1 while the subject with the larger error gets zero.5 Figure 1C presents a screen-

shot. Compared to the PRWL treatment, from round 6 onwards, the WTAT treatment not

only provides the win/lose information but changes the payoffs as well. So the WTAT

treatment incorporates payoff competition on top of the rank competition in the PRWL

treatment.

3.3. Research Questions and Hypotheses

In this section we formulate some hypotheses that will guide our analysis later in the

paper. We start by comparing our two pay schemes: piece rates and tournaments. Theory

predicts that the level of effort exerted under tournaments are identical to those under piece

rates, a finding confirmed empirically by Bull et al. (1987). We shall refer to this as Piece

Rate Equivalence. As such, we would expect the PR and WTAT treatments to perform

similarly and have similar forecast errors. This leads to our first hypothesis.

H1.

Piece Rate Equivalence holds, where the productivity of tournaments

are similar to that of piece rates. This implies: PR(Errors)

WTAT(Errors).

5 If the forecast errors of a particular pair are equal in particular round, then the tie is broken by randomisation.

12

Next we look at how rank competition affects productivity. In order to address this we

compare the PR and PRWL treatments. Both treatments are identical, except that the latter

provides additional information at the end of each round about whether subjects performed

better or worse than their partners. Subjects are paid piece rates in both treatments; the rank

information does not impact earnings, allowing us to isolate the effect of rank competition

from payoff competition.

If people are motivated solely by earnings, then we would not expect rank

competition to have any effect on productivity.6 On the other hand if people derive utility

(disutility) from winning (losing) a particular round, then we would expect them to work

harder to improve (reduce) their chances of winning (losing) – improving their forecast

errors. This effect may consist of an ex-ante anticipation effect (Blanes i Vidal & Nossol,

2011) that may be associated with preferences for status and respect (see Ellingsen &

Johannesson, 2007 for a review), or an ex-post revelation effect when people react to the

feedback received. Either of these effects or both, would lead better performance in the

PRWL treatment compared to the PR treatment. This leads to our second hypothesis.

H2.The provision of relative performance feedback improves productivity.

This implies: PRWL (Errors) < PR (Errors).

We now turn to the issue of competing for payoffs by comparing the PRWL and

WTAT treatments. Both treatments feature rank feedback, but differ in terms of incentives:

piece rates and rank-dependent prizes respectively. From Hypothesis H1, we expect

tournaments to perform similarly to piece rates according to Piece Rate Equivalence. Does

this equivalence still hold with our rank-augmented piece rates?

If piece rates perform no differently to tournaments (H1), and rank competition

induces higher performance from players (H2), then we would accordingly expect the

feedback-augmented piece rates (PRWL) to perform better than tournaments (WTAT).

Eriksson et al. (2009) provide evidence indicative of this; piece rates with relative feedback

perform better than tournaments featuring identical feedback. This suggests that the rank-

dependent payoffs inherent in tournaments do not perform as well as piece rates do once rank

6 Under a principal-agent framework, Aoyagi (2010) studies the effect of the principal providing feedback about the intermediate performance gap to agents in a dynamic tournament. He finds that the effect this feedback has on the effort exerted by agents depends on the nature of their cost of effort functions. If agents’ marginal cost of effort are concave (convex), then no (full) feedback induces low effort from agents. If marginal cost is linear, then effort is not affected by feedback policy. We caution applying the insights of Aoyagi to our study of rank competition, since the dynamics will be different according to the different setup (multi-stage tournaments in Aoyagi; finitely repeated single stage tournament here) and pay schemes (tournaments in Aoyagi; piece rates in our study).

13

competition has been controlled for. It is the motivating effect of rank competition that

underpins Piece Rate Equivalence. This leads us to our third hypothesis.

H3.

When relative feedback is added to piece rates, this will lead to better

performance than in tournaments. This implies PRWL (Errors) < WTAT

(Errors).

There are at least two reasons why tournament performance may be worse than that

under piece rates with rank feedback. First, there is an element of uncertainty in rank payoffs.

In order to get a payoff, one has to perform better than one’s peers while not knowing how

others will perform – due to random rematching of partners and since players are only

informed of whether they have won or lost, rather than on the margin of winning or losing.

Under winner-takes-all tournaments, the possibility of losing means that the exertion of

costly effort might not yield any return. For a given degree of risk aversion a person would

likely exert less effort under a tournament than a piece rate, for which payoffs are

deterministic.

The second reason why tournament performance may suffer relates to the concept of

shying away from competition (Niederle & Vesterlund, 2007). Those averse to competition

may find the tournament set-up challenging. While there is an element of competition in both

PRWL and WTAT, since relative performance feedback is provided in both, competition is

more salient in WTAT since payoffs are directly linked to rank. Competition avoidance could

further explain why productivity may be lower in WTAT than in PRWL. While we cannot

explicitly observe instances of this, we control for competition-aversion through trait anxiety

scores, which are elicited before participants learn of the task. See Segal and Weinberg

(1984) for a discussion on how trait anxiety can serve as a proxy for competition avoidance.

To summarize, our three hypotheses jointly predict forecast errors would be such that

PRWL (Errors) < PR (Errors) WTAT (Errors). Taken together this suggests that with the

provision of rank feedback, piece rates would perform better than tournaments, meaning that

competing for rank provides a stronger incentive than competing for payoffs.

3.4. Experimental Procedure

Sessions were conducted at a computer lab in our University using primarily first year

students in commerce. We report data obtained from 236 participants in the three treatments

14

described above. Participants are seated at computer cubicles with privacy partitions and are

cautioned about not communicating with any other subject. To start with, participants are

asked to fill out a questionnaire which elicits participants’ trait anxiety level. See Spielberger,

Gorsuch, Lushene, Vagg and Jacobs (1983). This is shown in the appendix. The

questionnaire consists of 20 questions that are answered on a 1 to 4 scale. Questions 1, 6, 7,

10, 13, 16 and 19 are reverse scored. The questionnaire is designed to measure a subject’s

general tendency to feel anxious rather than their current level of anxiety (McNaughton,

2011). A higher score generated from the pre-task questionnaire indicates a higher level of

trait anxiety associated with the individual. We use trait anxiety as a proxy of each subject’s

competitive preferences, so we can account for potential dropouts as a result of shying away

from rank or feedback competition (Niederle and Vesterlund, 2007). Segal and Weinberg

(1984) find that trait anxiety is higher among women than men which is attributed to

differences in competitive preferences.

Following this we hand out the instructions for the forecasting task. The instructions

are read out loud after subjects have had a chance to read them privately on their own. The

appendix contains a copy of the instructions.

These instructions describe the task, inform players that they will be earning piece

rates in the first five rounds, and that there might be a change in the way the game is played

in rounds 6 to 20. The instructions are identical for people in different treatments. They are

told that they will be provided further information prior to the start of round 6. Subjects are

also provided with ten examples for Cue A, Cue B and X. This is shown in Table 2.

Table 2: Cue Values given to subjects as practice examples

Single Cue Task Dual Cue Task

Cue A Cue B Actual Price Cue A Cue B Actual

Price150 92 117 12 64 54150 143 157 372 63 162150 379 321 179 109 137150 373 313 415 146 240150 240 220 116 186 175150 285 256 355 223 275150 187 188 145 286 255150 143 153 199 356 317150 191 185 439 354 372

15

150 361 311 73 442 345

In the PR treatment, following round 5, they are told that there are no further

instructions and they should continue as before. In the PRWL treatment, after round 5, they

are told that in going forward they will be paired with another player in each round, and

informed of whether they won or lost a round. They are also told that this rank information

has no bearing on their earnings, which still depend only on their absolute errors in any given

round. In the WTAT treatment they are told both about the pairing and that from round 6

onwards they will earn either $1 or nothing in each round. In both the PRWL and WTAT

treatments subjects are told that they will be randomly re-matched from one round to the

next.

At the conclusion of the task, participants are asked to fill out a post-task

questionnaire, which elicits information about participants’ intrinsic motivation, including

self-reports of how competent they felt at the task, how motivated they were, how interesting

they found the task, how much effort they exerted and how close to they felt to other

participants in the room. We also collected basic demographic information including gender,

age and ethnicity. We do not elaborate on the psychological questionnaires since we do not

exploit data from them for the purposes of this study.

After filling out the questionnaire, participants are privately paid their earnings from

the 20 rounds of the forecasting task in cash, plus a $5 show-up fee. On average, participants

earned $20 in each treatment.7

4. Results

Table 3 provides a broad overview of the forecast errors in different treatments along

with the number of subjects and sessions in each treatment.8 Not surprisingly the errors are

much smaller in the single cue task than the dual cue one. What is noticeable is that in both

the single and dual cue tasks, average forecast errors are highest in the WTAT treatment,

7 Under tournaments, the expected earnings in a round is $0.50 (50% chance of winning the $1 prize), which is much lower than the ex-post average earnings under piece rates in the PR and PRWL treatments. In the WTAT treatment, we paid participants an extra $4 to calibrate their average earnings with other treatments.

8 We had 36 subjects in the PRWL treatment but due to reasons beyond our control, one subject left early. Since we needed to put subjects in pairs from round 6 onwards, we discreetly replaced the departing subject with one of our experimental assistants (who had no prior experience with the game). We have excluded the choices made by this subject (and the replacement) from our analysis. However, given the random re-matching of subjects and the fact that subjects never get to see the ID numbers of their partners, we have retained the data for the remaining 35 subjects.

16

followed by the PR treatment. Errors are smallest in the PRWL treatment. To explore these

issues more rigorously we turn to regression analysis next.

As noted before forecast errors serve as our metric for performance (= |Predicted

stock value – Actual stock value|). Smaller forecast errors indicate higher productivity. In the

first three columns of Table 4 we present random effects regressions with forecast errors as

the dependent variable, with robust standard errors clustered by individual subjects. In

running these regressions, we pool the data from the single cue and dual cue tasks. We use a

random effects specification because we have both time-varying and time-invariant variables

among our regressors, where the time-invariant regressors are inestimable under fixed effects

regressions. In Column 4 of Table 4, we also present results of a quantile regression as a

robustness check.

17

Table 3: Average errors across treatments

TreatmentsSingle Cue Dual Cue Pooled Data

Session # N Average

ErrorSessio

n # N Average Error N Average

Error

Piece Rate (PR)

1 1610.2

1 2026.6 81 18.1

2 26 2 19

Piece Rate Win-Lose (PRWL)

1 229.6

1 2024.0 77 16.2

2 20 2 15

Winner Take All

Tournament (WTAT)

1 1610.0

1 1830.7 78 20.1

2 24 2 20

Total 124 112 236

Recall that in all treatments subjects play the PR treatment for the first 5 rounds and

the treatment (if any) is implemented only at the start of Round 6. Hence in running these

regressions we use data for rounds 6 through 20. In our simplest regression model, we regress

forecast errors against two dummy variables which represent the PRWL and WTAT

treatments (with the PR treatment serving as the reference category), as well as a linear time

trend denoted by Round which captures learning over time. Our second regression model

includes additional controls of each subject’s gender (= 1 for women and 0 for men) and trait

anxiety score. Our third regression model builds on the second, but better controls for

learning by allowing time trends to differ across treatments. Our fourth regression model

shares the same specification as Model 2, but instead of being estimated with random effects,

is estimated with a quantile regression. The bottom of Table 4 presents a pairwise Wald test

for each regression model comparing the performance of the PRWL and WTAT treatments.

18

Table 4: Pooled regressions for forecast errors; columns (1) to (3) present results for random effects regressions while column (4) presents results for a quantile regression

Dependent variable = forecast errors = |Predicted value – Actual Value|

Independent variables

Model 1 Model 2 Model 3 Model 4Pooled Pooled Pooled Pooled

Random Effects

Random Effects

Random Effects

Quantile Regression

PRWL-1.897(2.186)[0.385]

-4.388(2.425)[0.070]

-6.267(3.200)[0.050]

-1.318(0.743)[0.076]

WTAT2.052

(2.740)[0.454]

0.503(2.958)[0.865]

3.181(3.390)[0.358]

-0.442(0.750)[0.556]

Female ---8.317

(1.928)[0.000]

8.317(1.929)[0.000]

3.496(0.610)[0.000]

Trait anxiety ---0.162

(0.150)[0.281]

0.162(0.150)[0.281]

0.085(0.044)[0.052]

Round-0.136(0.075)[0.070]

-0.161(0.082)[0.049]

----0.093(0.070)[0.182]

PR X Round --- ----0.142(0.146)[0.331]

---

PRWL X Round --- ---0.003

(0.153)[0.984]

---

WTAT X Round --- ----0.348(0.119)[0.004]

---

Constant19.83

(1.878)[0.000]

10.90(6.164)[0.077]

10.65(6.630)[0.108]

5.915(2.052)[0.004]

R2 0.004 0.033 0.034 0.007(Pseudo R2)

Wald χ2 6.27 29.93 38.92 ---p > χ2 0.099 0.00 0.00 ---

No. of observations 3540 3180* 3180 3180No. of participants 236 212 212 212

PRWL = WTAT χ2 = 2.71p = 0.100

χ2 = 4.06p = 0.044

χ2 = 7.70p = 0.006

F = 1.43p = 0.231

19

* 15 subjects did not provide gender information, 7 subjects did not complete the trait

anxiety questionnaire properly, and 2 providing neither, thereby leading to a loss of 360

observations over 15 rounds. Standard errors in parentheses; p-values in square brackets.

In Model 1 of Table 4, we see that the PRWL treatment dummy is negative while that

for the WTAT treatment is positive. Neither of the coefficients are, however, significant at

conventional levels. When we compare the PRWL and WTAT treatments, the Wald test

shows that the PRWL treatment performs significantly better. From Model 1, we also see that

learning occurs, with forecast errors improving over time.

When we control for gender and trait anxiety in Model 2, we see that the average

forecast errors are significantly lower in the PRWL treatment compared to the reference PR

treatment. A Wald test also show that forecast errors are also smaller in the PRWL treatment

compared to the WTAT treatment. What is also noticeable is the large positive coefficient for

the gender dummy showing that, on average, women performed much worse than men across

the board.

In Model 3, when we allow for treatment-interacted time trends, we continue to see

similar results. The coefficient on the PRWL dummy continues to be negative and the WTAT

dummy remains insignificant, showing that performance in these treatments are, respectively,

better than and no different to the PR treatment. Comparing the PRWL and WTAT treatment

coefficients, the forecast errors continue to be smaller in the PRWL treatment. While the

WTAT treatment does not appear to perform well compared to the other treatments, it is

made up for by better learning. Of the treatment-interacted time trends, only the WTAT trend

is significant and in the direction that indicates improved forecasts over time. We will come

back to this shortly.

As a robustness check, in Model 4 we run a quantile regression, which is robust to

outlier values. However, quantile regressions do not accommodate the panel structure of our

data, nor do they allow us to estimate time trends. The quantile regression in Model 4

corroborates the previous findings.

On the basis of these results, we conclude that forecast errors in the PRWL treatment

are smaller than in the PR and WTAT treatments, with no difference in performance between

the latter two treatments. This provides support for each of our three hypotheses, where

PRWL (Errors) < PR (Errors) WTAT (Errors). We find support for Piece Rate

Equivalence, and we can infer from our results that it is driven by the motivating effects

20

associated with rank competition. When rank competition is controlled for in piece rates,

tournaments do not perform as well.

The regressions in Table 4 had pooled the single and dual cue treatments. We now

repeat our regressions, but separating single and dual cue treatments. Table 5 consists of two

random effects regressions for each of the single and dual cue tasks across rounds 6 to 20.

Similar to Model 3 of Table 4, each of the regression models in Table 5 regress forecast

errors against the treatment dummies (with PR as the reference treatment), gender (with

women = 1), trait anxiety, as well as time trends for each treatment. In disaggregating the

treatments by task difficulty, we are spreading the number of observations thin, effectively

using half the number of observations compared to the previous pooled regressions in Table

4.

In the single cue regression in Model 1 of Table 5, we see that although the PRWL

treatment dummy is insignificant, it carries a negative sign similar to what we observed in the

pooled regressions. The PRWL coefficient in the dual cue regression in Model 2 is also

negative and insignificant, but is larger in magnitude compared to the corresponding single

cue coefficient. While the signs on the PRWL coefficients for both single and dual cue

regressions are consistent with that from the pooled regressions, the fact that neither are

statistically significant suggests that the smaller number of observations plays a role.

Consistent with what we found earlier for the pooled regressions, the WTAT

treatment for both single and dual cue regressions remain insignificant. The Wald test shows

that the PRWL treatment performs no differently from the WTAT treatment in the single cue

regression in Model 1 of Table 5, while it performs better than the WTAT treatment in the

dual cue regression of Model 2.

The patterns of learning are also similar when we disaggregate our results by task

difficulty. In both single and dual cue regressions in Table 5, the time trends estimated for

each treatment show that it is negative and significant only in the WTAT treatment.

21

Table 5: Random effects regressions for forecast errors separated by Single and Dual Cue treatments

Dependent variable = Forecast errors = |Predicted value – Actual Value|

Independent variables Model 1 Model 2Single Cue Dual cue

PRWL-2.519(3.329)[0.449]

-7.224(4.935)[0.145]

WTAT0.243

(3.187)[0.939]

6.282(4.961)[0.205]

Female1.313

(1.432)[0.359]

11.86(2.811)[0.000]

Trait anxiety-0.128(0.116)[0.270]

0.252(0.249)[0.312]

PR X Round-0.158(0.140)[0.260]

-0.127(0.246)[0.606]

PRWL X Round-0.038(0.182)[0.836]

0.055(0.262)[0.833]

WTAT X Round-0.197(0.100)[0.050]

-0.494(0.212)[0.020]

Constant17.34

(5.728)[0.002]

12.34(11.05)[0.264]

R2 0.007 0.044Wald χ2 6.57 31.81p > χ2 0.475 0.00

No. of observations 1620 1560No. of participants 108 104

PRWL = WTAT χ2 = 0.92p = 0.338

χ2 = 6.01p = 0.014

Standard errors in parentheses; p-values in square brackets.

22

4.1. Heterogeneous Ability

In this section we disaggregate analyses by subjects’ ability. Since our MCPL task is

difficult, there could be a large degree of heterogeneity in players’ ability, which would in

turn adversely affect our statistical analyses. Furthermore, it is possible that rank and payoff

competition have different effects for subjects of different ability, which are not reflected by

our aggregated results.

Recall that participants in each treatment play under piece rates in the first five

rounds. Performance in those five rounds then provides us with a benchmark of participants’

ability. We use the median error in the first five rounds to categorise our participants into

“high” or “low” performers according to how they performed during the first five rounds. We

compute the median forecast error for each subject and those subjects whose median error

exceeded the aggregate median during the first five rounds are treated as “low” performers

while those whose median falls below the aggregate median are “high” performers.9 In Table

6 we report results of random effects regressions similar to the ones we presented earlier,

except we break this up by high and low performers as we have previously defined. As in

Table 4, we pool data from the two tasks (single cue and dual cue) for the regressions. We

report two regression specifications for each category of high and low performers. In the first

specification, the regressors include two treatment dummies PRWL and WTAT (with PR as

the reference category), controls for gender (Female = 1 for women, 0 for men) and trait

anxiety, as well as a linear time trend denoted by Round. In the second specification, we

include treatment interacted time trends instead of the aggregate trend.

9 The median error for all subjects across the first five rounds for the dual cue task is 21 while that for the first five rounds of the single cue task is 8. These serve as the ability thresholds which we use to categorise high and low performers. In the dual cue task a subject is classified as a high (low) performer if the median error for this subject during the first five rounds is less than or equal to (greater than) 21. In the single cue task a subject is classified as high (low) performer if the median error for this subject during the first five rounds is less than or equal to (greater than) 8.

23

Table 6: Pooled random effects regression for forecast errors broken up by high and low performers; columns (1) and 2 present results for high performers while columns (3) (4) presents results for low performers.

Dependent variable = Forecast errors = |Predicted value – Actual Value|

Independent variables

High performers Low performersModel 1 Model 2 Model 3 Model 4

PRWL-1.091(2.114)[0.606]

-3.528(3.423)[0.303]

-7.732(4.242)[0.068]

-9.312(5.743)[0.105]

WTAT1.250

(2.831)[0.659]

1.333(4.107)[0.746]

-0.308(4.993)[0.951]

5.292(5.336)[0.321]

Female5.429

(2.110)[0.010]

5.429(2.111)[0.010]

8.754(3.181)[0.006]

8.754(3.183)[0.006]

Trait anxiety-0.028(0.130)[0.830]

-0.028(0.130)[0.830]

0.275(0.215)[0.202]

0.275(0.216)[0.203]

Round-0.280(0.072)[0.000]

----0.022(0.155)[0.889]

---

PR X Round ----0.346(0.124)[0.005]

---0.088

(0.272)[0.746]

PRWL X Round ----0.158(0.108)[0.141]

---0.210

(0.319)[0.511]

WTAT X Round ----0.352(0.138)[0.011]

----0.343(0.200)[0.087]

Constant16.00

(5.416)[0.003]

16.85(5.810)[0.004]

10.41(9.304)[0.263]

8.981(10.06)[0.372]

R2 0.025 0.025 0.031 0.032Wald χ2 22.66 23.58 11.27 16.35p > χ2 0.000 0.001 0.046 0.022

No. of observations 1710 1710 1470 1470No. of participants 114 114 98 98

PRWL = WTAT χ2 = 0.79p = 0.373

χ2 = 1.89p = 0.169

χ2 = 3.48p = 0.062

χ2 = 5.71p = 0.017

24

What stands out is that the performance of the high performers does not change across

treatments. However, low performers perform better in the PRWL treatment than in the PR

and WTAT treatments. In Model 3, we note that the dummy for the PRWL treatment is

negative and significant (at 7%) indicating that the low performers committed smaller errors

compared to the PR treatment. Pairwise Wald tests for the equality of coefficients suggest

that the coefficient for the PRWL treatment is significantly smaller than that of the WTAT

treatment. In Model 4, the coefficient of PRWL narrowly misses conventional levels of

significance (p = 0.105). Once again, the pairwise Wald test suggests that compared to the

WTAT treatment, performance is better in the PRWL treatment.

In terms of learning, in Model 2 we observe significant learning for high performing

PR and WTAT subjects. For low performers, from Model 4 we observe learning only in the

WTAT treatment.

To summarise: high performing subjects performed equally well across the different

treatments, while low performing subjects performed particularly well in the PRWL

treatment compared to the PR and WTAT treatments. It appears that our results are mainly

borne out by low performers.

As a robustness check, we can instead separate players by their ex-post record of

winning. Since competition does not occur in the PR treatment, we will focus only on the

PRWL and WTAT treatments. We define subjects to be ‘winners’ if they have won 8 or more

of the 15 post-intervention rounds which feature an element of rank or payoff competition,

while ‘losers’ have won 7 or fewer rounds. Table 7 present similar regressions to the ones

before, where we regress forecast errors for winners and losers separately. In Models 1 and 3,

we regress forecast errors against the WTAT treatment dummy (with PRWL as the reference

category), the gender and trait anxiety of players as controls, as well as a time trend. In

Models 2 and 4, we replace the trend with time trends specific to the PRWL and WTAT

treatments.

25

Table 7: Pooled random effects regression for forecast errors broken up by winners and losers; columns (1) and 2 present results for winners while columns (3) (4) presents results for losers.

Dependent variable = Forecast errors = |Predicted value – Actual Value|

Independent variables

Winners LosersModel 1 Model 2 Model 3 Model 4

WTAT2.278

(2.202)[0.301]

2.145(3.362)[0.523]

9.843(4.560)[0.031]

16.17(6.262)[0.010]

Female4.217

(2.327)[0.070]

4.217(2.328)[0.070]

9.160(3.443)[0.008]

9.160(3.445)[0.008]

Trait anxiety0.074

(0.151)[0.624]

0.074(0.151)[0.624]

0.202(0.296)[0.494]

0.202(0.296)[0.494]

Round-0.267(0.091)[0.003]

----0.201(0.179)[0.261]

---

PRWL X Round ----0.273(0.118)[0.021]

---0.056

(0.304)[0.853]

WTAT X Round ----0.262(0.136)[0.054]

----0.430(0.195)[0.028]

Constant10.20

(6.503)[0.117]

10.27(6.527)[0.116]

5.985(13.66)[0.661]

2.635(13.74)[0.848]

R2 0.024 0.024 0.042 0.043Wald χ2 15.94 17.07 13.66 21.20p > χ2 0.003 0.004 0.009 0.000No. of

observations 990 990 1020 1020

No. of participants 66 66 68 68Standard errors in parentheses; p-values in square brackets.

26

From Models 1 and 2 of Table 7, we observe that winners in the PRWL and WTAT

treatments perform similarly to one another. Learning is observed for winners in both

treatments. Similar to what we have found with low performers, losers in the WTAT

treatment do not perform as well as losers in the PRWL treatment. For losers, we observe

learning for subjects in the WTAT treatment but not for the PRWL treatment.

5. Concluding Remarks

In this paper, we have distinguished the effects of rank competition and payoff

competition, each of which could motivate performance under tournaments. Rank

competition refers to the act of competing that is independent from monetary payoffs, while

payoff competition refers to competing for the tournament prizes. In a cognitively

challenging task, we find that rank competition plays a central role in tournaments, to the

extent that once its effect has been controlled for, tournaments actually perform worse than

rank-augmented piece rates (PRWL). We find that this effect is concentrated amongst low

performing participants, with high performers unaffected by the provision of relative

performance feedback.

Whether our findings generalize beyond the lab depends on the extent to which our

MCPL task mimics real-life work and which types of jobs are represented well by this task. It

appears to us that most jobs in real life require some cognitive effort and in that regard a task

like ours probably provides a better approximation of many work-places rather than the more

mechanical number adding type tasks used in many prior studies.

27

6. References

Aoyagi, Masaki. (2010). Information Feedback in a Dynamic Tournament. Games and Economic Behavior, 70(2), 242-260.

Azmat, Ghazala, & Iriberri, Nagore. (2010). The Importance of Relative Performance Feedback Information: Evidence from a Natural Experiment using High School Students. Journal of Public Economics, 94(7-8), 435-452.

Azmat, Ghazala, & Iriberri, Nagore. (2016). The Provision of Relative Performance Feedback: An Analysis of Performance and Satisfaction. Journal of Economics & Management Strategy, 25(1), 77-110.

Balzer, William K., Doherty, Michael E., & O'Connor, Raymond. (1989). Effects of Cognitive Feedback on Performance. Psychological Bulletin, 106(3), 410-433.

Bellemare, Charles, Lepage, Patrick, & Shearer, Bruce. (2010). Peer Pressure, Incentives, and Gender: An Experimental Analysis of Motivation in the Workplace. Labour Economics, 17(1), 276-283.

Blanes i Vidal, Jordi, & Nossol, Mareike. (2011). Tournaments Without Prizes: Evidence from Personnel Records. Management Science, 57(10), 1721-1736.

Brown, Paul M. (1995). Learning from Experience, Reference Points, and Decision Costs. Journal of Economic Behavior & Organization, 27(3), 381-399.

Brown, Paul M. (1998). Experimental Evidence on the Importance of Competing for Profits on Forecasting Accuracy. Journal of Economic Behavior & Organization, 33(2), 259-269.

Bull, Clive, Schotter, Andrew, & Weigelt, Keith. (1987). Tournaments and Piece Rates: An Experimental Study. Journal of Political Economy, 95(1), 1-33.

Cadsby, C. Bram, Engle-Warnick, Jim, Fang, Tony, & Song, Fei. (2010). Psychological Incentives, Financial Incentives, and Risk Attitudes in Tournaments: An Artefactual Field Experiment. Working Paper

Ehrenberg, Ronald G., & Bognanno, Michael L. (1990). Do Tournaments have Incentive Effects? Journal of Political Economy, 98(6), 1307-1324.

Ellingsen, Tore, & Johannesson, Magnus. (2007). Paying Respect. Journal of Economic Perspectives, 21(4), 135-149.

Eriksson, Tor. (1999). Executive Compensation and Tournament Theory: Empirical Tests on Danish Data. Journal of Labor Economics, 17(2), 262-280.

Eriksson, Tor, Poulsen, Anders, & Villeval, Marie-Claire. (2009). Feedback and Incentives: Experimental Evidence. Labour Economics, 16(6), 679-688.

Falk, Armin, & Ichino, Andrea. (2006). Clean Evidence on Peer Effects. Journal of Labor Economics, 24(1), 39-57.

28

Green, Jerry R., & Stokey, Nancy L. (1983). A Comparison of Tournaments and Contracts. Journal of Political Economy, 91(3), 349-364.

Hannan, R. Lynn, Krishnan, Ranjani, & Newman, Andrew H. (2008). The Effects of Disseminating Relative Performance Feedback in Tournament and Individual Performance Compensation Plans. The Accounting Review, 83(4), 893-913

Kluger, Avraham N., & DeNisi, Angelo. (1996). The Effects of Feedback Interventions on Performance: A Historical Review, a Meta-Analysis, and a Preliminary Feedback Intervention Theory. Psychological Bulletin, 119(2), 254-284.

Kräkel, Matthias. (2008). Emotions in Tournaments. Journal of Economic Behavior & Organization, 67(1), 204-214.

Kuhnen, Camelia M., & Tymula, Agnieszka. (2012). Feedback, Self-Esteem, and Performance in Organizations. Management Science, 58(1), 94-113.

Lazear, Edward P. (2000). Performance Pay and Productivity. American Economic Review, 90(5), 1346-1361.

Lazear, Edward P., & Rosen, Sherwin. (1981). Rank-Order Tournaments as Optimum Labor Contracts. Journal of Political Economy, 89(5), 841-864.

Mas, Alexandre, & Moretti, Enrico. (2009). Peers at Work. American Economic Review, 99(1), 112-145.

McNaughton, Neil. (2011). Trait Anxiety, Trait Fear and Emotionality: The Perspective from Non-Human Studies. Personality and Individual Differences, 50, 898-906.

Nalebuff, Barry J., & Stiglitz, Joseph E. (1983). Prizes and Incentives: Towards a General Theory of Compensation and Competition. Bell Journal of Economics, 14(1), 21-43.

Niederle, Muriel, & Vesterlund, Lise. (2007). Do Women Shy Away From Competition? Do Men Compete Too Much? Quarterly Journal of Economics, 122(3), 1067-1101.

Prendergast, Canice. (1999). The Provision of Incentives in Firms. Journal of Economic Literature, 37(1), 7-63.

Segal, Jan D., & Weinberg, Robert S. (1984). Sex, Sex Role Orientation and Competitive Trait Anxiety. Journal of Sport Behavior, 7(4), 153-159.

Spielberger, Charles D., Gorsuch, Richard L., Lushene, Robert E., Vagg, Peter R., & Jacobs, Gerard A. (1983). Manual for the State-Trait Anxiety Inventory. Palo Alto, CA: Consulting Psychologists Press.

Taylor, Beck A., & Trogdon, Justin G. (2002). Losing to Win: Tournament Incentives in the National Basketball Association. Journal of Labor Economics, 20(1), 23-41.

Tran, Anh, & Zeckhauser, Richard. (2012). Rank as an Inherent Incentive: Evidence from a Field Experiment. Journal of Public Economics, 96(9-10), 645-650.

Vandegrift, Donald, & Brown, Paul M. (2003). Task Difficulty, Incentive Effects, and the Selection of High-Variance Strategies: An Experimental Examination of Tournament Behavior. Labour Economics, 10(4), 481-497.

29

Vandegrift, Donald, Yavas, Abdullah, & Brown, Paul M. (2007). Incentive Effects and Overcrowding in Tournaments: An Experimental Analysis. Experimental Economics, 10(4), 345-368.

30

Appendix: Instructions

The University of XXXXInstructions for the Experiment

WELCOME.

PLEASE TURN YOUR CELL PHONES OFF NOW.

This is a study examining the manner in which people make decisions. The University of XXXX has provided the funds to conduct this research. If you follow the instructions and make good decisions you might earn a considerable amount of money.

At the beginning of the session each person will be given an Earnings Account with $5.00 in it. You will participate in a decision making task for each of 20 rounds. You will have the chance to earn money each round, with your earnings for each round being added to your Earnings Account. At the end of the experiment, the balance of your Earnings Account will be paid to you in cash.

___________________________________________________________________________DESCRIPTION OF THE TASK:

In each round you will be asked to predict the future value of a fictitious ‘stock’. The value of this stock is unknown to all participants, but you will be able to observe two CUES that can help you form your forecast. These cues can be used to predict the stock’s value much the same way that the amount of rainfall and the average temperature can be used to predict the quality of a corn crop, the number of unoccupied apartments and student enrolment this year can be used to predict next year’s rent increases, or the demand for sports cars can be used to predict their future price.___________________________________________________________________________In each round you will be shown the values for the two CUES.

NOTE: One of the CUE values will always be fixed at 150 for each of the 20 rounds.The other cue value will change each round. But the relation of the cue values to the stock’s price will remain the same.

Example:For example let the value of Cue A be fixed at 150. Suppose the values for the cues in a round were given as:

CUE A = 150CUEB = 100

You will be asked to predict the price of the stock given these two cue values. The next round one of the cues will take on a different value, such as:

CUE A = 150CUEB = 450

31

You will then predict that round’s price using these new cue values. Remember that even though the values of the cues change, the underlying relation between the cue values and the stock’s price remains the same. Thus, in order to make accurate forecasts you will need to determine the relation between the cues and the price of the stock.

___________________________________________________________________________

YOUR FORECASTING ERROR

After making your forecast, the computer will calculate the distance between your forecast and that round’s actual price (your absolute forecasting error). This amount will be your forecasting error.

Example:Suppose your forecast was 230. If the actual price of the stock was 200 then your forecasting error would be 30:

Your forecasting error = 230 – 200 = 30

Suppose your forecast was 148. If the actual price of the stock was 200 then your forecasting error would be 52:

Your forecasting error = 200 – 148 = 52___________________________________________________________________________

YOUR EARNINGS IN EACH ROUND:

In each round, your earnings will depend on your forecasting error. Your earnings in each round will equal $1 less your forecasting error for that round.

That is, your earnings (E) in each round will be given by E = $1.00 – (forecast error).

Example:Suppose your forecast error in a particular round is 30. Then you will earn $0.70 in that round. This is because:

$1.00 – $0.30 = $0.70

Suppose in another round your forecast error is 8. Then you will earn $0.92 in that round. This is because:

$1.00 - $0.08 = $0.92

Note that if your error is 100 or over, then you will earn nothing in that round. The minimum amount you can earn in a round is $0.00.

Suppose in another round your forecast error is 102. Then you will earn $0.00 because:

$1.00 - $1.00 = $0.00

32

SPECIFIC INSTRUCTIONS:

1. Before Round 1:You will be shown 10 examples of cues and stock prices. You will have 5 minutes in which to examine these examples.

2. Round 1:At the end of the 5-minute example round you will be shown the first two cue values and asked to forecast the price of the stock in Round 1. You will have 90 seconds to make your forecast.

3. End of Round 1:At the end of the 90 seconds all participants will have entered their forecasts. After all earnings have been calculated you will be shown your results for Round 1. The computer will then show you your earnings for the round, including:

Cue A Cue B Your Forecast

Actual Price

Forecasting Error

Earning this

Round

Total Earnings

Please record this information on the RECORD SHEET provided to you.

4. Beginning of Round 2:After examining and recording the earnings from round 1, you will be shown the values of CUE A and CUE B in Round 2. You will have 90 seconds to form your forecast.

5. Subsequent Rounds:

Each subsequent round proceeds in the same way and will be repeated for each of the 20 rounds. In each round, you will make a forecast based on two new cue values. At the end of round 20, you will receive a cash payment in the amount indicated by the earnings account.

However, after you have finished the first five rounds of play, we will have a pause. It is possible that there will be a change in the way in which you earn money for the subsequent rounds 6 through 20. If there is no change then we will tell you so and ask you to simply continue playing the game in the same manner as in the first five rounds. However, if there is a change in payment, then we will provide you with further instructions at that point and explain these changes and also answer any questions you may have.

Do not use calculators. Your forecasts and your earnings are your private information. It is important that you do not talk or in any way try to communicate with other people during the experiment. If you violate the rules, you will be asked to leave the experiment.

GOOD LUCK!!

33

PRWL Specific Instructions

Rounds 6 to 20

Rounds 6 to 20 are played exactly as rounds 1 to 5 but with the following exceptions:

Each period you will be paired with another participant in the session today. Your partner will change each round, so you will never be paired with the same partner for more than one consecutive rounds;

After you have made your forecast, the computer will compare your forecasting error to your partner’s forecasting error in that round;

Your results will show whether your forecasting error was greater or less than your partner’s for that round;

If your error is less than your partner’s, then you will be told you WIN that round. If your error is more than your partner’s, you will be told you LOST that round. If your error is equal to your partner’s, then the computer will randomly decide the winner and loser.

Your payment will remain unchanged. That is, each round you will continue to be paid:

Earnings = $1.00 – Forecasting Error

You will also be shown your partner’s forecast and forecasting error at the end of the round. That is, at the end of each round you will observe:

Cue A

Cue B

Your Forecast

Actual Price

Forecasting Error

Earnings this round

WIN or LOSE

Example: Suppose the actual price was 210, your forecast was 168, and your partner’s forecast was 163. Your forecasting error would be 42 and your partner’s forecasting error would be 47. You would see the following results for that round:

Cue A

Cue B

Your Forecast

Actual Price

Forecasting Error

Earnings this round

WIN or LOSE

168 210 42 $0.58 WIN

34

Example: Suppose the actual price was 210, your forecast was 168, and your partner’s forecast was 173. Your forecasting error would be 42 and your partner’s forecasting error would be 37. You would see the following results for that round:

Cue A

Cue B

Your Forecast

Actual Price

Forecasting Error

Earnings this round

WIN or LOSE

168 210 42 $0.58 LOSE

Do you have any questions?

35

WTAT Specific Instructions

Rounds 6 to 20

Rounds 6 to 20 are played exactly as rounds 1 to 5 but with the following exceptions: You will have an additional $4.00 added to your Earnings Account;

Each period you will be paired with another participant in the session today. Your partner will change each round, so you will never be paired with the same partner for more than one consecutive rounds;

After you have made your forecast, the computer will compare your forecasting error to your partner’s forecasting error in that round;

Your results will show whether your forecasting error was greater or less than your partner’s for that round;

If your error is less than your partner’s, then you will be told you WIN that round. If your error is more than your partner’s, you will be told you LOST that round. IF your error is equal to your partner’s, then the computer will randomly decide the winner and loser;

Your payment will depend upon whether your forecasting error is greater or less than your partners. That is, each round you will earn either $1.00 or $0.00. You will be paid either:

Earnings = $1.00 if you WIN

Or

Earnings = $0.00 if you LOSE

Example: Suppose your forecasting error was 42 and your partner’s forecasting error was 47. You would see the following results for that round:

Cue A

Cue B

Your Forecast

Actual Price

Forecasting Error

Earnings this round

WIN or LOSE

42 $1.00 Win

36

Example: Suppose your forecasting error was 42 and your partner’s forecasting error was 37. You would see the following results for that round:

Cue A

Cue B

Your Forecast

Actual Price

Forecasting Error

Earnings this round

WIN or LOSE

42 $0.00 LOSE

Do you have any questions?

37

Pre-Task Questionnaire

Player ID ____________

PLEASE ANSWER ALL OF THE FOLLOWING QUESTIONSA number of statements which people have used to describe themselves are given below. Read each statement and, using the scale below, tick the appropriate number indicating how you generally feel. There are no right or wrong answers. Do not spend too much time on any one statement but give the answer which seems to describe how you generally feel.

1 2 3 4 Almost Sometimes Often Almost never always

Alm

ost N

ever So

met

imes

Ofte

n

Alm

ost a

lway

s

1 2 3 41. I feel pleasant2. I tire quickly3. I feel like crying4. I wish I could be as happy as others seem to be5. I am losing out on things because I can’t make up my mind soon enough6. I feel rested 7. I am “calm, cool and collected”8. I feel that difficulties are piling up so that I cannot overcome them 9. I worry too much over something that doesn’t really matter 10. I am happy11. I am inclined to take things hard 12. I lack self-confidence 13. I feel secure14. I try to avoid facing a crisis or difficulty15. I feel blue16. I am content17. Some unimportant thoughts run through my mind and bother me18. I take disappointments so keenly that I can’t put them out of my mind 19. I am a steady person20. I get in a state of tension or turmoil as I think over my recent concerns and interests

38

Post-Task Questionnaire

Player ID ____________

PLEASE ANSWER ALL OF THE FOLLOWING QUESTIONS

A) For each of the following statements, please indicate how true the statement is for you using the following scale:

1 2 3 4 5 6 7 Not at all Somewhat Very

true true true

Not

at

al

l tru

e

Som

ewha

t tru

e

Ver

y tru

e

1 2 3 4 5 6 71. I enjoyed this activity very much2. I think I am pretty good at this activity3. I put a lot of effort into this activity4. I did not feel nervous at all which doing this activity5. This activity was fun to do6. I think I did pretty well at this activity, compared to other participants7. I did not try very hard to do well at this activity8. I felt very tense while doing this activity9. I thought this activity was boring10. After working at this activity for a while, I felt pretty competent11. I tried very hard on this activity12. I was very relaxed doing this activity13. This activity did not hold my attention14. I am satisfied with my performance at this task15. It was important to me to do well at this task16. I was anxious while working on this task17. I would describe this activity as very interesting18. I was pretty skilled at this activity19. I did not put much energy into this20. I felt pressured while doing this activity.21. I thought this activity was quite enjoyable.22. This was an activity that I could not do very well.23. While I was doing this activity, I was thinking about how much I enjoyed it.

39

B) The following items ask about how you felt about the other participants during the session.

Not

at

al

l tru

e

Som

ewha

t tru

e

Ver

y tru

e

1 2 3 4 5 6 71. I felt really distant to them2. I really doubt they and I would ever be friends3. I felt I could really trust them4. I’d really like the chance to interact with them more often5. I’d really prefer not to interact with them in the future6. I don’t feel like I could really trust them7. It is likely that they and I could become friends if we interacted a lot8. I felt close to them

C) How many of the people in this session did you know before the experiment? ____________

D) Basic information about you:

Your Gender (Male/ Female) ________

Age

Major: ______________________________________________

Year in School (e.g., Stage 2) _______________________________

Ethnicity (Please circle one): Maori Pacific Island NZ European

Asian Other _______________

Country where you were born? __________________

If you were born outside of New Zealand, at what age did you move here? _________

40