The political context around the 2016 Italian constitutional referendum

The use of a referendum for constitutional amendments, either as an absolute requirement or as an optional alternative, is common to several countries (Lijphart, Reference Lijphart2012). Referendums are usually prescribed in addition to legislative approval by simple or qualified majority, making constitutional amendments harder to adopt (Gallagher, Reference Gallagher1995).

Italy is an interesting case as the use of referendum represents an ‘alternative method’ to approve a constitutional reform without the qualified majority in Parliament. In fact, a constitutional reform approved by legislators with the mere simple majority can only be adopted if it obtains the support of the majority of Italian voters in a popular referendum without quorum (see art. 138 Cost.). Thus, former PM Renzi had two alternative paths: either reaching the qualified majority in Parliament with the key support of the center-right party FI or approving the reform with the simple majority, thus betting on his ability to mobilize the electorate in the confirmatory referendum.

Of course, each path implies different strategies. Reaching the qualified majority in both Chambers requires reform proponents to find a large parliamentary support. We call this strategy ‘consensual strategy.’ Reaching the simple majority plus winning a referendum asks instead reform proponents to bet on the support of the majority of voters. We call this strategy ‘popular vote strategy.’

But what makes one of the two strategies more likely to succeed? If we look back to the history of constitutional reforms in Italy, consensual strategies seem to have been more successful. In fact, major attempts to reform the Italian constitution (i.e. the Bozzi Bicameral Committee in 1983–1985, the De Mita-Iotti Bicameral Committee in 1992–1994, and the D’Alema Bicameral Committee in 1997) got sunk in Parliament under the conflicts between government and opposition parties (Ferrera and Gualmini, Reference Ferrera and Gualmini2004). In its attempt to reform the constitution in 2006, the center-right coalition led by FI leader Berlusconi decided to rely on the strength of its parliamentary majority only (Bull and Newell, Reference Bull and Newell2009). Thus, the subsequent confirmatory referendum saw the unanimous mobilization of center-left parties and the constitutional reform was rejected in popular vote. On the contrary, opposition parties were much more divided in the 2001 constitutional reform, which was proposed by the center-left coalition and was finally adopted through a popular referendum (Keating and Wilson, Reference Keating and Wilson2010).Footnote ³

According to Tsebelis (Reference Tsebelis2002, Reference Tsebelis2017), it is impossible to establish a priori which strategy (i.e. consensual strategy vs. popular vote strategy) is more likely to succeed in changing the status quo (i.e. in reforming the constitution) because the constitutional core (i.e. the set of provisions that can never be changed) depends on the actual preferences of the actors involved. The key difference between the two strategies is that the popular vote introduces an additional veto player, that is to say the population, with its own preferences.

With regards to the 2016 constitutional referendum, Tsebelis claimed that the popular vote strategy chosen by Renzi was more likely to succeed than the consensual strategy because Italian voters were more prone to reduce the power of the Senate (and the numbers of MPs) rather than the Senators’ (Tsebelis, Reference Tsebelis2017: 8).

However, Tsebelis’s argument requires citizens to manifest clear preferences on the content of the constitutional reform and to take part in a referendum to express their sincere opinions (systematic mode). Vice versa, referendums are often described in the literature as ‘second order elections,’ where citizens simply take the chance to express their judgments on contextual elements (e.g. Reif and Schmitt, Reference Reif and Schmitt1980), such as the economic situation, the performance of the incumbent government and the political élites (heuristic mode).

The election of Mattarella as President of the Republic in January 2015 without the agreement of FI is a turning point in the history of the 2016 constitutional referendum. Essentially, Renzi broke the political agreement he had signed 1 year before with FI leader Berlusconi,Footnote ⁴ thus abruptly dismissing the consensual strategy in favor of the popular vote strategy. More interestingly, by breaking up with Berlusconi, Renzi decided to give up FI’s support in the subsequent confirmatory referendum: he chose to undertake the referendum on his own, with the support of his cabinet and his party (i.e. the PD) only.

The infringement of this pact in order to elect President Mattarella has been labeled by the large majority of Italian media as ‘Renzi’s political masterpiece’ (e.g. La Stampa, Il Corriere della Sera, Il Foglio, and Huffington Post). Almost 3 years later, after losing the referendum and, consequently, the premiership, a simple question comes to mind: was the election of President Mattarella worth the trouble? In retrospect, it is rather easy to provide a negative answer. However, in the winter of 2015 Renzi chose to place a bet on his popularity, the positive evaluation of his cabinet, and the strength of his party. Was this choice what Homer would have called an act of hybris? To assess the degree of Renzi’s misjudgment, we need to understand which cognitive mechanisms drove the vote choices of citizens at the 2016 constitutional referendum.

Systematic vs. heuristic reasoning in referendums: literature review and hypotheses

On 4 December 2016, Italian citizens casted their ballots on a far-reaching constitutional reform in a referendum vote without quorum. Ideally, their voting behavior should have been based on an informed judgement on the policy content of the issue at stake (e.g. Lupia and Matsusaka, Reference Lupia and Matsusaka2004; Kriesi, Reference Kriesi2005; Hobolt, Reference Hobolt2007). That is to say, they were meant to be collecting accurate information, critically evaluate the pros and cons and, finally, come to an informed vote choice.

According to recent studies, this systematic mode of information processing is not as uncommon as one might think. In fact, voters are increasingly aware of the main policy arguments of the campaign (Colombo, Reference Colombo2016) and they often rely on policy arguments to make their vote choices, when and if these argument are available (Boudreau and MacKenzie, Reference Boudreau and MacKenzie2014; Colombo and Kriesi, Reference Colombo and Kriesi2016). Thus, we formulate our first hypothesis:

However, since the average level of interest in politics and political knowledge tend to be rather low, individuals are not naturally inclined to invest their time in collecting ‘encyclopedic’ political information (Lupia, Reference Lupia1994). The so-called dual process models of reasoning state that individuals follow the least-effort principle, that is they aim at reaching a vote choice that reflects their policy preferences as best as possible by minimizing informational and cognitive efforts (e.g. Kahneman, Reference Kahneman2011). Besides, the 2016 constitutional reform included a package of such complex funding provisions that even the most informed and motivated individuals could not fully grasp its overall consequences. In such circumstances, individuals may resort to peripheral considerations and cognitive shortcuts as substitutes for policy considerations to reach a reasonable vote choice (e.g. Lupia and McCubbins, Reference Lupia and McCubbins1998: 2). This decision-making strategy is labeled ‘heuristic mode of information processing’ (Lupia et al., Reference Lupia, McCubbins and Popkin2000; Kriesi, Reference Kriesi2005).

Since Fiorina (Reference Fiorina1981), scholars have underlined the key role played by the economic situation in shaping voting behavior. In a nutshell, if the economy is doing well, citizens are more likely to vote for governing parties, while if the economy is going poorly, citizens tend to punish incumbents (e.g. Lewis-Beck and Stegmaier, Reference Lewis-Beck and Stegmaier2007). This mechanism also works in the stance of referendums: citizens may punish (reward) governing parties for their performance in the economic sphere by rejecting (accepting) their reform proposal. For example, the economic conditions experienced by British people were key drivers of the vote in the 2016 Brexit referendums (Hobolt, Reference Hobolt2016). We argue that this mechanism may have played a role in our case study as well. To emphasize this simplification of voters’ reasoning, we could assert that if the government has managed the economy badly (well), the constitutional reform which was formulated and supported by the same government will also be likely to be badly (well) written. Thus, we formulate our second hypothesis:

Similarly, voters’ behavior in referendums may be affected by their assessments of the overall performance of the cabinet: ceteris paribus, a reform supported by a popular cabinet is more likely to be approved (Gallagher, Reference Gallagher2015: 282). Of course, citizens’ degree of approval towards the cabinet may be driven by both policy and valence issues (Stokes, Reference Stokes1963). For example, it is claimed that President Mitterand’s unpopularity nearly caused the defeat of the Maastricht Treaty referendum in France in 1992, while the personal popularity of the PM Gonzalez ensured that the Spanish cast their vote to stay in the NATO in 1986 (LeDuc, Reference LeDuc2003). Accordingly, we formulate our third hypothesis:

This mechanism is likely to have played a leading role in our case study. PM Renzi linked the fate of his government to the approval of the constitutional reform from the very beginning, thus transforming the referendum into a verdict on his personality. A survey experiment hints that this strategy may have been counterproductive: indeed, voters’ support for the reform weakens when it is explicitly linked to the government (Colombo et al., Reference Colombo, de Angelis and Morisi2016).

Moreover, when confronted with complex policy decisions, citizens may rely on signals from trusted political parties, partisan attachments and feelings (e.g. Slothuus and de Vreese, Reference Slothuus and de Vreese2010). Indeed, citizens are in lack of time, interest and cognitive capacity to evaluate the possible consequences of accepting or rejecting a referendum proposal (Borges and Clarke, Reference Borges and Clarke2008: 437; Clarke et al., Reference Clarke, Kornberg and Stewart2004: 346). Thus, following party recommendations can be an important heuristic helping citizens to determine what information is trustworthy (Hobolt, Reference Hobolt2009: 35–7; Lupia and McCubbins, Reference Lupia and McCubbins1998). However, empirical evidences on partisan effects are not clear-cut (e.g. de Vreese, Reference de Vreese2006; Neijens and van Praag, Reference Neijens and van Praag2006).

Since the majority of the PD frontbenchers campaigned for the Yes-side (together with the small NCD party), while all the other parties (i.e. FI, M5S, and Northern League – LN) stood by the No-side, we expect that:

The hypotheses listed so far will guide us in the learning of whether Italian voters in the 2016 constitutional referendum were driven mainly by substantive information about the content of the constitutional reform (Hypothesis 1) or whether peripheral dynamics (i.e. retrospective economic evaluation, the popularity of the Renzi cabinet and partisan attachments) prevailed (Hypotheses 2–4). Identifying the main drivers of voters’ choices in direct-democratic settings is crucial as they affect the quality of democratic decisions indeed.

However, this paper takes a step forward in order to understand whether peripheral dynamics also have a conditioning effect on the impact of substantive policy considerations on voting behavior. Was a negative evaluation of the economy (heuristic mode) able to nullify the effect of a positive evaluation of the complex provisions embedded in the constitutional reform (systematic mode)? Else, did a positive evaluation of the Renzi cabinet contain the effect of a skeptical content evaluation (systematic mode)? Finally, did Italian voters conform too readily to the views of their trusted parties (heuristic mode) even at the cost of neglecting their opinions on the issue at stake (systematic mode)? The empirical analysis described in the next section will help us test the hypotheses and answer these additional questions.

Research design

Model specification

We test the listed hypotheses through a nested logit model. It is a probabilistic choice model formulated by the Nobel Prize McFadden (Reference McFadden1978) to operationalize choice as a sequential process. Decision makers are assumed to be utility maximizers that evaluate the alternatives in consecutive steps. Therefore, the model is especially suitable for modeling decision-making processes in which, from a certain time on, only a subset of alternatives is still available.Footnote ⁸

Probabilistic choice models maintain that utility contains two systematic components and one random component. Systematic components include the attributes of the decision maker and those of the alternatives. The random component consists of unobserved attributes of the alternative, unobserved attributes of the decision maker, measurement error, and proxies or instrumental variables (Manski, Reference Manski1977). Due to the random component, utility maximization and consequently choice are probabilistic.

Usually, probabilistic choice models (e.g. Lucean-type models; Luce, Reference Luce1959) make rather stringent assumptions about the random component of the utility. In fact, they assume that the property of independence of irrelevant alternatives (IIA) is fulfilled, meaning that the random component of the utility should be identically distributed, independent, and homoscedastic. Instead, the nested logit model relaxes this assumption.

In a nested logit model, the researcher establishes the nesting structure, namely how many groups (nests) are included and which alternatives belong to which nest. Each nest should feature a subset of more similar options.

Our decision tree groups the alternatives in two nests (abstention) and (Yes vote, No vote).Footnote ⁹ We maintain that the property of belonging to the same nest is due to temporal considerations: the will to cast one’s vote precedes the decision of voting Yes or No. Indeed, since the constitutional referendum did not request the quorum, abstention cannot be seen as a strategic choice. Only if respondents participate voting Yes or No becomes relevant.

There are three alternatives in the structure of Figure 1: 1=abstention; 2=Yes vote, 3=No vote. Let Y _q =1, 2, 3 be the observed choice for decision maker q.

Figure 1

Decision tree.

Under a utility maximization assumption, Y _q =i if U _qi >U _qj ∀j≠i. For instance, Y _q =2 if U _{q 2} >U _{q 3} and U _{q 2} >U _{q 1} . We assume that for decision maker q, the utility of alternative i is conceptualized as random, consisting of a systematic component (V _{q i} ) and a random component (ε _q ):

$${\rm U}_{{qi}} {\equals}{\rm V}_{{qi}} {\plus}{\varepsilon}_{{qi}} $$

V _qi is a function of explanatory variables and parameters to be estimated and ε _qi a random error term that captures the unobserved attributes of the utility. If explanatory variables involve characteristics of the individual making the choice, the variable assumes the form of the individual characteristics interacted with alternative-specific dummy variables (Born, Reference Born1990).

The nested logit model assumes that the errors of similar alternatives are correlated, while the errors of dissimilar alternatives are constrained at 0. Since alternatives 2 and 3 are similar, and indeed they are grouped in the same nest, then ε _{q 2} and ε _{q 3} are correlated. Instead, alternatives 1 and 2 are dissimilar, therefore ε _{q 1} and ε _{q 2} are uncorrelated.

In addition, error terms are assumed to follow a Gumbel’s Type B extreme value distribution (Born, Reference Born1990). Resulting choice probabilities are of the following type:

$$\pi _1\,{\equals}\,{{{\rm exp}\left( {V_{1} } \right)} \over {{\rm exp}\left( {V_{1} } \right){\plus}\left( {\sum_{j\,{\equals}\,2}^{3} ({\rm exp}(V_{j} \,/\,\lambda _2)} \right)^{{\lambda _2}} }}$$

and

$$\pi _2\,/\,\pi _3 \,{\equals}\,{{{\rm exp}(V_{2} \,/\,\lambda _2)} \over {{\rm exp}(V_{3} \,/\,\lambda _2)}}$$

The π ₁ equation represents the marginal probability of abstaining rather than voting for any option.

The π ₂ /π ₃ equation represents individual conditional probability of voting Yes, given that he/she chose to participate to the referendum. The relative choice probabilities between alternatives 2 and 3 are entirely given by the utilities of those alternatives. No other alternative influences the likelihood of voting Yes or No. This would be true even adding other alternatives to the nest. In other words, for any two alternatives in the same nest, the ratio of probabilities is independent from the existence of other alternatives in the same nest: the IIA assumption holds within each nest. Conversely, for any two alternatives in different nests, the IIA assumption does not hold.

λ is a measure of the degree of independence in unobserved utility among the alternatives in a choice subset and it is called ‘dissimilarity parameter’. In π ₂ /π ₃ equation, λ ₂ is the dissimilarity parameter for the choices in the second nest (Yes vote, No vote). A high λ means greater independence and less correlation. A value of λ of 1 means complete independence in the nest. If λ is equal to 1 for all nests, the nested logit would collapse into the standard logit model. In our case study, a value of λ of 1 for all nests would mean that the options Yes voting and No voting are interchangeable with the choice of abstaining. Conversely, a value of λ of 0 would entail suggest the presence of separate decisions on whether and what to vote for. Testing the distance of λ from 0 determines the appropriateness of the choice of a nested logit model.