Journal Issue
Share
Article

A Governance Dividend for Sub-Saharan Africa?1

Author(s):
Amine Hammadi, Marshall Mills, Nelson Sobrinho, Vimal Thakoor, and Ricardo Velloso
Published Date:
January 2019
Share
  • ShareShare
Show Summary Details

I. Introduction

There is a renewed interest in improving governance and accentuating the fight against corruption globally. The African Union’s aptly chosen theme for 2018 is “Winning the fight against corruption”. Similarly, the new leaders coming to power in Angola, Ethiopia and South Africa, among others, have placed the fight against corruption at the top of their agenda. These developments come against the backdrop where SSA countries generally lag those in most other regions in terms of corruption perceptions and governance. The average score of SSA on corruption and governance indicators is similar to the MENAAP but lower than other regions (Figures 1A and 1B). Eighty percent of SSA countries (36 out of 45) score below the global average in the Transparency International’s Corruption Perceptions Index (CPI) and only 3 of the 30 SSA countries included in the International Country Risk Guide’s (ICRG) governance indicator have above average scores (Figure 1C). There is, however, significant intra-regional variation in scores (Figure 1D).2

Figure 1:Corruption Perceptions and Governance in SSA and the World

Sources: Transparency International, ICRG, and authors’ calculations. All indicators are normalized 0 – 100, lower (higher) score better for corruption (governance). Confidence intervals are estimated around 2017 scores using std. deviations for the 5-year period 2012-17.

The “institutionalist view” posits that post-independence institutions and administrations in SSA are under-developed and that rent-seeking behaviors are intertwined with weak governance and corruption (Mulinge and Lesetedi, 1998). At the same time, the low scores of SSA countries are not necessarily due solely to a lack of legislative and institutional frameworks, as many countries have adopted legislation that criminalizes corruption and related offences, improved their AML/CFT frameworks, and established specialized anti-corruption agencies. Instead, one can argue that it is likely a combination of limited institutional capacities and weak enforcement of these frameworks that contribute to the heterogeneity in governance performance across SSA.

Weak governance, which is strongly and positively correlated with high corruption perceptions, has been associated with poor economic and social performance. Corruption and weak governance hinder economic performance more directly through various channels, including: higher tax evasion and lower tax revenue, as well as increased central bank financing of government deficits; a shift both in the composition and quality of government spending; potentially poor lending practices and weak financial supervision; and lower investment (IMF, 2016; IMF, 2018). Hence, corruption and weak governance have been put forward as some of the key factors holding back growth and economic development in SSA (e.g., Ndulu and O’Connell, 1999; Collier and Gunning, 1999; and Kilish et al, 2013).

This paper investigates the correlation between overall governance and one crucial dimension thereof, namely corruption, and economic performance in SSA. Our main goal is to test whether weak governance and corruption have any impact on SSA growth. At the same time, it is important to acknowledge that low corruption and good governance are not the sole drivers of growth. Indeed, there are various examples of countries perceived as being poorly governed that have had episodes of strong growth driven by other factors, for example natural resource wealth. In other cases, countries with good governance have not necessarily benefited from strong growth. Our conjecture is that corruption tends more often than not to undermine economic growth, thus behaving more like sand in the wheels rather than oiling the engine. While this paper looks at relationships across a sample of countries, examining exceptions to the trends could help reveal the combinations of factors including weak governance that hold back growth, as well as factors that could mitigate its effects.

We estimate a standard growth model augmented for governance and corruption to assess the impact on GDP per capita growth for 190 countries using 5-yearly observations over the period 1984-2015 in a system generalized method of moments (SGMM) model. In the baseline regressions, we use two measures of governance—ICRG’s Political Risk Rating (ICRG) and an aggregate measure of Kaufmann and Kraay’s (Kaufmann et al, 2010) Worldwide Governance Indicator (WGI), and two indicators of corruption perceptions—the Kaufmann and Kraay’s Control of Corruption Indicator (CCI) and Transparency International’s CPI.3 We focus on SSA but also try to tease out whether growth in the region is impacted differently by weak governance and corruption relative to other regions and country groups. To the extent possible, we also account for the concerns relating to the definition of governance and corruption, non-linearities, and potential endogeneity problems. We conduct a range of sensitivity analysis to ascertain that the results remain robust to both changes in time periods and alternative governance indicators. We also split the sample to assess whether countries with weaker governance or higher corruption are any different from the rest.

The main findings point to an adverse correlation between weak governance/corruption and growth that is both statistically and economically meaningful and has a higher impact in SSA. These estimates are generally robust to various sensitivity checks, including to alternative measures of governance, sample period, country groupings, and specifications that control for omitted variables bias. Our baseline models also survive nearly all the tests performed to better control for endogeneity.

While the estimated correlations between governance/corruption and growth do not prove causation, they do suggest the following:

  • The impact of weak governance on GDP per capita growth is stronger in SSA relative to the rest of the world and bringing governance to the world average could increase GDP per capita by an estimated 1 to 2 percentage points. An improvement in governance of about one-standard deviation in the SSA sample4—which for an average SSA country would result in governance converging to the world average—is associated with an increase in GDP per capita growth of about 1–2 percentage points, depending on the governance indicator. This observed impact is two to three times larger than for the average country in the rest of the world and stronger than the impact estimated for other regions like Latin American (LAC) and North Africa/Middle East/Afghanistan/Pakistan (MENAAP) that are also perceived to have weak governance. Admittedly, the process of institutional reforms will take considerable time and effort, but the results indicate a dividend from these efforts.5 Results from other studies globally suggest a broadly similar impact from governance improvements in SSA.

  • Corruption also has a deleterious impact on SSA countries relative to the rest of the world and bringing corruption to the world average could increase GDP per capita by 1 percentage point. An improvement in corruption perceptions in the average SSA country by about one standard deviation in the SSA sample—which would bring it to the world average—is associated with an increase in GDP per capita growth of about 1 percentage point. This impact is comparable to that estimated for other regions like MENAAP that are also perceived to have high corruption perceptions.

  • There is a nonlinear correlation between governance/corruption and growth for both the overall and SSA samples, with stronger correlations with weaker governance or higher corruption perceptions. Unsurprisingly, growth is more adversely affected in SSA countries with more acute governance and corruption problems. SSA countries with governance (corruption perceptions) scores below (above) 60—about 75 percent of SSA country-years in both samples—stand to benefit the most by addressing governance and corruption problems. The non-linear models suggest that growth gains could be twice as large as the baseline for those SSA countries with very low (high) governance (corruption) scores (e.g. at 5th and 95th percentile of the respective distributions).

  • Other specific dimensions of governance such as rule of law, regulatory quality and government accountability—which we see as proxies for transmission channels of governance—are also positively correlated with growth, and their improvement may also yield sizeable gains. The estimated growth gains from a 10-point improvement in these specific components of governance range from ¾ percentage point for rule of law to about 1 percentage point for voice and accountability. These gains are non-additive, but the overall impact could be higher due to potential complementarities across the different channels.

Based on a review of the literature, this paper is probably among the first to exhaustively explore the link between governance and growth and provide comprehensive estimates of the correlation between multiple aspects of governance and growth for SSA. We also explore a wide set of governance indicators and rely on a comprehensive and updated sample to ensure robustness. Research has typically focused on selected dimensions of governance (e.g., political instability, as in Mo, 2001) and worldwide samples. More recently, Kilish et al (2013) found that the quality of regulation and the rule of law have a significant impact on SSA growth.

The evidence gathered in this paper suggests that the correlation between governance/corruption and growth in SSA could be stronger than in other regions. While not entirely conclusive, our non-linear results suggest that very low (high) levels of governance (corruption perceptions) are more detrimental to growth. Because corruption is perceived to be severe in SSA, our non-linear estimations point to a clear negative correlation between corruption and growth in most SSA countries. Saha et al (2017), among others, also provide similar findings.6 We conjecture that there are unmeasured characteristics reflecting factors such as the capacity of the institutions and rent-seeking in resource-rich sectors that partly explain the amplified growth impact of corruption and governance in SSA. The study of exceptions to these patterns could help reveal these explanatory factors but is beyond the scope of this paper.

Our baseline corruption-growth nexus estimates are in the ballpark of other empirical investigations, including those using more advanced techniques to control for model uncertainty (IMF, 2018). Gyimah-Brempong (2002) estimate that a one standard deviation improvement in corruption in African countries would increase per capita income growth by about ¾ percentage point. Roughly identical estimate is also obtained from Gyimah-Brempong and de Camacho (2006) who estimate a model somewhat similar to ours and also find the impact of corruption on growth being stronger in SSA and LAC than in other regions. Similarly, Ugur and Dasgupta (2011) find that a one-unit improvement in corruption perceptions (equivalent to roughly one standard deviation if measured by CCI) would increase annual GDP per capita growth by about 0.6 percentage point in low-income countries (LICs), and almost 0.9 percentage point in a “mixed” sample of LICs and non-LICs.

Our findings emphasize the importance of buttressing the fight against corruption and striving to improve governance. However, how countries should go about doing this is very much a country-specific issue reflecting the interaction between the various players— government, institutions, civil society, media and the private sector—and the determination of the authorities to credibly advance these reforms, as well as the institutional capacities of the anti-corruption agencies, judicial system, and accountability and oversight entities.

Researchers have shown that there is no “one-size-fits-all” approach to improving governance and reducing corruption (e.g., Klitgaard, 1988; Mungiu-Pippidi and Johnston, 2017). The experience of countries like Botswana, Chile, Estonia, and Georgia, suggests that multiple factors may have contributed to their success, including political will, measures to reduce opportunities of corruption (e.g., lower red tape and trade barriers), and measures to increase constraints on corrupt behavior (e.g., independent judicial system, pressure from civil society) and improve fiscal institutions (greater fiscal transparency and controls). Further study of the transmission channels and the exceptions to the general tendencies could help shed light on what measures would yield the biggest impact.

Consistent with the lessons from the success stories, the process could be facilitated by putting in place independent and credible institutions that improve both transparency and accountability, while allowing for a better monitoring and prosecution of corrupt behavior. Moreover, enforcing laws that are in line with best practices—including those that level the playing field and thus reduce the scope for rent-seeking—and an independent judiciary can support the process. Additionally, there are some basic governance principles that apply across countries—including improving the regulatory quality and government effectiveness and strengthening fiscal institutions—that could deliver significant gains. Irrespective of the specific path countries choose, this paper suggests that the dividends from improving governance and reducing corruption are likely to be significant.

An in-depth examination of how to fight corruption or improve governance, as well as how long it takes to reap the growth gains, is largely beyond the scope of this paper. However, it is addressed to some extent by IMF (2016, 2017, 2018), amongst others.

The rest of this paper is structured as follows. Section II provides a brief literature review; section III looks at some stylized facts; section IV focuses on the empirics using aggregate indicators of governance; section V unbundles governance and looks at specific governance channels; section VI discusses potential non-linearities in the growth-governance (corruption) nexus; and section VII concludes.

II. Brief Literature Review

Multiple channels have been put forward to explain how corruption is detrimental to growth and development. Corruption is often the result of weak governance coupled with distorted economic incentives and weak institutions. Shleifer and Vishny (1993) argue that corruption is the result of multiple government agencies imposing independent bribes on private agents seeking public services (e.g., permits), and is akin to a distortionary tax on the private sector. Hence, corruption tends to divert resources from human capital-enhancing expenditures (e.g., education and health) into less-growth enhancing areas (e.g., defense) or areas with greater opportunity for graft (e.g., public procurement), while discouraging entry, competition, and innovation.

Empirically, Mauro (1995) found that corruption adds to the cost of business and introduces significant uncertainty in the decision-making process. Corruption can also undermine the efficiency of fiscal institutions and lead to sub-optimal budgetary processes and spending compositions, with broader spillovers on the economy. The main channels relate to (also see IMF, 2018):

  • weakened state capacity to mobilize revenue by undermining tax compliance, for instance by avoiding paying taxes through bribes (Ghura, 1998; Baum et al, 2017);

  • distorted composition of government spending towards items that are wasteful/allow for easier capture of rents and bribes (e.g., military spending), at the detriment of investment in human capital such as education and health care and other growth-enhancing spending (Tanzi and Davoodi, 1997; Mo, 2001); and

  • depressed private investment—corruption increases uncertainty and acts as a distortionary tax on private investment (Mauro, 1995), while higher public investment, despite its positive impact on infrastructure, allows for the capture of more bribes and become highly inefficient when management practices are weak (Tanzi and Davoodi, 1997; also see Gupta et al, 2005, for a summary of earlier findings).

Corruption can also set in motion a vicious circle which further exacerbates economic and social vulnerabilities. For instance, by lowering spending on education and health, corruption disproportionately affects the poor who rely more on social services thus entrenching poverty and inequality (Gupta et al, 2002; Hindriks et al, 1998; Johnston, 1989; and Li et al, 2000). Corruption can also contribute to lax lending standards and weak supervision, which can contribute to poor lending practices, and lead to weakened bank portfolios and precipitate financial crises (IMF, 2016, 2018). In extreme cases, corruption exhausts fiscal capacity and the state can lose its legitimacy leading to fiscal and political crises (Mo, 2001). Similarly, disregard for property rights and expropriation of private capital could lead to collapse of investment and job creation, thus fueling social unrest.

The negative impact of corruption on growth is economically relevant. For instance, the elasticities estimated by Gyimah-Brempong (2002) using a sample of African countries imply that a one standard deviation improvement in corruption in Africa would increase per capita income growth by about ¾ percentage point annually. Gyimah-Brempong and de Camacho (2006), estimating a model somewhat like ours that interacts regional dummies with corruption, find that the impact of corruption on growth is stronger in SSA and LAC than in other regions. Ugur and Dasgupta (2011) conduct a meta-analysis of the impact of corruption on growth and find that a one-unit improvement in corruption perceptions— equivalent to a one standard deviation as measured by the CCI and ICRG corruption indicators—would increase annual GDP per capita growth by about 0.6 percentage point in low-income countries (LICs), and almost 0.9 percentage point in a “mixed” sample of LICs and non-LICs.

Poor governance or weaknesses stemming from specific aspects of governance have also been found to affect growth. Mauro (1995) finds a positive correlation between growth and a composite measure of bureaucratic efficiency (corruption, red tape, and efficiency of the judicial system). Kilish et al (2013) find that the quality of government regulation and the rule of law are positively correlated with economic performance in SSA. Weak governance associated with political instability, lack of political pluralism, and deficient public services have been long associated with low capital accumulation and growth in SSA (e.g., Collier and Gunning, 1999; and Ndulu and O’Connel, 1999).

III. Stylized Facts

SSA countries tend to score unfavorably on corruption perceptions and governance. With many SSA countries also among the poorest, there is a natural question as regard causality: does weak governance stunt growth and development, or does lower development hinder countries’ capacity to have the institutions needed to improve governance? In this section, we take a first look at these correlations and explore potential transmission channels at play.

Weaker governance and higher corruption are associated with lower levels of development

The data shows an unconditional positive (negative) correlation between governance (corruption perceptions) and real GDP per capita both in SSA and the rest of the world (Figure 2). The scatterplots confirm that on average SSA countries have weaker governance, higher corruption perceptions, and lower development. These facts would suggest that weak governance and corruption may partly explain the income differential between SSA and other regions.

Figure 2:Governance and Corruption Perceptions and Level of Development

Sources: Kaufmann and Kraay, Penn World Tables, and authors’ calculations. Orange dots denote SSA countries and blue dots denote the rest of the world. Dotted lines are fitted linear trends. Higher CCI scores mean higher levels of perceived corruption.

Weaker governance and higher corruption are associated with lower growth in SSA

The data show a positive (negative) unconditional correlation between governance (corruption perceptions) and growth in the SSA region (Figure 3). These correlations look weaker and even have the opposite sign for the rest of the sample. They also hold for SSA when controlling for income level but are about zero for the other countries. This suggests that weak governance (and corruption perceptions) may be preventing SSA from catching up with the rest of the world.

Figure 3:Governance and Corruption Perceptions and Growth

Sources: Kaufmann and Kraay, Penn World Tables, and authors’ calculations. Orange dots denote SSA countries and blue dots denote the rest of the world. Dotted lines are fitted linear trends.

Higher corruption is associated with worsened fiscal performance

Higher corruption perceptions are negatively associated with education spending and lower quality of public investment in the overall sample and more so in SSA, in line with the literature (Figure 4). Our regression model, presented in the next section, includes schooling and total investment as drivers of growth. Hence, it would capture the indirect impact of corruption on growth through these two variables.

Figure 4:Corruption Perceptions and Fiscal Outcomes

Sources: Kaufmann and Kraay, WEO, Fund staff, and authors’ calculations. In the first panel, orange dots denote SSA, blue dots denote rest of the world. In the second panel, “low”, “medium” and “high” denote the bottom, medium and top third of the distribution of CCI, respectively; green and gray areas are the second and third (from the bottom) quartiles of the sample distribution, respectively; and blue and orange diamonds denote the overall sample and the SSA average, respectively. The measure of public investment efficiency is described in IMF (2015a).

SSA lags behind relative to peers in most granular measures (‘channels’) of governance

Besides corruption, we also investigate the correlation between other granular measures of governance—voice and accountability, government effectiveness, regulatory quality, and rule of law—and GDP per capita growth.

SSA countries have low scores across most dimensions of governance may be suggesting some deeper region-specific driving force(s) at work, perhaps fragile institutions, or perhaps the fact that corruption might have become entrenched in everyday expectations (Figure 5).

Figure 5:Sample Distribution of Governance and Corruption Perceptions

We also note that these ‘governance channels’ are strongly correlated with measures of corruption perceptions (Table 1). Thus, improvements in specific aspects of governance that lead to lower corruption, would also mitigate the fiscal and social costs of corruption.7

Table 1:Correlation Between Governance and Corruption Perceptions in SSA, 1995-2015
ICRGWGIVoice and accountabilityGovernment effectivenessRegulatory qualityRule of lawCCICPI
ICRG1.00
WGI0.93***1.00
Voice and accountability0.81***0.91***1.00
Government effectiveness0.86***0.94***0.81***1.00
Regulatory quality0.84***0.91***0.81***0.91***1.00
Rule of law0.89***0.98***0.83***0.92***0.87***1.00
CCI-0.85***-0.91***-0.75***-0.87***-0.76***-0.91***1.00
CPI-0.79***-0.91***-0.79***-0.88***-0.76***-0.89***0.94***1.00
Sources: ICRG, Kaufmann and Kraay, Transparency International, and authors’ calculations.Note: ICRG = ICRG overall governance indicator, WGI = Worldwide governance indicator, CCI = Control of Corruption Indicator, and CPI = Corruption Perceptions Index.

denotes significant at 1-percent level.

Sources: ICRG, Kaufmann and Kraay, Transparency International, and authors’ calculations.Note: ICRG = ICRG overall governance indicator, WGI = Worldwide governance indicator, CCI = Control of Corruption Indicator, and CPI = Corruption Perceptions Index.

denotes significant at 1-percent level.

Other attributes, e.g., resource intensity, may be linked to higher incidence of corruption

Most SSA countries have weak governance scores and are LICs or resource intensive compared to other regions except for MENAAP and to a less extent LAC (Figure 6).8 As proposed in the literature on “resource curse” (e.g., Sachs and Warner, 1995), countries with abundant natural resources tend to grow slower than resource-scarce economies. The large economic rents generated by resource-rich sectors such as oil—often controlled by state-owned companies subject to political interference—can expose resource-rich countries to higher levels of corruption, especially when institutions are weak (OECD, 2014, 2016).

Figure 6:SSA Sample Attributes

Sources: WEO, and authors’ calculations. The country group description is provided in the Appendix.

IV. Empirical Analysis

A. Identification Strategy

We follow the literature and estimate a standard growth regression, augmented for governance and using an unbalanced panel comprising of 190 countries over the period 1984-2015.9 The baseline specification is the following:

where g is real GDP per capita growth, GOV is governance (or corruption perceptions), SSA is a dummy variable for the SSA region, X is a column vector of country-specific explanatory variables that for now are assumed to be strictly exogenous, τ denotes time-fixed effects, and μ and ν denote unobserved country fixed effect and error term, respectively. Subscripts i = 1, 2, …, N, and t = 1, 2, …, T index country and time, respectively.

Vector X includes controls that have been typically considered by similar studies in the literature: initial GDP per capita, investment, schooling, inflation, and terms of trade. But we also test the robustness of our model to alternative choices.

Because OLS does not account for potential endogeneity of the right-hand side variables, the baseline model was estimated using robust two-step SGMM (Blundell and Bond, 1998; Roodman, 2009). The system includes the level and difference versions of equation (1). We treat investment and lagged growth (included in alternative specifications) as endogenous and consider the remainder (initial income, education, inflation, terms of trade, governance/corruption, and time and regional dummies) as exogenous for the time being. For the differenced equation, the estimator uses as instruments lagged values of endogenous and predetermined variables and current and lagged values of differenced exogenous variables. For the level equation, it uses as instruments lagged values of differenced endogenous and predetermined variables and current and lagged values of exogenous variables.

In the above setup, endogenous regressors are valid instruments if their differences are uncorrelated with the fixed effects and the error term. Hence, for a variable xit we assume that

where each of the three terms on the right-hand side is equal to zero. Since this is equivalent to assuming that the error term is not serially correlated, we test for the absence of second-order serial correlation of the error term. We also report the Hansen test of over-identifying restrictions, of whether the instruments, as a group, are valid.

The use of level and differenced instruments and a two-step estimator increase efficiency. The SGMM estimator is suitable for panel data like ours that has a short time dimension (small T) and a large country dimension (large N) and contains country fixed effects and idiosyncratic errors that are possibly heteroskedastic and correlated within but not across countries. We also conduct additional tests to check for endogeneity of governance and corruption perceptions.

The coefficients of interest are β1 and β2. Both are expected to be positive for governance and negative for corruption perceptions. The stylized facts discussed in Section III suggest that β1 + β2 > β1 (in absolute terms) for SSA, that is, governance (corruption perceptions) could potentially have a stronger marginal effect on growth in the SSA region than in other regions.

Gyimah-Brempong and de Camacho (2006) argue that the likely differentiated impact of weak governance and corruption on growth in SSA (i.e., β2 is statistically significant), and possibly in other regions like LAC and MENAAP that are also perceived to have acute governance and corruption problems, could be related to the way corruption is practiced. In some parts of the world, corruption tends to be centralized and paying the bribe once would suffice to get the needed government service. However, in much of SSA, corruption is decentralized and uncoordinated, a bribe must often be paid at each stage of a transaction and there is no guarantee that the service will be provided even after paying multiple bribes. Therefore, decentralized corruption practices would typically lead to longer delays, higher transaction costs, lower output and hence be more harmful to growth.10

In our view, the interaction term encapsulates in a parsimonious way this feature and the complex set of attributes discussed in Section III that look particularly relevant for the SSA region. However, we also test the robustness of the baseline specification by including some of these attributes directly into the regression model. We note that interaction terms including categorial variables have also been used in the literature in somewhat similar setup as ours (e.g., Gyimah-Brempong and de Camacho, 2006; Naceur et al, 2017). Unlike difference GMM, system GMM can accommodate time-invariant regressors like the SSA dummy and thus allows to identify β3 and adhere to the good practice of including in the regression all constitutive terms of interaction (e.g., Brambor et al, 2006). Furthermore, it does not affect (asymptotically) the other coefficient estimates because all instruments for the equation in levels are assumed to be orthogonal to any time-invariant variable, nor does it affect the moment conditions (Roodman, 2009).11

Lastly, we note that because weak governance and high corruption lead to distorted social and capital spending and worse government policies overall, the covariates gross capital formation, education and inflation already capture some of the indirect impact of governance and corruption on growth. Therefore, coefficients β1 and β2 in our regression model should be interpreted as measuring only the direct effect of governance and corruption on growth, after controlling for human and physical capital, as well as the quality of the macroeconomic policies.

B. Data

The controls in vector X are measured as follows:

  • Initial GDP per capita: real GDP per capita PPP (in log) in the year immediately before each five-year period, to control for income convergence.

  • Gross capital formation: total investment, public and private (percent of GDP), to capture the contribution of capital accumulation to growth. To mitigate endogeneity problems, investment is measured in the first year of each five-year period.

  • Level of education: per capita years of secondary and tertiary schooling for the population aged 15 and above, to capture the contribution of human capital.

  • Dummy variable for high inflation: it takes value 1 if inflation is larger than 15 percent—at the top quartile of the sample distribution—and zero otherwise. It proxies for the quality of macroeconomic policies and macroeconomic volatility.

  • Terms of trade: percentage change in terms of trade, to reflect external shocks.

In our baseline regressions we use two aggregate indicators of governance and two indicators of corruption perceptions:

  • ICRG’s Political Risk Rating (ICRG). It covers several aspects of governance, including government stability, internal and external conflicts, corruption, law and order, ethnic tensions, democratic accountability, and bureaucracy quality. The time dimension (available since 1984) is one advantage of this index but it only covers two thirds of SSA countries.

  • An aggregate measure of the Worldwide Governance Indicators (WGI). This is constructed as the simple sum of the six WGI from Kaufmann and Kraay (see Kaufmann et al, 2010): voice and accountability, political stability and absence of violence/terrorism, government effectiveness, regulatory quality, rule of law, and control of corruption. It covers all SSA countries, but it is available only from 1996 (except for 1997, 1999, and 2001, where observations were interpolated).

  • Control of Corruption (CCI). This indicator of corruption perceptions is one component of WGI, as mentioned above, and is also available from 1996.

  • Transparency International Corruption Perceptions Index (CPI). This measures corruption perceptions from 1995, starting with a relatively small country coverage that was gradually expanded to cover all SSA countries more recently.

These indicators have been normalized to 0-100, with higher values denoting better governance or higher corruption perceptions.12 Despite differences in methodology and sample, these indicators are highly correlated with each other (Table 1).

It is important to note that the measures of governance and corruption perceptions are not without caveats. For instance, subjective measures of corruption perceptions may not fully capture the actual corruption in a country. Moreover, the perception of corruption may have inherent cultural differences. For instance, a gift to a government official may be considered a bribe in one country but may be customary in another. Lastly, the indicators may be measuring distinct attributes of corruption (e.g., petty corruption in one country, grand corruption in another) that may not be readily comparable across-countries or across-regions (for a detailed discussion on these issues, see Gyimah-Brempong and de Camacho, 2006). However, Hamilton and Hammer (2018) argue that the corruption perceptions measured by CCI and CPI are sufficiently comprehensive to capture all elements of corruption and hence are a good starting point of empirical analyses (also see Background Note III of IMF, 2017).

In line with the literature and to reflect the fact that changes in governance do not impact growth immediately, the annual data used in the baseline regressions and most robustness tests is averaged over nonoverlapping five-year periods. Table 2 below and Table A1 in the Appendix provide more details about the data used in the empirical analysis.

Table 2:Summary of Variables and Data Sources
VariableData descriptionData sourcesSample periodExpected sign
Dependent variable
Growth1GDP per capita growth rate, percentPenn World Tables 9.0, WEO1980-15
Right-hand side variables
GDP per capitaMeasured at PPP, logPenn World Tables 9.0, WEO1980-14(-)
InvestmentGross capital formation, percent of GDPPenn World Tables 9.01980-14(+)
Education2Secondary and tertiary education, yearsBarro-Lee database1980-15(+)
Inflation31 if inflation > 15 percent, 0 otherwiseWEO1980-15(-)
Terms of tradeAnnual percent changeWEO and WDI1980-15(+)
ICRGGovernance indicator, 0-100ICRG1984-15(+)
WGIGovernance indicator, 0-100Kaufmann and Kraay1996-15(+)
CCICorruption perceptions, 0-100Kaufmann and Kraay1996-15(-)
CPICorruption perceptions, 0-100Transparency International1995-15(-)

Observations with average annual changes of ±20 percent or more over any five-year period are excluded.

Observations for the 5-year period 2011-15 were obtained by extrapolating Barro-Lee data.

The dummy variable takes value 1 if average annual inflation exceeds 15 percent over a given five-year period, and zero otherwise.

Observations with average annual changes of ±20 percent or more over any five-year period are excluded.

Observations for the 5-year period 2011-15 were obtained by extrapolating Barro-Lee data.

The dummy variable takes value 1 if average annual inflation exceeds 15 percent over a given five-year period, and zero otherwise.

The first block of Table 3 presents summary statistics of the baseline sample data used in the governance regressions (results are similar for the sample based on corruption perceptions). For comparison, the second block depicts the same statistics for SSA. These show that most variables vary widely across countries in both samples, with some of the extreme values coming from SSA. There is also great variation in the governance and corruption perceptions indicators, with SSA having lower (higher) average scores on governance (corruption perceptions) than the overall sample (also see Figure 1). Note that one standard deviation in the SSA sample is just above 10 units for the four normalized indicators.

Table 3:Summary Statistics of Main Variables Used in the Regressions
VariableUnitMeanStd. Dev.Min.Max.MeanStd. Dev.Min.Max.
Overall sample (Obs = 470)SSA sample (Obs = 91)
GrowthPercent2.32.6-11.812.21.82.8-11.88.1
GDP per capitaLog (2011 PPP$)9.21.25.911.97.60.95.99.9
InvestmentPercent of GDP22.18.52.061.017.08.02.046.2
EducationYears3.51.80.18.61.51.00.14.1
InflationPercent6.89.0-6.484.28.210.6-6.484.2
Terms of tradePercent0.53.9-15.918.00.75.3-15.912.4
ICRG0-10067.014.425.596.455.911.125.778.7
WGI0-10055.521.93.299.139.713.53.270.4
CCI0-10053.422.81.995.568.611.935.195.5
CPI0-10057.722.81.991.872.210.641.084.8

C. Baseline Results

Before discussing the baseline, results using SGMM, we present OLS estimates. As expected, we find a positive correlation between governance and growth for the overall sample and, although the coefficient on the interaction term for SSA (β2) is not statistically significant, we reject the null that both coefficients are jointly equal to zero (Table 4).13 However, the OLS results do not account for potential endogeneity with the explanatory variables. Moreover, the sign on schooling contradicts theory and previous empirical evidence. Saha et al (2017) also find similar results for the correlation between corruption perceptions and growth and question the validity of a linear relationship. We address this issue later by testing a non-linear version of model (1).

Table 4:Governance and Growth in SSA: OLS Regressions
Dependent variable:(1)(2)(3)(4)(5)
Real GDP per capita growthNoICRGWGI
governanceNo interactionInteractionNo interactionInteraction
Governance0.087 **0.0540.134 ***0.096 *
Governance × SSA0.1050.120
Initial income per capita (ln)-7.664 ***-7.862 ***-7.871 ***-8.080 ***-7.954 ***
Investment (percent of GDP)0.0330.0310.034 *0.0240.023
Education (years)-0.238-0.227-0.248-0.248-0.230
High inflation dummy-2.956 ***-2.655 ***-2.642 ***-2.438 ***-2.413 ***
Change in terms of trade (percent)0.074 ***0.083 ***0.086 ***0.077 ***0.080 ***
Constant74.25 ***70.48 ***71.63 ***70.94 ***70.87
Observations470470470470470
R-squared119119119119119
Number of countries0.3260.3470.3550.3600.366
F-test for H0: β1 = 0 and β2 = 04.451 **7.722 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using OLS and include country and time fixed effects.The F-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using OLS and include country and time fixed effects.The F-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

Table 5 presents our baseline estimations using system SGMM. All coefficients on the core controls have the expected sign and almost all are statistically significant. The coefficient on the interaction term for SSA (β2) has the positive expected sign and is also statistically significant. The specifications pass the standard tests for absence of serial correlation and validity of instruments.14 We also reject the null hypothesis that the coefficient β1 and β2 are jointly zero.

Table 5:Governance and Growth in SSA: Baseline System GMM Regressions
Dependent variable:(1)(2)(3)(4)(5)
Real GDP per capita growthNoICRGWGI
governanceNo interactionInteractionNo interactionInteraction
Governance0.108 ***0.107 ***0.026 *0.036 **
Governance × SSA0.0890.087
Initial income per capita (ln)-2.031 ***-2.435 ***-2.765 ***-1.929 ***-2.554 ***
Investment (percent of GDP)0.225 ***0.1280.0390.150 ***0.134 ***
Education (years)0.622 ***0.440 ***0.354 **0.478 ***0.361 **
High inflation dummy-1.351 **-1.244 **-1.562 ***-1.901 **-1.760 **
Change in terms of trade (percent)0.0460.069 **0.064 *0.077 **0.080 **
Constant13.49 ***13.27 ***19.46 ***13.27 ***20.00 ***
Observations470470470470470
Number of countries119119119119119
Number of instruments1415171517
Serial correlation test (p-value)0.0030.1410.4560.410.458
Hansen test (p-value)0.0310.3270.2830.220.255
χ2-test for H0: β1 = 0 and β2 = 047.44 ***18.96 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity. The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity. The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

We note that these correlations should not be interpreted literally as causal effects. But mechanically, a 10-point improvement in governance—equivalent to a one standard deviation in the SSA sample and enough to move an SSA country from the bottom quartile to the median of the SSA distribution or bring the average SSA country to the world average— would be associated with higher growth by between 1¼ percentage point (WGI) and 2 percentage points (ICRG). This impact is about thrice and twice larger than for the overall sample, respectively.

As mentioned before, some of the covariates like education and inflation also capture some of the indirect effects of governance on growth. Therefore, the total impact of governance on growth—including the indirect effects through these variables—would be likely larger than β1+ β2.

We also run model (1) using difference GMM, which as explained earlier does not allow to identify the SSA dummy. Like in the OLS estimation, the individual coefficients on governance have the expected sign but are not statistically significant. However, the coefficients are still jointly significant and larger than those obtained from system GMM but are estimated more imprecisely, as reflected by the larger confidence intervals (Figure 7).

Figure 7:Overall Effect of Governance on Growth in SSA

D. Mitigating endogeneity

Lagging governance

Until now we have considered contemporaneous values of governance in equation (1). We lag the governance indicators by one period to further mitigate endogeneity bias. The interaction term has the expected sign and is statistically significant but the overall effect, although statistically significant, is smaller (Figure 8, “Lagged governance” column).

Figure 8:Overall Effect of Governance on Growth in SSA Controlling for Endogeneity

The chart shows β1 + β2 (orange diamonds) and 95-percent confidence intervals.

***, **, * Significant at 1, 5, and 10 percent, respectively.

Instrumental variables

Next, we explore several instrumental variables (IV) to further control for endogeneity. Because most IVs are time-invariant we average the data over the entire sample period and use these cross-sections using averages over 1984-2015 and IV estimations to estimate the correlation between governance and growth. The IVs relate to ethnic fractionalization, settler mortality, independence, and latitude. In all regressions, initial income (i.e. averages over 1980-1983), inflation, education, and terms of trade are treated as exogenous, and governance and investment as endogenous. Figure 8 presents the coefficients of interest and Table A2 summarizes the key estimation details.

Ethnic fractionalization

Mauro (1995) argued that using lagged governance and corruption on the right-hand side may not fully address endogeneity bias because these variables are autocorrelated. We follow Mauro (1995), Easterly and Levine (1997), and others, and use ethnic fractionalization as an instrument for governance. This variable measures the likelihood that two people drawn randomly from a country’s population will not belong to the same ethnolinguistic group. The coefficients of interest have the expected sign and β1 + β2 is statistically significant for both ICRG and WGI (Figure 8, “Ethnic fractionalization” column).

Settler mortality

Acemoglu et al (2001) use settler mortality rate in colonial times as an instrument for expropriation risk. Higher expropriation risk tends to be associated with weaker rule of law and property rights and higher corruption. Hence, settler mortality would be a sensible instrument for governance. In our sample, however, it is informative for WGI only (Figure 8, column “Settler mortality”), with the overall marginal effect statistically significant at 5 percent (Table A2).

Independence

Another instrument is the fraction of years a country has been independent since 1776. This variable has been often used in the literature to measure the extent to which a country has had enough time to develop its own economic institutions (e.g., Easterly and Levine, 1997, 2016; Beck et al, 2003). The coefficients on governance have the expected sign, and the sum for SSA is relatively large and statistically significant for both ICRG and WGI (Figure 8, “Independence” column).

Latitude

We also use the absolute value of latitude, i.e., distance from the equator, as an instrument for governance (e.g., Acemoglu et al, 2001; Easterly and Levine, 2016). This variable has been found to be correlated with good institutions. Easterly and Levine (2016) claim that latitude was especially relevant for European settlers since they might have been more attracted to lands with the same temperate climate as in Europe. As with the other instruments, the coefficients of interest have the right sign, and the overall effect for SSA is statistically significant for both ICRG and WGI.

Other instruments

We also explored other instruments, including incidence of malaria, legal origins, and religious and language fractionalization, which have also been used in the literature as proxies for the quality of institutions. Results using these instruments are weaker and less reliably estimated, including because of small samples.

Summary of endogeneity tests

Our model withstands most tests when sample size is not an issue. Better controlling for potential endogeneity problems, we still find a strong correlation between governance on growth in the SSA region. Moreover, we typically reject the null hypothesis that the coefficients on governance are jointly equal to zero (Table A2). Therefore, the estimated correlation seems to be genuine and not an artifact of the data. Overall, our results point to a broadly robust correlation between governance and growth in SSA and the magnitude of this correlation appears to be larger for SSA than for the full sample.

E. Robustness

Regional benchmarking

In this section, we take a closer look at the link between governance and growth in other regions and compare with the findings for SSA. We split the sample in three distinct ways. First, we run the baseline model using the overall sample but including one region at a time (model 2 below). Second, we use only the samples for each region (model 3). Lastly, we introduce multiple regions in the specification (model 4). We thus generalize our baseline model as follows:

where REGj is a dummy variable for region j. In line with the discussion in subsection B, we consider three relevant regional groupings: non-SSA developing and emerging countries (Non-SSA), LAC, and MENAAP. Table A1 in the Appendix describes all country groups.

Figure 9 and Table A3 present the results for models (2) and (3), while Table A4 shows the results for model (4). Most right-hand side controls have the expected sign and are statistically significant, and most models pass the standard statistical tests for autocorrelation and validity of instruments. Starting with model 2, we find a strong correlation between governance and growth for all regions considered. The evidence based on model (3) broadly holds but the precision of the estimates is affected by small sample sizes.

Figure 9:Overall Effect of Governance on Growth Across Regions

The charts show β1 + β2 (orange diamonds) and 95-percent confidence intervals. ***, **, * Significant at 1, 5, and 10 percent, respectively. LAC = Latin American and and Caribbean; MENAAP = Middle East, North African and Afghanistan and Pakistan; Non-SSA = Developing and emerging non Sub-Saharan African countries.

The evidence for model (4) is somewhat mixed. Except for SSA, the coefficient on the interaction term (β2SSA) is not statistically significant but it is for LAC and MENAAP when jointly considered with SSA in the case of WGI. However, we reject the null that all coefficients on governance are jointly equal to zero for both governance indicators. Overall, these findings suggest that weak governance also appears to hinder growth in other regions, but apparently to a lower extent than in SSA.

We also run two additional specifications that control for income group and quality of institutions to capture additional cross-country differences:

where GROUPk is a dummy variable for group k. We consider the following country groups: LICs, LICs in SSA, LOW (countries with governance scores below the sample median), and LOW in SSA.

The overall effect of governance on growth is statistically strong across in almost all country groups and even more so for LICs in SSA or SSA countries with below-median governance (Figure 10 and Table A5).

Figure 10:Overall Effect of Governance on Growth Across Country Groups

The charts show β1 + β2 (orange diamonds) and 95-percent confidence intervals. ***, **, * Significant at 1, 5, and 10 percent. LIC = low-income country, LOW = countries with governance below the sample median.

Alternative control variables

We investigate whether the baseline model survives the inclusion of alternative controls. We test for lagged growth, the level of inflation, the inflation dummy with a higher (20 percent) threshold, resource intensity, conflict, agriculture intensity, openness, and government consumption. Results are shown in Figure 11 and Table A6 in the Appendix. The specifications pass the tests for serial correlation and validity of instruments, and the overall effect remains significant and with the expected sign in all specifications.

Figure 11:Overall Effect of Governance on Growth in SSA Controlling for Other Factors

The chart shows β1 + β2 (orange diamonds) and 95-percent confidence intervals. ***, **, * Significant at 1, 5, and 10 percent, respectively.

Other robustness tests

We also tested the sensitivity of the baseline model to the choice of the averaging period. We considered 3-year instead of 5-year periods, and yearly data. The overall effect for SSA remains statistically significant at conventional significance levels and with the right sign in most cases (Figure 12 and Tables A7-A8).

Figure 12:Overall Effect of Governance on Growth Using 3-Year and 1-Year Averages

The chart shows β1 + β2 (orange diamonds) and 95-percent confidence intervals. ***, **, * Significant at 1, 5, and 10 percent, respectively.

LIC = low-income country, LOW = countries with governance below the sample median.

Finally, we also checked whether outliers could be driving our baseline results. We run model (1) for ICRG and WGI using SGMM and dropping one country at a time (2 x 119 countries = 238 regressions). The model remained remarkably stable and the overall effect was significant in all regressions. Thus, no single country is driving the baseline results.

V. Unbundling Governance

So far, we have shown the correlation between aggregate measures of governance and growth. In this section, we unpack the governance indicators and analyze the correlation between their main components and growth. We devote special attention to corruption given its large economic and social costs (see Section I). Analyzing the components of governance may shed light on the key transmission channels through which governance affects growth. Therefore, we interpret these granular measures as ‘governance channels’.

To start, we construct narrower variants of ICRG and WGI by excluding corruption perceptions—which in this section we will treat as a standalone measure of governance. We also exclude economic and social outcomes to mitigate endogeneity, political stability, and a few other components. Our measure of inflation already captures the political channel to some extent. For instance, inflation would likely be higher when there is high turnover of central bank officials reflecting weaknesses in monetary policy framework such as political interference, lack of central bank independence, and limited technical capacity.15

The narrower version of ICRG thus includes: military in politics, religions tensions, ethnic tensions, law and order, democratic accountability, and bureaucracy quality. The narrower WGI includes: voice and accountability (freedom of participation and expression, and free media), government effectiveness (quality of public policies and services), regulatory quality (ability of the government to formulate and implement business-friendly policies and regulations), and rule of law (respect for contract enforcement, property rights and law enforcement). In a nutshell, we see these components as proxies for the quality of political institutions and government regulations and policies. These measures are appropriately rescaled to reflect their weight in the original governance indicator.

The main properties of the estimated models based on these narrow measures (Table 6) are similar to their baseline counterparts (Table 5). The overall impact of narrower measures of governance on growth in SSA remains strong but the magnitude is slightly smaller than in the baseline. Because the exclusion of corruption would probably increase the correlation between the narrower measures of governance and growth (corruption is negatively correlated with governance, see Table 1) the smaller estimates are probably related to the exclusion of the other components of governance. However, these correlations though smaller are also less subject to concerns about endogeneity, as mentioned above.

Table 6:Alternative Measures of Governance: System GMM Regressions
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthICRGWGI
BaselineNarrowBaselineNarrow
Governance0.108 ***0.107 ***0.043 ***0.031 ***0.026 *0.036 **0.0220.027 *
Governance × SSA0.089 **0.177 **0.087 ***0.085 ***
Initial income per capita (ln)-2.435 ***-2.765 ***-1.958 ***-2.201 ***-1.929 ***-2.554 ***-1.884 ***-2.432 ***
Investment (percent of GDP)0.1280.0390.1140.0240.150 ***0.134 ***0.151 ***0.134 ***
Education (years)0.440 ***0.354 **0.472 ***0.460 ***0.478 ***0.361 **0.478 ***0.389 **
High inflation dummy-1.244 **-1.562 ***-1.604 ***-1.797 ***-1.901 **-1.760 **-1.951 **-1.879 **
Change in terms of trade (percent)0.069 **0.064 *0.054 *0.0450.077 **0.080 **0.078 **0.079 **
Constant13.27 ***19.46 ***13.21 ***18.96 ***13.27 ***20.00 ***13.05 ***19.20 ***
Observations470470470470470470470470
Number of countries119119119119119119119119
Number of instruments1517151715171517
Serial correlation test (p-value)0.1400.4600.1600.5500.4100.4600.4200.470
Hansen test (p-value)0.3300.2800.1800.2300.2200.2600.2100.230
χ2-test for H : β = 0 and β = 047.44 ***16.07 ***18.96 ***19.92 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

Table 7 shows that the WGI’s four selected channels have a strong correlation with growth in SSA. The overall effect for SSA is in the range 0.07-0.1, comparable to that for the overall and narrow measures of the WGI (Table 6).16 Because the four channels have similar weight on the WGI, these results suggest that improving even only one aspect of governance could be associated with higher growth. In practice, improving multiple aspects of governance could potentially yield larger payoffs, including because of complementarities across them (see Table 1).

Table 7:Governance Channels and Growth: System GMM Regressions
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthVoice and AccountabilityGovernment EffectivenessRegulatory QualityRule of Law
Governance0.0120.0040.036 **0.046 ***0.0340.036 *0.0060.017
Governance × SSA0.099 ***0.074 ***0.085 ***0.072 ***
Initial income per capita (ln)-1.713 ***-2.095 ***-2.079 ***-2.707 ***-1.973 ***-2.489 ***-1.705 ***-2.296 ***
Investment (percent of GDP)0.147 **0.119 **0.155 ***0.148 ***0.159 **0.145 ***0.145 **0.129 **
Education (years)0.504 ***0.495 ***0.464 ***0.338 *0.464 ***0.380 **0.534 ***0.434 **
High inflation dummy-2.062 **-2.130 **-1.914 **-1.835 **-1.854 *-1.796 **-2.062 **-2.014 **
Change in terms of trade (percent)0.073 **0.070 **0.081 **0.086 **0.079 **0.082 **0.074 **0.080 **
Constant11.98 ***17.15 ***14.16 ***20.74 ***12.99 ***18.84 ***12.19 ***18.44 ***
Observations470470470470470470470470
Number of countries119119119119119119119119
Number of instruments1517151715171517
Serial correlation test (p-value)0.4300.5700.4200.5100.3700.3700.4300.470
Hansen test (p-value)0.1700.2000.2600.2800.2100.2100.1600.200
χ2-test for H0: β1 = 0 and β2 = 015.75 ***23.91 ***17.41 ***15.15 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

We now turn our attention to corruption. We estimate model (1) using CCI and CPI. The results for the core part of the model (Table 8) are roughly in line with those for the governance indicators. The specifications for corruption perceptions also pass the tests for serial correlation and validity of instruments, as well as most of the robustness tests (available upon request) that we run for the governance models.

Table 8:Corruption Perceptions and Growth in SSA: System GMM Regressions
Dependent variable:(1)(2)(3)(4)(5)(6)(7)
Real GDP per capita growthNo corruptionWGI’s CCITransparency International’s CPI
No interactionInteractionJointNo interactionInteractionJoint
Governance-0.012-0.020 *-0.002-0.002-0.0120.0046
Corruption × SSA-0.093 ***-0.112 ***-0.090 ***-0.106 ***
Governance × LAC-0.037-0.031
Governance × MENAAP-0.073 **-0.057 **
Initial income per capita (ln)-1.654 ***-1.607 ***-2.118 ***-1.942 ***-1.420 ***-1.863 ***-1.658 ***
Investment (percent of GDP)0.1170.096 *0.096 *0.094 **0.0940.0610.053
Education (years)0.594 ***0.495 ***0.417 ***0.270 *0.473 ***0.391 ***0.237 *
High inflation dummy-1.249 *-1.525 *-1.567 **-1.396 *-1.278 *-1.498 **-1.503 **
Change in terms of trade (percent)0.0130.0200.0200.0220.0130.0190.021
Constant12.43 ***13.54 ***19.41 ***17.97 ***11.41 ***17.60 ***16.04 ***
Observations489489489489489489489
Number of countries136136136136136136136
Number of instruments14151721151721
Serial correlation test (p-value)0.0010.3720.3690.3350.0620.0390.040
Hansen test (p-value)0.0530.2790.2500.2030.1490.2350.228
χ2-test for H0: β1 = 0 and β2 = 011.63 ***19.37 ***8.71 **16.5 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

Unlike the findings for governance, we do not find supporting evidence of a statistically significant correlation between corruption perceptions and growth for the average country in the overall sample. However, we do find evidence of a statistically strong correlation between corruption perceptions and growth for the MENAAP region (columns 4 and 7), consistent with the stylized facts discussed before. Moreover, both corruption perceptions indicators point to a stronger and negative correlation with economic performance in SSA. The results do not show a differentiated impact for LAC though.

While these correlations should not be interpreted literally as causation, a 10-point improvement in corruption perceptions or about one standard deviation in the SSA sample— enough to move an SSA country from the bottom quartile to the median of the SSA distribution or bring the average SSA country to the world average—would be associated with higher growth by 0.4-0.6 percentage points in the long run. These magnitudes are comparable to those estimated by Ugur and Dasgupta (2011).

VI. Non-Linearity

So far, we have made two key assumptions on the link between governance (or corruption perceptions) and growth: the relationship is linear, and the expected correlation is unambiguously positive (negative) for governance (corruption perceptions). These assumptions have been debated in the literature. For instance, the ‘grease the wheels’ view argues that some amount of corruption could increase bureaucratic efficiency and improve growth by mitigating red tape in developing countries.

Cerqueti et al (2012) show in a game theory setup that a non-linear relationship between corruption and growth could arise depending on the degree of ethnic fragmentation in a country.17 Several empirical studies have also found a non-linear relationship between corruption and growth, including Swaleheen (2011), Saha and Gounder (2013), and Saha et al (2017). These studies typically find that a quadratic function fits the data well.

Accordingly, to test for a non-linear correlation between governance (or corruption perceptions) and growth we add a squared term as an additional explanatory variable in our baseline model:

The coefficients of interest are β1 and β2, with the expected signs being positive and negative, respectively. The marginal impact of governance (or corruption perceptions) on growth is β1 + 2β2GOVit and the level of governance (or corruption perceptions) above which growth declines is given by GOVit*=β1/2β. For parsimony, model (7) is estimated for the full and SSA samples separately.

Table 9 shows the results using the ICRG indicator of governance and the CCI indicator of corruption perceptions for both samples.18 With a squared term and splitting the samples, some of the core regressors are estimated more imprecisely, even though all specifications pass the standard tests for validity of instruments and serial autocorrelation. As expected, the coefficient on the linear term (β1) is positive and coefficient on the squared term (β2) is negative. Moreover, in all models but one we reject the null that these two coefficients are jointly equal to zero.

Table 9:Governance and Corruption Perceptions: Non-linear Models
Dependent variable:Governance (ICRG)Corruption (CCI)
Real GDP per capita growthFull sampleSSA sampleFull sampleSSA sample
Governance or Corruption0.284 ***0.743 **0.057 *0.442 **
Governance or Corruption squared-0.0017 ***-0.0053 *-0.0006 *-0.0040 **
Initial income per capita (ln)-1.831 ***-2.326 **-1.498 ***-1.785 **
Investment (percent of GDP)0.0880.156 ***0.092 *0.251 *
Education (years)0.554 ***0.8950.526 ***0.556
High inflation dummy-1.255 **-1.100-1.523 *-1.680
Change in terms of trade (percent)0.0480.0610.019-0.043
Constant4.560-9.11310.95 ***-1.811
Observations47091489119
Number of countries1192313630
Number of instruments16161616
Serial correlation test (p-value)0.4000.2200.3500.740
Hansen test (p-value)0.1300.5200.2400.380
χ2-test for H0: β1 = 0 and β2 = 016.21 ***16.50 ***3.57412.17 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

Figure 13 plots the overall marginal effect on growth. For any given level of governance, the effect for SSA is positive and stronger relative to the overall sample. Comparing the non-linear and linear estimates suggests that imposing linearity could lead to underestimation of the correlation between governance and growth for countries with low governance (governance is below 60 for over 75 percent of SSA country-years in the sample). Similarly, corruption seems to be particularly harmful to growth for countries with corruption perceptions above 60 (about 75 percent of all SSA country-years in the sample). The impact of corruption perceptions on growth would be less of a problem when corruption is moderate or small, in line with the findings of Saha et al (2017).

Figure 13:Overall Effect of Governance and Corruption Perceptions on Growth

The non-linear models suggest that growth gains for SSA countries could be larger than those implied by the linear model, especially for SSA countries with very low (high) governance (corruption perceptions) scores. For instance, the overall effect for a country at 5th (95th) percentile of the SSA distribution of governance (corruption perceptions)—equivalent to normalized scores around 30 and 85 respectively—would be about twice as large as that implied by the linear model.

Weak performers in the overall sample also stand to benefit from improving governance but the size of the growth gains is smaller compared to that for SSA countries. In the case of corruption, these gains would be around zero irrespective of a country’s starting position.

VII. Concluding Remarks

In this paper, we have found a strong correlation between governance/corruption perceptions and growth in sub-Saharan Africa. This finding remains largely robust to various specifications and sample choices, notwithstanding concerns about potential endogeneity problems, which we have sought to mitigate throughout. While these correlations do not necessarily imply causation, they suggest that there could potentially be significant economic gains to be achieved by improving governance and reducing corruption in SSA. While a demanding process, most SSA countries, particularly SSA LICs and those with the lowest performance in governance and perceived corruption, would likely gain the most from these improvements.

We expect the focus on reducing corruption and improving governance to remain high on the agenda of policymakers, particularly in SSA, as illustrated by the high-level discussions in the 2018 African Union Summit. Strengthening governance and fighting corruption are not easy tasks—and the process is often time-consuming and requires considerable political effort—but they are worth pursuing given the potentially large payoffs.

Appendix

Sample period

The baseline regressions are based on data spanning 20 years (1995–2015) from several commonly used macroeconomic databases. To mitigate the issue of data availability, the data used in the baseline regressions in annual frequency was averaged out to five (nonoverlapping) five-year periods.

Country coverage

The raw sample covers 190 countries, including advanced and emerging economies, and low-income countries. However, the regression samples are somewhat smaller depending on data availability and econometric specification. The main country groups are listed below, where SSA, EMDE, MENAAP, CIS, LIC stand for sub-Saharan Africa, emerging and developing, Middle East North African and Afghanistan and Pakistan, Commonwealth of Independent States, and Low-Income Country, respectively.

Table A1.Country and Regional Samples
CountryISO codeIFS codeSSASSA LICResource-RichLIC
AngolaAGO61411
BeninBEN638111
BotswanaBWA61611
Burkina FasoBFA7481111
BurundiBDI6181111
Cabo VerdeCPV6241
CameroonCMR6221111
Central African RepublicCAF6261111
ChadTCD6281111
ComorosCOM632111
Congo, Dem. Rep. of theCOD6361111
Congo, Republic ofCOG6341111
Côte d’IvoireCIV6621111
Equatorial GuineaGNQ64211
EritreaERI6431111
EthiopiaETH644111
GabonGAB64611
Gambia, TheGMB648111
GhanaGHA6521111
GuineaGIN6561111
Guinea-BissauGNB6541111
KenyaKEN664111
LesothoLSO666111
LiberiaLBR6681111
MadagascarMDG674111
MalawiMWI6761111
MaliMLI6781111
MauritiusMUS6841
MozambiqueMOZ688111
NamibiaNAM7281
NigerNER692111
NigeriaNGA6941111
RwandaRWA714111
São Tomé and PríncipeSTP716111
SenegalSEN722111
SeychellesSYC7181
Sierra LeoneSLE7241111
South AfricaZAF19911
South SudanSSD733111
SwazilandSWZ7341
TanzaniaTZA738111
TogoTGO742111
UgandaUGA746111
ZambiaZMB7541111
ZimbabweZWE698111
CountryISO codeIFS codeEMDE AmericaResource-RichLIC
Antigua and BarbudaATG3111
ArgentinaARG21311
Bahamas, TheBHS3131
BarbadosBRB3161
BelizeBLZ3391
BoliviaBOL218111
BrazilBRA2231
ChileCHL22811
ColombiaCOL23311
Costa RicaCRI2381
DominicaDMA3211
Dominican RepublicDOM2431
EcuadorECU24811
El SalvadorSLV2531
GrenadaGRD3281
GuatemalaGTM2581
GuyanaGUY33611
HaitiHTI26311
HondurasHND268111
JamaicaJAM3431
MexicoMEX27311
NicaraguaNIC27811
PanamaPAN2831
ParaguayPRY28811
PeruPER29311
St. Kitts and NevisKNA3611
St. LuciaLCA3621
St. Vincent and the GrenadinesVCT3641
SurinameSUR36611
Trinidad and TobagoTTO36911
UruguayURY29811
VenezuelaVEN29911
CountryISO codeIFS codeEMDE Europe & CISResource-RichLIC
AlbaniaALB9141
Bosnia and HerzegovinaBIH9631
BulgariaBGR9181
CroatiaHRV9601
HungaryHUN9441
KosovoXKX9671
Macedonia, FYRMKD9621
Montenegro, Rep. ofMNE9431
PolandPOL9641
RomaniaROU9681
SerbiaSRB9421
TurkeyTUR1861
BelarusBLR9131
MoldovaMDA92111
RussiaRUS92211
UkraineUKR9261
ArmeniaARM9111
AzerbaijanAZE91211
GeorgiaGEO9151
KazakhstanKAZ91611
Kyrgyz RepublicKGZ91711
TajikistanTJK92311
TurkmenistanTKM92511
UzbekistanUZB927111
CountryISO codeIFS codeAdvancedResource-RichLIC
AustraliaAUS19311
Hong Kong SARHKG5321
JapanJPN1581
KoreaKOR5421
New ZealandNZL1961
SingaporeSGP5761
Taiwan Province of ChinaTWN5281
AustriaAUT1221
BelgiumBEL1241
CyprusCYP4231
Czech RepublicCZE9351
DenmarkDNK1281
EstoniaEST9391
FinlandFIN1721
FranceFRA1321
GermanyDEU1341
GreeceGRC1741
IcelandISL1761
IrelandIRL1781
IsraelISR4361
ItalyITA1361
LatviaLVA9411
LithuaniaLTU9461
LuxembourgLUX1371
MaltaMLT1811
NetherlandsNLD1381
NorwayNOR14211
PortugalPRT1821
San MarinoSMR1351
Slovak RepublicSVK9361
SloveniaSVN9611
SpainESP1841
SwedenSWE1441
SwitzerlandCHE1461
United KingdomGBR1121
CanadaCAN15611
United StatesUSA1111
CountryISO codeIFS codeEMDE MENAAPResource-RichLIC
SomaliaSOM72611
AfghanistanAFG512111
AlgeriaDZA61211
BahrainBHR41911
DjiboutiDJI61111
EgyptEGY4691
IranIRN42911
IraqIRQ43311
JordanJOR4391
KuwaitKWT44311
LebanonLBN4461
LibyaLBY67211
MauritaniaMRT682111
MoroccoMAR6861
OmanOMN44911
PakistanPAK5641
QatarQAT45311
Saudi ArabiaSAU45611
SudanSDN732111
SyriaSYR46311
TunisiaTUN7441
United Arab EmiratesARE46611
YemenYEM474111
CountryISO codeIFS codeEMDE AsiaResource-RichLIC
BangladeshBGD51311
BhutanBTN51411
Brunei DarussalamBRN51611
CambodiaKHM52211
ChinaCHN9241
FijiFJI8191
IndiaIND5341
IndonesiaIDN53611
Lao P.D.R.LAO54411
MalaysiaMYS5481
MaldivesMDV5561
Marshall IslandsMHL86711
MicronesiaFSM8681
MongoliaMNG948111
MyanmarMMR51811
NepalNPL55811
PalauPLW5651
Papua New GuineaPNG853111
PhilippinesPHL5661
SamoaWSM8621
Solomon IslandsSLB813111
Sri LankaLKA5241
ThailandTHA5781
Timor-LesteTLS53711
TongaTON8661
TuvaluTUV86911
VanuatuVUT8461
VietnamVNM58211
KiribatiKIR82611

Mitigating endogeneity problems

Table A2.Overall Effect of Governance on Growth Controlling for Endogeneity
Dependent variable:(1)(2)(3)(4)(5)
Real GDP per capita growthLaggedEthnicSettler MortalityIndependenceLatitude
ICRG
z-test for H0: β1 + β2 = 00.079 **0.093 **0.1990.111 ***0.078 ***
χ2-test for H0: β1 = 0 and β2 = 020.61 ***5.272 *1.48530.14 ***16.46 ***
Observations460115115115132
Number of countries119115115115132
Data structurePanelCross-sectionCross-sectionCross-sectionCross-section
Estimation methodSGMMIVIVIVIV
WGI
z-test for H0: β1 + β2 = 00.059 **0.173 **0.147 **0.141 ***0.078 ***
χ2-test for H0: β1 = 0 and β2 = 06.411 **10.65 ***5.233 *17.19 ***16.46 ***
Observations407132132132132
Number of countries136132132132132
Data structurePanelCross-sectionCross-sectionCross-sectionCross-section
Estimation methodSGMMIVIVIVIV

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The system GMM (SGMM) regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instrumental variables (IV) regressions are estimated using robust two-stage least square estimator.The z-test shows the sum of coefficients and significance level from testing the null hypothesis that the sum of the coefficients on governance is zero.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The system GMM (SGMM) regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instrumental variables (IV) regressions are estimated using robust two-stage least square estimator.The z-test shows the sum of coefficients and significance level from testing the null hypothesis that the sum of the coefficients on governance is zero.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

Benchmarking against regions and country groups

Table A3.Overall Effect of Governance on Growth Across Regions (Models 2 and 3)
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthSSANon-SSALACMENAAPSSANon-SSALACMENAAP
ICRG Model (2)ICRG Model (3)
z-test for H0: β1 + β2 = 00.196 ***0.126 ***0.123 **0.128 ***0.194 ***0.096 ***0.101 **0.054
χ2-test for H0: β1 = 0 and β2 = 047.44 ***58.46 ***27.89 ***29.07 ***16.74 ***13.54 ***5.261 **2.200
Observations470470470470912438462
Number of countries11911911911923622116
Number of instruments1717171715151515
Serial correlation test (p-value)0.4560.2750.1290.1270.5970.8110.1310.438
Hansen test (p-value)0.2830.3440.3220.3440.3280.2800.0620.186
WGI Model (2)WGI Model (3)
z-test for H0: β1 + β2 = 00.123 ***0.063 ***0.068 **0.074 **0.16 ***0.063 ***0.0280.097
χ2-test for H0: β1 = 0 and β2 = 018.96 ***19.92 ***5.858 *6.835 **8.561 ***10.74 ***0.7031.967
Observations4704704704701192829266
Number of countries11911911911930722317
Number of instruments1717171715151515
Serial correlation test (p-value)0.4580.3670.3860.3640.5520.4960.3470.882
Hansen test (p-value)0.2550.2610.2120.2470.3680.1290.0190.035
The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The z-test shows the sum of coefficients and significance level from testing the null hypothesis that the sum of the coefficients on governance is zero.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.
The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The z-test shows the sum of coefficients and significance level from testing the null hypothesis that the sum of the coefficients on governance is zero.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.
Table A4.Overall Effect of Governance on Growth Across Regions (Model 4)
Dependent variable:(1)(2)(3)(4)(5)(6)
Real GDP per capita growthICRGWGI
SSANon-SSALAC & MENAAPSSANon-SSALAC & MENAAP
Governance0.107 ***0.163 ***0.097 ***0.036 **0.054 ***0.008
Governance × SSA0.089 **0.099 **0.087 ***0.107 ***
Governance × Non-SSA0.0260.005
Governance × LAC0.0290.088 **
Governance × MENAAP0.0290.089 ***
Observations470470470470470470
Number of countries119119119119119119
Number of instruments171921171921
Serial correlation test (p-value)0.4560.3290.4080.4580.3850.380
Hansen test (p-value)0.2830.3430.2540.2550.2420.164
χ2-test for H0: β1 = 0 and β2SSA = 047.44 ***18.96 ***
χ2-test for H0: β1 = 0 and β2NSSA = 053.25 ***23.16 ***
χ2-test for H0: β1 = 0 and β2SSA = 0 and β2LAC = 0 and β2MENAAP = 038.47 ***26.34 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.
Table A5.Overall Effect of Governance on Growth Across Groups (Models 5 and 6)
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthSSALICSSA × LICLOWSSA × LOWSSALICSSA × LICLOWSSA × LOW
ICRG Model (5)ICRG Model (6)
z-test for H0 : β1 + β2 = 00.197 ***0.142 ***0.168 ***0.132 ***0.216 ***0.194 ***0.143 *0.195 ***0.115 **0.243 ***
χ2-test for H0: β1 = 0 and β2 = 047.44 ***12.61 ***26.16 ***14.05 ***26.59 ***16.74 ***3.123 *14.68 ***13.33 ***40.25 ***
Observations470470470470470911187518068
Number of countries1191191191191192330195920
Number of instruments17171717171515151717
Serial correlation test (p-value)0.4560.3470.3990.1270.3130.5970.5540.0580.5360.130
Hansen test (p-value)0.2830.2070.2490.1730.1760.3280.06470.2560.4720.705
WGI Model (5)WGI Model (6)
z-test for H0 : β1 + β2 = 00.123 ***0.0420.109 ***0.072 *0.128 ***0.16 ***0.109 *0.205 ***0.120 **0.172 ***
χ2-test for H0: β1 = 0 and β2 = 018.96 ***1.19315.02 ***4.04515.92 ***8.561 ***3.222 *16.34 ***9.502 ***10.530 ***
Observations4704704704704701191619525990
Number of countries1191191191191193041247324
Number of instruments17171717171515151717
Serial correlation test (p-value)0.4580.4380.4530.3760.4730.5520.433 4.06e-050.5160.369
Hansen test (p-value)0.2550.1310.2090.1840.2010.3680.1580.3020.2720.180
The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The z-test shows the sum of coefficients and significance level from testing the null hypothesis that the sum of the coefficients on governance is zero.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.
The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The z-test shows the sum of coefficients and significance level from testing the null hypothesis that the sum of the coefficients on governance is zero.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

Testing additional control variables

The overall effect of governance on growth for SSA survives the inclusion of additional controls and remains significant and with the expected sign (Table A6). First, we include lagged growth in equation (1) and find weak evidence of persistence. The coefficient on lagged growth is positive but its significance is weak (Table A6, column 1). Next, we replace the inflation dummy by the level of inflation (column 2) and by a dummy that considers a higher (20 percent) threshold for inflation (column 3). The coefficient on inflation in both cases has the right sign and is statistically significant.

Table A6.Overall Effect of Governance on Growth Controlling for Additional Factors
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthLagged ConflictInflation LevelInflation DummyResource IntensityConflictAgriculture IntensityOpennessGovernment Consumption
ICRG
Governance0.101 ***0.107 ***0.108 ***0.070 ***0.114 ***0.112 ***0.105 ***0.108 ***
Governance × SSA0.068 *0.0790.084 *0.129 *0.0750.092 **0.091 **0.091 **
Alternative control variable0.107 *-0.041 ***-1.942 ***-9.38 **0.389-0.106 ***0.0020.020
Observations469468470470362451470470
Number of countries11911911911992117119119
Number of instruments1717171818181818
Serial correlation test (p-value)0.9410.6360.5290.6680.8270.1010.4680.442
Hansen test (p-value)0.6240.1610.1430.010.3430.0110.2820.276
χ2-test for H0: β1 = 0 and β2 = 045.59 ***31.62 ***39.75 ***19.95 *28.06 ***25.4 ***46.68 ***48.43 ***
WGI
Governance0.034 ***0.036 **0.039 **0.0170.051 **0.0280.035 **0.036 **
Governance × SSA0.072 ***0.079 ***0.078 ***0.077 **0.084 ***0.087 ***0.088 ***0.087 ***
Alternative control variable0.107-0.051 **-2.044 *-4.89 ***-1.068-0.0560.0010.022
Observations469468470470362451470470
Number of countries11911911911992117119119
Number of instruments1717171818181818
Serial correlation test (p-value)0.7160.4160.4770.4690.6980.6100.4670.431
Hansen test (p-value)0.3790.1570.1300.0640.3510.0010.2530.220
χ2-test for H0: β1 = 0 and β2 = 021.08 ***12.51 ***17.58 ***7.482 **23.07 ***12.3 ***18.87 ***17.8 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

We also test the hypothesis that SSA countries with abundant natural resources tend to grow more slowly than natural-resource-scarce economies (column 4). Perhaps the governance indicator in our regressions is capturing differences in natural resource intensity rather than differences in institutions. We therefore test this conjecture by interacting governance with a dummy variable for natural resource intensity. We identify resource-rich countries where 20 percent of exports or more are comprised by primary products such as oil, gas, minerals and agricultural commodities.

Controlling for resource intensity does not change the baseline results and the correlation between governance and growth remains statistically significant for SSA (column 5). The coefficient on the dummy for resource intensity is negative, suggesting that undiversified economies grow less rapidly in the long run.

Another possibility is that our indicators of governance may be instead capturing deeper problems such as state fragility or incomplete state formation. We test this conjecture by controlling for the extent to which these countries have been exposed or experienced armed conflicts in the past, using data on conflicts from the UCD/PRIO Armed Conflict Dataset v4 2015. Africa has a disproportionate number of fragile states, or half of the countries identified by the Fund as fragile states (IMF 2015b). Controlling for conflicts, which are known to undermine basic institutions such as the rule of law, we continue to find a statistically valid correlation between governance and SSA growth (column 5).

We also test the robustness of the model by including other traditional controls such as share of agriculture in value added, trade openness, and government consumption. The baseline results remain virtually unchanged (columns 6 to 8).

Testing alternative sample frequencies

Using 3-year averages allows to nearly double the number of observations. Most controls remain statistically significant and with the expected sign (not shown). Like in the baseline model, the overall effect of governance on growth is strong in the full sample and more so in the SSA sample regardless of how we slice the data (Table A7). Using annual data yields similar result for SSA but not for the overall sample (Table A8).19

Table A7.Overall Effect of Governance on Growth Using 3-Year Averages
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthICRGWGI
ALLSSASSA × LICSSA × LOWALLSSASSA × LICSSA × LOW
Governance0.103 ***0.099 ***0.104 ***0.050 ***0.030 *0.038 **0.038 **0.032 ***
Governance × SSA0.108 ***0.108 **0.149 ***0.071 **0.0730.115 ***
Observations822822822822822822822822
Number of countries119119119119119119119119
Number of instruments1618181816181818
Serial correlation test (p-value)0.01930.01920.02210.01870.03960.03850.03840.0370
Hansen test (p-value)0.3160.3920.4170.2530.6850.7410.6950.730
χ2-test for H0: β1 = 0 and β2 = 029.73 ***28.95 ***19.12 ***23.30 ***17.73 ***24.11 ***

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

These results should be taken with caution because we could not completely remove the presence of serial autocorrelation. In our view, using 5-year averages seems to be a reasonable choice given the data available, and it is also in line with the practice in the literature (e.g., and Gyimah-Brempong and de Camacho, 2006; Heckelman and Powell, 2010; Swaleheen, 2011; and Saha et al, 2017).

Table A8.Overall Effect of Governance on Growth Using Annual Data
Dependent variable:(1)(2)(3)(4)(5)(6)(7)(8)
Real GDP per capita growthICRGWGI
ALLSSASSA × LICSSA × LOWALLSSASSA × LICSSA × LOW
Governance-0.02 *-0.02 *-0.02-0.03-0.01-0.00-0.00-0.01 *
Governance × SSA0.073 ***0.070 **0.111 ***0.065 ***0.066 **0.078 **
Observations2,1972,1972,1972,1972,1972,1972,1972,197
Number of countries119119119119119119119119
Number of instruments3133333330323232
Serial correlation test (p-value)0.7790.7870.7990.7810.2680.3300.2690.281
Hansen test (p-value)0.001070.1000.1060.1350.005720.1100.1270.136
χ2-test for H0: β1 = 0 and β2 = 09.736 ***6.609 **13.99 ***8.174 **4.4575.674 *

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.

significant at 1, 5, and 10 percent, respectively. P-values are estimated with robust standard errors.

The regressions are estimated using robust two-step system GMM estimator and include country and time fixed effects.The instruments are the right-hand side controls shown above, lagged values of the dependent variable, and a dummy for resource intensity.The serial correlation test is for second-order serial correlation and the Hansen test is for overidentifying restrictions.The χ2-test shows the test statistics and significance level from testing the null hypothesis that the coefficients on governance are jointly zero.
References

We thank Suman Basu, Anja Baum, Claudia Berg, Arnildo Correa, Tito da Silva Filho, Davide Furceri, Carlos Gonçalves, Farayi Gwenhamo, Thomas McGregor, Priscilla Muthoora, Monique Newiak, Axel Schimmelpfennig, Abebe Selassie, Haimanot Teferra, Rhoda Weeks-Brown, Concha Yepes, Shu Yu, Zeine Zeidane and the participants at the IMF’s African Department Seminar for their valuable comments and suggestions. We are grateful to Miguel Pereira Mendes and Azanaw Mengistu for research assistance and to Li Tang for help with coding.

Similar results hold across a range of corruption and governance indicators. Following the literature, corruption is defined as “the abuse of public office for private gain”. Governance is defined as: “institutions, mechanisms, and practices through which government power is exercised in a country, including for the management of public resources and regulation of the economy. This includes processes at the country level, including institutions-level structural arrangements” (IMF, 2017). In this paper, governance indicators are normalized to 0 (worst) 100 (best), and corruption perceptions indicators to 0 (best) 100 (worst).

Section IV justifies this choice and gives a detailed description of these indicators. Also see http://info.worldbank.org/governance/wgi/index.aspx#home.

One standard deviation in the SSA sample for both governance and corruption perceptions is equivalent to 10 percentage points if all scores are normalized to 0-100.

All estimated growth dividends in this paper are long-term gains. Measuring the exact time that is required for these gains to materialize is beyond the scope of this paper, but it will depend to a large extent on the starting point of the countries and the overall political commitment to the process. Giavazzi and Tabellini (2005) estimate that economic and political liberalization, which is typically associated with improved governance, takes at least three years to affect governance and corruption perceptions.

Swaleheen (2011) also finds a non-linear relationship between corruption and growth but his evidence suggests that corruption is more beneficial to growth at higher levels.

The high correlation between corruption perceptions and overall measures of governance also arises by construction as corruption perceptions are included in most aggregate governance indices including ICRG and the WGI used in this paper (see further details in Section IV).

For a recent analysis of corruption problems and anti-corruption policies in Latin America see https://blogs.imf.org/2017/09/21/corruption-in-latin-america-taking-stock/.

The number of observations in the regression sample varies depending on data availability and model specification. The baseline sample covers 1995-2015.

This argument clearly applies to the so-called “bureaucratic” corruption. However, other forms of corruption that are also observed in SSA such as state capture involving high-level government officials acting on behalf of private interests (“grand” corruption) or kleptocratic regimes are also damaging to growth, including because of their scale and persistence.

While Roodman (2009) warns about considering time-invariant regressors in models like (1), the fact that the SSA dummy affects a significant fraction of our sample helps mitigate bias.

The normalization of CPI ensures that there is no level shift due to the methodological change introduced by Transparency International in 2012.

For presentational purposes, the tables only show the coefficients on the main covariates. To the extent possible, we use the same sample across governance indicators, but samples may vary due to data availability.

Throughout the estimations we have used a parsimonious number of instruments, restricted the number of lags to 2 and “collapsed” instruments to limit instrument proliferation (Roodman, 2009). To illustrate, the 17 instruments in the baseline model (columns 3 and 4 of Table 4) include: income, education, inflation, terms of trade, governance, interaction term on governance, SSA dummy, time dummies (3), lagged growth and investment (lags 1 and 2), and differenced lagged growth and investment (lag 1), and the constant term.

Another reason for excluding the political component is that stability and entrenched corruption have been prevalent in many long-lasting regimes in SSA. This is not to say that political stability (or the lack of) is irrelevant for explaining cross-country growth differentials. A large empirical literature—including Mo (2001), Aisen and Veiga (2006, 2013)—has found that political instability is associated with worse economic outcomes such as lower growth and higher inflation especially in developing countries.

The overall effect of Political Stability and Absence of Violence/Terrorism on growth is also statistically significant but somewhat smaller than those presented in Table 6. Results are available upon request.

Intermediate levels of ethnic diversity would be more supportive of checks and balances, while extreme ethnic homogeneity or fragmentation would be associated with higher corruption and lower growth. Thus, their model predicts an inverted U-curve for the relationship between corruption and growth.

The CPI is not statistically significant for the SSA sample. Results using the WGI governance indicator are roughly similar and hence are not reported.

For the 3-year regressions we interpolated Barro and Lee data on schooling, available at every five years, to mitigate the problem of missing observations. For yearly regressions we lag all control variables, except the inflation dummy, to mitigate endogeneity problems.

Other Resources Citing This Publication