Conditional cooperation, context and why strong rules work — A Namibian common-pool resource experiment

Conditional cooperation, context and why strong rules work — A Namibian common-pool resource experiment

Ecological Economics 129 (2016) 21–31 Contents lists available at ScienceDirect Ecological Economics journal homepage: www.elsevier.com/locate/ecole...

433KB Sizes 0 Downloads 37 Views

Ecological Economics 129 (2016) 21–31

Contents lists available at ScienceDirect

Ecological Economics journal homepage: www.elsevier.com/locate/ecolecon

Conditional cooperation, context and why strong rules work — A Namibian common-pool resource experiment Dirk Röttgers Institute for Environmental Economics and World Trade, Leibniz University of Hannover, Königsworther Platz 1, 30167 Hannover, Germany

a r t i c l e

i n f o

Article history: Received 12 August 2014 Received in revised form 22 February 2016 Accepted 13 March 2016 Available online xxxx Keywords: Common-pool resource experiment Conditional cooperation Framing Context Leadership Namibia

a b s t r a c t Common-pool resource experiments in the field have not yet fully explored cooperative decision-making in its dependence on established past behavior, explicit rules and environmental context. The purpose of this paper is to analyze these factors and their single as well as combined influence on cooperative behavior. The results show that rule enforcement changes the influence of past action as a determinant of cooperation: Past action retains relevance for decision-making but reveals a partly contrarian influence as long as rules are strong. Further, the introduction of explicit rules does not change the influence of personal relationships among participants, but renders the influence of leadership insignificant. Furthermore, the experiments show that context plays a decisive role, which suggests the possibility of tailoring experiments to specific uses: If the context-specific behavior of locals is relevant to policy decisions, then experiments can help capture these effects. © 2016 Published by Elsevier B.V.

1. Introduction and Literature Ecological economics is inherently anthropocentric. Ultimately, it must grapple not only with the ecological aspects of environmental problems but also their behavioral aspects. The inhabitants of an ecosystem will react not only to their surrounding environment and to the natural resources contained in it but also to each other and to the rules, implicit or explicit, by which they govern themselves. Since Ostrom (1990) conducted their seminal field work on community behavior in the presence of common-pool resources (CPRs), it has been clear that stakeholders exhibit cooperative behavior that prevents them from exploiting resources to the extent possible. Numerous experiments in the field in different parts of the world have since painted a similar picture with respect to overuse, although the outcomes of these experiments indicate diverse behavioral patterns (for overviews, see Agrawal (2003) and Sturm and Weimann (2006)). What is clear in all of these studies is that CPR extraction in the real world is not as exploitative as previously thought (Hardin, 1968) or as predicted by political or game theory (Ostrom et al., 1992). Much research has been conducted in communities in developing countries into how different forms of governance can solve problems of resource overuse, such as overfishing, deforestation and overgrazing (Agrawal, 1999; Cardenas et al., 2004; Cavalcanti et al., 2010; Janssen et al., 2013; Nordi, 2006; Ostrom, 2006; Travers et al., 2011). Many field experiment papers on CPRs focus on testing different rule

E-mail address: [email protected].

http://dx.doi.org/10.1016/j.ecolecon.2016.03.013 0921-8009/© 2016 Published by Elsevier B.V.

mechanisms but not on the factors involved in decision-making and on how such decision-relevant factors may be affected by rule changes. Indeed, the empirical literature on the determinants of decision-making in CPR experiments has been scarce to date. Velez et al. (2009) find that decisions hinge on expectations and result from the human tendency to conform to groups or to reciprocate1, i.e., individuals observing the behavior of others and making similar decisions. Furthermore, Hayo and Vollan (2011) attribute decisions to group cohesion, which leads, among other things, to conditional cooperation, of which reciprocity is one expression. These results partly explain decision-making in common-pool resource situations. Nevertheless, they do not fully answer the call of Anderies et al. (2011) to conduct more research on what they label “micro-situational variables and broader context”. Specifically, existing research does not enable us to determine why certain rules change decision-making and why others do not. Answering the implicit challenge of these findings, this paper helps fill the knowledge gap with regard to the determinants of decisions in cooperative field experiments with rule treatments. To my knowledge, this is the first paper to combine regression-based analysis with common-pool resource experiments under rule treatments.

1 There are, of course, various other-regarding tendencies that could, in principle, have the same effect. The most prominent of these is conformity (for an experiment on and discussion of conformity versus reciprocity, see Bardsley and Sausgruber (2005)). In this paper, other-regarding tendencies will be indicated exclusively by their effect, namely, conditional cooperation, because distinguishing among various other-regarding tendencies is neither useful nor possible given the present data, nor is it necessary to answer the research questions examined.

22

D. Röttgers / Ecological Economics 129 (2016) 21–31

For CPR use decisions, the relevant theoretical background is provided by Ostrom (1998). Ostrom's behavioral approach to rational choice theory identifies two important exogenous factors for cooperation: (i) the development and adoption of shared rules as well as (ii) past action (see Fig. 1). In this behavioral model of rational choice theory, both factors influence different parts of the self-reinforcing circular relationship among reputation, trust, reciprocity and cooperation. The development or adoption of shared rules can enforce cooperation in situations in which participants might otherwise shirk their responsibilities. For example, without punishment by a commonly agreed-upon fine, community members in a Namibian conservancy might not have an incentive to avoid detrimental slash-and-burn agricultural practices. Such shared rules are usually introduced with the explicit aim of increasing net benefits and therefore carry a plus sign in Fig. 1. The ambiguous sign of past action in Fig. 1 is due to the ambiguity of cues that community members might take from their own and their peers' past behavior. Past action can reinforce trust between group members either directly or by strengthening individuals' reputations, which builds trust. After all, if community members observe that their peers exploit an overused fishing ground less than they did before, thereby leaving more fish for others, an expectation of repetitive behavior might spur cooperation. Conversely, past action could also induce contrarian behavior: If community members know from past behavior that their peers are inclined to care for the ecosystem, they themselves can free-ride, or a community member could feel responsible for protecting the ecosystem because her peers evidently do not. Hence, Fig. 1 presents an ambiguous sign for past action. What this model of cooperative behavior fails to clarify, however, is the interplay between shared rules, on the one hand, and cooperative/ contrarian behavior as a reaction to past action, on the other. Both – rules and past actions – create expectations, if only subconsciously, of how people will behave. The question is which of these factors accounts for conditional cooperation and to what extent: the one that is more easily governed, i.e., rules, or the more elusive effects of past action? Accordingly, the first research question of this paper is ‘What is the relationship between past action and shared rules?’ Based on Ostrom's theoretical framework, I hypothesize that both past action and shared rules significantly affect behavior but that each has unique effects on conditional cooperation. Clarifying the interplay between rules and past action is the first main contribution of this paper. Knowledge of this relationship will make navigating community politics easier and facilitate decision-making about rule introductions in communities. In addition to the influence of implicit and explicit rules, it has become clear that increased group cohesion leads to increased

Fig. 1. A behavioral approach to rational choice theory, including additional factors of group decision-making. Adapted from Ostrom (1998).

cooperation (e.g. Dayton-Johnson, 2003; for an overview of behavioral psychology findings on this issue, see Evans and Dion, 1991). There are many possible drivers for this cohesion effect, even beyond past action and shared rules, (Tavoni et al., 2012; Velez et al., 2009). Following Ostrom (1998), such driving factors ultimately foster reputation and trust, including personal relationships and leadership (Moxnes and van der Heijden, 2003; Ostrom, 2006; van Dijk et al., 2003; Vedeld, 2000). If community members know each other well and have strong relationships, for example, they will trust each other more to cooperate, even when individuals are not monitored. As another example, obedience to the (actual or perceived) will of a strong leader might function just as a shared rule would and therefore increase cooperation. These other possible drivers of cooperation are also included in the analysis to determine their direct influence as well as their interplay with past action and rules. The second research question of this paper concerns a currently neglected aspect of artefactual field experiments: the framing of an experiment or the ‘context,’ as Anderies et al. (2011) term it. In general, framing affects experimental outcomes, and therefore, experimenters seek to test theories in abstraction from distorting outside factors to obtain results that are as generalizable as possible (Davis, 1993). However, if experimental results are used to justify policy measures that relate to a specific common-pool resource, experiments should approximate real-world ecological circumstances as closely as possible. Presumably, the specific ecological context influences real-life outcomes just as it would experimental outcomes. Therefore, the experiments described below aim at answering the question ‘How do different contexts influence the decision-making of participants?’. Using real-world contexts introduces more realism into the experiment so that the experimental results and their interpretations relate more clearly to real-world circumstances. The following analysis reveals the effects of framing experiments in ways that closely resemble participants' actual situations, in contrast to lab experiments and previous field studies. Making this distinction between lab and artefactual experiments and thereby enabling a reevaluation of the applicability of field experimental results to the actual policy challenges of communities is the second contribution of this paper. The field study site offers an opportunity to test such framing effects of context. Data were collected in September and October 2012 in the community-managed conservancy of Sikunga, an area in the Namibian Caprivi. Within the conservancy, fishing, farming and a fledgling tourism business are the primary income sources. Fishing and the relationship between wildlife tourism and subsistence farming present suitable CPRs as the context for CPR experiments. Sikunga's fishery resources are endangered as a result of overfishing, and the fledgling ecosystem-based tourism and hunting industries are endangered as a result of habitat destruction caused by slash-and-burn farming. While fishing has been extensively analyzed in CPR experiments (Bwalya, 2007; Cavalcanti et al., 2010; Fehr and Leibbrandt, 2008; Schnier, 2009; Velez et al., 2010), wildlife habitat maintenance has seldom been examined as a CPR issue (Travers et al., 2011). Nevertheless, a more complex and realistic context can bridge the gap between experimental research of a qualitative nature (e.g., Lankford et al. (2004)) and the purely quantitative but abstract experimental sphere. It is a unique feature of the dataset used in this paper that two relevant CPR contexts exist within one community. It is not simply that both ecosystem-based activities occur in the same area but that many participants depend on them economically and that all of the participants are familiar with the circumstances surrounding both activities. Accordingly, the two contexts of wildlife resources and fishery resources are the contexts that were used to frame the experiments. The characteristics of the dataset are further described in the next section along with other aspects of the experimental design. The results of the experiment under different rule settings are descriptively evaluated in Section 3. Sections 4, 5 and 6 present the regression analysis of decision-making, beginning with a section on the econometric method,

D. Röttgers / Ecological Economics 129 (2016) 21–31

which is followed by a description of the variables used and a discussion of the econometric results. The final section concludes. 2. Experimental Design The artefactual field experiment was run in forty sessions of groups of 5 in 7 different villages throughout the remote rural Sikunga conservancy in September and October 2012. The conservancy has a population of 2000 in 450 households of 7 villages (Davis, 2009). For each village, households were randomly drawn from a comprehensive household list proportionally to village size, with only one participant per household. Not only does the proportional sampling method guarantee a representative choice across villages, but it was also necessary to fairly distribute the winnings from the experiment throughout the conservancy, which was a sine qua non condition of the local traditional authorities for allowing the research in their territory. Replacements for absent household representatives were chosen from a list of oversampled households. When possible, the household's decisionmaker was asked to participate in the experiment, which was the case for 84% of the participants (econometrical robustness checks determined that participants identifying as household head did not influence results). Participants were not allocated fully randomly over treatments and contexts since fully randomized groups from all 7 villages would have meant hours of travel for some participants due to lack of transportation.2 Accordingly I was forced to allocate participants semirandomly to the 3 treatments and 2 framing contexts explained below to avoid systematic variation or a lack of variation due to common factors such as village membership, tribal allegiance and religious affiliation. That means that the randomly drawn participants were always in groups with participants from the same village. Further this means that in no village all groups of that village would participate in experiments with the same treatment or context combination to ensure variation and also ensure that e.g. a treatment dummy (see below) would not accidentally capture the effect of the village.3 Please note that the 3 treatments and 2 contexts overlap to save on already scarce observations. The following regression analysis will control for these effects. The sessions lasted between 2 and 3 h, depending mostly on the treatment variation, and were hosted by a local university student who was intensively trained to explain the experiment, guide the participants through the test rounds, encourage as well as answer questions and enforce the restrictions of the experiment, such as secret decision-making and keeping within the allotted time if the treatment allowed communication. Employing a local as the host was necessary because the participants often did not speak English well enough to understand the instructions in a language other than their native one. Except for the questions on environmental awareness explained below, all socioeconomic questions were asked after the experiment was completed. Under the experimental design based on Cardenas (2004), each of the 5 participants in each group decided on a personal level of resource extraction effort (henceforth effort or extraction effort) from 1 to 8, which the participant noted on his or her participant sheet (see Appendix) out of view of the other 4 group members. The minimum extraction effort was e = 1, and the maximum was e = 8. Each participant calculated their payoff for each round played, where the payoff depended on the decisions of both the participant and the other four players. Higher payoffs were obtained for higher individual effort, given the decisions of the other participants, whereas lower payoffs were obtained for higher cumulative extraction effort of the rest of the group, given the individual effort level. Every experimental group engaged in 20 2 After a few tests with full randomization (not reported), local authorities even threatened to stop cooperation due to the burdensome and long trip full randomization demanded. 3 The undesired randomized treatment/context allocations were redrawn until a combination fulfilling these criteria was found.

23

rounds. After all 20 rounds were completed, one round was chosen at random, and only the payoff for this round was paid out in Namibian Dollars. Note that the payoff equation of Cardenas (2004) was adjusted with a factor for currency conversion to Namibian dollars (cN$):   πi ¼ cN$  axi  0:5  b  x2i þ α  n  e  α∑xi :

ð1Þ

Given n = 5 players, a maximum effort e = 8, a conversion factor of cN$ = 5/64 and coefficients α = 20 , a = 60 and b= 5, Eq. (1) becomes4: πi ¼

X  5  xi :  60xi −0:5  5  x2i þ 20  5  8−20 64

ð2Þ

For payoff function (2), the Nash equilibrium (NE) is an extraction effort of 8, while the social optimum (SO) is 1, which is a corner solution.5 For the full payoff table given a participant's decision and the cumulative decisions of the four other participants rounded to the next dollar, see the Appendix. The participants were not allowed to communicate, they decided on extraction levels in secret for the first 10 of the 20 rounds played. They did not receive any feedback from the host except regarding the overall effort of the group, which was announced after all decisions were made. In the final 10 rounds, the experiment's host introduced one of three different rule treatments, which represented the only difference from the previous 10 rounds. As summarized in Table 1 below, under the three treatments — an open communication treatment and two punishment treatments — additional rules were introduced to 30 of the 40 groups. To create a baseline, 10 groups played all 20 rounds under the same conditions throughout the experiment (baseline treatment; BL). Of the remaining 30 groups with rule treatments, members of the 9 groups with open communication were allowed to speak to one another and coordinate their decisions after round 10. Before the 11th round, the communication treatment (CT) was introduced. Participants were allowed to speak for 10 min, and for rounds 12–20, they were allowed to speak for 5 min between rounds. In these 9 groups, however, each individual continued to make their own decision in secret, and none of the group discussions and agreements were binding. The 21 groups with a punishment treatment were told that the new rule only allows them to choose an effort level of 1. Although this is the socially optimal level, this fact was not revealed to the participants to separate the effect of the rule from the effect of new knowledge. Further, the participants were told that they would suffer a penalty of f = 4 Namibian dollars toward their payoff for every unit above the socially optimal effort level of xso = 1. After the rule was established following round 10, these participants continued to make their decisions in secret and in silence. In these groups, the participants' decisions were checked, and participants were penalized with two different likelihoods if they cheated: For 10 of the groups, the likelihood of being checked (‘being caught’) was P = 0.2 (weak enforcement (WE)); for the other 11 groups, the likelihood was P = 0.7 (strong enforcement (SE)). Punishment was purposely established without cost to the group. A lack of cost for punishment ensures that the participants decided to follow the rules without considering the effect that a possibly costly punishment would have on their peers. Otherwise, possible regard for peers would have been harder to 4 Because Namibia is a country with high income inequality, national or even regional income averages as well as industry minimum wages could not be used as a reference for the range of payoffs. Instead, the payoff table refers to local wages. Qualitative interviews prior to the experiments revealed that the two formal forms of work in the conservancy, at tourism lodges and at rice farms, were paid a typical day's wage of 50 and 60 Namibian Dollars, respectively. Accordingly, the payoffs in the experiment, which were intended to compensate an individual for at least half a day's work, were set so that they would compensate even for the low day-wage of participants in a group that played the NE, i.e., the worst outcome. Hence, the payoff of the NE was half of 50. 5 Maximizing (1) over xi yields the NE xNash = (a - α)/b, which is equal to 8, given i a=60,α=20 and b=5. Maximizing ∑xi over xi and exploiting the identical payoff functions for all i yields the social optimum, xSO i =(a-αn)/b, which is 1, given the above parameters and the fixed minimum extraction effort of 1.

24

D. Röttgers / Ecological Economics 129 (2016) 21–31

Table 1 Distribution of contexts and rules over experimental groups. Source: own. Rule/context

Wildlife

Fishery

No context

Sum

None Communication Weak enforcement Strong enforcement Sum

3 1 4 4 12

4 4 2 4 14

3 4 4 3 14

10 9 10 11 40

Note: Numbers are groups.

separate from the rule effect as it already is, and it also might have been harder to disentangle from the effects of the sole rule in the regression analysis of Sections 4 to 6. Adding terms for the penalty and the likelihood of being checked to Eq. (1) yields Eq. (3):   X 1 xi −P  f  ðxi −xSO Þ : ð3Þ π i ¼ cN$  axi −0:5  b  x2i þ α  n  e−α cN$ Accordingly, considering the same parameters as in Eq. (2), with weak enforcement (P = 0.2), the NE is xNash-WE = 6, and with strong eni - SE = 1, which again is a corner forcement (P = 0.7), the NE is xNash i solution.6 This difference in the likelihood of being caught is used to examine the effects of strong versus weak enforcement. Varying the likelihood P rather than the size of the penalty f simulates the local circumstances more realistically: Locals reported in previous stakeholder workshops that the likelihood of being checked and punished by game guards for abuse of the ecosystem has a strong influence on decisions about whether to transgress. Therefore, differences between weak and strong enforcement are captured by differences in the likelihood of being caught cheating rather than by differences in the severity of penalties. To simulate being checked and possibly caught, all of the participants randomly drew one marble from a bag of ten at the end of each round. If the marble was black, the participant's sheet was marked for cheating in this round, and the participant was penalized accordingly if this round was the randomly chosen payoff round. In the weak enforcement treatment, with a 20% likelihood of being checked, the bag contained two black and eight white marbles. Similarly, in the strong enforcement treatment, the bag contained seven black and three white marbles. To examine framing effects, groups engaged in the experiment in either of two different contexts or in a context-free baseline version. In the baseline version, participants were told that the experiment concerned an abstract natural resource, but they were not given any additional information. This treatment, which was employed for 14 of the 40 groups, kept the experiment neutral. In the two other contexts, the experiment was orientated toward two types of resource overuse that are prevalent in the area. The first of these overuse cases was overfishing, which is a classic CPR problem. Although only a small fraction of the Sikunga population identifies as fishermen7, the nearby fish stocks of the Zambezi River are noticeably depleted (Heider, 2012). To reflect this condition, 14 groups were told that the experiment concerned fishing resources, and the rules were explained using fishing as an example. The participants had to decide on effort levels in the form of sending between 1

and 8 of their hypothetical household members either to fish or to engage in subsistence farming on marginal land. The second resource overuse case was the use of slash-and-burn practices. Although farmland itself is not a CPR, slash-and-burn practices have an indirect but strong effect on wildlife in the area, as both the burned farmland and surrounding areas are wildlife habitats. While there are many ways to profit from wildlife, the one chosen for the experiment involves regularly issued and tradable hunting permits (quotas). Depending on the amount of wildlife prevalent at a given time, the conservancy regularly receives permits to shoot a given number of animals. Such permits can be resold to professional or recreational hunters for substantial financial return. This money has in the past been distributed among members of the conservancy. In this context, money earned from the sale of hunting permits is the CPR. Slash-andburn farming takes the role of an extractive action because, while it diminishes wildlife habitats and thus negatively impacts the equally distributed hunting quota, it increases the productivity of the land and therefore the yield and farming income. Again, as in the fishing context, farming on marginal land is an alternative to the exploitative choice. The tradeoff decision in this CPR context becomes the following: Slash-andburn farming for the farmers' own good versus more sustainable but marginal farming that preserves the valuable wildlife habitat. The remaining 12 groups engaged in the experiment in this context. The participants had to decide on the number of fields, between 1 and 8, that participants could farm using slash-and-burn practices. Table 1 shows the distribution of rule changes and context treatments over all 40 groups. For both contexts, the experiment's host used visual aids to make the respective context tangible (see Figs. 3 and 4 in the Appendix). The main visual aid was a wooden board with a map of the conservancy and the bordering Zambezi River, which contains the community's fishing grounds. Depending on the context applied, the experiment's host used either animal figures on the conservancy map or ellipsoid marbles representing fish on the Zambezi part of the board. Using these tools, the host illustrated the ecological conditions and the differences between large and small individual and group effort levels. The game host was trained to exclude value judgments regarding different strategies from his explanation. The experiment's host explained the setup in the local language and then answered the participants' questions, trained the participants to read the payoff table, explained how to indicate decisions on the provided player sheet and conducted 3 test rounds. He then again answered participants' questions (see the Appendix for the payoff table and a player sheet). After conducting the full experiment, the host administered a short questionnaire to collect the data described below.8

3. Descriptive Results Fig. 2 shows the average effort level chosen by the participants, ranging from 1 to 8, in each round by rule treatment. The drop in effort in all of the treatments after round 10 indicates an effect of the treatments on the participants' behavior. The baseline treatment hovers close to its mean value of 3.69. The baseline results, thus, are not only far below the baseline NE (8) but even below the NE in the weak enforcement case (6). Furthermore, the baseline effort levels do not significantly deviate from this value before or after round 11. Participants did not behave differently simply because they participated in more rounds or

6

The corresponding equation for the NE is xNash-WE/SE =(a-α-P⋅f)/b. i Only 6% of the participants indicated that fishing was their main occupation, although qualitative workshops revealed that a social stigma is attached to not living off one's own land. Accordingly, participants might have chosen to describe themselves as farmers rather than fishermen, not because their main occupation is farming, but because they did not wish to identify themselves as fishermen, although they may mainly subsist on fishing. 7

8 The two exceptions are the variables on awareness, for which data were recorded in a pre-experimental questionnaire, and data on self-reported slash-and-burn farming, which were collected using an accompanying household survey that was answered by the same respondents. The author will gladly provide all of the questionnaires as well as the wordfor-word transcript of the host's explanations upon request via e-mail.

D. Röttgers / Ecological Economics 129 (2016) 21–31

25

Fig. 2. Average effort under different treatments. Source: own depiction.

learned something which significantly made them change their strategy. Under the Communication treatment, there is a slow drop in effort levels over the final 10 rounds, which is similar to previously reported results, most importantly in Ostrom et al. (1994). While initially, after the first round in which talking is permitted, participants only slightly deviate from their strategies of rounds 1–10, they deviate further in the rounds that follow. Although the values fluctuate too much to provide clear implications, there appears to be a continuous downward trend leading to relatively low values of between 2 and 3. This downward trend might be a sign of sustainability under the rule change. With regard to the time limit for discussion under the Communication treatment, it should be noted that participants never had to be told to stop talking because their time had ended in any of the 10-minute discussion periods, and they rarely used the 5-minute slots at all after round 11. Under the punitive treatments, weak enforcement and strong enforcement, there was an initial drop in effort levels nearly to the SO followed by a slow but steady rise in the remaining rounds. However, under these treatments, the average effort never returned to the baseline values. Instead, the lines for both punishment treatments appear to approach their asymptotes, which are approximately 3.5 in the case of weak enforcement and just above 2 in the case of strong enforcement. Participants were likely initially intimidated by the prospect of punishment and then, as they experienced the new rule, they gained a sense of how costly it would be to break it. This effect is not unusual in behavioral experiments (Smith, 2010), and a tendency for participants to become more competitive, i.e., to test whether they can gain a greater share, has also been observed in repetitive CPR experiments (Gillet et al., 2009). In light of the Nash equilibria for the enforcement treatments, which are 6 and 1, respectively, this result has implications for the interpretation of participants' behavior under both punitive rules. Under the weak enforcement treatment, although participants disobey the new rules to some extent, they nevertheless behave in a more social manner than the NE predicts. Under strong enforcement, participants are even oriented toward effort levels above the SO and even above the new NE, which is equal to the SO. Over time, participants move even further from the SO. The regression analysis of Section 5 will provide explanations for this seemingly unreasonable behavior. The result that all of the treatments are effective is confirmed by a signed rank test. The results in Table 2 show that the group decisions made under the treatments are significantly different from decisions

made under the baseline treatment. The signed rank tests show the expected significant impact of treatments and no statistically significant changes in the baseline groups. 4. Econometric Model and Data To analyze the factors of participants' decision-making, the influence of past action, rule treatments, socioeconomic factors, cohesion factors, context-related factors and rules interactions are regressed on the effort levels that the participants chose. The regression model is as follows: 0

Decision ¼ α þ βOthers Effort t1 þ γDifferencet1 þ δRules þ ζ Context þ ηCohesion þ θSEcon þ ρRulesInteract þ ϵ:

ð4Þ

Others' Effortt − 1 captures the past actions of others in the group as the average effort of the other participants in the previous round, and Differencet − 1 similarly captures the individual participant's past deviation from the group average by subtracting Others' Effortt − 1 from the participant's own effort in round t − 1. Rules is a vector consisting of three dummies for the different rule treatments (Communication, Weak Enforcement, Strong Enforcement), Context is a vector of variables indicating context specificity, Cohesion is a vector of two variables that capture the social distance/closeness of the participants, SEcon is a vector of socioeconomic variables, and RulesInteract is a vector of interactions between the extraction effort of others in the previous round and the effects of treatments. In the following, the relevant variables and hypothesized outcomes are described. An overview, including basic statistical information, is provided in Table 3. The lag variable Others' Effortt − 1 captures the effect of the cumulative extraction effort of other members of the experimental group in the last round. This variable reveals the influence of other participants'

Table 2 Results of a signed-rank test for differences between treatments (within groups). Source: own. Tested treatment

z-Value

Significant difference

WE SE CT

4.673 4.687 4.673

Yes Yes Yes

Note: Observations are individual decisions.

26

D. Röttgers / Ecological Economics 129 (2016) 21–31

Table 3 Descriptive statistics of employed variables and additional demographics. Variable

Obs.

Mean

SD

Min.

Max.

Decision Others' Effortt − 1 Schooling Children Age Household Size Gender (1 = male) Personal Relationships Leadership Awareness Wildlife Awareness Fishery Wildlife Fishery Slash & Burn Farmers ∗ Wildlife Fishermen ∗ Fishery

3800 3800 188 200 200 200 200 200 200 200 200 200 200 195 200

3.44 13.88 9.18 2.84 35.64 4.66 0.34 3.19 0.30 2.75 4.84 0.30 0.35 0.22 0.10

2.24 5.64 2.73 2.16 13.81 3.18 0.47 1.25 0.46 3.15 5.59 0.46 0.48 0.42 0.29

1 4 1 0 14 1 0 0 0 1 1 0 0 0 0

8 32 17 9 80 32 1 4 1 25 25 1 1 1 1

Note: Decision and Others' Effortt − 1 are in units of effort (see above), Schooling and Age are in years, Children, Household size and Personal Relationships are in numbers of people, both Awareness variables are the product of the multiplication of two 5-step Likert scales and all of the other variables are binary.

behavior as a guide to decision-making and is thus an indicator that may be explained by a signal reinforcements similar to an information cascade, which in turn leads to herding behavior (Baddeley, 2010; Drehmann et al., 2005; Miller and Page, 2004), or by the above model of conditional cooperation. Similarly, the other lagged variable, Differencet − 1, tests the effect of a player's deviation from the average group decision in the previous turn. To test the effects of the introduced Rules on herding behavior, model 2 includes dummy variables for the rule treatments Communication, Weak Enforcement and Strong Enforcement. Please note that due to resource constraints only circa 50 individuals, i.e. 10 groups, are allocated to each treatment variation. Context introduces dummies for the experiment's context and also includes variables on the participants' environmental disposition with respect to the two different ecological contexts, wildlife and fishery. In particular, participants were asked about their degree of Awareness of Wildlife problems and their degree of Awareness of Fishery problems. Awareness of wildlife and fishery problems were each measured on two scales that ranged from 1 to 5. On the first scale, participants indicated how strongly they believed the number of fish or big game has changed in recent years. On the second scale, participants indicated the degree to which they believed the change negatively impacted locals in the conservancy. These results were multiplied to determine the awareness variables used in the regressions. Awareness is included to test for its influence on environmentally friendly behavior. The Context vector also contains variables that control for the experimental frame itself, i.e., dummies for the Wildlife context and the Fishery context. These two context dummies are regressed on their own and are used in interactions to more closely examine the relevance of context to the participants: The interaction between the Wildlife context and a dummy that indicates whether a given participant is used to destroying wildlife habitat through Slash-and-Burn Farming in real life captures a similar effect in the wildlife context. Self-reported real-life use of slash-and-burn practices, although condemned in many parts of the world (Kleinman et al., 1995), is unlikely to be underreported due to the social acceptance of such practices in the area. Similarly, the interaction between the fishery context and a dummy that indicates whether a participant identifies as a fisherman (Fishers) captures any additional impact that context may have on the behavior of an individual who might have an established behavioral pattern within the context. Cohesion includes two variables that pertain to the social structure of the experimental groups: Personal Relationships and Leadership. Personal Relationships is measured as the sum of the other participants with

whom a participant claims to have a personal relationship. The more personal relationships a participant has, the more socially a participant might behave to avoid undermining relationships with friends and family. Leadership is an indicator of the influence of hierarchy or of how strongly participants are influenced by the behavior of perceived leaders. Influence in the village captures one of the defining (and, in the Sikunga case, one of the few) properties of leadership (Rogers and Cartano, 1962): Elevated social status. Therefore, rather than employing the more common self-designation technique, participants were asked to rate their peers with respect to influence, as it is one's peers who grant an individual social status (compare Goldsmith and Desborde (1991) and Jacoby (1974)). The participants reported their opinions in secret regarding who among the others in the group had the most influence in their village. For each group in which the same individual was named by 3 or 4 participants as influential in decision-making, the dummy for Leadership is 1, indicating that a perceived opinion leader was present, and 0 otherwise. The SEcon variables included in the regression are the socioeconomically relevant variables of Schooling and Children. Educational attainment, measured in years, could equally be an indicator of how well participants understand the nature of common-pool resources and how to outcompete others in the experiment. Number of children in a participant's household is used as an indicator of the number of dependents the participant must care for. As noted in the results section below, many additional socioeconomic variables were tested but were found not to add explanatory power to the model. The final vector, RulesInteract, contains three interactions between the three rule treatments, on the one hand, and Others' Effortt − 1, on the other. These interactions show the influence that the perceived behavior of other players has under each rule treatment. While models 1 through 4 provide the foundation for the analysis of conditional cooperation – with model 1 establishing the herd effect, model 2 testing the influence of rules and models 3 and 4 refining these results – additional models add control variables and test for the effects of context and leadership. Among these, model 5 tests the robustness of conditional cooperation to the introduction of leadership, context variables and further controls. Models 6 through 8 examine groups of these additional variables separately to distinguish between possible separate effects of the variables in question. Model 9 tests for the interplay between rule treatments and the herd effect by introducing the interaction variables. Model 10 concludes the regression analysis by providing a full model for further robustness checks. The estimator is a right- and left-censored Tobit model with clusterrobust standard errors. The use of a censoring model is appropriate because both the number 0 and numbers higher than 8 for the dependent variable – decisions regarding extraction effort level – are gametheoretically possible and plausible in the real world, but the experiment restricts them to values between 1 and 8. Additionally, the regression must account for a possible relationship between the 20 consecutive decisions of a participant. Because it is plausible that an underlying decision-making process of a participant does not correlate with that of other participants, apart from the variables that are controlled for, the data require the use of cluster-robust standard errors clustered at the level of the individual participant. Furthermore, each model includes a set of group dummies to control for group fixed effects of different experimental groups, such as time of day, temperature, the field lab being in the open or in a closed room, other slightly altered practical conditions of the experiment, and outsiders unavoidably disturbing the group, among other possibly otherwise unrecorded disturbances. The sample of 4000 observations is trimmed by one observed round per participant due to the model's lagged variables. Additionally, some observations are discarded due to missing values for variables in some of the models (see Table 3), resulting in an overall sample size of between 3439 and 3800 observations (depending on missing observations) from between 181 and 200 participants.

D. Röttgers / Ecological Economics 129 (2016) 21–31

27

Table 4 Determinants of the common-pool resource game decision (1–8).

Constant Others' Extractiont − 1 Differencet − 1 Communication Weak Enforcement Strong Enforcement

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Basic

Rule treatments

Decision difference

Difference & rules

Other factors w/o rules

Awareness

Cohesion

Socioecon only

Rule interactions

All factors

⁎⁎⁎ 4.083 ⁎⁎⁎ 1.241 ⁎⁎⁎ 2.990 2.188 (0.540) (0.511) (0.394) (0.386) ⁎⁎⁎ −0.175 ⁎⁎⁎ 0.471 ⁎⁎⁎ 0.122 0.178 (0.067) (0.053) (0.061) (0.046) ⁎⁎⁎ 0.324 0.393 (0.041) (0.039) −1.420 ⁎⁎⁎ −1.164 (0.243) (0.203) ⁎⁎⁎ −1.205 −1.055 (0.290) (0.239) −3.391 ⁎⁎⁎ −2.884 (0.384) (0.329)

⁎⁎⁎ 1.763 (0.570) ⁎⁎⁎ 0.499 (0.065) ⁎⁎⁎ 0.375 (0.044) ⁎⁎⁎ ⁎⁎⁎ ⁎⁎⁎

⁎⁎⁎ 4.851 (0.376) ⁎⁎⁎ 0.125 (0.046) ⁎⁎⁎ 0.324 (0.039) −1.164 (0.203) −1.056 (0.240) −2.884 (0.330) ⁎⁎ −0.043 (0.023) 0.017 (0.015) ⁎⁎⁎ −2.293 (0.412) ⁎⁎ −1.878 (0.525) ⁎⁎ 0.738 (0.313) −0.082 (0.251) ⁎⁎

−0.048 (0.023) 0.009 (0.015) −2.120 (0.428) −1.171 (0.577) 0.724 (0.324) −0.141 (0.221) −0.165 (0.068) −1.268 (0.582) ⁎⁎⁎ 0.110 (0.041) ⁎⁎⁎ 0.104 (0.036)

Awareness Wildlife Awareness Fishery Wildlife Fishery Slash-and-Burn Farmers ∗ Wildlife Fishers ∗ Fishery Personal Relationships Leadership Schooling Children

⁎⁎⁎ 3.253 (0.437) ⁎⁎⁎ 0.121 (0.046) ⁎⁎⁎ 0.321 (0.039) ⁎⁎⁎ −1.164 (0.202) ⁎⁎⁎ −1.057 (0.240) ⁎⁎⁎ −2.886 (0.330) ⁎

⁎⁎⁎ 1.346 (0.600) ⁎⁎⁎ 0.147 (0.047) ⁎⁎⁎ 0.315 (0.041) ⁎⁎⁎ −1.045 (0.198) ⁎⁎⁎ −1.145 (0.255) ⁎⁎⁎ −2.921 (0.353)

6.2% 14,725 15,006 3800

5.9% 14,774 15,042 3800

7.5% 14,522 14,809 3800

6.5% 13,270 13,571 3439

⁎⁎⁎ ⁎⁎⁎ ⁎⁎⁎

⁎⁎

−0.130 ⁎⁎ (0.063) 0.097 (0.617) 0.107 (0.039) 0.120 (0.038)

7.7% 14,160 14,471 3705

⁎⁎⁎

⁎⁎⁎

Weak Enforcement Others' Extractiont − 1 ∗ 4% 15,063 15,326 3800

⁎⁎⁎

⁎⁎⁎

Others' Extractiont − 1 ∗ Communication Others' Extractiont − 1 ∗

Strong Enforcement Pseudo-R2 AIC BIC N

⁎⁎

7.6% 14,516 14,809 3800

8.1% 13,173 13,462 3477

⁎⁎⁎ ⁎⁎⁎

⁎⁎⁎ 3.882 (0.559) ⁎⁎ 0.155 (0.066) ⁎⁎⁎ 0.298 (0.041) ⁎⁎⁎ −1.390 (0.460) −0.030 (0.661) ⁎⁎⁎ −3.020 (0.575) −0.044 (0.022) 0.009 (0.015) −2.534 (0.437) −1.647 (0.587) 0.740 (0.321) −0.096 (0.229) −0.166 (0.068) −0.562 (0.616) 0.109 (0.040) 0.107 (0.037) 0.150 0.100 (0.106) (0.107) −0.436 ⁎⁎ −0.340 (0.178) (0.186)

3.021 (0.456) 0.136 (0.062) 0.315 (0.039) −1.673 (0.463) 0.364 (0.625) −2.798 (0.549)

−0.026 (0.171)

0.054 (0.176)

7.6% 14,513 14,819 3800

8.3% 13,028 13,366 3439

⁎⁎⁎ ⁎⁎ ⁎⁎⁎ ⁎⁎⁎

⁎⁎⁎ ⁎⁎

⁎⁎⁎ ⁎⁎⁎ ⁎⁎

⁎⁎

⁎⁎⁎ ⁎⁎⁎



Notes: The model is a Tobit I regression with cluster-robust standard errors and dummies for experimental groups; standard errors are in brackets; significance levels are ⁎p b 0.1, ⁎⁎p b 0.05, ⁎⁎⁎p b 0.01. Common household and individual variables and other possible factors for cohesion, such as the proxy for income and assets and variables for gender, education, household size, the participants being the household head, religious affiliation, ethnicity and village dummies, were tested in numerous robustness checks. Their exclusion did not result in any relevant differences.

5. Econometric Results Table 4 shows the results of the ten regressions described in the previous section, starting with a naïve model and then adding the vectors described above for additional insights. For all of the variables, a negative coefficient indicates decreasing extraction effort for an increase in the variable's value. A reduction in extraction effort can be interpreted as either environmentally friendly, since a smaller extraction effort per se leads to a more sustainable environment, or as social, since a smaller extraction effort leaves more resources for others to extract. Because the recorded data do not indicate which of these two rationales may motivate a participant's decision-making, the interpretation of each variable's influence is based on what appears to be most appropriate for that variable. A full interpretation and discussion of the regression results is carried out in Section 6. Model 1 shows that the herding effect/conditional cooperation, which are frequently observed in group experiments, are also present in this common-pool resource experiment (compare Hayo and Vollan,

2011), as shown by the positive coefficient for Others' Effortt − 1. This result is put into perspective in model 2. Model 2 includes the dummies for all three rule treatments, which causes the coefficient for Others' Effortt − 1 to become negative, while all three rule dummies are themselves negatively significant. The negative coefficients for the rule dummies show that all of the rules are effective: All three decrease the extraction effort of the participants, whereas the more sizable effect of Strong Enforcement compared with the other treatments, as shown in Fig. 2, is reflected in a coefficient that is more than 2.5 times larger. Models 3 and 4 introduce Differencet − 1, which itself is positive and, as model 4 shows, makes Others' Effortt − 1 stay positive even in the presence of the Rules variables. Notably, with the introduction of the difference between past player and average group decisions, the coefficient for Others' Effortt − 1 increases by a factor of 2.5, and the coefficient for the difference in past decisions is almost as high. This seeming independence of conditional cooperation and an effect of rules contains the answer to the first research question and is one of the main results discussed in Section 6.

28

D. Röttgers / Ecological Economics 129 (2016) 21–31

Model 5 shows that even in the presence of other controls (but in the absence of rule dummies), the effects of past behavior persist. Furthermore, model 5 includes context-specific variables. Of the two awareness variables, Awareness Wildlife and Awareness Fishery, only awareness of the deterioration of fish stocks induces a significant decrease in harmful behavior. Independently, the contexts themselves both have a negative effect on effort, with the effect of Wildlife notably stronger than that of Fishery. The significantly positive coefficient for the interaction between Slash-and-Burn Farmers and Wildlife further reflects that those who engage in environmentally unfriendly actions as part of their real-life economic activity also do so in the experiment, given a relatable context. Among the control variables for cohesion, having more Personal Relationships among group peers is associated with more social behavior, and the mere presence of Leadership in the group appears to compel participants to behave more socially. The socioeconomic controls Schooling and Children both show a positive coefficient, which indicates an unsocial effect that is non-negligible in size. Models 6 through 8 check the robustness of model 2's results by excluding groups of control variables. The only noticeable difference between the results of the first two models and models 6 through 8 is the insignificance of Leadership in model 7. Indeed, the variable is insignificant in all of the models that include the rule-treatment dummies (including the following models 9 and 10). Beyond this relationship between the rule dummies and Leadership, the coefficients and their significance levels are robust, thus confirming the previous results. Model 9 checks for an interaction between Others' Effortt − 1 as an orientation for cooperation and rules. It is here – with the results of model 9 – that Weak Enforcement reveals its separate effect. The interaction carries a negative and significant coefficient, meaning that given the result for Others' Extractiont − 1, under the Weak Enforcement treatment there is a separate, opposite effect of knowing what the group in general does. This effect is so decisive that it accounts for the full effect of Weak Enforcement, and the rule dummy on its own becomes insignificant. The all-encompassing model 10 includes all of the variables to further test the robustness of all of the previous models. Except for the aforementioned differences between the models, there are no additional noteworthy changes here. Thorough robustness checks (not reported) show that the results of all of the regressions are stable. Furthermore, other household and individual variables that are commonly used in microeconomics regressions, such as age, age2, household size, a proxy for income, a proxy for assets, occupation, gender, the participant being the household head, religious affiliation, ethnicity and village dummies, were tested in numerous robustness checks. Their exclusion did not result in any relevant differences. The same holds for the exclusion of a dummy set to 1 if participants had spoken with previous participants about the content of the experiment and 0 otherwise. The relatively low pseudo-R2 values of models 1 through 10 are due to randomness that is inherent in human behavior, which, after all, is the subject of economic experiments. The CPR experiments presented in this paper are no exception to this rule (for a close comparison of the explanatory power of the models, see Hayo and Vollan, 2011). Among the 10 models, those with additional variables yield lower values for the AIC and the BIC, which shows that the specifications were indeed useful and not simply a case of overfitting. Furthermore, comparing the cluster-robust Tobit results of Table 4 to Tobit random-effects and simple OLS specifications shows that the results cannot solely be caused by specifying, and thus possibly misspecifying, the regression as cluster-robust: The results of all specifications are highly similar. For instance, the correlation between the predicted values of the cluster-robust Tobit estimation and the random-effects estimation is beyond 0.99 (see Fischbacher and Gächter (2010) for this method of comparison). I report only the Tobit cluster-robust results because they are conceptually the safer choice, as noted above.

6. Interpretation & Discussion These results allow new insights regarding the interplay between past behavior, i.e., the variable Others' Effortt − 1, and the rules, i.e., the three rule dummies: both have an effect on their own. In search for an answer to the first research question on the relationship between past action and rules, the following sub-section on Past Action and Rules will therefore deal extensively with the interpretation and explanation of this relationship and its underpinnings. The sub-section The Influence of Context will resolve the second research question on the effect of context. 6.1. Past Action and Rules Regarding past action and rules, model 1 reveals that, without controlling for other determinants of decisions, participants take orientation from their peers: The positive coefficient implies that for every additional unit of cumulative peer extraction effort, participants increase their own effort, though only by a fraction of an effort unit. The coefficient can be interpreted as a size indicator for the utility of conforming to the informal norm implied by past action, or a positive delta parameter (Ostrom, 2009; Ostrom and Crawford, 1995). Ostrom (2009) argues, based on the behavioral approach to rational choice theory depicted in Fig. 1, that the payoff for this conditional cooperation depends on knowing more of the delta parameters of others. This interpretation could explain the results of introducing rule dummies in model 2 and again in model 4. All rules are effective in that they let the group come closer to making “[…]the presence, size, and sign of delta parameters […] common knowledge over time” (Ostrom, 2009). The introduction of rule dummies in the regressions relieves past behavior of the norm- and therefore behavior-defining role to make an indication about the delta parameter of other group members. This conclusion is consistent with both i) the theorized effect of knowing the delta parameters of the other participants and ii) Vollan's (2008) result that pro-social behavior can be crowded out by rules. Crowding out is caused by rules since participants use rules as indicators of how their peers will behave. Accordingly introducing the rule dummies leaves Others' Effortt − 1 to capture other effects, in this case seemingly a contrarian one, as indicated by the negative sign in model 2. The contrarian effect shows that, as long as there is a positive amount of utility derived from following rules, in model 2 and afterwards represented by the rules' coefficients, there seems to be a certain disutility from going along with the herd even further. Broken down specifically to the results of Table 4, participants behave in altruistically/environmentally friendly ways when others do not, and vice versa, given that the existing formal rules have a contrary effect. The fact that Others' Effortt − 1 carries both a cooperative effect if rules are not held constant and a contrarian effect if rules are included in the regression, is somewhat puzzling: Why would rules change the effect of peer-orientation? However, this puzzle is solved by adding Differencet − 1 as a factor in models 3 and 4. Comparing the results of these two models shows that, again, Others' Effortt − 1 only carries conditional cooperation, rules keep their significance and their intended negative effect, and the contrarian effect is only found in Differencet − 1. This shows that past action can have and indeed has two possible outcomes: The past action of others generally will trigger a conditional cooperation, but setting what others have done in relation to one's own past decision will reveal a contrarian effect. The presence of both types of effects is expected in situations in which benefits are conferred (Bowles and Gintis, 1998). In such situations, it is consistent with lab and field experimental results that both types of behaviors exist, but pro-social behavior is rewarded more often and is therefore the more prevalent behavior (Axelrod and Hamilton, 1981; Bowles and Gintis, 2000). This result is supported by a comparison of the coefficient size of models 3 and 4 for the two relevant variables: The average and range of values for Others' Effortt − 1 is much larger than for

D. Röttgers / Ecological Economics 129 (2016) 21–31

Differencet – 1's average of 0 and range between negative and positive 7, yet the coefficient in model 3 is approximately the same. Similarly, the difference between the two coefficients in model 4 does not make up for the difference in the average and range of the two. In addition, the rules take up much of the effect of Others' Effortt − 1 anyway. In essence, participants try to deviate but without pushing the boundaries of the acceptable too much, or, in the terminology of delta parameters, participants derive utility from the knowledge of getting further ahead of others once the urge to follow implicit and explicit rules is satisfied. Of course, this result is found using a large sample that might only contain few players who are as two-minded about their behavior as the coefficients make it seem. There is a possibility that there might be a fraction of participants who are more self-regarding and another fraction of participants who are less self-regarding and are therefore conditional cooperators. These two groups could lead to the effect in the regression that both variables are significant but show opposite effects: on the one hand conditional cooperation and on the other a tendency to deviate, which is likely a result that is driven by selfregarding behavior. These results even allow for a ranking of the three rules. All rules improve group results, but formal and informal rules have different impacts (Cardenas, 2004; Ostrom et al., 1994); there are differences in how efficient the rules are, as shown by the different sizes of the coefficients. The larger a rule's coefficient regarding effort decisions is, the higher participants' delta parameters are, or the more certain participants were of the correct estimation of others' delta parameters. Both effects lead to cooperation, the former through obedience, the latter through reciprocation. For a ranking, Weak Enforcement and Strong Enforcement are easily compared: Their coefficients mirror the hierarchy between these two rules along the lines of punishment probability, but only to a certain extent. The punishment probability of these two rules differs by a factor of 3.5 (cheaters are caught with a likelihood of 0.2 under Weak Enforcement and 0.7 under Strong Enforcement). The coefficients differ by not quite the same factor, but the difference is close enough to see the relationship (e.g., in model 2, the factor is 2.8, while in model 4, it is 2.7). Thus, the regressions show that stronger enforcement, i.e., in the present experiments, a higher likelihood of a fine, indeed has a stronger effect, but sub-proportionally so, which is an indication of increased delta parameters originating in stronger rules and possibly also of diminishing effectiveness of stricter enforcement. The less easily comparable Communication Treatment's coefficient is approximately the same size as that for Weak Enforcement. This shows that communication is approximately as efficient as the more easily undermined formal enforcement. This result means that under communication treatment and weak enforcement treatment, making assumptions about the presence, size and sign of delta parameters leads to equally sized and/or uncertain parameter estimates. Placing this insight from the regression analysis in the context of the progression of rule effectiveness under Weak Enforcement and Communication provides a fuller picture. As shown in Fig. 2, Weak Enforcement has a slightly upward trend that plateaus at approximately the baseline value, while communication continuously leads to lower effort levels. This difference can be explained by participants becoming more certain of the properties of others' delta parameters and possibly by the participants feeling more committed due to repeated communication (or at least the possibility for repeated communication). This type of progression does not occur for Weak Enforcement. The interactions in models 9 and 10 show that even under Weak Enforcement, participants have to revert to past behavior as a gauge for others' behavior and consequently as an orientation for their own behavior under Weak Enforcement. Thus, while the regression results of models 2 and 4 attribute approximately the same efficiency to these two methods, Fig. 2, in conjunction with results of models 9 and 10, demonstrates that communication leads groups on a more sustainable path. Communication provides them with a better chance to gauge others' delta parameters instead of reverting to taking their orientation from past actions. Weak Enforcement does not have

29

this property and therefore stops progressing to become more efficient as a rule. 6.2. The Influence of Context Regarding the second research question on the influence of context, models 5 and 6 show that context should not be neglected in commonpool resource games, as the context-relevant variables show that i) awareness of ecological problems, ii) context itself, and iii) real-life experience all make a significant difference, even when the participants participate in an otherwise unaltered experiment. The determining effect of all of these factors clearly concerns what the participants consciously or subconsciously associate with the common-pool resource in question. As the experiment and the significant results for these factors show, this effect is strong enough to lead to different participant behavior compared with a neutral context. This result is yet another example of how experiments, especially field experiments, reveal the bounded rationality of resource users (Cardenas et al., 2004; Ellingsen et al., 2012; Gsottbauer and van den Bergh, 2010). Regarding awareness of ecological problems, the experiments reveal only an influence for awareness of dwindling wildlife stocks. Awareness in the community at the time regarding the much-discussed question of fish stocks seems to be inconsequential. This is possibly due to a change in the real situation shortly before the experiments: Fishery guards were hired by the conservancy, which might have triggered the behavioral response to not have to do anything because someone else is. This belief might have found its way into the participants' decision-making during the experiment. The results for the context show that there appears to be a desire to preserve concrete natural resources relative to the abstract baseline case, as both significant context dummies are associated with less extractive efforts in concrete cases. While it is trivial to attribute the effort-decreasing effect of the sole contexts to the better visualization of and therefore confrontation with the damage done to the environment, it is less obvious why the coefficients differ by approximately 1 effort unit in models 5 and 10. It is possible that this result is an expression of a subconscious regard for the ecosystem in question. If this regard was not expressed in answers to the awareness questions, it might still have found an outlet through the decisions made in that context.9 Another interpretation could be that the social situation of the experiment adds pressure to recognize problems that the participants would not recognize when asked individually (Wang, 1996). Other than the effort-decreasing effect of awareness and context, models 5, 6 and 10 reveal a tendency for real-life overuse to carry over into the experimental results: The interaction between occupation and its appropriate context demonstrates an ingrained behavioral pattern that, along with the other positively significant variables, could explain the irrationally high effort (in excess of SO) under Strong Enforcement, as depicted in Fig. 2. The finding that this factor is driven by ingrained behavior and is dependent solely on the experience of participants and is otherwise independent of the experiment demonstrates the importance of field experiments, settings and subject pools. Models 5 through 8 generally show that spheres of rules and herding behavior on the one hand and socioeconomic factors, cohesion and context factors on the other hand seem to be separate. Unreported robustness checks showed a lack of change in one sphere by the introduction of the other, even beyond results reported in Table 4. The one notable exception is the result for Leadership. Leadership is significant in model 5, most likely because the participants follow the perceived will of the perceived authority figure, but it turns insignificant once the rules are introduced as determinants in model 7. Apparently, opinion leaders 9 It can be ruled out that awareness was only triggered by the context explanation before the experiment, which took place after the participants were asked the awareness questions. The same awareness questions were asked after the experiments, and including them in the regressions leads to similar results.

30

D. Röttgers / Ecological Economics 129 (2016) 21–31

fulfill part of the function of rules when rules are absent. However, rules, rather than Leadership, are more precise determinants of participant behavior. Conversely, in the absence of rules, the will of a leader, even if it is only perceived, influences the behavior of participants. Otherwise, although they do have an impact on behavior, the socioeconomic, cohesion and context variables that are examined in models 5 through 8 do not subtract from the magnitude of the impact of rules and past behavior — only the rules and past behavior change each other's results, save for small changes in the coefficients. As for the independent effect of the remaining two socioeconomic controls, the coefficient for Schooling indicates that better-educated participants behave less altruistically (compare Marwell and Ames, 1981; Selten and Ockenfels, 1998), as do participants with more Children, most likely because a larger number of dependents makes them more desperate or creates a type of social entitlement. 7. Conclusion The results of this paper allow us to comment on the modus operandi of implicit versus explicit rules in CPR experiments. The analysis of the interplay between explicit rules, such as the communication treatment and the punishment conditions, and implicit guidelines, such as the past actions of group members, personal relationships and leadership, leads to novel insights about decision-making. At a descriptive level, the experiment shows that although the communication treatment is not immediately effective, it might be preferable to punishment because cooperation under this treatment is more sustainable. Although the graph in Fig. 2 is inconclusive, it shows a tendency for communication to lead to a pattern of decision-making that is as sustainable as that associated with Strong Enforcement. The regression analysis based on the experimental data and the additional socioeconomic data shows that shared norms (i.e., explicit rules) and past actions (i.e., others' effort levels) indeed have their own separate effects, thus answering the research question on the interplay between shared rules and past action: Participants take what others are doing as a good predictor of behavior – and therefore as a factor in their own decisionmaking – as long as rules are absent. Once rules are present, the effect of the past action of others is diminished in size, although it is still present, and rules also eclipse some otherwise important factors in cohesion, as indicated by the insignificant influence of opinion leaders in the presence of rules. However, not all of the factors in group cohesion are negated, which provides yet more evidence of the importance of communities in CPR management. Other than the clear influence of material incentives and the impact of individual factors, the regressions further demonstrate the importance of a third layer of decision-making: The composition of the group (Cardenas and Ostrom, 2004). In this regard, the regression analysis shows that field experiments deserve a place beside lab experiments precisely because lab experiments cannot simulate the social environment of the field. The regression results provide evidence regarding the importance of this third layer of implicit interaction, such as the effect of personal relationships. Interpersonal interactions such as these stem from the broader community context outside of the experiment and cannot be separated from the participants' decisionmaking within the experiment or be replicated in a lab. The same may be true for pre-established social interactions such as leadership and conditional cooperation with a group of peers. Revealing these effects brings experiment-based policy advice closer to reality, which is the value added of field experiments. The policy implications of these results for a traditionally governed community, such as the case study area of Sikunga, are clear. In Sikunga, as in many other areas under community-based natural resource management, strong enforcement is costly and often underlies other constraints that lie outside the control of community leaders. An alternative low-cost approach might be to strengthen cooperation through the use of traditional authorities as role models, such as tribal

leaders or exemplary villagers, coupled with regular community meetings. The experiments show that this low-cost approach is in certain cases even preferable to costly formal half-measures, as demonstrated by the results for weak enforcement. The second contribution of this paper is equally relevant to community governance: the influence of context. The regression analysis shows that the results depend on framing, partly because established behavioral patterns within a given context matter. This result shows as opposed to lab experiments, results in the field not only depend on general human behavior, but also on the conditions the participants are placed in. Results might therefore show a more realistic picture of human behavior since the situation is more realistic to participants, as already shown in Castillo et al. (2011). This suggests that modeling an experiment to more specific situations, rather than using impartial outside lab subjects or generalizing over different communities, can provide greater insight into the behavior of the people actually affected in that setting. Using well-specified CPR experiments to tailor policy measures based on the results obtained in a specific area may therefore represent a new approach to designing and testing new governance measures. This approach increases the applicability of results to actual policy challenges. It thus possibly increases the relevance of artefactual field experiments not only for researchers but also for policy makers within and outside of the community as well as extension services working with communities. This relevance of experimental results to policy is perhaps best illustrated by the policy recommendations that our team provided at a meeting with the local traditional authorities after gathering our data. While we experienced polite agreement on our advice regarding community measures that can be taken to improve the general resilience of the ecosystem, the participants at the meeting were willing to contribute personal insights and were overall more receptive of our policy advice once we explained the idea using the example of wildlife habitat, i.e., providing context. Once the example was discussed, the traditional authorities and the conservancy management readily accepted our advice on instituting a tree nursery system. On the one hand, this demonstrates how the bare result of the significant context coefficient translates to higher readiness for action. On the other hand, a different result from the experiment could have helped steer policy as well. If we had found in the experiment that the community in general is dismissive of a concrete resource such as wildlife habitat, even though ecologically it is necessary to protect it, we could have provided different advice. For example, we could have urged the authorities to first start an awareness campaign, or we could have appealed to the formal authorities for a different approach, thus circumventing a community which would have demonstrated an unwillingness to save the resource in question. In summary, the analysis shows that i) reliance on past action to orient one's behavior leads to increased cooperation independent of rules, but such reliance is diminished by rules if the rules are effective; ii) factors of community cohesion matter, although some of these factors can be rendered superfluous by effective rules; and iii) the applicability of experiments to actual circumstances can be increased by choosing an appropriate context, if it is relevant. Specific context should be avoided, however, if the experiments are intended to produce generalizable results. Given the nature of this case study and the number of observations, further research of similar or larger scope could confirm and strengthen these results. Appendix A. Supplementary Data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.ecolecon.2016.03.013. References Agrawal, 1999. Greener Pastures: Politics, Markets and Community Among a Migrant Pastoral People. Duke University Press.

D. Röttgers / Ecological Economics 129 (2016) 21–31 Agrawal, A., 2003. Sustainable governance of common-pool resources: context, methods, and politics. Annu. Rev. Anthropol. 32, 243–262. http://dx.doi.org/10.1146/annurev. anthro.32.061002.093112. Anderies, J.M., Janssen, M.A., Bousquet, F., Cardenas, J.C., Castillo, D., Lopez, M.-C., Tobias, R., Vollan, B., Wutich, A., 2011. The challenge of understanding decisions in experimental studies of common pool resource governance. Ecol. Econ. 70, 1571–1579. Axelrod, R., Hamilton, W.D., 1981. The evolution of cooperation. Science 211, 1390–1396. Baddeley, M., 2010. Herding, social influence and economic decision-making: sociopsychological and neuroscientific analyses. Philos. Trans. R. Soc. B Biol. Sci. 365, 281–290. http://dx.doi.org/10.1098/rstb.2009.0169. Bardsley, N., Sausgruber, R., 2005. Conformity and reciprocity in public good provision. J. Econ. Psychol. 26, 664–681. http://dx.doi.org/10.1016/j.joep.2005.02.001. Bowles, S., Gintis, H., 1998. The moral economy of communities: structured populations and the evolution of pro-social norms. Evol. Hum. Behav. 19, 3–25. http://dx.doi. org/10.1016/S1090-5138(98)00015-4. Bowles, S., Gintis, H., 2000. Optimal Parochialism: The Dynamics of Trust and Exclusion in Networks (Working Paper) University of Massachusetts, Department of Economics. Bwalya, S.M., 2007. Dissertation abstract: the experimental analysis of the political economics of fisheries governance. Exp. Econ. 10, 181–182. http://dx.doi.org/10.1007/ s10683-006-9141-1. Cardenas, J.C., 2004. Norms from outside and from inside: an experimental analysis on the governance of local ecosystems. Forest Policy Econ. 6, 229–241. Cardenas, J.C., Ostrom, E., 2004. What do people bring into the game? Experiments in the field about cooperation in the commons. Agric. Syst. 82, 307–326. Cardenas, J.C., Ahn, T.K., Ostrom, E., 2004. Communication and co-operation in a commonpool resource dilemma: a field experiment. Adv. Underst. Strateg. Behav. Game Theory Exp. Bounded Ration 258–286. Castillo, D., Bousquet, F., Janssen, M.A., Worrapimphong, K., Cardenas, J.C., 2011. Context matters to explain field experiments: results from Colombian and Thai fishing villages. Ecol. Econ. 70, 1609–1620. http://dx.doi.org/10.1016/j.ecolecon.2011.05.011. Cavalcanti, C., Schläpfer, F., Schmid, B., 2010. Public participation and willingness to cooperate in common-pool resource management: a field experiment with fishing communities in Brazil. Ecol. Econ. 69, 613–622. http://dx.doi.org/10.1016/j.ecolecon. 2009.09.009. Davis, D.D., 1993. Experimental Economics. Princeton University Press. Davis, A., 2009. Namibia's Communal Conservancies: A Review of Progress and Challenges in 2009. NACSO (Namibian Association of CBNRM Support Organizations). Dayton-Johnson, J., 2003. Knitted warmth: the simple analytics of social cohesion. J. SocioEcon. 32, 623–645. http://dx.doi.org/10.1016/j.socec.2003.10.002. Drehmann, M., Oechssler, J., Roider, A., 2005. Herding and contrarian behavior in financial markets: an internet experiment. Am. Econ. Rev. 95, 1403–1426. Ellingsen, T., Johannesson, M., Mollerstrom, J., Munkhammar, S., 2012. Social framing effects: preferences or beliefs? Games Econ. Behav. 76, 117–130. http://dx.doi.org/10. 1016/j.geb.2012.05.007. Evans, C.R., Dion, K.L., 1991. Group cohesion and performance a meta-analysis. Small Group Res. 22, 175–186. http://dx.doi.org/10.1177/1046496491222002. Fehr, E., Leibbrandt, A., 2008. Cooperativeness and Impatience in the Tragedy of the Commons. Institute for Empirical Research in Economics - University of Zurich. Fischbacher, U., Gächter, S., 2010. Social preferences, beliefs, and the dynamics of free riding in public goods experiments. Am. Econ. Rev. 100, 541–556. Gillet, J., Schram, A., Sonnemans, J., 2009. The tragedy of the commons revisited: the importance of group decision-making. J. Public Econ. 93, 785–797. http://dx.doi.org/10. 1016/j.jpubeco.2009.02.001. Goldsmith, R.E., Desborde, R., 1991. A validity study of a measure of opinion leadership. J. Bus. Res. 22, 11–19. Gsottbauer, E., van den Bergh, J.C.J.M., 2010. Environmental policy theory given bounded rationality and other-regarding preferences. Environ. Resour. Econ. 49, 263–304. http://dx.doi.org/10.1007/s10640-010-9433-y. Hardin, G., 1968. The tragedy of the commons. Science 1243–1248. Hayo, B., Vollan, B., 2011. Group interaction, heterogeneity, rules, and co-operative behaviour: evidence from a common-pool resource experiment, in South Africa and Namibia. J. Econ. Behav, Organ. Heider, L., 2012. Fish Never Finishes Versus Shifting Baseline Syndrome (Master Thesis) Wageningen University, Wageningen. Jacoby, J., 1974. The construct validity of opinion leadership. Public Opin. Q. 38, 81–89. http://dx.doi.org/10.1086/268136.

31

Janssen, M.A., Bousquet, F., Cardenas, J.C., Castillo, D., Worrapimphong, K., 2013. Breaking the elected rules in a field experiment on forestry resources. Ecol. Econ. 90, 132–139. http://dx.doi.org/10.1016/j.ecolecon.2013.03.012. Kleinman, P.J.A., Pimentel, D., Bryant, R.B., 1995. The ecological sustainability of slash-andburn agriculture. Agric. Ecosyst. Environ. 52, 235–249. http://dx.doi.org/10.1016/ 0167-8809(94)00531-I. Lankford, B., Sokile, C., Yawson, D., Lévite, H., 2004. The River Basin Game: A Water Dialogue Tool. Iwmi. Marwell, G., Ames, R.E., 1981. Economists free ride, does anyone else? Experiments on the provision of public goods, IV. J. Public Econ. 15, 295–310. http://dx.doi.org/10.1016/ 0047-2727(81)90013-X. Miller, J.H., Page, S.E., 2004. The standing ovation problem. Complexity 9, 8–16. Moxnes, E., van der Heijden, E., 2003. The effect of leadership in a public bad experiment. J. Confl. Resolut. 47, 773–795. http://dx.doi.org/10.1177/0022002703258962. Nordi, N., 2006. Common property resource system in a fishery of the São Francisco River, Minas Gerais. Brazil. Hum. Ecol. Rev. 13, 1. Ostrom, E., 1990. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, Cambridge. Ostrom, E., 1998. A behavioral approach to the rational choice theory of collective action: presidential address, American Political Science Association, 1997. Am. Polit. Sci. Rev. 1–22. Ostrom, E., 2006. The value-added of laboratory experiments for the study of institutions and common-pool resources. J. Econ. Behav. Organ. 61, 149–163. Ostrom, E., 2009. Understanding Institutional Diversity. Princeton university press. Ostrom, E., Crawford, S., 1995. A grammar of institutions. Am. Polit. Sci. Rev. 89, 582–600. Ostrom, E., Walker, J., Gardner, R., 1992. Covenants with and without a sword: selfgovernance is possible. Am. Polit. Sci. Rev. 86, 404–417. http://dx.doi.org/10.2307/ 1964229. Ostrom, E., Gardner, R., Walker, J., 1994. Rules, Games, and Common-Pool Resources. Univ of Michigan Pr. Rogers, E.M., Cartano, D.G., 1962. Living research methods of measuring opinion leadership. Public Opin. Q. 26, 435–441. http://dx.doi.org/10.1086/267118. Schnier, K.E., 2009. Spatial externalities and the common-pool resource mechanism. J. Econ. Behav. Organ. 70, 402–415. http://dx.doi.org/10.1016/j.jebo.2009.02.004. Selten, R., Ockenfels, A., 1998. An experimental solidarity game. J. Econ. Behav. Organ. 34, 517–539. http://dx.doi.org/10.1016/S0167-2681(97)00107-8. Smith, V.L., 2010. Theory and experiment: what are the questions? J. Econ. Behav. Organ. 73, 3–15. http://dx.doi.org/10.1016/j.jebo.2009.02.008. Sturm, B., Weimann, J., 2006. Experiments in environmental economics and some close relatives. J. Econ. Surv. 20, 419–457. Tavoni, A., Schlüter, M., Levin, S., 2012. The survival of the conformist: social pressure and renewable resource management. J. Theor. Biol. 299, 152–161. http://dx.doi.org/10. 1016/j.jtbi.2011.07.003. Travers, H., Clements, T., Keane, A., Milner-Gulland, E.J., 2011. Incentives for cooperation: the effects of institutional controls on common pool resource extraction in Cambodia. Ecol. Econ. van Dijk, E., Wilke, H., Wit, A., 2003. Preferences for leadership in social dilemmas: public good dilemmas versus common resource dilemmas. J. Exp. Soc. Psychol. 39, 170–176. http://dx.doi.org/10.1016/S0022-1031(02)00518-8. Vedeld, T., 2000. Village politics: heterogeneity, leadership and collective action. J. Dev. Stud. 36, 105–134. http://dx.doi.org/10.1080/00220380008422648. Velez, M.A., Stranlund, J.K., Murphy, J.J., 2009. What motivates common pool resource users? Experimental evidence from the field. J. Econ. Behav. Organ. 70, 485–497. http://dx.doi.org/10.1016/j.jebo.2008.02.008. Velez, M.A., Murphy, J.J., Stranlund, J.K., 2010. Centralized and decentralized management of local common pool resources in the developing world: experimental evidence from fishing communities in Colombia. Econ. Inq. 48, 254–265. Vollan, B., 2008. Socio-ecological explanations for crowding-out effects from economic field experiments in southern Africa. Ecol. Econ. 67, 560–573. http://dx.doi.org/10. 1016/j.ecolecon.2008.01.015. Wang, X.T., 1996. Domain-specific rationality in human choices: violations of utility axioms and social contexts. Cognition 60, 31–63. http://dx.doi.org/10.1016/00100277(95)00700-8.