Turnout and voting behaviour in constitutional referendums: a regional analysis of the Italian case

ABSTRACT This paper investigates regional disparities of both turnout and voting behaviour in constitutional referendums. The analysis is undertaken at the NUTS-3 level and it considers the three constitutional referendums held in Italy in the period 2001–2016. It finds that turnout was lower in provinces with higher unemployment rates and where citizens had a stronger affiliation to opposition parties. These factors, along with level of government popularity, were important drivers of referendum results, especially in 2006 and 2016 when the referendums were rejected. In addition, while the three referendums implied different effects for rich and poor regions, mainly due to decentralization of powers, the local voting patterns did not reflect this. Overall, these findings suggest that the merit of the constitutional reforms played little part in explaining the outcome of the referendums.


INTRODUCTION
Referendums are considered a form of direct democracy, which give citizens the opportunity to participate actively in issues and strengthen their involvement in politics. The use of referendums is increasing around the world (Boyd, 2010). Referendums have been called on different topics, including territorial, constitutional, moral and other issues. Constitutional referendums are particularly important because they may affect national constitutions. However, although the use of referendums to settle major constitutional changes has sharply increased over time (Tierney, 2009), there have been very few studies on this type of referendum.
A large number of studies have examined turnout and voting behaviour in general elections and other types of referendums, but little attention has been paid to constitutional referendums. General elections and referendums, however, are different in three main ways, which may affect the voting decisions of citizens (Laycock, 2013). First, referendums concern issues rather than candidates or parties. Second, frequency and predictability of referendums varies greatly around the world, with some countries holding referendums far more frequently than others. Finally, rules governing referendums are often ad hoc, with considerable discretion left to politicians. Elections, however, are mandatory and their results binding. Constitutional referendums also have some specific features that make them unique among referendums. They usually concern very important and complex issues. Citizens often have a low level of knowledge about these issues and, therefore, find it difficult to make a decision that is consistent with their political values and economic interest. This can give rise to a high risk of mistakes in voting behaviour (Tierney, 2009). An empirical analysis by Pattie, Denver, Mitchell, and Bochel (1999) confirmed that the determinants of turnout and voting behaviours in constitutional referendums differ at least in part from those in general elections. For example, in the Scottish devolution referendum in 1997, turnout was affected by the factors that affect turnout in general elections, but also by feelings about devolution. Voting behaviour was also strongly related to partisanship and sense of national identity. These arguments make it interesting to investigate whether turnout and voting are different in constitutional referendums than in other types of referendums and general elections. This paper, therefore, focuses on constitutional referendums, and aims to shed a light on the factors that shape turnout and voting behaviour in local areas. More specifically, we tested the hypothesis that political and economic variables, as well as government popularity, were the main drivers of regional turnout and outcome in constitutional referendums, while the merit of the proposals had little impact on how local areas voted. Recent studies have recognized that geography is an important dimension affecting referendum results. Huggins (2018), for example, analysed how local areas voted in the 2016 Brexit referendum in UK, emphasizing that Scotland and Northern Ireland favoured 'remain'. Given that only a few papers have considered constitutional referendums, our analysis was explorative in nature and presented an initial investigation on the geographical drivers of turnout and outcome in constitutional referendums.
We considered the three constitutional referendums held in Italy (in 2001Italy (in , 2006Italy (in and 2016 and used econometrics to identify the determinants of their turnout and outcome. Italy is an interesting context of analysis to explain regional heterogeneity of citizens' attitude towards referendums because of the important socioeconomic disparities across Italian regions, which are particularly strong between the north and the south of the country. The three referendums were also on similar issues and had different consequences for different areas of the country.
As explanatory variables, we used the main factors identified by the literature on general elections and referendums, including socioeconomic variables, political recommendations, popularity of government and views on the referendum issues. The results suggest that the variables explaining turnout and outcome in general elections and other types of referendums are only partially relevant in constitutional referendums. We provide empirical evidence that turnout in constitutional referendums is greatly affected by unemployment, both short and long term, and political factors such as citizens' affiliation to opposition parties. Provinces with higher unemployment rates and strong affiliation to opposition parties showed lower turnout. The same two factors, as well as government popularity, determined the rejection of the proposals in the 2006 and 2016 referendums. In addition, despite that the constitutional proposals implied different consequences for rich and poor areas of the country, local voting patterns were not consistent with these effects. Taken together, these findings suggest that the two referendums were rejected because of discontent with the governments of the day, which were not very popular, rather than because the majority of citizens had a negative opinion of the issues proposed.
The 2001 referendum, however, received a majority because of very weak opposition by nongovernment parties. In this case, turnout was very low, and this allowed government parties to win a majority of 'yes' votes. Therefore, we suggest that if a government does not have popular support, a constitutional referendum may be rejected more for political reasons and discontent than on the merit of the questions.
Our findings are consistent with those of Pattie et al. (1999), who emphasized the differences between general elections and constitutional referendums, and recognized the importance of partisanship in shaping referendum results. Our analysis suggests that economic factors and government popularity also have a crucial role in determining turnout and outcome of constitutional referendums, but the merit of the proposed reforms is less important.
This study contributes to the literature in several ways. First, it sheds a new light on the factors affecting turnout and voting behaviours in constitutional referendums at regional level. It is not clear from previous studies whether or not the factors affecting turnout and outcome are different in referendums and general elections. Some studies have suggested that there are no or only marginal differences (e.g., LeDuc & Pammett, 1995), while others have emphasized that referendums are distinct from general elections (Laycock, 2013). Research has also shown that turnout varies from referendum to referendum, depending on the issues at stake (Neijens, van Praag, Bosveld, & Slot, 2007). In Switzerland, for example, the lowest turnout after the First World War was in a referendum on economics, while the highest was on retirement pensions (Kobach, 1994). Voting behaviours can also be affected by the type of referendum. Rational choice theory suggests that information costs on political issues are high compared with the benefits of a single vote (Laycock, 2013). Citizens lack the time, interest and capacity to access information about referendum issues and struggle to evaluate the consequences of their vote (Clarke, Kornberg, & Stewart, 2004).
Moreover, although there is a strong heterogeneity in how local areas voted in referendums, little attention has been paid to the geography of the vote. Among the few papers on the topic, Dhingra, Machin, and Overman (2017) focused on the local economic impact of Huggins (2018) investigated the impact of European Union regional spending on the outcome of Brexit referendum. This paper contributes to fill this research gap.
Our analysis also contributes to the debate on the appropriateness of referendums as a form of direct democracy. This concern is common to all types of referendums but is even more severe for constitutional referendums. This type of referendum often concerns very important and complex issues, and ordinary people may lack the knowledge and competences required to make informed decisions. We have, therefore, contributed to an open debate on referendums, and particularly whether they are opportunities to express preferences about the merit of the proposals or views about the government at the time.
Despite the increasing use of referendums worldwide, the literature has tended to focus on the countries that use referendums more extensively, such as Switzerland, the United States and the UK. Focusing only on these countries can bias the results, so it is important to investigate other contexts (Jacobs, 2018). In this paper, we consider Italy, which has received little attention. In recent decades, it has become a leader in direct democracy, and referendums have been called to decide major reforms including divorce, abortion and anti-corruption. It is, therefore, an interesting context to study (Boyd, 2010).
The paper is structured as follows. The next section reviews the literature on the drivers of both turnout and outcome in referendums. The third section describes the Italian constitutional referendums held in the period 2001-16. The fourth section presents the empirical strategy as well as the data and variables used in the regression models. The fifth section provides the empirical results. The sixth section concludes and suggests some policy implications.

LITERATURE REVIEW
Previous studies have argued that there are no general models of referendum voting, probably because of the infrequency and uncertainty of referendums (Brown, McCrone, Paterson, & Surridge, 1999;Laycock, 2013). In this section, we have reviewed the literature on the determinants of both turnout and voting behaviour. We considered studies on generic referendums, and also general parliamentary elections, because very few studies have investigated constitutional referendums.

The determinants of turnout in referendums
The literature on turnout and its determinants is essentially empirical. Many variables have been used in previous studies, but none is seen as indispensable. A vote in referendums, therefore, is often considered a 'step in the dark' (Pattie et al., 1999). This is partly because of a lack of theoretical models to drive the selection of explanatory variables. Empirical analyses also provide mixed results on the statistical significance of the individual variables (Geys, 2006). No solid theoretical framework explains why people choose to vote or abstain. This section, therefore, discusses the variables that have been used most often in previous empirical studies.
When considering the factors that affect turnout in referendums, it is useful to distinguish between voter and referendum characteristics. Within the first group of factors, political competence on the issues at stake increases participation in referendums (Neijens et al., 2007). This competence is measured mainly as knowledge about the referendum questions and consequences of the referendum outcomes, as well as ability to explain the vote (Kriesi, 2005). General interest and attitude towards politics is also important in shaping referendum turnout. High trust in the political system, high political interest and knowledge, and high participation in previous elections of representative bodies are positively linked to high turnout in referendums (de Vreese & Semetko, 2004;Johnston, Blais, Gidengil, & Nevitte, 1996;Neijens, Minkman, Slot, Saris, & de Ridder, 1993). The probability of voting is also linked to the perceived importance of the issues (Smets & van Ham, 2013;van Holsteyn, 1996) and to involvement in the referendum campaign (Slot, 1999).
Other socioeconomic variables have been associated with voter turnout in referendums. Several papers have used the resources model of participation in general elections proposed by Verba and Nie (1972), and considered age, education, occupational status and income as important factors affecting turnout (Matti & Zhou, 2017). In general, the evidence is mixed, as shown by Streicher, Schmidt, and Schreyer (2019). They analysed referendums on the decision to bid to host the Olympic Games. Using data from a representative online survey in 12 countries, they showed that respondent age was positively linked to the probability of casting a vote, and the levels of income and education had a weak impact on the turnout decision. 1 Lastly, some population characteristics have been included in turnout models, including size, density, stability and homogeneity (Geys, 2006). Theoretically, population size and density are expected to be negatively correlated with turnout. Decisions to vote are partly based on the expectation that an individual vote will be decisive, and this probability decreases as population increases. However, citizens in rural areas face higher social pressure to vote than those in urban areas because cities are considered more individualistic. Population stability and homogeneity are expected to increase turnout. A stable population increases both identification with the local area and knowledge of the issues at stake, increasing social pressure to vote. Social homogeneity is a requisite for the cohesion and solidarity of the community, which in turn strengthen social pressure to vote.

Theories and evidence on how citizens vote in referendums
There is a debate in the literature on the most important factors affecting outcomes in referendums. Some authors have suggested that a referendum represents a vote on the government of the day because it provides voters with an opportunity to vote against it. Franklin, Van der Eijk, and Marsh (1995) and Franklin (2002) argued that shifting attitudes toward domestic political actors, or the relative popularity or unpopularity of the government of the day, can sometimes provide a more plausible explanation of referendum outcomes than views on the issue at stake itself. They suggested that the rejection of the referendums in France and Denmark in 1992 to ratify the Maastricht Treaty was not evidence of a negative attitude towards Europe, but reflected the unpopularity of the ruling parties in both countries. A referendum at about the same time in Ireland, where the government was more popular, achieved a large majority, as did the referendum a year later in Denmark, after a more popular government had taken office. A similar explanation is particularly relevant at a time of general discontent and amid the rise of populism, when referendums can be seen as an opportunity to cast a vote against the 'elite' (Topaloff, 2017). A referendum may, therefore, represent a 'punishment trap' because it gives voters the opportunity to punish the government for poor performance (Schneider & Weitsman, 1996).
Lastly, the outcome of a referendum may be influenced by the evaluation of voters about the leaders of parties that promoted the referendum (Clarke, Sanders, Stewart, & Whiteley, 2009;Stevens & Banducci, 2013). Empirical evidence shows that when the decision to hold a referendum is taken by a governing political party, the results could be different from those expected. Generally, a governing party opts to announce a referendum in the expectation that it will win, or that its position on a particular issue will be sustained. This strategy is only sometimes successful. At other times, the referendum issues are not well understood, or voters may prefer to vote against the government (Pasquino & Valbruzzi, 2017).
Some authors have claimed that political parties' recommendations may be important cues that help voters to decide on referendum issues (Hobolt, 2009;Jenssen & Listhaug, 1999). Leduc (2002) argued that the relative weight of different factors can vary substantially with the context in which a referendum takes place. A referendum in which political parties take well-known and predictably opposing positions ought to hold the least potential for changes in opinion relative to the results of a general election. In this situation, voters who rely on partisan cues or party endorsement are influenced by the position of their preferred parties. Van der Brug, van der Meer, and Van der Pas (2018) noted that if a voter is a strong supporter of a party and the leader of that party suggests voting 'yes', then the voter is likely to decide to follow this advice.
A different line of thinking has emphasized the importance of the specific issues at stake in the referendum. This suggests that voters base their vote on attitudes to the referendum topic. An example is European Union referendums, where voters may decide based on their opinion of the European Union (Schuck & de Vreese, 2008).
Recently, some studies have suggested that economic factors can influence voting behaviour. Considering Olympic bid referendums, for example, Streicher, Schmidt, Schreyer, and Torgler (2016) provided evidence that the potential economic impact, in terms of benefits for the population deriving from revenue, expenditure and the subsequent use of the infrastructure created for the Olympic Games, plays a role when people decide to support a bid to host the Olympics. Coates and Wicker (2015) showed that communities with higher rates of unemployment had higher percentages supporting bids to host the Olympic Games.
Lastly, it has been argued that the referendum campaign plays a major role in determining voting behaviours. Referendums lack some of the long-term partisan and social features that shape general elections because of their infrequency (Laycock, 2013). Referendum outcomes, therefore, depend more extensively on short-term factors, such as campaigning (LeDuc & Pammett, 1995). The campaign provides voters with information to help them to form an opinion on referendum issues (Hobolt, 2009). It can also emphasize some of the consequences of the vote, which in turn can affect voters' decisions (LeDuc, 2003).

THE ITALIAN CONSTITUTIONAL REFERENDUMS
Three constitutional referendums were held in Italy between 2001 and 2016. Excluding the choice between monarchy and republic in 1946, these have been the only referendums held in Italy on constitutional issues. The three referendums were proposed by the central government and did not required any particular turnout to be considered valid.
The 2001 constitutional referendum was sponsored by a centre-left government. The opposition by the centre-right parties was not very strong and the electoral campaign was not very politicized. Turnout was only 34.0% and 64.2% voted 'yes'. Voters were asked whether they approved of amending the constitution to give more powers to the regions in several fields including agriculture, education, health and taxation. Theoretically, the decentralization of powers to regions favoured the wealthier regions of the north, but the percentage of the 'yes' vote in the poorer regions in the south was only slightly lower than in the centre-north, 59.5% versus 61.9%, respectively. Figure 1 shows that the 'no' vote was not geographically clustered in a given area of the country, while it characterized provinces in the north (i.e., Aosta 53.7%, Como 47.5%, Verona 46.9%), in the centre (Latina 46.4%, Isernia 45.4%, Frosinone 44.4%) as well as in the south (Messina 45.0%, Caserta 42.5%, Catania 42.3%).
The two referendums on constitutional reforms held in 2006 and 2016 were quite different. They were highly politicized with a stark contrast between government parties and opposition parties. Both proposals aimed to change the composition, power and size of the Italian parliament, and the allocation of powers between central government and local administrative authorities.  In 2006, citizens were asked whether they approved of amending 57 articles of the constitution (1) to give more power to the prime minister, who would be able to dissolve parliament and sack ministers; (2) to give more power to regions on education, healthcare and law and order; and (3) to separate the competences of the chamber of deputies and the senate: the first would have had the responsibility for foreign policy, defence and immigration, the second for federal law. In 2016, citizens were asked whether they wanted to amend the Italian constitution (1) to change the composition of powers of parliament; (2) to change the division of powers between state, regions and other administrative entities; (3) to reduce the number of members of parliament; and (4) to limit the costs of institutions and suppress the National Council for Economics and Labour.
There were similarities and differences between the two proposals ( Table 1). The 2006 and 2016 referendums aimed to eliminate the equal distribution of powers between the chambers and their very similar legislative competencies (symmetrical bicameralism), which were considered a source of inefficiency in approving laws. They also aimed to give more power to the executive. The main difference was that in 2006 the aim was to increase the powers of the regions, while in 2016 it was to decrease their powers, distributing powers between central and regional government to reduce the conflicts between the different levels of administrations.
Both proposals were rejected with a large majority: 61.3% in 2006 and 59.1% in 2016. The percentage of 'no' vote in the 2006 referendum was higher in the south (around 73%) than in the centre-north (61%). In all provinces of Calabria, for example, the percentage of 'no' was higher than 80%, and many other southern provinces scored just below this value (i.e., Nuoro, Agrigento and Matera).
Similarly, in the 2016 referendum the 'no' vote was higher in the south (68.6%) than in the rest of Italy (59.9%). The opposition to the referendum was particularly high in some southern regions, such as Sardinia (72.2%) and Sicily (71.6%). The only regions with a majority of 'yes' votes were Tuscany (52.5%), Emilia Romagna (50.4%) and Trentino (53.9%), all of which are in the centre-north. This territorial heterogeneity for 2006 and 2016 is shown in Figure 2, which shows the distribution of the 'no' vote across Italian provinces.   In both referendums, the campaign for the 'no' vote had the support of many constitutional experts (including the previous constitutional court chairman) who warned against the almost complete transformation of the constitution that the two proposals would have provided. A few constitutional experts supported the 'yes' vote, but they were in a minority. The interesting aspect was that the two referendums had similar objectives, but were supported by parties from different sides of the political spectrum: the 2006 vote was supported by centre-right parties, and the 2016 vote mainly by centre-left parties. The rejection of the two referendums also contrasts with the view that when a government has discretion on whether to call a referendum, doing so will usually strengthen the government (Butler & Ranney, 1994). Boyd (2010, p. 3) argued that: although representatives may misjudge what the results will be, they (referendums) are help primarily on questions officials believe they can win, that they feel need more legitimacy, to attack their opposition or to avoid issues that they find internally divisive.
Our analysis aimed to compare the three referendums to identify how differences in the socioeconomic and political context of geographic units affected the referendum outcomes. Survey data at individual level covering a sample of the same voters in the three referendums were not available, so the use of aggregate data was considered an appropriate approach for our research questions. Similar data are widely used in empirical studies on referendums, as documented by Geys (2006) in a meta-analysis on 83 studies with variables measured at district, municipality, state or country level.
Our econometric analysis was at provincial (NUTS-3) level, the third-level administrative division in Italy. It drew on data from the following sources: the historical archive of elections (Italian Ministry of Internal Affairs); the Italian National Institute of Statistics (ISTAT); the Institutional Quality Index (IQI) database developed by Nifo and Vecchione (2014); and the European Social Survey (ESS). The final sample consisted of 103 observations, corresponding to the administrative level of 'province'.
We estimated two econometric models. In the first, we considered the determinants of turnout, and the dependent variable was the ratio between the number of voters and the population within the age range eligible to vote (Turnout). In the second model, we analysed the outcome of the referendums, and the dependent variable was computed as the share of 'no' votes (Vote for no).
The explanatory variables were based on previous studies and computed in different years, depending upon the year in which the three referendums were held. The models included the rate of unemployment at two points in time: in the year before the referendums, to reflect the current labour market situation (Short-term unemployment), and as an average over the period from 10 to five years before the referendums, to measure long-term unemployment (Longterm unemployment).
The popularity of government is accounted for by using the following question from the ESS: How satisfied are you with the national government? We used data from the first, second and eighth rounds of the ESS to build a variable that ranged between 1 and 10, with high values associated with higher satisfaction with government (Government satisfaction). To control for other socioeconomic conditions, we included the IQI developed by Nifo and Vecchione (2014), a composite indicator that measures the quality of Italian institutions and assumes higher values for better institutions (IQI), 2 and the proportion of the adult population with a bachelor's degree or higher (Higher education).
The explanatory variables for political factors were 'No' parties and Regional council. 'No' parties was the proportion of votes for parties opposing the current government in the general election before the referendums (Table 1 shows the various parties). Regional council was a dummy indicating the regions governed by parties not forming the majority government or whose leader supported the 'no' vote. We also accounted for the general level of interest in politics by including diffusion of newspapers over population (Newspaper). The demographic variables measured the average age of the population (Age), the percentage of women in the province (Female) and population density (Population density). Lastly, the model considered the change in the foreign-born population in the three years before the referendums to capture possible discontent because of immigration (Foreign change), and dummy variables for north, centre and Mezzogiorno.
The empirical models were estimated using ordinary least squares (OLS) regression, which is a common strategy used to analyse turnout and the share of 'no' (or 'yes') votes when data are aggregated at the geographical level. In this context, the dependent variables are continuous, have a moderate variance and do not contain outliers, or observations that are substantially different from the majority. Table 2 shows the descriptive statistics for the variables used in the analysis.

Factors affecting turnout in the referendums
This section analyses the determinants of referendum turnout in the three constitutional referendums. The estimates are shown in Table 3. Each referendum was estimated with both short-(columns labelled 'a') and long-term unemployment (columns labelled 'b').
Our estimates suggest that turnout was affected mainly by unemployment and political factors, which were statistically significant at the usual levels in all regressions. The rate of unemployment, both in the short and long term, was negatively associated with turnout, meaning that citizens in provinces with high unemployment were less likely to vote. Similarly, citizens who supported opposition parties in the previous general election were less likely to vote in the referendums. Lastly, a higher diffusion of newspapers was linked to a higher referendum turnout, suggesting that a general interest in politics increases the probability that people will vote in referendums.
Other findings show that turnout was also higher in provinces with an older population and higher levels of institutional quality. Gender, education and population density were not important drivers of turnout, and those variables had low statistical significance. The R 2 of the model ranged from 0.80 to 0.92, suggesting that it has good explanatory power for referendum turnout.
Summing up, the results indicate that turnout in referendum is shaped only partially by the variables that explain turnout in elections and other types of referendums. Our analysis has highlighted other factors that matter in constitutional referendums, especially unemployment and affiliation to opposition parties.
Determinants of the outcome of the referendums This section analyses voting behaviour to identify the main factors that affected the outcome of the referendums. The estimates for the three referendums are shown in Table 4. As in the turnout model, each referendum was estimated with both short-(columns labelled 'a') and long-term unemployment (columns labelled 'b').  Notes: ***, ** and *Statistical significance at the 1%, 5% and 10% levels, respectively. Robust standard errors are shown in parentheses. A constant term is included but not shown.
As with turnout, the most important factors affecting voting outcome were unemployment, both in the short and long term, support for opposition parties, and satisfaction with government. The coefficients of the variables measuring unemployment were statistically significant in all regressions. In the 2006 and 2016 referendums, the coefficients were positive, meaning that the 'no' vote was stronger in provinces with higher rates of unemployment. For the 2001 referendum, however, the coefficients were negatively associated with the 'no' vote, meaning that citizens in provinces with higher unemployment were more supportive of the constitutional reform. However, it is worth noting that these coefficients are very small and the impact on the share of the 'no' vote was very limited. In all referendums, however, unemployment was crucial for the final outcome, as the constitutional proposal was approved in 2001, but rejected in 2006 and 2016.
The coefficients were slightly higher for short-term unemployment in 2006, but the difference from the coefficient for long-term unemployment was not statistically significant. In 2001 and 2016, however, the coefficients for long-term unemployment were higher and differed significantly from those for short-term unemployment, suggesting that long-term unemployment was more important in voters' decisions. The importance of unemployment is consistent with other analyses on Italian referendums. For the 2016 referendums, for example, David (2016) argued that the higher percentage of 'no' votes was stronger in southern Italy because of higher unemployment rates relative to other areas of the country.
Political factors measured by the variables 'No' parties and Regional council had a major role in shaping the outcome of the referendums. 'No' parties had positive and statistically significant coefficients, meaning that the proportion of opposition party supporters influenced the outcome because they tended to cast 'no' votes in each referendum. This suggests that the effect of party affiliation was quite strong, and that voting behaviour may not have been directly linked to the referendum question. In the 2016 referendum, the coefficients were smaller Notes: ***, **, and *Indicate statistical significance at the 1%, 5% and 10% levels. Robust standard errors in parentheses. Constant term included but not shown.
than in the other referendums, but still significant. This may be because of the effect of the other variable measuring political factors, that is, Regional council, which had a higher coefficient than in previous referendums. These results are in line with Istituto Cattaneo (2016) and Del Monte (2017), who suggested that the outcome of the 2016 referendum was linked to voters' adherence to particular parties. Pasquino and Valbruzzi (2017) emphasized that the campaign was particularly bitter and protracted, and suggested that the prime minister engaged in exaggerated political personalization, so that the vote became one on him rather than on the referendum questions. Regalia and Tronconi (2017) stressed the territorial differences in voting behaviour: in northern Italy, adherence to a party was more important, but unemployment was crucial in the south. The variable Government satisfaction was negatively correlated with the 'no' vote. In provinces where the government was unpopular, citizens voted against the referendum proposals. The estimated coefficients were statistically significant in all regressions, and their size increases in the last two referendums.
Citizens with higher levels of education had an important role in the outcome of the 2001 and 2006 referendums, but were not significant in 2016. In the 2001 referendum, educated people were more supportive of the constitutional reform, while in the 2006 referendum they opposed the proposal. As with unemployment, in both referendums education contributed to the final outcome. Gender and the share of foreign population had a weak impact on the referendum outcomes. The variable Female was only significant for the last referendum, showing that women were more likely to support the reform. The variable Foreign change was statistically significant only in 2006, when it was negative, suggesting that where foreign population increased, the share of 'no' vote was lower. For the other referendums, however, the coefficients were positive but not significant. This inconsistency casts doubt on the explanatory power of the variable, and suggests that anti-immigration sentiment had only a marginal role in the result of any of the three referendums.
The IQI negatively affected voting in the 2006 and 2016 referendums, and provinces with better institutions were more supportive of the constitutional reform. In the 2001 referendum, the variable was negative but not statistically significant. The R 2 of the model was relatively high (0.69-0.91), indicating that the covariates, taken together, explain a considerable amount of the variability in the 'no' vote.
An important aspect of the three referendums was the distribution of power between central government and the regions. The 2001 and 2006 referendums would have increased the level of decentralization, giving more power to local government. The opposite was true for the 2016 referendum, which reduced the powers of regions and increased centralization. Therefore, it could be expected that wealthier regions, which have fiscal revenues higher than expenditure, would support decentralization because it would allow them to use more of their fiscal revenues to improve the quality of public services. However, poor regions that obtain resources from central government to provide services for their constituency were expected to oppose decentralization of powers. In this case, the richest regions should have a negative impact on the share of 'no' votes in 2001 a 2006, and a positive impact in the 2016 referendum.
To test this idea, two variables were included in the model in Table 4: gross domestic product (GDP) per capita at provincial level (GDP per capita) and dummies for north, centre and Mezzogiorno. These variables were used alternatively, given the high correlation between North and GDP per capita (0.69). 3 The estimates are shown in Table 5.

CONCLUSIONS
This paper has analysed the determinants of turnout and voting behaviours in constitutional referendums, drawing on data from the three constitutional referendums held in Italy in 2001, 2006 and 2016. The first referendum was approved, while those in 2006 and 2016 were rejected. We found that turnout was affected particularly by unemployment and political identification with opposition parties. These two factors also had a considerable impact on the outcomes of the 2006 and 2016 referendums, in the sense that the 'no' vote was positively linked to unemployment and affiliation to opposition parties in particular regions. In addition, government satisfaction was negatively associated with the percentage of the 'no' vote, suggesting that the 2006 and 2016 referendums were probably rejected because the government of the day was unpopular, rather than because the majority of citizens were critical of the constitutional proposals.
The 2006 and 2016 referendums are important examples of voting in time of discontent because of the economic downturn and high unemployment rates that characterized Italy during these periods. The contribution of unemployment to opposition to the proposals in these referendums suggests that discontent with national government, especially because of economic and labour market conditions, played a major role in shaping voting behaviours. This is consistent with the descriptive statistics in Table 2 on government satisfaction, which on average decreased from 3.91 in 2001 to 3.11 in 2016.
Party cues also played an important role in determining referendum voting. We found that citizens supporting the parties opposing the current government in the general election before the referendums tended to vote against the proposal. This result was further corroborated by Table 5. Further results on the determinants of the share of 'no' votes. considering the party that led local government at the time of the referendum: the share of 'no' vote was higher in regions led by parties that opposed the national government. The 2001 referendum, however, was approved because of weak opposition by non-government parties. Turnout was low and government parties obtained the majority of 'yes' votes. These results suggest that if a government does not have popular support, a constitutional referendum may be rejected more for political reasons than on the merit of the questions.
The 2001 and 2006 referendums were in favour of decentralization of power to the regions, while that of 2016 increased centralization of power in central government. The richest regions should, therefore, have supported the first two referendums and opposed the 2016 referendum. The empirical results, however, showed that centralization or decentralization of powers had little impact on the referendum outcome.
Overall, our findings suggest that the merit of the constitutional reform can play a minor role in the outcome of a referendum. One important implication of this study is that the 2016 and 2006 Italian referendum probably did not reflect citizens' views on constitutional reform, an issue affecting the quality of institutions and economic development of the country. It is, therefore, important to establish legal and procedural conditions to ensure that voting behaviour better reflects views on referendum questions. The results also cast doubt on the suitability of using referendums as a way to judge attitude to constitutional changes. This point has also been debated in the literature. There is a question about whether voters have adequate knowledge and ability to make informed decisions about complicated issues such as electoral reform or division of power between different administrative levels. It may also be inappropriate for the rules of constitutional referendums to be decided by politicians, who may have a strong self-interest in the issues concerned.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.

NOTES
1 These variables should be interpreted with caution if aggregate data are used because of the problem of 'ecological fallacy' (e.g., Kramer, 1983). This describes the situation in which inferences about the nature of individuals are deduced from inferences about the group to which those individuals belong, rather than from individual data.