Adding noise to the institution: an experimental welfare investigation of the contributionbased grouping mechanism
Abstract
Realworld institutions dealing with social dilemma situations are based on mechanisms that are rarely implemented without flaw. Usually realworld mechanisms are noisy and imprecise, that is, which we call ‘fuzzy’. We therefore conducted a novel type of voluntary contributions experiment where we test a mechanism by varying its fuzziness. We focus on a range of fuzzy mechanisms we call ‘meritocratic matching’. These mechanisms generalize the mechanism of ‘contributionbased competitive grouping’, and their basic function is to group players based on their contribution choices—i.e. high contributors with high contributors, and low contributors with low contributors. Theory predicts the following efficiencyequality tradeoff as a function of the mechanism’s inherent fuzziness: high levels of fuzziness should lead to maximal inefficiency, but perfect equality; decreasing fuzziness is predicted to improve efficiency, but at the cost of growing inequality. The main finding of our experimental investigation is that, contrary to tradeoff predictions, less fuzziness increases both efficiency and equality. In fact, these unambiguous welfare gains are partially realized already at levels where the mechanism is too fuzzy for any highefficiency outcome to even be a Nash equilibrium.
1 Introduction
We conducted an experiment to investigate the welfare consequences of implementing a ‘fuzzy’—rather than a fully precise—mechanism in the context of voluntary contributions games. Our aim is to take a first step in the direction of improving our understanding of the consequences of implementing institutions based on inherently imprecise mechanisms.^{1} This inquiry is relevant before any potential policy recommendations can be made, because realworld implementations would rarely be without flaw, particularly in the context of private, voluntary contributions.
Most investigations of mechanisms in the context of social dilemmas presume that there is perfect observability. Inherently, however, introducing a mechanism in the real world will produce substantial fuzziness. This is due to the fact that players’ individual actions are often not perfectly observable, to the players and to the authority. ‘Imperfect public monitoring’ is ubiquitous in the real world (Abreu et al. 1990; Fudenberg et al. 1994), and there exists a rich body of theoretical investigations in various social dilemma contexts closely related to ours such as noisy prisoners’ dilemmas (Wu and Axelrod 1995) and team production without individual feedback (Alchian and Demsetz 1972).^{2} Previous experimental studies of social dilemma games with imperfect monitoring have revealed that noise may play a crucial, nontrivial role in determining the performance of a mechanism and in sustaining cooperation generally.^{3}
In this paper, we investigate a fuzzy mechanism for voluntary contributions games. Voluntary contributions games, as introduced by Marwell and Ames (1979, 1980), provide parsimonious models to capture the strategic interaction underlying public goods provisioning.^{4} In the baseline implementation of these games, individual players make private, costly contributions that create a public good which is then shared equally amongst all players. In the absence of suitable mechanisms, there are insufficient private incentives for contributing behaviors, and universal noncontribution is the unique Nash equilibrium. This is reflected in many economic experiments on voluntary contributions games by a decay of contributions over time (Ledyard 1995; Chaudhuri 2011).
Outcomes with high contributions can only be expected when a suitable mechanism is implemented. Several mechanisms are known.^{5} Successful mechanisms are able to change the incentive structure of the game in such a way that high contributions are stabilized. Numerous lab experiments have shown how voluntary contributions are stabilized through their introduction (Ledyard 1995; Chaudhuri 2011). An important difference between the typical lab setting and the real world is that a realworld implementation of the mechanism would rarely be without flaw. Instead, the realworld mechanism would be fuzzy, that is, subject to various sources of monitoring imperfections, mechanism imprecision and other environmental noise due to, for example, measurement error or enforcement issues.
In this paper, we aim to advance in the direction of understanding fuzzy mechanism implementation. We therefore conduct a novel type of voluntary contributions experiment where we test a mechanism by varying its degree of fuzziness. Our baseline is the mechanism of ‘contributionbased competitive grouping’, as was introduced in a recent, seminal paper by Gunnthorsdottir (2010). Under this mechanism, players are grouped based on their individual voluntary contributions.^{6} As a result, noncontribution is no longer a dominant strategy. Instead, players have incentives to contribute positive amounts if others do likewise in order to be matched with them. Contributionbased grouping changes the game’s entire incentive structure without requiring payoff transfers. Universal noncontribution, the least efficient outcome, continues to be a Nash equilibrium. However, new, more efficient equilibria may also emerge that feature high contribution levels. These new equilibria are characterized by an asymmetric strategy profile such that a vast majority of players contributes fully and a small minority of player freerides.^{7} Several recent lab experiments confirm coordination on the asymmetric highefficiency equilibria with high aggregate precision (Gunnthorsdottir 2010; Gunnthorsdottir and Thorsteinsson 2011; Gunnthorsdottir et al. 2010; Rud and Rabanal 2015).
The focus of this paper is on (i) the performance of ‘fuzzy’ variants of Gunnthorsdottir (2010)’s mechanism, and (ii) on their welfare consequences in terms of efficiency and equality. Fuzzy variants of Gunnthorsdottir (2010)’s mechanism were recently formulated by Nax et al. (2014) introduction of an additional parameter (interpretable as variance) that measures the degree of imprecision inherent to the competitive grouping’s basic functioning.^{8} This fuzzy generalization of contributionbased group matching is referred to as ‘meritocratic matching’. Compared to the basic mechanism, meritocratic matching works as follows: instead of grouping players based on actual contributions, actual contributions are ‘noised’ by variance \(\sigma ^2\) (measuring fuzziness/imprecision), and players are grouped based on these noised contributions. Meritocratic matching thus bridges the nomechanism case of random matching and contributionbased competitive grouping continuously: when \(\sigma ^2=0\) the mechanism corresponds to contributionbased competitive grouping; when \(\sigma ^2\rightarrow \infty \) the mechanism approaches random rematching as in a standard implementation (Andreoni 1988). Nax et al. (2014) show that highefficiency equilibria exist provided the mechanism is precise enough implying a bound on \(\sigma ^2\).
In this study, we investigate the welfare properties of meritocratic matching for a wide range of \(\sigma ^2\) values. Theory predicts existence of highefficiency Nash equilibria for some of these values, but not for all, and not for \(\sigma ^2\rightarrow \infty \). Our results summarize as follows. We confirm that the asymmetric highefficiency equilibria are coordinated upon with high aggregate precision when they exist, validating theory predictions and previous lab studies. Contrary to theoretical predictions, however, we provide novel evidence that higher levels of meritocracy also increase ex post equality. Moreover, we find that these unambiguous welfare gains are even realized when the mechanism is too fuzzy for theory to even predict existence of highefficiency equilibria.
We view high fuzziness as the realistic ‘default’ implementation of our mechanism. However, a policy maker could somehow choose to reduce the mechanism’s inherent noise (e.g. by investing into monitoring or implementation technologies). Viewing the mechanism’s inherent noise, therefore, as some sort of policy variable and without explicitly modeling the involved costs of ‘setting the noise’, we can draw the following main policy conclusion from our results: provided that there is ex ante equity (as there was in our experiment by design), it would be unambiguously welfarebeneficial to introduce meritocratic matching, even if the mechanism remains very fuzzy.
The remainder of this document is structured as follows. Next, we provide details of the model and of the experimental design. In Sect. 3, we present the results of our welfare investigation. Finally, we conclude in Sect. 4.
2 The experiment
2.1 Modified voluntary contributions games
 1.
Contributions. Each agent i simultaneously decides to contribute any number of coins \(c_i\) between zero and his budget \(B>0\). The amount \(Bc_i\) not contributed goes straight to his/her private account. The ensemble of players’ decisions is represented by the vector of actual contributions c.
 2.
Fuzz. Fuzz in the form of i.i.d. Gaussian noise with mean zero and variance \(\sigma ^2\ge 0\) is added to each actual contribution \(c_i\) which results in the noised contribution vector \(c'\).
 3.
Grouping. k groups of a fixed size \(s<n\) (such that \(s*k=n\)) are formed according to the ranking of the noised contributions \(c'\) (with random tiebreaking).^{9} That is, the highest s contributors according to \(c'\) form group \(G_1\), the next highest s contributors form \(G_2\), etc. The resulting group partition is \(\rho =\{G_1,G_2,\ldots ,G_k\}\).
 4.Payoffs. Finally, based on grouping and the actual (not noised!) contributions vector c, payoffs \(\phi \) realize as follows. Each player i, matched into \(G_i\) with \(j\ne i\), receives the standard linear public goods payoff of:where m represents the marginal per capita rate of return, and \(G_{i}\) indicates the members of group \(G_i\) excluding i.$$\begin{aligned} \underbrace{\phi _i(c)}_{\text {payoff}}= \underbrace{(B(1m)*c_i)}_{\text {return from private account}} + \underbrace{\sum _{j\in G_{i}} m * c_j,}_{\text {return from group account}} \end{aligned}$$(1)
2.1.1 Parameter choices
In our design, we consider games where 4 groups of size 4 form (\(n=16\) and \(k=4\)) with a budget of \(B=20\) each period and a marginal per capita rate of return \(m=0.5\). The choice of parameters ensures good comparability of results with, on the one hand, the literature on voluntary contributions games under random rematching (as reviewed by Ledyard 1995; Chaudhuri 2011), and, on the other hand, with contributionbased competitive grouping mechanisms. Our treatments vary with respect to values and orders of values of \(\sigma ^2\), that is, which mechanisms are being played and in which order.
Before performing the main experiment, we tested online on Amazon Mechanical Turk (AMT) the following \(\sigma ^2\)values: 0, 2, 4, 5, 10, 20, 50, 100, 1000, and \(\infty \). This way, we could quickly explore the behavior of the participants in the vast parameter space. Previous research has suggested that the data quality of AMT experiments is adequate and reliable in various settings (Wang et al. 2015; Hauser and Schwarz 2016; Arechar et al. 2017). In fact, in our study, the behavior of participants in our online sessions did not significantly deviate from the known patterns found in the literature, or from the behavior of participants in the closest corresponding laboratory sessions from our own study except for the \(\sigma ^2=\infty \) condition.^{10}
Meritocracy regimes with corresponding variance intervals and equilibrium structure
Merit. regime  Variance  Equilibria  Experiment 

PERFECT  \(\sigma ^2=0\)  ‘high’ and ‘zero’  LAB \((\sigma ^2=0,n=48)\) 
HIGH  \(\sigma ^2=(0,20)\)  ‘high’ and ‘zero’  LAB \((\sigma ^2=3,n=48)\) 
LOW  \(\sigma ^2=[20,75)\)  ‘high’ and ‘zero’  LAB \((\sigma ^2=20,n=48)\) 
NO (ZERO)  \(\sigma ^2=\infty \)  Uniquely ‘zero’  LAB \((\sigma ^2=\infty ,n=48)\) 
INSUFFICIENT  \(\sigma ^2=[75,\infty )\)  Uniquely ‘zero’  AMT \((\sigma ^2=[100,1000],n=59)\) 
Throughout the paper we will focus on the results from the laboratory experiments. We will use the results from online sessions with \(\sigma ^2\)values of 100 and 1000 pooled together to complement the analysis to a region of the parameter space not covered by our laboratory sessions.
2.1.2 Nash equilibrium play: existence and intuition
As characterized in Gunnthorsdottir (2010) and Nax et al. (2014), there exist two types of Nash equilibria. One equilibrium is such that all players contribute zero. This equilibrium always exists, for all specifications of the game and for all variance levels. It coincides with the outcome that generates the lowest total payoffs. Another outcome that is a candidate for Nash equilibrium is characterized by full contribution by a vast majority of players and by freeriding of a few. This outcome produces close to the maximum of possible total payoffs, and it exists for certain parameter and noise specifications, in particular when the marginal per capita rate of return is high enough and when the noise is not too large.
To get some intuition for these Nash equilibria one has to evaluate the expected payoff, \(\mathbf {E}\left[ \phi _i(c)\right] \), that the players foresee during the decision stage, i.e. before groups are formed. In Eq. (1), the first term on the righthand side, i.e. the privateaccount return, is completely determined by the agent’s own contribution choice. There is no uncertainty. The second term on the righthand side—i.e. the groupaccount return—however, may depend on the player’s own and on others’ contributions in a probabilistic way.
Special case of NOMERIT. Only in the case of zero meritocracy (i.e. random rematching with \(\sigma ^2=\infty \)), grouping is completely independent of the contribution decisions. Hence, \(\mathbf {E}\left[ \phi _i(c)\right] \) behaves as it does in the standard voluntary contributions game: payoff is strictly decreasing in each player’s own contribution. Hence, noncontribution is a strictly dominant strategy, and the only equilibrium is universal noncontribution (indicated by ‘zero’ in Table 1).
General case. For meritocratic matching with any finite \(\sigma ^2\ge 0\), the player’s contribution decision affects the probability of being matched into different groups. Deciding to make a positive/higher contribution comes with a tradeoff between the sure loss on the own contribution (private account) and the promise of a higher return from being matched with others’ contributions (group account). Noncontribution is no longer a strictly dominant strategy, provided the promise of a higher group return is likely enough, that is, if the variance is not too large. Crucially, the probability to be matched accurately according to one’s actual contribution is decreasing as \(\sigma ^2\) increases, and noised contributions become less and less accurate representations of actual contributions.
Nax et al. (2014) show that—and this generalizes the results by Gunnthorsdottir (2010)—if the level of meritocracy is sufficiently large in addition to a bound on m, there exist highefficiency purestrategy Nash equilibria where a large majority of players contributes the full budget B and a small minority of players contributes nothing (This equilibrium is indicated by ‘high’ in Table 1). The intuition behind that kind of equilibrium is that those who contribute do so in the hope of being matched with others doing likewise (which happens with a high probability), while the freeriders expect to be able to freeride on the unlucky contributors who end up being matched with them.
The outcome where all players contribute zero, indicated by ‘zero’ in Table 1, continues to be an equilibrium for any level of \(\sigma ^2\) too.^{12} This is because no players can individually benefit from the grouping mechanism, no matter how precise it is, if he/she is the only player contributing a positive amount.
Equilibria for our parameters. For our parameter choices (\(n=16\), \(k=4\), \(B=20\) and \(m=0.5\)), Nash equilibrium existence is summarized in Table 1: for treatments with \(\sigma ^2\in \{100, 1000, \infty \}\), the only equilibrium is ‘zero’; for treatments with \(\sigma ^2\in \{0, 2, 3, 4, 5, 10, 20, 50\}\), one equilibrium is ‘zero’ and in addition there are \({n\atopwithdelims ()2}\) asymmetric purestrategy ‘high’ equilibria where exactly 2 players freeride and 14 others contribute fully.^{13}
2.2 Experimental details
Altogether we ran 28 experimental sessions with a total of 434 participants using the new experimental software NodeGame (Balietti 2016). We shall now briefly summarize our experimental design. More details about the experiment can be found in Appendix A.2.
12 Experimental sessions with a total of 192 participants were run at the ETH Zürich Decision Science Laboratory (DeSciL). Each lab session lasted roughly one hour. There were 16 participants in each session. DeSciL recruited the subjects from the joint subject pool of the University of Zurich and ETH Zurich maintained by the University Registration Center for Study Participants (UAST). The experiment followed all standard behavioral economics procedures and meets all Ethics Committee guidelines. Decisions, earnings and payments were anonymous. Payments were administered by the DeSciL’s lab staff. In addition to a 10 CHF showup fee, each subject was paid according to a known exchange rate of 0.01 CHF per coin. Overall, monetary rewards ranged from 30 to 50 CHF, with a mean of 39 CHF. Each lab session consisted of two games, each of which was a 40round repetition of the same underlying stage game.^{14} The same fixed budget was given to each subject every period. Each game had separate instructions that were distributed at the beginning of each game. These instructions contained full information about the structure of the game and about the payoff consequences to themselves and to the other agents. After reading the instructions, all participants were quizzed to make sure they understood the task. The two games had different variance levels. There were four variance levels in our lab study, \(\sigma ^2=\{0,3,20,\infty \}\), and each game had equivalent instructions. We played every possible pair of variance levels in both orders to have an orthogonal balanced design. As the game went on, players learned about the other players’ previous actions and about the groups that formed.
The remaining 16 experimental sessions were run on Amazon’s Mechanical Turk (AMT) with a total of 242 participants. In each AMT session, all participants played only one game with one of the following variance levels from \(\sigma ^2=\{0,2,4,5,10,20,50,100,1000,\infty \}\). In order to mitigate dropout problems, AMT sessions were shorter with only 20 or 25 rounds. AMT sessions lasted 20 min on average, and subjects earned between 1.7 and 3.4 USD.
3 Results
Figure 2 shows the average contribution levels over rounds. Conditions LOW, HIGH and PERFECTMERIT sustain high levels of contributions, close to the highefficiency equilibrium. For the case of NOMERIT, the steadily declining contributions reflect the usual pattern (Ledyard 1995; Chaudhuri 2011). For INSUFFICIENTMERIT, unfortunately with slightly shorter series from AMT, we observe no such decline, but intermediate contribution levels throughout instead.
Our main analysis focuses on the laboratory data, that is, on conditions PERFECT, HIGH, LOW and NOMERIT with \(\sigma ^2=\{0,3,20,\infty \}\). First, we study the efficiency, inequality and fairness properties of the data by analysis of the first game played in each session. Subsequently, we analyze data from the second game to assess the participants’ sensitivity to changes in meritocracy levels.
3.1 Efficiency
Overall, we observe significant differences in the mean of realized payoffs among the four lab treatments (linear mixed model LMM: \(F_{3,8} = 36.95, P < 0.0001\)). Taking NOMERIT as a baseline, LOWMERIT led to an increase in the average realized payoff of 7.1611 (Likelihood Ratio Test LRT: \(\chi _{(1)} = 12.7, P = 0.0004\)), HIGHMERIT to an increase of 8.1964 (LRT: \(\chi _{(1)} = 17.48, P < 0.0001\)), and PERFECTMERIT to an increase of 8.8287 (LRT: \(\chi _{(1)} = 16.22, P < 0.0001\)). These levels correspond to roughly double those of NOMERIT. Computing the most conservative (Bonferroni) adjusted pvalues on all pairwise differences reveals that the treatment with variance \(\infty \) is significantly different (\(P < 0.0001\)) from the other three variance levels \(\sigma ^2=\{0,3,20\}\), which are themselves not significantly different from each other.
For intermediate meritocracy regimes \(\sigma ^2=\{3,20\}\), efficiency is significantly below the level implied by the payoffdominant equilibria,^{15} but the difference is small (two and eight percent respectively). Conversely, under full meritocracy \(\sigma ^2=0\), efficiency is above and within two percent of equilibrium.
3.2 Equity by design versus ex ante/ex post inequality
Note that our experimental design features ex ante ‘equity’ in the sense that the games we study are symmetric in every respect including budgets. Note also that the noncontribution ‘zero’ equilibrium (which exists for all \(\sigma ^2\)) is also characterized by perfect equality in outcomes, independent of whether evaluated ex ante (at the contribution decision stage) or ex post (after payoffs realize).
By contrast, the nearefficient ‘high’ Nash equilibria (which exist for \(\sigma ^2<75\)) are asymmetric and predict that 14 out of 16 players contribute fully and 2 players freeride. This asymmetry implies both ex ante (to a lesser extent) and ex post (to a larger extent) inequality.^{16} Ex post inequality in the highefficiency equilibria, in particular, is quite serious as the two freeriders who get matched with at least two fullcontributors are amongst the bestoff players, while some lucky fullcontributors are betteroff (those not matched with freeriders) than other unlucky fullcontributors (those matched with freeriders) who are substantially worseoff.
In this section, we shall show that laboratory evidence yields diametrically opposite results compared with what theory predicts regarding ex post equality; namely, contrary to theoretical predictions, higher meritocracy levels lead to outcomes that are more equal in terms of payoff distributions than lower meritocracy regimes. This is because players play less heterogeneously and more in line with equilibrium in high meritocracy regimes than in lower ones.
One can identify two measures of payoff inequality directly from the moments of the payoff distribution: (i) the payoff of the worstoff (Rawls 1971), \(\underline{\phi }=\min \{\phi _i\}\), and (ii) the variance of payoffs, \(\sigma ^2=\frac{\sum _{i\in N}(\phi _i\overline{\phi })^2}{n}\). A more sophisticated third alternative is (iii) the Gini coefficient. In terms of all measures, our analysis shows that equality increases with meritocracy. Note that the following results are also robust to other measures of inequality (Cowell 2011) (see “Appendix”).
Figure 4 summarizes our analysis. It highlights that, as with efficiency—but this time contrary to theoretical predictions—equality also increases from \(\sigma ^2=\infty \) (NOMERIT) through \(\sigma ^2=\{20,3\}\) to \(\sigma ^2=0\) (PERFECTMERIT). A significant difference in the variance of realized payoffs in each round among the four treatments is found (LMM: \(F_{3,8} = 7.27, P < 0.0113\)). When computing Bonferroni adjusted pvalues, the treatment with variance \(\infty \) was found significantly different (\(P = 0.0003 ; P = 0.0004 ; P = 0.0086\)) from the other three variance levels (\(\sigma ^2=\{0,3,20\}\)), which are themselves not significantly different from each other. Taking NOMERIT as a baseline, LOWMERIT led to a decrease in the variance of realized payoffs in each round of −13.546 (LRT \(\chi _{(1)} = 8.13, P = 0.0043\)), HIGHMERIT to a decrease of −16.914 (LRT \(\chi _{(1)} = 9.89, P = 0.0016\)), and PERFECTMERIT to a decrease of −17.122 (LRT \(\chi _{(1)} = 6.78, P = 0.0091\)).
3.3 Interpretations
We have found that Nash predictions fare well in approximating efficiency levels in the fuzzy regimes LOW and HIGHMERIT, and in the perfect implementation PERFECTMERIT. Nash equilibrium was neither implemented nor converged to in INSUFFICIENTMERIT. In NOMERIT, there was aggregate convergence toward equilibrium (decay of contributions). This section is dedicated to different behavioral explanations of these phenomena under the variously fuzzy mechanisms.
3.3.1 Fairness in meritocratic matching
In our analysis of Nash equilibria, we considered the theoretic case where all subjects were riskneutral and pursued an entirely selfish, linear payoff function. From a wide variety of experiments, however, we know that players are riskaverse and pursue distributional and otherregarding preferences that take into account how their decisions affect not only their own material payoff but also the material payoffs of others. In particular, human preferences have been shown to include ‘fairness’ considerations. Amongst the bestknown models of preferences for fairness are the models of Fehr and Schmidt (1999), Bolton and Ockenfels (2000) and Charness and Rabin (2002).
Fairness models have been used to explain why subjects in voluntary contributions games, in the absence of a mechanism such as our contributionbased meritocracy, may initially contribute differing positive amounts. As players contribute different amounts under random rematching, in particular, those contributing more earn less (which is disadvantageously unfair), and those contributing less earn more (which is advantageously unfair). Numerous experiments have shown that experiences of unfair outcomes lead to contribution adjustments, and that experiences of disadvantageously unfair outcomes, especially, lead to contribution reductions. Therefore, these kinds of fairness considerations, where contribution reductions due to disadvantageously unfair experiences outweigh the corresponding contribution increases due to advantageously unfair experiences, lead to a spiraling down of conditional cooperation (Fischbacher et al. 2001). This is a phenomenon we also observe in our NOMERIT baseline treatment.
Fairness motives are likely to be different when a mechanism is in place. Consequently, which fairness notions are relevant in evaluating outcomes may depend on the underlying principle of the mechanism. The basic principle of the mechanisms considered in this paper is to group contributors with contributors, and to group freeriders with freeriders. As this mechanism becomes more fuzzy (\(\sigma ^2\) increases), this principle is more frequently violated, and freeriders may even be matched into better groups than contributors. Subjects may find this unfair, in light of what the mechanism is designed to achieve in principle, and react to this by adjustments of their behavior. To explore these reactions we propose a notion of fairness in light of our mechanism, which we term ‘meritocratic fairness’. We contrast this notion with the aforementioned ‘payoff fairness’ concerns inherent to models such as Fehr and Schmidt (1999), Bolton and Ockenfels (2000) and Charness and Rabin (2002).
3.3.2 Contribution decisions: meritocratic fairness and strategic concerns
Contributions in our model play a doubly strategic role. On the one hand, they determine a player’s payoff within a given group. On the other hand, they also determine the group into which the player is matched. As regards individual contribution decisions, we conjecture that fairness considerations matter, and that the relevant fairness considerations are adapted to the functioning of our mechanism and inherent noise with which the mechanism is announced. Hence, we test the following two higherorder hypotheses. \((H_1)\) Players will adjust their contributions after experiences of unfair outcomes. \((H_2)\) What is considered unfair will depend on the mechanism that is in place.

Under PERFECTMERIT, starting at the nearefficient Nash equilibrium prediction, we do not expect significant departures from a strategic bestresponse state as there is no inherent meritocratic unfairness (by definition), and we expect standard fairness considerations to be less important.

For the intermediate meritocracy levels (HIGH, LOW, INSUFFICIENTMERIT), we expect contribution decreases in response to meritocratic unfairness experiences. This effect is expected to be weaker the higher the noise of the implementation. However, other than under NOMERIT, downward corrections of contributions would not necessarily need to trigger an overall downward decay of contributions, because of the different strategic incentives. We expect standard fairness considerations to be of limited importance, and of increasing importance as noise increases.

Under NOMERIT, we expect meritocratic fairness to play no role, as the mechanism has no such function. Instead, we expect standard fairness considerations to matter, which will lead to downward corrections and to an overall downward decay of contributions.
3.3.3 Meritocratic fairness: results
In PERFECTMERIT, there is zero meritocratic unfairness. Indeed, no bias in corrections that could reduce contributing behavior are predicted. Not even fullcontributors who are matched in the lowest group due to bad luck decrease their contribution in the next round. In fact, this happened only 3 times in 210 occurrences of such bad luck. This striking result is as predicted. In fact, experiences of meritocratic unfairness are not possible under PERFECTMERIT, and so participants do not perceive being placed in a lower group due to bad luck as unfair; it is part of the mechanism, and many subjects successfully manage to take turns in those unlucky positions.
In conditions HIGH, LOW, and NOMERIT, we studied how the level of meritocratic unfairness experienced in the previous round impacts the decision to contribute in the following round. To do so, we performed a multilevel regression of betweenrounds contribution adjustments with subject and session as random effects. Our findings reveal that disadvantageous unfairness leads to decreases in treatments LOWMERIT \(0.18^{***} (0.05)\), and NOMERIT \(0.25^{***} (0.03)\)). For HIGHMERIT the decrease is consistent in sign and size, but not statistically significant \(0.39 (0.21)\). However, if HIGHMERIT and LOWMERIT are pooled together the effect turns out to be significant \(0.25^{***} (0.03)\). Meritocratic disadvantageous fairness can, therefore, originate significant differences between the theoretical equilibrium predictions and experimentally observed behavior. Advantageous unfairness leads to increases under some but not under all regimes. The strengths of these effects varied and the evidence was not contrary to predictions.
We also performed additional regressions to compare meritocratic fairness to a standard notion of fairness, which we chose to be represented by Fehr and Schmidt (1999). As expected, applying the standard notion of distributional fairness yielded good results only for the case of NOMERIT (for which it was conceived). On the other hand, meritocratic unfairness proved a good predictor of the contribution adjustments between rounds across all other treatments. Standard fairness did not prove to be a good predictor of contributions in these treatments, and decreasingly so for higher levels of meritocracy. Therefore, meritocratic fairness can be seen as a natural extension of distributional fairness in games with positive levels of meritocracy. Additional details and full regression tables are available in the Appendix. We leave it as an avenue for future work to analyze alternative fairness measures too, and to develop a general theory of fairness (which fairness consideration matters when?).
3.4 Sensitivity
So far, we have shown that (i) both efficiency and equality increase with meritocracy, and that (ii) considerations of ‘meritocratic fairness’ may explain deviations from equilibrium predictions. In this section, we show that changes in the level of experienced meritocracy have significant implications as well. In particular, we test whether participants coming from a higher (lower) meritocracy level in part 1 are more (less) sensitive to meritocratic unfairness in part 2.
For this analysis, we used the data pertaining of part 2 of the experiment, controlling for which meritocracy level was played in part 1. We divided the dataset in two subsets, depending on whether participants in part 2 experienced a higher or lower meritocracy level than in part 1. In order to obtain a balanced design with respect to the direction of meritocracy changes, we further sampled the data from part 2 to include only the intermediate regimes of meritocracy (\(\sigma ^2=\{3,20\}\)). In this way, both conditions could be tested against perfect meritocracy, zero meritocracy, and one intermediate regime. We created a dummy variable for “contribution goes down” (0;1) and performed a multilevel logistic regression with subject and session as random effects. We used the level of disadvantageous meritocratic unfairness experienced in the previous round as a predictor of whether contribution is expected to go up or down in the next round.
Our main finding is that the distaste for meritocratic unfairness is exacerbated after having played a more meritocratic regime in part 1. That is, if a participant experienced meritocratic unfairness in the previous round, he/she is more likely to reduce the own contribution in the current round if the level of meritocracy in part 2 is lower than in part 1 (Logistic Mixed Regression LMR: \(Z = 2.521, P = 0.0117\)). The effect in the opposite direction—a lower meritocracy level in part 1 than in part 2—is not significant (LMR: \(Z = 1.522, P = 0.128\)).
The different sensitivity to meritocratic unfairness may explain the different levels of efficiency and equality overall. Sessions in part 2 with higher sensitivity to meritocratic unfairness—i.e. descending the meritocracy ladder—have significantly lower average payoff (Onesided KolmogorovSmirnoff KS: \(D^+ = 0.1531, P < 0.0001\)), and significantly higher inequality—measured by the average Gini coefficient per round (\(D^+ = 0.1583, P = 0.0494\)). These results confirm that, in our settings, increases in efficiency are followed by inequality reduction, and that meritocratic fairness considerations may help explain this dissolution of the predicted efficiencyequality tradeoff.
4 Discussion
Our aim was to investigate the welfare consequences of institutions implementing a fuzzy mechanism instead of a perfect one. We initiated this line of research by considering a novel mechanism where contributors have a tendency to be matched with contributors, and freeriders with freeriders. Under such a mechanism, it was predicted that zero meritocracy leads to maximal equality and minimal efficiency, while perfect meritocracy to the opposite. Regarding fuzzy implementations of meritocracy, theory predicted ‘leaky buckets’ in both directions: reducing meritocracy increases equality at the expense of efficiency, and increasing meritocracy increases efficiency at the expense of equality. These predictions reflect a tradeoff that is at the core of economic theory (Arrow 1951; Sen 1970; Okun 1975; Gauthier 1986; Arrow et al. 2000).
In our study, we analyzed the efficiencyequality tradeoff in a controlled laboratory experiment. We explored a range of intermediate ‘fuzzy’ meritocracy regimes motivated by the fact that realworld mechanisms would typically not be perfectly precise. We consider a highfuzziness implementation the default for a realworld implementation, but consider the possibility that a policy maker could make investments to reduce the noise of the monitoring/implementation. When we consequently interpret setting the level of precision of the mechanism—for the moment without considering the costs that such an activity would entail—as a policy choice, we obtained the following result. Surprisingly, we found that the tradeoff is dissolved behaviorally. Matching mechanisms that are more meritocratic and that, therefore, promise higher efficiency from a theoretic point of view, also turn out to benefit the worstoff and to improve overall distributional equality. Theory predicted otherwise. This result was true even in parameter ranges where the highefficiency equilibria did not exist. This suggests that any fuzzy version of meritocratic matching would be beneficial to implement.
Our results rely on two critical assumptions. First, our experiments feature ex ante equity. Indeed, this is an important prerequisite, as meritocratic matching generally does not enable highefficiency equilibria with heterogeneity amongst agents (Duca et al. 2016). Second, group sizes are fixed. While alternative models have been proposed (Cinyabuguma et al. 2005; Ehrhart and Keser 1999; Ahn et al. 2008; Coricelli et al. 2004; Page et al. 2005; Brekke et al. 2007; Brekke 2011; Charness and Yang 2014), exploring combinations of endogenous groupformation and meritocratic matching is left as an avenue for future research.
Footnotes
 1.
Importantly, we use the terminology of ‘fuzzy mechanism’ and ‘fuzzy institution’ to refer to the imprecision of the mechanism, not about mechanism design with ‘fuzzy preferences’. Fuzzy mechanisms in our setting can range from perfectly precise to completely imprecise and random. This is different to the use of the word for example in ‘fuzzy social choice where standard social choice theory based on ‘crisp preferences is extended to the case of ‘fuzzy preferences’ that is, imprecise preferences, e.g. Blin and Whinston (1973) and Dutta (1987).
 2.
 3.
 4.
See also Isaac et al. (1985) for early, alternative implementations.
 5.
 6.
Ties are randomly broken.
 7.
See Gunnthorsdottir (2010) Theorem 1 for the theorem by which to compute these highefficiency equilibria.
 8.
Nax et al. (2014) is a separate theory paper that develops the relevant predictions that are being tested here.
 9.
Note that, due to the i.i.d. random draws, the exact same two numbers are generated with probability zero, so that tiebreaking is only needed when \(\sigma = 0\).
 10.
Higher contribution levels were observed online than in the lab for the \(\sigma ^2=\infty \) condition. More details are available in the Appendix.
 11.
At \(\sigma ^2=3\) the system became more volatile, and at \(\sigma ^2=20\) the system became unstable.
 12.
 13.
Further details on our computations to obtain equilibrium conditions for our parameter values can be found in Appendix A.1.
 14.
In order to run two variance levels per person, we play half of the 80 rounds of Gunnthorsdottir (2010) each time.
 15.
Kruskal–Wallis ranksum test \(\chi _{(1)}= 4.35, p = 0.0369\). Unit of observation is one session (\(n1=n2=3\)).
 16.
Ex ante inequality for \(\sigma ^2=0\), evaluated at expected payoffs in equilibrium, there are 2 freeriders earning 40 coins, and 14 fullcontributors earning 37.1 coins. Ex post inequality for \(\sigma ^2=0\), evaluated at realized payoffs in equilibrium, we have 2 freeriders earning 40 coins, 12 (lucky) fullcontributors earning 40 coins, and 2 (unlucky) fullcontributors earning 20.
 17.
Additional analysis of the Gini coefficient with more statistical tests can be found in Appendix A.3.
 18.
Note also that \(MU_{Adv}\) (\(MU_{Dis}\))is (dis)advantageous, because the players enjoys higher (lower) contributions from others than he would if meritocratic matching was implemented without flaw. Note that, because of this definition, meritocratic unfairness cannot arise under PERFECTMERIT (when \(\sigma ^2=0\)).
 19.
Notes
Acknowledgements
HHN and SB contributed equally to the manuscript overall. ROM and DH helped analyzing the data and writing the paper. The authors acknowledge support by the European Commission through the ERC Advanced Investigator Grant ‘Momentum’ (Grant No. 324247). The authors thank Bary Pradelski, Anna Gunnthorsdottir, Michael Mäs, Stefan Seifert, Jiabin Wu, Yoshi Saijo, Yuji Aruka, Stefano Duca, and Guillaume Hollard for helpful discussion and comments on earlier drafts, and finally members of GESS at ETH Zurich as well as seminar participants at the Behavioral Studies Colloquium at ETH Zürich, at the 25th International Conference on Game Theory 2014 at Stony Brook, at the Choice Group at LSE, at the TOM Seminar at PSE, and at the Kochi University of Technology for helpful feedback. All remaining errors are ours.
Supplementary material
References
 Abreu D, Pearce D, Stacchetti E (1990) Toward a theory of discounted repeated games with imperfect monitoring. Econometrica 58(5):1041–1063CrossRefGoogle Scholar
 Ahn T, Isaac RM, Salmon TC (2008) Endogenous group formation. J Public Econ Theory 10(2):171–194CrossRefGoogle Scholar
 Alchian AA, Demsetz H (1972) Production, information costs, and economic organization. Am Econ Rev 62(5):777–795Google Scholar
 Ambrus A, Greiner B (2012) Imperfect public monitoring with costly punishment: an experimental study. Am Econ Rev 102(7):3317–3332CrossRefGoogle Scholar
 Andreoni J (1988) Why free ride? Strategies and learning in public goods experiments. J Public Econ 37(3):291–304CrossRefGoogle Scholar
 Aoyagi M, Fréchette G (2009) Collusion as public monitoring becomes noisy: experimental evidence. J Econ Theory 144(3):1135–1165CrossRefGoogle Scholar
 Arechar AA, Gächter S, Molleman L (2017) Conducting interactive experiments online. Exp Econ. doi: 10.1007/s1068301795272
 Arrow K, Bowles S, Durlauf S (2000) Meritocracy and economic inequality. Princeton University Press, PrincetonGoogle Scholar
 Arrow KJ (1951) Social choice and individual values. Yale University Press, YaleGoogle Scholar
 Atkinson AB (1970) On the measurement of inequality. J Econ Theory 2(3):244–263CrossRefGoogle Scholar
 Balietti S (2016) nodeGame: realtime, synchronous, online experiments in the browser. Behav Res Methods. doi: 10.3758/s134280160824z
 Blin JM, Whinston AB (1973) Fuzzy sets and social choice. J Cybern 3(4):28–36CrossRefGoogle Scholar
 Bolton GE, Ockenfels A (2000) ERC: a theory of equity, reciprocity, and competition. Am Econ Rev 90(1):166–193CrossRefGoogle Scholar
 Brekke K, Nyborg K, Rege M (2007) The fear of exclusion: individual effort when group formation is endogenous. Scand J Econ 109(3):531–550CrossRefGoogle Scholar
 Brekke K et al (2011) Playing with the good guys. A public good game with endogenous group formation. J Public Econ 95(9):1111–1118CrossRefGoogle Scholar
 Charness G, Rabin M (2002) Understanding social preferences with simple tests. Q J Econ 117(3):817–869CrossRefGoogle Scholar
 Charness GB, Yang CL (2014) Starting small toward voluntary formation of efficient large groups in public goods provision. J Econ Behav Org 102:119–132CrossRefGoogle Scholar
 Chaudhuri A (2011) Sustaining cooperation in laboratory public goods experiments: a selective survey of the literature. Exp Econ 14:47–83CrossRefGoogle Scholar
 Cinyabuguma M, Page T, Putterman L (2005) Cooperation under the threat of expulsion in a public goods experiment. J Public Econ 89(8):1421–1435CrossRefGoogle Scholar
 Coricelli G, Fehr D, Fellner G (2004) Partner selection in public goods experiments. Econ Ser 48(3):356–378Google Scholar
 Cowell F (2011) Measuring inequality. Oxford University Press, OxfordCrossRefGoogle Scholar
 Duca S, Helbing D, Nax HH (2016) Assortative matching with inequality in voluntary contribution games, working paperGoogle Scholar
 Dutta B (1987) Fuzzy preferences and social choice. Math Soc Sci 13(3):215–229CrossRefGoogle Scholar
 Ehrhart KM, Keser C (1999) Mobility and cooperation: on the run. Tech Rep s24. CiranoGoogle Scholar
 Fehr E, Gächter S (2000) Cooperation and punishment in public goods experiments. Am Econ Rev 90:980–994CrossRefGoogle Scholar
 Fehr E, Schmidt KM (1999) A theory of fairness, competition, and cooperation. Q J Econ 114:817–868CrossRefGoogle Scholar
 Fischbacher U, Gaechter S, Fehr E (2001) Are people conditionally cooperative? Evidence from a public goods experiment. Econ Lett 71(3):397–404CrossRefGoogle Scholar
 Fudenberg D, Levine D, Maskin E (1994) The folk theorem with imperfect public information. Econometrica 62(5):997–1039CrossRefGoogle Scholar
 von Gaudecker HM, van Soest A, Wengström E (2012) Experts in experiments. J Risk Uncertain 45(2):159–190CrossRefGoogle Scholar
 Gauthier DP (1986) Morals by agreement. Oxford University Press, New YorkGoogle Scholar
 Green EJ, Porter RH (1984) Noncooperative collusion under imperfect price information. Econometrica 52(1):87–100CrossRefGoogle Scholar
 Gunnthorsdottir A, Thorsteinsson P (2011) Tacit coordination and equilibrium selection in a meritbased grouping mechanism: a cross cultural validation study. doi: 10.2139/ssrn.1883465
 Gunnthorsdottir A, Vragov R, Shen J (2010) Tacit coordination in contributionbased grouping with two endowment levels. Res Exp Econ 13:13–75CrossRefGoogle Scholar
 Gunnthorsdottir A et al (2010) Nearefficient equilibria in contribution based competitive grouping. J Public Econ 94(11):987–994CrossRefGoogle Scholar
 Hauser DJ, Schwarz N (2016) Attentive turkers: MTurk participants perform better on online attention checks than do subject pool participants. Behav Res Methods 48(1):400–407CrossRefGoogle Scholar
 Isaac MR, McCue KF, Plott CR (1985) Public goods provision in an experimental environment. J Public Econ 26:51–74CrossRefGoogle Scholar
 Ledyard JO (1995) Public goods: a survey of experimental research. In: Kagel JH, Roth AE (eds) The handbook of experimental economics. Princeton University Press, Princeton, NJ, pp 111–194Google Scholar
 Levin J (2003) Relational incentive contracts. Am Econ Rev 93(3):835–857CrossRefGoogle Scholar
 Marwell G, Ames RE (1979) Experiments on the provision of public goods. I. Resources, interest, group size, and the freerider problem. Am J Sociol 84(6):1335–1360CrossRefGoogle Scholar
 Marwell G, Ames RE (1980) Experiments on the provision of public goods. II. Provision points, stakes, experience, and the freerider problem. Am J Sociol 85(4):926–937CrossRefGoogle Scholar
 Nax HH, Balietti S, Murphy RO, Helbing D (2015) Meritocratic matching can dissolve the efficiencyequality tradeoff: the case of voluntary contributions games. doi: 10.2139/ssrn.2604140
 Nax HH, Murphy RO, Helbing D (2014) Stability and welfare of ‘MeritBased’ groupmatching mechanisms in voluntary contribution games. SubmittedGoogle Scholar
 Okun AM (1975) The big tradeoff. Brookings Institution Press, Washington, DCGoogle Scholar
 Page T, Putterman L, Unel B (2005) Voluntary association in public goods experiments: reciprocity, mimicry and efficiency. Econ J 115(506):1032–1053CrossRefGoogle Scholar
 Rawls J (1971) A theory of justice. Belknap Press, CambridgeGoogle Scholar
 Rud OA, Rabanal JP (2015) Efficient investment via assortative matching: a laboratory experiment. Available at SSRN https://ssrn.com/abstract=2565196
 Sen A (1970) The impossibility of a paretian liberal. J Polit Econ 78(1):152–57CrossRefGoogle Scholar
 Wang S et al (2015) Mechanical Turkbased experiment vs laboratory based experiment: a case study on the comparison of semantic transparency rating data. In: Proceedings of the 9th Pacific Asia conference on language, information and computation (PACLIC), Shanghai (pp 53–62)Google Scholar
 Wu J, Axelrod R (1995) How to cope with noise in the iterated prisoner’s dilemma. J Confl Resolut 39(1):183–189CrossRefGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.