Public Choice (2005) 122: 355393

DOI: 10.1007/s11127-005-3900-y

Springer 2005

A logistic growth theory of public expenditures: A study of ve countries over 100 years
MASSIMO FLORIO & SARA COLAUTTI
Dipartimento di Economia Politica e Aziendale, Universit degli Studi di Milano, 20122 Milano, Italy; e-mail: massimo.orio@unimi.it Accepted 17 November 2003 Abstract. This paper offers a new theory and empirical testing of long-term trends of public expenditures for ve countries. While Wagners Law would imply an exponential growth process of the ratio between public expenditures and national income (G/Y), the law should be rejected both on theoretical and empirical grounds, because it disregards the role of ever increasing distortionary taxation. However, under some conditions, the combination of Wagners Law and the Pigous conjecture that the excess burden of taxation constrains the growth of public expenditures can be captured by a non-linear rst order differential equation. The equation is the Verhulsts logistic, originally invented to model Malthusian predictions on population growth. The integration of a Verhulst equation generates an S-shaped curve. This analytical framework combines intuitions from a welfare economics and a public choice perspective, and potentially offers a new research strategy on the dynamics of government expenditures. We offer preliminary econometric estimates on long run trends (around 1870 1990) of G/Y in U.S., U.K., France, Germany, Italy. These estimates conrm a pattern of similar trajectories, in spite of different national parameters, and suggest that the logistic view of growth of government is consistent with observed data.

1. Introduction This paper offers a new theory of the secular growth of public expenditures and empirical testing for ve countries. While Wagners Law would imply an exponential growth process of the ratio between public expenditures and national income (G/Y), the law should be rejected both on theoretical and empirical grounds, because it disregards the role of ever increasing distortionary taxation. But, under some conditions, the combination of Wagners Law and the Pigous conjecture that the excess burden of taxation constrains
The authors are grateful for helpful comments to Dr. George Tridimas, University of

Ulster, to participants at a seminar of the Department of Economics, Milan University, to participants in the annual meeting of the Italian Society for Public Economics (Pavia, 2002) including Professor Agnar Sandmo, and to two anonymous referees. A Grant of Ministry of the University, MURST 40% (University of Milan), nanced the research. Also MURST 60% funds were helpful. Dr Stefano Iacus and Prof. Achille Vernizzi gave very helpful suggestions for the estimation techniques. The usual disclaimer applies.

356 the growth of public expenditures can be captured by a non-linear rst order differential equation. The equation is the Verhulsts logistic, originally invented to model Malthusian predictions on population growth. The integration of a Verhulst equation generates an S-shaped curve. We offer econometric estimates on very long run trends (around 18701990) of G/Y in U.S., U.K., France, Germany, Italy. These estimates conrm a pattern of similar trajectories, in spite of different national parameters, and suggest that the logistic view of growth of government is consistent with observed data. This result offers a new framework for the analysis of long term trend of public expenditures. The growth of government in the last 150 years has been huge. A rough measure of this process is the increase in the ratio between public expenditures and national income (G/Y). This ratio was between 5% and 10% in the second half of 19th century in U.S., U.K., France, Germany, Italy, Sweden and for many other countries involved in modern economic growth (Mueller, 2003; Tanzi and Schuknecht, 2000; Peacock and Scott, 2000). At the end of the 20th century the ratio was typically in a range between 35-55%. Quite often public nance scholars and policy-makers have been interested in the level of G or G/Y; the long-run dynamics of the process, however, may be also very interesting. For example, and contrary to popular perceptions, since the beginning of this century the increase of G/Y has been greater in the U.S. and the U.K., while it was slower in some big-spender countries, as France and Italy, where the ratio was already high at the beginning of the century. Moreover, in most countries the growth of G/Y started to decrease after World War II, in the decades usually perceived as the booming of the welfare state (Table 1). Looking just at levels may be misleading in a long run process. Several theories have been proposed in order to explain the level of G or G/Y or public expenditures per capita. The long term dynamic growth mechanism is elusive. However, when we look carefully to available data (scant and imprecise as they are), this pattern can be depicted as a sigmoid curve. The rate of growth of G/Y increases continuously over time until a certain point, then the process is reversed: while G/Y still increases, its growth rate declines. At the end of the 20th century the latter is near to zero, indicating perhaps the convergence to a steady state. We explore this pattern for several countries and offer a simple theory that may explain it. The theory we propose assumes (Section 2) that virtually the growth of the demand for G/Y is an exponential process, as was rst suggested by Wagner (1893). It well may be true that most goods and services (including the administration of transfers) efciently offered by the state show demand elasticity to income greater than 1. However, if this were unconditionally true, the trajectory of G/Y over time we should observe would be an

357

Table 1. The ratio public expenditures/national income. Selected years and data sources.

358

Figure 1. Logistic curve

exponential curve. In fact we show this was approximately true at Wagners time and for subsequent decades but we suggest that Wagners law disregards the social cost of distortionary taxation (Section 3). Following a conjecture by Pigou (1947), who wrote at a time of changing trends, we assume the excess burden of taxation acts as a brake to the supply of publicly provided goods (and transfers). The increase over time of the excess burden of taxation is a function of the square of T/Y, where T is distortionary tax revenue. If G = T, or the budget is balanced in the long run, the combination of Wagners law and Pigous conjecture may be captured by a non linear differential equation in G/Y. The equation is Verhulsts logistic (1847), rst proposed to model Malthus population law, and than widely used by demographers (Peerl and Reed, 1920), mathematical biologists (e.g., Lotka or Volterra, 1931) and more recently in deterministic chaos theory (Section 4). The solution of the equation gives an S-shaped curve over time, with some convenient properties that lend them-

359 selves to interesting economic interpretation. We then study long term data on G/Y for ve countries: U.S., U.K., Germany, France, Italy (Section 5). We have selected a limited number of data sources out of a longer list of possible candidates (see References). The selection criterion was comparability of denitions and not the length of the time series. However there are some remarkable differences in levels of G/Y reported by different sources, reecting an array of denitions and estimates. In some cases we used time series comprising a very limited number of data, but we tested our model on a much longer set of sources with similar results. Simple estimation procedure is used to test an exponential process against a logistic one (Section 6). The econometric test, while a preliminary one, conrms that for all ve countries a logistic equation in G/Y offers a better t than an exponential one. Thus the data do not reject our simple theoretical model. We conclude the paper by claiming that the new theory is consistent with observation and offers a fruitful framework for further analysis of long-term dynamics of public expenditures.

2. Public expenditures theories The debate on the size of government, both at theoretical and empirical level, is old, controversial and often inconclusive. We are unable to review it here: others have provided excellent surveys. It would be helpful however to restate some aspects of the current debate. Mueller (2003: 501 ff.), starts his survey by presenting simple facts, such as the increase of federal expenditures in the U.S., from 1.4% of national income in 1799 to 23,7% of GNP in 1987, or comparative statistics for OECD countries; Tanzi and Schuknecht (2000) present secular data for several countries, observe a huge increase of G/Y ratio for most countries over time, but also differences in levels of G/Y, with the U.S. at the bottom of an international ranking by this ratio. They ask themselves which country is on the right spending track in welfare terms: big spenders, e.g. Scandinavian countries, where G/Y is around 50%, or the U.S., Switzerland and some other countries where public expenditures do not exceed, let us say, 30-40% of GDP. These examples of broad perspectives are representative of two big research families in the growth of government literature: the more positive approach, versus the more normative approach. Under the former heading there is the search for explanations for the size and growth of government, under the latter the question is about optimal size or desirable reforms. Mueller classies the explanatory literature in ve strands of research.

360 First, according to a popular view, government grows because there is an increasing demand for public goods and for the control of externalities. The median voter theorem is frequently invoked in this context, and a number of models have been proposed, where typically explanatory variables of G (or of its change over time) are the income of the median voter, a vector of taste parameters, and relative prices of public and private goods (see e.g. Borcherding, 1977). Most of the existing empirical literature in this framework could be seen as an inquiry into the elasticity of public expenditures to income, to tastes and to prices. The price component of the story was particularly emphasised in papers by Baumol (1967), Baumol, Blackman and Wolf (1985), and subsequent research (more on this below). Tastes have been perhaps less studied, while the evaluation of income elasticity is in the core of the wide Wagner empirical testing literature, discussed, and criticised by Peacock and Scott (2000). Recent examples of this literature include Ram (1987), Ahsan et al. (1996), Ansari et al. (1997), Bohl (1996), Payne and Ewing (1996), Courakis et al (1993), Afxentiou (1997), Legrenzi, Costas (2002). Second, classic contributors such as Musgrave (1965) have acknowledged that government is a powerful re-distributor of income and wealth, and some authors following Meltzer, Richard (1978) and others have proposed models where the rather extreme view is taken that all government activity has a redistribution element, and thus the extension of suffrage may explain the growth of government because low-income voters push for more and more re-distributive expenditures. Third, others have emphasised the role of interest groups able to capture government through the majority rule, and the earlier literature pioneered by Buchanan and Tullock has been more recently revived by the political economy approach to economics, both at national and local level (e.g. Tridimas, 2001). Fourth, bureaucracy may be seen as a special but important case of interest group, and the earlier work by Niskanen, or, fth, the classic study of scal illusion by Puviani (1903) can be seen now as part of a large body of research on information structures in the organisation of the state (the principal-agent framework for the theory of the state). In contrast to these strands of literature, the normative approach to public economics has offered models of the optimal provision of public goods and the study of desirable reforms. One typical aspect of this literature is, implicitly or explicitly, a cost-benet test. A welfare maximising social planner should expand public provision up to the point the marginal social cost of provision equals its marginal social benet. This approach involves the assumption and/or estimation of a social welfare function and a character-

361 isation of second best properties for economies where lump-sum taxation is unfeasible. In this framework there is often an element of dead-weight loss of taxation as counterpart of social benets of public expenditures, see e.g. Ng (2000). In what follows we contribute to the existing public expenditure literature by proposing an extremely simple model that combines what we see as a key aspect of the positive approach, namely the income driver of public expenditure, with the typically welfaristic idea that the excess burden of taxation should play a role in the assessment of costs and benets of public expenditures. In other words our frameworks blends in a straightforward way intuitions from a public choice and from a welfare economics perspectives, and opens the path to a research strategy on the dynamics and interactions of the drivers of public expenditures under distortionary taxation. Before going on with the presentation of the model, and empirical testing of it, it seems convenient to state the simple idea that we wish to explore. We assume that there is a social demand for public expenditures (conveniently dened), that this social demand is increasing with per capita income, and that the social good provided by public expenditure virtually is a superior good. At the same time we assume that there is only distortionary taxation available to government to nance public expenditures. Under these circumstances, the dynamics of G/Y has one driver and one brake, and there may be the case that over time the brake becomes powerful enough to stop the driver. This intuition can be easily framed in the old debate on the Adolph Wagner (18351917) Law of Increasing State Activity. According to a straightforward interpretation of Wagners work (1893),1 the ratio G/Y should increase over time because the goods and services offered by the State are superior goods. In other words, the increasing complexity of the society calls for an increasing role of the State. This development was so evident according to Wagner that he formulated it as a law of ever increasing Staatsttigkeit including both public expenditures and regulation. The examples he offers relate to transport, education, and technical progress: e.g. the steam technology will make easier for the state and municipality to organize its own production plants (Wagner, 1893: 902), more efciently than the private sector. According to Peacock, Scott (2000: 3): Wagner was aware that his association of growth in government activity with growth and development of emerging industrial societies raised the question of the time period over which his empirical uniformity was supposed to prevail. He also recognised that there must be an upper limit to the proportion of government growth which would be determined by the resistance that would come from taxpayers if their normal personal

362 consumption were to be diminished by tax burdens, but he was scornful of attempts to dene such a limit (see Wagner, 1883). A similar idea could be found in Pigou (1947) and thus two founding fathers views of public expenditures, one more positive the other one more normative, can be reconciled. We shall show that an extremely simple dynamic model potentially answers the questions implicit in the above citation: rst, we estimate the time span over which Wagner Law is approximately a good description of the demand and supply of public expenditures; second, we offer a the determination of a virtual upper limit of G/Y based on the welfare burden of taxation. Moreover, we can discover when there is a change of regime, from the realm of a quasi Wagner Law to a mature context where the burden of taxation reverses the process. In the spirit of the classical contributions in public nance doctrine, the theory should be as simple as possible, but more in line with contemporary need for empirical verication, the theory should also be testable. We do not propose predictions, or detailed models for the specic countries or type periods, but a law based on very few assumptions and a framework easy to implement for testing with a richer set of control variables.

3. The demand of public expenditures: Wagners Law restated Wagners intuition, in our interpretation, is that a) Some goods can be supplied efciently by the state b) The demand for these goods is such that as Y, or Y per capita, increases over time (rather than across countries), their demand increases even more. While many economists or politicians may disagree on both propositions, it is a matter of fact that in the last 100-150 years, and in many countries such goods as education, health and social insurance have been mostly delegated to the state, indicating in some way a social preference for this public provision. It may also be true, albeit less easy to test empirically, that many of these goods show for long periods of time income elasticity greater than 1, while many goods typically offered by the private sector are inferior goods in respect to income (food is a typical example). The above propositions do not need to be true for all the items of the public expenditures; it is sufcient that they are true for the average or typical item comprised in G. Having said this, the intuition behind Wagners Law may be modelled in different ways. According to Henrekson (1990, 1993), Flster and Henrekson (2001) at least ve different testable models can be found in the literature on this subject. We list these basic models and also the elasticity

363 values n of the dependent variable to the independent variable that may be seen as conrming Wagner Law: W.1) G/Y W.2) G W.3) G/N W.4) G W.5) G/Y = = = = = f(Y /N) f(Y) f(Y /N) f(Y /N) f(Y) 1 2 3 4 5 >0 >1 >1 >1 > 0.

In this context, G should typically be understood as total public expenditures, excluding transfers, of general government (both central and local), Y and Y as respectively nominal and real national income (probably GDP may be the appropriate variable), N is population. Henrekson shows that models W.1) and W.3) are equivalent for a monotonic transformation; so are W.2) and W.5), while W.4) is conceptually different (and the interpretation of the elasticity is also more loosely related to Wagners Law). Moreover it can be shown that the elasticity for W.1) and W.2) is the same if N is constant and Y/N is always increasing. We do not want at this stage to discuss in detail the most appropriate empirical denitions of variables for G, Y, N, an issue by far much more complex that it may look at rst sight (for our selection of empirical variables see below, Section 7). Consider for instance the issue of transfer in this framework. In our opinion there is little doubt that Wagner had in mind the increase of public consumption and investment relative to private demand, while he did not mention transfers (in fact quite low at the end of 19th century). However, we can now think that transfers are a way to provide redistribution and insurance, and that these may also be superior goods. For example, Rodrik (1998) suggests that open economies face risk of price and income volatility and shows empirical evidence that suggests that public consumption for developing countries is a substitute for social insurance programs, e.g. unemployment bents. Henrekson (1993), Flster and Henrekson (2001) suggest that there is not much insight to gain from cross-country analysis, and that long time series may be more revealing. However, in this case we need to model explicitly the dynamic process, while none of the above equations is dynamic. To restate Wagners Law over time, let us rst think of a progressive economy with stable population. Thus we assume: Y = Y(t); N = N(t); G = G(t) and dY/dt > 0, dN/dt = 0 Derivatives are over time. In this setting we want to test W.1 or equivalently W.2, implicitly then we test also W.3 and W.5 (W.4 is excluded from the scope of our study, because we think it is poorly related to Wagners intuition).

364 Our interest is to study the trajectory of G/Y. It is convenient to use the following notation: g= y= R = G/Y It easy to see that dR/dt = hR where h= gy In fact it will then be h(G/Y) = 1/Y(dG/dt) G/Y[1/Y(dY/dt)] The right side of the above is exactly the same thing as the derivative of G/Y dR/dt = [Y(dG/dt) G(dY/dt)]/Y2 (3) 1 dG d log G = dt G dt 1 dY d log Y = dt Y dt (1) (2)

Then h > 0 implies dR/dt > 0, for R0 > 0, which we can easily assume. There is also a simple relationship between and h: h = y( 1). (4)

Thus for y > 0 and constant by assumption, also any value of > 1 and constant over time implies a simple exponential process integrated by R = R0 eht . If y changes over time, the same exponential process to be generated needs an offsetting change in the opposite direction of the income elasticity of G. In the more general case h may vary over time, and the integration of the differential equation above may generate many different type of trajectories. However, in what follows we are going to conne ourselves to the case of constant h, but it is important to remember that this assumption is compatible either with log linear trend of GDP and xed elasticity of public expenditures to income, or alternatively with appropriate offsetting trends in these parameters. We come back on this later. Having said this, it is clear why in the long run the Wagner Law is both theoretically and empirically implausible. An exponential process implies that G/Y increases over time without limits, and this in turn implies either

365 an ever-increasing debt or a very costly process of transfer and tax. The latter is true for G=T, where T is tax revenue, when G/Y > 1, because in this case there is no other way to levy taxes than to tax public expenditures themselves (e.g. taxation of public pensions). The alternative for G > T is a debt rising without limits. And this also seems implausible. However, it may well be that Wagners Law is only a part of the long-run dynamics but that other forces put a brake on it. To this inquiry we turn now.

4. The excess burden of taxation: Pigou s conjecture One weak point of Wagners Law is that it simply ignores the social cost of nancing government expenditures by taxation. Public expenditures may be nanced by incomes directly accruing to government (e.g. statutory monopolies, rents of public land and forests, etc), by seignorage, by debt, and by taxation. Each of these sources of government revenue has its own welfare impact, but taxation is obviously associated with less scal illusion than the other sources, and often it caused greater social unrest and political opposition than public monopolies, seignorage and debt. However, out of necessity, the 19th century state moves increasingly away from non-tax sources of revenue. For a detailed presentation of the available evidence of this historical pattern see Schremmer (1989). He documents the steadily declining role of rents from state-owned lands, forests, and real estates as a source of revenues. For example, in France the latter accounted for 11.9% of public revenues in 180014 and for just 1,6% in 1913. In the U.K. the corresponding revenues from the Crown estates was 0,28% in 1913. In Prussia the non-tax revenues were sizeable in the early 19th century, around 28% of revenues came from property of forests, lands and monopolies (e.g., salt and post, in later years also railways). However, in 1913 public enterprises contributed only 16,1% of the revenues of the German Reich, and then decreased steadily. While debt still played an important role, taxation largely substituted for other sources of public nance. Pigou (1947), writing in decades of increasing scal burdens, was well aware of this pattern. In fact, public monopolies (e.g. railways in Germany, or Royal Mail in Britain) were still important in 19th century, but in subsequent decades covered a very thin proportion of general government expenditures. More recently, despite occasional crisis, particularly after the First World War and the Oil Shocks in the 1970s, seignorage and public debt nance were typically below 10% of GDP in developed economies. To sum up: in the last two centuries, the process of decrease of non-tax government revenues and of parallel increase of the scal pressure has been different by country, but the overall trend is clear: government expenditures in 20th century are mostly tax-based (including social security contributions).

366 Hence, at this stage of our inquiry, we shall ignore these other sources of government revenues (while we suggest that more complete models should allow for a mix of tax and non-tax revenues and for their differential impact on social welfare). Thus, we shall assume below that normally G = T or G/Y = T/Y. Having said this, in order to study the supply of good offered by the state we should think on their production and cost functions. We introduce here two further assumptions: a) We assume constant returns to production factors for the composite average good provided by the public sector; b) There is an extra-cost of taxation, its excess burden, and this is quadratic in T/Y. As for assumption a), only a detailed analysis of the production function of such goods as education, health, transport, social insurance, etc. may establish whether these goods show increasing, decreasing or constant returns. One traditional case for public provision has been that some or most of these industries show increasing returns or that they are natural monopolies. Turning to dynamic economies of scale, Baumol (1967, 1985) tried to show that under certain conditions, when one sector shows stagnating productivity while the other one is progressive, the rst one would absorb most of the expenditures of the economy if the demand is inelastic. He thought that this proposition was relevant in order to explain the growth of public expenditures. However, the pattern that would emerge if Baumols theory holds were again an exponential growth of G/Y and this is not plausible. Moreover, there is no evidence that on average and in the long run government activity is based on technologies that do not allow for (absolute or relative) rise in productivity. One may think of many counter-examples. Having said this, it seems that while it is possible that some government activities show decreasing returns, it is more likely that many others show scale economies. We assume here constant returns as a benchmark case, because we think that for reasonable empirical parameters our qualitative conclusions are unaffected by a perhaps more realistic (but difcult to test for the aggregate) assumption that government exhibits economies of scale (see below on evidence on this point, Section 6). Under constant production costs on average one extra Euro of inputs provides one extra Euro of government output in real terms. We may suppose that the gross marginal social benet of one

367 Euro of G is not less than its marginal production cost. This is approximately true for pure transfers (when we disregard administration costs), while it may be an understatement for such goods as education or health (standard national income accounting conventions make the assumption that one more Euro of public sector costs adds one more Euro to GDP, hence to national welfare, if we are not interested in income distribution: a particularly crude, but practical, way of measuring economic welfare). However, consumers of G typically do not pay an individual price for the goods they consume, they pay taxes that cover the aggregate cost of G. Because taxation is distortionary (quasi lump-sum taxes may be levied, but they play a minor role) a positive T/Y implies a cost in excess to the social benet of G/Y, the excess burden of taxation. How large is this cost? There are several ways to compute it, all of them based on specic assumptions on the measurement of consumer welfare: again this is not the place to study this in detail or to review a well developed literature.2 However it is easy to show that the excess burden for small increases of taxation should be a function of the square of T/Y. Starting with the simplest case, suppose there is an economy with one consumer, one private good and one publicly provided good. The quantity of the private good is x, while its xed unit production cost is p. If the only taxable good is the private one (leisure is untaxed, or there is no lump sum taxation), tax revenues for the state are T = px = T/px

where is the effective tax rate. If the public good is provided free, national income at consumer prices is Y = px where p is the consumer price. In this simple context suppose we start at t = 0 with = 0. Thus for a small tax dp = p = where = T/x is the tax for a conventionally small unit of good. The Marshallian demand price elasticity of the private good is = |(dx/dp)/(x/p)| and by substitution dx = x. Thus the simple standard denition of the excess burden of taxation implies: E = dxdp/2 = Y 2 /2 (5)

368 But because G = T and Y = px, the ratio of the aggregate excess burden to national income is then quadratic in the ratio of public expenditures to national income itself: E/Y = 2 /2 = (/2)(T/px)2 = (/2)(G/Y)2. (6)

We can think that the general form of a differential equation for the growth of public expenditures is thus: d(G/Y)/dt = f(G/Y, E/Y) where the expected sign for E/Y coefcient is negative. There is a simple theoretical case for this claim. A welfare maximising government should produce the publicly provided good, in a one consumer economy, up to the point where the social benets of provision to the consumer balance with the social costs to the taxpayer, including the excess burden of distortionary taxation. One can also think of a political economy argument for an economy with two types of consumers/voters: one type is a net beneciary of public provision, the other one is more concerned with taxation. The two types inuence the government decision to spend and equilibrium should be reached over time. More on this later. The simplest empirical specication of this equation is a linear combination of Wagners law (as restated above) and of Pigous conjecture of the increasing excess burden of taxation as a brake to public expenditures: d(G/Y)/dt = ( 1)(G/Y) (/2)(G/Y)2 (7)

where , are parameters that may reect institutional factors, e.g. the time needed by the political system to react to welfare changes of voters or the welfare weights attached by the social planner to different types. We are going to discuss the political economy of this public expenditure process below in Section 6, and we show that the model itself can be interpreted in a way consistent with the view that in the long run the consumer-tax payer-voter is able to inuence policy makers. The above result needs a number of qualications. First, consumer surplus measured on the Marshallian curve is generally an unsatisfactory basis for the evaluation of the excess burden of taxation, for well-known reasons. A better measure is a metric based on the equivalent variation on compensated or Hicksian demand curve. In this case, in contrast to the compensating variation measure, the tax revenue is the effective one and, this is particularly useful in our context. Second, moreover, we are interested in a dynamic process where taxes are already existing in previous periods (in fact we use a continuous process). It

369 is known that in such case rst order welfare effects will appear (Harberger, 1964; Auerbach, 1990; Feldstein, 1997). The Harberger triangle becomes a trapezoid. A second order Taylor expansion formula in such a case is E = (dE/dp) + 1/2(d2 E/dp2 ) + . . . where is the tax per unit of good, q = p + . For linear compensated demand x we have then the approximate value E = dx + 1/2(d dx) where d is a small change of the unit tax over time. Thus we can write: dx = (x/p)dp = x Thus the rst addendum is 2 Y and the second addendum is 1/2(Ydx). The approximate value of the excess burden relative to income, with preexisting taxes, remembering T = G is circa: E/Y = [(G/Y)2 + 1/2d(G/Y)]. (9) (8)

Thus again we have a quadratic term in the tax rate, plus an additional term. Even if we ignore the latter, it is clear that the excess burden coefcient in the equation above is now greater than in the benchmark case. For example, a country where G/Y is around 0.50, as e.g. Italy or France, and where the aggregate compensated demand elasticity to prices is, say, 0.20, would have an excess burden of 5% of GDP plus half the increase of G/Y.3 Having said all this, it seems that there is a theoretical case for a differential equation of public expenditure growth of the general form shown in equation (7), where the key parameters are two elasticities: respectively the income elasticity of public goods and the compensated price elasticity of private goods; plus the growth rate of national income, and institutional factors affecting the reactivity of the political system to the change in consumer and taxpayer welfare (not explored in this paper). The above equation is a non-linear rst-order ordinary differential equation that can be written as d(G/Y)/dt = (G/Y)(h k(G/Y)). (10)

where h = y( 1) and k = In this form the equation is a logistic process. This equation has some interesting properties that we shall discuss in the next section. Looking at the history of government through this perspective may give us a new perception of long-term trends of public expenditures.

370 5. Logistic growth Let us resume the above assumptions. Consumers-taxpayers want more and more G as their income increases, but they suffer an extra loss of welfare that increases with the square of T/Y. Because T/Y = G/Y, their excess burden will be a function of the square of G/Y itself. At the beginning of the process, for reasonable values of the key parameters, the welfare loss will be small enough and the growth of G/Y may look similar to an exponential process. However, after some time the growth process will be countervailed by the welfare loss, and the logistic trajectory will become more apparent. The curve has some well-known properties: a) let us denote R = G/Y; dR/dt > 0 only if h/k > R; with h > 0, k > 0 and xed parameters, their ratio is then the upper limit of R. When R = h/k, dR/dt = 0 and there is no more increase of G/Y ratio, i.e. the two variables increase at the same speed. At the beginning of the process, near t = 0 and R = 0, then dR/dt = hR, as in the exponential growth; b) the integral value of the function is Rt = K/(1 + Ceht ) where K = h/k, C is a constant term; c) the proportional derivative is a simple linear function: d log R/dt = (1/R)(h kR)R = h kR; (12) (11)

a) the ex point occurs at point t = (1/h)log C, and there the G/Y ratio is at half of its maximum value: i.e. K/2. At that point dR/dt reaches its maximum value or hK/4. The logistic process over time generates an S-shaped curve that at the beginning is similar to an exponential: the process may be seen as the prevalence of expanding forces, while at a certain point the brake given by countervailing forces reverses the process. Thus the ex-point divides the trajectory of government expenditures into two patterns: the rst part is convex to the horizontal axis, the later part is concave. In the rst part the growth is accelerating, in the second part it is still positive, but decelerating. At a certain point the process converges to a steady state. If the actual process governing G/Y is a logistic one, the data should lie on an S-curve. In fact, the empirical trajectories we observe are compatible with logistic growth, as we shall show below. Other sigmoid curves may t well with data, but at this stage we are not interested mainly in empirical analysis, but in a preliminary test of the theory.

371 6. The public choice perspective In this section we discuss the theoretical relationship of our framework with the public choice approach and we briey consider also differences and similarities with other contributions mentioned in Section 2. When we consider the public choice perspective, there is a clear relationship between our approach and the median voter literature, however with a crucial innovation. Mueller (2003, p. 506 and ff) summarizes the explanations for the size of and growth and government and considers as a rst benchmark framework the following set of hypotheses: a) government function is to provide public goods and alleviate externalities b) voters have a demand for publicly provided goods which is function of individuals income, the relative price of public to private goods, and perhaps other taste variables c) there is a majority rule, and voting is directly and uniquely on the level of government issues. Then, the median voter theorem implies that government expenditures are a function of the characteristics of the median voter. This lends itself to the following testable equation, the archetypal version of a wide empirical literature: lnG = a + lnPg + lnYm + lnZ + (13)

where G is government provision of public goods, Pg is the tax price of this provision, a is a constant term, Ym is the income of the median voter, Z a vector of taste explanatory variables, and an error term. In this framework the growth of government arises because the demand for public gods is inelastic to its own price and Pg grows more than the price of private goods (we can interpret here the tax price as a real variable, i.e. relative to the price index of private goods, more on this below). Alternatively, the demand for public goods is elastic and the relative prices of public goods are falling. Moreover, income may increase over time and income-elasticity of G may be grater than 1. Eventually tastes may change over time and favour public provision. If we ignore the change of tastes over time and consequently drop Z (this would be considered as a xed effect in a cross-country empirical analysis), take the time derivative of (13), rearrange terms and consider the expected minus sign for the tax price parameter, we have the following equation: dlnG/dt = (dlnYm /dt) (dlnPg /dt) (14)

372 There is a clear relationship between equation (14) and our equation (7). The dynamic version of the median voter equation (with xed tastes) says that G increases with income of the median voter and decreases with the relative tax price. In our framework G/Y increases with the growth rate of income y, and our parameter is multiplicative in the income elasticity of G and a country or group specic coefcient . Our second term in the left side of (7) captures the change over time of the tax price because of the increasing excess burden of taxation (without any special assumptions about the relative change of direct costs of public provision, as in the Baumol effect). In other words, we say that there is a quadratic impact of the tax price variable because of distortionary taxation. In fact, our equation (7), having in mind (2) and dividing both sides by G/Y can be re-arranged as follows, dlog(G/Y)/dt = a( 1)y (/2)(T/Y). This rearrangement says that the proportional derivative of the ratio of public expenditure and income is a function of the growth rate of the economy and of the ratio of tax on income. In our framework, we consider , as parameters that reect institutional factors, e.g. the time needed by the political system to react to welfare changes of voters or the welfare weights attached by the social planner to different types. In fact, while in the traditional median voter empirical literature the and parameters usually cannot discriminate between the institutional factors and the median voter demand drivers, in principle our approach (if appropriate data are available) would allow for a separate estimation of income and price elasticities and the coefcients for institutional reaction). This point is best considered with an ideal experiment. Suppose you have two countries, with a similar starting point in G/Y, the same rate of growth of the income of the median voter, same tastes, and the same relative price elasticity of demand of public goods. Our framework predicts that any difference in the dynamics of G/Y could be traced back to the capacity of the political system to translates in majority decisions and actual government spending behaviours the demand of the median voter. The actual shape of the sigmoid curve generated by the dynamic process can be inuenced by different starting points, different political systems, and of course different growth of income and changes in the relevant elasticities. To test this we would need a set of institutional variables, for example related to voting systems (and their change over time), something that goes far beyond the scope of the present paper.

373 Obviously, our framework is extremely simple and parsimonious in terms of variables it uses, but it easy to think how to enrich it with additional insights from the public choice literature. We very briey mention here some of these other variables, for future research: a) tastes: candidate explanatory variables include openness of the economy, volatility of employment, population density. All these would appear as additional parameters in the model. b) the Baumol effect: if this effect actually exists (something that is quite dubious) it would appear in our framework as an additional brake in terms of the tax price, because the tax price would increase over time and ceteris paribus it will further constrain the underlying logistic process c) interest groups and bureaucracy: the impact of lobbies, unions, or other interest groups, and of bureaucrats in the long run growth of public expenditure calls for moving away from the simple median voter hypothesis and needs an in depth analysis of the structure of the society and governmental machinery. This is not easily accomodated in our framework and any such effect will be captured by the institutional parameters (and the error term). d) scale economies: this candidate rationale for increasing public expenditures is not easily accomodated in an highly aggregated model, neither ir commands strong empirical support: see Reiter and Weichenrieder (1997) for a survey of the empirical literature and a restatement of the theoretical case. e) for similar aggregation reasons, our model is not particularly apt to consider redistributional aspects of the political process. In fact we see our model as broadly consistent with the intuition behind the median voter framework: government provides (superior) goods for which a demand exists, however there is a social cost for this process, because of (distortionary) taxation and other institutional features. The median voter (crudely proxied by the average taxpayer in empirical analysis) calculates its own costs and benets and, albeit imperfectly, with delays, and national specicities, is able to reach a dynamic equilibrium. Having said this, we move to a preliminary test of our framework, in which we focus on the simple dynamic process described in Section 5.

374 7. Data Secular data on public expenditures and GDP are difcult to nd, and we need to rely on previous work by economic historians, rather than on ofcial sources. The sources of data we have selected are the following one:4 France: Andr, Delorme (1984), Flora (1983), Maddison (19891995). Germany: Andr, Delorme (1979), Flora (1983), and Maddison (1989 1995). Italy: Brosio, Marchese (1981), Luzzati, Portesi (1984), Tanzi, Schuknecht (1995). United Kingdom: Maddison (19891995), Middleton (1996). United States: Andr, Delorme (1979), Maddison (19891995), Musgrave (1995). The available time series are incomplete and reveal wide variance among different sources for the same country and the same time period. However, in what follows we rely on the evidence already made available by the cited experts in quantitative economic history. Because data differ among authors, and because there are important data issues that should be solved, e.g. whether to exclude war years from time series, and how to control for volatility of Y, we decided that at this stage the empirical analysis should be as simple as possible. Thus, what follows should be regarded as a preliminary test. We repeat the test on different sources for each individual country, rather than focus on a more sophisticated estimation procedure on one specic data source. As we shall see, this approach allows us to test the theory in a simple way, in spite of data variations across sources. 8. Estimation procedure The empirical curve to be estimated can be linearized in the following way: We consider the function Rt = R0 + K/(1 + Ceht ) where R0 is the low asymptote and R, K, C > 0. It is convenient to consider the following transformations: (K + R0 Rt )/(Rt R0 ) = Ceht loge (K + R0 Rt )/(Rt R0 ) = loge Ceht loge (K + R0 Rt )/(Rt R0 ) = loge C + ht Let us now dene loge (K + R0 Rt )/(Rt R0 ) = Zt . This new variable, a transformation of Rt is linearly related to t through the parameters C and h:

375

Zt = a + bt with a = loge C and h = b. We need now to determine two values R and K that constitute the extremes of the above mentioned Z variable and are compatible with the linear relationship between the N observation pairs Zt and t. Thus, xed R and K we can estimate the other two parameters (C e h) with an OLS estimate of a and b. We can also use non-linear regression to estimate the parameters of the logistic function. In order to start the non-linear estimation algorithm, we must have initial values for the parameter. Unfortunately the results of non-linear estimations often depend on having good starting value for the parameters. Poor initial values can result in non-convergence, a local rather than global solution or a physically impossible solution. There are several methods for obtaining starting values. Sometimes a linear model can be derived and linear regression can then be used to obtain initial values. This is here the case: we specify the results of the linear regression described above as starting values in non-linear regression. In non-linear regression, just as in linear regression, we choose values for the parameters so that the sum of squared residuals is a minimum. There is not, however, a closed solution. We must solve for the values iteratively. There are several algorithms for the estimation of non-linear models. With SPSS we can use a Levenberg-Marquardt algorithm by default or we can try the sequential quadratic programming algorithm. For a particular problem, one algorithm may perform better than the other. Additional options are available in SPSS, when the sequential quadratic programming algorithm is used. We can supply linear and non-linear constraints for the values of the parameter estimates and we can specify our own loss function (by default, the loss function that is minimised is the sum of the squared residuals). In our estimates we used linear constraint only when the upper asymptote estimated by the Levenberg-Marquardt algorithm was greater than 100 or when the low asymptote was < 0 or greater than the value minimum of the series. For a non-linear model, the tests used for linear models are not appropriate. In this situation, the residual mean square is not an unbiased estimate of the error variance, even if the model is correct. For practical purpose we can still compare the residual variance with an estimate of the total variance, but the usual F statistics cannot be used for testing hypotheses. The R2 , however, can be applied in its conventional sense to a non-linear regression. It may be interpreted as the proportion of the total variation of the dependent variables around its mean that is explained by the tted model. For non-linear models,

376 its value can be negative if the selected model ts worse than the mean. In the case of non-linear regression it is not possible to obtain exact condence interval for each of the parameter. Instead, we must rely on asymptotic approximations. For a wider discussion of estimating techniques of the logistic curve see Oliver (1966, 1969, 1982), Nelder (1961), Pearl and Reed (1977), Ratkowsky (1986, 1993).

9. Main results Our main results for the countries we considered are the following ones: 9.1. United States Table 2 shows data and regression results for the data by Delorme (1979), Musgrave (1995), and Maddison (19841995). The number of observations is limited, respectively 19, 14, 11.5 The Delorme, Maddison, Musgrave sets of data are generally similar in levels: for example for 1902 the range of G/Y is between 6.8% and 7.9%; for 1913 it is 8-8.5%; for 1938 the range is 17.8-19.8; for 1970 it is%; 32.234.2. The latest year we consider is 1992, at 38.5% by Maddison (obviously we have recent data from other ofcial sources, but as said we do not want here to integrate various sources). The rst observation we have is for 1890, by Musgrave, where G/Y was 7.1%. The overall picture seems clear enough: G/Y increased by 3 times between the end of the 19th century and the 30s, then in the following 5060 years the increase was much less, around 2 times. The overall t (adjusted R-squared) of a logistic curve is only marginally better than an exponential one. However, the extrapolation by an exponential process gives a completely wrong trend for the most recent period, while the logistic captures well the history of G/Y at the end of the 20th century. Most interesting, the ex-point year in our estimates is 1944 for both Delorme and Musgrave, and 1954 for Maddison. The change in the rate of increase of G/Y greatly anticipates the perception of a change of attitudes towards public spending in the 80s. The value of the time coefcient h (linear regression) is very similar for Musgrave and Maddison: around 0.074, and 0.066 for Delorme series. The range of the same coefcients estimated with the non-linear regression is between 0.047 and 0.054. These are rather high values, which imply that, contrary to popular views, the U.S. have a history of sustained increase in public expenditures, and if their current level of G/Y is much lower than

377

Figure 2. United States. Logistic and exponential functions Musgrave G/Y data

in Europe, this is due to the low starting level and not to a quite different pattern of growth. 9.2. France We considered three data sources: Delorme (1984), Flora (1983), and Maddison (198495). We also have data from Brosio (1993) that are usually consistent with Delorme. Table 3 shows selected estimation results. Around 1870, G/Y in France was in the region of 11%, and by the end of the 19th Century it exceeded 14%: two times the corresponding level of the U.S.. The increase was around 2 times between 1870 and the early 1930s

378
Table 2. United States. Output regression summary Delorme, Maddison, Musgrave data

(while as said it was more than 3 times in the U.S.), then it more than doubled again until it nearly stabilised around 50% since the 1960s. As in the U.S., the dynamic stabilization in France greatly precedes the hot debates about budgetary issues in the 1980s. Unfortunately the three series of data we use are very different as for the number of observations: 55 for Delorme, 23 for Flora, just 10 for Maddison. This makes the estimations difcult to be compared. Moreover, there are quite different values for in-

379

Figure 3. France. Logistic and exponential functions Delorme G/Y% data

dividual years, probably due to different denitions of G or Y considered by the authors (but for some years the G/Y ratios are surprisingly similar). The t for a logistic (linear and non-linear) is slightly better than with an exponential with Delorme and Flora data, but not with Maddison (only the non linear estimation gives an R2 higher than the exponential function). Flex points in the estimated logistic are very different, from 1924 with Flora data (that stop much earlier than the other two series), to 1960 with Maddison data, with Delorme in between, 1941. As for the U.S., the superiority of the logistic as compared with the exponential is best seen looking at most recent years. When extrapolated to 2000,

380
Table 3. France. Output regression summary Delorme, Flora, Maddison data

the exponential widely overstates actual G/Y, while the logistic is very near to observed recent data for France (from ofcial sources). The estimates for h coefcient are in the region of 0.05-0.08.

381

Figure 4. Italy. Logistic and exponential functions Brosio/Others G/Y% data

9.3. Italy We considered here a number of sources including Brosio, Marchese (1986), Luzzati, Portesi (1984) and Oecd (2001), but we decided to combine different time series in a unique data set . According to this series, G/Y was around 14% in 1870, thus a starting point higher than France. The doubling time was around 60 years (in the early 1930s), than G/Y doubled again at the end of the 1970s. The t of the logistic is marginally better than the exponential, and the extrapolation test is less clear cut than in the previous cases. In the case of Italy data may support either an exponential

382
Table 4. Italy. Output regression summary Brosio/others series

process or a logistic one: if we accept the second one, it is interesting to remark that the ex point year is in the 60s, around 20 years later than France or U.S. However, the h coefcient is just 0.03, half than the U.S. and much less than France; Italy seems to be more a late comer than a big spender, starting from very high level of government expenditures at its birth as a state.

383

Figure 5. Germany. Logistic and exponential functions Flora G/Y% data

9.4. Germany The sources we have selected are Maddison, Flora, and Delorme. The number of observations for the rst author are just 9, but they are consistent more or less with the other authors. Flora has 39 years, Delorme 34, but these two authors end their series in 1974, while the last year for Maddison is 1985. There are also some surprising differences for individual years. The history these authors tell is the following one: G/Y in Germany was between 8-10% in 1881 (thus less than Italy or France, but more than in the U.S.). In the early 1920s the ratio doubled (in 50 years), then doubled again in the early 1970s (again in 50 years). In the following decades the increase was limited.

384
Table 5. Germany. Output regression summary Delorme, Flora, Maddison data

A logistic interpretation of this story is marginally better than an exponential one in all three cases in terms of overall t, however the extrapolation test to 2000 with an exponential implies that G/Y would exceed 60%, while the logistic is closer to observation (around 50%) for Maddison and Flora, while it underscores the target with Delorme data. The range of estimates for h is between 0.03 and 0.04. Again this is less than for the U.S., and in between the corresponding values for Italy and France.

385 9.5. United Kingdom The sources we have selected for the United Kingdom are Maddison (11 observations) and Middleton (23 observations). We have considered also Peacock, Galloway, Flora. Peacock data (45 years) end in 1955, and contain a high number of war years. As a result that series is of limited use. The history of G/Y in the U.K. that emerges from our sources is the following one: between 18701895 the ratio was between 910%, half way the case of the U.S. and the case of Germany. The increase was modest until the First World War (in 1913 G/Y was still around 12%). However in the 1920s the ratio doubled (offering some justication to the well-known Peacock-Wiseman argument), and it doubled again in the following 50 years, through the 1970s. Then the increase was limited. The overall t of a logistic is better than an exponential with Maddison data (both linear and non-linear estimation), but equal with the Middleton data in the case of linear regression. However, the extrapolation with an exponential to year 2000 largely overshoots the target, while the logistic is much nearer to it. The ex point for a logistic process is 1954 for Maddison and 1947 for Middleton, and the h coefcients are between 0.03 and 0.058.

10. Conclusions This paper has offered a new theory of the growth of government consistent with observation. We have suggested to look at long term trends of public expenditures as a simple dynamic process arising from the combination of two forces: the rst one is Wagners law, that would imply an ever increasing ratio between public and expenditures and national income, and the second is Pigous conjecture that the increase of taxation generates an ever increasing deadweight loss. This framework is new, and more comprehensive, than popular empirical approaches to testing Wagners law. The combined action of two forces, the demand for public expenditures and the excess burden of taxation generates an S-shaped trajectory, and this is what in fact we can observe in G/Y trends for several countries in the last 100150 years. We have suggested that because the excess burden of taxation is quadratic in G/Y, when we assume a long run balanced budget, the resulting differential non-linear equation can be integrated as a logistic process. Modelling the dynamics of government as a logistic process has a number of very attractive features. First, we focus on logarithmic derivatives of public expenditures and national income instead that on levels: this allows easier comparisons of countries at different stages of their development: instead of G/Y the key

386

Figure 6. United Kingdom. Logistic and exponential functions Middleton G/Y data

variable here is h = g y, or the difference in growth rates of government and the economy. Second, h can be seen also as the product of growth rate and income elasticity of public expenditures minus one: thus for any positive value of g, income elasticity should be more than one to generate positive h, which is consistent with the intuition by Wagner. More interesting however is that for a constant h to hold, it is possible to combine different offsetting changes of growth and income elasticity. However the resulting process is virtually exponential and this is implausible both on empirical and theoretical grounds.

387
Table 6. United Kingdom. Output regression summary Maddison, Middleton data

Third, as government becomes bigger and bigger, scal resistance increases, and this acts as a brake to Wagners law: the importance of this factor may vary across countries, as the income elasticity of public expenditures varies, but despite different country specicity, it is helpful to model a common underlying process. If scal resistance depends on the dead-weight loss of taxation, under the assumption of a balanced budget, Wagners law is replaced

388 by a more complex pattern: in fact Wagners laws is only approximately valid in the earlier stages, where the rst part of a logistic curve is similar to an exponential one, then the process changes around a ex-point, G/Y increase slows down and then it stabilises. Fourth, in this context along with parameter h, depending upon the income elasticity of publicly provided goods, the other key parameter, is k, depending upon the price elasticity of private goods. The two parameters in a logistic process have some nice properties that are easy to interpret: h/k is the upper limit of G/Y, thus if we can estimate them (and a set of country specic variables) we can predict where and when G/Y increase stops. Five, ex-point occurs at t = (1/h)logC, and there G/Y is half its upper limit. The study of ex-points allows us to look at the economic history of government under a most interesting angle: around the ex-point year Wagners law is reversed and Pigous effect starts to prevail. With countryspecic information on the circumstances of this reversal, our interpretation of economic history changes, because the ex-point occurs largely before the perception of a slowdown looking at variables in levels. Sixth, we have shown the relationship of our framework with the standard median voter demand equation: when tastes are xed over time, there is clear relationship between our theory and the median voter framework, however we do not need to assume that the relative tax price increases because of the Baumol effect, of because of increasing bureaucratic inefciency. Distortionary taxation is all what we need to generate brake to public spending, under the assumption that the median voter-taxpayer is able to calculate costs and benets, votes rationally, and the political system dynamically adjusts to these preferences. While we are condent that the theoretical framework is well grounded and goes beyond all the Wagner-testing literature, our empirical ndings are just exploratory: however, we think they are reassuring and interesting per se. Despite the obvious difculty of dealing with rather conjectural secular expenditure data, those available by quantitative economic history research, and some technical problems with estimation in this framework, we can conrm that for most data sets the logistic interpretation offers a better t than an exponential process. Moreover, it offers new insights. Table 7 and related gures presents selected results among the many combinations of sources/countries/estimation techniques (more than one hundred) we have tested.6 We turn now to comment some of these results as an illustration of the interest of the new approach. The U.S. started with a very low G/Y ratio as compared to other countries at the same time. This clearly reects the youth of government in that country, the nature of the state and the early aversion to widespread public interven-

389
Table 7. Selected results summary

tion. However, with h = 0, 05 it shows a rapid dynamics. The turning point was immediately after World War II, and a virtual steady state compatible with the estimated trajectory would x the maximum of G/Y around 37% 43% (according the source of data and the estimation approach). This is below observed data in most recent data, also because of the actual high growth rate of the U.S. economy in the 1990s: but it also would imply that at the end of the 21st century G/Y in the U.S. would virtually converge to values more similar to those prevailing now in Europe. In other words, the apparently smaller government of the U.S. as compared e.g. with Germany, is partially a matter of time of convergence, than of a totally different dynamic process: in fact the h parameter is similar for the U.S. and Germany. The United Kingdom, in spite of two decades of governments committed to public expenditures restraint, tends to converge to a G/Y ratio similar to Germany. France conrms of course to be a big spender, its long-term trajectory virtually converges to G/Y=58%, and the h parameter is considerably higher than other countries. Italy is also a big spender, and is the only country where an exponential process can challenge a logistic process as empirical trajectory: but, apart from the correction in recent years, when the admission to the European Monetary Union forced some scal discipline, it is also the country where the ratio between public debt and income increased more in the last decades. This can be seen rather as a conrmation of our theory: when the budget is not balanced for many years, the excess burden of taxation is shifted away to future generations and offers limited brake to the demand for ever growing public expenditures.

390 We conclude by some suggestions for further research. The framework we have proposed seems to be a convenient way to model long-term trajectories of G/Y, however it is based on some assumptions that should be removed in a more elaborate framework. First, population is not stable, and it would be interesting to include in the analysis a model of secular demographic processes. Originally, the logistic was invented in order to solve this problem, but in this paper we have not tried to integrate demographic factors in our framework. Moreover, growth of income as a long-term process should also be modelled. And there may be interactions as well among growth of income, population, and public expenditure.Thus a possible extension of our research is a dynamic system of simultaneous differential equations for these secular trends. This is a difcult but challenging task for future research.

Notes
1. The interpretation of Wagners Law is rather controversial and has a long history, see Peacok, Scott (2000). They show convincingly that Wagner included public enterprises in his denition of government. In Wagners own words: Der Staat speziell als Wirthschaft zur Frsorge der Bevlkerung mit gewissen Gtern, besonders Gemeingtern fr gewisse Bedrfnisse aufgefasst, wird dabei absolut immer wichtiger fr die Volkswirthschaft und fr die Einzelnen. Aber auch seine relative Bedeutung steigt . . . (Wagner, 1893: 893). 2. See Feldstein (1997). 3. The consideration of exible production prices and of many different consumers adds complexity to the evaluation of the excess burden of taxation. The literature on the subject offers different formulas to deal with this reacher framework of analysis. Unfortunately generally it is showed that E depends upon the initial income distribution among agents and we need then a social welfare function to consider this. We cannot dwell here on these additional difculties, important as they are: we must conne ourselves to a much simpler framework. Other difculties that we shall ignore at this stage arise from the distinction between different types of taxes (e.g. indirect versus direct ones) and between types of public expenditures (public consumption, investment and transfers). Both income taxes and income (or goods ) subsidies have distortionary effects and in principle generate an excess burden, but the welfare impact may be different. It seems worthwhile to avoid disaggregation of taxes and expenditures by categories at this very preliminary stage of analysis (and because no reliable empirical disaggregation of secular expenditure data are available). Readers should be aware of this point. 4. The data set used for this paper is available with the authors and will be delivered on request. 5. A larger number of observation (33) is available for U.S. Stat (U.S. Bureau of Census, 1975), however that set of data stops at 1970 and includes some strong outliers for Second World War years, with a bad t for either an exponential or a logistic process (however in table A1 we report U.S. STAT data for the years included in the other sources). 6. Available with the authors.

391 References

Andr, C. and Delorme, R. (1978). Lvolution de longue priode des dpenses publiques en France: 18721971. In H.C. Recktenwald (Ed.), Tendances long terme du secteur public, XXXIIe Session, Congrs dEdimbourg 1976. Paris: ditions Cujas. Andr, C. and Delorme, R. (1984). Materiaux pour une comparaison internationale des dpenses publiques en longue priode: Le cas de six pays industrialiss. In Statistiques & Etudes Financires, Ministre de lEconomie, des Finances et du Budget, Paris. Auerbach, A. and Feldstein, M. (Eds) (1990). Handbook of public economics, vol I. AmsterA dam: NHPC. Baumol, W. (1967). The macroeconomics of unbalanced growth: the anatomy of urban owsis . American Economic Review 57 : 415426. Baumol, W., Baty Blackman, S.E. and Woltt, E. (1985). Unbalanced growth revisited: Asymptotic stagnancy and new evidence. American Economic Review 75: 806817. Biswal, B., Dhawan, U. and Lee, H.J. (1999). Testing Wagner versus Keynes using disaggregated public expenditure data from Canada. Applied Economics 31: 12931298. Borcherding, T.E. (1977). One hundred years of public spending, 18701970. In T.E. Borcherding, (Ed.), Budgets and bureaucrats: The sources of government growth, 1944. Durham, NC: Duke University Press. Borcherding, T.E. (1977). The sources of growth of public expenditures in the United States, 19021970. In T.E. Borcherding (Ed.), Budgets and bureaucrats: The sources of government growth, 4570. Durham, NC: Duke University Press. Brosio, G. (1993). Economia e nanza pubblica. Roma: NIS. Brosio, G. and Marchese, C. (1986). Il potere di spendere. Storia ed economia della spesa pubblica dallunicazione ad oggi. Bologna: Il Mulino. Buchanan, J. and Musgrave, R.A. (1999). Public nance and public choice: Two contrasting visions of the state. Cambridge and London: MIT Press. Delorme, R. and Andr, Ch. (1979). tude comparative de lvolution des dpenses publiques en longue priode dans six pays industrialiss occidentaux. Paris: CEPREMAP. Delorme, R. and Andr, Ch. (1983). Ltat et lconomie. Un essai dexplication de lvolution des dpenses publiques en France (18701980). Paris: dition du Seuil. Ercolani, P. (1969). Documentazione statistica di base in Fu (a cura di) Lo sviluppo economico in Italia, Vol. III-Studi di settore e documentazione di base. Milano: Franco Angeli Editore. Feinstein, C.H. (1972). National income, expenditure and output of the United Kingdom, 18551965. Cambridge: Cambridge University Press. Feldstein, M. (1997). How big should government be. National Tax Journal 50: 197213. Flora, P. Kraus F., Pfenninf W. (1983). State economy and society in western Europe 1815 1975, Volume I: The growth of mass democracies and welfare states. Frankfurt: Cam pus Verlag. Flster, S. and Henrekson, M. (2001). Growth effects of government expenditure and taxation in rich countries. European Economic Review 45: 15011520. Henrekson, M. (1990). An economic analysis of Swedish government expenditure. Gothenburg: FIEF Trade Union Institute for Economic Research. Henrekson, M. (1993). Wagners Law: A spurious relationship? Public Finance 48: 406415. Luzzati, E. and Portesi, R. (1984). La spesa pubblica in AAVV. Storia della societ italiana dallunit ad oggi, Vol. 9: Cassese S. Lamministrazione centrale. Torino: UTET.

392
Maddison, A. (1984). Origins and impact of the welfare state 18831983. Banca Nazionale del Lavoro. Quarterly Review, March. Maddison, A. (1989). The world economy in the 20th century. Paris: OECD. Maddison, A. (1991). Dynamic forces in capitalist development. Oxford plce: Oxford University Press. Maddison, A. (1995). Monitoring the world economy, 18201992. Paris: OECD. Middleton, R. (1996). Government versus the market: The growth of the public sector, economic management and British economic performance, c. 18901979. Cheltenham U.K.: Edward Elgar Published Limited. Mueller, D.C. (2003). Public choice III. Cambridge: Cambridge University Press. Musgrave, R.A. (1969). Fiscal systems. New Haven/London: Yale University Press. Musgrave, R.A. (1995). Teoria delle crisi scali. Un saggio di sociologia scale. In R. Musgrave (Ed.), Finanza pubblica, equit, democrazia. Bologna: Il Mulino, 1995. Musgrave, R.A. and Culbertson, J.M. (1953). The growth of public expenditures in the United States, 18901948. National Tax Journal 6: 000000. Nelder, J.A. (1961). The tting of a generalisation of the logistic curve. Biometrics 17: 89110. Ng, Y.-K. (2000). The optimal size of public spending and the distortionary cost of taxation. National Tax Journal 53: 253272. OECD (2001). Economic outlook, n. 70 (June). Paris. Oliver, F.R. (1964). Methods of estimating the logistic growth function. Applied Statistics 13: 5766. Oliver, F.R. (1966). Aspects of maximum likelihood estimation of the logistic growth function. American Statistical Association Journal 78: 697705. Oliver, F.R. (1969). Another generalisation of the logistic growth function. Econometrica 37: 144147. Oliver, F.R. (1982). Notes on the logistic curve for human populations. Journal of Royal Statistical Society A 145 3: 359363. Peacock, A. and Scott, A. (2000). The curious attraction of Wagners Law. Public Choice 102: 117000. Peacock, A. and Wiseman, J. (1961). The growth of public expenditure in the United Kingdom. Princeton: Princeton University Press. Peacock, A. and Scott, A. (2000). The curious attraction of Wagners law. Public Choice 102: 117. Pearl, R. and Reed, L.J. (1977). On the rate of growth of the population of the United States since 1790 and its mathematical representation. In D. Smith and N. Keytz (Eds.), Mathematical demography. Selected Papers. Berlin plce: Springer Verlag. Pigou, A.C. (1947). A study in public nance. London: MacMillan. Ratkowsky, D.A. (1993). Handbook of non-linear regression models. New York/Basel: Marcel Dekker Inc. Ratkowsky, D.A. and Reedy, T.J. (1986). Choosing near-linear parameters in the fourparameter logistic model for radioligand and related assays. Biometrics 42: 575582. Recktenwald, H.C. (Ed.) (1978). The public sector in transition. In H.C. Recktenwald (Ed.), Tendances long terme du secteur public (Secular Trends of the Public Sector), XXXIIe Session, Congrs dEdimbourg 1976. Paris: ditions Cujas. Reiter, M. and A.J. Weichenrieder (1997). Are public goods public? A critical survey of the demand estimates for local public services (with Michael Reiter). Finanzarchiv 54: 374 408.

393
Rodrik, D. (1998). Why do more open economies have bigger government. Journal of Political Economy 106: 9971032. Schremmer, D.E. (1989). Taxation and public nance: England, France, and Germany. In P. Mathias and S. Pollard (Eds.), The Cambridge economic history of Europe, Vol VIII. Cambridge: Cambridge University Press. Tanzi, V. and Schuknecht, L. (2000). Public spending in the 20th century: A global perspective. Cambridge: Cambridge University Press. Tridimas, G. (2001). The economics and politics of the structure of public expenditure. Public Choice 106: 299316. Tridimas, G. (1993). Public expenditure growth in Greece and Portugal: Wagners law and beyond. Applied Economics 25: 125134. U.S. Bureau of Census (1975). Historical statistics of the United States, colonial times to 1970. Washington DC: U.S. Government Printing Ofce. Verhulst, P. (1977). A note on the law of population growth, 1838. In D. Smith and N. Keytz (Eds.), Mathematical demography. Selected papers. Berlin plce: Springer Verlag. Volterra, V. (1931). Theory of functionals and of integro-differential equations. London/Glasgow: Blackie & Son Limited. Wagner, A. (18921893). Grundlegung der politischen Oekonomie: 1. Theil, Grundlagen der Volkswirthschaft, 2 vols. Leipzig: Winter. Wagner, A. (1894). Grundlegung der politischen Oekonomie: 2. Theil, Volkswirthschaft und Recht, besonders Vermgensrecht, 5 vols. Leipzig: Winter; French translation: Les fondaments de lconomie politique. Paris: Girard &Brire, 1912.