Wagner ’ s Law versus Keynesian Hypothesis : Evidence from pre-WWII Greece

With data of over a century, 1833-1938, this paper attempts, for the first time, to analyze the causal relationship between income and government spending in the Greek economy for such a long period; that is, to gain some insight into Wagner and Keynesian Hypotheses. The time period of the analysis represents a period of growth, industrialization and modernization of the economy, conditions which are conducive to Wagner’s Law but also to the Keynesian Hypothesis. The empirical analysis resorts to Autoregressive Distributed Lag (ARDL) Cointegration method and tests for the presence of possible structural breaks. The results reveal a positive and statistically significant long run causal effect running from economic performance towards the public size giving support to Wagner’s Law in Greece, whereas for the Keynesian hypothesis some doubts arise for specific time sub-periods.

The theoretical and empirical relationship between government expenditures and economic growth turned to be an intense subject of analysis and controversy within the economic literature.It was raised as far back as in late 1800s and since then many empirical attempts have been proposed in order to explore the flow of causality between government size and economic development.Adolph Wagner (1893) was among the first who observed the overtime increasing tendency of public spending and the concomitant Wagner's Law, as presented in economic literature since then, states that economic performance has a fundamental positive impact on public sector's growth.However, an equally important strand in economic policy literature, rooted in Keynesian principles, argues for the opposite; that is public spending can form an exogenous tool of economic policy to enhance growth through its multiple effect on aggregate demand.As a result, many economists following the Keynesian tradition argue that an economy needs a Keynesian-type fiscal stimulus to be given temporarily in periods of recession by an active government (Phillip Arestis 2011), whereas, others, belonging to mainstream economics, argue that the government has to be small and not to replace the market mechanism.The evolution of the debate about the role of government in a society is presented in Daniel Yergin and Joseph Stanislaw (2002).
For the first time, the present paper attempts to gain some insight into the causal relationship between economic performance and government size in Greece for such a long period of over a century, 1833-1938.In so doing, we explore the dynamics of this highly debated relation in a developing country, since Greece during this period was gradually transformed into a modern state.It is worth noting that the majority of relevant studies refer only to Postwar Greece and use econometric techniques that may give rise to spurious results about the strength of the long-run equilibrium relationship between government expenditure and national income.Instead, in the context of our empirical analysis the use of Autoregressive Distributed Lag (ARDL) approach to cointegrated analysis (Hashem H. Pesaran and Yongcheol Shin 1999) and time series data from 1833-1938 allows us to arrive to more reliable conclusions.The major advantage of the ARDL method is that by employing an appropriate augmentation it avoids problems of serial correlation and of endogeneity experienced by other cointegration methods.In addition, this method avoids pretesting of the order of integration, which is associated with other cointegration techniques.Moreover, the present effort checks for structural breaks in the data set, since the time period is very long.Consequently, the results derived from our empirical analysis are expected to be more reliable and thus to provide us with a better understanding of the long run causal relation between government size and economic performance in Greece.
The remainder of the paper is organized as follows: Section 1 briefly presents the theoretical underpinnings of Wagner's Law and reviews the basic literature about the causal relation between government size and economic performance.Ιn Section 2, the ARDL approach to cointegration is presented; the section continues with the description of data and the presentation and discussion of the empirical findings.Finally, the concluding remarks and the proposals for future research are reported in Section 3.

Wagner's Law in Economic Literature
Through the years, several attempts have been made to explain the growth in public sector.Several propositions were put forward either by economists or economic historians.Walt W. Rostow (1960) suggested that the increase in public expenditure might be related to the pattern of economic growth and development of the various societies whereas Alan T. Peacock and Jack Wiseman (1961) argue that social crises cause the increase in public sector expenditure.William J. Baumol (1967) argues that the public sector, by being less productive than the private, is doomed to increase as the private sector increases, whereas Morris Beck (1976) suggests that the "relative price effect", meaning that public sector unit costs increase faster than those in private sector, is the cause for the rise in public sector expenditure.Nevertheless, Wagner's Law is among the first attempts made to explain the increase in public sector and mathematically can be formulated as: where G refers to the size of the public sector, Y stands for the level of economic performance, and t is for time.The various empirical studies use different indexes to approximate the size of the government, (i.e. the share of government spending over PANOECONOMICUS, 2013, 4, pp. 457-472 GNP, the growth rate of government spending, the per capita public spending, etc.) and the level of economic activity, (i.e. the GNP per capita, the growth rate of GNP, etc).
However, different interpretations of the hypothesis under testing inevitably led to various models modifications.For instance, Richard A. Musgrave (1969) uses total and partial public spending and GNP in logarithms as the dependent and independent variables, respectively, in an attempt to estimate the income elasticity of "public goods".Arthur J. Mann (1980), in his model for Mexico, estimates the income elasticity by taking GNP per capita as the independent variable and the share of public spending to GNP as the dependent variable.
Wagner's Law has been empirically asserted over the years and, not surprisingly, the reported results are highly diverse and conflicting.For example several studies -Subrahmanyam Ganti and Bharat R. Kolluri (1979); Rati Ram (1987) for a sample of 115 countries; Les Oxley (1994) for Britain; John Thornton (1999) for the 19 th century Europe; Kolluri, Michael J. Panik, and Mahmoud S. Wahab (2000) for the G-7 countries; Nikolaos Dritsakis and Antinios Adamopoulos (2004) for Greecehave presented results in favour of Wagner's Law.In contrast, the studies by Magnus Henrekson (1993) for Sweden and by George Hondroyiannis and Evangelia Papapetrou (1995) for Greece have reported empirical evidence that contradict Wagner's hypothesis.Moreover, mixed results have been reported by Ram (1986) for a sample of 63 countries, Michael Chletsos and Christos Kollias (1997) for Greece, Erkin I. Bairam (1995) for USA and by Chang Tsangyao, Wenrong Liu, and Stecen B. Caudill (2004) for a sample of ten industrialized countries.
In addition, studies exploring Wagner's hypothesis versus the Keynesian one have also reported highly diverse and conflicting results.Panos C. Afxentiou and Apostolos Serletis (1996) for six European countries showed no evidence for both propositions, whereas the study by Biswal Bagala, Dhawan Uruashi, and Hooi-Yean Lee (1999) for Canada and the work by Constantinos P. Katrakilidis and Persefoni Tsaliki (2009) for Postwar Greece find supportive empirical evidence for both hypotheses.The studies by Mohammed A. Ansari, Daniel V. Gordon, and Christian Akuamoah (1997) for three African countries, Anisul M. Islam (2001) for the USA and by Abdulrazak F. Al-Faris (2002) for the Gulf Cooperation Council countries find evidence supporting the Wagner hypothesis but not the Keynesian view, whereas the study by John Loizides and George Vamvoukas (2005) for UK, Ireland and Greece reports mixed results.
However, the econometric techniques engaged in the aforementioned empirical analyses render their results problematic.Recently, progressive economic studies allow the use of cointegration and other advanced techniques in order to check; first, the over time tendency of public spending and gross national product; second, the hypothesis of a long-run relationship between public spending and gross national product; and third, the underline causality of this relationship (Islam 2001;Al-Faris 2002).

The ARDL Cointegration Approach
In the present analysis, we use the ARDL approach to cointegration which we consider it as a more appropriate technique since it presents certain advantages over other conventional cointegration procedures such as: the long and short-run parameters of the model are estimated simultaneously; all variables are assumed to be endogenous and the estimates obtained are unbiased and efficient since they avoid the problems that may arise due to serial correlation and endogeneity (Pesaran, Shin, and Richard J. Smith 2001); inability to test hypotheses on the estimated coefficients in the long-run associated with the Engle-Granger method are avoided; the need to establish the order of integration amongst the variables is obviated, which is equivalent to saying that the method can be applied even in the case where the variables are I(0) or I( 1) or a mixture of the two but not I(2).Moreover, in case that cointegration is detected, the resulting Error Correction Model (ECM) can be used for Granger noncausality tests, as suggested recently by Joao R. Faria and Miguel Leon-Ledesma (2003).An advantage of using a bi-variate approach to test for causality is that it allows to test, on the one hand, the short-run causality through the lagged differenced explanatory variables and, on the other hand, the long-run causality through the Error Correction (EC) term.As Clive W. J. Granger, Bwo-Nung Huang, and Chin-Wei Yang (2000) suggest a significant EC term implies long-run causality from the explanatory variables to the dependent variable.The ARDL approach to cointegration involves the estimation of the following conditional EC versions: where G is a proxy for government expenditure, Y is a proxy for economic performance and d denotes first differences.Based on the above equations, we perform a "bounds test" for the detection of a long-run causal relationship between the two variables.The test involves an F-test on the joint null hypothesis that the coefficients of the level variables are jointly equal to zero (Pesaran and Shin 1999;Pesaran, Shin, and Smith 2001).The statistic displays a non-standard F-distribution and its value depends on whether the variables are individually I(0) or I (1).Instead of the conventional critical values, the test employed involves two asymptotic critical value bounds, depending on whether the variables are I(0) or I(1) or a combination of both.If the test statistic exceeds their respective upper critical values, it may be argued then that there is evidence of a long-run relationship.If the test statistic falls below the lower critical values, we cannot reject the null hypothesis of no cointegration.However, if the test statistic lies between the bounds, inference about cointegration is inconclusive.
PANOECONOMICUS, 2013, 4, pp.457-472 The conditional long-run model can be produced from the reduced form solution of the above equations, when the first-differenced variables jointly equal zero.The long-run coefficients and the corresponding ECM are estimated through an appropriate ARDL specification.The lag structure for the ARDL specification is determined by the Akaike's Information Criterion (AIC).Autocorrelation has also been considered.

The LM Unit Root Test of Lee and Strazicich with Break
Junsoo Lee and Mark C. Strazicich (2004) developed two versions of the LM unit root test with one structural break, Model A known as the "crash model" and Model C known as the "crash-cum-growth model".Model A allows for a one-time change in the intercept under the alternative hypothesis and is described as ´, where D t = t T B + 1, and zero otherwise.Model C allows for a shift in intercept and change in trend slope under the alternative hypothesis and is described as ´, where DT t = t -T B for t > T B + 1, and zero otherwise.
The one break LM unit root test statistics according to the LM (score) principle are obtained from the following regression: , with y 1 and Z 1 the first observations of y t and Z t respectively.The unit root null hypothesis is described in (1) by  = 0 and the LM t-test is given by ; where = tstatistic for the null hypothesis  = 0.The augmented terms j t S   ~, j = 1,...k terms are included to correct for serial correlation.The value of k is determined by the general to specific search procedure.To endogenously determine the location of the break (T B ), the LM unit root searches for all possible break points for the minimum (the most negative) unit root t-test statistic as follows:

Data
The empirical analysis engages annual data of the Greek economy (Antonios Antoniou 2004; Antoniou, George Kostelenos, and Ioannis Kaskarellis 2007) and the time period under investigation covers the period 1833-1938.At this point, we have to mention that for Greece, there are official published data since 1911 by the Statistical Service.Prior to 1911, there are no officially completed time series data: however, a reliable time series data set for the period under question has been constructed by two research institutions, namely the Centre of Planning and Economic Research and the Historical Archives of National Bank of Greece (Georgios Kostelenos et al. 2007).
As a proxy for the size of the public sector we employ the total real per capita government expenditure in logarithmic form (G), whereas for the economic performance we use the log of real per capita GDP (Y).The use of per capita indexes for public size and economic performance was imposed on us, since Greece from its formation as a state in 1833 until the outbreak of the World War II experienced significant changes in both population and territorial size.

Results
The first step in our empirical analysis is the application of the Dickey-Fuller (David A. Dickey and Wayne A. Fuller 1981) test in order to examine the integration properties of the series.Note: 95% critical value for the augmented Dickey-Fuller statistic without a trend is -2.8900 and with a linear trend is -3.4543.
We next report the results of the empirical analysis regarding the Wagner and in turn the Keynesian hypotheses for the entire period (Tables 2 and 3 below).An ARDL (1,1) model has been selected by means of AIC to test the Wagner hypothesis.From Table 2 (left panel), we see that there is a significant positive long-run relationship between Y and G. Actually, the long-run coefficient of the Y is 1.94 and significant (p-value<0.01)indicating that the income elasticity of government spending is elastic.Interpretation of Wagner's Law suggests that public expenditure grows faster than national income, indicating that the income elasticity for public goods is positive rendering their demand similar to the demand for normal goods.Plausible reasons explaining the value of elasticity for public goods can be found in Dritsakis and Adamopoulos (2004).
In addition, from Table 3 (left panel) the negative and statistically significant coefficient of the EC term (-0.44 with p-value<0.01)confirms the existence of a long-run causal relationship from Y towards to G, while the statistically significant dY variable (p-value=0.01)reveals a short-run causal effect running from income towards public spending.However, the applied Cumulative Sum of Recursive Residuals (CUSUM) and Cumulative Sum of Squares of Recursive Residuals (CU-SUMSQ) tests showed that the coefficients in the equation are not that stable over the PANOECONOMICUS, 2013, 4, pp.457-472 whole period suggesting the existence of at least two distinct periods in the data set and a possible significant structural break (Figures 1a and 1b).
We next proceed with the investigation of the Keynesian hypothesis for the entire period by employing an ARDL(1,1) model selected by means of AIC as well.Following Table 2 (right panel), there is a significant positive long-run causal effect (0.35 and p-value=0.014)running from G to Y.
In addition, from Table 3 (right panel), the ECM estimates confirm that there is a rather weak long-run relationship between the two variables (p-value for EC term=0.068),as well as short-run effects, running from dG towards dY (p-value for EC term<0.01).Once again, the applied CUSUM and CUSUMSQ tests showed instability of coefficients, suggesting the existence of at least two distinct periods in the data set and a possible significant break (Figures 2a and 2b).

Note:
The straight lines represent critical bounds at 5% significance level.

Note:
The straight lines represent critical bounds at 5% significance level.
Source: Authors' calculations.Wagner's Hypothesis (1833-1938) Plot of Cumulative Sum of Squares of Recursive Residuals      In order to gain a more clear insight about the presence of a structural break, we test the integration properties of both the spending and income series by applying the LM test proposed by Lee and Strazicich (2004).This test allows for a structural change to be determined "endogenously" from the data searching for possible significant structural breaks either in the mean or in the slope, or in both the mean and the slope of the examined series.The results, reported in Table 4, panel A, confirm that public spending is a non-stationary process with a t-significant break in both the mean and the slope, identified around the year 1876.Similar findings, reported in panel B of the same table, reveal that the income series is also non-stationary with a t-significant break in both the mean and the slope, identified around the year 1880, Following the above results and in order to avoid unreliable inferences, we proceed by splitting the whole period in two sub-periods, a first one from 1833 to 1880 while the second from 1881 to 1938.It is worth noting at this point that the two sub-periods, 1833-1880 and 1881-1938, indicated by the empirical analysis, correspond to two distinct economic and political periods of the Greek state.From its establishment till 1860's, Greece's fiscal policy was determined by the decisions of its monarch whose management was characterized by the total neglect on human capital and infrastructure concerns.As a matter of fact, during the period 1833-1843 (absolute monarchy) the bulk of public expenditure was directed towards the purchase and construction of buildings to host mostly monarch's luxury needs.During the constitutional monarchy period (1844-1862), we had the construction of new public buildings, as well as some public ex-penditure on transportation infrastructure.According to Maria Synarelli (1989), Antoniou (2011) and Thanassis Kalafatis and Evangelos Prontzas (2011) during the years 1833-1881 a very small portion of public expenditures was directed towards public infrastructure and public investment in general.Until the decade of 1860's, state revenues lacked significantly state expenditures and external borrowing was the main source to finance mainly the royal standards of living.After 1860's and the first Constitution, the political climate became more stable.In the late 1870's the Greek economy opened to foreign markets and applied active fiscal policy through spending in transport infrastructure (railroads, ports, etc.).During the same period, several sectors showed rapid growth, such as banks, trade, shipping and agriculture.Greek and foreign funds were invested in the economy providing in that way the ground for better economic performance (Lefteris Tsoulfidis 2003).

Figure 2a
Thus, the two distinct periods of the Greek state should appear in the analysis.As a matter of fact, the results, reported in Tables 5 and 6 below, for the period 1833-1880 confirm the existence of cointegration between the tested variables and reveal that the Wagner hypothesis holds.With G as the dependent variable the evidence reveal the existence of a long-run equilibrium relationship (p-value=0.013), with long-run causality running from Y to G (Table 5, left panel) providing support in favour of Wagner's Law.The ECM representation, reported in Table 6 (left panel), also confirms the existence of long-run causality running from Y to G, since the respective coefficient is statistically significant and have the correct negative sign (-0.488, p-value = 0.006).Besides, there is evidence of a short run causal impact running from dY to dG (p-value < 0.01) while the CUSUM and CUSUMSQ tests reject the instability of the estimated coefficients.
Turning now to the Keynesian hypothesis (Table 5, right panel) with Y as the dependent variable, we see that the evidence is not in favour of long-run causality running from G to Y (p-value = 0.751).That is, for that specific period of time, the Keynesian hypothesis is not vindicated.The same conclusion can be drawn from Table 6 (left panel) which refers to the EC estimates.Actually, we observe that there is a short-run causal effect running from dG to dY (p-value = 0.00) but there is no evidence of a long-run relation, since the EC term is not statistically significant (p-value=0.235).CUSUM and CUSUMSQ tests confirm stability of the estimated coefficients.As already mentioned, a possible explanation for the above results can be found in the total absence of public works and in the increase in royal spending absorbing all state revenues and loans (Synarelli 1989;Tsoulfidis 2003).In contrast, for the sub-period 1881-1938, the analysis confirms that both Wagner's Law and Keynesian hypothesis hold (Tables 7 and 8 below).When government spending G, is the dependent variable, the appropriate specification, by means of AIC, is an ARDL(1.1)model, which is used to investigate the long-run relation between the economic level and government spending.The left panel of Table 7 indicates the existence of a positive and statistically significant long-run causal effect (p-value<0.01) running from income towards government spending; hence, the results provide evidence in favour of Wagner's Law.Also, the ECM representation reported in Table 8 (left panel), also confirms the existence of long-run causality running from Y to G, since the respective error correction term is statistically significant and negative (-0.54, p-value<0.01).Besides, the statistically significant coefficient of dY (p-value=0.01)reveals a short-run causality running from dY to dG.
Turning now to the Keynesian Hypothesis with Y as the dependent variable the evidence reported in Table 7(right panel) is also in favour of a long-run equilibrium relationship with long-run causality running from G to Y (p-value=0.003)vindicating the Keynesian hypothesis.The same conclusion can be drawn from Table 8 (right panel) which refers to the EC estimates.In particular, there is evidence of significant causal effect running from public spending to income in both the long and short-run period (EC term=-0.50 with p-value=0.00and p-value=0.00for dG).It should be noted at this point that the employed ARDL models for all sub-samples were also tested for coefficients' stability by the CUSUM and CUSUMSQ tests and we found no evidence of coefficients' instability over the examined sub-periods.The analysis shows that whenever there is an active state following Keynesian policies, the growth performance of an economy improves.This becomes obvious from the results, since in the second sub-period 1881-1938 in which the state undertook a more active role in country's reform, the performance improved.These results imply that active Keynesian policies may help an economy's growth record and thus the austerity programs in the phase of a severe recession period such as the current one in Greece might not be the most appropriate economic policy to follow.However, the detection of Wagner's hypothesis, at the same time, could be justified by the increased needs of a newly formed State to build the necessary mechanisms (institutions) in order to control and organize effectively its functions.The same needs, someone may argue, are still present and have to be considered in the restructuring of the Greek State to overcome from the current deep recession conditions.
2,…T) and Z t is a vector of exogenous variables defined by the data generating process;is the vector of coefficients in the regression of The straight lines represent critical bounds at 5% significance level.Source: Authors' calculations.

Figure 1b
Figure 1bKeynesian Hypothesis Plot of Cumulative Sum of Recursive Residuals

Figure 2b
Figure 2bKeynesian Hypothesis Plot of Cumulative Sum of Squares of Recursive Residuals (Soren Johansen 1988;Johansen and Katarina Juselious 1990)for G and Y series provide rather vague results depending on the presence of a time trend variable; when testing the variables in first differences (dG and dY) the results clearly indicate stationary properties.In such a case and given that the conventional cointegration tests(Soren Johansen 1988;Johansen and Katarina Juselious 1990)are based only on I(1) variables, we employ the ARDL cointegration method which can be applied regardless of the stationary properties of the variables involved.

Table 2
Estimated Long Run Coefficients Using the ARDL Approach Source: Authors' estimations.

Table 3
Error Correction Representation for the Selected ARDL Model

Table 4
Lee and Strazicich min LM t-stat

Min t-stat Break date t-stat mean dummy t-stat slope dummy
Source: Authors' estimations.

Table 5
Estimated Long-Run Coefficients Using the ARDL Approach Source: Authors' estimations.

Table 6
Error Correction Representation for the Selected ARDL Model

Table 7
Estimated Long-Run Coefficients Using the ARDL Approach Source: Authors' estimations.