Does democracy enhance economic growth?
An extensive dynamic panel data analysis using a newly composed datasetAbstract - This study investigates the effect of a country’s level of democracy on economic growth. Using a newly composed dataset from the V-Dem project with data ranging from 1986-2010 for 100 countries, a dynamic panel data analysis is performed. Applying the Arellano-Bond one-step and two-step estimation methods, a model is estimated which is adequate according to the various dynamic panel data misspecification tests. It is found that a more authoritarian regime enhances short term economic growth more than a democracy, but performs worse in the long run. An explanation for this result might be that dictatorships make profitable short-term decisions but neglect the probable negative long-term effects. To investigate the aforementioned effect in more unstable countries, a similar analysis is performed on a subset of 72 lower-income countries. This results in similar, but less significant, coefficient estimates for democracy and the control variables, implying the model is quite robust. Although this study has methodological limitations, a variety of standard panel data techniques have been applied to a well-built dataset, and so the findings seem to merit further investigation.
Thesis MSc. Econometrics
Author: Mark de Boer Student number: 6035779 Supervisor: Prof. Dr. J.F. Kiviet Second reader: Dr. K.J. van Garderen June 16, 2017Statement of originality
This document is written by Mark de Boer who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Preface
I would like to thank Jan Kiviet for his supervision during the process of writing this thesis. From the first lectures about the wonderful world of panel data to the last meeting about the technical details of my thesis, he has always been a great help and very pleasant to work with. Furthermore, I would like to thank Kees Jan van Garderen for taking the time out to read my thesis.
The most important person I want to mention here is my father Roel. From primary school, where he made me enthusiastic for mathematics, to high school, when he spent (too) much time to keep me motivated, to university, where he was still able to assist me despite his lack of knowledge about econometrics; I could not have wished for a smarter and more patient person to support me in lots of ways. Although my school career has come to an end, I hope I can still make use of his (sometimes unsolicited) advice during my actual career.
”If God had meant there to be more than two factors of production, He would have made it easier for us to draw three-dimensional diagrams.”
Contents
1 Introduction 1
2 Literature overview 2
2.1 The Solow-Swan model and its extensions . . . 2
2.1.1 The mathematics of the model . . . 2
2.1.2 Shortcomings and extensions of the Solow-Swan model . . . 4
2.2 Previous literature . . . 5
3 Model & Methods 8 3.1 General model on economic growth . . . 8
3.2 Arellano-Bond/Blundell-Bond estimators . . . 9 3.3 Estimation methods . . . 11 3.4 Classification method . . . 12 3.5 Long-term effects . . . 13 4 Data 14 4.1 Growth Determinants . . . 14 4.2 Sources . . . 17
4.3 Constructing the dataset . . . 18
4.4 What is a democracy? . . . 19
5 Results 21 5.1 Model on economic growth . . . 21
5.2 Main results . . . 23
5.3 Long-term results . . . 28
6 Conclusion 28
References 31
INTRODUCTION
1
Introduction
The upcoming Brexit is an example of a very democratic process which results in a, most likely, negative economic growth for Great Britain. Contrary, since 1990 Sub-Saharan Africa has an increasing level of democracy and at the same time its economy is also increasing. There also exist examples of ’undemocratic’ countries with economies either expanding or contracting. Easy to find are examples of countries with an autocratic regime in control and meanwhile a negative economic growth. Countries at war (Yemen, Syria) or countries with substantial corruption problems (Brazil, Russia) are frequently subject to negative economic growth. On the other hand, China’s one-party system also scores poorly on the level of democracy, but this does not seem to affect its economy negatively. The well-known The New York Times journalist Tom Friedman explains this as follows:
“One-party autocracy certainly has its drawbacks. But when it is led by a reasonably enlightened group of people, as China is today, it can also have great advantages. That one party can just impose the politically difficult but critically important policies needed to move a society forward in the 21st century.” (Friedman, 2009)
In the aforementioned cases one should not see the link between democracy and eco-nomic growth as a causal relationship per se. There are obviously many more factors that influence economic growth. However, these examples do point out that the most straightforward line of thought - more democratic implies more economic growth - should be considered to be erroneous. In the 90’s and 2000s a vast literature investigating the potential effect of democracy on economic growth emerged. Modeling economic growth be-came more popular since Mankiw et al. (1992) proposed their extension of the Solow-Swan (1956) model. Nevertheless, a conclusive outcome of the potential effect of democracy on economic growth has not been given because most studies find mixed effects or no effect at all.
Previous research has shown that the causal mechanism of this relationship between democracy and economic growth is questionable, which might introduce an endogeneity problem. Economic growth can drive social mobilization, which enhances political mo-bilization and results in a change of regime type. Other studies find that demand for democracy decreases once a certain economic threshold has been reached. This study investigates the impact of the level of democracy on economic growth and contributes to the literature by using a newly constructed dataset by the V-Dem project. This dataset has, to the author’s knowledge, never been used to investigate the influence of democracy on economic growth.
LITERATURE OVERVIEW
growth and the research that has been conducted with these models to find a link between the level of democracy and economic growth. Section 3 gives the framework of the adopted estimation methods in this study. In section 4 the data that has been used in this study is described into detail. In section 5 the results are presented, including results where heterogeneity has been reduced. Section 6 concludes, states possible shortcomings and gives recommendations for future research.
2
Literature overview
In this section, first the framework of the neoclassical growth model will be explained, because almost all growth theory builds on this model originally proposed by Solow (1956) and Swan (1956). This paper will not be an exception. Second, an overview of the empirical literature regarding economic growth and its determinants is given. Remarkably, in the last decade the amount of research done in this field has decreased drastically.
2.1 The Solow-Swan model and its extensions
In the classical articles by Solow and Swan, they independently thought of a way to model long-term economic growth. It is an extension to the post-Keynesian Harrod-Domar (1946) model, which had very unstable solutions and various other shortcomings. Therefore, Solow and Swan proposed their famous, mathematically manageable, models that deliver more stable and convenient solutions.
2.1.1 The mathematics of the model
This section partially follows Barro & Sala-i-Martin (1990) and Mankiw et al. (1992)1. In the Solow-Swan model there are only two inputs, Labor, L(t), and physical capital, K(t). The model takes both these variables as exogenous. Physical capital is also referred to as
savings and hence the production function is
Y (t) = F [L(t), K(t)].
Here Y (t) is the total production at time t. This production function follows the Inada conditions, as Inada (1963) stated. We regard this production function neoclassical if the following properties hold:
1
From now on, the model proposed by Mankiw, Romer and Weil (1992) will be referred to as the M-R-W model
LITERATURE OVERVIEW
1. F (·) has positive and diminishing marginal products with respect to each input.
2. F (·) exhibits constant returns to scale.
3. Labor and capital satisfy the Inada (1963) conditions, which means that the elasticity of substitution asymptotically goes to 1.
Now another exogenous variable is introduced, namely the level of technology, A(t). This variable can also be interpreted as the ’knowledge’ of the labor force. By implementing
A(t), the newly formed production function by Solow and Swan is the Cobb-Douglas
production function. This is assumed to reasonably describe the economy.
Y (t) = K(t)α(A(t)L(t))1−α 0 < α < 1
The total number of workers, which is a proxy for labor, grows exogenously at rate n and similarly the level of technology grows at rate g:
L(t) = L(0)nt A(t) = A(0)gt
Therefore, the effective units of labor A(t)L(t) grow at a rate n + g. Now the model assumes that a fraction of the output is constantly consumed and invested. Here cY (t) is consumed so s = 1 − c is left for investments. Furthermore, the stock of capital per effective unit of labor, k = ALK , depreciates at rate δ . Now if y is the output per effective unit of labor, y = ALY the evolution of k can be formulated as
˙k(t) = sy(t) − (n + g + δ)k(t) = sk(t)α− (n + g + δ)k(t)
This equation shows that k is converging to the steady state value sk?α= (n + g + δ)k?, or k? = s n + g + δ 1−α1 (1)
As can be seen from this steady state value of capital to labor, savings have a positive impact and growth rate of the population a negative impact. This makes sense, because more savings in a country means richer inhabitants. On the other hand, a growth of the population means that more people have to share the same amount of savings, which makes everyone worse off.
LITERATURE OVERVIEW
real income. Therefore, equation (1) is substituted into the production function with logs taken on both sides, resulting in
ln Y (t) L(t) = lnA(0) + gt + α 1 − αln(s) − α 1 − αln(n + g + δ)
From the result two conclusions follow. First, as mentioned above, this model predicts the signs of the coefficients on saving and population growth. Second, it can be seen that the third property of the Inada condition holds, since the function is strictly increasing in s. Because the capital’s share in income (α) is roughly one third, this model implies that the elasticity of income per capita with respect to s is practically 0.5 and the elasticity with respect to n + g + δ is practically −0.5.
2.1.2 Shortcomings and extensions of the Solow-Swan model
The Solow-Swan model has been subject to various extensions. Some important ones are the contributions by Cass (1965) and Koopmans (1965). This paper will not elaborate on these extensions, but focuses on an important and well-cited paper by Mankiw et al. (1992).
As Mankiw et al. (1992) point out, the Solow-Swan model is consistent with the evidence to a certain point. Using their more recent data set, they find that the signs of the coefficients for savings rate and population growth are positive and negative, respectively. This corresponds to generally accepted expectations. However, they do have criticism towards the magnitudes of the impact of the savings rate and population growth in the model. The contributions of both, to the level of output, are too large and thus the model seems incorrectly specified. Furthermore, the Solow-Swan model does not explain why poor countries do not converge to the same level of wealth as rich countries. Kendrick (1976) already estimated that half of the United States’ physical capital consisted of human capital. Therefore, Mankiw et al. include accumulation of human capital, H(t), to the model, changing the original Cobb-Douglas production function into
Y (t) = K(t)αH(t)β(A(t)(L(t))(1−α−β) (2)
where the stock of human capital depreciates at the same rate δ as physical capital. Poor countries have less human capital than rich countries, which explains why different countries converge to different steady states. Now the fraction of output s is split up into the fraction invested in physical capital, sk, and the fraction invested in human capital,
LITERATURE OVERVIEW
defined as
˙k(t) = sky(t) − (n + g + δ)k(t),
˙h(t) = shy(t) − (n + g + δ)h(t),
Hence, assuming that α + β < 1 , the economy converges to the following steady state
k? = s 1−β k s β h n + g + δ !1−α−β1 h?= s α ks 1−α h n + g + δ ! 1 1−α−β
Substituting this in (1) and taking logs on both sides gives
ln Y (t) L(t) = lnA(0) + gt − α + β 1 − α − βln(n + g + δ) + α 1 − α − βln(sk) + β 1 − α − βln(sh)
From this equation it can be observed how the output per capita is derived from the population growth, physical capital and human capital in a country.
In their analysis, Mankiw et al. (1992) find that human capital is indeed correlated with both physical capital and population growth. Also, they find that 80% of the income per capita in a country can be explained by their model. Therefore one can say that, in hindsight, Robert Solow’s quote at the beginning of this study is based on an incomplete growth model and that he maybe did not foresee the potential of computational power. In the next section there will be an overview of how different authors used the M-R-W model and how they interpreted human capital.
2.2 Previous literature
One of the first studies that utilized the M-R-W model is Knight et al. (1993). They used a panel of time series cross-sectional data from 1960-1980 to increase the number of data, opposed to M-R-W, who used only one cross-section in their analysis. Hence, for M-R-W, the time-specific effect becomes irrelevant. Knight et al. come up with proxies for the level of technology, physical and human capital.2 They conclude that all these factors exert a
positive and significant effect on economic growth.
The main focus of our study is the impact of political (in)stability on economic growth. In the past, several papers have been written about this subject. Przeworski and Limongi
2For physical capital they use as proxy the ratio of real investment to real GDP, for human capital
they use the gross enrollment rate of secondary schooling and for the level of technology the amount of import-export products are the proxy.
LITERATURE OVERVIEW
(1993) investigated the outcomes of 37 different studies in this field, where in 18 of them they could not find any conclusive outcome. In 11 of the studies they examined, they found that an authoritarian regime facilitated economic growth better than democracies. In the other 8 democracies performed better. Some articles argue that democracy does not have a direct effect on economic growth. Helliwell (1994) states that including the level of democracy as a proxy for human capital is wrong because he finds that democ-racy has a small direct negative effect on economic growth. He argues that this effect can not be true and so the model must be misspecified. When checking for the indirect effect of democracy through education and investment as proxies for human capital, he finds a much larger, positive effect. These findings are similar to what Baum and Lake (2003) report, although some panel data techniques used in this paper are questionable. Przeworski et al. (2000) show that democracy has an indirect effect on economic growth through population growth.
Although the previously mentioned papers discuss the indirect effect of democracy on economic growth, by far most of the research in this field ignore this potential effect. All of them use an augmented M-R-W model. Barro (1996) analyzed a panel of 100 countries from 1960-1990 and concludes that the overall effect of democracy on growth is weakly negative. He suggests that in a country with a poor growth rate more democracy in-creases growth, but once a certain higher level of democracy is reached a further increase in democracy depresses the growth rate.
In the recent paper by Moral-Benito (2014) the potential effect of 9 variables on growth is investigated, where all the variables (including a variable for the level of democracy) are proxies for the components in the M-R-W model. Interestingly, he finds that none of the regressors has a significant effect on economic growth. This is in contradiction to what Durlauf et al. (2004) found in their survey of 43 studies on economic growth. They found a total of 145 proposed regressors as proxies and each of them had a significant effect in at least one of the studies, but many of them had also an insignificant effect in similar studies. This suggests that it is very difficult to identify the determinants of economic growth.
One of the obstacles for estimating economic growth per country, according to Durlauf
et al. (2004), is the small amount of data, e.g. the number of countries in the world.
Fur-thermore, many countries can not be used in the regressions for various reasons explained later. Harberger (1987) and Solow (1994) voiced their concerns about using the same linear model for a set of potentially very different countries. Durlauf et al. also warn for the problem concerning the interpretation of results. For example, when India and China,
LITERATURE OVERVIEW
together accounting for almost 40% of the world’s population, have a period of economic growth, more people benefit than when Kenya and England perform well. The point is that countries have different population sizes and that must be taken into account when interpreting results that have countries as units of observation.
The question arises whether a country’s current state of democracy has a direct effect on today’s economic growth. It does make sense to think that today’s economic growth is a result of the level of democracy over the passed years. Gerring et al. (2005) argue that too many studies look forward, instead of backwards. In their view, the level of democracy should be seen as a stock variable, a variable with a value that has accumulated over time, rather than a flow variable, a variable that is measured at only one certain point in time. It has been shown that the level of economic growth today can be dependent of the level of democracy years, or even centuries, ago (Collier and Collier, 2002; Mahoney, 2004). In their paper, Gerring et al. find a significant and robust effect of the stock of democracy on economic growth. This is also in line with Persson and Tabellini (2006), who argue that countries with a longer history of high democracy experience more economic growth.
In nearly every paper about economic growth, conditional convergence is discussed; countries with a poor economy will grow faster than a better performing country and will end up in the same level of wealth as any other country. One of the first to do research on this was Baumol (1986) who provided evidence of the existence of conditional convergence in the world. However, his findings have been criticized by many. As De Long (1988) puts it bluntly:
“The answer to this ex ante question – have those nations that a century ago appeared well placed to appropriate and utilize industrial technology converged? – is no”
So, although Barro and Sala-i-Martin (1995) show that in the neoclassical growth model the growth rates decrease as the level of output per capita increase, not every country ends up the same. An explanation, argues Barro (1996), is that in countries with a smaller eco-nomic growth sometimes new technologies can not be implemented due to the bad political situation, geographical restrictions or lack of infrastructure. The advantage of being a fol-lower economy do not apply in that case.
It is remarkable to see that in the last decade the amount of research done in the field of economic growth has decreased significantly. Most literature about this subject has been written in the 90’s and the beginning of the 2000s, presumably because of the very useful extension of the Solow-Swan model in 1992. Furthermore, there is not yet a final judgment about the relationship between democracy and economic growth. Therefore this paper will contribute to this field by analyzing a completely new and very recent data set
MODEL & METHODS
that has, to the author’s knowledge, never been used for this type of research.
3
Model & Methods
In this section, first a general model for economic growth is proposed. This model is the take-off point for the final model. Second, the Arellano-Bond (AB) and Blundell-Bond (BB) dynamic panel estimators are discussed, both using the generalized method of moments (GMM) approach. The AB and BB estimators are designed for “small T , large
N ” datasets and studies. Then the estimation methods are described, followed by the
classification method for the regressors. This section ends with the long-term estimation method.
3.1 General model on economic growth
In a study to investigate the effect of different variables on economic growth, it is advanta-geous to make use of a dynamic model. This means that one or more lags of the dependent variable or of other regressors are added to the right-hand side of the regression equation. If these dynamics are ignored, one assumes the system is time-invariant, and thus in an equilibrium. These so-called static models try to explain what happens when time passes, but time itself is not embodied in the model: one solely investigates the conditions of the equilibrium, not the path of change. However, when conducting a study like ours, it is better to use time-series data instead of taking a ’snap-shot’ of the situation in a country at a certain (arbitrary) point in time.
As mentioned before, several control variables and their lags will be used to create a set of variables that determine economic growth. Furthermore, one or more lags of the depen-dent variable are included since these will likely have a large effect on the current-period dependent variable. In previous research many variables have been included as control variables to model economic growth. As stated, Durlauf et al. (2004) found 145 possible significant regressors in 43 studies. However, in comparable studies it is shown that using approximately 10-20 variables and their possible lags is sufficient, and the inclusion of more variables drastically increases the probability of overfitting the model. As shown in the analysis of the previous literature and especially the M-R-W model, economic growth can be determined by the financial status, political condition, human capital and the level of technology. Certainly more variables exist that influence economic growth, however the list of explanatory variables in this study is fairly complete and proved to be comprehen-sive according to numerous studies in the past.
MODEL & METHODS
The general regression model on economic growth is presented as follows:
gi,t = αgi,t−1+ β10CVi,t+ β20CVi,t−1+ µt+ ηi+ i,t (3)
i = 1, . . . , N ; t = 2, . . . , T
where only the contemporaneous regressors and their first lags are included to represent the dynamics in the model. Obviously, more lags could be included at the right-hand side. In regression model (3), git is economic growth in country i at time t. CVi,t contains
the endogenous, predetermined (or weakly exogenous) and exogenous regressors. More information about the different types of variables will be given in section 3.2. µt are the
time specific fixed effects and ηi are unobserved time-invariant individual effects. i,t is
an error term. The effects of the lagged control variables (CV) are not only represented by β2 but are reflected by the lagged dependent variable too. The next section discusses
an estimator that resolves the problems of the inclusion of the fixed individual effects and the possible endogeneity that may be present in the model.
3.2 Arellano-Bond/Blundell-Bond estimators
One of the issues concerning regression model (3) is that it allows the country specific fixed effects ηi to correlate with CVi,t. Since it is very hard, or even impossible, to observe these fixed effects, one should find ways to eliminate these effects from the regression model. One way to deal with this is to subtract gi,t−1 from gi,t, e.g. by first differencing. This results in the following equation:
∆gi,t = ∆gi,t−1+ β01∆CVi,t+ β20∆CVi,t−1+ ∆µt+ ∆i,t (4)
i = 1, . . . , N ; t = 3, . . . , T
where ηi is differenced out since it is constant over time. In this model endogeneity is
introduced because ∆gi,t−1 correlates with ∆i,t via i,t−1.
As mentioned, CVi,tcontains the exogenous, predetermined and endogenous regressors.
Exogenous variables are uncorrelated with past, present and future errors, predetermined variables are variables uncorrelated with present and future errors and endogenous vari-ables are varivari-ables that are uncorrelated with future errors only. CV is split up in x, w and v, thus this classification implies the following conditions:
MODEL & METHODS E(xiti,s) = 0 E(witi,t+l) = 0 E(viti,t+1+l) = 0 ∀i, t, s, l ≥ 0
Here x contains the exogenous variables, w the predetermined variables and v the en-dogenous variables. Since there is endogeneity present in (4), one should make use of the instrumental variable technique. To find suitable instruments is extremely hard or even impossible for pure cross-section or time-series data. Therefore, Arellano and Bond (1991) introduced an instrumentation method that enables the use of internal instruments, i.e. lags of the explanatory variables, assuming there is no autocorrelation present in the er-rors. Exogenous regressors and their lags can be used as their own instruments in this way and, in model (4), the first and further lags of predetermined regressors are suitable instruments too, as well as second and further lags of endogenous regressors. Hence, nat-ural candidates as instruments when estimating model (4) are ∆gi,t−2 and ∆CVi,t−2. In
the Arellano-Bond framework other possible instruments are the traditional instruments, such as the exogenous time-dummies, or any other instrument that is uncorrelated with the error term.
Another estimator that should be considered in this type of research is the Blundell-Bond estimator, as proposed by Blundell and Blundell-Bond (1998). They developed a method, building on the work of Arellano and Bover (1995), that increases efficiency under an addi-tional assumption. This assumption is that the fixed effects are uncorrelated with the first differences of the regressor variables in model (3). In addition to Arellano and Bond, who difference (3) to remove the individual-specific fixed effects from the model and exploit lags of the regressors of (3) as internal instruments for equation (4), they use lags of the regressors of (4) as instruments for equation (3). The use of these extra moment conditions results in a higher estimation efficiency than when using the Arellano-Bond counterpart and, if the extra assumption holds, Blundell-Bond estimation should be preferred. Sim-ilar to Arellano-Bond, Blundell-Bond allows for the use of time-dummies as exogenous instruments. For more detailed information on the Arellano-Bond and Blundell-Bond es-timators, see Roodman (2009). In both estimation methods a unit root should be absent, meaning |α| < 1. The implications of a unit root will be discussed in section 3.5. The next section describes the application of these estimation methods.
MODEL & METHODS
3.3 Estimation methods
In this study, four different versions of the Arellano-Bond estimator will be utilized. First, Arellano-Bond using the one-step and two-step GMM estimators are used. These are abbreviated as AB1 and AB2 for the one-step and two-step estimation methods, respec-tively. Furthermore, the effect of reducing the number of instruments is tested. This is done by collapsing the instrument matrix, a technique that ’collapses’ some instruments into one column. Roodman (2009) elaborates on this technique more thoroughly. The one-step and two-step Arellano-Bond estimators with collapsed instrument matrix will be abbreviated as AB1c and AB2c and when the full instrument matrix is used as AB1f and AB2f. Reducing the number of instruments generally reduces the bias and thus improves the size properties of coefficient tests. However, as Kiviet et al. (2014) point out, this possibly implies power loss because efficiency suffers. Furthermore, collapsing the instru-ment matrix reduces the size problems of the Sargan-Hansen test, an overidentification test that will be discussed below.
When applying AB1 estimation, the standard errors can be made heteroskedasticity robust. AB2 already yields heteroskedasticity robust standard errors, however there is a possible downward bias of the estimated asymptotic standard error present. This is addressed by the Windmeijer correction, proposed by Windmeijer (2005). According to Roodman (2009a), AB2 with the Windmeijer correction is preferable over AB1 with robust standard errors. However, Kiviet et al. (2017) find that after collapsing the instrument set, the AB2 estimator with Windmeijer correction does not perform better than AB1 with robust standard errors. Therefore, throughout this study both the AB1 and AB2 estimators will be treated as equally important.
The one-step (BB1) and two-step (BB2) Blundell-Bond estimators are treated in the same way as their Arellano-Bond counterparts. Thus, heteroskedasticity robust standard errors are requested when performing BB1 estimation and the Windmeijer correction is applied for BB2. Furthermore, the instrument set can also be collapsed. The instrument set will contain more instruments than when Arellano-Bond is used, since not only the levels of variables are used as instruments, but also their first differences.
When performing a panel data analysis using Arellano-Bond or Blundell-Bond, three specification tests have to be considered. The first one is the Sargan-Hansen test of overidentifying restrictions (Sargan, 1958 and Hansen, 1982), which tests the validity of instruments.3 The null hypothesis of this test is validity of the instrument set, hence we do not want to reject. Something that should be approached with suspicion is the p-value
3
MODEL & METHODS
of this test when the number of instruments is large. Roodman (2009) points out that the Hansen test is weakened by many instruments and therefore the coefficient estimation re-sults should be interpreted thoughtfully. Moreover, a Hansen test that is barely exceeding a significance level of 0.05 is not comforting, in particular if the test is oversized. Rood-man also stresses that results should be tested for sensitivity to reduction in the number of instruments.
The second and third specification tests that have to be considered are the tests for presence of first- and second order autocorrelation, AR(1) and AR(2). AR(1) is expected in first differences, E(∆i,t∆i,t−1) 6= 0, since in this equation both factors share the com-mon factor i,t−1. AR(2) is not expected in first differences, E(∆i,t∆i,t−2) = 0, since the
first factor contains i,t and i,t−1 and the second factor i,t−2and i,t−3. If the AR(2) test rejects, that points into the direction of wrongly omitted variables from the model. The null hypothesis of both the AR tests is the nonexistence of autocorrelation of the residuals in first differences. Hence, the AR(1) test should reject and the AR(2) test should not reject. The extra assumption for Blundell-Bond estimation, the first-difference of lagged regressors being uncorrelated with the fixed effects, is also tested by the Hansen test.
3.4 Classification method
This section explains into detail how the classification of regressors is determined, e.g. whether a regressor should be considered as endogenous, predetermined or exogenous. This is done making use of the incremental Hansen test. The incremental Hansen test tests whether a specific group of instruments is valid, under the assumption that all other instruments are valid.
First: all unlagged regressors from the general model on economic growth (1) are treated as endogenous. This means that the instrument set consists of only the second and higher-order lags of the regressors. The furthest lag of any regressor that is used as instrument is the third lag and the instrument matrix may be collapsed. The reason for not using the full instrument set is that in this way a fine trade-off between the number of instruments and a relevant instrument set is made. Furthermore, the author of this study believes fourth and further lags to be of no or hardly any influence. The reason for this is that every lag is a time period of four years, which is elaborated on in section 4.3.
Second: as long as the specification tests still have acceptable p-values, regard one of the endogenous regressors to be predetermined. This is done by additionally including the first lag of this regressor as instrument. If the incremental Hansen test for these instruments has a p-value of greater than 0.3, this regressors may be considered to be predetermined, and
MODEL & METHODS
otherwise endogenous. This is done sequentially for all regressors in the model, starting with the variable with the highest p-value of the incremental Hansen test. Note that the results might be different if this sequence were started with another regressor. In the same manner, the regressors now regarded as predetermined will be examined to see if they are exogenous. The only difference is that in this case the contemporaneous realization of the regressors are used as instruments. Note that during this classification process, the overall Hansen and autocorrelation tests have to be satisfied.
The results of this method are given in section (5.1). The lagged dependent variable,
gi,t−1, is obviously endogenous and the time-dummies are exogenous and therefore treated
as ordinary instruments. On a final note: there exists no overall consensus on what the optimal approach for the classification of regressors is. The p-value of 0.3 is questionable and during the whole process one should keep in mind that every dataset has its own characteristics that possibly have to be controlled for.4
3.5 Long-term effects
As showed by Collier and Collier (2002), the level of democracy at this time can have an impact on the economic growth a long time from now. To show this, take a look at the following simplified model of (2):
gt= α + ρgt−1+ γxt+ t (5)
t = 2, . . . , T
Here xt directly effects gt, but xtalso has an effect on gt+1through the lagged dependent variable. The size of this effect is ργxt. In the same manner, the effect of xt on gt+2 is
ρ2γxt, the effect of xt on gt+3 is ρ3γxt, and so on. Now the instantaneous and delayed
effects can be summed up to infinity, which will give the cumulative effect of xt on gt.
Making use of the sum of this geometric series, the long-term effect of one unit change of xt on gt is γ/(1 − ρ). This method can be applied to models with more lags of the
dependent and independent variables as well, but the principle stays the same. To be able to calculate the total effect of a one unit increase of xt on gt, (5) is rewritten as follows:
∆gt= γxt− (1 − ρ)gt−1+ t (6)
t = 2, . . . , T
4
Choosing a higher p-value gives a higher probability of a type I error, whereas a lower p-value increases the probability of a type II error.
DATA
It can be seen that if ρ = 1, so a unit root is present, there is no convergence since 1 − ρ = 0. Contrary, if ρ = 0 and thus 1 − ρ = 1, there is no dynamic adjustment process since there is a steady state in each stage. gtis directly influenced by any fluctuation from
the long-term relation.
Hence, if 0 < ρ < 1, then lim
t→∞ρ
t = 0. This shows that the effect of a shock is
only temporary, since the effect decreases when time moves on. This can be observed in equation (5): if gt−1 is larger than its equilibrium, −(1 − ρ) < 0 adjusts for this. Using the STATA command nlcom the long-term effect of democracy on economic growth can be obtained, along with its 95% confidence interval. The calculations of this command are based on the ’delta method’, which is in large samples a convenient approximation.
4
Data
In this section first the main- and control variables from the dataset that will be used to determine economic growth are discussed. Second, the composition of the dataset is explained. Thirdly, which subset is analyzed and how is dealt with gaps in the data is elaborated on. Lastly, what is measured to determine the level of democracy in a country is described, following the seminal work from Przeworski (2000).
4.1 Growth Determinants
As mentioned previously, the M-R-W model takes into account four determinants of eco-nomic growth: physical and human capital, population growth and initial GDP growth. Durlauf et al. (2005) found over 140 potential regressors and all of them can be regarded as a proxy for one of these determinants.
The explanatory variables in this study consist of only a small subset of the growth determinants found by Durlauf et al. (2005), but is comparable to the number of regres-sors used in previous, similar research. As shown in Ciccone and Jarocinski (2010) and Moral-Benito (2012), the inclusion of many proxies for the same growth determinant leads to inaccuracies due to multicollinearity and measurement errors. Therefore, in this paper a maximum of four regressors proxy for the same determinant.
Durlauf et al. (2008) specify seven growth theories and associate each regressor to at least one of these.5 In this study, three of these categories will not be considered: religion, geography and ethnic fractionalisation. All the proxies that are commonly used for these categories are almost always time invariant and thus are captured by the fixed effects in
5The seven theories are: neoclassical growth theory, demography/health, macroeconomic policy,
DATA
the model.
Hence, the following variables are used in the model:
• GDP growth. The dependent variable in this study. As usual, GDP growth is calculated as gi,t = ln(GDPi,t) − ln(GDPi,t−1). In Figure 1 the changes in GDP over the years can be observed.
• Level of democracy. Following Baum and Lake (2003), this variable proxies for hu-man capital. Thus far, studies that investigate in what hu-manner the level of democracy influences economic growth point in different directions, or find no significant effect at all. Both the Freedom House and Polity measures for the level of democracy are transformed to a 0-10 scale and subsequently the average is taken.6 This value runs from 0 (dictatorial regime) to 10 (completely democratic). How the level of democracy is determined is explained in section 4.4.
• Log Life expectancy. In various studies, life expectancy is used as a proxy for human capital. Specifically, the log of life expectancy in years is taken as an indicator of the health status of a country, which usually has a positive impact on GDP growth. Bloom et al. (1998) show that a low life expectancy is associated with lower savings and government investment and thus a slower economic growth. Gallup and Sachs (2001) state that the extermination of malaria in sub-Saharan Africa will increase GDP by 2,6%.
• Corruption. The level of corruption in a country is measured on a scale from 0 (no corruption), to 1 (high corruption). The average is taken from the indexes of a) public sector corruption, b) executive corruption, c) legislative corruption and d) judicial corruption. Mauro (1995) finds evidence for a higher economic growth once corruption decreases. Furthermore, Welsch (2004) claims that, in particular for low-income countries, a lower level of corruption increases the economic as well as the environmental state of a country.
• Savings/GDP. Savings as fraction of GDP is one of the three financial variables in the dataset, along with government expenditure and inflation. Following the previously mentioned Harrod-Domar model, it is expected that the order of causality is more savings to more investment to more economic growth. Savings are calculated as gross national income less total consumption, plus net transfers. For ease of interpretation, this number is multiplied by 100.
• Government expenditure/GDP. Barro (1990, 1991) was the first to find a negative relationship between government expenditure and economic growth. Other authors
6
For more information about these two measures for democracy, visit www.freedomhouse.org and http://www.systemicpeace.org/polity/polity4.htm
DATA
used the percentage of government consumption of GDP and found similar results.7. Barro’s explanation behind these findings is that a higher government expenditure lowers savings and economic growth through the distorting effects from taxation or government-expenditure programs. Government expenditure is constructed as fraction of GDP.
• Export/GDP. This variable is a proxy for the level of technology of a country. In Alcala and Ciccone (2004), it is found that a higher fraction of export to GDP has a significant positive effect on economic growth. However, this variable has also been subject to various similar studies where no significant effect was found, for example Rodrik et al. (2004) observe an insignificant negative effect. For ease of interpretation, this number is multiplied by 100.000.
• Inflation. Many authors find a negative effect of inflation on economic growth. Bruno and Easterly (1998) find that countries that endure high inflation shocks have a large decrease in economic growth, but recover remarkably strongly once inflation falls. Inflation is used as the GDP deflator in annual percentages.
• Years of education above 15 years. Along with secondary school enrollment, this variable represents the educational level of a country’s workforce, and thus is a proxy for human capital in the neoclassical growth model. Theoretically one would suspect that it would have a positive effect on economic growth, however many authors fail to find such an effect.
• Secondary school enrollment. Similar to the Years of education above 15 years, the overall consensus is that this variable should have a positive effect but only few stud-ies have found a significant effect. Secondary school enrollment is operationalized as the percentage of the secondary school-aged population that is enrolled in secondary school.
• Armed conflict. Among others, Sala-i-Martin (1997a,b) includes a war dummy in his dataset and finds a significant negative effect. The dummy has a value of 1 when a country participated in an internal or external war, or endured a civil war in that given time-period.
• Mortality rate. Another proxy for human capital. Mortality rate, as stated by Kalemli-Ozcan et al. (2000), is one of the most influential aspects of the process of economic growth. In the V-Dem dataset, the infant mortality rate is measured as the number of deaths prior to age 1 per 1000 live births in a year.
7
DATA
• Population growth. Following from the neoclassical growth model, a growth in pop-ulation should have a negative effect on economic growth since every investment is shared among more people. Bloom et al. (1998) show that a large contributer to the slow economic growth in Africa is the low willingness to lower the fertility rate. Many other authors find similar results. Population growth is calculated as the difference of the logs of population size in subsequent periods.
1987 1990 1993 1996 1999 2002 2005 20082010 0 80 160 240 320 400 480 560 640 Real GDP x 10 9
(1a) Average GDP per year
1987 1990 1993 1996 1999 2002 2005 20082010 −0.05 0 0.05 0.1 0.15 0.2 Real GDP rate of change
(1b) Average GDP rate of change per year
Figure 1: The graph on the left shows average real GDP in USD in levels. On the right is the rate of change per year plotted, calculated as the difference of the natural logarithm of real GDP. Both graphs are plotted using a dataset of N = 100 countries.
4.2 Sources
The main contributer to the dataset used in this study is the dataset from the V-Dem project. The V-Dem dataset has been created in a collaboration between the Department of Political Science at the University of Gothenburg and the Kellogg Institute at the University of Notre Dame. The most actual version, v6.2, contains 173 sovereign or semi-sovereign countries from 1900-2014.
Variables from this project can be divided into two categories: 1) data collected from various other sources, such as the World Bank, Eurostat etc.; and 2) indicators coded by country experts. More than three thousand experts from all around the world contributed to this project by grading each of the variables per year and per country.8
The second largest contributor to the dataset is the databank of the World Bank. Even though the V-Dem researchers used the World Bank databank to construct their dataset, the latter contains data not used by the Dem researchers. For some variables the
V-8
All the experts live in the countries they code, and over 90% have a postgraduate degree. This makes the V-Dem dataset one of the most reliable datasets in this field. For more information, visit the V-Dem website: https://www.v-dem.net
DATA
Dem dataset has data only for certain countries, e.g. GDP per country, life expectancy, population size and secondary school enrollment ratio. The databank of the World Bank provides these missing numbers.9 The data on inflation is very limited in both the V-Dem dataset and the World Bank databank. Therefore, data on inflation has been collected from the website of the International Monetary Fund (IMF).10By combining the V-Dem, World Bank and IMF data, still a far from complete and unbalanced dataset results. The next section explains how these problems are resolved.
4.3 Constructing the dataset
Adding the data from the World Bank and the IMF to the original V-Dem dataset results in an unbalanced panel dataset for 173 countries and annual data ranging from 1900-2014, with more than 40% blank spaces. However, lacking data occur mainly because the data from V-Dem is incomplete from 1900-1960 and the World Bank data are only available from 1960 onwards. Even then, more than 20% of the data is missing, but by disregard-ing the years 1960-1985 and the years 2011-2014 the amount of missdisregard-ing data is reduced drastically. What is left is a dataset that consists of data from 1986-2010.
Following the aforementioned manipulations to the dataset, many data are missing for a large number of countries. This is due to various reasons. For example, for some countries the coding in the V-dem dataset started later than 1985 and therefore fewer data is available at the beginning of the dataset.11 Other countries ceased to exist in or before the time frame of this research and are therefore excluded from the dataset.12 In most comparable research, including this study, certain countries are excluded because those countries have been in a state of war for a long period.13 This study excludes countries for several reasons as explained above. Most of the excluded countries have characteristics which are fairly hard to control for when performing a panel data analysis. Therefore, it should be noted that the results of this study can only be applied to the countries in the dataset. Next, countries that have two or more missing data points for the measure of democracy or the GDP are excluded, unless the missing data is at the beginning or end of the time interval. Finally, countries with less than 300.000 inhabitants are ignored, as was frequently done in previous research.
Following the exclusion of several countries, a dataset consisting of 100 countries
re-9
For more information on the World Bank Data, see: http://data.worldbank.org/
10See: http://www.imf.org 11
Most of these countries are post Soviet states, who have data starting in 1989 or 1990, or countries that gained independence later than 1985, such as Slovenia or South Sudan.
12
For example, South Yemen united with the Yemen Arab Republic in 1990 and the Republic of Vietnam (or South Vietnam) was taken by force by North Vietnam in 1975.
DATA
mains. However, for a number of countries, data for some of the control variables in some years are still missing. This forced us to use linear interpolation between years. Although the existence of a linear trend in the data is a fairly strict assumption, the use of this technique is performed mainly because the exclusion of the countries with missing data would cripple the dataset excessively. In the end, only less than 0.2% of the datapoints are interpolated and when looking at the dataset in many cases a linear trend seems ap-proximately correct. The dataset consists of the 25 years from 1986-2010. Reducing the likelihood of the presence of autocorrelation in the errors results in smaller measurement errors and a better focus on the main patterns in the data. For this reason, the years 1987-2010 are divided into six intervals of four years each, e.g. 1987-1990, 1991-1994, etc.14 Each data-point in this dataset is the average value of the data from the four years it represents. For the growth variables in the dataset, the growth in period 1 (1987-1990) is the growth from 1986 to this period. The dummy variable for armed conflict in a period is set to 1 if one or more armed conflicts occurred in the time frame of the corresponding interval.
Some countries have the same value for the democracy variable throughout the years. These 11 countries, all of which are high-income countries according to the World Bank, always get a 10 on a 0-10 scale, which means they cannot be more democratic.15 The
World Bank classifies high income countries as countries where the Gross National Income (GNI) per capita in 2016 is $12,476 or more. In this dataset, 28 countries are classified as high-income according to these numbers. For ease of use, the remaining countries will be called lower-income countries.16 These lower-income countries have more variation in the level of democracy. Therefore, the analysis in this study will be two-fold. First, the complete dataset is analyzed. Second, only lower-income countries are analyzed to obtain more insightful results on countries with a more fluctuating level of democracy. The list of countries can be found in Appendix A.
4.4 What is a democracy?
Since democracy is the explanatory variable of major interest in this study, more details are given here about how this variable is constructed. The Polity IV and Freedom House measures are used to value the democracy in a country, but first will be explained what
14
Serial correlation is more likely to be present when using subsequent years for the analysis, because the assumption of E(tt−1) = 0 is expected to be violated.
15
These countries are: Australia, Austria, Canada, Denmark, Ireland, The Netherlands, New Zealand, Norway, Sweden, Switzerland and The United States.
16
This means that in this study’s lower-income countries are upper-middle, lower-middle and low income countries, using the World Bank’s terminology.
DATA
the general consensus is on this topic. The four rules in the seminal work by Przeworkski
et al. (2000) are frequently used in the political sciences to determine whether a country
is a democracy or a dictatorship. If at least one of the following four conditions hold, a country is considered to be a dictatorship:
1. The chief executive is not elected.
2. The legislature is not elected.
3. There is no more than one party.
4. (Applies only to the regimes that have passed the previous three rules.) The
incum-bents will have or already have held office continuously by virtue of elections for more than two terms or have held office without being elected for any duration of their cur-rent tenure in office, and until today or until the time when they were overthrown they had not lost an election.
This way, political scientists are able to create a binary measure of democracy and dicta-torship for each country for each year.
However, this uses a scale variable for the level of democracy, which is created by combining the widely used Polity IV and Freedom House measures. The Polity IV index is determined as described on its website:17 ”The Polity scheme consists of six component measures that record key qualities of executive recruitment, constraints on executive au-thority and political competition. It also records changes in the institutionalized qualities of governing authority.” On the other hand, the Freedom House index focuses more on
political freedom and not democracy, but this measure is widely used among researchers to determine the level of democracy in a country. The Freedom House index consists of two main categories: political rights and civil liberties.
The two aforementioned measures for the level of democracy are compared by H¨ogstr¨om (2013). He finds that both indices rate many countries’ levels of democracy differently, but neither should be preferred over the other. For this reason, this study utilizes the av-erage of both indices as the level of democracy, both variables are included in the V-Dem dataset.
17
RESULTS
5
Results
In this section the results are presented. First, the regressors are classified to be either endogenous, predetermined or exogenous and with this result the composition of the final model is described. Second, the results of the different estimation techniques on the full dataset and the subset are presented. Lastly, the long-term impact of the level of democracy on economic growth in the Arellano-Bond setting is examined.
5.1 Model on economic growth
Unreported results show that the use of second or higher order lags of economic growth and the control variables result in unsatisfactory p-values for the Hansen test and autocor-relation tests. Therefore, only the first lag of economic growth and the control variables are used, similar to the general model (3). To keep a convenient balance between the number of instruments and the number of observations, up until the third lag of the re-gressors are used as instruments and the instrument set may be collapsed. Now, model (3) with the aforementioned characteristics is estimated. The p-value of the overall Hansen test is 0.874 and the p-values for the first-order and second-order autocorrelation tests are 0.009 and 0.420, respectively. To obtain these results, all instrument sets of the variables are collapsed except for the instrument subsets of savings/GDP and government
expendi-ture/GDP, since collapsing the entire instrument set gives a p-value for the AR(1) test of
greater than 0.05. Furthermore, and also important, no unit root is present.
Following the strategy outlined in section 3.4, the classification of the regressors is obtained. During this classification process, the p-values of the three classification tests gave no reason to consider making adjustments to the model. This classification process is performed using AB1 and the resulting classifications are also used when performing AB2 estimation. The p-values of the two AR tests and the Hansen test are still satisfactory in this case. Table 1 presents the classifications of the variables in the dataset with all countries included. It can be observed that export/GDP and log life expectancy are the only two endogenous variables in this dataset, since the p-values of the Hansen test, which tests for predetermindness of the variables, is smaller than 0.30. This dataset consists of five predetermined variables, of which armed conflict and government expenditure/GDP have a Hansen test p-value of higher than 0.30 when testing whether those variables are exogenous. Since the borderline of 0.30 remains questionable, it is chosen to treat these variables as being predetermined instead of exogenous. At this point, the number of ex-planatory variables is high compared to the number of instruments, resulting in highly insignificant coefficients. Through trial and error, and as long no problems occur
concern-RESULTS
ing the specification tests, the most insignificant variables are excluded.
Table 1: Classification of the regressors, full dataset.
Variable Classification p-value stage 1 p-value stage 2
Export/GDP Endogenous 0.217 x
Log life expectancy Endogenous 0.188 x
Inflation Predetermined 0.622 0.276
Democracy Predetermined 0.865 0.113
Savings/GDP Predetermined 0.375 0.207
Armed conflict Predetermined∗ 0.457 0.391
Government expenditure/GDP Predetermined∗ 0.659 0.409
Mortality rate Exogenous 0.377 0.489
Education under 15 years Exogenous 0.739 0.497
Secondary school enrollment Exogenous 0.454 0.612
Corruption Exogenous 0.882 0.538
Population growth Exogenous 0.887 0.502
Note: p-value stage 1 is the resulting p-value of the incremental Hansen test which tests, in this case, whether an assumed endogenous regressor is actually predetermined. Similarly, p-value stage 2 tests whether a predetermined regressor should be treated as exogenous. The (*) following the clas-sifications of Armed conflict and Government expenditure/GDP means these variables are exogenous according to the Incremental Hansen test but will be treated as predetermined.
In Table 2 the results of the classification process of the variables in the dataset excluding high-income countries is displayed. This whole process is performed a second time because for most high-income countries the level of democracy barely fluctuates and thus the dynamics of this dataset are different than the dynamics in the full dataset. Lower-income countries have a higher varying level of democracy, presumably because lower-income countries are in many cases less stable in political organization. The same instruments as in the full dataset are used, which results in an AR(1)-test p-value of 0.075. To deal with this, additionally not collapsing the instrument subset of democracy gives satisfactory results for all specification test. In Table 2 it can be observed that the classification of several variables is different than in the full dataset classification process. For example,
savings/GDP was considered a predetermined variable in the full dataset, but treated as
endogenous in the subset. An example of a more extreme change of classification is the variable export/GDP, which is endogenous is the full dataset and considered exogenous in
RESULTS
the subset. In the next section these classifications are used to obtain results of estimating model (4).
Table 2: Classification of the regressors, high-income countries excluded.
Variable Classification p-value stage 1 p-value stage 2
Savings/GDP Endogenous 0.245 x
Log life expectancy Endogenous 0.273 x
Inflation Endogenous 0.182 x
Government expenditure/GDP Endogenous 0.176 x
Democracy Predetermined 0.998 0.091
Armed conflict Predetermined 0.451 0.291
Education under 15 years Predetermined 0.959 0.078
Secondary school enrollment Predetermined 0.942 0.073
Export/GDP Exogenous 0.860 0.432
Corruption Exogenous 0.952 0.981
Population growth Exogenous 0.773 0.387
Mortality rate Exogenous 0.579 0.587
Note: p-value stage 1 is the resulting p-value of the incremental Hansen test which tests, in this case, whether an assumed endogenous regressor is actually predetermined. Similarly, p-value stage 2 tests whether a predetermined regressor should be treated as exogenous.
5.2 Main results
The Blundell-Bond estimation procedure was applied to the data, but unreported results show that the p-values of the Hansen test for the BB1 and BB2 regressions reject the assumption that the instruments in difference form are uncorrelated with the fixed effects (p = 0.01). Hence, Blundell-Bond estimation is disregarded from this study. Table 3 presents the estimation results of AB1c, AB1f, AB2c and AB2f for the full dataset. These regressions estimate 27 coefficients, the number of observations is 374 and the number of countries is 100. The number of instrumental variables for the uncollapsed estimation and collapsed estimation techniques is 155 and 45, respectively. The p-values for the AR(1) test are all smaller than 0.05, ranging from 0.000 for AB1f and AB2f to 0.019 for AB2c. This implies first-order autocorrelation is present, as expected. Second-order autocorrelation seems not present in all cases, with AR(2) p-values ranging from 0.554 for AB2c to 0.871 for AB2f. The validity of instruments is never rejected, since the p-values of the Hansen test are 0.274 for AB1c and AB2c and 1.000 for AB1f and AB2f. A formal test for the
RESULTS
Table 3: Complete dataset results of Arellano-Bond robust one-step and Windmeijer cor-rected two-step estimation method.
AB1c AB1f AB2c AB2f
Variable Lag Estimate SE Estimate SE Estimate SE Estimate SE
Economic growth 1 0.253** 0.118 0.133** 0.063 0.220 0.138 0.137* 0.072 Population growth 0 18.800*** 7.319 13.016*** 4.783 20.374*** 9.903 12.871** 5.692 Inflation 0 -0.063*** 0.021 -0.041*** 0.01 -0.061*** 0.026 -0.046*** 0.016 Inflation 1 0.026*** 0.011 0.018*** 0.003 0.021*** 0.009 0.018*** 0.004 Log Life expectancy 0 -1.457 1.686 -2.062*** 0.699 -0.566 2.207 -1.702** 0.828 Log Life expectancy 1 -3.272 2.007 0.157 0.757 -3.590 2.491 -0.267 0.952 Armed conflict 0 -0.082 0.073 -0.002 0.050 -0.046 0.070 -0.009 0.045 Armed conflict 1 -3.272 2.007 -0.035 0.049 -0.150 0.084 -0.070 0.052 Education > 15 years 0 0.200** 0.094 0.211*** 0.066 0.242** 0.108 0.175** 0.074 Education > 15 years 1 -0.042 0.092 -0.080 0.064 -0.076 0.127 -0.021 0.064 Secondary schooling 0 -0.009** 0.004 -0.008*** 0.003 -0.009** 0.005 -0.008** 0.003 Secondary schooling 1 0.005 0.004 0.002 0.003 0.005 0.006 0.002 0.003 Corruption 0 -1.071** 0.466 -0.548* 0.251 -1.047** 0.655 -0.368 0.353 105∗Export/GDP 0 1.997 9.906 -6.270* 3.477 2.957 11.908 -6.252** 2.986 105∗Export/GDP 1 18.721*** 5.581 13.123*** 3.385 18.701*** 6.195 12.801*** 3.021 Democracy 0 -0.129** 0.060 -0.033 0.021 -0.128** 0.081 -0.042* 0.026 Democracy 1 0.005 0.015 0.028** 0.013 0.016 0.018 0.030** 0.013 102∗Savings/GDP 0 0.029** 0.011 0.012*** 0.004 0.031*** 0.010 0.013*** 0.004 102∗Savings/GDP 1 -0.012*** 0.004 -0.007** 0.003 -0.015*** 0.005 -0.008** 0.003 Govern. exp./GDP 0 0.001 0.020 -0.016* 0.009 0.014 0.028 -0.015 0.012 Govern. exp./GDP 1 -0.017 0.011 -0.017** 0.007 -0.024 0.014 -0.014** 0.007 Mortality rate 0 -0.004 0.005 -0.005* 0.003 -0.004 0.006 -0.005 0.003 Mortality rate 1 -0.006 0.006 -0.000 0.003 -0.005 0.007 -0.001 0.004 Time dummy period 3 -0.024 0.051 -0.005 0.032 -0.045 0.064 -0.030 0.097 Time dummy period 4 -0.263*** 0.086 -0.231*** 0.042 -0.283*** 0.093 -0.049 0.076 Time dummy period 5 0.017 0.133 0.098* 0.059 -0.020 0.136 -0.286*** 0.054 Time dummy period 6 -0.030 0.148 0.097 0.074 -0.091 0.163 0.028 0.047
Number of countries 100 100 100 100
Number of observations 374 374 374 374
Number of instruments 45 155 45 155
p-value AR(1) test 0.001 0.000 0.019 0.000
p-value AR(2) test 0.677 0.819 0.554 0.871
p-value Hansen test 0.274 1.000 0.274 1.000
(***), (**), and (*) indicate statistical significance at the 1%, 5%, and 10% level, respectively. AB1 is Arellano-Bond one-step estimation. AB2 is Arellano-Bond two-step estimation.
RESULTS
presence of a unit root seems unnecessary since one is not included in the 95% confidence intervals of the coefficients of lagged economic growth.
Comparing the results between the robust 1-step and Windmeijer corrected 2-step es-timation techniques, it can be observed that the standard errors of the 2-step estimator are always larger than the standard errors of the 1-step estimator. This is confirmed by Windmeijer (2005), who states that the standard error of the 2-step estimator is similar to or larger than the standard error of the 1-step estimator. As mentioned before, AB1 and AB2 perform similarly and although collapsing the instrument set generally reduces the bias, it could imply power loss because efficiency suffers. Something else that can be observed from Table 3 is the difference in p-values for lagged economic growth, ranging from 0.032 for AB1c to 0.111 for AB2c.
Time dummies are included in each regression, but only in a few cases the coefficients are significant. The dummy for the fourth time period (1999-2002) is negative and highly significant in three out of the four regressions. The high value for the coefficient for
population growth seems surprising, but the value of population growth is on average 15
times larger than economic growth. The results show that, for every regression,
popula-tion growth exerts a positive, significant effect. The reason behind this is that the more
people live in a country, the more money can be made and thus GDP growth is higher. The contemporaneous and lagged estimation outcomes of inflation offset each other, being significant negative and significant positive, respectively. An increasing inflation results in more investments which takes some time, one period according to the estimation out-comes, to result in economic growth. The reason for the negative sign for the effect of current inflation is most likely because an increasing inflation reduces the value of money. Contradicting previous studies, log life expectancy has a negative sign in most regressions. The reason for this could be that when people live longer, they cost more money for the society because they get more pension and more health care cost have to be made. It should be noted, however, that most coefficients are not or barely significant. Both lags of armed conflict show, in most cases, the expected negative sign but is insignificant in all regressions. The contemporaneous effect of education >15 years is significant and pos-itive at the 5% level in all regressions. This is as expected, since an increased educated workforce implies a better performing economy. The contemporaneous effect of secondary
schooling is significant and negative, whereas a positive sign is expected. This could be due
to small predictive power of the variables or measurement errors in the data. The influence of corruption on economic growth is (mostly) significant and negative, so an increase in the level of corruption reduces economic growth in a country. Similar to the findings of
RESULTS
Alcala and Ciccone (2004), the first lag of export/GDP exert a significant, positive effect, implying that exporting more at this point leads to an increasing economic growth in the next period. The current and lagged effect of savings/GDP offset each other, being significant positive and significant negative, respectively. The positive effect of savings in the current period is most likely because of the boost in economic development. The negative lagged effect is possibly caused because money that is saved can not be invested and therefore opportunity costs might be made. Government expenditure/GDP mostly show the expected, negative sign, although hardly significant in the regressions. To con-clude the interpretation of the control variables, the coefficients for mortality rate show the expected, negative sign. A higher mortality rate is mainly a characteristic of poor, undeveloped countries which frequently have many years of negative economic growth.
The explanatory variable of major interest in this study is democracy. For AB1c and AB2c the contemporaneous effect is significant at the 5% level and negative. This is in accordance with the idea that an increasing level of democracy slows down fast implemen-tation of unpopular, but growth increasing, choices. The lagged effect of democracy is significant and positive for AB1f and AB2f and not significant but also positive for AB1c and AB2c. The reason for this could be that an authoritarian regime governs in a way such that a positive current economic growth is immediately obtained, but the longer term effect of their decisions are negative. This long-term effect is discussed more thoroughly in the next section.
In Table 4 the results of the same estimation techniques on the subset of only lower-income countries are presented. Because the instrument set is somewhat altered com-pared to the full dataset estimation, the number of instruments is smaller than before. The smaller number of countries, 72, reduces the number of observations which makes it harder to find an acceptable, trustworthy, specification. However, both the autocorrela-tion tests and the Hansen test give nothing to worry about. The significance of lagged
economic growth runs from significance at the 5% level for AB1c to insignificant for AB1f
and AB2f. Again, no formal test for the presence of a unit root is required since the 95% confidence intervals of lagged economic growth are not even close to 1.
The sign of nearly all coefficients of the control variables in the estimation of this subset are similar to the signs of the coefficients in Table 3. This indicates that there are no large differences between the full dataset estimation and the estimation of this subset. However, there are somewhat fewer significant outcomes in the subset, which could be due to the decreased number of observations since this leads to larger standard errors. Focusing on the explanatory variable of interest, democracy, tells us that the signs are again negative
