political instability, investment and economic growth in sub-saharan africa

Political Instability, Investment and Economic Growth inSub-Saharan Africa

Kwabena Gyimah-Brempong and Thomas L. Traynora1

University of South Florida and aWright State University

This paper explores the relationship between political instability andeconomic growth in Sub-Saharan African nations. A more comprehensivemeasure of political instability than has previously been developed is usedin combination with a simultaneous equations model and dynamic panelestimation approach to produce several interesting inferences. First, thestatistically significant inverse relationship between political instabilityand economic growth identified by earlier studies is confirmed by the esti-mates presented here. Second, the estimated system of equations indicatesthat economic growth and political instability are jointly endogenous.Third, in addition to the direct impact that political instability hasupon growth, estimates confirm the hypothesis that political instabilityindirectly decreases economic growth by decreasing long-run capitalaccumulation. Fourth, failure to account for the dynamic nature of growthequations as well as the endogeneity of explanatory variables may producebiased effects of political instability on growth. Fifth, the broad measure ofpolitical instability we use in this study can better capture the effects ofpolitical instability on economic growth than ‘elite’ instability that hasbeen used by earlier researchers. Finally, principal components estimationis used to develop a measure of political instability that can serve as analternative to the arbitrary weighting scheme used in previous research.

JOURNAL OF AFRICAN ECONOMIES, VOLUME 8, NUMBER 1, PP. 52–86

© Centre for the Study of African Economies, 1999

1 This paper has benefited from very helpful comments of two anonymousreferees. We thank Manuel Arellano and Francesco Caselli for providing us withthe DPD program used to estimate the model. We also thank Gabriel Picone forgenerous programming help. The usual disclaimer applies.

1. Introduction

Since independence, most Sub-Saharan African countries haveexperienced declines in all measure of economic performance (WorldBank, 1984). Accompanying this poor record of economic performancehas been a high degree of political instability (McGowan and Johnson,1984). Besides the generally poor record of political stability andeconomic stagnation, there is a high degree of variation in economicperformance and political stability across countries in this part of thedeveloping world. Despite these observations and the fact that severaltheoretical analyses have argued that institutional weakness is at thecore of slow economic development in Less Developed Countries(LDCs) (e.g., see Hirschman, 1956; Powelson, 1972; Scully, 1988; North,1990), empirical growth studies have generally neglected institutionalfactors such as political stability and have focused on the quantitiesand quality of inputs in determining the pace of economic growth.This neglect has occurred in spite of the arguments of endogenousgrowth theory that economic policy formulation and implementationis an important determinant of long-term economic growth. If theinstitutional environment is important in determining the pace ofeconomic growth and an unstable political environment may preventLDCs from successfully embarking on sustained economic growth(Kuznets, 1966; Scully, 1988; Wells, 1992), then excluding it fromempirical studies of economic growth may result in biased estimatesand result in policy implications which neglect an important deter-mining element of economic growth in LDCs.

This paper uses time-series cross-national data, a simultaneousequations model and a dynamic panel estimation method to investi-gate the effects of political instability (PI) on growth rate of aggregateGDP in Sub-Saharan Africa within an augmented production functionframework. While a few studies have included PI in their analysesof economic growth, they have generally treated it as exogenous.However, empirical evidence from the political science literaturesuggests that, at least in Sub-Saharan Africa, poor economic perform-ance has a direct impact on PI, implying that it is endogenouslydetermined. Additionally, given the uncertainty generated by PI, wehypothesise that investment is endogenously determined by PI as wellas other variables.

We define political instability as situations, activities or patternsof political behaviour that threaten to change or actually change the

Political Instability, Investment and Economic Growth in Sub-Saharan Africa 53

political system in a non-constitutional way. These politically unstableevents often bring about sudden radical changes in property rightslaws and the rules governing business conduct. The key attribute ofpolitical instability, as defined here, is that these events generateuncertainties about the stability of the present political system and/orgovernment, and this uncertainty negatively impacts the authorityand effectiveness of the government. In this framework, politicalinstability need not involve a change in government or take a violentform. For example, acts of secession or prolonged anti-governmentdemonstrations could result in political instability without causing theincumbent government to fall from power. On the other hand, therecould be changes in government (e.g., through the ballot) withoutpolitical instability. Because of this, only non-constitutional changes ingovernment are included in our measure of political instability. In thissense, our definition of political instability differs from others, such asDeaton and Miller (1995), who have used changes in government,constitutional or otherwise, as their measure of political instability.Most previous research has focused narrowly on ‘elite’ politicalinstability and measured it as the occurrence of coups d’état, plots,political assassinations or purges (McGowan and Johnson, 1984;Londregan and Poole, 1990; Barro, 1991; Fosu, 1992; Svensson, 1993;Knack and Keefer, 1995). However, elite instability is not the onlyform of PI that may influence economic growth. For example, alengthy secession attempt may have a more deleterious impact on theeconomy than a coup d’état. For this reason, our more comprehensivemeasure of political instability, which includes all forms of politicalinstability — elite, communal or mass; peaceful or violent — mayfurther illuminate the relationship between growth and politicalinstability (Sanders, 1981).

In a study that included PI as a determinant of economic growth inLDCs, McGowan and Johnson (1984) found a negative correlationbetween economic growth and political instability in Sub-SaharanAfrica (although they did not report any significance tests). However,by analysing the relationship between PI and growth within a politicalscience framework, their paper fails to control for the widelyacknowledged economic factors that influence economic growth inLDCs. Fosu (1992) used cross-national data and a single equationmodel to investigate the effects of PI on economic growth in Sub-Saharan Africa over the 1960–86 period and found a significantlynegative direct relationship between PI and economic growth after

54 Kwabena Gyimah-Brempong and Thomas T. Traynor

controlling for other economic variables. Separate single equationestimates of the relationship between PI and capital and exports failedto provide statistically significant support for the hypothesis that PIindirectly impedes economic growth through reductions in capitalaccumulation and exports. The failure to confirm this intuitive hypo-thesis may be attributed to Fosu’s use of separate single equationmodels to evaluate the question, a problem which is rectified, as weshall see, by the use of simultaneous analysis. Recent research on therelationship between property rights and economic growth argues thatpolitical instability induces governments to refrain from legal reformsthat strengthen property rights, and the lack of strong property rightsprotection leads to lower economic growth in LDCs (Svensson, 1993;Knack and Keefer, 1995).

There is another mechanism through which PI can negatively affecteconomic growth in LDCs besides the direct effect PI has on the outputfrom existing resources or its indirect effect through the reduction ofresource accumulation. Edwards and Tabellini (1991) have argued thatpolitical instability weakens a government, making it difficult, if notimpossible, for the government to develop and implement theeconomic policy reforms necessary for long-term economic growth.Because of political weakness resulting from political instability,governments are likely to implement policies that encourage rent-seeking activities as a means of survival. This leads to economicstagnation. Devereaux and Wen (1996) argue that PI induces incum-bent governments to leave fewer assets to their successors, therebyforcing them to raise taxes on capital. The expected high tax on capitalinduces capital flight, low investment and slow economic growth.

The studies reviewed above assume that the relationship betweeneconomic growth and PI is unidirectional — PI affects economicgrowth but is not affected by economic growth. The assumption thatPI is exogenous is, however, inconsistent with the political scienceliterature dealing with coups d’état and other manifestations of politicalinstability in LDCs. While there is no consensus as to whether sloweconomic growth causes political instability (see Jackman, 1978;O’Kane, 1981, 1983; McGowan and Jognson, 1984; Johnson et al., 1984;Jenkins and Kposowa, 1990; Londregan and Poole, 1990) or whetherrapid economic growth is destabilising (Olson, 1963), there is a generalagreement that PI is influenced by economic performance. Londreganand Poole (1990) used a two equation model and time-series cross-national data to indicate that the probability of a coup d’état is


significantly affected by past and present income growth, but thatcoups d’état have no significant impact on income growth. This despitethe fact that they concluded that coups d’état and economic growthcannot be treated as exogenous variables in their model. While theirapproach represents a first effort at addressing the issue of simul-taneity, their growth equation does not include any economic variablesother than past growth rate and the level of per capita income.This of course raises important questions about possible modelunderspecification leading to an upward bias of the impact thatpolitical instability has on economic growth. While a few studies haveused simultaneous equations models to investigate the relationshipbetween PI and economic growth and found a negative relationshipbetween the two, they measure PI narrowly as ‘elite’ PI (Ozler andRoderik, 1992; Alesina and Perotti, 1994; Alesina et al., 1995). Besides,none of these studies investigates the indirect effect PI has on eco-nomic growth through reduced investment. Gyimah-Brempong andTraynor (1996) found that PI has negative impacts on savings andinvestment in Sub-Saharan Africa but did not extend the analysisto growth effects. Besides, the paper did not use dynamic panelestimation methods in its analysis.

Recently, a number of studies have used panel data and panelestimation methodology to estimate the neoclassical growth model(Knight et al., 1993; Islam, 1995; Caselli et al., 1996). These studiesconclude that previous studies employing single equation modelshave produced biased and inconsistent estimates of the growth–PIrelationship since they do not take into account the endogeneity ofsome of the regressors as well as the correlation between fixed countryeffects and some of the right hand side (RHS) variables. They generallyconfirm the convergence hypothesis and conclude that the speed ofconvergence is much faster than has been calculated by previousstudies employing single equations and cross-national data. However,these studies, like most empirical growth models before them,emphasise the convergence hypothesis and, as such, pay scantattention to the role of political instability in economic growth, whenit is included at all as an explanatory variable. This paper, on the otherhand focuses on the effects of political instability on economic growthin Sub-Saharan Africa.

Our study differs from previous studies of the relationship be-tween PI and economic growth in several ways. First, we provide asystematic economic framework within which to explore the issue of


simultaneity in the relationship between PI and economic growth, andestablish the direction of causation, something that single equationmodels using cross-national data cannot do. Second, the estimation ofa multiple equation model allows us to explore both the direct andindirect channels through which PI affects economic growth. Third,we develop a more comprehensive measure of PI than the ‘elite’ PIthat has been used in earlier research, permitting the measurementof the impact of non-elite forms of political instability on economicdevelopment in Sub-Saharan Africa. Fourth, lagged values of PI areincluded to measure the inter-temporal relationship between PI andeconomic growth. Fifth, we use the principal components methodto develop a more comprehensive measure of political instability inour study than has been used before. Finally, the use of panel esti-mation techniques allows us to correct for correlated country effectswith regressors as well as explore the dynamics of the PI–growthrelationship. We hope the results of our study will shed further lighton the relationship between PI and economic growth in LDCs.

The rest of the paper proceeds as follows: in the next sectionwe develop a three equation augmented production function growthmodel that endogenises investment and political instability. Section3 discusses the data used to estimate the model, while Section 4discusses econometric issues associated with the estimation of dy-namic panel growth models, and presents and discusses the statisticalresults as well as some specification tests. Section 5 summarises andconcludes the paper.

2. Model

While political scientists have argued that rapid economic growthgenerally has a significantly positive impact on political stability,economists, on the other hand, have argued that political instabilityhas a deleterious effect on economic growth. In most empiricalresearch, political scientists treat economic growth as exogenous whileeconomists treat PI as exogenous. The question of whether or not PIand economic growth are jointly endogenous is a testable hypothesis.Therefore, in modelling the relationship between PI and growth, wecombine these two strands of literature and treat PI and growth asjointly endogenous.

We use an augmented production function approach to investigatethe relationship between political instability and economic growth in


Sub-Saharan Africa. We model economic growth as a function of thegrowth rates of capital, labour and export (Balassa, 1978; Krueger,1980; Feder, 1983; Ram, 1987; Edwards, 1998). As argued by Fosu(1992) and others, economic growth is directly influenced by pro-duction losses created by PI, which is included as an additionalexplanatory variable. We expect this output loss, and hence decreasedeconomic growth, even if PI has no impact on investment. Addition-ally, lagged per capita real GDP is included to test the convergencehypothesis. The growth equation we estimate is given as follows:

(1) g = α0 + α1k + α2l + α3x + α4yt–1 + α5PI + e

where g is the growth rate of GDP, l is the growth rate of labour, k isinvestment, x is the growth rate of exports, PI is political instability, yt–1is lagged per capita real GDP and e is a stochastic error term. Since k isa normal input, we expect its coefficient to be positive. In neoclassicalgrowth theory, an increase in population growth is expected todecrease the growth rate of per capita GDP, all things being equal.However, because the growth rate of aggregate GDP is the dependentvariable in our model, we expect the coefficient of l to be positive in thegrowth equation. Since the economic growth of nations in Sub-Saharan Africa has historically varied positively with exports, α3 isexpected to be positive. As discussed earlier, PI is expected to directlydisrupt the production process (apart from indirectly disruptinggrowth through a reduction in investment as detailed below), andis therefore expected to have a negative coefficient. The coefficient ofyt–1 is expected to be negative in accordance with the convergencehypothesis.

As discussed earlier, there are reasons to question the assumptionthat PI is exogenous in the growth equation. It is reasonable to expectthat poor economic performance increases the clamour to change theincumbent government or political system. In developed countrieswhere there are well established institutional processes for changinggovernments, this dissatisfaction manifests itself through the electionof a new government or reforms of the economic and political systems.In Sub-Saharan Africa, where there is little chance of changing anincumbent government or political system through formal processes,frustration with poor economic performance manifests itself inpolitically unstable events such as coups d’état, riots and secession-ist movements. Theoretical as well as empirical evidence from the

.

.


political science literature supports this proposition (McGowan andJohnson, 1984; Londregan and Poole, 1990).

Following the political science literature, we represent PI as afunction of economic growth, the strength of the military’s role in civilsociety (mil), the legislative selection process (lgsl), the type of head ofstate (pres, gen) and the extent of fractionalisation in the polity (frac).Formally, the PI equation is given as:

(2) PI = β0 + β1g + β2mil + β3frac +β4 pres + β5gen + β6lgsl +β7yt–1 + β8PIt–1 + µ

where µ is a stochastic error term and all other variables are as definedabove. The political science literature suggests that the military tendsto intervene in politics when it has a strong, high profile role in society.This may be because the military sees itself as the arbiter betweencompeting interest groups in the country or because it sees an easyopportunity to seize power. The coefficient of mil is therefore expectedto be positive in the PI equation. frac is expected to have a positivecoefficient as increased fractionalisation may either create oppor-tunities for military involvement in political matters or create greaterfrustration and uncertainty among the public in the polity. pres andgen are dummy variables for constitutionally elected and militaryheads of state respectively (Monarchy is the excluded category). Thecoefficients of pres and gen cannot be signed a priori since their signswill depend upon how they affect PI relative to that of a monarchy. Onaverage, the more democratic and open the legislative selectionprocess is, the lower PI will be, all things being equal. We thereforeexpect the coefficient lgsl to be negative in the PI equation. The fastereconomic growth is, the lower PI will be, all things being equal.Therefore the coefficient of g is also expected to be negative in the PIequation. We have included a lagged value of PI based on thehypothesis that countries experiencing political instability develop aninter-temporal ‘culture of political instability’ (Londregan and Poole,1990). If this hypothesis is correct, then we expect the coefficient ofPIt–1 to be positive. We have included per capita real GDP to testwhether political instability is correlated with level of development.

PI can affect economic growth in two ways: directly by reducingthe output of existing resources, and indirectly by reducing theavailability (or growth) of inputs, such as capital. Investment incapital requires long-term planning and guarantees of property rights.Drastic and frequent changes in property right laws, rules governing


repatriation of profits, as well as the increased uncertainty that oftenaccompanies political instability makes this long-term planningimpossible. In Sub-Saharan Africa, the drastic increase in the ‘braindrain’ in the last two decades has been attributed in part to politicalinstability (Apraku, 1991). This implies that investment is, in part,determined by PI and the savings rate (s). The expected negativeimpact of PI on investment is consistent with the model devel-oped by Ozler and Roderik (1992), Devereaux and Wen (1996), andGyimah-Brempong and Traynor (1996). We write our investmentequation as:

(3) k = γ0 + γ1g + γ2PI + γ3PIt–1 + γ4m + γ5rds + γ6s + γ7kt–1 + ε

where m is the imports/GDP ratio, rds is the real foreign debtservice/GDP ratio, s is the savings ratio, ε is a stochastic error term andall other variables are as defined above. We expect the coefficients of gand s to be positive. In Sub-Saharan Africa, where almost all invest-ment goods are imported, investment is constrained by importcapacity, hence a relaxation of that constraint will increase investment.We therefore expect the coefficient of m to be positive. High foreigndebt constrains investment because of the need to divert resourceswhich would otherwise have gone for investment to service externaldebt. Therefore, the coefficient of rds is expected to be negative. Wehave included kt–1 in the investment equation to investigate thedynamic nature of investment in Sub-Saharan Africa.

The system of equations we estimate in our model is given asfollows:

git = α0 + α1kit + α2lit + α3xit + α4yit + α5PIit + eit

kit = γ0 + γ1git + γ2PIit + γ3PIit–1 + γ4mit + γ5resit + γ6sit + γ7kt–1 + εit

Piit = β0 + β1git + β2milit + β3fracit + β4presit + β5genit +β6lgslit + β7yit–1 + β8PIit + µit

The system of equations to be estimated can be written in compactmatrix format as:

(4) Γy + X′β + µ = 0

where y, X and β are the vectors of endogenous, exogenous andcoefficients to be estimated respectively, while µ is the vector ofstochastic error terms. The error term in equation (4) consists of a

.


country specific component, a time component and white noise.Formally, the error term is:

µit = αi + λt + vit

where αi is the country specific component, λt is the time specificcomponent and vit is white noise. E(αi) ≠ 0 and the same can be said ofλt. In general, αi and λt will be correlated with the regressors in theequation.

From the model presented in this section, the total current year effectof one standard deviation change in PI on economic growth dg/dPI isthe sum of the direct and indirect current year effects which is givenas:

dg/dPI = ∂g/∂PI + (∂g/∂k)(∂k/∂PI) = α5 + α1γ2.

Thus, our model allows for the fact that the direct effect of PI ongrowth may only be a partial determinant of the overall relationshipbetween PI and growth. Additionally, the 1 year inter-temporal impactof PI on growth is represented by

dg/dPIt–1 = (∂g/∂k)(∂k/∂PIt–1) + (∂g/∂PI)(∂PI/∂PIt–1) = α1γ3 + α5β8.

This inter-temporal impact shows the effects of PI on economic growth1 year after the occurrence of PI. Alternatively, one can interpret this asthe effects of earlier episodes of political instability on economicgrowth in the current period. Note that this relationship between PIand growth does not represent the usual multiplier effect; it simplydisplays the total effect that PI has an economic growth, regardless ofthe source of variation in PI.

3. Data

The endogenous variables in our model are g, PI and k. We measure gas the annual growth rate of aggregate real GDP in a country, andfollow Ram (1987) and others by measuring k as the investment/GDPratio. There is no simple way to measure PI. For one thing, our notionof political instability is multidimensional, taking such diverse formsas political assassinations, secession movements, guerrilla warfare,coups d’état and purges. For another, these events can occur withdifferent degrees of frequency in different countries at different times.For example, there may be more coups d’état in one country thananother but there may be more guerrilla attacks in the second country


than the first. The traditional approach to measuring political in-stability in the extant political science and economics literature isto create a political instability index by assigning arbitrary weights tovarious politically unstable events. The resulting political instabilityindex reflects the subjective evaluation of political events of theresearcher, and is therefore not the kind of objective measure thatwould be better suited for such research. Moreover, most of these PIindexes do not include the effects of ‘less serious’ politically unstableevents as they use only ‘elite’ PI events.

Since the many different politically unstable events which make upPI in this analysis are likely to be highly correlated, we measure PI asa single weighted index of politically unstable events in a countryduring a single calendar year. In an effort to measure political instabil-ity in a less arbitrary manner than those developed for previousresearch, the principal components method is used to create aweighting mechanism for PI. This method creates a variable whichis weighted such that it maximises the correlation between itselfand the individual politically unstable events, thus creating a singlerepresentative measure of political instability.2 In calculating thisindex, we normalised the variables so the resulting principal com-ponent indicate the deviation of PI from its mean — in effect, the PIsare standardised z scores. The use of the principal components methodfor the development of a measure of PI allows for the use of astatistically determined variable in our analysis which, althoughnot based on economic and political science theory, is free from thepotentially arbitrary hands of the researcher in its development.

Variables used in the construction of this broad index of politicalinstability are: successful and attempted coups d’état, guerrilla warfare,secession movements, political assassinations, revolutions, riots,major government crises, purges, constitutional crises, large-scaleanti-government demonstrations, politically motivated strikes, consti-tutional changes and plots. In addition to this broad measure of PI, wealso created and index of elite political instability using principalcomponents methodology and used the resulting PI measure to


2 Principal components creates a composite variable which has the highestpossible correlations with the individual types of political instability. This isaccomplished by choosing the vector of weights, b, which maximise the varianceof b′x subject to: b′b = 1 (see Dunteman, 1984).

re-estimate the model. We do so in order to compare the estimatesusing elite PI events with those using the more comprehensivemeasure detailed above to see if the index used makes a difference inthe results. The events used to create the ‘elite’ PI index are successfuland attempted coups, revolutions and plots.

To capture a nation’s import ability, we measure m as the import/GDP ratio. rds is measured as the ratio of external debt service to GDPwhile y is measured as real per capita GDP. It is difficult to quantify therole the military plays in society. However, as the political scienceliterature suggests, the extent of military involvement in civiliansocieties can be captured by the amount of resources the militarycommands in the society. We follow this approach and measure mil asthe number of military personnel per thousand in the population.3 fracis the party fractionalisation index as defined by Rae (1969) and ismeasured as frac = 1 – Σm

i=1ti2, where ti is the proportion of members of

the legislature associated with the ith political party. lgsl is a scalevariable that equals 0 if no legislature exists, 1 if there is a non-electedlegislature and 2 if the legislature is elected either directly or indirectly.pres is a dichotomous variable that equals 1 for a constitutionallyelected head of state, 0 otherwise, while gen is a dummy variable thatequals 1 if the head of state is either a military leader or a civilianimposed by the military, 0 otherwise. The excluded category ismonarchy, so the effects of pres and gen on PI are relative to that of amonarchy.

Data for the calculation of the socio-political variable (PI, frac, pres,gen, lgsl) were obtained from Banks (1993), while the economic data (g,l, k, m, s, y and rds) were obtained from the World Bank (1992, 1995).4

Data for MIL were obtained from Arms Control and DisarmamentAgency (ACDA), World Military Expenditures and Arms Transfers


3 It is possible to use defence burden (defence expenditure/GDP ratio) as ameasure of mil. However, given the problems of getting accurate measures ofdefence expenditure in LDCs, we chose to use defence personnel as a measure ofmil. For a discussion of the problems involved in measuring defense spending inLDCs, see Ball (1984).4 Nuxoll (1994) argues that because of index number problems, it is advisable touse GDP measured in domestic prices to calculate growth rate and to use PennWorld Tables to measure levels of GDP. However, we feel that it makes sense touse the same sources of data to measure levels as well as to calculate growth. Atleast that way data weaknesses may derive from one source rather than severalsources.

(various years). The data consist of annual observations for the1975–88 period for 39 Sub-Saharan African countries.5 The choice ofsample countries and period of coverage were constrained by theavailability and completeness of data.

Summary statistics of the data are presented in Table 1. As thestandard errors indicate, there is a wide variation in the values of thevariables across countries and through time, with the data showingrelatively high and variable levels of political instability as well as lowaverage rates of economic growth during the sample period. We notethat with aggregate GDP growing at 2.942% and population growing

Table 1: Summary Statistics of Data

Variable Mean Standard deviation

g (%) 2.942 6.774k (%) 21.125 10.760PI –0.00009 0.999y 485.727 569.696l 2.761 1.437mil 3.080 3.352pres 0.084 0.283frac 0.083 0.193gen 0.782 0.414lgsl 1.419 0.897m (%) 41.272 24.758s (%) 6.975 19.436rds (%) 42.26 39.15x 6.224 28.513elite –0.000005 1.00002

N = 546. Means reported here are unweighted averages.


5 The countries in our sample were: Benin, Botswana, Burkina Faso, Burundi,Cameroon, Cape Verde, Central African Republic, Chad, Comoros, Congo,Ethiopia, Gabon, Gambia, Ghana, Guinea, Ivory Coast, Kenya, Lesotho, Liberia,Madagascar, Malawi, Mali, Mauritania, Mauritius, Mozambique, Niger, Nigeria,Rwanda, Senegal, Sierra Leone, Somalia, Sudan, Swaziland, Tanzania, Togo,Uganda, Zaire, Zambia and Zimbabwe.

at 2.76% per annum, GDP per capita, on average, stagnated in Sub-Saharan Africa during the sample period. Therefore in addition, themodel is estimated with annual data which may be influenced bycyclical movements. In addition to estimating the model with annualdata, we also estimated the model with three year averages of the data.

4. Estimates

4.1 A Dynamic GMM Estimator

The three equation simultaneous equations model is estimated withdata from 39 countries over a 14 year period. In panel estimation,consistent estimation of the structural coefficients depend crucially onthe stochastic properties of the error terms, particularly whether theyare correlated with the regressors and whether the error terms areserially correlated. As argued above, the error term in equation (4) hasa country-specific component, a time-specific component and an idio-syncratic component. The error term is given as:

(5) µit = αi + λt + vit

where vit is i.i.d. with mean zero and a constant variance. If the RHSvariables are othorgonal to µit, a GLS estimator will be consistent. Onthe other hand, if the RHS variables are strictly exogenous with respectto vit but not to αi, a within group estimator [Fixed Effects (FE)] will beconsistent. In our model, there is no reason to believe that eithercondition will hold since there are endogenous regressors and thedynamic nature of growth equations implies correlation of error termswith regressors. Besides the endogeneity of some of the RHS variables,the inclusion of lagged endogenous variables in the PI and k equationsimplies that the orthorgonality condition will not be satisfied even forthe FE estimator, whether it is estimated in levels or in differences. Anestimator that is capable of taking care of these problems is thereforeneeded.

Chamberlain (1983) has proposed a Generalised Method of Moment(GMM) estimator (the Π matrix estimator) that allows transformationof the variables to achieve orthogonality between the country fixedeffects and the regressors (Holtz-Eakin et al., 1988; Caselli et al., 1996).In the absence of autocorrelation, the Π matrix GMM estimator isconsistent and efficient. The Π estimator, however, does not allow oneto control for the endogeneity of regressors. The Π matrix GMM


estimator is the estimator employed by Knight et al. (1993) and Islam(1995) to estimate the neoclassical growth model. From the model wedeveloped in Section 2, it is clear that there are endogenous regressorsin our model, hence the use of the Π matrix estimator will producebiased and inconsistent estimates. A consistent estimator is thereforeneeded.

To account for the endogeneity of regressors as well as addresscorrelated country effects, we use the dynamic panel estimatorproposed by Arellano and Bond (1991) and used by Caselli et al. (1996)to estimate their growth equation to estimate the model in this study.This estimator optimally exploits the linear moment restrictionsimplied by a dynamic panel data model. The estimator requires thatthe variables be measured as deviations from their period meansand equations be estimated in differenced form. We therefore writeequation (4) as follows:

(6) Γ∆y + ∆X′β + ∆µ = 0

where y and X are X and y centred on their period means and ∆ is thedifference operator. Measuring the variables as deviations from theirperiod means and first differencing them removes λt and αi from theerror term. Differencing the variables also implies that all timeinvariant dummy variables drop out of the equation, hence pres andgen drop out of the PI equation in our model.

Equation (6) cannot be estimated directly by least squaresprocedures because of the endogeneity of some of the regressors.Second, the lagged endogenous regressors are correlated with thecomposite error term through the contemporaneous error terms inperiod t – i. Therefore an instrumental variables (IV) estimator isrequired to estimate the model. The Arellano and Bond GMMestimator we use here is an IV estimator that uses all past values of theexplanatory variables as well as all strictly exogenous variables asinstruments. The dynamic GMM estimator is given as:

(7) θ = (X′ZANZ′X)–1XZANZ′y

where θ is the vector of coefficient estimates on both the endogenousand exogenous regressors, X and y are the vectors of first differences ofall the explanatory variables in equation (6), Z is the vector of instru-ments and AN is a vector used to weight the instruments. This GMMestimator is an instrumental variable estimator that is equivalent to anefficient Three Stage Least Squares (3SLS) estimator. The dynamic


GMM panel estimator requires that there be no serial correlationamong the error terms. Because we estimated the model in firstdifferences of the variables and the model includes lagged values ofsome endogenous variables, we had a total of 429 usable observations(39 countries over 11 years).

Arellano and Bond proposed two — one-step and two-step —estimators, with the two-step estimator being optimal. The differencebetween the two estimators is the weighting matrix used to obtain theestimates. The one-step estimator is obtained when the weightingused is the average covariance matrix of Zvi given by AN =(N–1ΣiZ′iHZi)–1 where H is a (T-2) square matrix with 2s in the maindiagonal, –1s in the first sub-diagonal and 0s otherwise. The two-stepestimator replaces the H matrix with an estimated variance–covariancematrix formed from the residuals of a preliminary consistent estimatorof θ. The optimal choice of AN for the two-step estimator is given as:

AN = VN = N–1ΣiZviviZi

where v are the residuals obtained from a preliminary consistentestimate of θ. The one- and two-step estimators will be asymptoticallyequivalent if and only if the errors are spherical. Given the nature ofour model with endogenous regressors, possible correlated fixedeffects and a heteroskedastic error structure, we suspect that theconditions for spherical errors and hence the equivalence of the two-and one-step estimators will not be met. In light of these reasons, ourdiscussions of the results will be based on the estimates from thetwo-step estimator which is the optimal estimator. For the purposes ofcomparison, we present the one-step estimator in Appendix A.

In estimating the model, we lagged all explanatory variables by twoperiods to ensure that yt–2 and yt–1 could be treated as exogenous andthat the error terms would not be serially correlated. We made twoidentifying assumptions. First, we assumed that there is no first- orsecond-order serial correlation among the error terms. Second,endogenous regressors are not considered predetermined for vit butare considered so for vit+2. This allows us to use all xi up to xt–2 as wellas yt–2 as valid instruments for xt. The consistency of our estimateshinges crucially on the assumption of lack of autocorrelation of theerror terms, so we test for the existence of first- and second-order serialcorrelation. We also perform Sargan’s test of overidentifyingrestrictions as well as the suitability of the instruments used toestimate the model (for details see arellano and Bond, 1991). If all

^

^

^


regressors in the model are not strictly exogenous, the dynamic panelestimator is consistent while the Π matrix estimator is not. On theother hand, if the strict exogeneity assumption holds, both estimatorsare consistent but the Π matrix estimator is more efficient than thedynamic panel estimator. If all regressors are strictly exogenous, thencontemporaneous as well as lagged values of all regressors are validinstruments. This gives us a way to test for the exogeneity of re-gressors. We test the strict exogeneity assumption by conducting aHausman test that compares our simple GMM estimator with one thatinclude current values of the regressors as additional instruments(Hausman, 1978). Rejection of the null implies the rejection of the strictexogeneity assumption.

4.2 Results

Table 2 presents the estimates of the two-step dynamic GMMestimator. Column 2 presents estimates of the growth equation andcolumn 3 the estimates of the investment equation, while column 4presents the estimates for the PI equation. Asymptotic ‘t’ statisticscalculated from the heteroskedastic consistent standard errorsassociated with the GMM estimator are reported in parentheses. Table2 also presents statistics for the Sargan test, a first- and second-orderserial correlation test, as well as the Hausman exogeneity test. Inaddition, it presents Wald test statistics to test for the significance of asubset of regressors. As the estimates indicate, the model goodness-of-fit statistics are relatively strong. In all three equations, we reject thenull hypothesis that the variation in the dependent variables cannotbe explained by the variation in all of the independent variables atα = 0.01 or better as indicated by the χ2 statistics of the joint test ofsignificance.

Before discussing the coefficient estimates and their growthimplications, we discuss some specification tests since the validity ofour results depend upon the consistency of the dynamic GMMestimator we use here. In our analysis, the key moment condition weexploit is lack of serial correlation among the error terms. Table 2presents test statistics for first- and second-order serial correlation(m2 statistics), the Sargan stability test, and the Wald test to test thesignificance of a set of regressors. Finally, Table 2 also includesHausman tests of the exogeneity of all regressors. The test statisticsshow the absence of any first- or second-order serial correlation. It is


also clear from the Sargan test for overidentifying restrictions thatthese restrictions are correct and that the set of instruments we used toestimate the equations are the correct set of instruments. The Hausmantest of the null of exogeneity of all regressors is soundly rejected for allequations. This implies that the Π matrix estimator would produceinconsistent estimates. The Wald tests indicate that all the tested

Table 2: Dynamic Panel GMM Estimates of Model Coefficient Estimates

EquationVariable g k PI

g – 0.1203 (3.1136)a –0.0385 (2.5508)k 0.2453 (2.9076) – –kt–1 – 0.2598 (10.621) –PI –0.1013 (2.0843) –0.2053 (1.8032) –PIt–1 – –0.9619 (2.2637) 0.1786 (4.5931)x 0.0552 (6.3759) – –l 0.8541 (1.6965) – –yt–1 –0.0018 (1.3088) – 0.0022 (2.7303)s – 0.5971 (16.1072) –m – 0.6191 (37.7909) –rds – –0.2223 (4.9794) –mil – – 0.0886 (1.0769)lgsl – – –0.2647 (4.1091)frac – – 0.4118 (1.6774)Joint testsignificance

39.059 (5) 4200.72 (6) 76.940 (6)

Sargan 12.305 (19) 21.973 (18) 16.303 (19)1st order serialcorrelation

1.050 1.007 1.217

2nd order serialcorrelation

0.591 0.371 0.427

Hausmanexogeneity test

95.6989 76.9745 399.023

Wald test 16.342 (PI) 23.72 (PI, PIt–1) 13.021 (g)

aAbsolute value of asymptotic t-statistics is given in parentheses.


variables (in parentheses) are significantly different from zero atconventional levels.

We now turn to a discussion of the coefficient estimates from thedynamic GMM estimator. In the growth equation presented in column2, k and l have positive coefficients as expected but only the coefficientof k is significantly different from zero at α = 0.01, while that of l issignificant at α = 0.05.6 The positive coefficient of these variablesindicates that they are normal inputs into the growth process. The factthat the coefficient of l is less than unity is consistent with theneoclassical hypothesis that increased population growth decreasesthe growth of per capita income. The coefficient of x is positive andsignificantly different from zero at α = 0.01, as expected. This result isconsistent with those of researchers who find a positive relationshipbetween export growth and economic growth in LDCs. The coefficientof yt–1 is negative as expected, but it is insignificant at conventionallevels. There is some evidence of the convergence hypothesis; how-ever, the evidence is very weak as the coefficient of yt–1 barely missedsignificance at the 10% level. PI has a negative coefficient as expected,and is significantly different from zero at α = 0.05 or better in thegrowth equation. A one standard deviation increase in politicalinstability directly decreases economic growth in Sub-Saharan Africaby 0.1 percentage points. This is a relatively large effect given that theaverage growth rate of aggregate GDP in Sub-Saharan Africa duringthe sample period was only 2.9%. This implies that political instabilityhas a significant direct negative impact on economic growth in Sub-Saharan Africa after controlling for other factors that affect economicgrowth in a system of equations model. This result qualitativelyconforms with the results of other researchers who have explored therelationship between PI and growth.

In reviewing the estimates for the PI equation, g has a negativecoefficient that is significantly different from zero at α = 0.01, implyingthat, all things being equal, good economic performance increasespolitical stability in Sub-Saharan Africa — a result which is consistentwith the political science literature and the results obtained byLondregan and Poole (1990), Alesina and Perotti (1994) and Alesina etal. (1995). A 1% increase in economic growth rate decreases political

.


6 In the cases where a definite sign for the parameter estimate is hypothesised, aone-tailed t-test was conducted. Otherwise, two-tailed t-tests were used.

instability by 0.04 standard deviations. An alternative way to state thisresult is that economic deterioration leads to political instability inSub-Saharan Africa. The implications of the negative coefficient ofgrowth in the PI equation is that governments — democratic as well asautocratic — can achieve political stability with improved economicperformance. The coefficient of lgls is negative, relatively large andsignificantly different from zero at the 1% level of significance.Thus, an open legislative selection process reduces political instability.frac and mil have coefficients that are positive but only that of fracis significantly different from zero at α = 0.05, indicating thatfractionalisation of the legislative body increase political instabilityin Sub-Saharan Africa. PIt–1 has a positive coefficient which issignificantly different from zero at α = 0.01, suggesting that Sub-Saharan African countries that experience political instability tend todevelop a ‘culture of political instability’. The negative and significantcoefficient of PI in the growth equation together with the negative andstatistically significant coefficient of g in the PI equation are compatiblewith our hypothesis of simultaneity in the relationship between PI andeconomic growth.

In the investment equation, the coefficients of g, m and s are positiveas expected, and are significantly different from zero at α = 0.01. Thelarge, positive and significant coefficient of m in the investment equa-tion indicates that import capacity is a very important determinant ofinvestment in Sub-Saharan Africa. The coefficient of rds is negativeand significant at α = 0.01 as expected. This is an indication thatexternal debt overhang decreases investment in Sub-Saharan Africa.The coefficient of kt–1 in the investment equation is positive andsignificantly different from zero at α = 0.01. This indicates that lastyear ’s investment is positively correlated with investment in thecurrent year. The coefficients of PI and PIt–1 are both negative,relatively large and significantly different from zero at α = 0.05 orbetter. These significantly negative coefficients imply that politicalinstability has deleterious effects on investment both contempor-aneously as well as with a lag. A one standard deviation increase inpolitical instability decreases investment rate by 0.21 percentagepoints contemporaneously. Investment in the current year is alsodecreased by 0.9619 percentage points as a result of one standarddeviation increase in political instability in the previous year. Thisnegative and statistically significant coefficient of current and laggedPI in combination with the positive and significant estimate of the


parameter for k in the growth equation indicates that PI negativelyinfluences growth indirectly through investment. This result is con-sistent with the hypothesis that PI impacts economic growth throughinvestment by altering incentives through changes in property rights,rules supervising business conduct, increasing uncertainty aboutthe outcomes of investment decisions, and through increased capitaltaxation.

The estimates presented in Table 2 indicate that PI negatively affectseconomic growth in two ways: directly and in the current year througha reduction in output as well as indirectly and inter-temporallythrough its impact on present and future levels of investment. Thetotal current year effect of PI on economic growth, as shown above, isgiven as: dg/dPI = α5 +α1γ2, which is –0.1517, made up of –0.1013 directeffects and –0.0504 indirect effects through reduced investment, whilethe 1 year inter-temporal effect is dg/dPIt–1 = α1γ3 + α5β8, which is equalto –0.253, a much larger effect than the current year effect. This is avery significant effect given the poor growth record of Sub-SaharanAfrican countries. The inter-temporal growth effect of PI implies thatan increase in PI in an earlier period has a negative and significantimpact on economic growth in the current period through reducedinvestment, making the PI effects on growth longer lasting thanthe direct current year effect estimated by earlier researchers. Theinter-temporal effect arises from two sources: a decrease in investmentand hence a negative impact on growth later, and through additionalepisodes of PI in the future generated by the current episode of PI lead-ing to further economic deterioration. The implied direct PI growthelasticity for the current period, calculated at the means of thevariables, is –0.056. These calculations indicate that political instabilityhas a large, negative and statistically significant effect on economicgrowth, both directly and indirectly through reduced investment, inSub-Saharan Africa. Failure to estimate a dynamic structural modelwould imply missing this lagged effect of PI on economic growth. Thisimplies that studies that do not account for this indirect effect mayseriously underestimate the effects of PI on economic growth inSub-Saharan Africa and, more importantly, miss the mechanismsthrough which PI affects economic growth. The inter-temporal natureof the investment effects of PI implies that the growth effect will bemagnified over time.

It is possible that the annual data used to estimate the model issubject to cyclical movements, making our results dependent solely on


the cyclical movements in the data. To investigate this possibility,we re-estimated the model using 3 year averages of the data. With14 years of data for each country, averaging over 3 years gave us fiveobservations for each country. Taking first differences and laggingthe differenced variables by one period left us with a total of 117observations (39 countries over 3 years) to estimate the model.Coefficient estimates, together with regression statistics, are presentedin Table 3. The regression statistics as well as the test statistics indicatethat the model fits the data relatively well. In the growth equation, thecoefficients of k, x, l and yt–1 are positive as expected but only those ofk and x are significant at conventional levels. The coefficient of PIis negative as expected, relatively large and significantly differentfrom zero at α = 0.01, indicating that political instability has a directlynegative impact on economic growth in the current period in Sub-Saharan Africa. In the investment equation, the coefficients of g, kt–1, sand m are positive as expected and all are significantly different fromzero at α = 0.05 or better. The coefficients of PI, PIt–1 and rds arenegative and significantly different from zero at α = 0.05. This impliesthat investment in the current period is negatively impacted bypolitical instability in the current period as well as political instabilityin the previous period. In the PI equation, the coefficients of PIt–1 andfrac are positive and significant at α = 0.05 while those of g and lgsl arenegative and significant at α = 0.01. The coefficients of yt–1 and mil areinsignificant.

The growth effect of PI in the current year from the estimatespresented in Table 3 is –0.4153 divided into –0.227 direct effect and–0.188 indirect effect through reduced investment. The total inter-temporal effect of PI on economic growth calculated from theseestimates is –0.2398. These estimates are relatively large, given that theaverage growth rate of aggregate GDP in the sample period was only2.9%. The results presented in Table 3 are qualitatively the same as thosepresented in Table 2. In both tables, PI has a negative effect on growthdirectly and indirectly through decreased investment, and an inter-temporal effect on growth through decreased investment in additionto the current year effects. The only major difference between theresults presented in the two tables is that the estimates using the 3 yearaverages are quantitatively larger in absolute magnitude than thosebased on annual data and the estimates in Table 3 are generally lessefficiently estimated than their counterparts in Table 2. Given that theannual data may have more noise than the 3 year average data, it is

..


not surprising that the estimates differ in absolute magnitude. Theconclusion we draw from a comparison of estimates in the two tablesis that there is no qualitative difference in our results whether we useannual data or the 3 year averaged data, hence our results are notbeing driven by cyclical variation in the data.

How do our results differ from the results of earlier researchers whoeither treat PI as exogenous or do not employ the dynamic panelestimator? To investigate this question, we estimate single equation

Table 3: Dynamic Panel GMM Estimates of Model: Averaged Data CoefficientEstimates


g – 0.0898 (2.346)a

k 0.2710 (2.484) – –kt–1 – 0.4062 (2.299) –PI –0.2270 (2.662) –0.6949 (1.948) –PIt–1 – –0.6033 (2.345) 0.3733 (3.761)x 0.0048 (5.246) – –l 0.6183 (1.430) – –yt–1 –0.0068 (1.368) – –0.0001 (0.519)s – 0.4148 (3.856) –m – 0.4475 (7.766) –rds – –0.7534 (2.744) –mil – – –0.0356 (1.128)lgsl – – –0.3067 (2.612)frac – – 1.5935 (2.9748)Joint testsignificance

52.133 (5) 897.278 (7) 65.903 (6)

Sargan 6.504 (9) 6.903 (10) 6.461 (10)1st order serialcorrelation

0.607 0.087 0.308

Hausmanexogeneity test

56.297 68.9745 89.683




growth equations allowing for fixed effects and a 3SLS version of ourstructural model and compare the results to the estimates presented inTable 2. Table 4 presents the results of the 3SLS estimates; Table 5presents the single equation estimates. Column 2 of Table 4 presentsthe estimates for the growth equation and column 3 estimates for theinvestment equation, while column 4 presents the estimates for the PIequation. All the 3SLS coefficient estimates in all equations, except thatof PIt–1 in the investment equation, have the expected signs, but thecoefficients of l in the growth equation and gen and pres in the PIequation are statistically insignificant. The coefficient of PIt–1 in theinvestment equation is positive, relatively large and significantlydifferent from zero at α = 0.05. This is contrary to expectation and

Table 4: 3SLS Estimates of Economic Growth in Sub-Saharan Africa

Coefficient estimatesVariable g k PI

g – 0.0299 (2.040)a –0.0595 (6.315)k 0.1173 (3.356) – –kt–1 – 0.5172 (15.867) –PI –1.2182 (2.1247) –4.1868 (3.661) –PIt–1 – 1.2258 (2.553) 0.3745 (8.764)x 0.0687 (7.209) – –l 0.1652 (0.4315) – –yt–1 –0.0012 (2.178) – –0.0002 (2.369)m – 0.1473 (8.437) –s – 0.0707 (3.802) –rds – –0.1996 (2.610) –mil – – 0.0194 (1.668)pres – – 0.0248 (0.170)gen – – –0.1229 (1.169)frac – – 0.3658 (1.983)lgsl – – –0.0392 (2.518)N 546 546 546R2 0.1598 0.6821 0.2041DW 1.9161 2.011 2.089

aAbsolute value of t-statistics is given in parentheses.


different from the results in Tables 2 and 3 above. In addition to thewrong sign of the coefficient of PIt–1 in the investment equation, thecoefficient of PI in the growth equation implies that a one standarddeviation increase in PI decreases growth rate by about 1.35percentage points. The total growth effect of a one standard deviationincrease in PI is –1.709 in the current period and –0.313 in the nextperiod. These estimates are far larger than those estimated from thedynamic GMM estimator. Given the average growth rate of 2.9%during the sample period, these effects are likely to be overestimatesof the growth effects of PI. Table 5 presents estimates for the singleequation models. Column 2 presents the OLS estimates, while column3 presents GLS panel estimates of the growth equation. In both cases,the coefficient of PI is negative and significant. They indicate that a onestandard deviation increase in PI decreases growth by 0.65 percentagepoints, estimates that are larger in absolute terms than the total growtheffect calculated from the dynamic GMM estimates in Table 2.

The estimated total effects of PI on growth from the 3SLS and singleequation growth equations far exceed those calculated from the

Table 5: Single Equation Estimates of Growth Equation

Coefficient estimatesVariable OLS Panel

k 0.1366 (4.870)a 0.1355 (4.407)x 0.0685 (7.002) 0.0706 (7.289)PI –0.6460 (2.266) –0.6797 (2.289)l 0.2509 (0.904) 0.3320 (1.0222)yt–1 –0.0004 (0.701) –0.0002 (0.360)R2 0.1435 0.1475F 19.262 –RMSE 6.451 6.2906Variance components

Cross sections – 1.5962Times series – 0.7930

Hausman test for random effects 12.083 (df = 5)

aAbsolute value of t-statistics is given in parentheses.


dynamic panel estimator. Given our conclusion that the Π matrixestimator is biased and given the fact that neither the 3SLS estimatornor the single equation estimator is able to solve the problem ofcorrelated fixed effects and account for the dynamic nature of thegrowth equations, it is most likely that the larger calculated growthimpacts from these estimates is a reflection of upward biases impartedby the correlated country effects. The elasticity of growth with respectto PI calculated from Tables 4 and 5 are –0.728 and –0.291 for the 3SLSand the single equation estimates respectively. These estimates seemto be on the high side. On the other hand, our elasticity estimate of–0.056 is closer to the elasticity obtained by Caselli et al. (1996) whenthey used assassinations as their measure of PI.

Earlier researchers have used ‘elite’ PI as the measure of politicalinstability in their empirical analysis. To what extent do the resultsfrom the broad measure of PI differ from the results based on ‘elite’ PI?For the sake of comparison, we re-estimated the dynamic GMM modelusing principal components estimates of an elite PI index based on thePI events (successful and attempted coups, revolutions, and plots) thathave been used in most earlier studies. The results are presented inTable 6. As in Table 2, column 2 presents the OLS estimates of thegrowth equation and column 3 the estimates for the investmentequation, while column 4 presents the estimates for the PI equation. Asin Table 2 above, the coefficients of k, x and l in the growth equationare positive and statistically significant. The coefficient of PI in thisequation is statistically insignificant.7 In the investment equation, allcoefficients have the expected signs and a large number of them arestatistically significant. The coefficient of PI is insignificant, althoughit has the expected sign. In the PI equation, almost all coefficients havethe expected signs, with the exceptions of PIt–1 and MIL, althoughneither is statistically significant.

The estimates presented in Table 6 are qualitatively similar to thosepresented in Table 2. Except in a few cases, both sets of coefficientshave the same signs. However, the coefficients in Table 6 are generallyless precisely estimated than their counterparts in Table 2. There arealso some marked differences between the two sets of estimates aswell. The coefficient of elite PI in the growth equation and in the


7 We note that Caselli et al. (1996) find a statistically positive relationship betweenPI and growth when they use ‘revolutions’ as their measure of PI.

investment equation are statistically insignificant while the coef-ficients of the broad measure of PI in these equations are significant.This suggests that elite political instability has no direct impact oneconomic growth or on investment in the current year. The onlyimpact elite PI has on growth is its inter-temporal effect throughreduced investment. This may explain why research that use cross-national data based on single equation models and measure PI as elite

Table 6: Estimates of the Growth Model with Elite PI

Coefficient estimatesVariable g k PI

g – 0.1247 (1.7623)a –0.0774 (6.679)k 0.3594 (2.212) – –kt–1 – 0.5259 (7.576) –PI 0.5756 (1.274) –0.0729 (0.126) –PIt–1 – –1.9918 (1.823) –0.0873 (0.695)x 0.0454 (4.1805) – –l 0.5946 (1.655) – –yt–1 –0.0185 (0.5714) – 0.0028 (3.302)s – 0.2408 (2.378) –m – 0.5015 (10.038) –rds – –0.1478 (3.4864) –mil – – –0.0308 (1.033)lgsl – – –0.3447 (4.446)frac – – 0.4130 (1.2418)Joint test forsignificance

29.2196 334.171 72.346

Sargan test 6.8854 16.846 16.62121st order serialcorrelation

1.154 –1.022 0.018


0.613 0.112 0.011

Wald test 2.33 (PI) 7.857 (PI, PIt–1) 35.772 (g)Hausman exogeneitytest

52.49 98.276 176.217



PI has produced mixed results. It is clear that elite PI cannot captureall the growth effects the broad measure of PI we use is able to capture.Our more comprehensive measure of PI, on the other hand, reflectsinstability in political and social institutions that is likely to lead to bothshort- and long-term economic stagnation. This economic effect islikely to be more widespread. Elite PI, however, only reflects executiveinstability which economic agents may learn to discount over time.The implication is that a comprehensive measure of PI is more capableof capturing the deleterious effects it has on growth than elite PI.

Our results that PI affect economic growth both directly andindirectly through reduced investment in the current period as well asinter-temporally have important policy implications for economicdevelopment in LDCs. If political instability has a negative impact oneconomic growth, then it should be considered as an integral part ofeconomic development policy formulation and implementation. Atthe least, policy makers should make sure that economic policy doesnot disrupt political stability. For example, structural adjustmentprogrammes imposed by the IMF and the World Bank may createshort-run hardships which in turn may increase political instabilityand ultimately have a negative impact on short-term economicgrowth. On the other hand, failure to carry out the necessary economicreforms will create long-term economic stagnation and hence worsenand prolong the period of political instability. An implication from ourresults is that these programmes may best be accompanied by effortsto elicit political support for these policies. Alternatively, economicreforms can be linked to political reforms to reduce uncertainty aboutpolitical stability, which is necessary for the successful implementationof economic policies. This can reduce the likelihood that resources willbe misallocated to doomed development efforts. After all, goodpolicies are useless unless they are effectively implemented, andeffective implementation of policies may not be possible withoutpolitical stability. Another policy implication is that it is in thebest interest of all governments (democratic or autocratic) to promoteeconomic growth as a way of ensuring political stability.

The results indicate that simultaneous equations models usingdynamic panel estimation methodology provide insights into thechannels through which PI affects economic growth. For example, ourestimates indicate that, while the direct effect of PI on growth isstatistically significant, there also exists a statistically significantnegative indirect inter-temporal effect that PI has on economic growth


through reductions in capital accumulation. Simultaneous equationsmodels appear to be useful tools in uncovering the indirect mechan-isms through which PI affects economic growth. More important,dynamic panel estimation methods allow the researcher to account forfixed effects and the endogeneity of regressors, as well as explore thedynamic nature of the growth–PI relationship. Our results also showthat economic stagnation leads to political instability which in turnleads to economic stagnation — a bi-directional causal relationshipbetween PI and economic growth. The results of this study also implythat a more comprehensive measure of PI which includes non-elite PIevents such as riots and strikes will more completely capture therelationship between PI and economic growth in LDCs than the elitepolitical instability events that have been used in previous research.

5. Conclusion

This paper uses a simultaneous equations model, pooled cross-national time-series data and dynamic panel estimation methodologyto investigate the effects of political instability on economic growth inSub-Saharan Africa. Using a more comprehensive measure of politicalinstability than has been used in earlier studies, we confirm the resultsof researchers who find that political instability has a negative andstatistically significant impact on economic growth in Sub-SaharanAfrica. Moreover, the estimates presented here provide evidence thatpolitical instability affects economic growth both directly as well asindirectly through a reduction in capital formation. Using 3 yearaverages of the data to estimate the model indicates that our results arenot being driven by cyclical movements in the data. We are also able toconfirm the conclusions of both political scientists and economists. Wefind that there is a bi-directional causal relationship between economicgrowth and political instability — slow economic growth causespolitical instability which in turn leads to further economic stagnation.The implication of these results is that economic policy makers inSub-Saharan Africa should consider the impact of policy changes onpolitical stability when formulating and implementing economicpolicy. Better still, economic policy could be combined with policiesdesigned to increase political stability. Additionally, we have de-veloped a measure of political instability via principal componentsmethodology which does not depend on arbitrary weights. Thebroader measure of political instability we use here is better able to


capture the effects of political instability on economic growth andinvestment than the elite political instability that has been used byother researchers.

We hope that our efforts will generate research interest in therelationship between institutional development generally (and polit-ical instability in particular) and economic growth in LDCs. Areas forfuture research include identification of additional indirect channelsthrough which political instability influences economic growth inLDCs, identification of additional institutional factors which impacteconomic growth in LDCs, and the development of a widely acceptedmeasure of political instability which can be used by policy makersand researchers alike.

References

Alesina, A. and R. Perotti (1994) ‘The Political Economy of Growth: ACritical Survey of the Literature’, The World Bank Economic Review, 8(3): 351–71.

Alesina, A.S., N. Ozler and P. Swagel (1995) ‘Political Instabilityand Economic Growth (revised)’, NBER Working Paper No. 4173,Cambridge MA.

Apraku, K.K. (1991) African Emigres in the United States: A Missing Linkin Africa’s Social and Economic Development, New York: Praeger.

Arellano, M. and S. Bond (1991) ‘Some Tests of Specification for PanelData: Monte Carlo Evidence and Application to EmploymentEquations’, Review of Economic Studies, 58: 277–97.

Balassa, B. (1978) ‘Exports and Economic Growth: Further Evidence’,Journal of Development Economics, 5: 181–9.

Ball, N. (1984) Third World Security Expenditures: A Statistical Compen-dium, Stockholm: National Defense Research Institute.

Baltagi, B. (1995) The Econometric Analysis of Panel Data, New York:Wiley.

Banks, A.S. (1993) Cross-national Time-series Data Archives, Center forSocial Analysis, State University of New York at Binghampton,Binghampton NY.

Barro, R.J. (1991) ‘Economic Growth of Cross-section of Countries’,Quarterly Journal of Economics, 106, 407–43.

Caselli, F., G. Esquivel and F. Lefort (1996) ‘Reopening the ConvergenceDebate: A New Look at Cross-country Growth Empirics’, Journal ofEconomic Growth, 1: 363–89.


Chamberlain, G. (1983) ‘Panel Data’, in Z. Griliches and M.D.Intriligator (eds), Handbook of Econometrics, Vol. 2, Amsterdam: NorthHolland.

Deaton, A. and R. Miller (1995) International Commodity Prices, Macro-economic Performance, and Politics in Sub-Saharan Africa, PrincetonStudies in International Finance, No. 79.

Devereux, M. and J.F. Wen (1996) ‘Political Uncertainty, CapitalTaxation and Growth’, Mimeo, University of British Columbia.

Dunteman, G. (1984) Multivariate Analysis, Beverly Hills: SagePublications.

Edwards, S. (1998) ‘Openness, Productivity, and Growth: What do weReally Know?’, Economic Journal, 108: 383–98.

Edwards, S. and G. Tabellini (1991) ‘Political Instability, PoliticalWeakness,and Inflation: An Empirical Analysis’, NBER WorkingPaper No. 3721, Cambridge MA.

Feder, G. (1983) ‘On Exports and Economic Growth’, Journal ofDevelopment Economics, 12: 59–73.

Fosu, A.K. (1992) ‘Political Instability and Economic Growth: Evidencefrom Sub-Saharan Africa’, Economic Development and Cultural Change,40 (4): 823–41.

Gyimah-Brempong, K. and T. Traynor (1996) ‘Political Instability andSavings in Less developed Countries: Evidence from Sub-SaharanAfrica’, Journal of Development Studies, 32 (5): 695–714.

Hausman, J.A. (1978) ‘Specification Tests in Econometrics’, Econo-metrica, 46 (6): 1251–71.

Hirschman, A.O. (1956) A Strategy of Economic Development, NewHaven: Yale University Press.

Holtz-Eakin, D., W. Newey and H. Rosen (1988) ‘Estimating VectorAutoregressions with Panel Data’, Econometrica, 56: 1371–95.

Islam, N. (1995) ‘Growth Empirics: A Panel Data Approach’, QuarterlyJournal of Economics, 110: 1127–70.

Jackman, R.W. (1978) ‘The Predictability of Coups d’état: A Model withAfrican Data’, American Political Review, 72: 1263–75.

Jenkins, C.J. and A.J. Kposowa (1990) ‘Explaining Military Coups d’état:Black Africa 1957–1984’, American Sociological Review, 55 (6): 861–75.

Johnson, T., R.O. Skater and P. McGowan (1984) ‘Explaining AfricanCoups d’état, 1960–1982’, American Political Science Review, 78: 622–40.

Knack, S. and P. Keefer (1995) ‘Institutions and Economic Performance:Cross-country Tests using Alternative Institutional Measures’,Economics and Politics, 7: 207–27.


Knight, M., N. Loayza and D. Villanueva (1993) ‘Testing the Neo-classical Theory of Economic Growth: A Panel Data Approach’, IMFStaff Papers, 40 (3): 512–41.

Krueger, A.O. (1980) ‘Trade Policy as an Input to Development’,American Economic Review Papers and Proceedings, 70: 288–92.

Kuznets, S. (1966) Modern Economic Growth: Rate, Structure, and Spread,New Haven: Yale University Press.

Londregan, J.B. and K.T. Poole (1990) ‘Poverty, The Coup Trap, and theSeizure of Executive Power’, World Politics, 42 (2): 151–83.

McGowan, P. and T.H. Johnson (1994) ‘African Military Coups d’état andUnderdevelopment: A Quantitative Historical Analysis’, Journal ofModern African Studies, 22 (4): 633–66.

North, D.C. (1990) Institutions, Institutional Change and Economic Per-formance, New York: Cambridge University Press.

Nuxoll, D.A. (1994) ‘Differences in Relative Prices and InternationalDifferences in Growth Rates’, American Economic Review 84 (5):1423–36.

O’Kane, R. (1981) ‘A Probabilistic Approach to the Causes of Coupsd’état’, British Journal of Political Science, 11: 287–308.

– (1983) ‘Towards and Examination of the General Causes of Coupsd’état’, European Journal of Political Research, 11: 27–44.

Olson, M. (1963) ‘Rapid Growth as a Destabilizing Force’, Journal ofEconomic History, 23 (4): 529–52.

Ozler, S. and D. Roderik (1992) ‘External Shocks, Politics, and PrivateInvestment: Some Theory and Empirical Evidence’, Journal ofDevelopment Economics, 39: 141–62.

Powelson, J.P. (1972) Institutions of Economic Growth, Princeton:Princeton University Press

Rae, D. (1968) ‘A Note on the Fractionalization of Some European PartySystems’, Comparative Political Studies, 1: 413–18.

Ram, R. (1987) ‘Exports and Economic Growth in DevelopingCountries: Evidence from Time Series and Cross Section Data’,Economic Development and Cultural Change, 36 (1): 51–72.

Sanders, D. (1981) Patterns of Political Instability, New York: St Martin’sPress.

Scully, G.W. (1988) ‘The Institutional Framework and EconomicDevelopment’, Journal of Political Economy, 96 (3): 652–62.

Svensson, J. (1998) ‘Investment, Property Rights and PoliticalInstability: Theory and Empirical Evidence’, European EconomicReview, 42 (7): 1317–41.


Wells, K. (1992) ‘South African Violence Wounds Economy’, Wall StreetJournal, September 30.

World Bank (1984) Towards Sustained Development in Sub-Saharan Africa:A Joint Program of Action, Washington DC: World Bank Publications.

World Bank (1992) African Economic Indicators, Washington DC: WorldBank Publications.

World Bank (1995) *STARS* World Data Tapes, Washington DC: WorldBank Publications.

Appendix

The results presented in the body of the paper are based on the estimates fromthe two-step estimator which is the optimal dynamic GMM estimator. Theone- and two-step estimators are asymptotically equivalent if and only if theerror terms are spherical. Given our data, it is most likely that the errorstructure is heteroskedastic even if they are not correlated over time. Wepresent and briefly discuss here estimates from the one-step estimator andcompare them with those from the two-step estimator. The estimatespresented are based on the broad measure of PI and correspond to theestimates in Table 2.

Coefficient estimates from the one-step estimator are presented in Table A1.Column 2 presents the estimates for the growth equation and column 3 theestimates for the investment equation, while column 4 presents the estimatesfor the PI equation. It appears that the one-step estimator provides reasonableestimates for the model. As in the two-step estimates, there is evidence of theexistence of endogenous regressors in all three equations and PI negativelyimpacts growth both directly and indirectly through investment. There is alsoan indication that fast economic growth decreases political instability.

In the growth equation, the coefficients of k, x and l are positive andsignificant while the coefficients of PI and yt–1 are negative and significant.The signs of these coefficient estimates are in accord with expectations. In theinvestment equation, the coefficients of g, kt–1, s and m are all positive asexpected, and significantly different from zero at α = 0.10 for a one-tailed test.The coefficients of PI, PIt–1 and rds are negative but that of PI is not significant.In the PI equation, the coefficients of PIt–1 and frac are positive and significantas expected while those of g, yt–1, mil and lgsl are negative. However, onlythose of g and lgsl are significant. In addition, the PI equation appears not tobe well estimated in terms of signs and precision of coefficients by theone-step estimator.

The estimates from the one-step estimator presented in Table A1 are similarto those of the two-step estimator in Table 2. However, there are somedifferences. In general, the two sets of coefficients differ in terms of magnitudeand a few of the coefficients also differ in terms of signs. For example, the


coefficient of k in the growth equation from the one-step estimator is about30% less than its counterpart from the two-step estimator while the estimatesof PI and yt–1 are larger than their two-step counterparts. With the exceptionof l, coefficient estimates of the growth equation from the one-step estimatorare less precisely estimated than their counterparts from the two-stepestimator. In the k equation, the coefficients from the one-step estimator aregenerally larger than those of the two-step estimator. However, all estimatesfrom the one-step estimator in this equation are less precisely estimated than

Table A1: Dynamic Panel GMM Estimates of Model: One-step Estimator CoefficientEstimates


g – 0.1580 (1.554) –0.0098 (1.7891)k 0.1689 (1.9428)a – –kt–1 – 0.2636 (8.318)PI –0.3809 (1.8972) –0.2901 (0.9032) –PIt–1 – –0.8081 (1.6245) 0.1466 (3.4937)x 0.0588 (5.0891) – –l 0.6626 (2.2681) – –yt–1 –0.0276 (3.6085) – –0.0001 (1.1414)s – 0.7051 (7.8987) –m – 0.7033 (17.9169) –rds – –0.3448 (3.1663) –mil – – –0.0048 (1.1089)lgsl – – –0.3792 (3.6113)frac – – 0.4045 (1.8634)Joint test significance 40.927 (5) 222.675 (6) 63.921 (6)Sargan 16.131 (19) 45.626 (18) 14.5657 (19)1st order serialcorrelation

–4.075 3.380 –4.108


–2.101 0–2.247 –1.673

Hausman exogeneitytest

89.6341 72.342 341.860




their two-step counterparts. Similar differences can be observed in theestimates of the PI equation. Besides differences in magnitude and statisticalsignificance of the coefficient estimates, there are sign reversals in the PIequation when one compares the one-step and the two-step estimates. Theone-step coefficients of mil and yt–1 are negative contrary to the expectedpositive signs as in the two-step estimates.

Although the one-step estimates on these coefficients are statisticallyinsignificant, the unexpected negative signs are troubling. The one-stepestimates of the PI equation are less precise than their two-step counterparts.These differences in the estimates from the two estimators leads us toconclude that while the one-step estimates tell a similar story as their two-stepcounterparts, there are some quantitative and qualitative differences in thestories told by the two sets of estimates.


political instability, investment and economic growth in sub-saharan africa

Documents