INTRODUCTION
Can a nation ever find itself regretting discovering some of its natural resources? Many have posed this question after witnessing the hardship in resource-rich countries such as Libya or Iraq. Sachs and Warner (Reference Sachs and Warner1995) and a large literature have presented considerable evidence that natural resources can indeed be a curse for a variety of reasons, including diminished income growth, lower income, or adversely affected institutions (Frankel Reference Frankel2010; Ross Reference Ross2001).
The focus of our study is the recently proposed, refined version of the resource curse, whereby the negative consequences of natural resources are restricted to economies with weak institutions (Acemoglu and Robinson Reference Acemoglu and Robinson2006; Acemoglu et al. Reference Acemoglu, Verdierand and Robinson2004; Al-Ubaydli Reference Al-Ubaydli2011; Bulte and Damania Reference Bulte and Damania2008; Hodler Reference Hodler2006; Mehlum et al. Reference Mehlum, Moene and Torvik2006; Olsson Reference Olsson2006; Reference Olsson2007; Robinson et al. Reference Robinson, Torvikand and Verdier2005; Torvik Reference Torvik2002; Vicente Reference Vicente2010). We test a modified version of the Torvik (Reference Torvik2002) model in a laboratory virtual world experiment. Players allocate their labor between production, which is positive-sum, and rent-seeking in the natural resource sector, which is zero-sum. If players play the symmetric Nash equilibrium, then resource booms attract labor away from the productive sector toward rent-seeking to such a degree that the aggregate income decreases—the resource curse. However, if players can establish institutions that promote cooperation (North Reference North1990; North et al. Reference North, Wallis and Weingast2009), they can realize the benefits of a resource boom.
We vary two factors: natural resource income and the players’ ability to communicate and monitor one another—the latter being a key determinant of the quality of institutions. We find strong support for the prediction that the resource curse arises only in economies with weak institutions. In general, we find evidence of a mild version of the resource curse, whereby groups with weak institutions tend to squander all the benefits of a resource boom and may even be harmed by it.
Previous empirical studies have failed to obtain exogenous variation in both of these explanatory variables, instead relying on potentially endogenous, naturally occurring variation (studies such as Tsui Reference Tsui2009, and Acemoglu et al. Reference Acemoglu, Johnson and Robinson2001, rely on plausibly exogenous variation but do not use it to study economic performance; also see Compton et al. Reference Compton, Giedeman and Johnson2010). In fact, the empirical debate over how to best handle endogeneity has led to Alexeev and Conrad (Reference Alexeev and Conrad2009) questioning whether the aggregate data even support the existence of a resource curse.
This article has two main contributions. First, by using a laboratory experiment, we can implement randomized control in the explanatory variables of interest, guaranteeing exogenous variation (see Leibbrandt and Lynham Reference Leibbrandt and Lynham2013, for a complementary experimental investigation). Second, by allowing players to interact in a visceral environment (the Second Life™ virtual world), we can more fully explore the nuances of strong versus weak institutions. By providing clean evidence of the refined, institutions-mediated model of the resource curse, our article represents a significant step in our understanding of the economic impact of natural resources.
MODEL
The Torvik (Reference Torvik2002) model is well suited for theoretically exploring the effect of natural resources on societal welfare. However, in its original form, the large number of players and the environment's complexity render it ill suited for a laboratory experiment. We modify the model, drawing heavily upon Morgan and Sefton (Reference Morgan and Sefton2000) and the rent-dissipation literature (Baye et al. Reference Baye, Kovenock and De Vries1994; Knapp and Murphy Reference Knapp and Murphy2010; Nti Reference Nti1997; Potters et al. Reference Potters, de Vries and van Winden1998; Tullock Reference Tullock, Buchanan, Tollison and Tullock1980; Walker et al. Reference Walker, Gardner and Ostrom1990). More generally, we are only testing one of several distinct causal mechanisms proposed by resource curse scholars. We restrict our attention to the institutions-mediated version of the resource curse to sharpen our focus and because of its suitability for laboratory experimentation.
The economy is populated by identical players, indexed by i ∈ {1, 2, . . ., n}, interacting for one period. Each has one unit of time to allocate between rent-seeking pursuit of the natural resource (xi
) and production (1 − xi
). The economy-wide income from the natural resource is R > 0, and i's share is equal to her share of the aggregate rent-seeking effort
$\mathop \sum \limits _j x_j$
. (If all choose xi
= 0, then each gets (1/n)R.) This implies a negative externality to allocating resources to the natural resource sector: i can only increase her share by decreasing that of others.
Production has a positive externality: For each unit of time that i allocates to production, she receives α > 0 and each player j ≠ i receives β > 0. Thus i's payoff is
\begin{equation*}
y_i = \alpha \left( {1 - x_i } \right) + \beta \mathop \sum \limits _{j \ne i} \left( {1 - x_j } \right) + R\frac{{x_i }}{{\mathop \sum \nolimits _j x_j }}.
\end{equation*}
Analogous results can be obtained by eliminating the positive externality in production and having a stronger negative externality in the natural resource sector, or by having increasing returns in the productive sector, similar to Torvik. (See the online Appendix for more on this.)
Why should we expect strong negative externalities in the natural resource sector? Economists have long regarded a substantial proportion of natural resource income to be economic rent (Sachs and Warner Reference Sachs and Warner1995), which invites zero-sum rent-seeking behavior (Nti Reference Nti1997; Tullock Reference Tullock, Buchanan, Tollison and Tullock1980). The rent-seeking can become wasteful socially as the groups in control of the rent erect barriers to secure their position (Auty Reference Auty2001), and the conflict can become violent (Frankel Reference Frankel2010). Our precise specification is motivated by a desire to keep payoffs as simple as possible (for the purposes of running an experiment), subject to retaining the spirit of Torvik (Reference Torvik2002).
Equilibrium
Let
$Y = \mathop \sum \limits _i y_i$
denote the gross domestic product (GDP). Each player chooses xi
∈ [0, 1] to maximize yi
, given (x
1, . . ., x
i − 1, x
i + 1, . . ., xn
). The maximand is smooth and strictly concave; assuming an interior solution, the unique symmetric Nash equilibrium is (see the online Appendix for all proofs)
\begin{equation*}
x^{NE} = \frac{{\left( {n - 1} \right)R}}{{\alpha n^2 }},Y^{NE} = \frac{1}{{\alpha n}}\left( {\left( {\alpha - \beta \left( {n - 1} \right)^2 } \right)R + \alpha ^2 n^2 + \alpha \beta \left( {n - 1} \right)n^2 } \right).
\end{equation*}
In contrast, the symmetric Pareto efficient outcome is
\begin{equation*}
x^{PE} = 0,Y^{PE} = \left( {\alpha + \beta \left( {n - 1} \right)} \right)n + R > Y^{NE} .
\end{equation*}
We refer to players who play the Nash equilibrium as non-cooperative players, players who play the Pareto efficient equilibrium as cooperative players, and an increase in R as a resource boom.
Proposition 1: In response to a resource boom
-
(a) non-cooperative players increase rent-seeking, ∂xNE /∂R > 0;
-
(b) cooperative players do not alter their rent-seeking, ∂xPE /∂R = 0.
Resource booms increase the marginal private return to rent-seeking, but the marginal social return to rent-seeking is always zero, since rent-seeking is purely redistributive. The marginal private and social costs to rent-seeking are both positive and unaffected by R.
Proposition 2: In response to a resource boom
-
(a) for a sufficiently large β, GDP in a non-cooperative economy will fall, ∂YNE /∂R < 0;
-
(b) GDP in a cooperative economy always increases one-on-one, ∂YPE /∂R = 1 > 0.
The marginal social cost to rent-seeking, α + (n − 1)β, exceeds the marginal private cost, α. When β is sufficiently large, the increased rent-seeking brought about by a resource boom actually decreases GDP in a non-cooperative economy, as too many resources are shifted from the positive-sum production sector to the zero-sum natural resource sector. In contrast, a cooperative economy reaps the full benefits of a resource boom.
Our goal is to investigate the possibility that an economy regrets a resource boom. The above version of the resource curse uses levels of GDP rather than, say, GDP's growth rate. We chose levels to simplify the theoretical exposition and to facilitate laboratory testing; however the thrust of the argument in Torvik (Reference Torvik2002)—and hence our argument—does not depend upon any particular outcome variable.
If this game is repeated finitely with period-by-period payoff information, by backward induction, the symmetric Nash equilibrium, efficient play, and all the predictions are retained. The symmetric Nash equilibrium remains an equilibrium even if the game is repeated infinitely.
Institutions
The model's predictions hinge upon whether the players play cooperatively versus non-cooperatively. Torvik (Reference Torvik2002), Hodler (Reference Hodler2006), Mehlum et al. (Reference Mehlum, Moene and Torvik2006), Olsson (Reference Olsson2006, Reference Olsson2007), Bulte and Damania (Reference Bulte and Damania2008), and Al-Ubaydli (Reference Al-Ubaydli2011) argue that this is determined by the quality of institutions (North Reference North1990). Other models that predict an ambiguous effect of natural resources are Acemoglu et al. (Reference Acemoglu, Verdierand and Robinson2004), Robinson et al. (Reference Robinson, Torvikand and Verdier2005), Acemoglu and Robinson (Reference Acemoglu and Robinson2006) and Andersen and Aslaksen (Reference Andersen and Aslaksen2008), but they explore different mechanisms.
Players recognize that rent-seeking is inefficient, and they want to establish formal and informal rules that eliminate it. The rent-seeking effort can be extrajudicial violence or (corrupt) lobbying. In both cases, the government can play a key role in curtailing the rent-seeking by, for example, enforcing property rights, operating transparently, establishing rules that minimize corrupt practices and generally managing conflict in an effective manner, all of which equate to high-quality institutions (North et al. Reference North, Wallis and Weingast2009).
How do good institutions come about? We will not tackle this sizeable topic comprehensively in this article (Acemoglu and Robinson Reference Acemoglu and Robinson2006; Greif Reference Greif2006). For the purposes of a laboratory experiment, a good departure point is Ostrom (Reference Ostrom2000), which emphasizes cooperation and coordination. She concisely summarizes her findings in the form of several design principles:
-
• Clear boundary rules (Kimbrough et al. Reference Kimbrough, Smith and Wilson2008; McCabe et al. Reference McCabe, Twieg and Weel2011)
-
• Clear usage rules
-
• Participation by the individuals affected by the rules in rule-making
-
• Selecting and holding the monitors accountable by the stakeholders
-
• Graduated sanctions for rule-violation
-
• Access to conflict-resolution arenas
A common thread is the need for communication channels and monitoring mechanisms. Ostrom's fieldwork is complemented by a large experimental literature on the benefits of communication to cooperation and coordination (Blume and Ortmann Reference Blume and Ortmann2007; Brandts and Cooper Reference Brandts and Cooper2005; Burton and Sefton Reference Burton and Sefton2001; Charness and Grosskopf Reference Charness and Grosskopf2004; Cooper et al. Reference Cooper, DeJong, Forsythe and Ross1992; Duffy and Feltovich Reference Duffy and Feltovich2002).
For the purposes of the present model, we can think of good institutions as being those that increase the likelihood of the players playing cooperatively. Good institutions include a clear allocation of property rights, which decreases the usefulness of violence and lobbying as a means of securing a larger share of the pie. They also include a strong sense of community and mutual affection among stakeholders, which limits selfish rent-seeking. Beyond this, they include punishment mechanisms that can further deter rent-seeking. In the online Appendix, we formally integrate institutions into our model and explain how communication channels can bring about good institutions. For now, in the interests of brevity, we focus on the result.
Proposition 3: In repeated play, cooperative play is more likely the greater the opportunity to communicate and monitor.
EXPERIMENTAL DESIGN
Procedure
Each session has 12 players and proceeds in the following manner:
-
• Virtual world training
-
• Round 1: Play the main game in groups of 4 for 15 minutes with natural resource income R = r 1.
-
• Round 2: Reassign to new groups of 4 and play the main game for 15 minutes. For the first
$7^{^{1}\hspace*{-1.0pt}/\hspace*{-1.0pt}_{2}}$
minutes, R = r
1; for the next
$7^{^{1}\hspace*{-1.0pt}/\hspace*{-1.0pt}_{2}}$
minutes, R = r
2 ≠ r
1. -
• Round 3: Reassign to new groups of 4 and play the main game for 15 minutes with R = r 2.
To minimize noise, we made the experiment long enough to allow players to explore their environment and experiment with strategies. Players were anonymously assigned unique in-game names that they retained throughout the session. We reassigned groups to minimize the reputational transfer across rounds.
Parameters
Players played a continuous time version of the game in our model under a neutral frame (e.g., “Activity X” instead of “production”) and with the following parameter values:
-
• Group size n = 4
-
• Private production return α = $800/minute
-
• External production return β = $500/minute
-
• Natural resource income: high rH = $4, 000/minute and low rL = $1, 200/minute
-
• The support of x transformed to integers in the range [0, 100]
\begin{eqnarray*}
x_L^{NE} &=& 28,\quad \quad \frac{1}{n}Y_L^{NE} = \$1,953/\min ,\quad \quad \frac{1}{n}Y_L^{PE} = \$2,600/\min \\
x_H^{NE} &=& 94,\quad \quad \frac{1}{n}Y_H^{NE} = \$1,144/{\rm min},\quad \quad \frac{1}{n}Y_H^{PE} = \$3,300/{\rm min}
\end{eqnarray*}
Under non-cooperative play, the resource curse will materialize (YNE L > YH NE ). Under cooperative play, GDP per capita should increase by $700/min (27%) when R = rH . In each session, we randomly assigned r 1 = rH , r 2 = rL or the reverse. At the start, players learn r 1, and they are informed that it may change, with any change being preceded by an announcement.
Information
Players receive continuous information about their payoff broken down into the amount they are earning from their production, the amount they are earning from others’ production, and the amount they are earning from rent-seeking. This allows players to identify when others are rent-seeking without knowing the rent-seeker's identity. To facilitate payoff calculations, players are continuously told how their payoff will change if they allocate one more/less unit to rent-seeking. In principle, making this information so saliently available could induce players to change their rent-seeking with higher frequency than would be rational (especially in the no-comms treatment below, where there is less for players to do during a session).
We were unconcerned by this prospect because optimal rent-seeking was dependent on the choices of other players, which were themselves changing in real time, and so technically, it was optimal for players to regularly change their rent-seeking. Given the ample time afforded for experimentation, subjects will plausibly have inferred the importance of continually fine-tuning their rent- seeking.
Treatments
The first explanatory variable is natural resource income. All sessions experience within variation as it is shocked up or down halfway through round 2. We also have between session variation, with half the sessions starting with r 1 = rH = $4, 000 and the other half starting with r 1 = rL = $1200.
The second explanatory variable is the opportunity to communicate and monitor. Our baseline treatment is no comms, where players have no way of communicating. This corresponds to the poorest institutions and is designed to induce non-cooperative play.
Our second treatment is partial comms. Players can communicate with fellow group members via a live chat interface for 2 minutes prior to the start of each round. It is common knowledge that once the round actually starts, the players will revert to the conditions in the baseline. This gives players an opportunity to establish some informal norms and to discuss optimal strategies. This treatment corresponds to (the opportunity for) an intermediate quality of institutions and is more likely to result in cooperative play than the no-comms treatment.
Our third treatment is full comms. In addition to the pre-round communication of the partial-comms treatment, players can chat throughout the 15 minutes of the round. Each player also has the ability to intermittently monitor the rent-seeking activities of two out of her three fellow group members. This gives players much more information on the distribution of rent-seeking, allows them to administer targeted verbal sanctions, helps create a sense of community and mutual affection, and provides a forum for discussing the best response to the natural resource shock in the middle of round 2. This corresponds to (the opportunity for) high-quality institutions and the best chance of cooperative play.
We chose not to run the additional treatments necessary to parse the role of communication vis-à-vis monitoring because it was not relevant to our main goal: demonstrating that under sufficiently poor institutions, the resource curse appears, and that under sufficiently good institutions, the resource curse is absent. The precise configuration of institutions at which the resource curse pendulum swings is a topic for future research.
Virtual world environment
Each session was conducted using the virtual world of Second Life™ (Atlas and Putterman Reference Atlas and Putterman2011; Chesney et al. Reference Chesney, Chuah and Hoffmann2009; Fiedler et al. Reference Fielder, Haruvy and Li2011). Second Life™ places individual users into the role of controlling a graphical avatar and navigating and interacting with a three-dimensional spatial environment (Figure 1).
This picture shows a subject's avatar, two monitoring beacons (the pillars) which can give information on rent-seeking, and the experimental HUD which displays information and allows for investments in rent-seeking to be made during the session. The monitoring beacon closest to the avatar lies in the player's area, and the other player's area lies behind a semi-transparent barrier (the horizontal lines). Thus neighboring players can see each other and can see when they are within range of the monitoring beacons (and hence are likely to be monitoring).
Second Life™ was employed in this study in order to naturalistically implement many of the features outlined in Ostrom (Reference Ostrom2000). It is possible to study how communication and monitoring interact with the institutions-mediated resource curse in an environment devoid of the visceral, visual elements inherent in Second Life™, for example, using zTree (Fischbacher Reference Fischbacher2007), though to the best of our knowledge, no such experiment has been done.
We regarded the visual aspects of Second Life™ as desirable because they reduce the social distance between subjects relative to traditional laboratory designs (Bente et al. Reference Bente, Rüggenberg, Krämer and Eschenburg2008; Fielder et al. Reference Fielder, Haruvy and Li2011; Kang et al. Reference Kang, Watt and Ala2008; Nowak et al. Reference Nowak, Watt and Walther2005). Low levels of social distance help establish the cooperative institutions that can overcome natural resource management problems (Charness and Gneezy Reference Charness and Gneezy2008; Charness et al. Reference Charness, Haruvy and Sonsino2007).
The spatial layout of the experimental environment is depicted in Figure 2. Each group interacted within a separate copy of this environment. It is divided into a central area, in which subject training occurs prior to the first experimental round, and four outer areas, each inhabited by one subject. Each subject is able to view its two adjacent neighbors and can check on these neighbors’ rent-seeking at various times, depending on the treatment. Only one neighbor can be effectively monitored at a time, however, as subjects must travel from one side of their area to the other in order to monitor their other neighbor (this takes approximately 15 seconds).
Subjects are trained in the inner area and then travel to one of the four outer areas prior to the session's commencement. The outer areas are each one corner of the outer square, and the white transparent barriers can be seen emerging from the outer corners of the inner diamond. Each transparent barrier is straddled by a pair of monitoring beacons (see the ones at the bottom of the figure; the others are invisible from this range because of rendering limitations). During play, players are free to move around their area and see their two immediate neighbors, but they do not interact visually with the neighbor from the opposing corner.
This monitoring process is made especially salient via the fact that it not only involves the conveyance of information about rent-seeking behavior but also visual contact between subject avatars; previous studies have demonstrated the potential significance of visceral, visual cues of monitoring to cooperative decision-making (Bateson et al. Reference Bateson, Nettle and Roberts2006). Furthermore, the ability to monitor only one neighbor at a time allowed for well-organized groups to devise collective monitoring strategies—in one group in particular it was suggested that each subject constantly monitors his or her left-hand neighbor to ensure that each subject was constantly monitoring (and being monitored by) another subject.
Even in sessions where the monitoring and communication options of subjects were limited, subjects would extensively move around and explore their respective areas. Furthermore, this movement ability afforded subjects a rudimentary capability for communication insofar as movement patterns could be used to communicate information to neighbors. For example, it was common for subjects to approach each other and mimic each other's movement patterns, perhaps as a sort of pro-social display (Lakin et al. Reference Lakin, Jefferis, Cheng and Chartrand2003). After one round in the partial-comms treatment, one subject commented, “i [sic] am pretty sure i [physically interacted]Footnote 1 with a fellow avatar last round through a transparent fence.”
Rent-seeking is selected using the heads-up display (HUD) depicted in the top-left corner of Figure 1. The HUD also displays several additional pieces of information described above (absent in Figure 1 because the screenshot is from training).
Research Hypotheses
Our data permit us to examine the effect of a resource boom and a resource bust. For brevity, we formulate our hypotheses only in terms of booms.
Prediction 1: In response to a resource boom
-
(a) under no comms, rent-seeking will increase;
-
(b) under full comms, rent-seeking will be unchanged;
-
(c) the increase in rent-seeking will be larger in no comms vs. partial comms and in partial comms vs. full comms.
Prediction 2: In response to a resource boom
-
(a) under no comms, GDP will decrease;
-
(b) under full comms, GDP will increase one-on-one;
-
(c) the increase in GDP will be larger in full comms vs. partial comms and in partial comms vs. no comms.
Studies in the rent dissipation literature have presented data collected from similar environments. Our study differs in that we examine the comparative statics of the size of the prize, and we consider the relationship between rent seeking and social welfare, and how it is mediated by institutions.
RESULTS
We ran 12 sessions in 2011. The sessions took place in the George Mason University Krasnow Institute laboratory. The sessions took 2 hours, and the average earning was $25.
We collected data every 10 seconds. Table 1 contains the main descriptive statistics where the averages are across players and time. Figure 3 gives us a sense of the overall dynamics of rent-seeking averaging across players in each treatment.
The top (bottom) panel is for sessions where natural resource income started at the low (high) value. Each data series is based on averaging across 24 players (across 6 groups). Period denotes 10 second intervals, so that each 15 minute round has 90 periods. The middle of the second round sees the unannounced shock to natural resource income.
Sample Means and Standard Deviations (in Parentheses) for Rent-Seeking and Earnings
Note: Rent-seeking and Earnings are calculated by averaging across players and time for the three time periods that correspond to each row. Nash equilibrium rent-seeking is 28 under low natural resource income and 94 under high natural resource income, whereas efficient rent-seeking is always 0. Nash equilibrium earnings are $1953/min under low natural resource income and $1144/min under high natural resource income, while the respective figures under efficient play are $2600/min and $3300/min.
Rent-seeking is low at the start of a round because its default value is 0. Each panel of Figure 3 supports our main hypotheses based on within variation: In the absence of communication (solid line), when natural resource income is higher, so too is rent-seeking; when communication is permitted (dotted line), rent-seeking is low and unresponsive to natural resource income. Comparing the two panels of Figure 3 also supports our main hypotheses based on between variation. However the panels of Figure 3 are based on averages, and they conceal variation around the mean, to which we turn our attention now.
Halfway through round 2, natural resource income R is shocked from its initial value. For each group, we can calculate GDP per capita when R = $4, 000 and compare it to GDP per capita when R = $1, 200. In Figures 4(a)–4(b), we plot this difference by treatment for each of the 12 groups that experience this shock in each treatment.
For example, according to Figure 4b, in the partial-comms treatment, 8 out of 12 groups had a higher GDP per capita when R = $4, 000 than when R = $1, 200. In all treatments, if rent-seeking does not change in response to a resource boom, then GDP per capita should increase by $700. Some groups experience a greater increase than $700 because they decrease their rent-seeking in response to a resource boom. We will further discuss Table 1 and Figures 3 and 4 later in our formal results.
Data Structure and Empirical Strategy
We construct six observations per player per session by averaging each player's 45 minutes of data to two observations per player per round: one for the first
$7^{^{1}\hspace*{-1.0pt}/\hspace*{-1.0pt}_{2}}$
minutes and one for the last
$7^{^{1}\hspace*{-1.0pt}/\hspace*{-1.0pt}_{2}}$
minutes. Let t ∈ {1, 2, . . ., 6} denote the period; see Table 2. (See the online Appendix for a fuller discussion, including the alternatives.)
The Definition of Time Periods
Our main dependent variables are rent-seeking, xit , and earnings, yit , where i denotes player across all sessions, that is, i ∈ {1, 2, . . ., 144}; Fj i is a player fixed effect that takes the value 1 when i = j and 0 otherwise; Ts t is a time fixed effect that takes the value 1 when t = s and 0 otherwise; g(it) is a function denoting which of the 108 groups i finds herself in period t.
Let Hit be a dummy variable that takes the value 1 when R = $4, 000 and 0 when R = $1, 200. Let D Partial it be a treatment dummy for partial comms, and let D Full it be a treatment dummy for full comms. We estimate the following econometric models:
\begin{eqnarray*}
x_{it} &=& a + b^H H_{it} + b_{{\rm Int}}^{{\rm Partial}} H_{it} D_{it}^{{\rm Partial}} + b_{{\rm Int}}^{{\rm Full}} H_{it} D_{it}^{{\rm Full}}\\
&& +\, \mathop \sum \limits _{j = 2}^{144} b^{j}_{F} F_i^j + \mathop \sum \limits _{s = 2}^6 b_T^s T_t^s + u_{g\left( {it} \right)} + \varepsilon _{it} ,\\
y_{it} & =& a + b^H H_{it} + b_{{\rm Int}}^{{\rm Partial}} H_{it} D_{it}^{{\rm Partial}} + b_{{\rm Int}}^{{\rm Full}} H_{it} D_{it}^{{\rm Full}}\\
&& +\, \mathop \sum \limits _{j = 2}^{144} b_F^j F_i^j + \mathop \sum \limits _{s = 2}^6 b_T^s T_t^s + u_{g\left( {it} \right)} + \varepsilon _{it} ,
\end{eqnarray*}
where ε it is a white noise error term and u g(it) is a group cluster to capture correlation between decisions within a group. The player fixed effect is included to correct for within-player correlation. We exclude communication treatment variables because we include player fixed effects.
Main Results
Result 1a: Under no comms, rent-seeking increases in response to a resource boom.
Table 1 and Model 1 in Table 3 (which contains the regressions) confirm that under no comms, per capita rent-seeking increases by 35 (p < .01), which is economically significant (larger than a standard deviation).
Regression Models
Note: Asterisks denote significance: * = 10%, ** = 5%, *** = 1%. All regressions include player fixed effects, time fixed effects, and clustering by group.
Result 1b: Under full comms, rent-seeking slightly increases in response to a resource boom.
Table 2 and Model 1 in Table 3 confirm that under full comms, per capita rent-seeking increases by 5.1 (p < .1 using a Wald test), which is economically quite small (less than half a standard deviation). (In Model 1 of Table 3, the treatment effect is obtained by summing the coefficient on ‘High natural resource income’ with the coefficient on ‘High natural resource income × Full comms’.) Both pre- resource boom and post-resource boom, rent-seeking on average does not exceed 10 (out of a possible 100); overall, rent-seeking is very low under full comms.
The chat transcripts reflect the players’ success in limiting rent-seeking (we hired research assistants who were blind to the experiment's goal and its treatments to classify chat statements; to ensure accurate classification, we used multiple research assistants for each of the approximately 4,500 messages exchanged in the experiments). In the two minutes prior to the start of each round, over 6% of the 640 messages exchanged by players were explicit attempts at coordinating on zero rent-seeking (and many of the remaining messages were expressions of affirmation, and so the figure 6% is a substantial understatement); for example, “Lets [sic] go all X”, “everyone press 0 in y,” and “Do what's best for everyone and dont [sic] increase Y.” (In the experiment, production was labeled ‘Activity X’ and rent-seeking was labeled ‘Activity Y’.)
In line with Ostrom (Reference Ostrom2000), successful avoidance of rent-seeking was down to a combination of overt coordination, monitoring, and sanctions. In the 36 groups in the full-comms sessions, monitoring stations were used 1099 times (2.5 times per player per round). There were 18 occasions where one player explicitly and correctly accused another player of selfishly rent-seeking. If this number seems low, it is because most accusations were not directed at individual players; more often they were of the form “someone is putting tokens in Y!!!” or “stop cheating!!!” The (correctly) accused's subsequent level of rent-seeking almost always went down in response to the accusation (p < 2% level, using a Wilcoxon paired-values test). Accusations were typically aggressive in tone and constituted verbal sanctions; for example, “QUIT BEING GREEDY!!!!!!!!”
Though harder to quantify, chat transcripts clearly reflected some groups’ success in forging a sense of community and shared goals, which presumably helped limit rent-seeking.
Result 1c: The increase in rent-seeking in response to a resource boom is larger in no comms versus partial comms and in partial comms versus full comms.
Table 1 and Model 1 in Table 3 confirm that under partial comms, per capita rent-seeking increases by 19.3, which is below the estimated increase under no comms and larger than that under full comms. Using Wald tests, we can reject the hypotheses that it is equal to either (p < .01).
Similar to the full-comms sessions, players in the partial-comms sessions took advantage of the (limited) opportunity to chat: 8% of the 577 messages constituted an explicit attempt to coordinate on zero rent-seeking.
Result 2a: Under no comms, on average, GDP per capita does not change in response to a resource boom. However there is substantial likelihood that it decreases.
Table 1 and Model 2 in Table 3 confirm that under no comms, per capita GDP decreases by $48 (p = .6), which is statistically and economically insignificant (below a quarter of a standard deviation). Thus on average, we have a mild version of the resource curse, whereby most groups simply do not gain from a resource boom (when they should be gaining $700 per capita).
However if we look at Figure 4a, we can see that 8 out of 12 groups suffered a decrease in GDP per capita, and for 5 of them, this decrease was around $500 per capita, which exceeds a standard deviation in size—a veritable curse. In aggregate, their cursedness is masked by the high success of four groups in fully reaping the benefits of a resource boom.
Result 2b: Under full comms, GDP per capita increases one-on-one in response to a resource boom.
Table 1 and Model 2 in Table 3 confirm that under full comms, per capita GDP increases by $641 (p < .01), which is almost the predicted $700. A Wald test fails to reject its equality to $700 (p = .49). In Figure 4c, we can see that 11 out of 12 groups improve their GDP per capita as a result of a resource boom. Under full comms, resources are anything but a curse.
Result 2c: The increase in GDP in response to a resource boom is larger in full comms versus partial comms and in partial comms versus no comms.
Table 1 and Model 2 in Table 3 confirm that under partial comms, per capita GDP increases by $250, which is above the estimated increase under no comms and smaller than that under full comms. Using Wald tests, we can reject the hypotheses that it is equal to either (p < .01). The intermediacy of the partial-comms treatment is reflected in Figure 4b, where large gains and losses can be seen in response to resource booms.
In the online Appendix, we demonstrate the robustness of our results to changing various aspects of the data structure and estimation technique, as well as some ancillary results concerning how close the players’ actions are to a Nash equilibrium.
CONCLUSION
Much empirical evidence on the resource curse has been gathered, but there are still serious econometric questions about the curse's existence. We present clean evidence that resource booms can be completely squandered and that, on occasion, more can indeed be less. The key ingredient in ensuring the rigorousness of our evidence is our deployment of randomized control, in contrast to the literature's typical dependence on observational data. Our data also support the institutions-mediated version of the resource curse, whereby groups with good institutions reap the full benefits of resource booms.
Although we could have explored the same questions in a conventional laboratory experiment (Leibbrandt and Lynham Reference Leibbrandt and Lynham2013), we chose to allow participants to interact in a three-dimensional virtual world. We did this to more accurately capture the features of good institutions that Ostrom (Reference Ostrom2000) argued characterized successful natural resource management, while retaining the simplicity necessary to test our model.
Historians (Engerman and Sokoloff Reference Engerman, Sokoloff and Haber1997) and economists (Acemoglu et al. Reference Acemoglu, Johnson and Robinson2001) have offered important insights into how natural resources interact with economic and political institutions and how such interactions can have effects that last many centuries. By demonstrating that the mechanism underlying the institutions-mediated version of the resource curse is sound, we hope that the literature can focus on more accurately diagnosing its role in the woes of many resource-rich countries, such as Libya.
We believe that a particularly interesting path of inquiry would involve the use of small-scale randomized field experiments to bridge between the somewhat abstract virtual world we used and the complex environments in which the resource curse operates (Banerjee and Duflo Reference Banerjee and Duflo2009). For example, in our experiments, the economies were composed of a small number of people, whereas in practice, the resource curse is typically analyzed as a macroeconomic phenomenon. We leave such considerations for future research.
SUPPLEMENTARY MATERIAL
To view supplementary material for this paper, please visit http://dx.doi.org/10.1017/S2052263014000049.
