Inequality and Democracy Christian Houle∗

Abstract Recent theoretical research has found that inequality strongly affects both the likelihood that non-democracies transit to democracy and the likelihood of survival in already established democracies. Empirically, however, these claims either rest on a small number of case studies or samples subject to severe selection bias. In this paper, I present a model showing that inequality does not systematically affect the probability that non-democracies transit toward democracy, but that, once established, equal democracies are unlikely to collapse. Using a Markov transition model and a dataset covering about 2400 country-years between 1950 and 2001, I indeed find that – contrary to what the previous theoretical research has predicted – inequality has no systematic effect on democratization. Inequality does, however, destabilize already established democracies.

∗

Thanks to Kevin Clarke, Hein Goemans, Gretchen Helmke, Stuart Jordan, Michael Peress, G. Bingham Powell, Curtis Signorino, Randall Stone and especially, Alexandre Debs and Mark Kayser for helpful comments. I am also grateful to Arjun Jayadev for sharing his dataset on labor shares. I stake sole claim to all errors.

1

1 Introduction Recent theoretical research has found that inequality strongly affects both the likelihood that non-democracies transit to democracy (democratization) and the likelihood of survival in already established democracies (consolidation). Two theories are especially influential. Boix (2003) argues that inequality inhibits both democratization and consolidation. Acemoglu and Robinson (2006), for their part, instead suggest an inverted U-shaped relationship between inequality and democratization, thus predicting that only countries with intermediate levels of inequality democratize. They also argue that unequal democracies are less likely to consolidate. However, the relationships between inequality and democratization, on the one hand, and inequality and consolidation, on the other, have yet to be tested adequately. Acemoglu and Robinson (2006) do not perform statistical analysis, but only present four case studies: Singapore, Britain, Argentina and South Africa.1 Boix (2003) does test his model. But his empirical analysis is largely flawed, notably because he uses the inequality dataset of Deininger and Squire (1996), which suffers from both non-comparability between observations and non-representative coverage.2 This paper attempts to bridge the gap between the empirical and theoretical literatures by proposing and testing a new model. I argue that inequality does not systematically affect the probability that nondemocracies transit toward democracy, but that, once established, equal democ1

They however cite two papers that find preliminary support for their model. The first is Epstein et al. (2004). However, these authors use infant mortality to measure inequality. This is a poor measure because it is likely to capture the level of development instead of inequality. For instance, Nepal – a very poor but equal country – has an infant mortality rate several times larger than the one of Brazil, which is among the most unequal countries. The second is Papaioannou and Siourounis (2004). But, in a subsequent version of their working paper they find that their previous results are not robust to the inclusion of new observations. 2 These observations are not comparable because they are drawn from different sources. The dataset has observations based on expenditure and income, on net and gross income and on household and per capita surveys. This is likely to bias the results because, for instance, inequality is systematically lower with net than gross income. Moreover, it only contains about 600 observations for the full post-World War II period. The coverage is non-representative, because observations are overwhelmingly drawn from rich democracies.

1

racies are unlikely to collapse. The major challenge in unequal countries is not to establish democracy, but to sustain it once it is created. This is particularly clear in the history of Latin America, the most unequal region. While most countries transited to both democracy and dictatorship several times, only Costa Rica and Uruguay maintained stable democracies throughout nearly the entire post-World War II period.3 Interestingly, these are the two most equal Latin American countries.4 Other equal democracies, like India, Jamaica and Mauritius, have also been remarkably stable, despite low levels of development. Many highly unequal countries, such as Nigeria, Peru and Turkey, have also been able to establish democratic regimes, suggesting that – contrary to what both Boix (2003) and Acemoglu and Robinson (2006) argue – highly unequal countries can become democratic. However, these regimes have been extremely fragile. For instance, between 1950 and 2001, Turkey has experienced 3 democratic breakdowns in only 16 years of democracy. My model consists of two games – one about democratization and another about consolidation – that build on the ones of Acemoglu and Robinson (2001, 2006) and Boix (2003). In the democratization game, a country democratizes when it is more costly for the elites to maintain the authoritarian regime than to establish a democracy. The population affects the choice of the elites by selecting a contestation level. To maintain the authoritarian regime, the elites have to counter contestation through repression. The idea is that to be maintained any authoritarian regime must rely on some level of repression. I find that while inequality makes democracy more costly to the elites – by increasing redistribution – it also 3

However, Uruguay was a dictatorship between 1973 and 1983. Other countries, such as Columbia and Venezuela have also maintained fairly stable democracies. But only Costa Rica and Uruguay are widely considered as consolidated democracies in Latin America (e.g., Diamond 1999). Both Venezuela and Colombia have moreover recently experienced democratic instability. Acemoglu and Robinson (2006) also argue that low inequality is largely responsible for the stability of the Colombian democracy. 4 The link between equality and democratic stability in both countries has been noticed by many authors (e.g., Karl 2000).

2

increases the cost of maintaining a dictatorship – because it increases contestation. The net effect is thus generally ambiguous. So, in my model, contrary to the one of Boix (2003), the cost of repression is not independent of inequality. My model is also more general than the one of Acemoglu and Robinson (2006).5 Indeed, dictatorships are always costly to maintain and can be deposed by the population. The consolidation game is similar to the ones of previous authors, where the elites stage coups when redistribution is sufficiently large. I find that coups are more likely when inequality – and thus redistribution – is greater. The difference between the democratization and consolidation games is driven by the assumption that while transitions to democracy are the results of compromises between the elites and the population, transitions away from democracy only depend on the interests of the elites.6 Whereas transitions toward democracy are consensual, those toward dictatorship are not. Using a Markov transition model and a dataset covering about 2400 countryyears between 1950 and 2001, I indeed find that – contrary to what the previous theoretical research has predicted – inequality has no systematic effect on democratization. My empirical approach differs from the previous works because it combines three aspects. First, I test for the possibility of a non-linear relationship to capture the argument of Acemoglu and Robinson (2006). Second, instead of looking at transitions only between dictatorships and democracies, I include a third category: partial democracies. This captures the distinction between high and low 5

In Acemoglu and Robinson (2006), the cost of repression also depends on inequality but for different reasons. Whereas in my model this is driven by the assumption that there is more contestation in unequal dictatorships, in Acemoglu and Robinson (2006) it is driven by the assumption that the cost of repression is proportional to the income of the elites. 6 Acemoglu and Robinson (2006) also argue that transitions from democracy to dictatorship proceed differently than those from dictatorship to democracy. Indeed, ”There is one difference between the way we are modeling the transition from nondemocracy to democracy and the transition from democracy to nondemocracy: in the first case, the citizen had the option to undertake a revolution, and the elites created a democracy to prevent it. Here, the elites actually use their political power to mount a coup and change the system. [...] in most instances, democracy resulted from the elites democratizing, whereas the move from democracy to dictatorship is almost never consensual” (p.225).

3

quality democracies. The three regime types are constructed from the Polity scores. Lastly, I use a better measure of inequality, the labor share of the total GDP. This dataset is collected by the United Nations Industrial Development Organization (UNIDO). Because my dataset is incomplete, I impute the missing labor share observations using other measures of inequality and the distribution of education attaintment as instruments.

2 Theory 2.1 Democratization Game In this section, I show that inequality does not systematically affect the probability of democratization. This is because inequality makes it more beneficial for both the population and the elites to control the political institutions. In my model, a country democratizes when it is more costly for the elites to maintain the authoritarian regime than to establish a democracy. The population can make it expensive for the elites to keep the dictatorial regime by generating instability. Therefore, democratization is, in the end, conceded by the ruling elites, but initiated by the masses through social unrest. On the one hand, as argued by Acemoglu and Robinson (2006), the elites have no incentives to concede democracy unless the population demands it. The elites establish democracy only when maintaining a dictatorship is too costly. The importance of the population during democratization has, for example, been highlighted by Collier (1999), Wood (2000) and Lee (2002). On the other hand, as argued by O’Donnell and Schmitter (1986), transitions toward democracy are almost always ultimately conceded by the elites. Direct seizures of power by the population are rare and they usually result in the establishment of socialist dictatorships, not democracies. This can be illustrated by the case of the South African transition of 1994. Social unrest pushed the elites to democratize, 4

but the protestors did not directly seize power. Democratization is thus a compromise between the elites and the population. Inequality affects democratization through the population by determining its demand for democracy. In unequal dictatorships, the population has more to gain from democratization, because redistribution would be relatively large. It thus generates more unrest. The empirical literature indeed finds a positive association between inequality and political instability and civil unrest (e.g., Venieris and Gupta 1986; Alesina and Perotti 1996; Figueroa and Barron 2005). Maintaining unequal dictatorships is thus more costly for the elites. These costs take many forms. First, there is a direct cost needed to use repression, for instance large police forces are costly. Wood (2000) shows that, before the fall of the Apartheid, the South African government needed ever expending public spendings to defend the regime. Social unrest also has a human cost, for both the elites and the population, as illustrated by the 12 years civil war of El Salvador. Repression also has an economic cost, for example because it hurts investors confidence and often leads to capital flight. According to Wood (2000), in South Africa ”elite compromise was impelled by a crisis of confidence among investors in the institutions governing the economy of South Africa, and the resulting sustained decline in investment. Eroding confidence was largely the result of sustained mobilization by workers and others in the anti-Apartheid movement” (p.144). The unwillingness of the elites of El Salvador to continue paying the economic costs of the civil war also largely explains the democratic compromise they reached with the insurgents in 1992 (Wood 2000). In fact, empirical studies find that repression and social unrest are associated with less growth and less FDI inflows (e.g., Butkiewicz and Yanikkaya 2005; Campos and Nugent 2003). In addition, social unrest and massive repression is often linked with international sanctions or reduced foreign aid (e.g., Meernik, Krueger and Poe 1998). For instance, the threat of ceasing foreign aid

5

flows appears to have played a key role in African transitions in the early 1990s (Dunning 2004). So, unequal dictatorships are relatively costly to maintain. However, the cost to the elites of establishing a democracy is also larger in unequal countries (Meltzer and Richard 1981). Indeed, by the median voter theorem, unequal democracies redistribute more than equal democracies, because the difference between the mean income and the income of the median voter is larger. So, both the cost of maintaining a dictatorship and the cost of adopting a democracy increase with inequality. Since inequality has the same effect on both costs, its overall effect on the cost-benefit calculation is unclear. 2.1.1 Set Up There are two groups of agents, the poor and the rich. Each agent maximizes his utility, which is simply equal to his consumption. There is a continuum 1 of agents. The income of the rich is y r and the one of the poor y p , with y p < y r . These are defined as yr =

θy (1 − λ)

(1)

yp =

(1 − θ)y λ

(2)

where λ is the proportion of agents that are poor. I assume that the poor represent the majority of the population, i.e. λ > 12 . The average per capita income is given by y and inequality by θ ∈ (1 − λ, 1). Thus, θ ' 1 − λ denotes complete equality and θ ' 1 complete inequality. As θ rises inequality increases. There is a deadweight cost of taxation C(τ )y, such that for any tax rate τ , tax revenues are a fraction τ − C(τ ) of output. I assume C(τ ) is twice continuously differentiable and that C(0) = 0, C 0 (0) = 0, C 0 (τ ) > 0 for all τ > 0 and C 00 (τ ) ≥ 0 for all τ > 0. So, the deadweight cost is increasing and convex in the tax rate. The

6

preferred tax rate of agent i is determined as τ i = argmaxτ ∈[0,1] [(1 − τ )y i + (τ − C(τ ))y]

(3)

The elites are in power at the start of the game, but both groups jokey for power in subsequent periods. First, the elites have the option to democratize. If they maintain the dictatorship, they enter in conflict with the poor. The poor select a level of contestation. Higher contestation increases the probability of success. However, contestation is costly to the poor. Similarly, the rich set a level of repression. The elites are more likely to win the contest when they repress heavily, but repression is also costly. If the elites win the regime remains authoritarian, if they lose the poor take power and install a socialist regime. There are three states R ∈ {A, D, S}, where R = A indicates authoritarian, R = D democracy and R = S socialist. Transitions between states happen as follow. In the beginning of the game R = A. In the first stage the elites set d ∈ {0, 1}. If d = 1 the elites create a democracy and the state becomes R = D. Otherwise, the dictatorship is maintained. Then, if R = A, the population selects a level of contestation ζ ∈ [0, 1] and the elites a level of repression κ ∈ [0, 1]. With probability p(ζ, κ) the poor win the contest and the state becomes R = S. The probability that the regime gets overthrown is increasing in contestation and decreasing in repression, i.e.

∂p(ζ,κ) ∂ζ

≥ 0 and

∂p(ζ,κ) ∂κ

≤ 0.

The sequences of the game are as follow. 1. The elites select d ∈ {0, 1}, where d = 1 indicates that they democratize and d = 0 that they do not. 2. If they maintain the authoritarian regime, the poor set a level of contestation ζ ∈ [0, 1] and the elites a level of repression κ ∈ [0, 1]. 3. If the dictatorship is maintained, the regime is overthrown and replaced by a 7

socialist regime with probability p(ζ, κ). 4. The tax rate is selected. If the country is authoritarian, the rich set the tax rate. If it is democratic, the median voter – who is poor – selects the tax rate. In a socialist regime the elites are eliminated and their assets redistributed equally among the poor. The utility of the poor and the elites under each regime are given as follow

U i (τ, R) =

    (1 − τ )y i + (τ − C(τ ))y     y

if R ∈ {A, D}, ∀i if R = S, i = p

λ       0

(4)

if R = S, i = r

In democracies and dictatorships agents receive there after-tax income plus their share of the total redistribution. In a socialist regime, each poor receives an equal share of the total wealth. The utility of the elites in a socialist regime is normalized at zero. The values of the game for each player are V p = dU p (τ, D) + (1 − d)[−ζ + p(ζ, κ)U p (τ, S) + (1 − p(ζ, κ))U p (τ, A)]

(5)

V r = dU r (τ, D) + (1 − d)[−κ + (1 − p(ζ, κ))U r (τ, A)]

(6)

where −ζ and −κ respectively indicate the costs of contestation and repression. 2.1.2 Analysis I solve the model using backward induction. First, I find the optimal tax rate in each regime. In an authoritarian regime, the elites set the tax rate. Since, y r ≥ y, they always select τ r = 0. In a democracy, the median voter selects the tax rate. Since λ > 12 , the median voter is always poor. The tax rate under democracy is the

8

one that solves equation (3) with y i = y p . θ+λ−1 λ

C 0 (τ p ) =

(7)

Thus, under a democracy, the tax rate increases with inequality. Next, I need to find the optimal levels of contestation and repression, when the elites decide not to democratize. These are given by ζ ∗ = argmaxζ∈[0,1] [−ζ + p(ζ, κ)U p (τ, S) + (1 − p(ζ, κ))U p (τ, A)]

(8)

κ∗ = argmaxκ∈[0,1] [−κ + (1 − p(ζ, κ))U r (τ, A)]

(9)

Two necessary and sufficient conditions for the second order conditions to be satisfied are that

∂ 2 p(ζ,κ) ∂ζ 2

≤ 0 and

∂ 2 p(ζ,κ) ∂κ2

≥ 0. A key part of my argument is that

when inequality increases, both the poor and the rich have stronger incentives to take power. I thus need to find the effect of inequality on the optimal levels of contestation and repression, i.e.

dζ ∗ dθ

and

dκ∗ . dθ

The direct effects of inequality on

contestation and repression are given by ∂p(ζ,κ)

∂ζ ∗ ∂ζ = − ∂ 2 p(ζ,κ) ≥ 0 ∂θ θ( ) 2

(10)

∂ζ

∂p(ζ,κ)

∂κ∗ = − ∂ 2∂κ ≥0 ∂θ θ( p(ζ,κ) ) 2

(11)

∂κ

Thus, inequality always has a positive direct effect on both contestation and repression. However, inequality also has an indirect effect. Indeed, the optimal level of contestation is also a function of the optimal level of repression and vice versa. So, inequality affects the level of contestation (repression) by influencing the level of repression (contestation). The effect of repression on contestation and

9

of contestation on repression are ∂ 2 p(ζ,κ)

∂ζ ∗ ∂ζ∂κ = − ∂ 2 p(ζ,κ) ∂κ 2

(12)

∂ζ

∂ 2 p(ζ,κ)

∂κ∗ ∂ζ∂κ = − ∂ 2 p(ζ,κ) ∂ζ 2

(13)

∂κ

The sign of both expressions is ambiguous. If and

∂κ∗ ∂ζ

≤ 0, but if it is negative then

∂ζ ∗ ∂κ

≤ 0 and

∂ 2 p(ζ,κ) ∂ζ∂κ ∂κ∗ ∂ζ

is positive then

∂ζ ∗ ∂κ

≥ 0

≥ 0. This means that when,

for instance, a rise by the same amount in the levels of contestation and repression increases the probability that the poor win the contest, i.e.

∂ 2 p(ζ,κ) ∂ζ∂κ

> 0, the poor

always increase contestation when repression rises. The elites, on the other hand, decrease repression whenever the poor increase contestation. When contestation and repression completely offset each other – i.e.

∂ 2 p(ζ,κ) ∂ζ∂κ

= 0 – there is no indirect

effect. This suggests that the balance of power between the poor and the rich, and particularly their ability to influence p(ζ, κ), is a key factor in determining their behavior. Also notice that the capacity of each group to affect p(ζ, κ) is likely to vary both across countries and within countries over time. I now turn to the total effect of inequality on contestation and repression, which combines the direct and indirect effects. 2

2

p(ζ,κ) ∂p(ζ,κ) ∂p(ζ,κ) ∂ p(ζ,κ) − ∂ ∂κ∂ζ dζ ∗ ∂ζ ∂κ2 ∂κ = 2 p(ζ,κ) 2 p(ζ,κ) ∂ 2 p(ζ,κ) ∂ ∂ 2 dθ θ[( ∂ζ∂κ ) − ∂ζ 2 ] ∂κ2 2

2

∂ p(ζ,κ) ∂p(ζ,κ) p(ζ,κ) ∂p(ζ,κ) − ∂ ∂κ∂ζ dκ∗ ∂ζ 2 ∂κ ∂ζ = ∂ 2 p(ζ,κ) ∂ 2 p(ζ,κ) ∂ 2 p(ζ,κ) 2 dθ ] θ[( ∂ζ∂κ ) − ∂ζ 2 ∂κ2

Again, the sign of

∂ 2 p(ζ,κ) ∂ζ∂κ

(14)

(15)

is central. When it is positive inequality always in-

creases contestation but has an ambiguous effect on repression. If it is negative inequality increases repression but now has an ambiguous effect on contestation. 10

Lastly, if the cross-partial derivative is zero, then inequality increases both contestation and repression, since there is no indirect effect. This result is intuitive. If an increase in contestation and repression by the same amount augments the probability that the poor win – i.e.

∂ 2 p(ζ,κ) ∂ζ∂κ

≥ 0 – the poor always increase contesta-

tion when inequality increases. The behavior of the elites, however, is ambiguous because the direct and indirect effects of inequality go in opposite directions. I now turn to the central step of the model, i.e. the decision of the elites about whether or not to democratize. Let us denote Uar = −κ + (1 − p(ζ, κ))U r (τ, A). The elites democratize whenever U r (τ, D) ≥ Uar . The effect of inequality on the probability of democratization depends on its effect on each side of this equation. If

dU r (τ,D) dθ

>

dUar dθ

inequality promotes democratization, whereas if

dU r (τ,D) dθ


− ∂U for all values of θ. The second part of the second term of ∂ζ ∗ dθ

equation (17) is given by equation (14), while the first part is given below. ∂Uar ∂p(ζ, κ) θy =− ∗ ∂ζ ∂ζ (1 − λ)

(19)

Putting equations (14), (16), (18) and (19) together and simplifying we get that

12

inequality inhibits democratization if and only if ∂p(ζ,κ) ∂ 2 p(ζ,κ)

∂ 2 p(ζ,κ) ∂p(ζ,κ)

− ∂κ∂ζ ∂κ dτ d ∂p(ζ, κ) ∂ζ ∂κ2 [1−λ−(1−λ)C 0 (τ d )−θ]−τ d < −p(ζ, κ)− ( ∂ 2 p(ζ,κ) 2 p(ζ,κ) ∂ 2 p(ζ,κ) ) dθ ∂ζ ( ∂ζ∂κ )2 − ∂ ∂ζ 2 ∂κ2 (20) Depending on the specific functional forms of C(τ ) and p(ζ, κ), and the sign of its cross-partial derivative – all of which diverge across countries and within countries over time – we cannot find a single prediction. In order to show that we can have different predictions, I give an example leading to different relationships between inequality and the probability of democratization. In the case below, I √ √ a assume that C(τ ) = τa with a ≥ 1, and that p(ζ, κ) = m ζ(1 − κ) with 0 ≤ m ≤ 1. Both of these functions meet all the requirements derived above. As Boix (2003), I further assume that y = 1. With these specific functional forms equation (20) becomes

(

p √ 1 1 θ + λ − 1 2−a θ+λ−1 θ + λ − 1 a−1 )( ) a−1 [1−λ−(1−λ)( )−θ]−( ) < −2m ζ(1− κ)2 (a − 1)λ λ λ λ (21) Then, using equations (8) and (9), I find ζ ∗ and κ∗ for all θ. These are given

in figures 1 and 2, for different values of m, under the assumption that λ = 0.95. Here, m can be interpreted as a measure of the balance of power between the poor and the elites. The higher is m, the more powerful the poor are relatively to the elites. Because the cross-partial derivative of this p(ζ, κ) function is negative, the optimal level of repression should always increase with inequality, whereas the effect of inequality on the optimal level of contestation should be ambiguous. In figure 1, we indeed see that, depending on m, inequality has different effects on contestation. When the poor are weak comparatively to the elites (m = 0.1 or m = 0.5), inequality always increases contestation. However, when they are relatively pow-

13

0.020 0.010 0.000

0.005

Contestation

0.015

m=0.1 m=0.5 m=1

0.2

0.4

0.6

0.8

1.0

Inequality

Figure 1: Inequality and the Optimal Level of Contestation. Large values of m indicate that the poor are relatively powerful. erful (m = 1), inequality first increases contestation, but when θ attains about 0.45, the relationship reverses and inequality decreases contestation. This is because at that point, as illustrated in figure 2, repression becomes extremely high. The indirect (negative) effect of inequality then outweighs its direct (positive) effect. One interesting finding is that at high levels of inequality (here, at about θ = 0.6), contestation is higher when the poor are moderately powerful (m = 0.5) than when they are very powerful (m = 1). As expected, for all values of m, inequality always increases the optimal level of repressions, as shown in figure 2. Figure 3 plots both sides of equation (21) for different values of m, at a = 1.5. When the LHS is larger than the RHS, inequality promotes democratization. When instead the RHS is larger, inequality has a negative effect on democratization. Even with these specific functional forms, we cannot find a single relationship between inequality and the probability of democratization. First, for small values of m

14

0.8 0.6 0.4 0.0

0.2

Repression

m=0.1 m=0.5 m=1

0.2

0.4

0.6

0.8

1.0

Inequality

Figure 2: Inequality and the Optimal Level of Repression. Large values of m indicate that the poor are relatively powerful. (m = 0.1), the RHS of equations (21) is always greater than its LHS, and thus we have the relationship of Boix (2003) that inequality always decreases the probability of democratization. If the poor are unlikely to take power the first effect, in equation (17), outweighs the second. An increase in inequality increases the cost of democratization more rapidly than the one of maintaining the authoritarian regime, because the poor are unable to effectively contest – i.e. m is small. At middle and high values of m (m = 0.5 and m = 1), the relationship between inequality and democratization follows an inverted U-shaped curve, as argued by Acemoglu and Robinson (2006). For values of θ lower than about 0.2 at m = 0.5 and 0.3 at m = 1, the LHS is greater than the RHS and therefore, inequality promotes democratization. But, when θ passes these thresholds, the RHS is now larger than the LHS, such that inequality inhibits democratization. Furthermore, when a increases – i.e. the cost of taxation diminishes – the LHS

15

0.0 −0.1 −0.2 −0.3 −0.4 −0.5

LHS RHS (m=0.1) RHS (m=0.5) RHS (m=1) 0.2

0.4

0.6

0.8

1.0

Inequality

Figure 3: Inequality and Democratization. Inequality promotes democratization whenever the LHS of equation (21) is larger than its RHS, and inhibits it whenever its RHS is larger than its LHS. Large values of m indicate that the poor are relatively powerful. curve moves toward the left, such that when it is sufficiently high (about a = 2.8) inequality always has a negative effect on the probability of democratization, for all values of m. When a decreases the turning point on the inverted U-shaped curve increases. For instance, when a = 1.2 and m = 0.5, inequality has a positive effect on democratization until θ attains about 0.45. Thus, as taxation becomes less costly – for instance because the economy relies heavily on natural resources (instead of physical capital) – inequality is more likely to have a negative effect on democratization. Figure 4 illustrates the relationship between inequality and the probability that the poor win the contest and establish a socialist regime. As in the case of democratization, the shape of the relationship depends crucially on m. When the poor are very weak (m = 0.1), inequality always increases the probability of establishing a socialist regime, but is low. When the population is at least moderately strong 16

0.10 0.06 0.04 0.00

0.02

Probability

0.08

m=0.1 m=0.5 m=1

0.2

0.4

0.6

0.8

1.0

Inequality

Figure 4: Inequality and the Probability of Transition to Socialism. Large values of m indicate that the poor are relatively powerful. (m = 0.5 and m = 1), there is an inverted U-shaped relationship between inequality and the probability of socialism. At very high levels of inequality, repression is very high, such that authoritarian regimes are unlikely to be overthrown by the population. Interestingly, at high levels of inequality (above 0.6) dictatorships are more likely to collapse when the poor are moderately powerful (m = 0.5) than when they are very strong (m = 1). In this section, I have shown that some key factors – that vary both across countries and within countries over time – shape the relationship between inequality and democratization. The results of both Boix (2003) and Acemoglu and Robinson (2006) turn out to be just special cases. Here are the most important determinants: • The functional form of p(ζ, κ) and the sign of its cross-partial derivative. • The balance of power between the poor and the elites. As illustrated in figure 3, when the poor are relatively powerful (m is high) inequality is less likely 17

to always decrease the probability of democratization. • The functional form of C(τ ). If the deadweight cost of taxation increases (a decreases) there is a smaller set of m such that inequality always inhibits democratization. 2.1.3 Discussion As shown in the previous section, both the burden of democracy and the cost of maintaining a dictatorship increase with inequality. The specific effect of inequality is mainly determined by the deadweight cost of taxation and balance of force between the elites and the citizens. Authoritarian regimes that can effectively repress their population, for instance because of the quality of their police force, will respond differently to inequality than those that are ineffective. Moreover, dictatorships are not all equally dependent on the world economy. In countries particularly vulnerable to capital flight, it may be extremely costly to repress since social conflicts create uncertainty and hurt the confidence of investors. This seems to have played a key role in the democratization of South Africa, where repression and social unrest led to capital flight, international sanctions and a sharp decrease in the real value of the rand (Wood 2000). Similarly, inequality has different effects on democratization in countries where the economy is based on natural resources than in those where it is not. Contrary to what Acemoglu and Robinson (2006) argue, keeping an authoritarian regime in equal countries may be more costly than democratizing, because there is little redistribution in equal democracies. This can be illustrated by the cases of the democratic transitions of Nepal, Bangladesh, Mongolia, Indonesia, South Korea and Taiwan at the end of the 1980s and 1990s, all of which are among the most equal countries. These transitions were all responses to popular protests (Lee 2002). Moreover, in all cases the elites gave up power relatively easily, with18

out major confrontations. According to Lee (2002), ”As political turmoil raged out of control, the regimes in power chose finally not to suppress the opposition movements” (p.835). These examples show both that pro-democracy protests can occur even in the most equal countries and that once they are faced with such opposition the leaders tend to democratize without major opposition, because democracy has only a trivial cost for the elites. There are also many examples of transitions toward democracy in unequal countries, where dictatorial regimes were no longer able to maintain themselves via repression. For instance, in Bolivia the authoritarian regime first stayed in power through repression, but at one point it became too costly and the regime fell in 1982 (Collier 1999). Even in South Africa – used as a case study by Acemoglu and Robinson (2006) – democratization was caused by significant social unrest which eventually became costly to repress (Wood 2000). So, there is no reason to believe that in unequal countries the cost of redistribution is always larger than the cost of repression, since both increase with inequality. Thus, the effect of inequality on democratization is ambiguous.

2.2 Consolidation Game Whereas inequality does not systematically affect transitions from dictatorship to democracy, it increases the likelihood of transitions from democracy to dictatorship. This is because democratic breakdowns follow a different pattern than transitions toward democracy. Because of its small size and important resources, and often the support of the military, the elites are able to directly seize power, without the agreement of the population. In such transitions, the incentives of the population to keep the democracy have little impact, because there is no need for consensus. In unequal societies, even if the population has a lot to gain from redistribution, it cannot effectively prevent the elites from violently overthrowing the 19

regime. Here, only the cost of redistribution is relevant. Therefore, when inequality, and thus the redistribution burden, is high, democratic regimes are more likely to be deposed through coups. 2.2.1 Set Up Again there are two groups of agents: the poor and the rich. Their income and the tax rate are defined as above. However, the consolidation game is substantially different than the democratization game. This is because while democratization are consensual processes, democratic breakdowns are not. In the democratization game it is the elites that ultimately decide whether or not to democratize in order to prevent a revolution. But in the consolidation game it is not those that initially hold power that decide the faith of the regime, but the elites. In that sense, transitions from democracy to dictatorship resemble more those from dictatorship to socialism than those from dictatorship to democracy. The sequences of the game are as follow 1. The elites select c ∈ {0, 1}, where c = 0 means that they do not stage a coup and c = 1 that they do. 2. The tax rate is selected. If the country is authoritarian, the rich select it. If it is democratic, the median voter – who is poor – sets the tax rate. The utility functions are given as in equation (4). The value of the game for the rich is V r = c[−ϕ + U r (τ, A)] + (1 − c)U r (τ, D)

(22)

where ϕ ∈ [0, 1] is the cost of staging a coup.7 7 One could argue that coups are not always successful and that we should thus include a term capturing the probability of success. Moreover, whether or not a coup is successful may depend on the effort of the population to maintain the regime, such that in unequal democracies coups may be less successful. However, even if this is plausible, it is unlikely that the population can affect

20

2.2.2 Analysis As in the democratization game, I assume that y = 1. The elites stage a coup whenever −ϕ + U r (τ, A) is greater or equal to U r (τ, D). This condition is satisfied when8 ϕ≤

τ θ + (1 − λ)(τ − C(τ )) 1−λ

(23)

We see that the RHS of equation (23) is increasing in θ, such that coups become more likely as inequality increases.9 So, as most previous authors have argued, democracies are more likely to collapse in unequal countries. 2.2.3 Discussion This argument can be illustrated by the history of most Latin American countries – the most unequal region. These have experienced many coups, often directly aimed at preventing redistribution, like the Chilean one of 1973 (Acemoglu and Robinson 2001). Uruguay and especially Costa Rica have been the only fairly stable Latin American democracies. According to Karl (2000), these regimes have endured largely because of their relatively equal resource distribution, that made the seizure of power unappealing to the elites. Furthermore, not only does inequality promote coups, it also leads to the erosion of the quality of democratic institutions. When inequality is high the elites have more incentives to buy off politicians or to limit participation, and more genthe probability of success in the same way as the elites affect p(ζ, κ) in dictatorships. Democracies are far less repressive than dictatorships. Furthermore, the elites, because of their organizational capacity and their link with the military, can more easily affect this probability than the poor under a dictatorship. So, if I had included a probability of success it would have mainly be determined by the effort of the elites, not the population. Since the elites are more likely to exert effort when inequality is high, the qualitative results would be unchanged. One important factor in determining whether a coup succeeds is the cohesion of the elites and military (Barracca 2007). These are more likely to be united if inequality is high. 8 This condition is nearly the same as the one derived in Acemoglu and Robinson (2006) and has no claim of originality. Rather it illustrates the asymmetry between the democratization and consolidation games. 9 This is because the set of ϕ where the elites do coups increases with inequality.

21

erally to bias the political institutions toward themselves. Such regimes are partial democracies. Latin American countries are again good examples. Even if most of them are relatively safe from returning to dictatorship, many regimes are only low quality democracies. For example, Karl (2000) argues that, because of inequality, ”powerful economic and political elites have bent laws to their bidding, enfeebled courts, violated rights, corrupted politicians, and run roughshod over constitutions and contracts. [...] Inequality’s pernicious undermining of democratic aspirations, institutions, and rules is the greatest threat facing democracy in the Americas today” (p.155-56). Brazil is a good example of an unequal democracy that struggles to consolidate. For example, Linz and Stepan (1996) find that among a subset of seven Southern European and South American countries, Brazil is the farthest from full democracy. According to them, this is partially because ”Brazil has by far the most unequal distribution of income [...] of any of the southern European or South American or, for that matter, Central European countries in our set” (p.166). Democratic governments can try to prevent coups by limiting redistribution. But, the elites may still decide to overthrow the regime, for instance, because they fear redistribution in the future when they might not be strong enough to take power.10 If inequality is sufficiently high, limiting redistribution may not be enough to prevent coups. This has been the case for example in the Philippines where the democratic government was victim of 6 coup attempts between 1986 and 1989, despite limiting redistribution. These coups were directly aimed at preventing future redistribution policies, especially land reforms (Casper 2000). It is striking that most of the poor stable democracies are very equal. Some examples include India, Jamaica, Mauritius, Papua New Guinea and Mongolia since the fall of the communist regime. India is a particularly puzzling case. Despite 10

As argued by Acemoglu and Robinson (2006), redistribution today is not a credible commitment for redistribution in the future.

22

high poverty and illiteracy rates, and extreme ethnic and religious fractionization, it has been, with the exception of the 1975-77 period, surprisingly stable. However, India is equal and, contrary to countries like South Africa, neither of the two main groups (the Hindus and the Muslims) is significantly richer than the other. Lijphart (1996) argues that this absence of large inequalities between Hindus and Muslims is among the key factors explaining the stability of the Indian democracy. Most East European countries also represent examples of stable equal democracies. Greskovits (1997) indeed argues that one of the reasons why democracies are more stable in Eastern Europe than in Latin America is that the former are much more equal.

3 Data Dependent Variables: I construct the regime type variable with the polity scores of the Polity IV dataset. This variable ranges from -10 to 10, where -10 is given to the most authoritarian countries and 10 to the most democratic ones. It is based on three components: constraints on the executive, political competition and political participation. I define three regime types: dictatorship (-10 to 0), partial democracy (1 to 7) and democracy (8 to 10). These are the cutpoints used by Epstein et al. (2006).

Independent Variables: I measure inequality with the labor shares of GDP. It indicates the percentage of the GDP that goes to the labor, such that higher values denote more equality.11 This dataset is collected directly by the United Nations Industrial Development Organization (UNIDO) and covers more than 2400 countryyears and 123 countries between 1950 and 2001. It enables us to capture inequality 11

The correlation between the labor shares and the Gini coefficients of Deininger and Squire (1996) is -0.5191.

23

between classes in a manner consistent with my model, and those of previous authors. I thus measure the proportion of the income going to the population as the labor share of GDP.12 Most previous empirical studies on inequality and democracy use Gini coefficients. However, because the Gini coefficients are calculated from national surveys, that differ substantially across countries, they often lack comparability. The labor shares are collected by a single source, the United Nations, using the same definitions and methods, so they are comparable. Moreover, because the observations are collected by the United Nations – not the countries themselves – their quality is independent of other important variables, such as wealth or the regime type, again facilitating comparison across countries. The dataset also covers a far larger proportion of the country-years than in previous studies. Whereas the labor shares dataset contains 48.96% of the observations for the period 1950 to 2001, the one of Deininger and Squire (1996) only has 10.66%. In addition, because the data are not collected directly by the country, the sample is likely to be less biased toward rich democracies than, for example, with the Deininger and Squire (1996) dataset. For instance, while only 49 observations (about 8 %) in the Deininger and Squire (1996) dataset are from SubSaharan Africa, the labor share dataset contains 517 observations on Sub-Saharan Africa (about 19 %). The bias associated with the use of an incomplete, and nonrepresentative, dataset is reduced. The labor shares dataset has two other important caveats. First, it does not include the black market, which is important in poor countries. The GDP also omits the black market and so it is as well absent from the denominator in the calculation of the labor shares. Since the black market income goes to the lower class, inequality in poor countries will be overstated. This problem is also present in other in12

The labor shares are a direct measure of θ in the model of the previous section.

24

equality measures, such as the Gini coefficients, since by definition the black market is not covered in national statistics. Second, the earnings of the self-employed are not considered as labor revenues and are thus classified as belonging to the upper class. This is especially problematical in the cases of small entrepreneurs and subsistence farmers that should be included in the lower class. But, as argued by Baudey (2003), the necessary data to correct for this are not available for developing countries. However, the labor shares remain accurate estimates of inequality since agrarian countries – where the labor shares are low – are almost always extremely unequal. Indeed, in developing countries, peasants rarely own the land on which they work, such that most of the output goes to the landowners in the form of rent. To make sure that my results are robust to the inequality measure used, I test the relationship with different Gini coefficients and land distribution data in the robustness section.

Control Variables: My first control variable is GDP per capita, taken from the Penn World Tables, which is highly correlated with democracy (e.g., Burkhart and Lewis-Beck 1994). Wealth and inequality may also be linked, for example, through the Kuznets curve.13 I have included GDP per capita growth rates, taken from the Penn World Tables. Dramatic changes in wealth can destabilize political regimes (Haggard and Kaufman 1995). Growth may also affect inequality, since economic crisis or booms tend to affect diverse segments of the population differently. The level of urbanization also influences democracy and thus, I have included the percentage of the population that lives in urban areas. These data are taken from the World Bank. Religion is also thought to influence the regime type. I have included a variable 13

The Kuznets curve proposes an inverted U-shaped relationship between development and inequality, where inequality is at its height at intermediate levels of development.

25

measuring the percentage of the population that is protestant, which is argued to promote democracy. These data are taken from the dataset of Vanhanen (1997). Divided countries are also less likely to establish and maintain democratic institutions. We also see that inequality may not affect democracy the same way for instance in ethnically heterogenous and homogenous societies, as shown by the case of South Africa. I have used the measure of ethnic fractionization reported by Alesina et al. (2002). These indicate the probability that two individual selected randomly will not be from the same ethnic group. I have also included a dummy variable for former British colonies, which are argued to have inherited institutions particularly conducive of democracy. Finally, major oil producers are less likely to establish and maintain democracies. In such countries the elites are particularly vulnerable to taxation, because, unlike human or physical capital, natural resources are completely immobile (Boix 2003). Since major oil exporters also tend to be more unequal, the exclusion of this variable could bias the results. I use a dummy variable for OPEC membership.

4 Descriptive Statistics Before doing the statistical analysis I take a preliminary look at the data. One striking observation is that democratic countries tend to be more equal. Indeed, the mean labor shares in dictatorships, partial democracies and democracies are respectively 35.81, 40.05 and 48.51. This observation is consistent with several stories linking inequality and democracy: inequality may inhibit democratization, it may inhibit consolidation, or inequality itself may be affected by the political regime. We could also have a combinations of these possibilities. I argue that the observed association can mainly be accounted for by the second story. Tables 1 presents the probabilities of regime change at low, intermediate and

26

high labor share values, respectively. In each regime type, the observations are divided into three groups containing the same number of country-years. I report the probabilities of transition and the number of transitions within each group. The first row looks at the probabilities that dictatorships transit toward partial democracy or democracy. Contrary to what Boix (2003) argues, it is unequal dictatorships that are the most likely to become partial democracies or democracies. My model and those of previous authors consider communist and non-communist dictatorships as different regime types. The former are created by the masses after revolutions to redistribute assets, while the later are established and maintained by the elites to prevent redistribution. So, in the second row, I look at the same relationship, but without communist countries. The relationship is largely unchanged. As illustrated by the fourth row, equal partial democracies are not more likely to become democracies. I thus conclude that the hypothesis that transitions toward democracy are more likely in equal countries finds no support. The hypothesis of Acemoglu and Robinson (2006) does not find any support ei-

A to PD or D A to PD or D*

Labor Shares Low Middle High .049 (15) .0261 (8) .0411 (13) .0509 (15) .0237 (7) .0363 (11)

PD to A PD to D

.0763 (10) .0602 (8)

.0152 (2) .0227 (3)

.0222 (3) .0222 (3)

15 14

398 398

D to PD or A D to PD or A**

.0272 (11) .0449 (7)

0 (0) .0256 (4)

0 (0) 0 (0)

11 11

1229 473

# Transitions 36 33

N 929 893

Table 1: Probabilities of Transition per labor Share levels in each Regime Type (19502001). Number of transitions are in parentheses. A denotes authoritarian regimes, PD partial democracies and D democracies. * excludes communist countries and ** excludes Western countries and Japan. ther. There is even evidence of a U-shaped relationship between inequality and the probability of democratization, where middle inequality countries are less likely to 27

become democracies. There is also no inverted U-shaped relationship between inequality and the probability that a partial democracy becomes democratic. Therefore, the hypotheses of both Acemoglu and Robinson (2006) and Boix (2003) are not supported by the data. Inequality, however, has a strong positive effect on the probability that a democracy moves toward dictatorship. In fact, as shown in row 5, all the democracies that fell were among the most unequal tiers. One potential problem here is that the effect may be completely driven by rich developed democracies that are both stable and equal. So, in the last row I exclude Western countries and Japan. Once again unequal democracies are unstable. The most unequal democracies are nearly twice as likely to fall than those with intermediate inequality levels, while equal ones never broke down. I thus conclude that inequality indeed destabilizes democracies. These findings mean that the results obtained in the regression analysis below are not completely driven by my empirical methodology.

5 Results To analyze the relationship between inequality and democracy, I use a Markov transition model. This model is appealing because it enables us to distinguish between the effect of inequality on democratization and consolidation. Using a model that simply looks at the overall association between inequality and democracy would not have allowed me to differentiate the effect of inequality on democratization from its effect on stability once democracy has been created. Furthermore, this procedure distinguishes between the effect of inequality on democratization (or consolidation) in different regimes. For instance, inequality may affect the probability of transition toward democracy in dictatorships but not partial democracies. The Markov transition model provides different estimates for such

28

transition patterns.14 The Markov transition model generalizes the dynamic probit model, used by Boix (2003) and Pzreworski et al. (2000), to allow for more than two states, here regime types. Both of these authors only differentiate between democracies and dictatorships. I also include a partial democracy regime type. The Markov model estimates the probability of moving from any state to another state in a single year. I estimate the probability that, for instance, a country that starts as a dictatorship in period t will be a partial democracy in period t + 1.15 I estimate the following equation, derived by Epstein et al. (2006):

F [P r(Yit = b|Zit−1 = zit−1 )] = θb +

P1 l=0

αlb zit−1l + xit (β +

P1 l=0

γl zit−1l )

where i is the country, t the year and F [.] is the probit function. Also, b = 0, 1, 2 indicates the regime type in period t, where 0 is authoritarian, 1 partial democracy and 2 democracy. Let a = 0, 1, 2 be the regime type in period t − 1. Define zit−10 = 1 if a = 0 and 0 otherwise, and zit−11 = 1 if a ≤ 1 and 0 otherwise. To capture the distinct effect of the independent variables on each regime type, they are all interacted with both zit−10 and zit−11 , as show in the last part of the equation above. The effect of that variable is then calculated as the sum of the relevant estimated coefficients. Thus, the effect of, say, inequality in dictatorships is the sum of the coefficients on the inequality variable, inequality interacted with zit−10 and inequality interacted with zit−11 (β + γ0 + γ1 ).16 Its effect on partial democracies is the sum of the coef14

There are six transition patterns: dictatorship to partial democracy, dictatorship to democracy, partial democracy to dictatorship, partial democracy to democracy, democracy to dictatorship and democracy to partial democracy. 15 When estimating the effect of inequality on democracy, endogeneity may be problematical because not only does inequality affects democracy but democracy also affects inequality, for example through redistribution. Notice that here there is no endogeneity problem. I estimate, for instance, the effect of inequality on the probability that a country that was authoritarian in period t becomes democratic in period t + 1. But, then we cannot say that the democratic regime in period t + 1 influenced inequality in period t, while the country was still authoritarian. 16 This is because in dictatorships, both z0 and z1 take the value 1.

29

ficients on the inequality variable and inequality interacted with zit−11 (β + γ1 ). In democracies, the effect of inequality is simply given by its coefficient (β). The tables below only report the sum of the relevant coefficients and thus, the distinct effect of the variables in each regime type. A positive coefficient indicates that the corresponding variable promotes democratization – if the regime is dictatorial or partially democratic – or democratic stability – if it is democratic. The summary of the Markov estimates for the effect of labor shares on regime transition probabilities are presented in table 2.17 All independent variables are lagged.

0.30

Column 1 of table 2 reports the estimates when all countries are included. As

0.20 0.15 0.10 0.00

0.05

Predicted Probabilities

0.25

Non−Imputed Imputed

10

20

30

40

50

60

70

80

Labor Shares

Figure 5: Predicted Probabilities of Transition from Democracy to Dictatorship or Partial Democracy. expected, higher labor shares promote stability in democracies. Figure 5 plots the predicted probabilities of democratic breakdowns from column 1, when other variables are at their means or medians.18 The effect is substantial, especially in 17 18

I only report the sum of the relevant coefficients instead of the original regression results. The probabilities reported are the sum of the probabilities of moving to dictatorship and partial

30

All countries

A PD D A Labor Share2 PD D A Protestant PD D A log GDP pc PD D A Growth PD D A Urban PD D A Ethnic PD D A British PD D A OPEC PD D Labor Share

N

(1) -.004 -.006 .049***

(2) -.005 .002 -.548** .00001 -.00009 .008** .013 .012 -.002 -.001 .006 .002 .009 .009 -.182 -.189 1.161*** 1.407*** -.018* -.018* -.006 -.005 .095*** .114*** .003 .002 .015** .015** -.018 -.024* -.391 -.391 -.173 -.139 1.066 1.703* -.564** -.563** .525** .512** -.435 -.519 -.339 -.34 -.963** -.969** .646 .509 2403

Without Communist Region and Decade Dummies (3) (4) (5) (6) -.005 -.001 -.006 -.019 -.006 .002 -.001 -.019 .049*** -.551** .048*** -.654*** -.00005 .0001 -.0001 .0002 .008** .009*** .01 .01 .004 .004 -.001 -.001 .004 .005 .005 .002 -.003 -.008 .023 .022 -.06 -.046 -.185 -.193 -.408* -.391* 1.163*** 1.412*** 1.498*** 1.778*** -.015 -.015 -.016 -.016 -.006 -.006 -.004 -.005 .095*** .115*** .101*** .126*** .002 .002 .002 .003 .015** .015** .018** .018** -.017 -.024* -.028* -.037** -.318 -.32 -.105 -.101 -.176 -.14 .057 -.021 1.073 1.716* .874 1.515 -.519** -.52** -.415 -.388 .535** .522** .629** .684** -.433 -.519 -.416 -.499 -.349 -.346 .008 .02 -.981** -.987** -1.01** -1.023** .651 .514 .416 .443

2403

2367

2367

2367

2367

Table 2: Summary of Markov Estimates for Non-Imputed Labor Shares (1950-2001). Coefficients give the effect of each independent variable on each regime type. A denotes authoritarian regimes, PD partial democracies and D democracies. Positive coefficients indicate higher probability of moving toward democracy or staying democratic. All independent variables are lagged. ***p