Politics versus Economics in the Explanation of Government Size in Canada: 1870 20001

by

J. Stephen Ferris*
stephen_ferris@carleton.ca

Soo-Bin Park*
soo-bin_park@carleton.ca and

Stanley Winer**
stan_winer@carleton.ca

Second Draft April 10 2004

* Department of Economics
** School of Public Policy Carleton University Ottawa, Ontario, K1S 5B6 CANADA

Abstract
Politics versus Economics in the Explanation of Government Size in Canada: 1870 2000
by J. Stephen Ferris, Soo-Bin Park, and Stanley Winer Carleton University, Ottawa, Canada This paper uses annual data from 1870 and 2000 in Canada to test whether overtly political variables interact with macroeconomic variables through government size. We begin by asking whether Canadas macro data is consistent with political cycles, i.e., the hypothesis that macro cycles have been caused by overtly political influences such as the timing of elections, the political ideology of the governing party, the size of the winning majority, and/or whether there was a minority government. After finding some evidence of a correlation between political variables and output growth (but not inflation), the paper explores whether the transmission mechanism for these cycles could be through government size. To test for this relationship, the analysis uses an error correction model constructed under the base case assumption that political variables have no separate influence on government size. Competition among political parties is assumed to lead government size to converge on an equilibrium that depends only on the underlying tastes and technology of the community. The addition of political variables to this structure then allows us to assess whether explicit political considerations can still explain sympathetic variations in real government size once a complete long and short run model of the economic factors at play has been fully specified. Key Words: Political cycles, government size, rational opportunistic and partisan political hypothesis, politics and growth. JEL Categories: H1

Politics versus Economics in the explanation of government size in Canada: 1870 2000

I.

Introduction
Alesina, Roubini, and Cohen (1997) among many others argue that the macroeconomic data

across many countries is broadly suggestive of the cycles predicted by rational opportunistic and/or partisan political theories. These hypotheses argue that overtly political variables, such as the timing of elections and the change in the ideology of the party in power affect systematically such aggregate macroeconomic time series as the rate of growth of real output and/or the inflation rate. 2 Implicitly they argue for the centrality of distinctly political factors in the determination of aggregate economic outcomes (Bartels and Brady, 2003). In this paper we examine the relationship between political and economic variables in Canada by first asking whether such political cycles can be found in Canada. We do so using annual data covering 1870 to 2000, a time period that spans the entire post-Confederation political era (running from 1867 to the present). 3 Given that annual cycles can be found in the data, we proceed to examine in detail one potential transmission mechanism. That is, a more compelling case for political factors causing macro cycles can be made (rather than have the political outcome result from the cycle) if we can identify the route through which political influence is exercised. The second part of the paper then takes up the Bartels and Brady challenge to better investigate plausible transmission mechanisms between political and economic outcomes. We do so by formally examining whether political control could have been exercised through variations in government size.4 To test whether overtly political variables can explain the required changes in government size in Canada, we first ask whether the data can distinguish between two competing hypotheses called, for convenience, the economic convergence and political nonconvergence hypotheses. If

2 they can, we then ask whether the form that political influence took would imply the cycles in the first stage analysis. For our purposes, economic convergence is taken to be the hypothesis that government size will converge to the level dictated by the communitys tastes and technology alone. In contrast, the competing hypothesis of nonconvergence means that opportunism or partisanship will prevent convergence to that level. To operationalize our test, we use an error correction model to identify as nonconvergent situations where at least part of the short run variation about the long run size can be explained by the partisan nature of politics (which party is in power, the size of the majority received by the winning party, whether the party in power had a majority or not) or by the opportunistic nature of political competition (the specific timing of events during the electoral cycle). While nonconvergence could arise in either the long or the short run, our focus on cycles leads us to be concerned primarily with the adjustment process about long run size.

II.

Overview of the Canadian Macroeconomic Data and Political Variables Used

The data used in this study comes primarily from two sources: Canadian Historical

Statistics for the earliest time period (1870 through 1921) and from Cansim, the statistical database maintained by Statistics Canada, for the later time period (1921-2001). Because annual variables used in the pre 1921 period are often interpolated between census dates, 1921 becomes a natural time interval across which we test for the robustness of our findings. In addition, we use the 1945 to 2000 time period to see if political variables have come to matter more (or less) in contemporary times. In the appendix we list the specific data definitions and their sources. Here we motivate the choice of variables and present their descriptive statistics, particularly in relation to their time series

3 properties. Figures 1and 2 illustrate the two stages of our problem. Our first test follows the traditional approach of examining the influence of political variables on the growth rate of real output and the inflation rate and these two series are illustrated in Figure 1. Output growth (GROWTH) is measured as the rate of growth of real GNP and inflation (INFLATION) as the rate of growth of the GNP deflator. As the diagram illustrates, both series are somewhat correlated (but not perfectly), with both series exhibiting more variability in the earlier rather than later time period. 5 Figure 2 presents the dependent variable of our second test. Here government size is defined as non-interest federal government expenditures as a fraction of Gross National Product (GSIZE). The federal level is chosen because of the transmission process to output runs from federal political parties through elections to government control to fiscal policy. In choosing federal size, however, we recognize that at least part of federal expenditures relate to the comparative advantage that the federal government has in redistributing tax revenues to and among provincial and, to a lesser extent, local governments. Hence we control for changes in the intergovernmental federal transfers by netting out equalization payments and other grants from federal expenditures (GRANTS). The difference this makes to our measure of government size is shown as the difference in the two government size lines in Figure 2. As Figure 2 illustrates, the size of Canadas federal government has grown more or less constantly over the entire 1870 -2000 time period. Only since 1992 has there been some sign that this upward trend may be ending. What is perhaps most striking in the diagram is the dramatic effect of the two world wars. In response to WWII, for example, the federal government grew in size from roughly ten percent of GNP in 1939 to almost fifty percent by 1944. Most of that

4 increase was reversed by 1949. In the public finance literature, however, there is the widespread perception that WWII has represented a watershed in public attitudes towards government that in turn permitted a ratchetting upwards in government size (Peacock and Wiseman 1961). Below we test for the Peacock and Wiseman effect by including a dummy for the post WWII time period. In effect, we ask whether the second world war represents a structural break in an otherwise stable equilibrium relationship.6 Canada is a small open economy located immediately adjacent to the US. As such, the Canadian economy has always been integrated with and strongly influenced by the US. Because of that dependence, we use US growth and inflation rates as our control for nonpolitical reasons for output and inflation changes in Canada. US output growth is measured as the growth rate of the US Index of Industrial Production (USGROWTH) and US inflation (USINFLATION) as the rate of growth of the GDP price deflator. To test for the presence of nonconvergence, we use combinations of a set of five overtly political variables: a dummy variable representing the year of each federal election (ELYEAR); a discrete variable representing the years since the last election (ELAPSE); a dummy variable representing the years in which the ideology of the party in power was more liberal (LIBERAL); a dummy variable representing the time period in which the governing party was a minority (MINORITY); and the fraction of the seats in parliament held by the winning political party (SEATS). The descriptive statistics for these variables are presented in Table 1a. Note, in particular, the adjusted Dickey Fuller test statistics for each series. 7 All of these variables, including the set of political variables, are stationary in their levels and so are I(0). To capture the determinants of government size, we need a set of variables that both span the time period covered (1870 - 2000) and reflect the deeper structure of the Canadian economy.

5 The variables used for this purpose have been widely used in other studies of government size (see, Kau and Rubin (1981), Borcherding (1985), Mueller (1989) and Ferris and West (1996)). The starting point is almost always Wagners Law, the hypothesis that government size will increase as society grows in scale and complexity, and this is interpreted as implying an income elasticity (of size) that is positive. Hence our first explanatory variable is real income per capita (RYPC). Second, government may exist to provide particular types of goods, such as public goods. Here the publicness of government output is accounted for by including population (POP) as a variable. Next, public choice theories of government size suggest that the fraction of the population in agriculture (AGRIC) and the fraction that are young (YOUNG) may proxy either changes in the structure of the economy or the relative strength of interest groups with high demands for publically provided goods and services. To capture other structural features that may promote more (or less) government involvement, we include immigration rates (IMRATIO) and the openness of the economy through its reliance on foreign trade (OPEN; Rodrik, 1998). Finally because both GSIZE and some of our explanatory variables (IMRATIO, AGRIC, YOUNG, and OPEN) are all restricted to the range between zero and one, we transformed our variables into logarithms. In summary, the variables used to explain government size are: real gross national product per capita (LNRYPC), the percentage of the population in agriculture (LNAGRIC), the percentage of the population that is young (LNYOUNG),8 the fraction of new immigrants in the population (LNIMRATIO), and the degree of openness in the economy (as reflected in the sum of exports and imports as a fraction of GNP (LNOPEN)).9 The descriptive statistics for these variables are presented in Table 1b. Note that in contrast to Table 1a, all variables are nonstationary in levels and become stationary only after first differencing. The exception is LNPOP that becomes stationary only after a second

6 differencing. With this exception, all the variables utilized as economic determinants of government size are I(1).

III.

Are there Political Cycles in annual Canadian Macroeconomic Data? The evidence for political cycles in annual Canada macro data is presented in Tables 2 and 3

below. These tables represent the results of regressing the rates of growth of real output, GROWTH, and inflation, INFLATION, against a short list of constructed political variables suggested by either opportunistic or partisan theories of political cycles (see Alesina, Roubini, and Cohen, pp. 36 and 62), with some control for other nonpolitical influences. With respect to the political variables used, traditional opportunistic political theories suggest that incumbent political parties will seek to increase real output in the time period immediately before an election to gain votes. Rational opportunistic theories, on the other hand, emphasize that for such spending to affect aggregate demand and hence output and/or prices, that spending would need to be unanticipated. Hence rational opportunistic theories would not expect predictable pre-election spending to change real output growth or inflation. To test whether the data is consistent with traditional or rational pre-election opportunism, we look for a positive coefficient on ELYEAR lagged. 10 Partisan political theories, on the other hand, suggest that parties on the left (i.e., the Liberal Party in Canada, LIBERAL) will spend more when in power than the conservative alternative. Again because anticipated changes will be accounted for in individual behaviour, rational partisan theories refine this hypothesis by predicting that as long as the election outcome is not certain, the realization of the election of a leftist political party will generate some boost to aggregate output (and/or inflation) in the period following the election. The size of the effect should then depend upon the

7 degree of surprise in the election result. In addition, whatever the surprise, the initial stimulative effect would be expected to dissipate throughout the tenure of the party in power. 11 The exact opposite would be expected for the election of the more conservative party (1- LIBERAL). Here a larger electoral surprise would be expected to result in a larger fall off in growth and/or inflation rates. Once again as time eliminates surprise, this contractionary effect should dissipate. In Tables 2 and 3 we test for both the surprise and the duration hypotheses by including a dummy variable that distinguishes between the time periods when each party type was in power and measures electoral surprise as one minus the fraction of seats won by the governing party, (1SEATS). Hence rational partisan theory suggests that the coefficient on the composite variable {(1SEATS)*LIBERAL - (1-SEATS)*(1- LIBERAL)} should be positive. One additional modification is made to help capture the distinctive parliamentary nature of Canadas political system. That is, during time periods when parliament is controlled by a minority government, political behaviour is often anomalous. Because minority governments are required to maintain cooperation among a coalition of often disparate political parties, conflicting political agendas are likely to be resolved in ways that are different (for spending) than when control is exercised through a majority. For this reason we use the dummy variable (1-MINORITY) to isolate time periods when the surprise hypothesis is expected to be most relevant. Finally, the duration hypothesis is tested through the use of another composite variable, {(LIBERAL*ELAPSE) - (NOT LIBERAL)*ELAPSE}. The size of the differing surprise effects generated by the election of a liberal (not liberal) government are predicted to dissipate as time in power elapses.

The results of these tests for Canada are the OLS regression equations presented in Tables 2

8 and 3. Before reporting on these results, however, we note that in both tables, the same equations were run using different measures of some of the political data (columns labelled A versus B). That is, different interpretations of the size of the winning majority sometimes arise in Canada because closely associated independent candidates often run unopposed by the winning political party and, once elected, tend to vote together with the winning party. Hence whether they are part of the winning coalition is often problematic. For our purposes, this creates most ambiguity when trying to interpret whether or not there is a minority government. 12 The results presented in the first two columns then follow one convention (Beck and Campbell) while the second two columns follow another (the Canadian Parliamentary Guide). Finally, all equations control for economic influences on growth by using the US growth equivalent. The first of the two sets of columns in each table represent our findings for the entire 1870-2000 time period, while the second represents the 19212000 subperiod.13 A quick glance at Table 2 illustrates that despite the importance of economic variables for explaining the growth rate of Canadian output, as reflected in the significance of USGROWTH and USGROWTH(-1), the set of political variables do assist in explaining real output growth in one way or another in all forms of the test for Canada. A Wald test rejects the hypothesis that the political variables as a group are jointly insignificant (at five percent or better) in three of the four cases. By inspection it can also be seen that the different definitions of the size of the winning majority and minority status make for little difference in practice, the two equations generating virtually identical regression coefficients and standard errors. Table 3, on the other hand, does indicate that political variables have much less success explaining variations in inflation. The control variable for economic influences, USINFLATION, is highly significant (producing the

9 equations overall high explanatory power), but none of the designated political variables matter. That is, even though each coefficient has its expected sign, none of the coefficients (either individually or as a group) are significantly different from zero. 14 Overall, then, the macro data are consistent with the hypothesis of political cycles in annual real output growth in Canada over the 1870 to 2000 time period but give only limited support to the hypothesis that the inflation rate exhibits political cycles. We turn next to the question of which political hypotheses are consistent with cycles in real output growth. First, the traditional hypothesis that opportunistic pre-election spending will produce a significant positive effect on output growth is not supported by the data. Even though all four coefficients have their expected positive sign, all are insignificantly different from zero. In this sense, the data is more consistent with the non-significance predicted by rational opportunistic theo ry. According to that hypothesis, any predictable pre-election behaviour would be anticipated by and adjusted to by both individuals and the market so that its effect on behaviour and output would be negligible. The surprise hypothesis predicted by rational partisan theory is broadly consistent with the data. The surprise coefficient, {1-MINORITY}{(1-SEATS)*LIBERAL-(1-SEATS)*(NOT LIBERAL)} is positive in all equations and is significantly different from zero at five percent in two of the four equations (and at ten percent in the remaining two). 15 Narrow LIBERAL victories are associated with larger increases in the output growth, while narrow conservative victories correspond to periods with larger fall offs in output growth. The corresponding political tenure hypothesis, on the other hand, is suggested but not confirmed by the data. That is, in three of the

10 four equations the predicted negative sign appears, but in all cases the coefficients are insignificantly different from zero. In summary, whether or not the data reject the hypothesis that political variables produce cycles in the inflation rate, our results are consistent with the hypothesis that there are political cycles in the annual growth rate of real output. Individually, the data does not support traditional theories of opportunism or partisanship in their hypotheses that spending immediately prior to an election can affect real output growth or in traditional partisan directions. However, the data is consistent with the implication of rational opportunism and the hypothesis that surprise partisan political victories will change output growth rates in their predicted direction. If we use these results to motivate the search for a mechanism by which political control could have been exercised, the relative strength of the link between political events and real output growth (compared to the link through inflation) suggests that search might well begin with fiscal rather than monetary policy. For this reason we proceed and test for one plausible channel by which fiscal policy could affect real output growth. In particular, we ask whether the data is consistent with the hypothesis that political variables can explain cycles in government size. If we find these variables to be significant contributors to such cycles, we can then ask whether the changes found are consistent with the observed cyclical effects on real output growth. Both are needed to support the joint hypothesis of a political cause for such macro cycles.

IV.

Methodology Because partisan and opportunism models argue for the additional importance of overtly

political variables in explaining government size and/or its rate of change, we begin by setting out a

11 base case that would attribute changes in government size only to those nonpolitical variables that are thought to capture the underlying tastes and technology of the community. This base case model we then interpret as the economic convergence hypothesis. To represent the convergence hypothesis we develop an error correction model of government size. More specifically, we begin by using the public choice literature to indicate a set of long run variables as structural determinants of government size and then investigate whether the data indicates the existence of stable linear relationship among the levels of these variables across time in Canada. Typically, the variables used in this regression are nonstationary in their levels so that their estimated relationship could well be spurious. Nevertheless, if the residuals of the estimated equation in levels are stationary, we can interpret the estimated relationship as evidence of cointegration, i.e., evidence of a long run equilibrium relationship linking these variables with government size (Engle and Granger, 1987). This outcome we interpret as reflecting the underlying tastes and technology of the community in the long run. Given that a stationary long run relationship can be found, the residuals from this estimated relationship will be I(0) and can be used as the error term in an error correction model of short run adjustment about the long run equilibrium. That is, the lagged values of the residuals of the long run equation can be incorporated within a model of first differences. This allows us to estimate the short run adjustment process that best explains the annual variations in actual government size about its long run predicted size. This error correction model represents our taste and technology explanation of short run adjustment. Together with the cointegrating equation, the two relationships represent our base case convergence hypothesis. It follows that if the addition of a set of overtly political variables also results in the finding

12 of cointegration, we can infer the existence of an expanded long run equilibrium relationship and the corollary that the political variables do form part of the long run explanation of government size. This would be convincing evidence of nonconvergence. In the current case, however, the political variables chosen are all I(0) in their timing and/or sequencing of political events. This reflects their development for use in relation to cyclical or short run influences on inflation and/or economic growth. Hence using these political variables in relation to variations in government size, implies that we are looking primarily for evidence that the introduction of these overtly political variables will improve the predictive power of the error correction part of the model. In general, however, either type of evidence would contradict the maintained convergence hypothesis and provide support for the hypothesis that partisanship or opportunism works through government size to produce a political cycle.

V.

Cointegration and the Long Run Model of Government Size With this background we present our base case convergence model of long run government

size as the first and second columns of Table 2.16 In these equations we argue that a cointegration relationship does exist among the five I(1) variables with, or perhaps without, one structural break arising at the end of WWII. To derive this conclusion, we used the following procedure. First we began with a regression equation that included the complete list of explanatory variables as set out in Section II above. The residuals of that OLS regression equation, however, gave no evidence of cointegration. We then proceeded to eliminate variables successively, beginning with what seemed the least significant variable in the equation (first LNOPEN and then D(LNPOP)). At each stage the adjusted Dickey Fuller (ADF) test statistic on the equations residuals was compared with the

13 critical value of the modified MacKinnon criteria. This took us to the equations represented by columns (1) and (2). There the interpretation of the appropriate cointegrating equation became ambiguous. That is, in equation (1), an ADF statistic of -4.63 on the residuals (for five interrelated variables) falls between the MacKinnon critical value at one percent (indicating rejection) and the five per cent criteria (suggesting that cointegration is present). As discussed earlier, however, the Peacock and Wiseman hypothesis suggests that we should find a structural break at the end of the second world war, one whose presence might resolve the ambiguity suggested by this marginal finding of cointegration. To test this hypothesis, we added WW2aftermath, a dummy variable taking the value 1 in the years following WWII, to allow the relationship to incorporate a one-time break (jump) in the level of the series at that date. This led to a new set of residuals whose ADF statistic of -5.25 allowed the interpretation of cointegration. 17 While either the base case equation in column (1) or (2) is consistent with a long-run equilibrium relationship existing among the variables (with one structural break at the end of WWII), it is highly likely that innovations among the I(1) variables will be correlated. This implies that the standard errors of the equations coefficients will be inconsistent and not valid significance testing. Hence to go further and discuss the significance of each individual variable, we need to follow Saikonnen (1991) and adjust the equations and standard errors. [This is done in Table 5 below]. Nevertheless, because the OLS coefficient estimates are themselves super consistent, we can still utilize the equation as a whole as a reliable test of whether a cointegrating relationship exists. Equations (1) and (2) can then be seen to indicate that over the full 1870 to 2000 time period a cointegrating relationship can be found among a subset of the key variables indicated by public choice theory as determinants of government size. What is more uncertain is whether

14 cointegration requires the presence of WWIIaftermath in that relationship. In equations (3) and (4), we repeat equations (1) and (2) over the shorter 1921 - 2000 time period.18 What is interesting here is that both equations fit the data better than before, both in the sense of lowering the Akaike information statistic and providing stronger evidence of cointegration. Moreover, in the shorter time period, the structural break at the end of WWII is no longer needed to find convincing evidence of a cointegrating relationship among the subset of five economic and public choice variables. In both cases the adjusted Dickey Fuller test statistics exceed by some margin (in absolute terms) the MacKinnon critical value at the one percent level. Because we have a greater interest in the results over the longer time period, we continue to use the cointegrating equations in columns (1) and (2) as our base case and interpret the short period results as providing additional support to the cointegrating relations found in (1) and (2). However, the shorter period results do cast doubt on the hypothesis that there is a structural break in the cointegrating equation following WWII. The next step is to test for the importance of our list of political variables. Before doing so, we wish to emphasize two key factors: one a characteristic of the data and the second, a feature of the new test. First, because the political variables for the first stage of the analysis were collected in order to test for recurring cycles in the data, the political variables are all I(0) variables. Then since government size grows through time and does not revert to its mean, it seems most unlikely that any of the I(0) political variables could contribute to the explanation of long run government size (done through a collection of I(1) variables). For this reason we expect that if political variables were significant in relation to explaining government size that evidence of this relationship would be more likely to be revealed in relation to the error correction model than through the cointegrating

15 equation. Second, as we move from a test for the effect of politics on output to a test for their significance in relation to government size, the same hypotheses require the variables to enter in a different way. That is, unlike the growth equation, where rational expectations implies the importance of surprise, surprise is not needed for the effect on size. In particular, for government size to have been the mechanism by which output was affected through surprise, the change in size must simply be present in order for the further effect on output to have been possible. In this sense, then, the test for the effect of political factors on size is much more direct and straightforward than is the same test for the effect on output growth. With this background, then, all fiscal opportunism theories (traditional or rational) predict that government size will be larger leading into elections. 19 Thus ELYEAR(-1) should be positively related to government size.20 Partisan theories, on the other hand, suggest that Liberal victories should always increase government size (compared to the NONLIBERAL alternative). Similarly, given the existence of partisan effects, any real diminution (expansion) of that effect over the tenure of different governments should depend on the type of party elected. This we test for through the composite variable {LIBERAL*ELAPSE - NONLIBERAL*ELAPSE}. For completeness we add the remaining two political variables at our disposal, MINORITY and SEATS. Intuitively, minority governments might be expected to resolve potential conflict by spending more and so that MINORITY might be associated with larger size. Similarly, larger electoral majorities, SEATS, may lead all parties to spend either more (or less) independent of party affiliation. These last two hypotheses, however, are ad hoc in the sense that they are not associated directly with any partisan or opportunistic theory linking politics to cycles in the macro data. What is presented in columns (5) and (6) of Table 4 are the final stage results of our search

16 for cointegration among the enhanced set of economic and political variables. Following the same procedure as set out earlier, we first added all five political variables to the base case equation (in column (1) and then tested for cointegration. The absence of cointegration then led us to eliminate the least significant variable and to retest. This continued until all the political variables but SEATS were eliminated and the equations in columns (5) and (6) were isolated. Despite initial appearances, however, these findings give no support to the hypothesis that any of the political variables should be included in the previously discovered cointegrating relationship. That is, although SEATS appears to be significant in the full 1870 - 2000 time period, the equations residuals are not stationary so that the hypothesis of cointegration can be rejected. On the other hand, cointegration would seem to be indicated in equation (6) even though the SEATS variable itself does not appear to be significant.21 Here it seems likely that cointegration was found only because of the strength of prior cointegration relationship in (4). That is, the addition of SEATS to the equation only lowers the adjusted R2, the Akaike statistic, and the (absolute value of the) ADF statistic.22 As argued earlier, finding that political variables that indicate the timing of political events or the binary switching of power between partisan opposites do not enter the cointegrating equation might have been expected on the basis of the differences in the order of integration in the two sets of variables. From this perspective, the primary role of this section has been to find the long run cointegrating relationship on which we can build our error correction model of dynamic short run adjustment. To complete that process, we present in Table 5 the Saikonnen adjustment for serial correlation among the error terms of the set of I(1) variables in the cointegrating equation. Their presence has the effect of making the coefficient standard errors inconsistent, tending to overstate

17 the size of the t-statistics that arise under OLS estimation. For our purposes, what is particularly interesting about the Saikonnen result of Table 5 is that the WWIIaftermath dummy, that appeared to be significant in Table 4, can now be shown to be insignificantly different from zero. Together with the shorter period results in Table 4, the Saikonnen results resolve the ambiguity over whether a Peacock and Wiseman structural break was needed for cointegration. The Saikonnen results allow adoption of equation (1) as the cointegrating equation used in the error correction model below.

VI.

An Error Correction Model of Government Size While one might be surprised if the timing and/or partisan nature of political events

mattered for the long run size of government, it would be much less surprising to expect these political factors to matter in relation to explaining short run cyclical variations about that long run size. Indeed, the power of partisan and opportunistic political theories is their implicit reliance on the strategic use of transitory spending to influence individual behaviour and aggregate output. To explore this question, we present, in Table 5, the error correction model implied by our earlier analysis.23 In columns (1) through (3), we present the base case error correction model standing for the economic convergence hypothesis. The three columns reflect the same equation run over different time periods. To isolate the variables in the final equation used, we again followed the methodology of incorporating in the initial estimating equation all the first and second differences of the economic and public choice variables (as described in Section II). In addition, we included a variable to reflect changes in the scale of federal transfer payments made to other levels of government (D(LNGRANT_SHARE). The idea was that short run changes in the real size

18 of federal transfers could speed up or hold back competing federal programs and so influence the ability of the federal government to exercise spending discretion in relation to the cycle. With this enhanced set of variables, the full equation was run and those variables that were least significant were dropped until the variables remaining were all significantly different from zero. The result was a convergent error correction model that could explain sixty five percent of the short run variation in government size about its long run equilibrium path for both the 1870 - 2000 and 1921 - 2000 time periods. In the shorter 1945-2000 time period, the model explained somewhat more, roughly eighty percent, of the variation in government size. In each equation, the error correction term was negative as expected and was significantly different from zero at one percent for the entire 1870 - 2000 time period, and at five and ten percent for the shorter more recent time periods. The small size of the estimated error correction coefficient implies that deviations from long run size are corrected only over very long time periods. In terms of its economic meaning, the significant feature of the error correction model is that the coefficient estimate on the contemporaneous change in income is significantly negative in all equations and the lagged income coefficient, while positive, is considerably smaller than the long run coefficient estimates found in Tables 4 and 5. This is evidence consistent with a strong counter-cyclical role for government size. That is, government size first increases contemporaneously with a fall in income and then decreases over time at a rate that is slower than the rate of contraction will be in the long run. The other notable feature is that increases in the intergovernmental transfer share of federal government spending is always associated with a fall in government size. This seems to represent short run changes in the political strength of the federal government versus the provinces and so reflect a form of political competition. While our

19 explanation for this effect is somewhat ad hoc, the results consistently suggest that this effect is substantive and pervasive. It was significantly negative in every form of the test run. The final step is to test for the significance of politics on short run variations in size by adding the set of political variables to the error correction model. The remaining three columns of the table then present the alternative nonconvergence hypothesis for the same three time periods. In column (4) alone we present the complete set of results for the political variables. In succeeding periods the same pattern of results was found so that only the final equation estimate was presented. As column (4) indicates, the addition of the five political variables has no real effect on the explanatory power of the equation as a whole. The adjusted R2 and Akaike test statistics are largely unchanged and the previous economic coefficient estimates are also largely unaffected. This finding is reinforced by the individual political results that imply, with the exception of SEATS, that none of the political variables are significantly different from zero. This same pattern of findings is repeated in columns (5) and (6), where in the 1921 to 2000 time period, even the size of the winning majority had no significant effect on size. For the 1921-2000 time period, then, the data would lead us to reject the nonconvergence hypothesis in favour of convergence. In the other two time periods, however, the data does support the hypothesis that at least one characteristic of the political process, the size of the winning electoral majority, does matter in explaining variations in government size. The key question is not whether SEATS affects size, but whether evidence that the size of the winning majority is the only significant determinant of short run government size is consistent with the political cycle hypotheses tested in the first part of this paper. There we found tentative support for both the rational opportunism and rational partisan theories of the cycle. For asymmetric

20 information to imprison attempts at electoral opportunism in a political equilibrium (as implied by rational opportunism), government size would need to rise in the period leading into each election. Our results, however, give no evidence that the change in government size is any larger than usual in the period before elections and hence contradict the hypothesis of opportunism in either form through government size.24 Partisanship, on the other hand, suggests that liberal governments will increase government spending and hence government size relative to their conservative alternative in their tenure. This hypothesis is also not consistent with the data nor is the hypothesis that the real size of government diminishes (or expands) through tenure in the direction suggested by partisan ideology. It follows that while we do find evidence that politics matters for government size, the size variations implied are not consistent with either political theory of the cycle. Rather, evidence that large majorities result in temporarily larger governments is more consistent with the hypothesis that active competition among political parties is needed to prevent temporary overexpansion of government size relative to the level implied by technology and the tastes of the community. In line with our initial convergence hypothesise, then, it is the presence of political competition that leads to convergence; in its absence, politics matters.

VII.

Conclusion In this paper we have investigated the effect of the timing and partisan nature of Canadian

politics on economic growth, inflation and government size over the entire post-Confederation time period, 1870 - 2000. The evidence found, however, tends to pose more questions than it answers. For example, while a persistent political cycle does appear to be present in the data for real output growth in Canada between 1870 and 2000, the mechanism by which politics influences output is

21 not readily apparent. In this paper we investigate only government size as the mechanism by which political influences could be transmitted into variations in real output and/or inflation. There are, of course, a large number of other methods by which political control could be exercised changes in taxes, the size of the deficit, the composition of spending, and/or monetary policy to name only a few. However, to the extent that political influence on output requires sympathetic movements in government size, the data does not support any of the more established theories that would explain that interconnection. In particular, the timing of political events and the partisan nature of party politics does not seem to prevent the convergence of government size to its fundamental long run value nor does there seem to be any more evidence for the hypothesis that the timing or partisan nature of party politics matters in relation to cyclical variations about that long run size. There are two final aspects of our work that merit some emphasis. The first is that our paper not only distinguishes empirically changes in size that arise over the cycle from those that arise in the long run, but sees this distinction as derivable only as part of an analysis that jointly determines the two activities under political competition. In earlier work (Ferris and Winer 2003), we used an explicit probabilistic voting model to derive estimating equations for the set of spending, taxation, and debt policy choices made by government in both the long and the short run. In the current case we exploit cointegration theory and the error correction framework to distinguish among the many reasons for changes in government size to isolate the contributions made by different theories of counter-cyclical fiscal activity and political intervention to the short run. What is common to both approaches, however, is our belief that some encompassing theory is needed to interpret and disentangle the web of connections that link decision-making in the short and the long run. Finally, the particular measure of government size used in the paper uses federal spending

22 net of intergovernmental transfers as the instrument by which political influence affects aggregate demand and so output. In doing so it attributes greater power to spending on final goods and services than to redistributive spending through intergovernmental transfers. On the other hand, we have also seen that the error correction model of government size is highly responsive to the share of intergovernmental transfers in final spending. This we have interpreted as a reflection of the transitory consequences of political competition among the different levels government. Whether this latter interpretation is valid or not, our finding for the significance of this variable raises an important new set of questions for the nature of intergovernmental transfers and their effect on both government size and aggregate output.

Appendix 1. List of Economic Variable Names and Data Sources: AGRIC = proportion of the labor force in agriculture. 1871-1926: Urquhart, (1993), 24-55; 19261995 Cansim series D31251 divided by D31252; 1996-2001: Cansim II series V2710106 divided by V2710104. Exports and Imports = exports and imports. 1870-1926, Urquhart, (1993) Table 1.4; 1927-1960, Leacy, et al, 1983, Series G383, 384; 1961-2001: CANSIM series D14833 & D14836. OPEN = openness. Calculated as (Exports + Imports)/GNP. GNP = gross national product in current dollars. 1870-1926: Urquhart (1993), pp. 24-25 (in millions); 1927-1938: Leacy et al (1983), Series E12, p.130; 19391960 Canadian Economic Observer, Historical Statistical Supplement 1986, Statistics Canada Catalogue 11-210 Table 1.4. CANSIM D11073 = GNP at market prices. 1961-2001 Cansim I D16466 = Cansim II V499724 (aggregated from quarterly data). GOV = total government expenditure net of interest payments.1870-1989: Gillespie (1991), pp.284-286; 1990-1996: Public Accounts of Canada 1996-97: 1997-2000: Federal Government Public Accounts, Table 3 Budgetary Revenues Department of Finance web site, September 2001. To this we add the return on government investment (ROI) and subtracted out interest paid to the private sector. Data on ROI: 1870 to 1915: Public Accounts 1917 p.64; 1915-1967: Dominion Government Revenue and Expenditure: Details of Adjustments 1915-1967 Table W-1; 1916-17 to 1966-67: Securing Economic Renewal - The Fiscal Plan, Feb 10, 1988, Table XI; 1987-88 to 1996-97: Public Accounts 1996, Table 2.2. Interest on the Debt (ID) was subtracted out (with adjustment for interest paid to the Bank of Canada (BCI) ultimately returned to the government). Data on ID: 1870-1926: Historical Statistics of Canada, Series H19-34: Federal Government budgetary expenditures, classified by function, 1867-1975; 1926-1995: Cansim D11166. 19962000: Cansim D18445. Finally, data for BCI: copied by hand from the Annual Reports of The Bank of Canada, Statement of Income and Expense, Annually, 1935-2000. Net Income paid to the Receiver General (for the Consolidated Revenue Acct). Note: all government data had to be converted from fiscal to calendar years. GSIZE = government size, calculated as (GOV-GRANTS)/(GNP) GRANTS = Transfers to Provinces and Local Governments; 1870 - 1912: From Rowell-Sirois Commission, Subsidies and Grants Paid to Provinces Since Confederation, Table II; 1913-1935: From Rowell-Sirois Commission, Dominion-Provincial Subsidies and Grants, Statistical Appendix, p. 186; 1926 - 2001: Cansim label D11164 and D11165. IMMIG = immigration numbers. 1868 1953: Firestone (1958), Table 83, Population, Families, Births, Deaths; Updated by Cansim D27 (1955 to 1996). Cansim Sum of X100615 (Females) plus X100614 (Men) for 1954;1997-2001, Cansim D27 (sum of quarters). IMMRATIO = IMMIG/POP.

24 IPIUS = Index of Industrial Production for the United States. 1870-1929: Table A15. NBER, Nutter; 1930-1970, Table A16. (BEA) Bureau of Economic Analysis;1971-1995: Cansim D360048 (1987=100);1996-2000, U.S. Department of Commerce, Business Cycle Indicators, Index of Industrial Production 1992=100. LN = the log operator. PRCNTYNG = percentage of the population below 17. 1870-1920 Leacey et al (1983). Interpolated from Census figures Table A28- 45 sum of columns 29, 30, 31, and 32 all divided by 28 (adjusted to make 1921 the same); 1921-2001 Cansim C892547. P = GNP deflator before 1927 and GDP deflator after (1986 = 100). 1870-1926: Urquhart, (1993), 24-25;1927-1995 (1986=100): Cansim data label D14476; 1996-2001 Cansim D140668. All indexes converted to 1986 = 100 basis. POP = Canadian population. 1870-1926: Urquhart, (1993), 24-25; 1927 - 1995: CANSIM data label D31248; 1995 - 2001: Cansim D1 (average of four quarters). RYPC = real income per capita = GNP/(P*POP). 2. List of Political Variable Names and Data Sources: ELYEAR = year of election, 1 if election year 0 otherwise. LIBERAL = 1 if governing party was the Liberal Party, 0 if the Conservative Party MINORITY = 1 if the governing party was part of a minority government, 0 otherwise SEATS = percentage of the seats won by the governing party ELAPSE = the number of years since the last election. Sources: Canadian Parliamentary guide, 1997, 2002; Thirty Seventh General Election 2000, Elections Canada 2001. Beck, Murray, J., 1968, Pendulum of Power Prentice Hall of Canada, Scarborough. Scarrow, Howard A., 1962, Canada Votes: A Handbook of Federal and Provincial Election Data, Hauser Printing Company, New Orleans. Campbell, Colin, 1977, Canadian Political Facts 1945-1976, Methuen: Toronto.

26

Table 1a Descriptive Statistics of Political Variables and Growth Variables

1870 - 2000
ELYEAR ELAPSE MINORITY LIBERAL SEATS GROWTH INFLATION USGRWTH USINF Mean Median Maximum Minimum Std. Dev. ADF levels (4) 5% c.v. -2.88 0.273 0.000 1.000 0.000 0.447 -5.94* 1.561 1.000 5.000 0.000 1.297 -4.84* 0.1288 0.000 1.000 0.000 0.336 -4.03* 0.553 1.000 1.000 0.000 0.499 -3.34 0.5916 0.5880 .7900 .4100 0.093 -3.67* 3.69 4.23 16.5 -12.7 5.10 -5.68* 2.39 2.23 16.7 -12.3 4.68 -3.68* 4.06 5.23 22.9 -24.8 9.64 -7.05* 1.95 1.73 17.98 -13.89 4.71 -3.40

* significant at 1%

Table 1b Descriptive Statistics for the Government Size Variables

1870 - 2000
LNGSIZE LNRYPC LNAGRIC LNIMRATIO LNYOUNG LNOPEN LNPOP
9.29 9.29 10.33 8.19 0.670 -0.011 0.641 -2.50 -7.00*

Mean Median Maximum Minimum Std. Dev. Skewness

ADF stat (4) levels 5% c.v. -2.88 ADF stat (4) 1 st Dif 5% c.v. -2.88 ADF stat (4) 2 nd Dif

-2.523 -2.501 -0.869 -3.571 0.605 0.291 -2.82 -6.97*

8.668 8.466 10.10 7.39 0.809 0.164 0.144 -5.95*

-1.635 -1.091 -0.540 -3.612 1.02 -0.676 1.65 -3.11

-4.90 -4.96 -2.94 -6.91 0.880 -0.422 -2.88 -5.90*

3.59 3.65 3.88 3.14 0.191 -0.842 -0.186 -2.55**

-0.830 -0.872 -0.132 -1.175 0.219 0.987 -0.739 -5.13*

* (**) significant at one (ten) per cent

Table 2 The Effect of Political Variables on the Growth Rate in Canada

(White Heteroskedasticity-Consistent Standard Errors in brackets)

Definition of Political variables

DEPENDENT VARIABLE CONSTANT ELECTION YEAR(-1)
{1-MINORITY}{LIBERAL*(1SEATS) - (1-LIBERAL)*(1-SEATS)}

A
GROWTH of RGNP 1870 - 2000 1.38* (0.480) 0.992 (0.805) 2.68** (1.21) -0.025 (0.240) 0.309* (0.048) 0.148* (0.045) 129 0.460 2.31 5.57 0.02**

A
GROWTH of RGNP 1921 - 2000 1.78* (0.543) 0.071 (0.728) 2.68** (1.16) -0.115 (0.252) 0.347* (0.054) 0.132** (0.056) 80 0.462 2.21 5.18 0.13

B GROWTH of RGNP 1870-2000

1.39* (0.481) 1.04 (0.805) 2.46*** (1.26) 0.009 (0.247) 0.309* (0.048) 0.148* (0.045) 129 0.436 2.29 5.58 0.04**

B GROWTH of RGNP 1921 - 2000

1.82* (0.546) 0.146 (0.727) 2.45*** (1.25) -0.075 (0.262) 0.349* (0.054) 0.132** (0.056) 80 0.600 2.25 5.19 0.005*

{ELAPSE*LIBERAL (1-LIBERAL)*ELAPSE)} USGROWTH USGROWTH(-1) Statistics: No. of Observations R2 D.W. Akaike info criterion Wald Prob [c(2) = c(3) = c(4) = 0]

* (**)(***) significantly different from zero at 1% (5%) (10%).

A - Fraction of Seats won by the governing party and minority status as determined by Beck (1968): 1870-1944; Campbell (1977): 1945 - 1978; Canadian Parliamentary Guide (2002): 1979-2000. B - Fraction of Seats won by the governing party and minority status as determined by the Canadian Parliamentary Guide (2002): 1870-2000.

Table 3 The Effect of Political Variables on the Inflation Rate in Canada

(White Heteroskedasticity-Consistent Standard Errors in brackets)

Definition of Political Variables DEPENDENT VARIABLE

CONSTANT ELECTION YEAR(-1)
{1-MINORITY}{ LIBERAL*(1-SEATS) - (1-LIBERAL)*(1-SEATS)}

A Inflation Rate 1870 - 2000

0.721** (0.315) 0.005 (0.427) 0.592 (0.839) -0.017 (0.149) 0.836* (0.051) 130 .723 1.78 4.71 0.833

A Inflation Rate 1921-2000

0.433** (0.235) 0.381 (0.475) 0.846 (0.980) -0.007 (0.180) 0.850* (0.067) 80 .786 1.04 4.35 0.387

B Inflation Rate 1870 - 2000

0.723** (0.313) 0.012 (0.424) 0.722 (0.798) -0.033 (0.140) 0.834* (0.051) 130 0.723 1.79 4.71 0.768

B Inflation Rate 1921-2000

0.451*** (0.236) 0.391 (0.470 1.167 (1.017) -0.035 (0.170) 0.840* (0.069) 80 .788 1.05 4.34 0.330

{LIBERAL*ELAPSE (1-LIBERAL)*ELAPSE} USINFLATION Statistics: No. of Obs R2 D.W. Akaike Info criterion WALD Prob (c(2) = c(3) = c(4) = 0)

* (**)(***) significantly different from zero at 1% (5%) (10%).

A (B) fraction of Seats held by the governing party and minority governments defined as in Table2.

Table 4 Long Run Model of Government Size

Canadian Federal Government Expenditures as a fraction of GDP: 1870 - 2000
(t-statistics in brackets) (1) LNGSIZE: Base Case 1870 - 2000 (2) LNGSIZE: Base Case 1870 - 2000
-16.25 (7.00) 1.263 (8.18) 0.648 (5.32) 0.724 (1.86)

(3) LNGSIZE Base Case 1921 - 2000 -35.08 (11.91) 3.140 (10.46) 1.47 (7.90) 1.573 (5.03) -0.349 (8.02)

(4) LNGSIZE Base Case 1921- 2000 -31.16 (9.34) 2.818 (8.70) 1.449 (7.99) 1.181 (3.38) -0.374 (8.56) 0.319 (2.30)

(5) LNGSIZE: Political 1870 - 2000 -20.45 (10.33) 1.371 (9.70) 0.450 (4.28) 1.404 (4.25) -0.201 (6.14)

(6) LNGSIZE: Political 1921 - 2000 -34.91 (11.65 3.11 (9.09) 1.45 (7.39) 1.58 (5.02) -0.343 (7.33)

Constant LNRYPC LNAGRIC LNYOUNG LNIMRATIO WWII Aftermath SEATS

-18.97 (9.28) 1.379 (9.26) 0.514 (4.69) 1.184 (3.46)

-0.220 (6.46)

-0.224 (6.68) 0.340 (2.34)

1. 25 (3.84)

0.132 (0.386)

Equation Statistics: No. of Observations Adj. R2 D.W. Akaike info criterion A.D.F. statistic on residuals (MacKinnon critical value at 5 % for 5 vars is -4.17) (c.v. at 5 % for 6 vars -4.52)

131 .716 0.279 0.612

131 .726 0.308 0.584

80 .708 0.589 0.313

80 .724 0.632 0.269

130 0.744 0.410 0.516

80 .704 0.593 0.336

-4.63

-5.25*

-5.77*

-6.72* -3.768 -5.83*

Note: The standard errors in these regressions are inconsistent because of correlations arising among the random components of the I(1) variables and hence are unreliable for use as significance tests. In the companion table to follow Saikkonens method is used to adjust for the inconsistency of the standard errors of the base case estimators by including the contemporaneous, lagged and led values of the first differences of all variables (both left and right hand side variables) in the estimating equation and readjusting the standard errors. * significantly different from zero at 1%. Sources: See Appendix

Table 5 Saikonnen Supplement to Table 4

Canadian Federal Government Expenditures as a fraction of GDP: 1870 - 2000
(Saikonnen adjusted t-statistics in brackets) #

(1) LNGSIZE: Base Case from Table 4, Col. (1) 1870 - 2000 Constant LNRYPC LNAGRIC LNIMRATIO LNYOUNG WWIIAftermath Equation Statistics: No. of Observations Adj. R2 D.W. Akaike info criterion
Saikonnen adjustment factor

(2) LNGSIZE: Base Case from Table 4, Col. (2) 1870 - 2000 -16.64* (4.09) 1.06* (4.16) 0.307 (1.13) -0.228* (3.99) 1.21*** (1.64) 0.176 (0.552)

-17.59* (4.75) 1.06* (4.17) 0.198 (1.05) -0.233* (4.11) 1.43** (2.26)

128 .856 0.483 0.179 .563

128 0.831 0.484 0.186 .564

# Saikkonens estimator adjusts for inconsistency in the standard errors of the I(1) variables in the cointegrating equation by including the contemporaneous, lagged and led values of the first differences of both left and right hand side variables (with the exception of the dummy variable WWIIaftermath that tests for a break in the relationship). Only the coefficients of the level terms are relevant and so presented. In addition, the standard errors and hence tstatistics, had to be adjusted for the presence of correlation among the innovations of the I(1) variables by a factor formed by the ratio of two standard errors: the standard error of the augmented equation divided by the long run standard error. The latter is calculated as the square root of the variance plus two times the weighted sum of the significant autocovariances among the residuals (Saikonnen, 1991). In this particular example, Saikonnens adjustment led to the estimated t-statistics in column (1) being adjusted by the factor .610 while the t-statistics in (2) were adjusted by the factor .618. *(**) [***] significantly different from zero at 1(5) [10] %

Table 6 Error Correction Model of Government Size

The Change in Canadian Federal Government Expenditures as a fraction of GDP: 1870 - 2000
( t-statistics in brackets)

(1) Base Case 1871 - 2000

Dependent Variable Constant D(LNRYPC) D(LNRYPC)(-1) D(LNIMRATIO) D(LNGRANT_ SHARE) Error Correction term D(LNGSIZE)(-1) ELYEAR(-1) LIBERAL MINORITY
ELAPSE*LIBERAL -ELAPSE*NONLIB

(2) Equations 1921 - 2000

D(LNGSIZE)
0.002 (0.172)

(3) 1945 - 2000

D(LNGSIZE)
0.023*** (1.76) -0.949 (3.09) 0.562 (1.55)
-0.109* (4.17) -0.599* (1.62) -0.069*** (1.62) 0.406* (8.95)

(4) Political 1871 - 2000

D(LNGSIZE)
-0.158** (1.94) -0.599* (3.31) 0.636* (3.09)
-0.120* (4.70) -0.370* (6.83) -0.112* (3.62) 0.316* (5.37) -0.010 (0.530) 0.011 (0.392) 0.032 (1.03) -0.002 (0.240) 0.260** (2.03)

(5) Variables 1921 - 2000

D(LNGSIZE)
-0.075 (1.09) -0.556** (2.18) 0.741* (2.57)
-0.146* (3.89) -0.287* (4.33) -0.092* (2.50) 0.382* (5.59)

(6) 1945 - 2000

D(LNGSIZE)
-0.065 (1.53) -0.961* (3.25) 0.502 (1.45)
-0.091* (3.46) -0.597* (9.26) -0.084** (2.01) 0.402* (7.27)

D(LNGSIZE)
0.002 (0.239) -0.544* (3.14) 0.729* (3.77)
-0.127* (5.04) -0.374* (7.04) -0.097* (3.32) 0.333* (5.79)

-0.527** (2.08) 0.810* (2.86)

-0.158* (4.37) -0.291* (5.79) -0.080** (2.27) 0.393* (5.79)

SEATS
Equation Statistics: No. of Observations Adj. R2 D.W. Akaike info criterion

0.134 (1.14) 80 0.655 2.31 -1.73

0.153** (2.18) 56 .845 1.49 -2.96

129 .654 2.28 -1.79

80 .654 2.29 -1.74

56 0.834 1.40 -2.90

129 0.652 2.29 -1.76

* (**) [***] significant at 1% (5%) (10%)

32

References
Alesina, A., N. Roubini, with G.D. Cohen (1997) Political Business Cycles and the Macroeconomy, Cambridge: MIT Press. Bartels, L.M., and H.E. Brady (2003) Economic Behavior in Political Context, American Economic Review, 93: 156 - 161. Borcherding, T. E. (1985) "The Causes of Government Expenditure Growth: A survey of the U.S. Evidence", Journal of Public Economics, 28: 359 - 382. Borcherding, T. E., J. S. Ferris and A. Garzoni (forthcoming 2004) Changes in the Real Size of Government Since 1970", in Juergen Backhaus and Richard Wagner (eds.), Kluwer Handbook in Public Finance New York: Kluwer Academic Press. Engle, R.F. and C.W.J. Granger (1987) Co-integration and Error Correction: Representation, Estimation, and Testing, Econometrica 55, 2: 251-276. Ferris, J. S. and E. G. West (1996) "Testing Theories of Real Government Size: U.S. Experience, 1959 1989," Southern Economic Journal, 62: 537-553. Ferris, J. S. and S. Winer (2003) Searching for Keynes: With Application to Canada, 19870 - 2000" CESifo Working Paper No. 1016, 35 pages. Gregory, A. W. and B. E. Hansen (1996) Residual-Based Tests for Cointegration in Models with Regime Shifts, Journal of Econometrics 70, 1: 99-126. Heckelman, J.C. (2002) Electoral Uncertainty and the macroeconomy: the evidence from Canada, Public Choice 113: 179-189. Kau, J. B. and P. H. Rubin (1981), "The Size of Government", Public Choice 37: 261-74. MacKinnon, J. G. (1996) Numerical Distribution Functions for Unit Root and Cointegration Tests, Journal of Applied Econometrics 11, 6: 601-618. Mueller, D.C. (1989) Public Choice II, Cambridge: Cambridge University Press. Peacock, A. and J. Wiseman (1961) The Growth of Public Expenditure in the United Kingdom, Princeton: Princeton University Press. Rodrik, D. (1998) Why do more open economies have bigger governments? Journal of Political Economy 106, 997-1032. Saikkonen, P. (1991) Asymptotically Efficient Estimation of Cointegration Regressions, Econometric Theory 7, 1: 1-21. Sapir, A. and K. Sekkat (2002) Political cycles, fiscal deficits, and output spillovers in Europe Public Choice 111: 195-205.

33

Notes

1. An earlier version of this paper was given at the Public Choice Society annual meetings in Baltimore, March 2004. We would like to thank our discussant Anna Rubinchik-Pessach and Christian Bjornskov for their comments. 2. The unemployment rate is another variable that is often used. In Canada, however, unemployment rates are available only from the 1920's onward and so do not span our time period. For this reason we restrict our attention to the rate of growth of real income/output and the inflation rate. 3. Most other Canadian studies cover distinctly shorter time periods. For example, one of the most recent studies of the interaction between electoral outcomes and the macroeconomy in Canada (Heckelman, 2002) covers only the 1965 through 1996 time period. 4. After reporting on the effects of partisan differences in the U.S., Bartels and Brady (p.159) write, (o)ne might imagine economists reacting to Hibbs work by launching a major effort to understand the processes by which partisan politics shapes economic policies and performance. Unfortunately, no such effort seems to be forthcoming... 5. The standard deviation of the inflation rate fell from 5.24 over the 1870 to 1938 time period to 3.26 in the 1945 to 200 time period. The fall is even larger for the growth rate, with the standard deviation of the growth rate falling from 6.17 to 2.73. The correlation coefficient over thje whole period is .127. 6. Note that the post world war two time period also corresponds to the adoption of Keynesian counter-cyclical fiscal policy and the emergence of the welfare state. At the same time the growth in media coverage through television could have revolutionized the impact of politics on economic life. For an analysis that distinguishes between at least two of these possibilities, see Ferris and Winer (2003). 7. All unit root tests use four lags. 8. Canadian data allows no complete series for the percentage of the population older than sixty five between 1870 and 2000. However, we were able to find the opposite, i.e., the percentage of the population that is young. For this reason we expect to find the exact opposite prediction. 9. In our preliminary work we also tried using the share of federal transfers in total federal spending (LNGRANT_SHARE) as an additional explanatory variable. However, although this variable was consistently negative in its effect on federal size (and significantly so), its proposed inclusion in the cointegrating vector was never consistent (with cointegration). For this reason we excluded it from our later work. 10. Because our data is annual and because most elections took place mid-year, it is not clear that the hypothesized boost to aggregate demand prior to an election would have taken place in

34 the previous as opposed to the current year (i.e., in January - May for a June election). Hence all equations were rerun without the lag in ELYEAR. This made a difference only in relation to inflation, see footnote seven below. 11. Our test does not include an incumbency effect where the size of the surprise is viewed as biassed against the incumbent governing party winning reelection. See Heckelman (2002) who argues that the Canadian data (1965-1996) is more consistent with symmetry across parties. 12. To give one example of the difference between the two series in our data, in the election of 1872 the Conservative party was elected where its majority (minority) consisted either of 99 Conservatives out of 200 seats (as determined by the Canadian Parliamentary guide) or as 99 Conservatives plus three Independent Conservatives plus one independent (103) out of 200 seats (as determined by Beck). Different political scientists then read the effective size of the majority in different ways. 13. The equations were also rerun over the 1945 to 2000 time period with no appreciable change in results. That is, unlike the analysis to follow, there is no suggestion in the data that political cycles are more prevalent now than they were earlier. 14. Use of the current election year, rather than the year prior to an election, did produce significantly positive coefficients in two of the inflation equations (for 1921-2000) and hence does give some support to traditional opportunism. This improvement, however, did not spillover into similarly better results for the other two hypotheses. 15. Note that the difference in result is due to the narrower political definition of the winning margin. 16. The tests of nonconvergence in these sections use as political variables only the first (Beck, Campbell and Canadian Parliamentary Guide) definition of minority governments and sizes of winning majorities. The tests were rerun for the competing definition (Canadian Parliamentary Guide throughout) with no appreciable change in results. 17. As far as we are aware, there are no tables of critical values for cointegration relationships with structural breaks occurring at known break points. However, Gregory and Hanson (1996) do give approximate critical values for the cointegration test of Engle-Granger type cointegration with a structural break arising at an unknown point as -5.56 (for m = 4). The appropriate critical value should then be somewhere between the critical value of -4.77 (no break) and the critical value for a cointegration relation with one unknown break (-5.56). Our adjusted Dickey Fuller test statistic of -5.25 is somewhat closer to the lower of these bounds. 18. The year 1921 begins the period after which annual observations were available for each variable in the full data set (and not constructed by interpolation across census dates --usually every 5 to 10 years). 19. Note that the same theories of asymmetric information that suggest that opportunistic cheating cannot be effective in the long run also find that under asymmetric information such unnecessary spending behaviour will necessarily be embodied in the resulting equilibrium. That is, if incumbent parties do not spend (as expected) before an election, then aggregate

35 demand would be affected adversely with resulting party losses. Asymmetric information traps both political parties and the electorate into an equilibrium that is second best (compared to a world where such inefficiencies could be eliminated costlessly). 20. Again because the discreteness of the year means that there will be ambiguity over which year represents pre-election spending, we use both years. There are no differences in our findings if ELYEAR is used rather than ELYEAR(-1). 21. Although the standard errors and t-statistics of the OLS equations are inconsistent, the Saikonnen correction only decreases rather than increases the coefficient t-statistics. Hence correction for correlation among the variables will not improve the probability that SEATS is a significant determinant of size. 22. Finally, note that the reason why SEATS would appear in a cointegrating relationship of this type is not clear from the hypotheses presented by the political cycle literature. While the presence of SEATS in the cointegrating relationship would suggest the plausible hypothesis that larger winning margins produce larger government sizes, it is not clear how that hypothesis would be attached to opportunism or partisanship. 23. Once again the equations were rerun for the Canadian Parliamentary guide definitions of MINORITY and SEATS. There were no significant change in our findings. 24. Once again, ambiguity in the timing of elections led us to run the equations with both ELYEAR(-1) and ELYEAR with no appreciable difference in results.