The causal effects of primary elections on politicians ideologies

The causal effects of primary elections on politicians ideologies Andrea Cintolesi European University Institute In this work I identify the causal effects of the introduction of primary elections on the elected politicians ideologies using two different datasets. Since 1914 state level laws impose primary elections for all US Congress elections. Senators from New York State are an exception, as they have been chosen by primaries only since 1968. I set up a diff-in-diff analysis exploiting the roll call behaviour records as a measure of ideology. Moreover, I run a similar analysis on the atypical experience of primaries for the Italian municipal elections. After several years of contradicting theoretical investigations, results show that in both datasets, primaries have a null causal effect on politicians average ideologies, but significantly decrease the ideological volatility of the elected politicians. In both my samples the impact on the volatility is very large and accounts for an overall reduction of more than the 50%. In the end I discuss the possible mechanisms that can trigger such effects. Economics Department - Mail: andrea.cintolesi@eui.eu I am grateful to my advisors Andrea Mattozzi and Andrea Ichino, to the Candidate and Leader Selection standing Group for the data and the support and to all the participants of the micro-econometrics working group at EUI for the helpful comments. Any mistake is exclusively of my own responsibility. 1

1 Introduction In the last 20 years many scholars tried to think on the effect of primary elections on politicians ideology, delivering a variety of different and contradictory conclusions. Indeed, there are good reasons to think that primary elections deliver more extreme candidates, but there are also valid reasons to argue the opposite. Some researchers tried to show that since primaries pools of voters are more extreme, candidates are required to be extreme enough to get the nomination. On this line, many political scientists claimed that candidates respond to primary election proposing more extreme platforms, eventually enhancing the political polarisation. However, other scholars disagreed and concluded that primaries favour more moderate platforms. A first reason that they propose is that usually campaigns with primaries are longer, and this has been shown to be correlated with a convergence of the candidates toward moderate positions. A second and more convincing reason is that primary elections can solve conflicts within the parties, preventing more candidates from the same party to run in the same general election. Following this stream of theories, primaries would clean up the competition on the proper political side, quash any incentive to be extreme and leave candidates free to converge on moderate stances to have more chance to get elected. To the best of my knowledge, beside many theories and conjectures there are no clear evidence of any causal effect of primary elections on politicians ideologies. On one hand, given the richness of types of primaries and of countries that adopt them, a causal investigation on this question would be particularly devalued by the absence of external validity: it would be hard to define whether the results are specific to a particular country or type of primary, and it would be even more difficult to understand which theories apply to different cases. On the other hand, an investigation that provides results that are robust across country and level of elections would deliver an appreciable contribution for two reasons. Firstly, it may highlight the true implications of the introduction of primaries on a very important aspect as the politicians ideologies is in a moment where many countries are debating about introducing primaries. Secondly, it may point out which theories on primaries are empirically supported and which are the force triggered by the introduction of primaries on the light of very robust and not country specific results. Henceforth, to understand which is the true effect of primary elections on politicians ideology a robust quasi experimental investigation is strongly needed. In this paper I consider US Senators from New York in the 70s and Italian mayoral election in the late 2000s. I assess in a causal framework which is the effect of primary elections on politicians ideology, and I show that 2

the results are the same despite the significant differences between the two datasets, suggesting that they have a robust external validity. Before moving ahead, I need to distinguish my work from a similar attempt by Ansolabehere, Hansen, Hirano and Snyder (2010), from were I borrow some details better explained in the empirical section on US senators. They study the introduction of primaries in seven states, looking for ideological differences of US senators and representatives in the four congresses before and after the primaries were introduced. However, I disagree with their identification because primaries effects can take a while before showing up. Indeed, in their sample most of the elected candidates after the introduction of primaries are actually the incumbents that were chosen for the precedent election without primaries and that kept being elected. For example, considering their sample for senators, I find that in the first congress after the introduction of primaries only three out of fourteen candidates are not the incumbents; they become five out of fourteen in the second congress and eight out of fourteen in the third. All the others are politicians that were elected before primary were introduced and kept being elected in the following congresses, most likely taking some advantage from their incumbency. Unsurprisingly, they find a null effect of primaries on politicians ideology. I instead propose a causal design that considers a very long time range, focusing on the state of New York and with a dependent variable that accounts for time varying preferences of the constituencies, and I indeed find that primary elections halve the ideological volatility of the elected politicians, even though the effect on the average ideology is null. I validate my results with a number of placebo estimations, and I show that data on Italian municipal primaries provide the same results of comparable magnitudes. This work relates to the stream of literature that tries to measure the effect of primaries on candidates electoral strength (Carey and Polga-Hecimovich (2006), Ramiro (2013), Hirano, Snyder (2014)) and to the group of studies that investigates empirically the different types of primary elections (Gerber and Morton (2008), Abramowitz (2008)); it is also related to the numerous theoretical investigations of politicians behaviour under primaries (Hummel (2010), Hirano, Snyder and Tinga (2008),Snyder, Hirano, Lenz and Pinkovskiy (2015), Adams and Merrill (2008)). Finally, this work is build on the literature on primary elections in Italian municipalities (among the numerous work: De Luca (2014), Seddone and Venturino (2013), Bernardi and Seddone (2013),Venturino (2009), Valbruzzi (2005)). The remainder of the paper is organised as following: in Section 2 I develop a causal inference with data on senators from New York. In Section 3 I exploit data on Italian municipalities. In Section 4 I discuss the possible channels through which the effect I find can be transmitted and I finally draw some conclusions. 3

2 US Senators from New York US senators are elected statewide. Every state elects two senators that sit in three congresses holding the office for six years. House of Representatives members are instead elected districtwide 1 every two years. States with an higher density have more districts and therefore have more representatives. Neither senators nor representatives are term limited. In 1913 the Seventeenth Amendment to the United States Constitution established the popular election of United States Senators by the citizens of the state; and since 1914, direct primaries are compulsory in almost any state both for representatives and for senators. The introduction of direct primary in the United States was driven by a series of state level laws that left out some difference between states. Most of these differences regard minor details or very short period of time and cannot be exploited to investigate the effect of primary elections on politicians ideologies. However, what happened in the state of New York gives an opportunity for a first and very clear identification strategy. House of Representatives members from New York had to pass through primary elections since the 1914, as a consequence of a State level law. The same law required senators to do primaries. However, in 1920 primaries for senators were halted. This anomaly lasted almost fifty years, when primaries were reintroduced also for senators from New York in 1968. Why such a difference between the members of the House and the senators in New York, and why for such a long time? Primaries supporters had large victory in most of the states in the 1914, right after the Seventeenth Amendment became law, but the euphoria for primary elections vanished very soon and counter reform movement arose. New York State was special because of the strength of these movements, and by 1921 opposition in New York to the use of direct primaries for state wide offices was so great that their use was partially halted. The counter reform made a distinction considering the level of the election: primaries for state wide offices, including senators, were retracted, whereas district level elections as the ones for representatives were left untouched. In most of the other states, the opposition to primaries didn t manage to have such a victory, and primaries remained the way candidates for senators and for representatives were chosen. Moreover, it is not surprising that this anomaly lasted so long. The reason is that it was simply not worthwhile expending political capital to try to reenact primaries for senators: there was simply insufficient support to correct it - given the potential opposition within the 1 New York State always had a large number of districts and therefore a large number of elected representatives. Districts are drawn on the base of population density and therefore their number varies across time. In 1950 there were 45 districts in New York State, and through decennial census they have been decreasing to the actual level of 27. 4

most powerful county organisations in New York State. That situation would not change until the power of the latter was much reduced. 2 Along the years, both for voters and politicians, this difference became a minor issue. Finally, in 1968, New York state approved a law that reintroduced primary election also for state level election, and so also for senators. New York legislators opted for closed primary, the most conservative type of direct primaries, where only people registered to the party have the right to vote. The particular features of the introduction of primary election in New York State provide the possibility to compare the behaviours of the senators from New York respect to senators from a control state. I choose New Jersey as the control state of the diff-in-diff analysis explained below because it is the most reasonable choice in terms of geographical proximity, however the results are indifferently robust to any choice. 2.1 Data and measure of Ideology I got data on senators and house of representatives members from New York together with a measure of their ideologies from the well known and widely used DW-NOMINATE database. This measure of ideology derives from the multidimensional scale method developed by Keith T. Poole and Howard Rosenthal in the early 1980s. To propose a measure that allows to compare politician ideologies across congresses they make use of voting behaviour of any politicians in US congresses, assuming that politicians are sincere and rational. 3 The DW-Nominate measure of ideology for senators from New York ranges between -1.284 and 1.469, where -1.284 is the most liberal senator and 1.469 is the most conservative one. In figure 1 I plot the average ideology for Republicans and Democrats members of the House in New York State since the 1900. The two dashed lines sign the introduction of primary elections for representatives (1914) and the reintroduction of primaries for senators (1968). Two ideological streams can be clearly identified: the upper stream is the ideological pattern of Republican representatives, the lower stream is the same for Democrats. In figure 2 I add the senators ideology (in this case I don t consider the average, as the empirical specification will consider the difference between the single senators ideology and the average representative ideology). As it is easy to note, after 1968 senators and representatives ideologies are very close, while before when candidates for senators were not chosen by primaries there are considerable ideological differences between them. 2 Alan Ware, The American Direct Primary: Party Institutionalisation and Transformation in the North, Cambridge University Press; 1 edition (April 30, 2009) 3 For further reference to that: http://voteview.com/dwnomin.html. 5

ideology -1 -.5 0.5 1 1887 1914 1969 2015 year av. democratic representatives av. republican representatives Figure 1: Average ideology of House of Representatives members in every congress ideology -1 -.5 0.5 1 1887 1914 1969 2015 year democratic senators av.democratic representatives republican senators av.republican representatives Figure 2: Average ideology of House of Representatives and Senators members in every congress 6

To generate a quasi-experiment from the events that I described above, the DW measure doesn t work as it is. My aim is to compare the difference between senators ideologies and representatives ideologies in New York and in New Jersey before and after the 1968. In particular I want to compare the differences between senators ideologies and the average house of representatives ideology within the same party and congress before and after the reintroduction of primaries for senators in New York. In this way, representatives average ideologies act as a reference point for the party ideology in every congress from where I can establish whether senators ideologies are more extreme or moderate. Moreover representatives work as control for time varying party and state specific preferences of the constituencies. To implement my identification strategy, I firstly need to construct a single variable to consistently compare senators and average representatives ideological distances. Here I partially follow Ansolabehere, Hansen, Hirano, Snyder (2010): note that for Republican senators, being more moderate respect to their colleague in the House means to have a lower DWN; while for Democratic senators it is exactly the opposite. Henceforth,if S i,p,t is the DWN for senator i of party p at time t, and R p,t is the average DWN for representatives of party p at time t, my measure of ideological distance between senator i and the average representatives of party p in congress t is: S i,p,t R p,t if p=republican distance i,p,t (1) (S i,p,t R p,t ) if p=democrats A negative distance i,p,t means that senator i of party p in congress t, was more moderate than the average of the representatives of his party in the same congress. A positive distance instead means that the senator was more extreme. In figure 4 I plot the time series of distance for senators from New York. The dashed lines indicate the introduction of primaries for representatives and for senators as before. Moreover, I build a measure for the volatility of distance, that it is simply its square value 4. volatility i,p,t (S i,p,t R p,t ) 2 (2) The interpretation of volatility is straightforward: it measures the difference between every senator and the representatives average ideology within party and congresses regardless to the sign of distance. I plot the time series of volatility in Figure 5. In Figures 6 and 7 I plot the time series for distance and volatility for New Jersey. Finally, in Figure 3 I plot the histogram 4 Indifferently I could consider the absolute value 7

of the ideologies of senators and representatives from New York elected before and after the introduction of primaries. -.4 -.2 0.2.4 -.4 -.2 0.2 Figure 3: histogram of distance for senators elected with primaries(left) and without (right) distance -1 -.5 0.5 1 1887 1914 1969 2015 years democratic senators republican senators Figure 4: Distance for senators from New York 8

volatility 0.1.2.3.4.5 1887 1914 1969 2015 year democratic senators republican senators Figure 5: V olatility for senators from New York distance -1 -.5 0.5 1 1887 1914 1969 2015 year democrats senators republican senators Figure 6: Distance for senators from New Jersey 9

volatility 0.1.2.3.4.5 1887 1914 1969 2015 year democrats senators republican senators Figure 7: V olatility for senators from New Jersey 10

2.2 Empirical Strategy I set up a diff-in-diff strategy to compare the effect of primaries on the variables distance and volatility in New York respect to New Jersey. The years after the 1968 are the treatment period, and in my case the control group is always treated. Following Bertrand, Duflo, Mullainathan (2003) I collapse the time dimension of the data in a dummy that identifies whether the observation refers to a year before or after the treatment. This is necessary to preserve the standard errors and to avoid over rejection of the null hypothesis of null coefficients. Call After t the time dummy just described, and NewY ork s a dummy that identifies whether the observation refers to New York State or is otherwise about New Jersey. The specifications I use for senator i state s, party p and congress t are: distance i,p,t,s = β 0 + β 1 NewY ork s + β 2 After t + β 3 After t NewY ork s + β 4 X i,p,t,s + ɛ i,p,t,s (3) volatility i,p,t,s = β 0 + β 1 NewY ork s + β 2 After t + β 3 After t NewY ork s + β 4 X i,p,t,s + ɛ i,p,t,s (4) My causal coefficient of interest is β 3. Note that even if there isn t any observation referred to house of representatives members. However, the variables distance and volatility contain the representatives average ideology for state s, party p and congress t that behave both as a reference point and as a control for time varying party and state specific preferences. The set of controls contains individual characteristics of the senator, namely the age, the number of years as a senators and the party affiliation. A very important control for specification (4) is the direction of the deviation from the mean: as the distribution of the variable distance is slightly asymmetric, it turns out to be crucial to control to which side the absolute deviation leans. 5 2.3 Results I run the regressions making use of three different time ranges to show that results are not driven by the choice of the sample. I pick 24, 28 and 32 years around the 1968. I don t shrink the interval more because senators are not term limited and usually are elected many times. As said in the introduction, senators chosen without primaries right before the introduction of primaries, are still selected as a candidate and reelected in the years after the reform, most likely just because they are the incumbent. Hence to be able to see the effects of primary 5 If a distribution is skewed, a deviation from the mean (level of volatility) can imply higher or lower volatility depending on which side respect to the mean the ideology is located 11

Table 1: Results for New York Senators (32 years) (28 years) (24 years) (32 years) (28 years) (24 years) distance distance distance volatility volatility volatility Af ter*p rimaries -0.006 0.008 0.004-0.075* -0.100** -0.111** (0.051) (0.057) (0.063) (0.040) (0.044) (0.049) Af ter 0.048 0.042 0.037-0.021-0.015-0.011 (0.037) (0.041) (0.044) (0.027) (0.031) (0.035) NewY ork -0.092** -0.098** -0.082* 0.066*** 0.086*** 0.106*** (0.038) (0.043) (0.046) (0.023) (0.025) (0.026) Moderate 0.069*** 0.080*** 0.104*** (0.020) (0.022) (0.022) Constant -0.071*** -0.075*** -0.093*** 0.082*** 0.071*** 0.045** (0.025) (0.028) (0.028) (0.019) (0.021) (0.021) Observations 118 101 85 118 101 85 R 2 0.1295 0.1227 0.0888 0.2191 0.2621 0.3104 p 0.002 0.010 0.088 0.000 0.000 0.000 Regression for New York and New Jersey senators of specification (3) and (4) without controls.after is a time dummy to identify observation after the introduction of primaries, NewY ork is a dummy that identifies senators from New York, Moderate identifies whether the deviation from the average representatives ideology leans on the moderate or on the extreme side. In parenthesis I report robust standard errors. election purified of any incumbency effect we need a time interval large enough. 6 Table 1 reports the results without senators individual controls. Columns 1,2, and 3 show that primaries have null effect on the average politicians ideology as the coefficients of interest are not significant and are very close to zero in every specifications. Instead columns 4,5 and 6 show that primaries exhibit a strong negative effect on the volatility of the elected politicians ideologies. The reduction of the volatility ranges between the 36% and the 57% depending on the specification. Table 2 shows the results with senators individual controls, that fully confirm the findings just described. Also in these specifications, the average ideology is not affected by the introduction of primary election, but the ideological volatility reduces of about the 70%. Finally, figure 3 confirms that under primary election the ideological distribution of the distance between representatives and senators shrinks, reducing ideological volatility of senators respect to the average representatives ideology. Overall, the results of specifications (3) and (4) together with the graphical analysis, show that primary elections don t affect the average ideology of the elected politicians, but signifi- 6 In New York State primaries for senators were introduced in 1968, but the first senator chosen for his first time in office by primaries is observed in the 1974. Also from figure 4 and 5 we can see that any effects of primaries on politicians ideologies takes some time to become observable. 12

Table 2: Results for New York Senators (32 years) (28 years) (24 years) (32 years) (28 years) (24 years) distance distance distance volatility volatility volatility Af ter*p rimaries 0.066 0.069 0.058-0.141*** -0.156*** -0.149*** (0.044) (0.049) (0.052) (0.030) (0.034) (0.037) Af ter 0.016 0.007 0.002 0.042* 0.049** 0.039 (0.038) (0.042) (0.047) (0.021) (0.023) (0.026) NewY ork -0.112*** -0.116*** -0.101*** 0.088*** 0.104*** 0.112*** (0.030) (0.032) (0.035) (0.020) (0.021) (0.024) Democrats 0.160*** 0.169*** 0.175*** -0.127*** -0.127*** -0.122*** (0.025) (0.028) (0.033) (0.016) (0.017) (0.021) Age 0.004** 0.004** 0.003** 0.001** 0.001** 0.002* (0.001) (0.002) (0.002) (0.001) (0.001) (0.001) Experience -0.009*** -0.010*** -0.010*** 0.001 0.002 0.002 (0.002) (0.003) (0.003) (0.001) (0.001) (0.002) Moderate 0.051*** 0.057*** 0.076*** (0.017) (0.019) (0.021) Constant -0.291*** -0.270*** -0.272*** 0.040 0.019-0.000 (0.080) (0.084) (0.088) (0.044) (0.047) (0.052) Observations 118 101 85 118 101 85 R 2 0.4582 0.4734 0.4929 0.5484 0.5665 0.5978 p 0.000 0.000 0.000 0.000 0.000 0.000 Regression for New York and New Jersey senators of specification (3) and (4) including controls. After is a time dummy to identify observation after the introduction of primaries, NewY ork is a dummy that identifies senators from New York, Age is the age of the senator when the congress start, Experience is the number of years in office as a senator, Democrats is a dummy to identify the party affiliation, Moderate identifies whether the deviation from the average representatives ideology leans on the moderate or on the extreme side. In parenthesis I report robust standard errors. cantly decrease the volatility around the average. The magnitude of the effect on the ideological volatility is quite surprising: in five out of six specifications volatility gets at least halved by the introduction of primary elections. Moreover, this effect become even more interesting on the light of the null effects on the average ideology. In the next section I show that also Italian data confirm the dimension of this effect. To check the robustness of the results I propose a falsification exercise. I run specification (4) for the states that border with New York. These seven states are not affected by any kind of treatment in correspondence to the introduction of primaries in New York State, henceforth if my analysis is correct I expect a null effect whenever I replace New York with anyone of them in specification (4). Results are reported in Table 3, 4 and 5: any coefficient is indeed not significant, and the point estimations are much closer to zero respect to the ones of New York, reported in the first column of any Table. The falsification exercise definitely confirms the genuineness of the results. 13

However, as explained above, as the types of primaries are many and the countries where they are implemented are very different, a solid causal inference might not be very useful per se: in the next section I look for a different country and very different elections to check whether the results can be taken as general and extended to other countries or are just specific to the case of New York State elections for senators. Table 3: Placebo results: 24 years range for volatility (1) (2) (3) (4) (5) (6) (7) (8) New York Connecticut Maine Massach. New Hamp. Rhode Island Vermont Pennsylv. After*P rimaries -0.111** -0.024-0.033-0.008-0.045 0.013-0.003 0.031 (0.049) (0.041) (0.039) (0.039) (0.047) (0.036) (0.039) (0.052) After -0.011-0.018-0.017-0.016-0.018-0.016-0.015-0.018 (0.035) (0.035) (0.035) (0.035) (0.035) (0.035) (0.035) (0.035) State 0.106*** -0.005-0.022-0.011 0.152*** -0.046* -0.013 0.025 (0.026) (0.026) (0.025) (0.025) (0.031) (0.024) (0.025) (0.030) Moderate 0.104*** 0.034* 0.048** 0.050*** 0.037 0.057*** 0.065*** 0.036 (0.022) (0.018) (0.020) (0.018) (0.023) (0.021) (0.019) (0.025) Constant 0.045** 0.096*** 0.086*** 0.084*** 0.094*** 0.080*** 0.074*** 0.095*** (0.021) (0.022) (0.022) (0.021) (0.024) (0.021) (0.021) (0.024) Observations 85 85 72 85 82 83 76 85 R 2 0.3104 0.0756 0.1341 0.1122 0.2855 0.1408 0.1421 0.0767 p 0.000 0.052 0.001 0.015 0.000 0.002 0.008 0.047 Regression for states that borders with New York, and New Jersey senators of specification (4) in a 24 years range around 1968.After is a time dummy to identify observation after the introduction of primaries, NewY ork is a dummy that identifies senators from New York, Moderate identifies whether the deviation from the average representatives ideology leans on the moderate or on the extreme side. In parenthesis I report robust standard errors. 14

Table 4: Placebo results: 28 years range for volatility (1) (2) (3) (4) (5) (6) (7) (8) New York Connecticut Maine Massach. New Hamp. Rhode Island Vermont Pennsylv. After*P rimaries -0.100** -0.009-0.022-0.016-0.005 0.007-0.023 0.053 (0.044) (0.036) (0.033) (0.034) (0.043) (0.031) (0.036) (0.046) After -0.015-0.025-0.022-0.021-0.026-0.021-0.017-0.023 (0.031) (0.031) (0.031) (0.030) (0.031) (0.031) (0.031) (0.031) State 0.086*** -0.025-0.036* -0.014 0.109*** -0.044* 0.002 0.009 (0.025) (0.023) (0.022) (0.022) (0.029) (0.023) (0.025) (0.026) Moderate 0.080*** 0.021 0.039** 0.043** 0.017 0.045** 0.068*** 0.031 (0.022) (0.017) (0.018) (0.017) (0.021) (0.020) (0.019) (0.022) Constant 0.071*** 0.112*** 0.100*** 0.097*** 0.115*** 0.095*** 0.080*** 0.105*** (0.021) (0.020) (0.020) (0.020) (0.022) (0.020) (0.020) (0.022) Observations 101 100 88 101 98 99 88 102 R 2 0.2621 0.0728 0.1481 0.1245 0.2194 0.1299 0.1583 0.0851 p 0.000 0.044 0.000 0.001 0.000 0.001 0.001 0.016 Regression for states that borders with New York, and New Jersey senators of specification (4) in a 28 years range around 1968.After is a time dummy to identify observation after the introduction of primaries, NewY ork is a dummy that identifies senators from New York, Moderate identifies whether the deviation from the average representatives ideology leans on the moderate or on the extreme side. In parenthesis I report robust standard errors. Table 5: Placebo results: 32 years range for volatility (1) (2) (3) (4) (5) (6) (7) (8) New York Connecticut Maine Massach. New Hamp. Rhode Island Vermont Pennsylv. After*P rimaries -0.075* -0.006 0.018-0.029 0.030 0.005-0.032 0.061 (0.040) (0.033) (0.035) (0.030) (0.040) (0.028) (0.032) (0.042) After -0.021-0.030-0.025-0.025-0.035-0.026-0.021-0.029 (0.027) (0.027) (0.027) (0.026) (0.027) (0.027) (0.027) (0.027) State 0.066*** -0.026-0.045** -0.002 0.071** -0.037* 0.013 0.008 (0.023) (0.022) (0.019) (0.020) (0.029) (0.021) (0.022) (0.023) Moderate 0.069*** 0.025 0.048*** 0.048*** 0.001 0.043** 0.071*** 0.030 (0.020) (0.016) (0.017) (0.015) (0.020) (0.018) (0.017) (0.020) Constant 0.082*** 0.112*** 0.096*** 0.096*** 0.128*** 0.100*** 0.081*** 0.109*** (0.019) (0.018) (0.018) (0.017) (0.020) (0.018) (0.018) (0.019) Observations 118 116 103 119 115 114 101 119 R 2 0.2191 0.0891 0.1288 0.1689 0.1626 0.1304 0.1945 0.1001 p 0.000 0.008 0.001 0.000 0.000 0.000 0.000 0.004 Regression for states that borders with New York, and New Jersey senators of specification (4) in a 32 years range around 1968.After is a time dummy to identify observation after the introduction of primaries, NewY ork is a dummy that identifies senators from New York, Moderate identifies whether the deviation from the average representatives ideology leans on the moderate or on the extreme side. In parenthesis I report robust standard errors. 15

3 Italian Municipalities Let me consider a database on Italian municipal elections. This exercise allows me to show that the causal results that I found for US senators from New York are robust across country, across level of election and across time. In Italy there are 8,010 municipalities. They are governed by a mayor that together with a team of local ministers, has a substantial and decisive role in determining policies. Italian electoral law guarantees the mayor a large majority in the city council (independently from the margin of votes on the runner up) that gives him a lot of freedom to implement almost any policies. In municipalities with more than 15.000 inhabitants the majority bonus guarantees the party of the major to have at least the 60% of the seats in the councils, and at least the 67% in municipalities with less than 15.000 inhabitants. Hence, governability is almost never an issue for Italian mayors, and cases where the mayor loses the majority in the council and is obliged to resign happen but are really rare. Since 1999, one of the most important decision that a mayor is annually required to make is to set a local income tax. Until 2006 this tax could range between the 0% and the 0.5% and had to be a fix amount equal for any person resident in the municipality. Starting from 2007 two things changed: firstly, the upper limit of the local income tax reached the 0.8% of the gross personal income, and secondly mayors were allowed to switch to a progressive tax scheme on the basis of the personal income, and to give exemptions to particular category of people (such as people with disabilities, elderly etc.). The municipal tax sums up to a regional and a national tax rate, where the former is roughly the double of the municipal tax, and the latter represents at least the 90% of the overall direct income taxation for the lowest tax bracket, and the 96% for the highest bracket. Italian municipalities have a particular experience of primary elections for mayoral candidates. The first primaries were spontaneously organised in 2004 by the lefty coalition, and there is not any national law that regulates them. Since then, local primaries are held only if a local party (or eventually a coalition) decides to organise one to choose its candidate, and even though primaries are more common in bigger cities, in principle they are a completely endogenous phenomena, definitely driven by a very large list of political and geographical unobservables characteristics. Whenever a primary election takes place, the rules are established by a regulation written by 16

Table 6: Summary statistics on italian primary elections primaries lefty average turnout average inhabitants winner 2004 4 4 19.27 8336.5 4 2005 2 2 10.63 35074.5 1 2006 29 28 10.62 82857.41 12 2007 48 48 10.62 64010.52 14 2008 37 33 10.22 33107.97 15 2009 177 170 13.02 21859.28 103 2010 37 35 10.36 35154.22 18 2011 58 53 11.02 95445.03 36 2012 117 97 9.57 49780.14 64 2013 94 84 11.34 60695.55 52 number of primaries, number of primaries of lefty party or coalition, average turnout and inhabitants and how many candidates from primary won the general election the local party management that organise the primary election, and in principle after primaries have been done the party is not bounded by any law to choose as a candidate the winner of the primaries. The most common type of municipal primaries is the open primary with a single round election held in a single day in the party venues. Runoff primary elections are really rare at municipal level. Since 2004 7, in Italy there have been 821 municipalities were at least a party held a primary elections, out of them 775 were called by a lefty party or coalition. In Table 6 I report some summary statistics on Italian primaries for mayor. This numbers show clearly that local primary elections have been so far almost an exclusive characteristics of the lefty coalition, henceforth I run my analysis on lefty mayors focusing on 2009 municipal elections, where the sample of primaries is larger and the behaviour of the mayors is now observable for a satisfactory amount of years. 3.1 Data and measure of Ideology I get the data on municipal primaries from the C&LS 8, the data on elections results from the Italian Ministry of Interior and the data on municipal income tax since 2002 from the Italian Ministry of Economy and Finance. I proxy the ideology of a mayor with the local income tax rate that they set. This approach is very similar to Bordignon, Nannicini and Tabellini (2013) with the only difference that they use a slightly different local tax on business property, whereas I use the local income tax. I do not consider municipalities with progressive taxation, and I just keep the municipalities where the tax rate is fix for the majority of the population, allowing 7 until the 31/12/2015 8 Candidate and Leader Selection standing group 17

Table 7: Partisanship in local income tax rate between 2009 and 2014 lefty mayors average tax lefty mayors righty mayors average tax righty mayors difference 2009 304.4542763 413.4030387.0512376*** 2010 302.4512583 413.4016828.0495755*** 2011 298.4683893 411.4155353.052854*** 2012 234.5688462 367.5044687.0643775*** 2013 213.5979812 351.5550997.0428815** 2014 198.6218182 349.5827221.0390961* for minor exemptions for particular categories 9. Firstly, I need to show that this proxy is indeed a good measure for politicians ideologies: looking at Table 7 is possible to note that there is a clear partisans effect on the tax rate set since 2009; in particular in a subsample where the partisanship of the mayor is identifiable, local taxes are significantly higher for lefty mayors respect to righty mayors. The differences are strongly significant in every year, and are considerable also in terms of magnitude as they range between the 6% and the 12% of the average local income tax. I make use of this partisanship in tax rate policies to proxy the ideology of the mayors: the higher the municipal income tax a mayor sets, the more his ideology is lefty; viceversa the lower is the tax the more righty is the mayor. This findings can be easily justified by different preferences of the mayors on the level of redistribution: the more lefty the mayor the more he would favour redistribution with an higher taxation of personal income, allocating the income taxes proceeds to a variety of public good or services as daycare schools, sports, local health districts, social assistance and any kind of public infrastructures. Viceversa, the more righty the mayor, the more he would privilege competition and supporting business activities without distorting the market with heavy taxes. 3.2 Empirical Strategy I analyse the elections in 2009 when 4.098 municipalities renewed their mayors. I set up once again a diff-in-diff analysis between lefty municipalities that did primaries and lefty municipalities that did not. I reduce the sample to the lefty municipalities meaning the municipalities that were governed by lefty mayors 10 both before and after the year of the election: this is 9 from the municipal balance sheets it is quite clear that exemptions don t affect the finance of the municipalities, and that the main tax rate would be alternatively be the same without them. 10 I consider lefty all the mayors affiliated to Democratici di Sinistra, Italia Dei Valori, Verdi, La Margherita, SDI, L Ulivo, L unione, Lista Arcobaleno, Partito Democratico, Rifondazione Comunista, PDS, Sinistra Democratica. If there are more than one candidate from a lefty party I drop the municipality from the sample. 18

because I need to get rid of any possible different type of selections between the two groups based on the identity of the former mayor. In other words, consider as an example that primary candidates are stronger whenever the former candidate is righty, this would confound the effect of primary elections with the effect of a switch in the party that is in power. Moreover, as primaries usually takes place when the former mayor is term limited, I need to restrict the sample to municipalities where the incumbent mayor is term limited, to avoid that other effects due to the difference in term limited mayors kick in at the same time of the election and pollute the results. In many municipalities (especially in the smallest), the partisanship of the mayor is not identifiable as they are not affiliated to national parties but run in the election with local lists, and therefore the sample of lefty municipality whom mayor is term limited and did election in 2009 is relatively small. Overall, I end up with a sample of 68 municipalities among which the exact half did primaries. This narrow sample is guaranteed to be clean from any possible disturbance. To set up the diff-in-diff, I separate the municipalities that did primaries for 2009 election from the municipalities that did not, and I compute the average ideology for each group as the average tax in every year. Similarly I pin down the ideological volatility as the tax variance within each group in every year. In Figures 8 and 9 I draw the time series of the average ideologies and of the ideological volatility of the mayors in the two groups in every year since the 2002. Consider those figures: both groups share the same trend for the average ideology and for the ideological volatility up to the year of the election. Afterwards, the average ideologies still share the same pattern, while ideological volatility sharply decreases for the group of municipalities that did a primary election respect to the group of municipalities that did not. Moreover, the graphs suggest that the 2007 reform of the upper limit of the local income tax crisply change the two variables that I am considering; therefore to purify the results from the change in tax upper limit I will include a dummy in the algebraical specification. I test the common trend assumption as usual by regressing the pre-treatment period on a time dummy allowing for different time trend between groups. Then I identify the effects of primaries on the average ideology and on the ideological volatility with a specification very similar to (3) and (4). In particular, I regress the group average ideology and the group ideological volatility on a constants, on a time dummy af ter that identifies whether the observation takes place before or after the year of the election, on a dummy primaries that tells whether the municipality did primaries for 2009 election, on their cross effect and on the dummy Ref orm to control for the change in the upper limit of the local income tax. The specifications for group 19

s at time t of the elected mayor i is the following. As before, the causal coefficient of interest is delivered by the cross-effect dummy, β 3. average ideology t,s = β 0 + β 1 P rimaries s + β 2 After t + β 3 After t P rimaries s + β 4 Reform t + ɛ t,s (5) ideological volatility t,s = β 0 + β 1 P rimaries s + β 2 After t + β 3 After t P rimaries s + β 4 Reform t + ɛ t,s (6) 3.3 Results Table 8 shows that the pre treatment common trend requirement is fulfilled for the sample of election in 2009. This make us sure that any unobservable differences between group is now taken into account and don t vary the proper effect on the average ideology and on the ideological volatility between groups across time, as also the graphical analysis of Figures 8 and 9 would suggest. Therefore I can exclude the inference of any unobservable covariates that can possibly drive at the same time the decision on whether to do a primary or not and the outcome variables. Results for the estimation of (5) and (6) are in Table 9. Column 1 shows that average ideology is not affected by primaries, as the coefficient of interest is largely not significant. However, Column 2 reports that the ideological volatility sharply reduces under primary elections: the first column reports a significantly negative coefficient of -0.011, with a reduction of approximately the 57% of the group ideological variance. municipalities at the election in 2009 average ideology.3.4.5.6.7 2001 2003 2005 2007 2009 2011 2013 2015 year no primaries primaries Figure 8: average ideologies of municipalities at the election in 2009 20

municipalities at the election in 2009 ideological variance.02.03.04.05.06.07 2001 2003 2005 2007 2009 2011 2013 2015 year no primaries primaries Figure 9: ideological variance of municipalities at the election in 2009 Table 8: Test for the common trend assumption. (1) (2) av.ideology ideological vol. P rimary*y ear -0.007 0.001 (0.006) (0.001) Y ear 0.031*** 0.002** (0.007) (0.001) Constant 0.214*** 0.025*** (0.028) (0.003) Observations 14 14 p 0.003 0.008 Table 9: Results of (5) and (6) for Italian 2009 elections. (1) (2) average ideology ideological volatility Af ter*p rimaries -0.019-0.011* (0.035) (0.006) After 0.099*** 0.013** (0.029) (0.005) P rimaries -0.033 0.004 (0.022) (0.004) 2007Ref orms 0.155*** 0.012*** (0.023) (0.004) Constant 0.298*** 0.027*** (0.018) (0.003) Observations 26 26 p 0.000 0.000 21

4 The Possible Mechanisms The two datasets exploited in this paper regard two very different frameworks. The elections of senators from New York between the 50s and the 90s are different in many aspects respect to the mayoral Italian elections in the late 2000s. Firstly, the political histories of the two countries are deeply different, and in particular US elections have always been dominated by a two party system. Instead, Italian electoral system both at national and local level favoured the coexistence of a high number of parties, that often decided to form coalitions for single election. Secondly, also the kinds of primaries are radically different: candidates for senators from New York were chosen with closed primaries, a type of primary where the right to vote is given only to people registered to the party; wheras Italian mayors faced open primaries, a type of primaries where any citizen allowed to vote in the general election is also allowed to vote in the primary election. A third significant difference between the two datasets is the level of the election: US Senators are state wide elected for a national office, and the senatorial office is a typical path through which aspirant presidents try to make inroads into the White House. Italian mayoral elections instead are a local matter: municipalities are often very small, and candidates are often strongly linked somehow to the municipality where they decide to run. Moreover, only the big cities mayoral offices give to the mayor the opportunity to have a career at national level. Instead, for small towns mayoral offices are often the peak of the careers of the mayors, and some of them directly go back to their previous jobs when their offices end. A fourth and obvious difference is the political timing: the political debate between the 50s and the 90s was centred on the post-war policies and then on the Cold War, with a traditional media information and the absence of instrument as the internet. Instead, in the late 2000s Italy was involved in trying to solve the economic crisis mostly talking about unemployment and austerity measures (that were sharply imposed to many municipalities with a cut to the transfers from the national government), and the newest media and the social networks were largely in use. Despite such significant differences between the two datasets, the two reduced form estimations deliver the same results. Henceforth, it seems that the results exposed above are independent of the type of primaries, the period, the level of the election and the country that adopts primaries. Here below I propose an investigation of the possible mechanisms that drive the results. As a matter of fact, none of the theories I am aware of deliver the results shown above. Therefore, to 22

enlighten the mechanism that trigger the reduced form results I consider any possible change that primary elections carry on to the political competition, and I analyse them separately. I find five main aspects that get significantly changed whenever primary elections are introduced: the type of campaign that a politician needs to face; the possible coordination on candidates identities among different elections between the people that choose the candidate, the so called selectorate; the capacity to gather and elaborate information of the selectorate; the capacity of being strategic of the selectorate; and the pool of people that sort in as aspirant candidate. I make conjectures on all these possibilities, relating to the existing literature whenever possible. I ll be able to rule out the first of them. The second channel is plausible but data don t seem to support it. Furthermore, I show that the third and fourth channels in the list might be at work and can be responsible for the results that I got in the reduced form estimations, even though I am not able to disentangle them. Unfortunately, I won t be able to say much on the last. 1. The first hypothesis, from Hummel (2010), follows from the fact that with primaries campaigns are longer and initially addressed to a different pool of people. Hummel (2010) conjectures that candidates from primaries flip-flop between the primary and the general election. Primary electorates are definitely more extreme on average than the electorates of the general elections. Hence, candidates may find convenient to propose more extreme policies during the primaries and then reconsider their stances after primaries (even if they need to pay for a cost in term of electoral outcome) moving back to more moderate platforms to please the more moderate general electorate. In Hummel (2010), depending on the magnitude of the parameters of the model, this flip-flopping behaviour introduced by primary elections can deliver more moderate or extreme pool of elected politicians. However, there is no way that this hypothesis triggers the results I find in the empirical section: even if politicians would follow such strategies, all candidates would find convenient to behave in the same way. Regardless whether they would overall end up on more moderate or extreme stances, primaries would generate an effect on the average ideology and not on the ideological volatility. 2. The second possibility concerns the possible coordinations over candidates identities of the selectorate. Parties are formed by factions, and coalitions are formed by parties. 23

When candidates are not chosen by primaries, the subgroups (factions or parties) might have incentives to coordinate. A subgroup can decide to support a candidate from another subgroup in an election to have his proper candidate supported by the other subgroup in the next election. Nevertheless, such coordination can happen between a number of elections that take place at the same time: subgroups leaders can split the candidates and support them all reciprocally. Such coordination would be impossible when primary elections are in force: primary voters don t have any incentive to coordinate with other voters across time, neither have incentive to coordinate with voters of other elections that take place at the same time. However, as primary voters are from all the subgroups, the candidates that win the primaries would have to be ideologically somewhere in the middle between the factions or the parties ideologies. Said in a more formal way, the intra party bargaining over the candidates ideologies with office motivated subgroups can result in pareto optimal candidates allocations where the candidates are ideologically more volatile when primary election are not performed. This could indeed deliver more volatile candidates without primary elections, without necessarily influencing the average ideology of the candidates. Even if it is quite suggestive, this hypothesis is not confirmed by the data. Considering the dataset on US senators from New York, their ideological records don t exhibit any regular pattern of alternation when primaries don t take place, as this hypothesis would require. Nevertheless, considering the dataset on Italian municipalities, the decisions over the identity of a candidate is to a large extent a very local matter, that wouldn t be of any interest for the national leaders. Indeed, just for very big municipalities the national party leaders are interested to influence the decision. For medium small municipalities as the ones that dominate my sample, the local party or coalition has full freedom and the most of the interests to decide who is going to run as a candidate. Furthermore, it would be extremely difficult to sustain a coordination between the myriad of local parties of all the municipalities at an higher level to divide candidates between subgroups. All the same, Italian data show that there is not any regular division of candidates ideologies across different years of election. 3. The third explanation looks at the difference in the capacity of the selectorate to have information on the tightness of the general election. In particular, I argue that primary voters are not able to get information as party leaders are, and this can at least partially 24