The Effect Of Corruption On Economic Growth In The BRICS .

channels, has a totally negative impact on growth. All these studies were cross-sectional innature.Among the panel data studies carried out, Gyimah-Brempong (2002) estimated the impact ofcorruption on growth and income distribution on African countries using a dynamic panel dataestimator. Using a panel of 13 countries and a sample period between 1993 and 1999, the authorfound that corruption reduced GDP growth and per capita income by between 0.75 and 0.9percentage points and 0.39 and 0.41 percentage points per year, respectively. Using dynamicpanel estimators, Swaleheen (2011) investigated the impact of corruption on growth for a panelof 117 countries over the period 1984 to 2007 and concluded that corruption has a directlynegative effect on growth. D’agostino, Dunne and Pieroni (2012) evaluated the impact ofcorruption on government spending and economic growth. Using a sample of African countriesand a sample period between 1996 and 2007, these authors estimated a panel data model andconcluded that corruption has a negative impact on growth.In light of the previous studies on the impact of corruption and economic growth, the presentpaper aims to contribute to the existing literature by considering the impact of corruption oneconomic growth across BRICS countries using different panel techniques. The next sectiondiscusses the measurement issues of corruption.2. Measuring corruptionEmpirical studies on the impact of corruption on economic growth are plagued by disagreementover the correct measure of corruption. The contention is that corruption cannot be measuredobjectively. Without an objective measure, scholars and researchers have resorted to subjectivemeasures of corruption, such as corruption perception indices. These indices are usually in theform of surveys targeted at individuals, households, firms or experts, and asking them abouttheir experience of corruption, either in the private or public sector or both. The problem,however, is that corruption perception indices are known to be a poor reflection of realcorruption experiences (Kauffman, Kraay and Mastruzzi, 2006; Gonzales, Lopez-Cordova &Valladares, 2007; Olken, 2009).One of the reasons perception indices are a poor reflection of real corruption experience is thatperceptions are inherently biased. In this regard, it is helpful to note the observation made byOlken (2009) that perceptions are biased because individuals’ beliefs are biased. Similarly,Gonzalez et al. (2007) observe that, because each respondent has his own reference point,which is unlikely to be shared by many, perceptions tend to be plagued by a contextual problem.Another concern that has been raised about perceptions is the perceptions’ convergenceproblem, which relates to the idea that peoples’ perceptions of corruption will tend to convergesince they receive news from the same mass media and hear their friends’ opinion aboutcorruption (Cabelkova & Hanousek, 2004).Kauffman et al. (2006) argue that while some of the concerns raised about the validity ofcorruption perception indices remain valid, some have no merit. For instance, one of thearguments is that subjective measures of corruption are too unreliable. These authors argue thatno measure of corruption can be 100% reliable in the sense of giving precise measures ofcorruption owing to the measurement error present in any forms of data, both subjective andobjective. Another objection to the use of corruption perceptions is that they are generic andvague rather than a reflection of reality. Once again, Kauffman et al. (2006) note that surveyquestions on corruption have become specific, focused and quantitative. In this regard, it is alsoinstructive to note the observation made by Olken (2009) in his study of corruption in the5

context of a road-building programme in rural Indonesia. In the study, Olken (2009) observesthat villagers were sophisticated enough to distinguish between general levels of corruption inthe village and corruption in the particular road project examined.Given the issues around corruption perception indices, two questions need to be answeredrelating to their relevance. The first one is whether or not to abandon corruption perceptionindices completely. Kaufman et al. (2006) argue that corruption perception indices remain theclosest way to measure corruption. This is because it is difficult to measure the real corruptionexperience owing to the secretive nature of corruption and the fact that corruption is known notto leave a paper trail.Following from the first question is whether or not the valid limitations inherent in corruptionperception indices can be dealt with adequately enough to give proper insight into the natureand implications of corruption for the economy. We argue that it is possible to deal with theselimitations.The first way to deal with these limitations is by noting the bias inherent in whichever measureof corruption used and dealing with each bias in the interpretation of results. By taking note ofthese biases and how they shape these experts’ perceptions of corruption in different countries,it becomes possible to adjust this bias downward in the interpretation of results. Secondly,based on who is being asked, particular surveys are likely to provide better information oncorruption than others. For example, investment analysts, interviewed as experts, will give abetter view of how corruption affects their ability to invest in a country relative to an individualwho has no resources to invest in the real economy. Similarly, individuals with politicalconnections will likely provide a more informed view of how corruption affects their incentivesto accumulate human capital (relative to political capital) than investment analysts.In essence, based on the respondents of each survey, particular surveys are likely to reflectcorruption in particular spheres of society better than others. Thus, using a variety of corruptionperception indices may be a useful way of reducing bias inherent in a particular index as wellas providing proper insight into corruption. Lastly, a distinction ought to be drawn at all timesbetween what corruption perception indices measure and how they differ from actualcorruption experiences in order to make proper inferences and recommendations from studiesthat make use of them. To address these shortcomings, Kauffman and Kraay (2008) advise thatit is important for researchers to rely on a variety of data sources as measures of corruption.However, due to the lack of data, this study employs only one measure of corruption.3. MethodologyThis study seeks to investigate the effect of corruption on economic growth in the BRICScountries using a battery of panel4 data techniques. The major attraction of panel datatechniques stems from the ability of such models to address serious econometric issues such asheterogeneity, endogeneity and the persistence of shocks in dynamic models, which cannot beefficiently addressed in pure time-series and pure cross-sectional models. Accordingly, besidesthe benchmarks fixed effects model (FEM) and/or random effects model (RAM), the ArellanoBond first difference and the system generalized method of moments estimators are consideredto account for the dynamic nature of the growth model.4These includes Fixed effects, Random Effects and GMM estimators6

Specifically, the baseline model is given as:𝑃𝐸𝑅 𝐶𝐴𝑃 𝐺𝐷𝑃𝑖𝑡 𝛼𝑖 𝛽 𝐶𝑂𝑅𝑅𝑖𝑡 𝛾 𝐶𝑂𝑅𝑅𝑆𝑄𝑖𝑡 𝛿 ′ 𝜒𝑖𝑡 𝜀𝑖𝑡𝑖 1, . . , N (country); 𝑡 1, . . , T (time) ,,(3.1)where: 𝛼𝑖 signifies country-specific fixed effects; 𝐶𝑂𝑅𝑅𝑖𝑡 and 𝐶𝑂𝑅𝑅𝑆𝑄𝑖𝑡 are the corruptionvariable and the square of the corruption variable; 𝜒𝑖𝑡 is a vector of control variables, whichincludes investment, literacy rate, population growth, government consumption, openness andpoliticalstability(𝐼𝑁𝑉𝑖𝑡 ; 𝐿𝐼𝑇𝑖𝑡 ; 𝑃𝑂𝑃𝐺𝑅𝑂𝑊𝑖𝑡 ; 𝐺𝑂𝑉𝐶𝑂𝑁𝑖𝑡 ; 𝑂𝑃𝐸𝑁𝑖𝑡 ; 𝑃𝑂𝐿𝑆𝑇𝐴𝐵𝑖𝑡 )respectively; and 𝜀𝑖𝑡 is the error term.3.1.1. HeterogeneityThe significance of heterogeneity bias in the literature on corruption and economic growth hasbeen emphasized by Gyimah-Brempong (2002); Swaleheen and Stansel (2007) as well asAhmad, Ullah and Arfeen (2012), among others. They argued that time-invariantheterogeneity - in terms of religion, culture and institutions - has an important role to play inexplaining cross-country differences in the relationship between corruption and economicgrowth. Hence, failure to omit country and time-specific effects that exist among crosssectional units and time series units could result in inconsistent parameter estimates (Hsaio,2003).In fact, the use of different political systems represents one source of heterogeneity amongBRICS countries. China is a one-party state; Russia has a centralized government; while Brazil,India and South Africa are democracies. As North (1991) noted, institutions both formal(constitutions, laws and property rights) and informational (customs, traditions and taboos)have a role to play in the economic performance of nations. Furthermore, differences, in termsof the importance of the various determinants of economic growth of the BRICS countries, arelikely to lead to heterogeneity. In this regard, it is worth noting that BRICS countries also havevarying levels of economic development, with China outpacing the rest of the other countriesin terms of economic size, growth and trade. The member countries are also differently situatedin terms of resources, absolute consumption and energy intensity and have differentdemographic trends. For instance, Brazil has a predominantly urban population, while India islargely rural. Russia has an ageing population while South Africa is still young (Saran, Singh& Sharan, 2012).A potential solution to such heterogeneity bias is the use of fixed effects or random effectsmodels, which adequately control for unobserved time-invariant heterogeneity (Hsiao, 2003).A key assumption of the fixed effects model is that the explanatory variables are independentof the error term, 𝜀𝑖𝑡 . The parameter estimates are obtained by performing the regression indeviations from individual means. In effect, the fixed effects model eliminates country-specificeffects, 𝛼𝑖 , by transforming the data known as ‘demeaned’ or ‘within transformation’. And theordinary least squares technique is implemented on the transformed data to obtain theparameter estimates known as the ‘within estimator’ or ‘fixed effects estimator’ (Verbeek,2004). A fundamental assumption of the fixed effects model is that of strict exogeneity,wherein a strictly exogenous variable is not dependent on current, future and past values of theerror term (Verbeek, 2004).A shortcoming of the fixed effects model is its assumption of strict exogeneity, which may nothold in certain instances. In such circumstances, the random effects model is the mostappropriate. Contrary to the fixed effects model, the random effects model assumes that the7

country-specific effects,𝛼𝑖 , are random factors that are independently and identicallydistributed over individual countries. The error term consists of two components: an individualspecific component, which is time-invariant, and a remainder component, which is assumed tobe uncorrelated over time; while 𝛼𝑖 and 𝜀𝑖𝑡 are assumed to be mutually independent andindependent of the explanatory variables. In light of these assumptions, the OLS estimator forthe country-specific effects and the parameters is unbiased and consistent.However, the error components structure implies that the composite error term, 𝛼𝑖 𝜀𝑖𝑡, exhibits a particular form of autocorrelation (unless 𝜎𝛼2 0). As a result, routinelycomputed standard errors for the OLS estimator are incorrect and a more efficient generalizedleast squares (GLS) estimator is obtained by exploiting the structure of the error covariancematrix (Verbeek, 2004).Verbeek (2004) notes that it may be preferable to use a fixed effects estimator wherein interestlies in the country-specific effects, 𝛼𝑖 . Furthermore, the fixed effects may be the appropriatemodel to use when the country-specific effects, 𝛼𝑖 , and the explanatory variables, 𝑥𝑖𝑡, arecorrelated, since the fixed effects model eliminates the individual effects of 𝛼𝑖 and the problemsthey cause. The random effects approach, however, because it ignores the correlation betweenindividual effects ( 𝛼𝑖 ) and the explanatory variables (𝑥𝑖𝑡 ), may lead to inconsistent estimatorsif such an assumption holds. Therefore, a formal testing procedure, namely, the Hausman testhas been proposed as the test for choosing between the fixed effects estimator and the randomeffects estimator.The Hausman test essentially compares two estimators: one which is consistent under both thenull and alternative hypotheses; and one which is consistent and typically efficient under thenull hypothesis only. The test compares the random effects model against the fixed effect modelunder the following hypothesis:𝐻0 : 𝛼𝑖 𝑖𝑠 �� 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑒𝑑 𝑜𝑓 𝜒𝑖 (random effects);𝐻1 : 𝛼𝑖 𝑖𝑠 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑 𝑤𝑖𝑡ℎ 𝜒𝑖 (fixed effects).If the p-value is 0.05, the null hypothesis is rejected and the fixed effects model is chosen;but if the p-value is 0.05, the null hypothesis is not rejected, meaning the random effectsmodel is more consistent and efficient.While fixed effects and random effects models are capable of addressing problems ofheterogeneity, such models are likely to suffer from a number of shortcomings, notablyendogeneity, especially in the context of economic growth.3.1.2. EndogeneityEndogeneity refers to the correlation of explanatory variables and the disturbances in a model.This may be caused by the omission of relevant variables, measurement error, sampleselectivity, self-selection or other reasons. Endogeneity results in inconsistent ordinary leastsquares (OLS) estimates (Baltagi, 2005).One source of endogeneity in the corruption and economic growth literature is simultaneitybias. Simultaneity refers to the dual causality that exists between the dependent and one ormore of the explanatory variables, that is, corruption, investment and the rate of economicgrowth. In other words, random shocks that affect economic growth may simultaneously affect8

corruption and investment as well as other explanatory variables. (Gyimah-Brempong, 2002;Swaleheen, 2011; Islam, 1995; Baltagi; 1995: Swaleheen & Stanesel, 2007).The second source of endogeneity is omitted variable bias. This is because some variables thathelp explain economic growth might not be included in the model due to the lack of consistentdata. For example, human capital is regarded as a determinant of economic growth. Thisvariable is commonly proxied by the use of gross enrolment ratios in primary and secondaryeducation. However, due to the lack of data for the BRICS countries, this study had to rely onadult literacy rates as a proxy for human capital. Other omitted variables include those relatingto rule of law or property rights, which were excluded from the present study due to the lackof data but have been considered by other scholars as determinants of economic growth.Another source of endogeneity is measurement error, which arises from the use of survey data.In this instance, the corruption data used in the present study is based on perceptions ofcorruption in BRICS countries based on individuals’, households’ and experts’ perceptions ofcorruption in these countries. In this regard, corruption perceptions data is known to be biasedon the basis of economic development, religious beliefs and democratic institutions (Donchev& Ujheyli, 2014). Further, they are known to be a poor reflection of corruption experience(Kauffman, Kraay & Mastruzzi, 2006).A third source of endogeneity in the corruption and economic growth literature is attributed tothe dynamic structure of economic growth models. Economic growth models include, as anadditional explanatory variable, a lag of the dependent variable, that is, economic growth inprevious periods, to account for the persistence of economic growth (Islam, 1995; GyimahBrempong, 2002; Swaleheen, 2011). However, the addition of the lagged dependent variablecauses correlation between the lag dependent variable and the error term, resulting in biasedestimates of parameters (Hsiao, 2003; Judson & Owen, 1999).To address the problems of endogeneity in the model, the study proposes the use of thegeneralized method of moments estimators (Caselli, Esquivel & Lefort, 1996). Morespecifically, the model is presented as follows:𝐿𝑂𝐺 𝑃𝐸𝑅𝐶𝐴𝑃 𝐺𝐷𝑃𝑖𝑡 𝛼𝑖 𝛽 𝐶𝑂𝑅𝑅𝑖𝑡 𝛾𝐶𝑂𝑅𝑅𝑆𝑄𝑖𝑡 𝛿 ′ 𝜒𝑖𝑡 𝜆𝐿𝑂𝐺 𝑃𝐸𝑅𝐶𝐴𝑃 𝐺𝐷𝑃𝑖(𝑡 1) 𝜀𝑖𝑡(3.2)i 1, . . , N(country); t 1, . . , T (time),where: 𝛼𝑖 is country-specific fixed effects; 𝐶𝑂𝑅𝑅𝑖𝑡 and 𝐶𝑂𝑅𝑅𝑆𝑄𝑖𝑡 are proxies for corruptionand the square of corruption; 𝜒𝑖𝑡 is a vector of control variables, which includes𝐼𝑁𝑉𝑖𝑡 , 𝐿𝐼𝑇𝑖𝑡 , 𝑃𝑂𝑃𝐺𝑅𝑂𝑊𝑖𝑡 , 𝐺𝑂𝑉𝐶𝑂𝑁𝑖𝑡 , 𝑃𝑂𝐿𝑆𝑇𝐴𝐵𝑖𝑡 and 𝑂𝑃𝐸𝑁𝑖𝑡 .The variable 𝐿𝑂𝐺𝑃𝐸𝑅𝐶𝐴𝑃 𝐺𝐷𝑃𝑖(𝑡 1) is the lag of the logged dependent variable; and 𝜀𝑖𝑡 is theerror term.This study employs the Arellano-Bond first difference estimator as proposed by Arellano andBond (1991). To obtain the Arellano-Bond estimator, the growth regression is first rewrittenas a dynamic model, as in equation (3.2). Secondly, the dynamic model is differenced in orderto eliminate individual effects. Thirdly, the right-hand side variables are instrumented using alllagged values of endogenous and predetermined variables as well as the current and laggedvalues of exogenous regressors as instruments in the differenced equation. The last stepeliminates the inconsistency arising from the endogeneity of the explanatory variables, whilethe differencing removes the omitted variable bias. The model thus appears as follows:9

𝐿𝑂𝐺 𝑃𝐸𝑅𝐶𝐴𝑃 𝐺𝐷𝑃𝑖𝑡 𝛼 𝛽 𝐶𝑂𝑅𝑅𝑖𝑡 𝛾 𝐶𝑂𝑅𝑅𝑆𝑄𝑖𝑡 𝛿Δ′𝜒𝑖𝑡 𝜆Δ𝐿𝑂𝐺 𝑃𝐸𝑅𝐶𝐴𝑃 𝐺𝐷𝑃𝑖(𝑡 1) Δ𝜀𝑖𝑡(3.3)i 1, . . , N (country); t 1, . . , T (time).However, in circumstances where the lagged dependent variable and explanatory variables arepersistent, the lagged instruments of the Arellano-Bond first difference estimator are weak,thus compromising the asymptotic precision of the estimator. Furthermore, the first differencesused in the Arellano-Bond estimator worsen the bias due to measurement errors in variables(Blundell & Bond, 1998; Felbermayr, 2005; Swaleheen, 2011).3.1.3. Persistence of economic growthA solution to problems caused by the persistence of the lag of the dependent variable andexplanatory variables is the use of system generalized method of moments estimators asproposed by Blundell and Bond (1998). The Blundell and Bond GMM system estimator jointlyestimates the Arellano-Bond GMM first difference estimator in first differences and levels,using different sets of instruments for each part. The instruments for the equations in firstdifferences are lagged level values of the endogenous variables and first differences of theexogenous variables. The instruments for the equations in levels are lagged di

has found corruption to have a detrimental impact on economic growth, the growth impact of corruption does indeed decreases with the level of corruption. This suggests a possible corruption level from which, the relation might lead to opposite effects. Keywords: Corruption, Economic Gr