Political Accountability and the Size of Government: Theory and Cross-Country Evidence

David Dreyer Lasseny Economic Policy Research Unit and University of Copenhagen November 2001

Abstract This paper explores the e ect of political accountability on the size of government. Political contestability and transparency increases political accountability, increasing the attractiveness of public goods provision, which in turn makes voters vote for higher taxes. The hypothesis is strongly supported by robust empirical evidence from a cross section of 62 democratic countries in 1995: Higher political accountability is associated with higher general government tax revenue.

Keywords: political accountability, control of politicians, size of government, democracy JEL-classi cation: D72, E62, H00

I am grateful to Toke Aidt, James Alt, John Ferejohn, Laurent Martin, Sren Bo Nielsen, Christian Schultz, Kenneth Shepsle, Torsten Slk, and participants in seminars, workshops and conferences at CPEG 2000 in Vancouver, EPRU at University of Copenhagen, Harvard, IIPF Annual Conference in Seville, and PET 2000 in Warwick for helpful comments, discussions and suggestions, and to Pablo Zoido-Lobat and Freedom House for answering questions about the data. The views presented, n as well as errors and omissions, are my responsibility. The paper was partly written while visiting Harvard University; I gratefully acknowledge the hospitality of Harvard and nancial support from EPRU, the Danish Research Academy, the COWIfoundation, and a Sasakawa International Fellowship. The activities of EPRU are nanced by a grant from the Danish National Research Foundation. y Economic Policy Research Unit, Institute of Economics, University of Copenhagen, Studiestr de 6, DK-1455 Copenhagen K, Denmark. Phone: +45 3532 4412. Fax: +45 3532 4444. E-mail: David.Dreyer.Lassen@econ.ku.dk.

 Where the control of o cials is weak, one would not expect citizens to allow o cials to command many resources. John Ferejohn (1990, p. 8).

Introduction

The size of government di ers greatly across democracies. In 1995, tax revenue was 8.0 percent of GDP in Guatemala, and 49.5 percent of GDP in Sweden.1 At the same time, the scope for holding politicians in check di ered greatly. In 1995 there were, according to the classi cation by Freedom House based on political rights and civil liberties, 76 free countries, while 62 were partly free and 53 were free2 This paper explores , not . the link between the degree of political accountability in democracies and the size of government. The basic argument of the paper is that increasing political accountability increases public control of politicians. Politicians concerned with being reelected will respond by increasing public goods provision for a given level of tax revenue. This increase in productivity, in turn, makes voters vote for higher taxes. Political accountability measures the degree to which voters can control elected leaders in a principal-agent relationship. Two central features of political accountability are political contestability and transparency. Political contestability makes it possible for voters to punish incumbents, come election time, if the incumbents actions or the resulting outcomes fail voter standards. Transparency makes it possible for voters to distinguish whether bad outcomes are due to corrupt politicians or adverse external circumstances. Both transparency and political contestability are necessary for control of politicians: Lack of transparency makes it harder for voters to infer the behavior of politicians from observables, in which case the degree of political contestability is less important. On the other hand, lack of contestability makes the issue of transparency moot, as it does not matter whether voters can observe the politicians actions if they cannot punish him at the polls anyway. Ferejohn (1999) shows that greater transparency in public decision making increases
1 2

Data from this paper; see appendix B. Karatnycky (1996, p. 3)

voter control over political agents which, in turn, increases the size of government demanded by voters in a democracy, but he does not test his hypothesis empirically. In this paper, I present a theoretical model showing that political contestability has similar e ects: by increasing political accountability it increases public control of politicians which, in turn, increases the share of resources allocated by voters to public goods production. Together, the models of Ferejohn (1999) and this paper implies the testable hypothesis that more political accountability should lead to larger governments, measured by tax revenue to GDP. I test the hypothesis that greater political accountability in democracies leads to larger governments on a sample of 62 democratic countries in 1995. The size of government is measured by tax revenues of general government to GDP. Existing studies measure the size of government by tax revenues or expenditures of central government; using a variety of sources, I add data for sub-national government tax revenue to get a consistent measure of the size of general government corresponding to that of the theoretical model. The predictions of the model nd strong and robust support: Countries with a higher degree of political accountability, measured by a composite index estimated by Kaufmann, Kraay and Zoido-Lobat (1999a,1999b), have signi cantly higher genn eral government tax revenues, controlling for a variety of other potential determinants of public sector size. In the literature, the general result (see below) has been that democracy measures turn out insigni cant. The di erent result obtained in this paper stems from two things: First, I focus on degrees of political accountability within democratic countries, rather than a binary measure of democracy vs. autocracy. Second, the use of general, rather than central, government data: Results are robust in the former but not in the latter case, suggesting that the institutional unit of analysis should be chosen with care in empirical studies of public nance. The paper is organized as follows: The next subsection brie y discusses related literature. Section 2 presents the theoretical model. Section 3 sets up the empirical speci cation, instruments and data, and section 4 presents the empirical results and their robustness. Section 5 concludes. 2

1.1

Relation to the Literature

There exists a large literature, and several excellent surveys thereof, seeking to explain the growth of government over time as well as cross-country di erences in the size of government from economic, political and cultural factors; see Holsey and Borcherding (1997) for a recent overview. A number of empirical studies, among the more recent ones Easterly and Rebelo (1993), Becker and Mulligan (1998) and Cheibub (1998), have found no statistically signi cant di erences between democratic and autocratic regimes with respect to the size of government. These studies employ a dichotomous de nition of democracy: either a country is democratic, or it is not. The model set up in this paper is concerned only with democratic regimes, but suggests that, given that a country is democratic, di erent degrees of political accountability can help explain di erences in the size of government; to the best of my knowledge, this paper is the rst to study this issue. Becker and Mulligan (1998) nd that tax systems that are more e cient, i.e. induce less economic distortions, lead to larger government. The results in this paper can be seen as extending this argument to political e ciency: Political regimes that are more e cient, i.e. reduce agency costs, lead to larger government. Persson, Roland and Tabellini (1997, 2000) and Persson and Tabellini (1999) focus on the e ects of political institutions, in particular whether a country is presidential or parliamentary, on the size of government. They identify two e ects: An indirect e ect, where presidentialism increases political accountability, which, in turn, decreases the size of government; and a direct e ect in which legislative cohesion, typically found in parliamentary democracies as a consequence of vote of con dence procedures, leads to larger governments. Persson and Tabellini (1999) present empirical evidence supporting the hypothesis that presidential democracies have smaller public sectors.3
Empirically, the correlation between presidentialism and the political accountability measure is -.45, such that presidential systems are associated with lower political accountability (see also Linz (1994)). This means that the theory presented in this paper can account for the empirical nding that presidential systems have smaller governments: If presidential systems in reality have less political accountability, voters have less control of politicians, leading to smaller governments, ceteris paribus, in presidential regimes.
3

The hypothesis of Persson et al. (1997) that more political accountability leads to smaller public sectors is in contrast to the result of the model presented in this paper, where more political accountability increases the size of government. The di erent predictions stem from di erences in the speci cation of preferences. When accountability increases, rents to politicians decrease, which, other things equal, decreases the size of government. However, in this paper the decrease in rents is o -set by the willingness of voters to supply more funds for public goods. Survey evidence supports the idea that more control of the public sector can increase voters willingness to pay taxes: A Harris poll of U.S. voters on the federal income tax (Wall Street Journal, January 8, 1997) reported that 64% [of voters] said they would be willing to pay more if they were sure the money would be spent ectively on public needs. Similarly, a survey of Tanzanian e tax payers shows that more than 90 % of respondents indicated willingness to pay more taxes if services were improved (Fjeldstad and Semboja, 2001). The sign of the e ect of political accountability is the subject of the empirical analysis of sections 4 and 5. Further, the hypothesis of a direct e ect of presidentialism on the size of government is tested by including it in the regression analysis; see the section on robustness of the empirical results.

The Model

The relationship between citizens and their leaders can be seen as a principal-agent relationship. Voters, the principals, choose leaders, the agents, who in turn rule the principals. Barro (1973), and later Ferejohn (1986) and Persson et al. (1997), examine the possibilities for voters to control elected politicians. Compared to the elaborate incentive schemes found in standard economic principal-agent problems, the main instrument available to voters for controlling their leaders is somewhat blunt: Elections. If performance is unsatisfactory,voters can vote for someone else. Thus, one central feature of elections is to provide a check on elected leaders by holding them accountable for their decisions and the outcomes resulting from these

decisions. If the possibilities for holding leaders accountable through elections are limited, voters lose control over their leaders and, as suggested by Ferejohn (1990, 1999), this can in uence the degree to which citizens are willing to entrust their leaders with authority and resources. In the model, a representative voter gets utility from a private and a public good.4 The political agent uses taxes paid by voters to nance public goods provision. The interests of the voters and the politician are not perfectly aligned; indeed, this is what gives rise to a principal-agent problem. Further, there is asymmetric information. The politician privately observes the realization of a stochastic variable which in uences the rate of transformation from tax revenue to public goods. Come election day, voters evaluate the outcome and reelect the politician if the outcome is satisfactory. The relationship between politicians and voters continues inde nitely. The model is one of purely retrospective voting: politicians can make no credible commitments to platforms or honesty, but are judged solely by their performance in the current term.5 A feature of the model sets it apart from standard principal-agent models of government: It is voters, not politicians, who decide the size of government. When choosing the size of government, voters have an outside option: they want also to consume private goods. Therefore, the agent the government is constrained in its actions, as the outside option creates competition for funds. That voters choose the size of government is a de ning feature of democracies, according to Przeworski and Limongi (1993, p. 58). They classify political regimes along two dimensions: Who decides the size of government, and who has the rights to the scal residuum that is left after public expenditures have been paid for, and they de ne a democracy as a regime in which citizens both decide the size of government and have a right to the scal residuum. Thus, the theory developed in this paper applies only to democratic regimes, but argues that political accountability di ers within the group of democratic regimes.
This can easily be generalized to a population of voters, di ering in income or their relative preferences for public goods. This does not yield additional insights, though. 5 There is no adverse selection in the model potential politicians are all equally good at providing public goods.
4

While voters decide the size of government, it is politicians and bureaucrats who determine expenditure allocations; in this case between private rents and the public good. This assumption re ects the observation that tax rates and burdens are directly observable, and the implementation of new tax laws has to approved by the legislature. Furthermore, taxes and the size of government are generally thought to be salient issues in elections; expenditure decisions, on the other hand, require detailed information collection and administration, as well as expertise by politicians and bureaucrats; therefore, expenditure decisions are normally less salient and delegated to politicians and government administration.

2.1

Voters and Politicians

The voter gets utility in each period from a public good x and a private good c according to the per period preferences U = x + H (c) It is assumed that H is increasing and strictly concave, and that limc!0 H 0 (c) = 1. The voter has income I in each period and pays taxes leaving c = I to be used for , plus future utility EU = xe + H (c) + V (1) private consumption. In any period, expected utility can be written as current utility

where V is the discounted expected continuation value and xe is the expected amount of public goods provided. In a stationary state, EU = V . Tax revenue is transformed into the public good according to the public budget constraint x = (r) where is a stochastic variable distributed according to a distribution function F , and r denotes rents appropriated by the politician. The shock captures the fact that voters 6

are not always able to thoroughly monitor the public machinery and that politicians can use this leeway for their own bene t. The politician observes before choosing rents; rents, therefore, are a function of , r (). Voters observing a low realization of x cannot determine whether this is due to a bad shock (low ) or a high amount of rents appropriated (high r). The expected amount of public good provided is Z e x = (r ()) dF () . The politicians preferences are given by EU A = W r + p (x) V A where p (x) 2 [0,1] is the probability that the politician is reelected and can enjoy the

discounted expected continuation value, V A , and W 2 [0,1] is the fraction of rents

appropriated by the politician that can actually be used for consumption; this suggests that there is a leaky bucketin transferring public funds to be used for the politicians own consumption. In the model, voters and politicians have con icting interests over rents, capturing the real world phenomenon that preferences of politicians and voters are not perfectly aligned. Rents should be thought of broadly as representing the loss to voters (consumers) from policies enacted by politicians for their own self-interest.6 < Figure 1 around here > The timing of the model is shown in Figure 1. First, the representative voter determines the share of resources to be used for public goods (in a standard voting model, this would be the median voter). Then, the politician observes the shock and chooses a level of rents r (). Thereafter, payo s are realized for both voters and politician and, nally, elections take place on whether to replace the politician.

In this simple model, rents drive a wedge between taxes paid and public goods received. However, trade protection, e.g. import quotas, enacted to receive campaign funds from protected industries see Grossman and Helpman (1994) increases rents to the politician by a ecting consumer prices. In this case, abolishing trade protection would bene t consumers, increasing income and increasing consumption of both private and public goods in a model with general preferences. A famous example is the U.S. sugar industry, see Dixit (1996) for this and other examples.

2.2

Voting

Aiming for a stationary subgame perfect Nash equilibrium, the model is solved backwards. At election day, voters choose whether to vote for the incumbent politician or an opponent. Since voters cannot observe neither r nor , they cannot condition their vote on these variables. As is standard in the literature (see, e.g., Banks and Sundaram (1998)), we assume that voters employ a cut-o rule when deciding whether to vote for the incumbent or not, voting for the incumbent if public goods provision x is above some threshold x.7 The probability that the voter votes for the incumbent politician is given by p (x) = 8 > < 1 if x x

However, in this formulation it is implicitly assumed that voters can costlessly vote for an identical opponent. To capture that this is not always the case, we posit the following voting rule: p (x) = 8 > < 1 if x x

> : 0 if x < x.

If the incumbent politician satis es the threshold level set by voters, he is reelected. If, however, he does not satisfy this threshold, he may be reelected anyway, depending on the degree of political contestability ( or, more broadly, degree of democracy), denoted E. Formally, (E) 2 [0,1] is the probability that there is no change of power even though to full democracy, and it is assumed that (0) = 1, 0 (E) < 0 and limE!1 (E) = 0.8
7 That voters are able to control the politician at all hinges on the policy space being one-dimensional. If the policy space has two or more dimensions in a model with many voters, problems akin to those of applications of the median voter theorem in multiple dimensions arise, since the politician, using personalized transfers, can e ectively pit the voters against each other, leaving the voters with no or very little control of the politician (see Ferejohn, 1986). Empirical research, however, suggests that people vote on the basis of aggregate variables (the state of the economy, say) rather than pure pocketbook considerations (Lewis-Beck (1988)). 8 An implicit assumption, made for tractability, in this formulation is that is the same regardless of how much smaller x is compared to the threshold x .

> : (E) if x < x.

voters vote against the incumbent. E can take on values in [0,1), from no democracy

The probability should not necessarily be interpreted as the probability that the incumbent will nullify the election should he lose, but rather that, for example if political competition is low, it may not be possible for voters always to nd viable alternatives to the incumbent, in which case the incumbent stays in power. Examples of less than ideal elections in democracies abound. For example, Przeworski and Limongi (1993, p. 58) note that [p]arty competition must be easily the most protected industry in the United States. In the US Presidential Election 2000, the Libertarian Party sued the Federal Election Commission, contending that rules for participation in presidential debates are so restrictive that they unfairly exclude minor9 party candidates and independents.

In the Mexican presidential contest in 2000, which was thought to be the fairest and most open yet, pro-government media executives banned opposition commercials for weeks, visas for foreign election observers were delayed, and campaign nance control was stymied. It was the rst election with voting booths that had curtains and that t only one person.10 In the 2000 presidential election in Yugoslavia, the o cial election result of the rst round was such that the incumbent, president Milosevic, was to have a run-o against the main opponent, Vojislav Kostunica, despite reports that the opposition had won by a clear majority. In the end, however, the opposition leader took o ce without a second round run-o . The de facto voting rule can perhaps be thought of in terms of barriers to entry in the political market. In the Besley-Coate model of representative democracy (Besley and Coate, 1997), citizen-candidates can enter the political market at some utility cost. Obviously, the higher is this cost, the fewer will contemplate candidacy. Further, the cost is presumably lower for incumbents (Tullock (1965); Crain (1977)). Entry costs are in uenced by the degree of political rights and civil liberties. In societies with infringements of such rights and liberties, either de jure or de facto, the costs of entering the political markets can be high, perhaps prohibitively so, leading to limited political
9 10

The Campaign Trail, New York Times, May 9, 2000. Mexico Pledges a Clean Vote, But the Old Ways Die Hard,New York Times, May 9, 2000.

contestability. Given this voting rule, the politician must choose between diverting all tax revenue as rents and winning with probability (E) and diverting an amount r () and being reelected with probability 1.

2.3

Equilibrium

Consider rst the case of symmetric information. In this case, voters observe, before they decide on the size of government, the realization along with the politician, and we have Result 1 Assume that the realization of the stochastic shock is observed by both

voters and the politician. In a stationary equilibrium, the politician diverts rsym = (1 ) > 0 for > 0. 1

and is reelected with probability 1.11 Even under complete information, the politician is able to divert some rents. This was also the message of Persson et al. (1997, Proposition 1). Note, however, that the presence of imperfect electoral contestability ( > 0) increases equilibrium rents; the threat that the politician can divert all tax revenue and still be reelected with positive probability increases the politicians expected utility from not providing, which increases the amount that voters must give up to obtain a positive supply of public goods. Consider now to the case of asymmetric information, where the realization is observed only by the politician. Let be the threshold of the realization of the shock such that the public budget constraint is satis ed with equality given x and r, r = x/.
11

(2)

Proof: In the case of symmetric information, voters know the utility of the politician, V A : Hence, the politician must be allowed to divert rents rs such that W rs +V A = W + V A , W rs +(1 ) V A = W where (in stationary equilibrium) V A = W rs + V A , V A = W rs /(1 ) since the politician is reelected with certainty if r = rs . Inserted above, we have W rs + (1 ) W rs /(1 ) = W , which, rearranging, gives the result.

10

The politician will choose to go for r and be reelected with certainty if W r + V A W + V A which, using (2), translates into a threshold value for the stochastic shock = x . V A (1 ) /W (3)

If the realization of the shock is greater than the politician diverts r () and stays in power with certainty. If the realization is less than , the politician diverts all tax revenue and stays in power with probability . A choice of a threshold x thus corresponds to a choice of a threshold value for the shock, . Given the politicians behavior, the voter chooses x to maximize current period utility: If the threshold is set too high, the politician will divert the maximum amount of rents and forego election in most cases. If the threshold is too low, the politician will be able to capture rents from asymmetric information and still be elected in most cases. The optimal threshold maximizes EU i = 1 [x Prob f g + 0 Prob f g + H (I )] 1

which, using (3), yields the rst order condition (see appendix A for a full derivation of the model) 1 F () = . f ()

The right hand side of this expression is the inverse of the hazard rate of the distribution function F . The inverse hazard rate is decreasing monotonically in for a number of common distributions, including the normal and the uniform, and, therefore, this expression has a unique solution and, therefore, a unique threshold x. For the remainder of the paper, we will, to get an explicit solution, assume that is uniformly distributed on [0,1]. This yields = 11 1 2

and, using (3) again, this provides an expression for the optimal threshold level of the public good, x = V A (1 ) /W . 2 (4)

which reduces to

Invoking stationarity EU A = V A , the incumbents utility V A is given by Z 1 Z 1 2 x A A A W W + V dF () V = W + V dF () + 1 0 2 VA = 2W . 2 [1 ln 2 + (1 + ln 2)] (5)

Invoking stationarity once again (EU = V ), we nd, using (1) and (4), voter utility to be (1 ) 1 + H (I . ) EU = 1 4 2[1 ln 2 + (1 + ln 2)] (6)

This is the utility function of the voter when choosing the size of government in the beginning of the period. The voter chooses to maximize utility in a stationary equilibrium. The rst order condition is (1 ) = H 0 (I . ) 4 2[1 ln 2 + (1 + ln 2)] (7)

The interpretation of this is completely standard: at an optimum, the marginal utility of public goods must equal the marginal utility of private goods. Solving for we nd (1 (E)) = I h . (8) (4 2[1 ln 2 + (E) (1 + ln 2)])

where h (H 0 )1 .

2.4

Comparative statics

The comparative statics of the representative voters preferred choice of tax rate, , follows directly:12
12

The comparative statics of this expression follows directly by di erentiating (8) in the text: ! @ 2(1) 0 = h < 0; @ (4 2[1 ln 2 + (1 + ln 2)])2

12

Proposition 1 The amount that a voter wishes to invest in the public sector increases as the degree of political contestability (E) increases. The greater is political contestability, the greater is the electoral threat and, hence, the control of politicians. This, in turn, leads to larger government. Note that it is not necessarily the case that the problem has an interior solution (i.e. > 0). From the ,c assumption on the shape of the H function, voters will always want to consume private goods. However, they will not always want to consume public goods. Only if (1 ) > H 0 (I) , 4 2[1 ln 2 + (1 + ln 2)] (9)

that is, if the marginal utility of public goods is greater than the marginal utility of private goods at c = I, will the voter wish to consume a positive amount of public goods. As noted in the introduction, Ferejohn (1999) nds a similar e ect of transparency: increasing transparency reduces the informational rents enjoyed by the politician, leading voters to invest more in government. Ferejohn, though, makes no attempt at testing his model. Ideally, one would want to evaluate the two theories separately; however, in practice the available proxies for political contestability and transparency are very highly correlated (with pair wise correlation coe cients of almost 0.9). Including them in the same regression, therefore, would lead to severe multicollinearity problems. However, having shown that the two components of political accountability in uence voter control of politicians in the same way, I can test the hypothesis that increasing political accountability increases the size of government. To do this, I use an aggregated index of political accountability which treats the observed proxies for political contestability and transparency as di erent realizations of the common, latent, concept
noting that h0 < 0 since H is strictly concave, that limz!0 h (z) = 1 and that (4 2[1 ln 2 + (1 + ln 2)]) > 0; the strict parenthesis is monotonically increasing in : At = 1, the entire parenthesis becomes 4 4 > 0 as < 1. The numerator is negative as < 1. Relating the result to the underlying political contestability parameter E we nd that @ @ 0 = (E) > 0. @E @

13

of political accountability. The next section sets up the empirical model and presents the data.

3
3.1

The Econometric Model

Speci cation, Identi cation, and Estimation

The relationship between the size of government and political accountability is given by SIZEGOV = S + S P OLACCT + S XS + ", (10)

where P OLACCT is a measure of political accountability and XS denotes other, exogenous variables a ecting the size of government (see below). There are two reasons why a simple OLS speci cation could be inappropriate in this case. First, there is the problem of measurement error: as will become clear below, the index measuring accountability is a composite of several general and expert opinion surveys. Both surveys in themselves as well as the aggregation procedure give rise to imprecision in the independent variable; letting P OLACCT be the observed measure of political accountability, what we observe is P OLACCT = P OLACCT + u (11)

where u is a random measurement error. It is well known that measurement error in an independent variable leads to attenuation bias, a downward bias of the OLS estimate of that variable. Second, there could be a problem of endogeneity or feed-back. In the theory outlined above, the causal mechanism is from accountability to the size of government. If there is a feed-back e ect from the size of government to accountability, however, OLS estimates will be inconsistent (the simultaneous-equations bias). Such a feed-back could, perceivably, be both positive and negative. If the government suddenly expands its mode of operation to issues it had not in uenced before, a demand for greater electoral control

14

could arise.13 On the other hand, an increase in the volumeof government operations to new issues could in itself decrease accountability, as the dimension of the issue space would increase (cf. footnote 7). In the case of endogeneity, the model (10) and (11) must be extended with P OLACCT = P + P SIZEGOV + P XP + e. (12)

In addition to the size of government, other factors a ect political accountability as well; these are captured in XP , and will be discussed below. Using (10) and (11) together, we obtain SIZEGOV = S + S P OLACCT + S XS + , " S u which together with (12) constitutes the econometric model. This simultaneous equations model is identi ed by assuming that the composite error term in (13) is uncorrelated with the other determinants of political accountability captured in XP , i.e. that E [XP "] = E [XP u] = 0. Under this assumption, which measure P OLACCT when estimating (13). The identifying assumption imposes the restriction that the subset of XP used as instruments does not a ect the size of government directly, but only through its e ect on political accountability (E [XP "] = 0), and that the instruments are uncorrelated with the measurement error (E [XP u] = 0). This will be taken up in the discussion of the instruments below. will be tested, any subset of XP can be used as instruments for the true accountability (13)

Below, I begin by estimating (13) by two-stage least squares, correcting for heteroscedasticity in the errors. The identifying assumption is tested by a speci cation test for overidentifying restrictions. From there, I then estimate the full system (12) and (13), explaining simultaneously the degree of political accountability and the size of government.
For example, it has been argued that the rati cation of the Maastricht Treaty of the European Union, which increased the power of the European Commission (the executive branch of the EU) signi cantly, resulted in a democratic de citas the Commission is only very indirectly accountable to citizens of the EU.
13

15

3.2

Controls

The agency model set up in the previous section is a very bare-bones one. Therefore, when estimating the model, a number of variables suggested in the literature is included in XS in (13): In addition to the standard income level of a country, I include the following controls: Openness of the economy (Cameron, 1978, and Rodrik, 1998), population size (Alesina and Wacziarg, 1998), the dependency ratio de ned as the ratio of old and young to the working population (Rodrik, 1998; Alesina and Wacziarg, 1998), and the degree of urbanization (e.g. Oates, 1985, and North, 1985). Rodrik (1998) argues that in more open economies voters want more social insurance from the government to shield them from the increased variability owing to openness; along this vein, Alesina and Wacziarg (1998) argues that it could be country size that determines both openness and the size of government: large countries are less open and have smaller public sectors. The dependency ratio is included to control for demographic structure; whether this or other measures of the age-distribution should be used is discussed in more detail below. Finally, urbanization has long been held to in uence the size of government: One version of the argument (Oates, 1985) is that increased urbanization increases the demand for public goods, increasing the size of government.

3.3

Instruments

The requirements for the instruments (XP in (12)) are that they are (i) correlated with the measure of political accountability, and (ii) uncorrelated with the disturbance term . The instruments used are (1) distance from the equator, (2) whether sub-national governments have substantial scal authority and (3) the fraction of the population who are Protestants. These instruments are correlated with political accountability. As argued by Hall and Jones (1999), Western European in uence is correlated with distance from the equator: Western European colonization was likely to take place in sparsely populated regions and in regions where the climate was similar to that of Western Europe, both of which

16

points to regions away from the equator. There seems to be no reason why distance from the equator, in itself, would in uence the size of government. Federalism has often been linked to democracy (Lipset, 1959; Vanhanen, 1990). Federalism, it has been argued, tends to reduce (ethnic) political con ict, helping sustain democracy (Lipset, 1994). Further, sub-national governments provide division of powers, serving as checks and balances on the central government (Lijphart, 1999) and on each other (Gordon and Wilson, 2001), increasing political accountability. Rather than considering (constitutional) federalism per se, I follow the Polity III database (Jaggers and Gurr, 1995) in classifying countries as federal if there exist sub-national governments with substantial scal authority. Does the existence of lower levels of government a ect the size of government directly? The large literature on Leviathan remains inconclusive (e.g. Oates, 1985), possibly because there are many con icting e ects of decentralization. Finally, a large literature argues that religious tradition is a major di erentiating factor both in the transformation to, and in the maintenance of, democracy. Huntington (1991) and Lipset (1959, 1994), among others, note that Protestantism (as opposed to Catholicism, Orthodox Christianity, Confucianism and Islam) and democracy have been positively interlinked historically, due to the much greater emphasis on individualism found in Protestantism and to traditionally strong links between religion and state in the other main religions. Barro (1999) provides recent empirical evidence that, indeed, the share of Protestants in the population is positively associated with the degree of democracy. Given the relatively weak link between religion and state in Protestant countries, there should be no reason why Protestantism should in uence the size of government directly. The arguments above imply that the instruments in uence the size of government only through the e ect on political accountability. Due to the presence of overidentifying restrictions in the model, this hypothesis can be tested and, it turns out, cannot be rejected. To further investigate the robustness of results regarding the choice of instruments, the three instruments were excluded one at a time and instead included as an additional explanatory variable when estimating the size of government equation by 17

two-stage least squares. This did not a ect results on political accountability and the variables included in the estimating equation were never close to signi cant; see the section on robustness below.

3.4
3.4.1

Data
Measuring accountability

Most aspects of political accountability, including the concepts of transparency and political contestability considered in this paper, are di cult to measure. Political accountability is measured by an index of political accountability and voice constructed by , Kaufman et al. (1999a,1999b). This index is constructed by combining subjective indicators with survey evidence. It is cross-country only and based on data for 1997 and 1998. The construction of the index is grounded in the idea that although various subjective measures of accountability do not focus on the exact same characteristics of the political regimes surveyed, they can be seen as re ecting (random) variations in an underlying, fundamental concept of accountability. Aggregating di erent measures in this way, using an unobserved components model, allows for (i) constructing a larger data set and (ii) assessing the precision of the various measures of governance.14 The accountability index captures political contestability and transparency in the following way: The surveys and subjective indicators, described in detail in Appendix B, that constitutes the index are: Orderly political transfers, Transparency and Fairness of the Legal System (The Economist Intelligence Unit); Civil Liberties, Political Rights, Free Press (Freedom House); Military in Politics, Democratic Accountability (Political Risk Services); and Information to Business about Government Policies; Voice to Business to express concerns about policies (World Development Report). The accountability index is normalized such that, for the entire sample, it can take on values from -2.5 to 2.5, it has a mean of zero and a standard deviation of one.
See Bollen (1980, 1993) for a thorough discussion of such models sometimes called structural models with latent variables and their statistical properties in the context of political democracy.
14

18

Three of the components, Free Press, Information to Business about Government Policies, and Voice to Business, measure if and how information about government actions and policies are made available to the public. Rose-Ackerman (1999, p. 162) argues that [t]he public can be an important check on the arbitrary exercise of power by government. [...] [T]his check can operate only if the government provides information on its actions. Information will only be e cient in controlling public o cials, however, if individuals or the media is able to convey this information to the public without fear of reprisal. Thus, free press is an essential check, [...] [a]nd if elections are important, a the media is also crucial. (op.cit., p. 166). The components Orderly Political Transfers, Civil Liberties, Political Rights, Military in Politics and Democratic Accountability concern what we have tried to capture in the E parameter in the model. Finally, Transparency and Fairness in the Legal System is important for electoral control of politicians as corrupt judiciary is costly for a democracy because it cannot credibly play the role of watchdog on constitutional values or monitor the honesty of the other branches of government. (op.cit., p. 156). The advantage of subjective indicators and surveys like these is twofold: First, as noted above, objective measures for political contestability and transparency may be di cult, if at all possible, to obtain. Second, economic and political decisions are taken based on subjective opinions of people involved and these opinions may di er from objective measures of political and economic institutions, although they are certainly framed by such institutions. The ip side is that subjective measures, by de nition, fail to meet standards of reproducibility and further are imprecise, due to the polling , and survey nature of the data. As noted above, the problem of measurement errors is addressed by using instrumental variables. The main problem in using the accountability index is that several di erent e ects work as one and therefore it can be di cult to disentangle exactly what the index , is measuring. To address this problem, I perform robustness checks below using single components of the aggregate index. However, it may be that it is exactly the interaction between transparency and political contestability that is essential: Examining the role 19

of the media in government responsiveness in India, Besley and Burgess (2000, p. 22) nd that the formal institutions of political competition (such as open elections) are not su cient to deliver a responsive government unless voters have the real authority to discipline poorly functioning incumbents. This requires voters to have the necessary information. Thus, according to Besley and Burgess, political contestability and transparency are complements. This is captured by using the aggregated index when the underlying indicators are positively correlated. 3.4.2 Democratic countries

The model is developed within the context of democratic countries. To operationalize democracy the classi cation suggested by Freedom House is used. On their scale of , political rights and civil liberties, countries are assigned a value between 1 (the most free) and 7 (the most authoritarian); further, Freedom House classi es countries with a value greater than 5 as free Hence, in the present study a country is classi ed not . as democratic if in the period 1990-1995 it did not take on a value greater than 5 in neither the political rights nor the civil liberties index at any time. The selection of countries from that pool of democratic countries was based entirely on data availability; in particular, public nance data are often di cult to obtain. This procedure results in a sample of 62 democratic regimes. 3.4.3 Measuring the size of government

The size of government is de ned as tax revenue of all levels of government to GDP in 1995. The revenue data is from Government Finance Statistics Yearbook 2000 (2001), the most detailed source of public nance data available for a large number of countries. A problem with this data, however, is that for many countries regional and local government nances are not reported. The problem is potentially serious as the lack of reports on regional and local government nances is not evenly spread across the sample but concentrated among the countries with low accountability. Failing to take into account local government nances

20

thus underestimates the size of government in these countries, and could therefore bias the results in favor of the hypothesis. Leaving out revenue collected by sub-national governments is also problematic, however, as the theory implicitly like most theories of the size of government concerns the level of general, not only central, government. To deal with these problems, I have collected additional data on state and local government tax revenues from a variety of sources.15 Figure 2 shows the covariation between the degree of centralization (the share of central government tax revenue to total tax revenue) and the degree of political accountability. < Figure 2 around here > As is clear from the gure, sub-national government tax revenue is substantial in some countries: In Canada, Switzerland and Brazil, sub-national tax revenue accounts for more than 40 % of total tax revenue. The gure shows a tendency, without implying causality, for countries with higher political accountability to be more decentralized. This has profound e ects on the results. When including the full range of control variables, political accountability is sometimes insigni cant when the size of central government is the dependent variable; however, it becomes signi cant once we estimate the (correct) model using the size of general government; see the section on robustness (below).16 3.4.4 Controls

GDP per capita (1995) is taken at PPP, and openness of the economy is de ned as the volume of trade to GDP in 1995. The dependency ratio is the population aged 0-14 and
15 The data and sources on sub-national government revenues is described in appendix B. It was possible to determine tax revenue of state and local governments for all the democratic countries for which political accountability and central government tax revenues existed. Hence, the availability of local government tax revenue data did not determine the sample of countries for the empirical analysis. 16 An additional issue is the measurement of GDP. The existence of informal sectors implies that o cial GDP is a lower limit for true GDP. The size of the informal sector is inversely related to political accountability. Using informal sector data from Friedman, Johnson, Kaufman and Zoido-Lobat n (2000), a simple regression of the size of the informal sector on the index of political accountability reveals a signi cant negative relationship. Thus, an estimate of the size of government as resources controlled by the state relative to economic activity will be (more) overestimated in less accountable countries, implying that the e ect of political accountability probably is larger than that found here.

21

65+ to the population aged 15-64, and the degree of urbanization is de ned as the share of the population living in urban areas. These variables are from World Development Indicators, as is population in 1995. Data on latitude is from the Global Demography Project, used by Hall and Jones (1999). Fiscal authority of sub-national government is taken from Polity III (Jaggers and Gurr, 1995), and the share of Protestants from La Porta et al. (1999). For further information about the data, including sources and descriptive statistics, see the appendix.

4
4.1

Results
Basic results

Figure 3 plots the size of government against the measure of political accountability for the 62 countries in the sample. The simple correlation between the political accountability index and the size of general government is .79. < Figure 3 around here > Estimating the univariate relationship of gure 2 yields SIZEGOV R
2

= 15.970 + 12.992P OLACCT ,

(1.227) (1.268)

= .62, F (1,60) = 105.03,

where robust standard errors are in parenthesis. In this simple case, political accountability (P OLACCT ) has a coe cient of 12.992, meaning that a one standard deviation increase in political accountability increases the size of general government by 13 percentage points of GDP.17 Political accountability is signi cant at the .1 % level (with a t-statistic of 10.250), and it explains 62 percent of the variation in the size of government.
In all regressions, the dependent variable is the ratio of tax revenue to GDP. As this is constrained to be between zero and one, a more correct econometric speci cation would employ the logistic transformation log(tax/(1-tax)) as used by, e.g., Oates (1985). However, the statistical results are una ected by this transformation, so the standard formulation has been used to facilitate interpretation of the results.
17

22

This is a strong relationship, but, of course, it is necessary to control for other potential causes of government size. Table 1 presents the results when the full set of controls described above have been included: log(income), log (openness), log(population), urbanization, and a demographic variable. < Table 1 around here > In the literature, a demographic measure is often included as an explanatory variable; however, there is no consensus on how to measure the impact of a countrys demographic composition on the size of its government. For example, Rodrik (1998) and Alesina and Wacziarg (1998) use the dependency ratio, whereas Persson and Tabellini (1999) use the shares of young and old in the population separately. To show that the general thrust of my argument does not depend on the choice of measurement with respect to demographics, I estimate the model using both measures separately. The results of table 1 lend strong support to the hypothesis that more political accountability leads to larger government sectors. The coe cient on political accountability is positive and estimated with signi cant precision, being signi cant at the .1 percent and the 1 percent levels.18 Regarding controls, the general picture is that openness, urbanization and population size enter positively and signi cant, while income and the demographic variables except the share of the population aged over 65 are insigni cant.19 That population size is positively signed is in contrast to the results of Alesina
18 The results shown in table 1 do not depend on outliers; the method suggested by Hadi (1994) identi ed Singapore as the only outlier in the multivariate analysis; this is due to Singapore having a very high degree of openness (imports + exports are more than 300 % of GDP). Removing Singapore from the sample does nothing but strenghten the openness variable. 19 In general, the share of the population over 65 enters positively and is strongly signi cant when estimating by OLS, but not in instrumental variables regressions. In contrast, the dependency ratio is never signi cant. Furthermore, R2 of the OLS regression increases considerably when including the demographic shares seperately. However, the demographic composition of the population is almost certainly an endogenous variable with respect to income, and probably also with respect to the size of government. In particular, the so-called demographic transition (de ned by decreasing mortality rates followed by decreasing fertility rates) has lead to to the current demographic structure with relatively many elderly people found in advanced economies. This demographic transition is argued to be a consequence of economic development; see Galor and Weil (2000). The implications of this for empirical studies of public nance is left for future research.

23

and Wacziarg (1998); a possible reason is the di erent and smaller sample employed in this study. However, population size and openness are generally only signi cant when they are both included in the regression; when one is dropped, the other becomes insignificant. This, again, could stem from the use of general, rather than central, government revenue gures. Columns two and four report the results of two-stage least squares estimates of the size of government.20 It is notable that the estimated coe cients approximately double in size compared to those of the OLS estimations. This suggest (and is indeed con rmed by formal tests) that accounting for simultaneity and measurement errors is appropriate. Also reported in table 1 are the F -test statistics from the rst stage of the two-stage least squares estimation. Many standard results in instrumental variables regression rely on the instruments not being weak that is, the partial correlation between instruments , and the included endogenous variable P OL should not be too low. As a rule of thumb, Staiger and Stock (1997) suggest that the F -statistic of the rst stage regression should be greater than 10 for standard results to apply; in the table, the F -statistic from the rst stage regression is 25.62, suggesting that the instruments are su ciently strong for the 2SLS estimates to be reliable.21 Finally, table 1 reports for each of the instrumental variables regressions the p-value associated with the test of the no overidenti cation hypothesis. As argued above, the e ect of the instruments is not directly on the size of government, but rather through political accountability. Since the number of instruments (3) is greater than the number of endogenous variables (1), the equation is overidenti ed. If the test for overidenti cation fails, one or more of the instruments should be included as an exogenous variable
These estimation were carried out using GMM-estimation methods, as this method automatically accounts for heteroscedasticity in the errors when computing the test for overidenti cation (Hansen, 1982). 21 The estimated rst stage regression is P OLACCT = .080 + 1.564LAT IT U DE + .465 SU BN AT IONAL + .005 P ROT EST AN T
(.360) (.132) (.003) 20

with standard errors in parentheses and R2 =.57. Latitude and subnational are signi cant at the .1 % level, while the share of Protestants is signi cant at the 5 % level.

24

directly in the estimated equation. As the p-values are all high, the hypothesis of no overidenti cation cannot be rejected. Below, I check the robustness of the choice of instruments in more detail. One thing to note about the instrumental variables regressions: the coe cient of income is negative (though insigni cant), although the simple correlation between income per capita and the size of general government is .68. This may be due to multicollinearity (see below), as the correlation between income per capita and the political accountability index is .70. The next section tries to address this question in part by following Barro (1999) in letting income be a determinant of the degree of democracy.

4.2

System estimation

Although my main focus is the size of government, setting up the simultaneous equation system consisting of (12) and (13) allows for estimating the size of government and the degree of political accountability jointly, letting the size of government have a feedback e ect on the degree of political accountability. This is done using three-stage least squares. Potentially, this can have an impact on results by allowing a more general covariance structure that makes use of cross-equation correlations of the disturbances.22 The estimated model is similar in structure to the 2SLS model estimated above, except that log(income) is included as an explanatory variable for political accountability, following Barro (1999). Table 2 reports the results. < Table 2 around here > In general, the standard errors of most estimates somewhat larger than in the 2SLS estimations reported above, perhaps owing to the relatively small sample size. However, the message regarding political accountability is the same: The political accountability index is positive and signi cant, with the level depending on the demographic speci cation. When the age-shares are included separately, the size of government a ects the
On the other hand, although full system estimation is asymptotically more e cient than singleequation estimators such as 2SLS, it is less stable in the presence of speci cation errors. Also, in nite samples one might not reap the bene ts of 3SLS suggested by asymptotic e ciency results. See Greene (1997, ch. 16).
22

25

degree of political accountability in a positive and signi cant way; however, this results disappears when using the dependency ratio.

4.3

Robustness

To ensure that the results reported above are robust, a number of di erent speci cations were used. First, the political accountability index is replaced with the civil liberties and political rights indicators from Freedom House. These indicators are a part of the political accountability index, and therefore necessarily highly correlated with it, but, as these indices are often used on their own (e.g. Barro, 1999, and La Porta et al., 1999), it would be interesting to see whether the results apply also in this case. The average of each index over the period 1990-1995, which is the period in which the countries should be democratic to be included in the sample, is used. < Table 3 around here > The results are shown in table 3.23 Generally, both the civil liberties and political rights indices are signi cant the one exception is the OLS estimation using political rights when age shares are included separately (column 7). Both political rights and civil liberties are signi cant when accounting for their possible endogeneity using 2SLS. Above, correcting for endogeneity doubled the coe cient on political accountability; here, the coe cients on the political indices in some cases more than triple compared to the OLS estimates, suggesting, again, that instrumental variables should be employed when using these measures. The rst stage F -statistic is su cient and the hypothesis of no overidenti cation is not rejected. It would be interesting to separate the e ects of political contestability and transparency: although they in uence the size of government in a similar fashion, estimating these e ects separately could give some indication as to their importance, and would
Note that the sign on the indices is reversed, such that in the regression a higher value of CL9095 and P R9095 corresponds to more civil liberties and political rights, respectively. In the instrumental variables regressions, I use only latitude and subnational governments as instruments, as the share of Protestants is no longer signi cant in the rst stage regression. Full system estimation, using 3SLS as in the previous section, yields identical results.
23

26

make it possible to test whether transparency and contestability are in fact complements as suggested by Rose-Ackerman (1999) and Besley and Burgess (2000). However, the widest available measure of transparency used in constructing the index, Freedom Houses Index of Press Freedom, is very highly correlated with the measures of political rights and civil liberties indeed, this is why constructing the index makes sense in the rst place. The partial correlation coe cients are almost .9, which means that multicollinearity will be a severe problem, in particular when using instrumental variables.24 Including Press Freedom on its own, without the measures of contestability, gives results similar to those of political rights and civil liberties. To get some bearing, however, on this issue, I include a measure of availability of alternative sources of information than that from the state. This is a part of Dahls (1971) concept of polyarchy, and the measurement was made by Coppedge and Reinicke (1990). The measure, coded here such that a higher value of the index means more access to alternative information sources, is out of sync with the other data, as it pertains to the late 1980s; however, having been measured over a long period time, most of this was probably true also in the beginning of the 1990s. I consider only OLS estimates in this analysis; instrumental variables regression yield no signi cant variables at all due to multicollinearity. < Table 4 around here > Included alongside the civil liberties and political rights indices, the access to alternative information sources (ALTINF in the table) has the correct positive sign, but p-values around .2, while the political indices are signi cant as before. However, when I include an interactive term of the political variable multiplied with ALTINF, both the political variable, ALTINF and the interactive term are strongly signi cant and positive. The positive sign on the interaction variable con rms the argument by Rose-Ackerman
A con rmatory factor analysis (see Bollen (1989, ch. 7)) of civil liberties, political rights, and Press Freedom in 1995 strongly suggests that they should be seen as realizations of one common latent variable and, hence, should be aggregated into one index. Using such an index yield results identical to those of the political accountability index.
24

27

(2000) and Besley and Burgess (2000) that formal political institutions cannot be seen independently from transparency and access to alternative information sources; this, I believe, further supports the use of the composite political accountability index including both transparency and political contestability. I also examined the impact of additional explanatory variables. As already discussed above, Persson and Tabellini (1999) nd that presidential systems have smaller governments; further, Alesina, Baqir and Easterly (1999) nd that the degree of ethnic fragmentation leads to lower public expenditures for US states, while Becker and Mulligan (1998) argues that the share of the labor force not employed in agriculture constitutes a proxy for the e ciency of the tax system, which they nd to increase the size of government. I include these three variables one at a time, and the results are shown in table 5, where I have reported only the coe cients of political accountability and the additional variables.25 The regressions were run using the full set of controls as above: log(openness), log(income), log(population), urbanization and dependency ratio. < Table 5 around here > Including a dummy for presidential systems does not in uence results; political accountability remains signi cant at the .1 % level. The dummy for presidential systems has the correct sign in the OLS case, but is not signi cant. The sign reverses when 2SLS is used. A potential explanation for the insigni cance of the presidentialism dummy can be seen from gure 2: Some of the countries with presidential systems also have large tax revenues generated by sub-national governments; therefore, leaving out sub-national governments can in uence results. Indeed, when regressing public sector size only on political accountability and presidentialism, political accountability is positive and strongly signi cant when considering both general and central government tax revenues, but presidentialism is signi cant only in the central government tax revenue case.
The measure of ethnolinguistic fractionalization is taken from Easterly and Levine (1997), the classi cation of political regimes into presidential and parliamentary systems is based on Beck et al. (2000) and Lane and Ersson (2000), and the share of the labor force not employed in agriculture is from WDI (1999). The regression including tax e ciency is based on 61 observations.
25

28

Ethnolinguistic fractionalization has the expected negative sign, with p-values around .10, while the measure for tax e ciency (the share of the labor force not employed in agriculture) is not signi cant, and further, causes the coe cient on urbanization to become insigni cant, as these are (almost) mirror images of each other. In sum, including these additional variables have no e ect on results. Results do not change (compared to those found in table 1) if age-shares were included separately, and neither if the additional variables were included together. To check robustness of the results regarding the choice of instruments, additional robustness analysis were performed. Although the test for overidenti cation with respect to instruments never rejected the assumption of no overidenti cation, I also excluded each of the instruments one at a time and instead included them in the main regression. The results are shown in table 6. < Table 6 around here > The rst column in table 6 is simply reproduced from table 1 for comparison. As above, the regressions were run with the full set of controls and, again, the results do not depend in the choice of demographic control variable. In all cases the F -test statistic of the rst stage regression is very high, and the instruments included in the main regression are never close to signi cant. As mentioned above, results are generally weaker when central government tax revenue is used as dependent variable rather than tax revenue of general government. When demographics are measured by the dependency ratio, both the political accountability index, as well as civil liberties and political rights, remain signi cant. However, when the age-shares are included separately, the measures of accountability, political rights and civil liberties generally cease to be signi cant, in particular in the 2SLS speci cations. This suggests that care should be taken when interpreting results regarding degrees of democracy and the size of government when the analysis is based on central government data.

29

Conclusion

This paper has investigated the e ect of political accountability on the size of government in a democracy. In a principal-agent model of government it was shown that increases in political accountability, re ecting transparency and political contestability, leads to increases in the size of government. This, perhaps surprising, claim was investigated empirically on a cross section of 62 democratic countries using a recent index of political accountability, and was found to have strong and robust support in the data. Further, the results of this paper show that it is important to account for state and local governments in cross-country analyses of the public sector; I show that political accountability signi cantly increases the size of general government, but not that of central government. This suggests that determining the correct unit of analysis is important when evaluating theories of the public sector empirically.26 La Porta, Lopez-de-Silanes, Shleifer and Vishny (1999) examine the determinants of quality of governmentunderstood as the good-for-capitalist-developmentand they nd consistently ... that the better performing governments are also larger, and collect higher taxes. Poorly performing governments, in contrast, are smaller and collect fewer taxes ... [This] tells us that identifying big government with bad government can be highly misleading. [p. 266]. The analysis and empirical results in this paper can be seen as providing an explanation for this phenomenon: In democracies, bad governments do not get to be big. The results of this paper also provides an explanation of why increasing decentralization can have ambiguous e ects on the size of government. One the one hand, decentralization can result in a race to the bottom with tax competition resulting in , smaller governments. On the other hand decentralization typically increases political accountability, which by the logic advanced in this paper can lead to increases in the size of government, ceteris paribus. This could help explain the mixed nding of the Leviathan literature, e.g. Oates (1985).
Potentially, measuring the size of the public sector by general rather than central government could also a ect results in, for example, the literature on the e ects of scal policy on economic growth.
26

30

Future work should look at alternative measures of transparency and political contestability in more detail, and examine their impact on scal outcomes, preferably in environments where one or the other can be kept relatively xed. Along these lines, preliminary results of Alt, Lassen and Skilling (2001) suggest that scal transparency increases the size of government in a cross-section of U.S. states.

31

Derivations and Proofs

EU i = 1 [x Prob f g + 0 Prob f g + H (I )] i 1 i 1 V (1 ) =W
A

The optimal threshold maximizes

which yields the rst order condition

(1 F ()) = V A (1 ) =W f () and, rearranging, = 1 F () . f ()

Assuming that is uniformly distributed (in which case F () = and f () = 1) on [0,1] yields = 1 , = 1 2

which, using (3), provides an expression for the optimal threshold level of the public good, x = The politicianss utility is given by V
A

V A (1 ) =W . 2 Z 1 x A + V dF () W W 1 2 V A (1 ) [ln ]1 1 2 2

1 2

V A = W + V A F ( ) + (1 F ()) V A and, rearranging, VA =

W + V A dF () +

2W . 2 [1 ln 2 + (1 + ln 2)]

Inserting this expression into the utility function directly yields (6) in the text, and from this straightforward maximization yields equations (7) and (8) in the text.

32

Data
Description Tax revenue of general/central government to GDP at market prices Source(s) GFS, WDI Kaufman et al. (1999b) WDI WDI Hall and Jones (1999)

Variable SIZEGOV POLACCT LNINC LNOPEN LAT

Political accountability index Log of income per capita (at PPP in international $) Log of volume of trade (imports + exports) to GDP Latitude of the center of the county or province with the largest number of people

FEDERAL PROT PRES CL9095 PR9095 PRESS ALTINF URBAN

=1 if sub-national governments have substantial scal authority Share of protestants in population = 1 if presidential system Average of Civil Liberties index, 1990-1995 Average of Political Rights index, 1990-1995 Press Freedom Index Availability of alternative sources of information Fraction of midyear population living in areas de ned as urban in each country

Polity III, Gurr et al. (1995) La Porta et al. (1999) Beck et al. (2000) www.freedomhouse.org www.freedomhouse.org www.freedomhouse.org Copperidge and Reinicke (1990) WDI

DEPEND

Age dependency ratio = (population < 15 and > 64) / (population 15-64)

WDI WDI WDI WDI

>65 <15 LNPOP95

Share of population aged 65 or over Share of population aged 14 or less Log of population

33

* Government Finance Statistics Yearbook 2000 (2001). ** World Development Indicators, 1970-1995. *** From the Global Demography Project at University of California, Santa Barbara, used by Hall and Jones (1999). Normalized to [0,1]. **** Based on the classi cation in Beck et al. (2000).

B.1

Revenue data

To ensure that 1995 is not an exceptional year regarding public nances, one could have taken averages of several years. However, in the democratic regimes focused on here, tax revenues to GDP are relatively stable. Further, averaging over countries in the 1990s would be di cult due to a wave of standardization of national accounts that has created structural breaks in the time series of many countries. All gures, except where noted, are for 1995. The discrepancy between the dating of public nance data and accountability data is entirely due to data availability. Also, the selection among democratic countries is entirely due to data availability. Missing data excludes, e.g., Honduras, Jamaica and PNG. Countries included are: Argentina, Australia, Austria, Belgium, Bulgaria, Belize, Bolivia, Brazil,a Botswana, Canada,a Switzerland, Chile, Colombia,a Costa Rica, Denmark, Germany, Dominican Republic, Ecuador,a Spain, Finland, France, Great Britain, Greece, Guatemala, Hungary, India, Ireland, Iceland,b Israel, Italy, Jordan, Japan,b South Korea, Sri Lanka, Luxembourg, Madagascar, Mexico, Malta, Mongolia, Mauritius, Malaysia, Nicaragua, Netherlands, Norway, Nepal, New Zealand, Pakistan, Panama, Philippines, Poland, Portugal, Paraguayb , Singapore, El Salvador, Sweden, Thailand, Trinidad and Tobago,Turkey, Uruguay, United States, Venezuela, South Africa. (Superscripts a and b indicate that data are from 1994 and 1993, respectively). Table B.2 presents the additional data on sub-national government tax revenues. The table includes information for countries for which no local government data was available from Government Finance Statistics. Tsub =Ttotal is sub-national government tax revenue in percent of total tax revenue.

34

COUNTRY Tsub =Ttotal BLZ COL CRI DOM ECU GRC GTM JOR JPN KOR LKA MDG MLT MUS MYS NPL PAK PAN SGP SLV TTO TUR VEN 0 19.90 2.6 0.9 5.0 1.2 3.7 0 24.9 19.0 3.6 3.0 0 2.1 15.6 3.0 3.0 2.5 0 3.0 4.19 12.4 3.2

Year 1994 1995 1995 1992 1997 1993 1997 1997 1986 1995 1995 1995

Source [1] [2] [3] [3] [4] [5] [3] [6] [5] [5] [3] [7] [1] [3] [3]

1990s [8] 1987 1994 1990 1995 1997 1989 [9] [3] [1] [4] [3] [5] [2]

Sources and notes: 35

[1 ] No local governments with tax assignment exist. Source: Government Finance Statistics 2000, Institutional tables. [2 ] Garman, Haggard and Willis (2001). [3 ] http://www1.worldbank.org/publicsector/decentralization/ scalindicators.htm. [4 ] Nickson (1995). [5 ] OECD Revenue Statistics 1999 edition, table 129. [6 ] No revenue collected by municipalities. Source: UNDP, http://magnet.undp.org/Docs/dec/monograph/FiscalAdmin&CSRJOR.htm [7 ] Brosio (2000). [8 ] Kelly (1998). [9 ] UNDP (1993).

36

B.2

The measurement of political variables

This section gives a description of the political accountability index used in the analysis. The source of both the political accountability index and the description of the variables is Kaufmann et al. (1999a, 1999b). 1. Orderly Political Transfers (EIU) 2. Transparency and fairness of the legal system (EIU) 3. Political rights (FH) 4. Civil liberties (FH) 5. Freedom of the Press (FH) 6. Military in Politics (PRS) 7. Democratic Accountability (PRS) 8. Information to Business about Government Policies (WDR) 9. Voice to Business to express concerns about policies (WDR) (1) and (2) are subcomponents of indices for Political Stability and Political E ectiveness, constructed by the for-pro t The Economic Intelligence Unit, which produces forecasts of political, economic and business environment in more than 180 countries. The assessments are based on regular contributions from a global network of more than 500 information gatherers. A panel of regional experts checks the accuracy, consistency and impartiality of these assessments. Political stability asks whether the political scene in free of internal and external threats to security, including if change in government takes place without social unrest (1). Political e ectiveness examines the quality of governance, including if the legal system is transparent and fair (2). http://www.eiu.com. (3), (4) and (5) are indices constructed by the not-for-pro t institution Freedom House (FH). Freedom House experts evaluate political rights, civil liberties and press freedom around the world, on a yearly basis. They de ne political rights as those freedoms that enable people to participate freely in the political process, civil liberties as the freedom to develop views, institutions, and personal autonomy apart from that of the state, and press freedom as freedom of the press from political (including legal) and economic in uence over media content. Additional information is available from http://www.freedomhouse.org. (6) and (7) are produced by the commercial Political Risk Services (PRS). The indicators are part of the PRS International Country Risk Guide which provides assessments of political, economic and nancial risks in developed and developing countries. The assessments are based on analysis of a worldwide network of experts, and is subject to peer-review at subject and regional levels. (6) asks if there is military participation (or threat thereof) in politics, as this reduces accountability since the military is not elected. 37

(7) quanti es the responsiveness of government, including popularity of government and the existence of free and fair elections. http://www.prsgroup.com. Finally, (8) and (9) is from a survey by the World Bank for the 1997 World Development Report of approximately 3000 enterprises in 69 countries. The survey was designed to measure the perception of rms regarding the constraints imposed on them by government. Available from http://www.worldbank.org/wbi/governance/ent_surveys.htm.

38

References
Alesina, Alberto, Reza Baqir and William Easterly. 1999. Public Goods and Ethnic Divisions. Quarterly Journal of Economics 114:1243 1284. Alesina, Alberto and Romain Wacziarg. 1998. Openness, Country Size and Government. Journal of Public Economics 69:305 321. Alt, James E., David Dreyer Lassen and David Skilling. 2001. Fiscal Transparency, Popularity, and the Scale of Government: Evidence from the States. Presented to "The State of the . States: State Politics and Policy Conference", Texas A&M University, College Station, March 2001. Banks, Je rey S. and Rangarajan K. Sundaram. 1998. Optimal Retention in Agency Problems.Journal of Economic Theory 82:293 323. Barro, Robert J. 1973. The Control of Politicians: An Economic Model. Public Choice 14:19 42. Barro, Robert J. 1999. Determinants of Democracy.Journal of Political Economy 107:S158 S183. Beck, Thorsten, George Clarke, Alberto Gro , Philip Keefer and Patrick Walsh. 2000. New Tools and New Tests in Comparative Political Economy: The Database of Political Institutions. World Bank Working Paper No. 2283, February. . Becker, Gary S. and Casey B. Mulligan. 1998. Deadweight Costs and the Size of Government. . NBER Working Paper 6789, November. Besley, Timothy and Robin Burgess. 2000. Does Media Make Government More Responsive? Theory and Evidence from Indian Famine Relief Policy. Manuscript, London School of . Economics, April. Besley, Timothy and Stephen Coate. 1997. Economic Model of Representative DemocAn racy.Quarterly Journal of Economics 112:85 114. Bollen, Kenneth A. 1980. Issues in the Comparative Measurement of Political Democracy. American Sociological Review 45:370 390. Bollen, Kenneth A. 1989. Structural Equations with Latent Variables. New York: John Wiley and Sons. Bollen, Kenneth A. 1993. Liberal Democracy: Validity and Method Factors in Cross-National Measures.American Journal of Political Science 37:1207 1230. Brosio, Giorgio. 2000. Decentralization in Africa. Manuscript, IMF, October. . Cameron, David R. 1978. The Expansion of the Public Economy: A Comparative Analysis. American Political Science Review 72:1243 1261. Cheibub, Jose A. 1998. Political Regimes and the Extractive Capacity of Governments: Taxation in Democracies and Dictatorships.World Politics 50:349 376.

39

Coppedge, Michael and Wolfgang H. Reinicke. 1990. Measuring Polyarchy.Studies in Comparative International Development 25:51 72. Crain, W. Mark. 1977. the Structure and Stability of Political Markets. Journal of On Political Economy 85(4):829 842. Dahl, Robert A. 1971. Polyarchy: Participation and Opposition. New Haven: Yale University Press. Dixit, Avinash. 1996. The Making of Economic Policy. Cambridge: MIT Press. Easterly, William and Ross Levine. 1997. Africas Growth Tragedy: Policies and Ethnic Divisions.Quarterly Journal of Economics 112:1203 1250. Easterly, William and Sergio Rebelo. 1993. Fiscal Policy and Economic Growth: An Empirical Investigation.Journal of Monetary Economics 32:417 458. Ferejohn, John. 1986. Incumbent Performance and Electoral Control.Public Choice 50:5 25. Ferejohn, John. 1990. Information and the Electoral Process. In Information and Democratic Processes, ed. John Ferejohn and James Kuklinski. Urbana: University of Illinois Press pp. 1 19. Ferejohn, John. 1999. Accountability and Authority: Towards a Theory of Political Accountability. In Democracy, Accountability, and Representation, ed. Przeworski, Stokes and Manin. Cambridge: Cambridge University Press pp. 131 153. Fjeldstad, Odd-Helge and Joseph Semboja. 2001. Why People Pay Taxes: The Case of the Development Levy in Tanzania.World Development 27(12). Friedman, Eric, Simon Johnson, Daniel Kaufman and Pablo Zoido-Lobat 2000. n. Dodging the Grabbing Hand: The Determinants of Uno cial Activity in 69 Countries.Journal of Public Economics 76:459 493. Galor, Oded and David N. Weil. 2000. Population, Technology, and Growth: From Malthusian Stagnation to the Demographic Transition and Beyond. American Economic Review 90:806 828. Garman, Christopher, Stephan Haggard and Eliza Willis. 2001. Fiscal Decentralization: A Political Theory with Latin American Cases.World Politics 53:205 236. Gordon, Roger H. and John D. Wilson. 2001. Expenditure Competition. NBER Working . Paper 8189, March. Government Finance Statistics Yearbook 2000. 2001. Washington, D.C.: International Monetary Fund. Greene, William H. 1997. Econometric Analysis, Third Edition. Upper Saddle River: PrenticeHall. Grossman, Gene M. and Elhanan Helpman. 1994. Protection for Sale.American Economic Review 84:833 850.

40

Hadi, Ali S. 1994. Modi cation of a Method for the Detection of Outliers in Multivariate A Data.Journal of the Royal Statistical Society, Series B 56:393 396. Hall, Robert E. and Charles I. Jones. 1999. Why Do some Countries Produce so Much More Output Per Worker Than Others? Quarterly Journal of Economics 114:83 116. Hansen, Lars Peter. 1982. Large Sample Properties of Generalized Method of Moments Estimators.Econometrica 50:1029 1054. Holsey, Cheryl M. and Thomas E. Borcherding. 1997. Why Does Governments Share of National Income Grow? An Assessment of the Recent Literature on the U.S. Experience. In Perspectives on Public Choice, ed. Dennis C. Mueller. Cambridge: Cambridge University Press pp. 562 589. Huntington, Samuel P. 1991. The Third Wave: Democratization in the Late Twentieth Century. Norman: University of Oklahoma Press. Jaggers, Keith and Ted R. Gurr. 1995. Polity III: Regime Change and Political Authority, 18001994 [Computer le]. Boulder, CO: Polity Data Archive, http://weber.ucsd.edu/ kgledits/Polity.html. Karatnycky, Adrian. 1996. The Comparative Survey of Freedom 1995-1996. In Freedom in the World 1995-1996. New York: Freedom House pp. 3 13. Kaufman, Daniel, Aart Kraay and Pablo Zoido-Lobat 1999a. n. Aggregating Governance Indicators. World Bank Discussion Paper No. 2195, November. http://www.worldbank.org. . Kaufman, Daniel, Aart Kraay and Pablo Zoido-Lobat 1999b. n. Governance Matters. World . Bank Discussion Paper No. 2196, November. http://www.worldbank.org. Kelly, Roy. 1998. Intergovernmental Revenue Allocation Theory and Practice: Application to Nepal. DDP No. 624, Harvard Institute for International Development, February. . Lane, Jan-Erik and Svante Ersson. 2000. The New Institutional Politics: Performance and Outcomes. New York: Routledge. LaPorta, Rafael, Florencio Lopez-de-Silanes, Andrei Shleifer and Robert Vishny. 1999. The Quality of Government.Journal of Law, Economics and Organization 15:222 279. Lewis-Beck, Michael S. 1988. Economics and Elections. Ann Arbor: University of Michigan Press. Lijphart, Arend. 1999. Patterns of Democracy. New Haven: Yale University Press. Linz, Juan J. 1994. Presidential or Parliamentary Democracy: Does It Make a Di erence? In The Failure of Presidential Democracy. Vol. 1: Comparative Perspectives, ed. Linz and Valenzuela. Baltimore: Johns Hopkins University Press pp. 3 87. Lipset, Seymour M. 1959. Some Social Requisites for Democracy: Economic Development and Political Legitimacy.American Political Science Review 53:69 105.

41

Lipset, Seymour M. 1994. The Social Requisites of Democracy Revisited.American Sociological Review 59:1 22. Nickson, R. Andrew. 1995. Local Government in Latin America. Boulder: Lynne Rienner Publishers. North, Douglass C. 1985. The Growth of Government: An Economic Historians Perspective. Journal of Public Economics 28:383 399. Oates, Wallace E. 1985. Searching for Leviathan: An Empirical Study.American Economic Review 75:748 757. OECD. 1999. Revenue Statistics, 1965/1998. Paris: OECD. Persson, Torsten, G rard Roland and Guido Tabellini. 1997. Separation of Powers and Political Accountability.Quarterly Journal of Economics 112:1163 1202. Persson, Torsten, G rard Roland and Guido Tabellini. 2000. Comparative Politics and Public Finance.Journal of Political Economy 108:1121 1161. Persson, Torsten and Guido Tabellini. 1999. The Size and Scope of Government: Comparative Politics with Rational Politicians.European Economic Review 43:699 735. Przeworski, Adam and Fernando Limongi. 1993. Political Regimes and Economic Growth. Journal of Economic Perspectives 7 (Summer):51 69. Rodrik, Dani. 1998. Why Do More Open Economies Have Bigger Governments? Journal of Political Economy 106:997 1032. Rose-Ackerman, Susan. 1999. Corruption and Government. Cambridge: Cambridge University Press. Staiger, Douglas and James H. Stock. 1997. Instrumental Variables Regression with Weak Instruments.Econometrica 65:557 586. Tullock, Gordon. 1965. Entry Barriers Into Politics.American Economic Review 55:458 466. UNDP. 1993. Human Development Report 1993. New York: United Nations. Vanhanen, Tatu. 1990. The Process of Democratization: A Comparative Study of 147 States, 1980-88. New York: Taylor and Francis. World Development Indicators CD-ROM, 1975-1997. 1999. Washington, D.C.: World Bank.

42

Table 1. Political accountability and the size of general government. N = 62 Tax revenue of general government / GDP OLS P OLACCT log (IN C)
(1.843)

TSLS
(3.384)

OLS
(1.739)

TSLS 17.360
(6.157)

11.016 21.534 5.955 1.405

(1.790) (1.959)

3.417
(2.880)

(1.481)

.312

2.638
(2.689)

log (OP EN ) 3.494x DEP EN D < 15 > 65 U RBAN log (P OP ) R2 F (1st ST) p(OVER)
(.047) (11.047)

5.499
(2.092)

4.551
(1.843)

5.404
(1.864)

1.025

13.438
(12.910)

33.924
(23.302)

43.106
(29.171)

168.069
(38.858)

92.721
(63.370) (.059)

.105
(.621)

(.057)

.152
(.961)

(.040)

.077 x
(.512)

.130

1.387 .71

2.674

1.134 .81

2.246
(1.023)

25.62 .620

25.62 .735

Estimated using Stata 7.0. A constant term (not reported) was included in all regressions. Robust standard errors are in parentheses, ***, **, * and denote signi cance at the .1%, 1%, 5% and 10 % levels, respectively. F (1st ST) is the F -statistic from the rst stage of the TSLS regression. p(OVER) is the p-value associated with the null hypothesis of no overidenti cation .

Table 2. Joint estimation of the size of government and political accountability. N = 62 Dep. variable P OLACCT SIZEGOV log (IN C) log (OP EN ) DEP EN D < 15 > 65 URBAN log (P OP ) LAT IT UDE SUBN AT IONAL P ROT EST AN T \R2 "
2 (.053)

System 1 SIZEGOV 21.050

(3.789)

System 2 SIZEGOV 13.558

(5.984)

P OLACCT

P OLACCT

.000
(.028)

(.014)

.025 x

3.746
(2.585)

(.133)

.283

1.965
(2.180)

(.091)

.179

4.430
(2.218)

2.949
(1.404)

.102
(9.354) (17.327)

7.801

89.517x
(50.845)

.085

(.043)

.028 .445

1.389x
(.754)

(.657)

1.336
(.849) (.156)

(.410)

.581 .124 .003

.211 .003

(.106)

(.002)

(.002)

.51 107.7

.64 114.5

.70 174.3

.69 134.0

Estimated by three-stage least squares using Stata 7.0. A constant term (not reported) was included in all regressions. Standard errors are in parentheses, ***, **, * and denote signi cance at the .1%, 1%, 5% and 10 % levels, respectively.

Table 3. The size of government, civil liberties and political rights. N = 62 OLS CL9095 P R9095 log(INC) 2.204
(2.021) (1.894) (1.245)

Tax revenue of general government / GDP 2SLS

(2.414)

OLS
(1.223)

2SLS 9.217
(4.066)

OLS

2SLS

OLS

2SLS

5.491 12.962 2.304x

5.321 15.536 1.836

(1.616) (3.187) (1.397)

10.397x
(5.724)

4.570
(3.078) (2.694)

(1.803)

.713

3.365
(2.872) (2.433)

3.535
(2.359) (2.241)

3.445
(3.255) (4.048)

1.325
(1.899) (2.243)

2.649
(2.817)

log(OP EN ) 5.198 DEP END < 15 > 65 URBAN log(P OP ) R2 F (1st ST) p(OVER)
(.053)

10.011 5.289
(1.989) (16.683)

8.184 7.186 18.023 5.824

(13.340)

13.601
(5.371)

1.348
(12.555)

7.613

2.256

23.149
(19.443)

32.632
(24.734) (38.528)

22.042
(32.399) (66.098)

37.561
(26.118) (39.324)

51.237
(38.523) (65.264)

186.132 90.937 .074

(.070)

199.207 127.777
(.062)

.129 x

(.043)

.056

(.061)

.099

.072

(.084)

.146 x
(1.550)

(.044)

.053

(.071)

.115

2.233
(.762)

5.087
(1.249)

1.372
(.658)

3.779
(1.583)

2.190 6.071
(.733)

1.222x
(.619)

4.279x
(2.229)

.64 30.34 .167

.79 30.34 .197

.61 25.27 .238

.78 25.27 .239

Estimated using Stata 7.0. A constant term (not reported) was included in all regressions. Robust standard errors are in parentheses, ***, **, * and denote signi cance at the .1%, 1%, 5% and 10 % levels, respectively. F (1st ST) is the F -statistic from the rst stage of the TSLS regression. p(OVER) is the p-value associated with the null hypothesis of no overidenti cation .

Table 4. Robustness: Civil liberties, political rights, and alternative information sources. N = 58 Tax revenue of general government / GDP OLS CL9095 P R9095 ALT IN F CL9095 ALT IN F P R9095 ALT INF R2 .69 .74 .68 3.013
(2.343) (3.601) (1.351)

OLS
(2.049)

OLS

OLS

4.080 10.397 4.065 13.853

(1.636) (2.757)

16.193 2.896
(2.243)

18.013
(3.437)

4.492
(1.156)

6.003
(1.279)

.76

Estimated using Stata 7.0. A constant term (not reported) was included in all regressions. Robust standard errors are in parentheses, ***, **, * and denote signi cance at the .1%, 1%, 5% and 10 % levels, respectively. Controls included in all regressions but not shown are: log(open), log(income), log(population), urbanization, dependency ratio.

Table 5. Robustness: Additional explanatory variables. N = 62 OLS P OLACCT P RESIDENT IAL ELF T AX EF F ICIEN CY R2 F (1st ST) p(OVER) .71 25.62 .621 .72 25.62 .769
(1.895)

Tax revenue of general government / GDP 2SLS

(3.288)

OLS
(1.781)

2SLS
(3.217)

OLS
(1.838)

2SLS
(3.331)

10.638 21.794 10.957 20.691 9.852 20.813 2.319

(1.959) (2.904)

.346

5.275x
(3.151)

5.642
(3.612)

.125
(.085)

(0.151)

.011

.71 25.62 .584

Estimated using Stata 7.0. A constant term (not reported) was included in all regressions. Robust standard errors are in parentheses, ***, **, * and denote signi cance at the .1%, 1%, 5% and 10 % levels, respectively. F (1st ST) is the F -statistic from the rst stage of the TSLS regression. p(OVER) is the p-value associated with the null hypothesis of overino denti cation Controls included in all regressions but not shown are: log(open), log(income), . log(population), urbanization, dependency ratio.

Table 6. Robustness: Instruments. N = 62 Tax revenue of general government / GDP 2SLS P OLACCT LAT IT UDE SUBN AT IONAL P ROT EST AN T S F (1st ST) p(OVER) 25.62 .620 22.25 .950 27.05 .468
(3.384)

2SLS
(5.747)

2SLS
(4.100)

2SLS
(4.652)

21.534 15.802 22.728 23.118 11.000

(9.322)

1.931
(3.160)

.027
(.047)

34.56 .467

Estimated using Stata 7.0. A constant term (not reported) was included in all regressions. Robust standard errors are in parentheses, ***, **, * and denote signi cance at the .1%, 1%, 5% and 10 % levels, respectively. F (1st ST) is the F -statistic from the rst stage of the TSLS regression. p(OVER) is the p-value associated with the null hypothesis of overino denti cation Controls included in all regressions but not shown are: log(open), log(income), . log(population), urbanization, dependency ratio.

Figure 1: The structure of a period

Voters choose b by majority vote

Politician observes S

Politician Payoffs are Elections chooses r realized

Figure 2: Political accountability and the degree of tax revenue centralization Centralization 1 .9 .8 .7
IND TUR MYS COL MEX BOL MNG KOR JPN GTM PAK PRY DOM SGP JOR SLV NPL VEN MDG LKA THA ECU NIC BWA PAN PHL CHL URY BGR BLZ MLT GRC MUS IRL CRI PRT GBR NLD TTO BEL ZAFISR LUX NZL HUNITA POL FRA ESP ISL AUT NOR AUS DEU FIN USA ARG DEN SWE

.6 .5 .4 .3 -1 -.5 0

CHE

CAN

BRA

.5

1.5

Political accountability

Figure 3: Political accountability and the size of general government

Tax revenue of general government / GDP 50

FRA SWE DEN BEL LUX NLD ITA AUT DEU NOR FIN CHE

40

POL

ISR HUN CAN ESP GBR NZL IRL ISL PRT

30
NIC MYS JOR

BGR BRA

URY ZAF MNGTTO KOR JPN GRC BLZ MUS

MLT USA AUS

CRI

20
TUR PAK

ARG CHL LKA COL MEX SGP DOM SLV THA ECU VEN BOL PANBWA IND PHL

10 -1

GTM

PRY

NPL

MDG

-.5

0 .5 1 Political accountability

1.5

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 -----------------------------------------------------------------------------log: c:\account\data\main3sup.log log type: text opened on: 6 Nov 2001, 11:19:53 /* Additional regression results provided for referees. David Dreyer Lassen "Political accountability and the size of government: theory and cross-country evidence" */

/* logistic transformation */ /* table 1 */ reg loggtax polacct lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 29.72 0.0000 0.7401 .33235

-----------------------------------------------------------------------------| Robust loggtax | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | .6092419 .1036905 5.88 0.000 .4014416 .8170422 lnopen | .1951172 .1074726 1.82 0.075 -.0202627 .410497 lninc | .0392416 .0959452 0.41 0.684 -.153037 .2315201 depend | -.2254321 .6423763 -0.35 0.727 -1.512783 1.061919 urban | .0065546 .0025653 2.56 0.013 .0014136 .0116957 logpop95 | .0744518 .0318109 2.34 0.023 .0107015 .1382022 _cons | -4.275577 1.509038 -2.83 0.006 -7.299757 -1.251397 -----------------------------------------------------------------------------. ivgmm0 loggtax (polacct = lat federal prot) lnopen lninc urban depend logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 0.3757 Hansen J = 1.1114 Chi-sq( 2) P-val = 0.57366 -----------------------------------------------------------------------------| GMM loggtax | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 1.054551 .1682547 6.27 0.000 .724778 1.384324 lnopen | .2875832 .1055046 2.73 0.006 .0807979 .4943684 lninc | -.1714796 .1375773 -1.25 0.213 -.4411261 .0981669 urban | .0084589 .002713 3.12 0.002 .0031416 .0137763 depend | .3297325 .6413822 0.51 0.607 -.9273535 1.586818 logpop95 | .1312855 .0465774 2.82 0.005 .0399956 .2225755 _cons | -4.500449 1.682975 -2.67 0.007 -7.79902 -1.201878 -----------------------------------------------------------------------------Instrumented: polacct Instruments: lnopen lninc urban depend logpop95 lat federal prot -----------------------------------------------------------------------------. reg loggtax polacct lnopen lninc oldp youngp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 54) = Prob > F = R-squared = Root MSE = 62 38.09 0.0000 0.8041 .29121

-----------------------------------------------------------------------------| Robust loggtax | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | .3879475 .1086251 3.57 0.001 .1701673 .6057276 lnopen | .241566 .1036238 2.33 0.024 .0338127 .4493193 lninc | -.0048448 .0926637 -0.05 0.958 -.1906242 .1809346 oldp | 6.726689 2.458615 2.74 0.008 1.797462 11.65592 youngp | .8432878 1.484143 0.57 0.572 -2.13224 3.818816 urban | .0052936 .0023993 2.21 0.032 .0004832 .010104 logpop95 | .0638564 .0278578 2.29 0.026 .0080049 .119708 _cons | -4.600987 1.497147 -3.07 0.003 -7.602585 -1.599388 -----------------------------------------------------------------------------. ivgmm0 loggtax (polacct = lat federal prot) lnopen lninc oldp youngp urban logpop95; Instrumental Variables Estimation via GMM Number of obs Root MSE Page 1 of 12 = = 62 0.3229

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 Hansen J = 0.6533 Chi-sq( 2) P-val = 0.72132 -----------------------------------------------------------------------------| GMM loggtax | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | .8325471 .2986052 2.79 0.005 .2472916 1.417803 lnopen | .2841953 .0970435 2.93 0.003 .0939936 .4743971 lninc | -.1295337 .1328154 -0.98 0.329 -.3898471 .1307798 oldp | 3.985511 3.229742 1.23 0.217 -2.344667 10.31569 youngp | 1.304317 1.492278 0.87 0.382 -1.620493 4.229128 urban | .0074504 .002809 2.65 0.008 .0019449 .0129559 logpop95 | .1086083 .0487484 2.23 0.026 .0130631 .2041534 _cons | -4.768983 1.659095 -2.87 0.004 -8.020748 -1.517217 -----------------------------------------------------------------------------Instrumented: polacct Instruments: lnopen lninc oldp youngp urban logpop95 lat federal prot -----------------------------------------------------------------------------. /* table 2 */ > reg3 (loggtax polacct lnopen lninc urban depend logpop95) (polacct lat lninc prot federal loggtax); Three-stage least squares regression ---------------------------------------------------------------------Equation Obs Parms RMSE "R-sq" chi2 P ---------------------------------------------------------------------loggtax 62 6 .3790079 0.6189 132.1534 0.0000 polacct 62 5 .3940193 0.6621 123.2066 0.0000 --------------------------------------------------------------------------------------------------------------------------------------------------| Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------loggtax | polacct | 1.026477 .1832218 5.60 0.000 .667369 1.385585 lnopen | .2171018 .102482 2.12 0.034 .0162409 .4179628 lninc | -.1701533 .1289395 -1.32 0.187 -.42287 .0825634 urban | .0050602 .0026789 1.89 0.059 -.0001903 .0103107 depend | -.3110489 .487434 -0.64 0.523 -1.266402 .6443042 logpop95 | .0655786 .0364416 1.80 0.072 -.0058457 .1370029 _cons | -2.51868 1.245373 -2.02 0.043 -4.959565 -.0777945 -------------+---------------------------------------------------------------polacct | lat | 1.199443 .7021492 1.71 0.088 -.1767443 2.57563 lninc | .2586649 .135843 1.90 0.057 -.0075824 .5249122 prot | .0029867 .0018812 1.59 0.112 -.0007004 .0066738 federal | .1918196 .1358523 1.41 0.158 -.0744459 .4580852 loggtax | .093999 .4475864 0.21 0.834 -.7832542 .9712523 _cons | -1.996318 1.912562 -1.04 0.297 -5.744871 1.752235 -----------------------------------------------------------------------------Endogenous variables: loggtax polacct Exogenous variables: lnopen lninc urban depend logpop95 lat prot federal -----------------------------------------------------------------------------. reg3 (loggtax polacct lnopen lninc urban depend logpop95) (polacct lat lninc prot federal gtax95c); Three-stage least squares regression ---------------------------------------------------------------------Equation Obs Parms RMSE "R-sq" chi2 P ---------------------------------------------------------------------loggtax 62 6 .4704686 0.4128 138.3688 0.0000 polacct 62 5 .3928272 0.6641 147.8159 0.0000 --------------------------------------------------------------------------------------------------------------------------------------------------| Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------loggtax | polacct | 1.284632 .1866612 6.88 0.000 .9187828 1.650481 lnopen | .0883354 .0831881 1.06 0.288 -.0747103 .251381 lninc | -.2560027 .128704 -1.99 0.047 -.5082579 -.0037474 urban | .0033698 .0024091 1.40 0.162 -.0013519 .0080916 depend | -.1457946 .4145404 -0.35 0.725 -.9582788 .6666896 logpop95 | .0386454 .0338662 1.14 0.254 -.0277311 .1050219 _cons | -.9637522 1.180083 -0.82 0.414 -3.276672 1.349167 -------------+---------------------------------------------------------------polacct | lat | -.0372484 .2325858 -0.16 0.873 -.4931081 .4186114 Page 2 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 lninc | .178234 .077584 2.30 0.022 .026172 .3302959 prot | .0005046 .0013058 0.39 0.699 -.0020548 .003064 federal | -.0101157 .0746208 -0.14 0.892 -.1563698 .1361385 gtax95c | .0400774 .0060449 6.63 0.000 .0282296 .0519253 _cons | -1.852759 .601589 -3.08 0.002 -3.031852 -.6736661 -----------------------------------------------------------------------------Endogenous variables: loggtax polacct Exogenous variables: lnopen lninc urban depend logpop95 lat prot federal gtax95c -----------------------------------------------------------------------------. reg3 (loggtax polacct lnopen lninc urban youngp oldp logpop95) (polacct lat lninc prot federal loggtax); Three-stage least squares regression ---------------------------------------------------------------------Equation Obs Parms RMSE "R-sq" chi2 P ---------------------------------------------------------------------loggtax 62 7 .306148 0.7514 199.4694 0.0000 polacct 62 5 .3671748 0.7066 141.7899 0.0000 --------------------------------------------------------------------------------------------------------------------------------------------------| Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------loggtax | polacct | .631959 .2927287 2.16 0.031 .0582213 1.205697 lnopen | .1610573 .0731214 2.20 0.028 .0177419 .3043726 lninc | -.0687797 .1103962 -0.62 0.533 -.2851523 .1475928 urban | .0021807 .0022961 0.95 0.342 -.0023196 .006681 youngp | -.1607529 .9535808 -0.17 0.866 -2.029737 1.708231 oldp | 4.197482 2.470198 1.70 0.089 -.644018 9.038982 logpop95 | .0225707 .0344515 0.66 0.512 -.0449531 .0900944 _cons | -2.502211 1.051985 -2.38 0.017 -4.564064 -.4403573 -------------+---------------------------------------------------------------polacct | lat | .6041235 .4032011 1.50 0.134 -.1861361 1.394383 lninc | .1548173 .0968041 1.60 0.110 -.0349153 .3445499 prot | .0029891 .0017411 1.72 0.086 -.0004233 .0064015 federal | .1322485 .1045131 1.27 0.206 -.0725933 .3370903 loggtax | .470936 .2606931 1.81 0.071 -.0400131 .9818851 _cons | -.3999498 1.198852 -0.33 0.739 -2.749657 1.949758 -----------------------------------------------------------------------------Endogenous variables: loggtax polacct Exogenous variables: lnopen lninc urban youngp oldp logpop95 lat prot federal -----------------------------------------------------------------------------. reg3 (loggtax polacct lnopen lninc urban youngp oldp logpop95) (polacct lat lninc prot federal gtax95c); Three-stage least squares regression ---------------------------------------------------------------------Equation Obs Parms RMSE "R-sq" chi2 P ---------------------------------------------------------------------loggtax 62 7 .4996525 0.3377 113.1646 0.0000 polacct 62 5 .3903149 0.6684 146.9464 0.0000 --------------------------------------------------------------------------------------------------------------------------------------------------| Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------loggtax | polacct | 1.356002 .2975938 4.56 0.000 .7727291 1.939275 lnopen | .0567712 .0944282 0.60 0.548 -.1283046 .2418471 lninc | -.255376 .1509757 -1.69 0.091 -.551283 .0405309 urban | .0026427 .002919 0.91 0.365 -.0030784 .0083637 youngp | -.7280829 1.281891 -0.57 0.570 -3.240543 1.784377 oldp | -1.740749 2.471625 -0.70 0.481 -6.585045 3.103547 logpop95 | .0271831 .0428141 0.63 0.525 -.0567311 .1110973 _cons | -.3899163 1.465286 -0.27 0.790 -3.261825 2.481992 -------------+---------------------------------------------------------------polacct | lat | .0407367 .2322856 0.18 0.861 -.4145346 .496008 lninc | .1773358 .0777653 2.28 0.023 .0249186 .3297529 prot | .0005538 .0013193 0.42 0.675 -.002032 .0031396 federal | .0061361 .0730856 0.08 0.933 -.1371091 .1493813 gtax95c | .0385835 .006041 6.39 0.000 .0267434 .0504237 _cons | -1.842129 .603564 -3.05 0.002 -3.025092 -.659165 -----------------------------------------------------------------------------Page 3 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 Endogenous variables: loggtax polacct Exogenous variables: lnopen lninc urban youngp oldp logpop95 lat prot federal gtax95c -----------------------------------------------------------------------------. /* table 3 */ > reg loggtax cl9095m lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 18.49 0.0000 0.6628 .37851

-----------------------------------------------------------------------------| Robust loggtax | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | .2897177 .0696027 4.16 0.000 .1502306 .4292047 lnopen | .2798274 .1067584 2.62 0.011 .0658788 .493776 lninc | .0947211 .114577 0.83 0.412 -.1348964 .3243386 depend | -.3811257 .7234305 -0.53 0.600 -1.830913 1.068661 urban | .0047163 .0030261 1.56 0.125 -.0013481 .0107807 logpop95 | .1157839 .0394444 2.94 0.005 .0367355 .1948323 _cons | -4.390823 1.729145 -2.54 0.014 -7.856107 -.925539 -----------------------------------------------------------------------------. ivgmm0 loggtax (cl9095m = lat federal ) lnopen lninc depend urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 0.4495 Hansen J = 1.7910 Chi-sq( 1) P-val = 0.18080 -----------------------------------------------------------------------------| GMM loggtax | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | .6476909 .1256905 5.15 0.000 .401342 .8940398 lnopen | .502976 .143938 3.49 0.000 .2208627 .7850893 lninc | -.2296064 .1553305 -1.48 0.139 -.5340486 .0748358 depend | .0738939 .8784033 0.08 0.933 -1.647745 1.795533 urban | .007447 .003628 2.05 0.040 .0003362 .0145577 logpop95 | .2537993 .0644578 3.94 0.000 .1274643 .3801343 _cons | -4.207947 2.201893 -1.91 0.056 -8.523578 .1076835 -----------------------------------------------------------------------------Instrumented: cl9095m Instruments: lnopen lninc depend urban logpop95 lat federal -----------------------------------------------------------------------------. reg loggtax pr9095m lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 16.58 0.0000 0.6336 .39458

-----------------------------------------------------------------------------| Robust loggtax | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | .2752095 .0955083 2.88 0.006 .0838067 .4666123 lnopen | .378582 .1283846 2.95 0.005 .1212936 .6358704 lninc | .1681879 .1369378 1.23 0.225 -.1062417 .4426174 depend | -.2047364 .7674792 -0.27 0.791 -1.742799 1.333326 urban | .0046121 .0034713 1.33 0.189 -.0023444 .0115687 logpop95 | .1112739 .0397076 2.80 0.007 .0316981 .1908497 _cons | -5.644703 1.715008 -3.29 0.002 -9.081656 -2.207749 -----------------------------------------------------------------------------. ivgmm0 loggtax (pr9095m = lat federal ) lnopen lninc depend urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 0.5087 Hansen J = 1.3071 Chi-sq( 1) P-val = 0.25292 -----------------------------------------------------------------------------| GMM loggtax | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------Page 4 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 pr9095m | .7712741 .1660475 4.64 0.000 .445827 1.096721 lnopen | .9016074 .2133162 4.23 0.000 .4835153 1.319699 lninc | -.1742316 .1688554 -1.03 0.302 -.5051821 .1567188 depend | .8158692 1.035716 0.79 0.431 -1.214096 2.845835 urban | .0084362 .0043098 1.96 0.050 -.0000107 .0168832 logpop95 | .301074 .078989 3.81 0.000 .1462583 .4558896 _cons | -7.689965 2.688201 -2.86 0.004 -12.95874 -2.421189 -----------------------------------------------------------------------------Instrumented: pr9095m Instruments: lnopen lninc depend urban logpop95 lat federal -----------------------------------------------------------------------------. reg loggtax cl9095m lnopen lninc oldp youngp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 54) = Prob > F = R-squared = Root MSE = 62 29.21 0.0000 0.7686 .31645

-----------------------------------------------------------------------------| Robust loggtax | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | .1397377 .0736235 1.90 0.063 -.0078686 .2873439 lnopen | .2842837 .1119489 2.54 0.014 .0598397 .5087277 lninc | .0266977 .1152294 0.23 0.818 -.2043234 .2577188 oldp | 8.019462 2.487251 3.22 0.002 3.032824 13.0061 youngp | .7544922 1.577013 0.48 0.634 -2.407229 3.916214 urban | .0038864 .0026562 1.46 0.149 -.001439 .0092118 logpop95 | .0756097 .035367 2.14 0.037 .0047031 .1465162 _cons | -4.597475 1.632132 -2.82 0.007 -7.869703 -1.325248 -----------------------------------------------------------------------------. ivgmm0 loggtax (cl9095m = lat federal ) lnopen lninc oldp youngp urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 0.3605 Hansen J = 1.3779 Chi-sq( 1) P-val = 0.24046 -----------------------------------------------------------------------------| GMM loggtax | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | .4472819 .2231283 2.00 0.045 .0099585 .8846053 lnopen | .4242343 .1312391 3.23 0.001 .1670104 .6814582 lninc | -.1728851 .1527881 -1.13 0.258 -.4723443 .1265742 oldp | 4.110565 3.650363 1.13 0.260 -3.044016 11.26515 youngp | .4212637 1.755718 0.24 0.810 -3.01988 3.862407 urban | .0060987 .0032442 1.88 0.060 -.0002599 .0124572 logpop95 | .1860022 .0850467 2.19 0.029 .0193137 .3526908 _cons | -4.131047 1.935712 -2.13 0.033 -7.924973 -.3371204 -----------------------------------------------------------------------------Instrumented: cl9095m Instruments: lnopen lninc oldp youngp urban logpop95 lat federal -----------------------------------------------------------------------------. reg loggtax pr9095m lnopen lninc oldp youngp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 54) = Prob > F = R-squared = Root MSE = 62 31.19 0.0000 0.7581 .32358

-----------------------------------------------------------------------------| Robust loggtax | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | .1082678 .0901505 1.20 0.235 -.0724732 .2890087 lnopen | .3139261 .1301985 2.41 0.019 .0528938 .5749584 lninc | .0648307 .120448 0.54 0.593 -.1766529 .3063144 oldp | 8.833261 2.545837 3.47 0.001 3.729166 13.93736 youngp | 1.043516 1.636772 0.64 0.526 -2.238015 4.325047 urban | .0036854 .0027661 1.33 0.188 -.0018602 .009231 logpop95 | .0653858 .0355232 1.84 0.071 -.0058339 .1366055 _cons | -5.161968 1.579554 -3.27 0.002 -8.328782 -1.995153 -----------------------------------------------------------------------------. ivgmm0 loggtax (pr9095m = lat federal ) lnopen lninc oldp youngp urban logpop95; Page 5 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 0.3918 Hansen J = 1.1797 Chi-sq( 1) P-val = 0.27742 -----------------------------------------------------------------------------| GMM loggtax | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | .5094298 .3073351 1.66 0.097 -.092936 1.111796 lnopen | .6848113 .2869573 2.39 0.017 .1223853 1.247237 lninc | -.1205956 .1556757 -0.77 0.439 -.4257145 .1845232 oldp | 5.731225 3.619624 1.58 0.113 -1.363107 12.82556 youngp | 1.835567 2.139132 0.86 0.391 -2.357054 6.028189 urban | .0067731 .0038209 1.77 0.076 -.0007157 .0142618 logpop95 | .2105401 .1191422 1.77 0.077 -.0229743 .4440544 _cons | -6.7501 2.712209 -2.49 0.013 -12.06593 -1.434267 -----------------------------------------------------------------------------Instrumented: pr9095m Instruments: lnopen lninc oldp youngp urban logpop95 lat federal -----------------------------------------------------------------------------. /* central vs. general government */ > reg ctax95 polacct lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 17.59 0.0000 0.6114 6.2931

-----------------------------------------------------------------------------| Robust ctax95 | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 8.43328 1.735308 4.86 0.000 4.955644 11.91092 lnopen | 5.272744 1.906744 2.77 0.008 1.451544 9.093943 lninc | .569553 1.930061 0.30 0.769 -3.298376 4.437482 depend | -.7200438 9.913941 -0.07 0.942 -20.58802 19.14794 urban | .0817325 .0469181 1.74 0.087 -.0122934 .1757585 logpop95 | .8258836 .6304457 1.31 0.196 -.4375578 2.089325 _cons | -29.73964 27.39156 -1.09 0.282 -84.63354 25.15426 -----------------------------------------------------------------------------. ivgmm0 ctax95 (polacct = lat federal prot) lnopen lninc depend urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 6.3845 Hansen J = 8.6263 Chi-sq( 2) P-val = 0.01339 -----------------------------------------------------------------------------| GMM ctax95 | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 13.0868 2.624057 4.99 0.000 7.943742 18.22986 lnopen | 6.133602 1.609403 3.81 0.000 2.979231 9.287974 lninc | -1.946891 2.559707 -0.76 0.447 -6.963825 3.070042 depend | 3.23876 9.647052 0.34 0.737 -15.66912 22.14663 urban | .082661 .0468614 1.76 0.078 -.0091857 .1745077 logpop95 | 1.523614 .7636478 2.00 0.046 .0268914 3.020336 _cons | -28.78261 27.05398 -1.06 0.287 -81.80743 24.24221 -----------------------------------------------------------------------------Instrumented: polacct Instruments: lnopen lninc depend urban logpop95 lat federal prot -----------------------------------------------------------------------------. reg ctax95 polacct lnopen lninc youngp oldp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 54) = Prob > F = R-squared = Root MSE = 62 29.03 0.0000 0.7330 5.2642

-----------------------------------------------------------------------------| Robust ctax95 | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 3.696012 1.666962 2.22 0.031 .3539541 7.03807 lnopen | 6.255547 1.507488 4.15 0.000 3.233216 9.277878 Page 6 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 lninc | -.4500245 1.704588 -0.26 0.793 -3.867518 2.967469 youngp | 27.19091 20.64595 1.32 0.193 -14.20173 68.58356 oldp | 152.6583 28.52074 5.35 0.000 95.47771 209.839 urban | .0547967 .0391156 1.40 0.167 -.0236253 .1332187 logpop95 | .5883933 .5048879 1.17 0.249 -.423846 1.600633 _cons | -36.68228 24.19451 -1.52 0.135 -85.18934 11.82479 -----------------------------------------------------------------------------. ivgmm0 ctax95 (polacct = lat federal prot) lnopen lninc youngp oldp urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 4.9979 Hansen J = 6.5640 Chi-sq( 2) P-val = 0.03755 -----------------------------------------------------------------------------| GMM ctax95 | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 2.806564 4.808287 0.58 0.559 -6.617506 12.23063 lnopen | 6.739064 1.509398 4.46 0.000 3.780699 9.697429 lninc | .0943642 2.087671 0.05 0.964 -3.997395 4.186123 youngp | 25.07461 18.18489 1.38 0.168 -10.56713 60.71634 oldp | 144.1881 44.05067 3.27 0.001 57.8504 230.5259 urban | .0501254 .0431432 1.16 0.245 -.0344338 .1346845 logpop95 | .5376107 .7008818 0.77 0.443 -.8360924 1.911314 _cons | -40.74165 22.29972 -1.83 0.068 -84.4483 2.965004 -----------------------------------------------------------------------------Instrumented: polacct Instruments: lnopen lninc youngp oldp urban logpop95 lat federal prot -----------------------------------------------------------------------------. reg ctax95 cl9095m lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 12.72 0.0000 0.5606 6.6918

-----------------------------------------------------------------------------| Robust ctax95 | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | 4.225508 1.286374 3.28 0.002 1.647556 6.80346 lnopen | 6.592135 1.842371 3.58 0.001 2.899942 10.28433 lninc | 1.163257 1.98584 0.59 0.560 -2.816455 5.14297 depend | -2.497949 11.18496 -0.22 0.824 -24.9131 19.9172 urban | .0575614 .0484133 1.19 0.240 -.0394611 .1545838 logpop95 | 1.482453 .8306134 1.78 0.080 -.1821333 3.14704 _cons | -31.53362 30.1649 -1.05 0.300 -91.98542 28.91818 -----------------------------------------------------------------------------. ivgmm0 ctax95 (cl9095m = lat federal ) lnopen lninc depend urban logpop95;

Instrumental Variables Estimation via GMM

Number of obs = 62 Root MSE = 8.0209 Hansen J = 3.3323 Chi-sq( 1) P-val = 0.06793 -----------------------------------------------------------------------------| GMM ctax95 | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | 10.69346 2.26395 4.72 0.000 6.256201 15.13072 lnopen | 10.57989 2.3286 4.54 0.000 6.015915 15.14386 lninc | -4.852179 3.0154 -1.61 0.108 -10.76225 1.057897 depend | 2.911416 14.70911 0.20 0.843 -25.91791 31.74074 urban | .0960327 .0614242 1.56 0.118 -.0243564 .2164219 logpop95 | 3.733897 1.220417 3.06 0.002 1.341924 6.12587 _cons | -20.70647 37.47701 -0.55 0.581 -94.16007 52.74712 -----------------------------------------------------------------------------Instrumented: cl9095m Instruments: lnopen lninc depend urban logpop95 lat federal -----------------------------------------------------------------------------. reg ctax95 cl9095m lnopen lninc youngp oldp urban logpop95, robust; Number of obs = F( 7, 54) = Prob > F = R-squared = Page 7 of 12 62 27.71 0.0000 0.7200

Regression with robust standard errors

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 Root MSE = 5.3914

-----------------------------------------------------------------------------| Robust ctax95 | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | 1.35898 1.26551 1.07 0.288 -1.178214 3.896175 lnopen | 6.676668 1.585067 4.21 0.000 3.498801 9.854536 lninc | -.1639622 1.946112 -0.08 0.933 -4.065681 3.737757 youngp | 26.35729 21.5369 1.22 0.226 -16.8216 69.53618 oldp | 164.6654 28.3221 5.81 0.000 107.883 221.4478 urban | .041552 .0406982 1.02 0.312 -.0400429 .1231469 logpop95 | .7103621 .6821092 1.04 0.302 -.6571845 2.077909 _cons | -36.65954 25.58146 -1.43 0.158 -87.94728 14.6282 -----------------------------------------------------------------------------. ivgmm0 ctax95 (cl9095m = lat federal ) lnopen lninc youngp oldp urban logpop95;

Instrumental Variables Estimation via GMM

Number of obs = 62 Root MSE = 5.5913 Hansen J = 2.5354 Chi-sq( 1) P-val = 0.11132 -----------------------------------------------------------------------------| GMM ctax95 | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------cl9095m | 4.823629 3.231311 1.49 0.135 -1.509623 11.15688 lnopen | 8.056173 2.002373 4.02 0.000 4.131594 11.98075 lninc | -3.078245 2.580563 -1.19 0.233 -8.136056 1.979566 youngp | 13.86968 21.67239 0.64 0.522 -28.60743 56.34679 oldp | 117.3879 42.63996 2.75 0.006 33.8151 200.9607 urban | .0638048 .0434851 1.47 0.142 -.0214244 .149034 logpop95 | 1.718596 1.286152 1.34 0.181 -.8022145 4.239407 _cons | -18.20809 24.61719 -0.74 0.460 -66.4569 30.04072 -----------------------------------------------------------------------------Instrumented: cl9095m Instruments: lnopen lninc youngp oldp urban logpop95 lat federal -----------------------------------------------------------------------------. reg ctax95 pr9095m lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 11.85 0.0000 0.5612 6.6868

-----------------------------------------------------------------------------| Robust ctax95 | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | 4.606506 1.624213 2.84 0.006 1.351511 7.861501 lnopen | 8.67077 2.201441 3.94 0.000 4.258984 13.08256 lninc | 1.887718 2.201739 0.86 0.395 -2.524665 6.300101 depend | 1.54826 11.74975 0.13 0.896 -21.99877 25.09529 urban | .0595139 .054052 1.10 0.276 -.0488086 .1678365 logpop95 | 1.651788 .7967345 2.07 0.043 .0550962 3.24848 _cons | -53.09743 30.01073 -1.77 0.082 -113.2403 7.045423 -----------------------------------------------------------------------------. ivgmm0 ctax95 (pr9095m = lat federal ) lnopen lninc depend urban logpop95;

Instrumental Variables Estimation via GMM

Number of obs = 62 Root MSE = 8.5680 Hansen J = 2.3377 Chi-sq( 1) P-val = 0.12627 -----------------------------------------------------------------------------| GMM ctax95 | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | 12.85857 2.715477 4.74 0.000 7.536332 18.18081 lnopen | 17.22345 3.415348 5.04 0.000 10.52949 23.91741 lninc | -4.170137 3.101334 -1.34 0.179 -10.24864 1.908367 depend | 15.07615 16.10054 0.94 0.349 -16.48032 46.63263 urban | .1210096 .0704945 1.72 0.086 -.017157 .2591763 logpop95 | 4.572609 1.352171 3.38 0.001 1.922402 7.222817 _cons | -77.45576 42.83946 -1.81 0.071 -161.4196 6.508038 -----------------------------------------------------------------------------Instrumented: pr9095m Instruments: lnopen lninc depend urban logpop95 lat federal Page 8 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 -----------------------------------------------------------------------------. reg ctax95 pr9095m lnopen lninc youngp oldp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 54) = Prob > F = R-squared = Root MSE = 62 32.21 0.0000 0.7223 5.3686

-----------------------------------------------------------------------------| Robust ctax95 | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | 1.646388 1.47374 1.12 0.269 -1.308284 4.60106 lnopen | 7.518767 1.908949 3.94 0.000 3.691554 11.34598 lninc | .0162575 1.965897 0.01 0.993 -3.925129 3.957644 youngp | 31.09372 22.32292 1.39 0.169 -13.66103 75.84848 oldp | 168.4829 27.94428 6.03 0.000 112.458 224.5079 urban | .0429701 .0417975 1.03 0.309 -.0408288 .126769 logpop95 | .8313974 .6519514 1.28 0.208 -.4756865 2.138481 _cons | -45.54011 25.82622 -1.76 0.084 -97.31857 6.238354 -----------------------------------------------------------------------------. ivgmm0 ctax95 (pr9095m = lat federal ) lnopen lninc youngp oldp urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 5.6236 Hansen J = 2.2583 Chi-sq( 1) P-val = 0.13290 -----------------------------------------------------------------------------| GMM ctax95 | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------pr9095m | 5.528809 4.090494 1.35 0.176 -2.488412 13.54603 lnopen | 10.97524 4.147973 2.65 0.008 2.845358 19.10511 lninc | -2.579722 2.503277 -1.03 0.303 -7.486053 2.32661 youngp | 28.97177 24.2759 1.19 0.233 -18.60813 76.55167 oldp | 133.819 37.87582 3.53 0.000 59.58378 208.0542 urban | .0782096 .0487311 1.60 0.109 -.0173016 .1737208 logpop95 | 1.984226 1.650456 1.20 0.229 -1.250609 5.219061 _cons | -46.58726 32.66268 -1.43 0.154 -110.6049 17.43041 -----------------------------------------------------------------------------Instrumented: pr9095m Instruments: lnopen lninc youngp oldp urban logpop95 lat federal -----------------------------------------------------------------------------. /* Press freedom */ > reg gtax95 press95m lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 55) = Prob > F = R-squared = Root MSE = 62 25.33 0.0000 0.6374 7.1267

-----------------------------------------------------------------------------| Robust gtax95c | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------press95m | .3663389 .0910807 4.02 0.000 .1838092 .5488686 lnopen | 5.858966 2.154607 2.72 0.009 1.541036 10.1769 lninc | 2.455589 2.211406 1.11 0.272 -1.976168 6.887345 depend | -3.530796 12.02177 -0.29 0.770 -27.62297 20.56137 urban | .0931973 .0498741 1.87 0.067 -.0067526 .1931472 logpop95 | 1.977253 .6333432 3.12 0.003 .7080045 3.246501 _cons | -45.04733 28.9791 -1.55 0.126 -103.1227 13.02809 -----------------------------------------------------------------------------. ivgmm0 gtax95 (press95m = lat federal prot) lnopen lninc urban depend logpop95;

Instrumental Variables Estimation via GMM

Number of obs = 62 Root MSE = 10.7125 Hansen J = 1.2881 Chi-sq( 2) P-val = 0.52515 -----------------------------------------------------------------------------| GMM gtax95c | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------press95m | 1.116793 .232303 4.81 0.000 .6614871 1.572098 Page 9 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 lnopen | 15.60724 3.807479 4.10 0.000 8.144718 23.06976 lninc | -6.232368 4.209358 -1.48 0.139 -14.48256 2.017822 urban | .19494 .0735204 2.65 0.008 .0508427 .3390372 depend | 11.40041 14.66653 0.78 0.437 -17.34546 40.14628 logpop95 | 5.826626 1.466236 3.97 0.000 2.952857 8.700395 _cons | -62.52618 41.23794 -1.52 0.129 -143.351 18.29869 -----------------------------------------------------------------------------Instrumented: press95m Instruments: lnopen lninc urban depend logpop95 lat federal prot -----------------------------------------------------------------------------. reg gtax95 press95m lnopen lninc oldp youngp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 54) = Prob > F = R-squared = Root MSE = 62 44.75 0.0000 0.7955 5.4015

-----------------------------------------------------------------------------| Robust gtax95c | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------press95m | .1869642 .0760191 2.46 0.017 .0345552 .3393733 lnopen | 5.969783 1.886749 3.16 0.003 2.187079 9.752487 lninc | .4213827 1.715328 0.25 0.807 -3.017642 3.860408 oldp | 185.9121 38.55549 4.82 0.000 108.613 263.2112 youngp | 31.99957 23.24865 1.38 0.174 -14.61117 78.61032 urban | .0688439 .0402527 1.71 0.093 -.011858 .1495458 logpop95 | 1.437254 .5657649 2.54 0.014 .3029639 2.571544 _cons | -49.75312 26.50307 -1.88 0.066 -102.8886 3.382332 -----------------------------------------------------------------------------. ivgmm0 gtax95 (press95m = lat federal prot) lnopen lninc oldp youngp urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 62 Root MSE = 8.5028 Hansen J = 0.9258 Chi-sq( 2) P-val = 0.62945 -----------------------------------------------------------------------------| GMM gtax95c | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------press95m | .8590918 .3730616 2.30 0.021 .1279046 1.590279 lnopen | 13.13111 4.101555 3.20 0.001 5.092214 21.17002 lninc | -4.702575 4.08845 -1.15 0.250 -12.71579 3.31064 oldp | 100.9522 82.1908 1.23 0.219 -60.13886 262.0432 youngp | 40.92982 32.98504 1.24 0.215 -23.71967 105.5793 urban | .1557036 .073665 2.11 0.035 .0113229 .3000843 logpop95 | 4.615078 1.857766 2.48 0.013 .973924 8.256232 _cons | -65.15607 36.32516 -1.79 0.073 -136.3521 6.039931 -----------------------------------------------------------------------------Instrumented: press95m Instruments: lnopen lninc oldp youngp urban logpop95 lat federal prot -----------------------------------------------------------------------------. > > > > > > > /* reg gtax95 cl9095m press95m lnopen lninc depend urban logpop95, robust; reg gtax95 pr9095m press95m lnopen lninc oldp youngp urban logpop95, robust;*/ /* outliers */ /* Hadi, in J. of Royal Statistical Society, 1992, B 54:761-771 and 1994, 56;393-396 */ hadimvo gtax95 polacct lnopen lninc depend urban logpop95, gen(t1); 62 8 35 61 1

Beginning number of observations: Initially accepted: Expand to (n+k+1)/2: Expand, p = .05: Outliers remaining: . drop if t1; (3 observations deleted)

. /* Redo Table 1 without outliers*/ > reg gtax95 polacct lnopen lninc depend urban logpop95, robust; Regression with robust standard errors Number of obs = F( 6, 54) = Prob > F = Page 10 of 12 61 30.21 0.0000

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 R-squared Root MSE = = 0.7264 6.2041

-----------------------------------------------------------------------------| Robust gtax95c | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 9.545465 1.534668 6.22 0.000 6.468642 12.62229 lnopen | 5.149979 1.768087 2.91 0.005 1.605178 8.694781 lninc | 1.881267 1.674464 1.12 0.266 -1.475832 5.238366 depend | -1.607393 10.83635 -0.15 0.883 -23.33297 20.11818 urban | .1224105 .0475467 2.57 0.013 .0270851 .217736 logpop95 | 1.530855 .5831192 2.63 0.011 .3617717 2.699939 _cons | -51.45574 24.89608 -2.07 0.044 -101.3694 -1.542103 -----------------------------------------------------------------------------. ivgmm0 gtax95 (polacct = lat federal prot) lnopen lninc urban depend logpop95; Instrumental Variables Estimation via GMM Number of obs = 61 Root MSE = 7.6573 Hansen J = 1.1202 Chi-sq( 2) P-val = 0.57115 -----------------------------------------------------------------------------| GMM gtax95c | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 20.9843 4.004463 5.24 0.000 13.1357 28.83291 lnopen | 5.536802 2.281087 2.43 0.015 1.065954 10.00765 lninc | -3.113786 2.919043 -1.07 0.286 -8.835005 2.607433 urban | .1491637 .0545482 2.73 0.006 .0422512 .2560763 depend | 12.7068 13.31804 0.95 0.340 -13.39608 38.80968 logpop95 | 2.609001 .92848 2.81 0.005 .7892133 4.428788 _cons | -45.42261 34.05128 -1.33 0.182 -112.1619 21.31667 -----------------------------------------------------------------------------Instrumented: polacct Instruments: lnopen lninc urban depend logpop95 lat federal prot -----------------------------------------------------------------------------. reg gtax95 polacct lnopen lninc oldp youngp urban logpop95, robust; Regression with robust standard errors Number of obs = F( 7, 53) = Prob > F = R-squared = Root MSE = 61 45.36 0.0000 0.8209 5.0676

-----------------------------------------------------------------------------| Robust gtax95c | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 4.907398 1.563134 3.14 0.003 1.772148 8.042648 lnopen | 5.835053 1.762934 3.31 0.002 2.299055 9.371051 lninc | .6667912 1.464732 0.46 0.651 -2.271089 3.604671 oldp | 154.0074 39.77382 3.87 0.000 74.23121 233.7836 youngp | 24.90029 22.92468 1.09 0.282 -21.0808 70.88137 urban | .0918005 .0397941 2.31 0.025 .0119836 .1716173 logpop95 | 1.244278 .4878031 2.55 0.014 .2658692 2.222687 _cons | -54.48444 25.53262 -2.13 0.037 -105.6964 -3.27249 -----------------------------------------------------------------------------. ivgmm0 gtax95 (polacct = lat federal prot) lnopen lninc oldp youngp urban logpop95; Instrumental Variables Estimation via GMM Number of obs = 61 Root MSE = 6.3528 Hansen J = 0.6523 Chi-sq( 2) P-val = 0.72170 -----------------------------------------------------------------------------| GMM gtax95c | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------polacct | 16.24397 6.177927 2.63 0.009 4.135459 28.35249 lnopen | 5.554211 1.969946 2.82 0.005 1.693187 9.415234 lninc | -2.280283 2.58943 -0.88 0.379 -7.355473 2.794907 oldp | 95.44844 57.21769 1.67 0.095 -16.69616 207.593 youngp | 40.58171 29.84457 1.36 0.174 -17.91258 99.07599 urban | .1278467 .0512984 2.49 0.013 .0273037 .2283897 logpop95 | 2.15193 .9184661 2.34 0.019 .3517693 3.95209 _cons | -52.49734 31.66913 -1.66 0.097 -114.5677 9.573015 -----------------------------------------------------------------------------Instrumented: polacct Page 11 of 12

C:\harvard\account\data\addresults.log, 13:03, 17-12-01 Instruments: lnopen lninc oldp youngp urban logpop95 lat federal prot -----------------------------------------------------------------------------. log close; log: c:\account\data\main3sup.log log type: text closed on: 6 Nov 2001, 11:19:55 -------------------------------------------------------------------------------------------------------------------------------------------------

Page 12 of 12