Dec 12, 2014 - Natalia Alvarado and Gabriel Tarriba ..... However, ex ante we don't know if Fixed Effects is the appropriate method because it is possible that ...
Revisiting the natural resource curse: does the relationship between natural resource rents and corruption persist into the 21st Century? Natalia Alvarado and Gabriel Tarriba 12 December 2014
Contents 1 Introduction
2
2 Research Question and Hypothesis
2
3 Literature Review
3
4 Data Sources and Variables
4
5 Descriptive statistics
5
6 Empirical framework and methods
6
7 Empirical Results
8
7.1
Basic models: Pooled OLS, Random Effects and Fixed Effects . . . . . . . . . . . . . . . . . .
8
7.2
Robustness tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
7.3
Alternative methods with lagged dependent variable . . . . . . . . . . . . . . . . . . . . . . .
9
8 Conclusions
12
9 Appendix
13
9.1
Description of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
9.2
Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
9.3
Diagnostic Tests and Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
9.4
Packages used for the elaboration of this paper . . . . . . . . . . . . . . . . . . . . . . . . . .
21
1
1
Introduction
Corruption is one of the most intensely debated phenomena in the field of public policy and development. Although there is no universally accepted definition of corruption, the consensus in the literature is that it occurs when “public official (A), acting for personal gain, violates the norms of public office and harms the interests of the public (B) to benefit a third party (C) who rewards A for access to goods or services which C would not otherwise obtain” (Brown [2006]). There is substantial empirical evidence that corruption obstructs efforts to reduce poverty and inequality, to accelerate economic growth and to promote sustainabile development [see Ades and Tella, 1996, Mauro [1998], Azfar et al. [2001] and Gupta and Abed [2002]]. This is why most countries avow their commitment to fighting and avoiding corruption. And yet, corruption remains an endemic problem across the developing world. The question of which factors are conducive to higher levels of corruption is central to this discussion. Since many national governments, international organizations and development agencies have stepped up their efforts to fight corruption Michael [2009], it is crucial to know what policies and institutional characteristics are worth funding in order to fight corruption. In recent years, for instance, many countries have set up anti-corruption agencies, appointed corruption ombudsmen and deepened transparency laws, in line with alleged international best practices. Nevertheless, more empirical research is needed in order to identify the factors that enable the spread and persistence of corruption. In particular, a better understanding of the ways in which corruption takes hold in public institutions could aid policymakers to determine in which areas to concentrate their efforts. This includes understanding the interplay between economic activity, rent-seeking behavior and corruption. In this context, it is worth keeping in mind that empirically, developing countries with large stocks of natural resources have tended to develop more slowly than developing countries with less natural resources (the ‘resource curse’). Could it be that natural resources is the missing link -that is, that the dependence on natural resources leads to higher corruption levels, and that these hinder economic development?
2
Research Question and Hypothesis
This paper seeks to answer a simple question: How strong is the relationship between a country’s degree of dependence on natural resource rents and its corruption level? As we will explain in more detail in the next section, in the 1990s scholars found such a relationship, particularly for developing countries or countries with fragile institutions. However, this relationship has not been revisited in the academic literature since then. We want to find out if the corruptive power of natural resources has waned. Moreover, we want to find out if the relationship is mediated by a third variable -for instance, whether the effect of natural resource rents on corruption depends on the GDP per capita of the country, such that in richer countries the effect is less than in poorer ones. Our hypothesis is that the relationship has not changed, because the basic mechanism that related resource wealth to corruption remains unchanged (namely the incentives for rent-seeking behavior -see among others Leite and Weidmann [1999] and Sala-i Martin and Subramanian [2003]). However, we are also aware that since the begin of the current millenium many resource-rich dependent countries (particularly in Africa) have experienced very high rates of economic growth. It is unclear whether this growth has been accompanied by a decrease in corruption or whether it has been independent of it.
2
3
Literature Review
In the last two decades, a significant amount of research in the field of development economics has sought to explore and understand the nature of the relationship between natural resource exploitation, good governance and economic development. In this section, we will describe the key findings, hypothesis and knowledge gaps relevant to our research question. The Natural Resource Curse The trigger for much of the interest in this topic was the empirical discovery made in 1995 by Sachs and Werner that countries with abundant natural resources tend to experience slower economic growth than countries with scarce natural resources (Sachs and Warner [1995]). This phenomenon was termed the “Natural resource curse”. Nevertheless, the hypothesis that Sachs and Werner put forth to explain this empirical relationship is purely economical: they conjectured that the exploitation and export of natural resources might provoke an over-valuation of a country’s currency, thus making its other exports uncompetitive (i.e. the Dutch disease). Sachs and Werner further develop this thesis in “The curse of natural resources” [2001]. Explaining the Natural Resource Curse: the Rent-Seeking Hypothesis Following Sachs and Werner’s original paper, many scholars have hypothesized that the culprit for the low growth rates of resource-rich nations might be not overvalued currencies but weak and ineffective institutions. This is what Mehlum et al call the - rent-seeking hypothesis -, through which “resource abundance leads to a deterioration of institutional quality in turn lowering economic growth” (Mehlum et al. [2006]). The idea is that as governments in countries with relatively weak institutions become more dependent on natural resources, they engage in rent-seeking behavior that is detrimental for the development and functioning of institutions. There is substantial empirical evidence for this institutional hypothesis. For instance, Leite [1999] concludes that the availability of natural resources is one of the factors on which corruption depends, along with government policies and the concentration of bureaucratic power. Similarly, analyzing the case of Nigeria, Sala-i-Martin and Subramanian [2003] conclude that “some natural resources - oil and minerals in particular exert a negative and nonlinear impact on growth via their deleterious impact on institutional quality”. One crucial fact highlighted by the empirical literature is that not all types of extractive industries have the same impact on the quality of governance and economic development: some researchers (notably Pendergast [2007] and Shaxson [2007]) find that the exploitation of fuel resources have a more negative impact on governance than the exploitation of other natural resources. Another important lesson from the empirical literature is that the effect of natural resource reliance on corruption levels depends on the quality of democratic institutions (see for instance Bhattacharyya and Hodler [2010]). Looking at the special case of the oil industry, Ross [2012] observes that oil-rich countries with strong governance structures (such as Norway, Canada or the US) are not more corrupt than non-oil-rich countries, but he highlights the prevalence of what he calls “rent seizing” (when the political elite seizes the natural rent to use it arbitrarily) in oil-rich countries with weaker governance structures. This is associated with patronage links and corruption (Ross [2012]). Resource dependence and governance An interesting framework to examine the relationship between corruption and natural resource rents is provided by Robinson and Acemoglu in their book “Why nations fail” (Robinson and Acemoglu [2012]). At the core of this framework is the idea that a country’s developmental path in terms of the extent to which its economic and political institutions are either extractive or inclusive. Extractive economic institutions are those designed to extract incomes and wealth from one subset of society to benefit a different subset, while extractive political institutions “concentrate power in the hands of a narrow elite and place few constraints on exercise of this power” (Robinson and Acemoglu [2012]). Moreover, extractive economic institutions “inherently depend on extractive political institutions for their survival” (Robinson and Acemoglu [2012]).
3
Since natural resource rents have historically been a key element of countries with extractive economic institutions, and considering that extractive economic institutions go hand in hand with extractive political institutions (where corruption is rife), the connection between the dependency on natural rents and corruption seems plausible, perhaps even likely. Summary of the existing academic literature To summarize the findings of the literature review, there seems to be a consensus that higher levels of dependence on natural resource rents is associated with higher levels of corruption, but this relationship is not equal for all types of natural resource rents. A recurrent theme in the literature on the topic is the difficulty to measure corruption as well as identification and causality issues: being dependent on natural resources might favor corruption, but highly corrupt regimes might also choose to rely more heavily on natural resources (rather than diversifying their sources of income by promoting economic growth, for instance). Moreover, from the framework of Acemoglu and Robinson [2012] we expect to see extractive economic institutions (typically resource-dependent) accompanied by extractive political institutions (where corruption tends to be more significant than in fully democratic countries).
4
Data Sources and Variables
In order to be able to empirically investigate our research question, we needed data for our dependent variable (corruption level), our independent variable of interest (natural resource rents) and control variables that are relevant explanatory variables of our dependent variable. Our dataset was composed from data gathered from two projects of the World Bank, the World Governance Indicators (WGI) (variable for corruption level) and the World Development Indicators (WDI) (all other data). We gathered our data using R (R Core Team [2015]). Because we want to explore the hypothesis that oil and gas rents are particularly detrimental for good governance (see among others Ross [2012]; Shaxson [2007]), we collected data not only on overall natural resource rents but also on oil and gas rents. Table 1 below displays a summary of the variables that composed our final dataset (see the Appendix for a detailed description of the variables): Table 1. Summary of variables in dataset Variable
Measurement
Source
Corruption Control Total Natural Resource Rents Oil Rents Natural Gas Rents GDP per capita Unemployment rate
Index, -2500 to 2500 % of GDP % of GDP % of GDP 2005 constant USD, PPP % of workforce
WGI WDI WDI WDI WDI WDI
The resulting dataset has thus six variables and 1164 observations, corresponding to 165 countries and 7 years (2005 to 2011). However, for purposes of our analysis, we created some extra variables from these ones such as lags (for the dependent variable Corruption), or the logged GDP per capita (since this variable is highly skewed to the right). Moreover, we also created interaction variables to investigate the interplay between the dependent variable, the independent variable and the control variables. Data limitations and challenges Unfortunately, due to the prevalence of missing data, we were not able to include in our dataset variables that the academic literature identifies as important for the determination of corruption. Among these are indicators such as the Gini coefficient (income inequality), education or conflict. This is an important limitation of our data. We faced a trade-off between number of countries in our sample and number of variables, and decided to try and increase the number of countries in order to be able to analyze data for all regions of the world. Nevertheless, the six variables that we have are all relevant for our analysis so the fact that we could not gather other variables is not a major setback for our research purposes. 4
5
Descriptive statistics
Corruption in the world
Total Natural Resource Rents in the world
Before we begin with the econometric analysis of our data, it is worth examining the general trends observable in both the dependent variable, corruption, and the independent variable of interest, natural resource rents. The above maps show the level of corruption and the dependence on natural resource rents in 165 countries, respectively. In the corruption map, darker hues indicate higher corruption while in the natural resource rents map, darker hues indicate higher dependence on rents (as % of GDP). Seen side by side, the maps suggest that there is a considerable overlap between highly corrupt countries and natural rent-dependent countries. Many African, Middle Eastern and former Soviet countries fit into both categories. Inversely, few Western countries are either highly corrupt or highly dependent on natural resource rents. These patterns, however, are not enough to draw conclusions because there might be a third variable (for instance, GDP per capita) driving both corruption and natural resource dependence. A limitation of the maps shown above is that they do not show the values of the variables. In order to have a better overview of the data, we created a table with basic descriptive statistics for all variables. Table 2 below shows the mean, standard deviation, minimum and maximum for all variables, as well as the number of observations (which is 1,164 as we have data for 165 countries and 7 years):
5
Table 2: Descriptive statistics Statistic Obs corrupest gasrents gdppc oilrents totrents unemp gni population l_gdp gdp_rents icor
6
N
Mean
St. Dev.
Min
Max
1,164 1,164 1,164 1,164 1,164 1,164 1,164 1,164 1,164 1,164 1,164 1,164
582.50 −52.12 1.78 12.00 6.41 12.74 8.45 11,995.64 39.62 1.38 13.70 52.12
336.16 1,006.32 5.76 17.12 14.08 17.37 6.04 17,116.75 141.71 1.61 40.72 1,006.32
1 −1,637.17 0.00 0.12 0.00 0.00 0.00 120 0.27 −2.12 −77.13 −2,552.69
1,164 2,552.69 69.18 90.27 78.25 89.22 37.60 90,270 1,344.13 4.50 245.97 1,637.17
Empirical framework and methods
In order to investigate our research question empirically and test our hypothesis, we need to operationalize it according to the characteristics of our panel dataset. Thus, from the research question we derive an empirical question: To what extent does the level of dependence on natural resource rents of a country explain the level of corruption, after controlling for other relevant variables? Approach 1 - Poooled OLS Regression The starting point of our empirical analysis is the simplest of methods, namely a Pooled OLS regression which does not distinguish among countries nor across time. This type of regression is useful for identifying general trends in time. The basic form of our Pooled OLS regression is as follows. Corrit = β0 + β1 N RRit + β2 lnGDP pcit + β3 lnGDP pcxN RRit + β4 U nemit + it Where Corr stands for the Control of Corruption Index, NRR stands for Natural Resource Rents, lnGDPpc stands for the natural logarithm of GDP per capita, lnGDPpcxNRR is the interaction term of GDP per capita and Natural Resource Rents, and Unemp stands for Unemployment. To find out if there are different effects when the rents originate from oil or natural gas, we also run this regression substituting Total Natural Resource Rents for Oil rents and Gas rents. Please note that the interaction term between natural resource rents and logged GDP per capita has been added to allow for a better fit: we think that as countries get richer, the effect of natural resources on corruption should decrease. Hence, we expect a positive coefficient on the interaction term. However, given that the variables Oil Rents and Gas Rents are potentially very correlated with Total Rents, we run a Variance Inflation Factor test and choose the model according to its results. Given that we have a panel structure, we also consider running Fixed Effects and Random Effects to account for possible time-invariant country-specific characteristics included in the error term and that may lead to endogeneity problems (in which case Pooled OLS will deliver inefficient and biased results (i.e. covariance between the independent variables and the residuals). Approach 2 - Fixed Effects regression In case that Pooled OLS has problems of endogeneity, we can try to estimate a fixed-effects (FE) regression. This method allow us to control for the time-invariant country-specific characteristics (i.e. the fixed effects), such that we will only be analyzing the variation of a variable within a country across time. However, FE does not consider the variation between countries. 6
To run a Fixed Effects regression, we will subtract from each term its mean. Thus, the regression equation will be: ∗ ∗ ∗ Corrit = β1 N RRit + β2 lnGDP pc∗it + β3 lnGDP pcxN RRit + β4U nem∗it + αi +
The terms in the equation above are time-demeaned, such that: ∗ N RRit = N RRit − N RRi
lnGDP pc∗it = lnGDP pcit − lnGDP pci ∗ lnGDP pcxN RRit = lnGDP pcxN RRit − lnGDP pcxN RRi
U nem∗it = U nemit − U nemi Where the overbar denotes that it is the panel-specific mean across all time periods. Meanwhile, the αi denotes the country-specific fixed effect. An F-Test is necessary to decide whether Pooled OLS or FE is the more appropriate technique. Please see the section of Diagnostic Tests and Robustness Checks to see the results of the F-Test in the annex. Approach 3 - Random Effects regression However, ex ante we don’t know if Fixed Effects is the appropriate method because it is possible that the panel specific components αi are actually random time-invariant terms within the residual such that it = αi + uit (i.e. the error term has a time-invariant element and a time-variant element, both of which are random). If these residuals are random, we are better off with a Random Effects regression because it exploits not just the variance within panel units (countries in this case) but also between them. Therefore, statistically RE is more efficient as it produces smaller standard errors and thus more reliable estimators. With Random Effects, our regression is:
Corrit − λCorri = β1 (N RRit − λN RRi ) + β2 (lnGDP pcit − λlnGDP pci ) + β3 (U nemit − λU nempi ) + it In the above equation, the overbar denotes that the term has been time-demeaned, just as in the case of Fixed Effects. Meanwhile, λ is a parameter that can take values between 0 and 1. It is estimated as follows: λ=1−
σu2 T σα2 + σu2
Where σu2 is the variance of the time-variant part of the residual ui t and σα2 is the variance of the time-invariant part of the residual, αi . To verify which of the two methods is more appropriate -that is, whether the αi is a random disturbance or reflects time-invariant country-specific characteristic- we need to conduct a Hausman Test. And to see whether Fixed Effects is more appropriate than Pooled OLS, we conduct an F-test. The results of the F-Test are in the Annex. Other approaches Should we find evidence of endogeneity in our model, we might undertake other approaches such as Dynamic Linear Panel Models (including a lagged dependent variable) and/or First-Difference models. The goal is to devise a regression model that deals with the two problems that are typical to panel data, namely endogeneity and the residual structural issues (heteroskedasticity, panel dependence, auto- or serial correlation). However, we will choose these alternative models according to the results we obtain from our basic models.
7
7 7.1
Empirical Results Basic models: Pooled OLS, Random Effects and Fixed Effects
We run the three models described in the previous section and obtain the following results: Table 3: Regression Results: Pooled OLS (1), Fixed Effects (2) and Random Effects (3) Dependent variable: Corruption Control (1)
(2)
(3)
Natural Rents
−5.94 (1.40)
2.05 (0.99)
2.11∗∗ (1.04)
Log GDP per capita
531.37∗∗∗ (12.77)
50.38∗∗ (21.36)
234.02∗∗∗ (19.01)
Log GDP x Natural Rents
−5.32∗∗∗ (0.64)
−1.66∗∗∗ (0.57)
−3.49∗∗∗ (0.58)
Unemployment rate
−17.65∗∗∗ (2.67)
−4.55∗ (2.56)
−2.93 (2.64)
(Intercept)
−488.42∗∗∗ (37.13)
Observations R2 Adjusted R2 F Statistic
∗∗∗
∗∗
−341.01∗∗∗ (59.55)
1,164 0.70 0.70 685.88∗∗∗ (df = 4; 1159)
3.74∗∗∗
1,164 0.01 0.01 (df = 4; 991) ∗
Note:
40.89∗∗∗
p
