Direct democracy and the representative system

The relationship between direct democracy and public spending is the subject of a rich literature. According to (Vatter et al. Reference Vatter, Rousselot and Milic2019, 177–178), there are basically two schools of thought. Highlighting the veto player function of direct democracy, especially in the form of financial referendums (cf. Hug and Tsebelis Reference Hug and Tsebelis2002), several scholars observe a negative effect on public spending (e.g. Romer and Rosenthal Reference Romer and Rosenthal1979; Matsusaka Reference Matsusaka1995; Wagschal Reference Wagschal1997; Obinger Reference Obinger1998; Feld and Matsusaka Reference Feld and Matsusaka2003; Funk and Gathmann Reference Funk and Gathmann2011; Funk and Gathmann Reference Funk and Gathmann2013). Other scholars, in the tradition of Downs (Reference Downs1957), argue that more direct participation, especially in the form of popular initiatives, may lead to more state intervention due to democracies’ inherent tendency to redistribute (e.g. Freitag and Vatter Reference Freitag and Vatter2006; Blume et al. Reference Blume, Müller and Voigt2009; Blume and Voigt Reference Blume and Voigt2012; Asatryan Reference Asatryan2016; Asatryan et al. Reference Zareh, Baskaran, Grigoriadis and Heinemann2017; Walter Reference Walter2019). In an encompassing literature review, Matsusaka (Reference Matsusaka2018) finds that financial referendums are consistently observed to have a negative effect on state intervention. In contrast, the evidence is more mixed in the case of popular initiatives (see also Morger and Schaltegger Reference Morger and Schaltegger2018). Finally, comparatively little is known about the effect of law referendums.

In this section, we develop a series of theoretical expectations about how popular initiatives, law referendums and financial referendums relate to public spending. Most notably, we explore how these direct democratic institutions interact with the representative system. The literature has repeatedly shown the effect of direct democratic institutions to depend on features of the representative system (e.g. Gerber Reference Gerber1996b; Hug Reference Hug2009; Boehmke et al. Reference Boehmke, Osborn and Schilling2015; Leemann and Wasserfallen Reference Leemann and Wasserfallen2016; Matsusaka Reference Matsusaka2018). Yet, there is comparatively little research focused specifically on understanding which features of the representative system matter. We examine a specific feature, which is of particular importance in shaping political decision-making processes: the size of the governing coalition, here understood as the number of parties forming the government.

The existing literature has shown coalition size to matter for public expenditure (e.g. Volkerink and de Haan Reference Volkerink and de Haan2001; Perotti and Kontopoulos Reference Perotti and Kontopoulos2002; Bräuninger Reference Braeuninger2005; Bawn and Rosenbluth Reference Bawn and Rosenbluth2006; Persson et al. Reference Persson, Roland and Tabellini2007). Conceiving of the political space as being composed of a limited number of societal groups, this literature argues that the groups’ welfare is a function of public spending. Government participation provides political parties with the possibility to channel resources to their constituencies. Yet, spending decisions are made under cost constraints that change with the size of a party’s support base. In small coalitions, parties represent the interest of multiple societal groups that constitute a large part of the population. As a result, they have to balance the demand for different public goods with limited scope to externalise costs (assuming that the financial burden for public spending is borne by all societal groups). In contrast, parties in large coalitions depend on smaller constituencies, respectively, with more homogeneous preferences for spending, allowing them to externalise costs to a larger population share. This simple but powerful argument implies that public spending increases with the number of parties partaking in the governing coalition. ^{Footnote 4}

How does coalition size interact with the relationship between direct democratic institutions and public spending? In principle, we can distinguish between direct and indirect effects of direct democratic institutions (Neidhart Reference Neidhart1970; Gerber Reference Gerber1996a; Matsusaka and McCarthy Reference Matsusaka and McCarthy2001; Matsusaka Reference Matsusaka2014). Consider the case of law referendums, which allow citizens to collect signatures against a newly legislated law. By definition, this institution’s direct effect is to block new legislation and, by doing so, limit the creation of new government action with spending implications (e.g. Immergut Reference Immergut1990; Hug and Tsebelis Reference Hug and Tsebelis2002). Yet, such institutions might also have indirect effects. In the case of law referendums, indirect effects may result from the anticipation of referendum challenges. If governments and legislators are strategic, governments may be enticed to forge oversized coalitions to avoid defeat at the ballot box (e.g. Neidhart Reference Neidhart1970; Gerber Reference Gerber1996a). In this case, the indirect effect of law referendums might, in fact, increase public spending.

We expect direct democratic institutions to differ in the way they interact with the representative system. Therefore, we discuss the interaction between direct democratic institutions and coalition size for each institution separately. We start with popular initiatives. Subsequently, we discuss law referendums and, finally, financial referendums.

Case selection, data and research design

We examine the conditional effect of direct democratic institutions on public spending using subnational (cantonal level) data from Switzerland. Two features make Switzerland a good choice for our analysis. First, all Swiss cantons know direct democratic institutions. Yet, there is considerable variation across time and space. While some cantons introduced these instruments early, others were laggards (Leemann Reference Leemann2019). Similarly, while some cantons make the recourse to direct democratic instruments easy, others feature more restrictive rules.

Second, most public spending is located at the subnational level. Despite several decades of fiscal centralisation, only 32% of public spending is located at the federal level (data for 2016 from the Eidgenössische Finanzverwaltung 2017). In contrast, 41% of public spending is located at the cantonal level (the remaining 27% is primarily located at the municipality level). In addition, there is considerable variation across time and space, as we have shown in Figure 1.

Figure 1.

The development of public spending in Swiss cantons, 1830–2015.

We are, of course, not the first to focus on Swiss cantons to examine the link between direct democratic institutions and public policies. Several researchers have used similar research designs to examine the effect of direct democracy on public spending (e.g. Kirchgässner et al. Reference Kirchgässner, Feld and Savioz1999; Feld and Matsusaka Reference Feld and Matsusaka2003; Freitag and Vatter Reference Freitag and Vatter2006; Funk and Gathmann Reference Funk and Gathmann2011) or other socio-economic outcomes (e.g. Frey Reference Frey1994; Feld and Savioz Reference Feld and Savioz1996; Freitag and Vatter Reference Freitag and Vatter2000).

We have engaged in a considerable data collection effort. First, we have collected public spending data in 5-year intervals from the public budget reports of the cantonal governments, covering the period 1830–1930 (Emmenegger et al. Reference Emmenegger2020). ^{Footnote 8} Subsequently, we have merged our data with that provided by Funk and Gathmann (Reference Funk and Gathmann2011), which covers the period 1890–2000 and were obtained from federal records (most notably the Federal Office of Statistics). Note that no federal records are available before 1890. Finally, we have extended the dataset to cover the period 2000–2015, again based on federal records.

Both data collection strategies have their specific strengths and weaknesses. Cantonal budget reports are closer to the “source”. Yet, there is a risk that cantonal definitions differ as to what belongs to public spending (e.g., separate budgets for cantonal monopolies on the salt trade). Federal records might provide more comparable information. Yet, federal records are more valid only if the data provided by the cantons correspond to the federal guidelines. The overlap of 40 years (1890–1930) in our data and that provided by Funk and Gathmann (Reference Funk and Gathmann2011) allowed us to explore such differences. In some cases, we went back to the cantonal archives to find additional budgets that were not part of the official cantonal budget reports (e.g., separate budgets for public schools in Obwalden). In other cases, we found that cantons reported incomplete data to the federal administration (e.g., Appenzell Innerrhodes reported data only for its finance department, but not public spending by other cantonal departments). Once we had all discrepancies cleared, we merged the data to obtain a dataset covering the period 1830–2015. For the sake of consistency, we use all data in 5-year intervals.

To make the data comparable over time and across cantons, we employ official exchange rates and the consumer price index (prices in 2005 Swiss francs). ^{Footnote 9} In addition, we divided public spending by the cantonal population to account for different canton sizes. Finally, we transformed the data with the natural logarithm. In Figure 1, we plot the growth of spending in all cantons with representative systems. Unsurprisingly, the figure shows a clear upward trend in the period 1830–2015. Yet, we also observe considerable variation across time and space, which we will explore in the empirical section.

Second, we rely on data on the institutionalised rules of direct democratic instruments for the period 1830–2015. For the purpose of the article, we distinguish between popular initiatives, law referendums and financial referendums. We rely on all cantons that had a representative system throughout the period of investigation. We, therefore, exclude the six small cantons that rely on citizen assemblies (Glarus, Nidwalden, Obwalden, Uri and both Appenzell).

Direct democratic institutions vary with regard to the ease to which they can be put to use. The existing literature has used this variation to identify the effect of direct democratic institutions on political outcomes. Researchers have employed proxies for the economic costs of using direct democratic institutions related to variations in signature requirements and collection periods to capture these institutions’ effects (e.g. Stutzer Reference Stutzer1999; Fehr and Gächter Reference Fehr and Gächter2000; Torgler Reference Torgler2002; Frey and Stutzer Reference Frey and Stutzer2010). More precisely, the higher the signature requirements (relative to the population of the relevant political unit) and the shorter the collection period, the higher the economic costs of using direct democratic institutions.

Our measures for popular initiatives and law referendums are composite indices and come from Leemann (Reference Leemann2019). Both indices range from 0 to 6 with higher values displaying lower “costs” of using direct democratic instruments, i.e. lower signature requirements and longer signature collection periods. Figure 2 shows the extent to which cantonal constitutions grant the citizenry participatory rights. Red lines capture law referendums, while blue lines capture popular initiatives. Higher values indicate that it is easier to force a ballot vote on a policy that has been passed in parliament (law referendum) or force a popular vote on a specific proposal (popular initiative). We provide more detailed information on the indices in the appendix (see Sect. 6.2 in supplementary material).

Figure 2.

The extent of direct democratic institutions in Swiss cantons, 1830–2015. Note: Red lines show values for law referendums and blue lines show values for popular initiatives.

We also examine the effect of financial referendums on public spending. Existing analyses have often used rather simple measures of financial referendums, typically a dummy variable capturing its existence (e.g. Funk and Gathmann Reference Funk and Gathmann2011). Instead, we use the deflated monetary threshold per capita for financial referendums (for a similar approach, see Feld and Matsusaka Reference Feld and Matsusaka2003). Put differently, we take advantage of the fact that the “stinginess” of financial referendums varies across cantons and time, with some cantons enabling financial referendums in the case of small expenditures, while others rule out the financial referendum even for relatively large expenditures. Figure 3 shows how the deflated monetary thresholds per capita developed in the period 1830–2015. In line with the literature, we expect the financial referendum to have a stronger effect in the case of low monetary thresholds (deflated and per capita). As a consequence, we expect the coefficient of our indicator of financial referendums to be positive (because high thresholds indicate a comparatively permissive financial referendum).

Figure 3.

Deflated monetary thresholds of financial referendums (per capita) in Swiss cantons, 1830–2015. Note: The canton of Neuchatel (NE) has a different scale on the y-axis.

This operationalisation implies that we are not working with an index comparable to the ones used for popular initiatives and law referendums. The reason is that in many Swiss cantons, a popular vote is mandatory if a policy surpasses a certain predefined monetary threshold. As a consequence, in these cases, there are no signature requirements and collection periods. Instead, we capture variation across cantons and over time by means of the deflated monetary threshold per capita triggering a financial referendum.

In the case of financial referendums, we are, however, confronted with a challenge. If the financial referendum has not been adopted (yet), we have no data for the monetary threshold. This is not a problem with our indices of popular initiatives and law referendums, which conceptualise the recourse to popular initiatives and law referendums in terms of associated “costs”. In this operationalisation, the absence of popular initiatives and law referendums, respectively, corresponds to the highest possible costs on the index (score is 0). In the case of financial referendums, we follow two strategies to deal with such cases. First, in the absence of a financial referendum, we assign a value of 0 to the monetary threshold. Second, we add a binary indicator that takes on the value of 1 when a canton does not have the financial referendum institution. This is the modified zero-order regression described by Greene (Reference Greene1993) and Maddala (Reference Maddala1977) as a solution to partial missingness.

Third, we rely on data on the political composition of cantonal governments for the period 1848–2017 (Walter and Emmenegger Reference Walter and Emmenegger2019). ^{Footnote 10} In line with the previous literature (e.g. Bawn and Rosenbluth Reference Bawn and Rosenbluth2006), we use the number of parties in government as a theoretically relevant predictor of public spending. In addition, we control for the seat share of left parties to control for partisan effects on public spending (Schmitt and Zohlnhoefer Reference Schmitt and Zohlnhoefer2019). Finally, the previous literature has linked parliamentary fragmentation to higher government spending, especially in political systems in which governments are directly elected (Eslava and Nupia Reference Eslava and Nupia2017). To proxy parliamentary fragmentation, we use a binary variable to capture whether the proportional representation is employed to elect representatives for the legislative.

In addition, we use a number of control variables to account for socio-economic influences on government spending. Given that GDP data is not available, we use the size of the first and second sectors, infant mortality and physician density as proxies. Furthermore, we measure demographic pressures with the dependency ratio, i.e. the share of the population below the age of 15 and above 64. We also include the population’s size, transformed by the natural logarithm to capture changes in the denominator of our dependent variable. Lastly, we use the logarithm of federal subsidies per capita. A table with summary statistics is presented in the appendix (see Sect. 6.1 in supplementary material).

The combined data covers the period 1860–2015 and is measured in 5-year intervals. To estimate the effect of direct democratic instruments on public spending, we rely on an error correction model and employ the following specification. First, we include canton and year fixed effects (FE) to control for unobserved constant heterogeneity and common shocks. Second, we add a lag of the outcome variable (public spending) to all specifications (LDV). Finally, standard errors are clustered by cantons.

Empirical tests

In this section, we examine our expectations. All models that we employ can be described by the following equation:

$$\Delta {y_{i,t}} = {\beta _{LDV}}{y_{i,t - 1}} + \beta {X_{i,t}} + {\alpha _i} + {\gamma _t} + {\varepsilon _{i,t}},$$

whereas $${\alpha _i}$$ are unit FE, $${\gamma _t}$$ are time FE and $${\varepsilon _{i,t}}$$ is a normally distributed error term. Our main independent variables as well as our control variables are captured by the stacked matrix $${X_{i,t}}$$ with a vector of coefficients $$\beta $$ . This specification helps to isolate the relationship between cantonal spending and the institutional setting. To shield against spurious correlations due to common trends, we add unit as well as time FE. General levelling effects are absorbed by lagged per capita spending $${\beta _{LDV}}{y_{i,t - 1}}$$ . ^{Footnote 11}

Model 1 in Table 1 includes all variables as well as the interaction of the popular initiative with the number of parties in government. In Model 2, the interaction is set up with the law referendum. Model 3 looks at the interaction with the financial referendum. Finally, Model 4 includes all three interactions simultaneously. The findings are consistent across all specifications displayed in Table 1.

Table 1.

Direct democracy and government spending, 1860–2015

***p < 0.001, **p < 0.01, *p < 0.05, ^p < 0.1.

Three observations can be made based on Table 1. First, there is clear empirical support for the argument that the association of popular initiatives and public spending is dependent on the number of parties in government. When only a few parties are in government, popular initiatives have a positive effect on public spending, but this positive effect becomes weaker as the number of parties in government increases. Second, we cannot observe any statistically significant relationship between law referendums, coalition size and public spending. Third, the models suggest that the effect of financial referendums is not dependent on the number of parties in government. In all models, the association is significant with the expected (positive) sign, whereas the interaction terms in Models 3 and 4 do not reach conventional levels of statistical significance.

In the appendix, we present a number of robustness checks in Sect. 6.4 in supplementary material. We reestimate the same models but also include the small citizen assembly cantons in Table 6 in supplementary material. The inclusion of these small cantons – which also have very extensive direct democratic rights – does not alter the substantive results. In Table 7 in supplementary material, we drop the highly urbanised cantons Basel and Geneva from our sample. In addition, we exclude all observations for Neuchâtel after 1970 (see Table 8 in supplementary material) because the canton displays unusually high values for the financial referendum (see Figure 3). These additional model specifications yield substantively identical results. Put differently, the estimation results are in line with the results presented here.

In addition, we also explore whether these estimates are stable over 155 years or whether there is any temporality to them (Matsusaka Reference Matsusaka2004). Below, we show that these results remain stable in the face of three independent modelling strategies, all aimed at uncovering temporal change in the estimates. However, first, we take a closer look at the interaction terms.

Interaction terms

We now turn to the interpretation of the interaction terms. In this section, we focus on illustrating the short-run effects of direct democratic institutions for different coalition sizes. Our discussion of the magnitude of the long-run effects under different scenarios can be found in the appendix (Sect. 6.5 in supplementary material).

We first examine the effect of popular initiatives. Across all specifications presented in Table 1, we find a conditional effect of popular initiatives on public spending. Figure 4 displays the marginal effect of popular initiatives conditional on the number of parties in government. We find that for small governments consisting of only one or two parties, there is a positive marginal effect for the popular initiative. This positive effect turns negative as the number of parties in government increases. This result is in line with our theoretical argument.

Figure 4.

Marginal effect of popular initiatives. Note: Based on estimates from Model 4. Simulated marginal effect with 90 and 95% confidence intervals.

Figure 5 displays the marginal effect of law referendums conditional on coalition size. The marginal effect increases for larger coalitions. Interestingly, the effect is significant for large coalition governments. These results provide some limited support to the argument that in anticipation of law referendums, governments try to foster large bill-specific coalitions, which in turn are linked to higher spending.

Figure 5.

Marginal effect of law referendums. Note: Based on estimates from Model 4. Simulated marginal effect with 90 and 95% confidence intervals.

Finally, Figure 6 displays the marginal effect of financial referendums conditional on coalition size. The marginal effect appears to be decreasing, but the interaction parameter is not distinguishable from 0. In addition, the marginal effect is straddling the threshold for significance. In the case of two-party coalition governments, the effect is significantly different from zero. Together with the results displayed in Table 1, this finding suggests that financial referendums are linked to lower public spending, which is consistent with our theoretical expectations (remember that we examine the monetary threshold triggering a referendum). However, we do not find strong evidence that the effect of financial referendums is conditional on coalition size.

Figure 6.

Marginal effect of financial referendums. Note: Based on estimates from Model 4. Simulated marginal effect with 90 and 95% confidence intervals.

Error correction models have a dynamic component. Yet, the shown estimates, as well as the visualisations of the interaction terms, display only the short-run effects. In the appendix, we, therefore, discuss the long-run effects of all direct democratic instruments (see Sect. 6.5 in supplementary material).

Stability over time

Analysing data over more than 150 years immediately leads to the question of whether these estimates are stable over time or whether one is missing part of the historical processes by imposing a too strict model specification (Wawro and Katznelson Reference Wawro and Katznelson2014). The models presented so far all rely on fixed parameters over time, thus assuming that the effect of direct democratic institutions remained stable over time. At first, this might seem like a heroic assumption, but it is in line with what we typically assume when we analyse the effects of electoral rules on strategic behaviour (cf. Cox Reference Cox1997).

However, rather than just asserting this point, we also conduct a number of empirical tests that are likely to show if these effects vary over time. In total, we employ three different strategies to show this overtime stability. First, we explicitly interact explanatory factors with time variables (and the square thereof) to allow for different functional forms. Second, we estimate the model only on slices of the data and test whether estimates on early data are different from estimates on later data (Egger et al. Reference Egger, Nigai and Strecker2019). Finally, we use hierarchical models and model parameters as random effects and inspect how their time disturbances cluster (Wawro and Katznelson Reference Wawro and Katznelson2014).

Interacting with year variable

In the first step, we interact the interaction terms of the baseline specification in Model 4 with a binary indicator for the 19th century. This simple test should allow us to see if there is any difference between the 19th and the 20th centuries. As an alternative and more flexible approach, we also interact the term with the year and in subsequent models with squared years.

None of these interactions with century dummies, years or years squared are significant (see Table 2). There is thus no indication that the effects vary over time. While this is a somewhat restrictive approach, as we impose a clear functional form, we also present two more flexible approaches in the next two subsections.

Table 2.

Temporal heterogeneity tests (controls and constitutive terms omitted)

***p < 0.001, **p < 0.01, *p < 0.05, ^p < 0.1.

Differences between periods

Following Egger et al. (Reference Egger, Nigai and Strecker2019), we divide data into different folds and test whether the parameter estimates are different. We split the time up into an early and into a late period. We then reestimate Model 4 from Table 1 on both folds and compare the two estimates for the interaction term (delivering $${\hat \beta _{int,{\rm{early}}\,{\rm{fold}}}}$$ and $${\hat \beta _{int,{\rm{late}}\,{\rm{fold}}}}$$ ). In the second step, we test whether these two parameters are different from each other. We repeat these two steps while moving the cut-off between early and late periods from 1880 up to 2000, each time moving ten years forward per step.

For all three interactions, we analyse whether there are any differences between estimates based on early folds and late folds. The results of this explorative exercise are displayed in Figure 7. The top panel shows the results for popular initiatives, the middle panel shows the results for law referendums and the bottom panel shows the results for financial referendums. Across all three panels, there is no clear evidence that the mechanism changes considerably over time.

Figure 7.

Overtime stability. Note: Top panel: popular initiative; middle panel: law referendum; bottom panel: financial referendum. The cut-off year (separating early and late fold) is shown on the x-axis. For each test, we take 1,000 draws from the posterior vector and then compare the two draws. The figure shows the 95 and the 99% confidence interval of the difference of the two coefficients (early versus late fold).

For the initiative, there might be a difference depending on the applied $$\alpha $$ level threshold. However, as soon as we exclude the three WWII time periods, no single comparison is significant (see Sect. 6.6 in the appendix in supplementary material). From this second approach, which is less restrictive than the first one, we take that there is no empirical evidence that these mechanisms change over time. We move next to an alternative nonrestrictive approach, where we try to impose a less functional form.

Random coefficient modelling

Our final test for stability builds on a hierarchical modelling strategy, where we include the relevant parameters as random effects. Following Wawro and Katznelson (Reference Wawro and Katznelson2014), we use a hierarchical modelling approach to see if there is a temporal drift in the disturbance estimates of the relevant parameters. Specifically, rather than estimating the interaction term as a fixed $${\beta _{int}}$$ , we model it as a random coefficient $$\beta _{int}^*$$ , whereas $$\beta _{int,t}^* = \beta _{int,t}^0 + {\nu _t}$$ and $${\nu _t}$$ follows a normal distribution with mean 0 and a variance that is estimated by the model.

Figure 8 shows the realisation of the random component of the random coefficient of the interaction term for the three institutions. The top panel shows the effects for popular initiatives, the middle panel for law referendums and the bottom panel for financial referendums. We see neither any shock nor any drift over time.

Figure 8.

Overtime stability in random shocks to parameters. Note: Top panel: initiative; middle panel: law referendum; bottom panel: financial referendum. Plot shows realisation of random component of the interaction term when estimated as a hierarchical model using the BRMS package (Bürkner Reference Bürkner2018).

In this section, we have provided three tests on whether the main findings of this article change over time. We have employed three different strategies that vary in how much the model imposes on the data. None of these tests indicates that there is a change over time. The observed effects remain stable from the mid-19th century over the 20th century into the 21st one.