Biased Democracies: The Social and Economic Logic of Interest-Based Voting

Transcription

1 0 Based Democraces: The Socal and Economc Logc of Interest-Based Votng Torben Iversen Department of Government Harvard Unversty Davd Soskce Department of Government London School of Economcs Earler versons of ths paper were presented at the Unversty of Essex, November 2008, at the Annual Meetngs of the Amercan Poltcal Scence Assocaton 2009, at the Educaton, Relgon, and the Gender-Vote-Gap Conference, Unversty of Bremen, October 14-15, 2011, and at the Poltcal Economy of Sklls and Inequalty workshop, July 26 th and 27 th 2012 at Unversty of Konstanz, Germany,. We wsh to thank the partcpants at these conferences for many useful comments, and n partcular Geoffrey Brennan, Geoff Evans, Rob Franzese, Ncola Lacey, Rakeen Mabud, Phlp Manow, Davd Rueda, Vera Troeger and Paul Whteley.

2 0 Abstract Whle many people are ratonally gnorant about poltcs, most poltcal economy models of preferences and votng assume that people are well-nformed about ther nterests. Here, we endogenze the ndvdual ncentve to acqure poltcal nformaton, wth stark mplcatons for understandng partsan poltcs. We argue that the ncentve to acqure poltcal knowledge s a byproduct of other ncentves, both prvate and socal, and that these ncentves are unequally dstrbuted across groups. For those who lack ncentves to be nformed the ratonal strategy s to vote centrst. We test the model on publc opnon data from 16 advanced democraces and show that the couplng between economc nterests and poltcal choce depends on socal networks that have weakened over tme. Because those wth low nformaton vote centrst we can explan the declne n class votng as reflectng a declne n the socal ncentves to be poltcally nformed, lnked to declnng unon membershp.

3 1 1. Introducton In the past decade and a half a rch lterature n the comparatve poltcal economy of advanced socetes has emerged on ndvdual economc preferences and votng, and the consequences for redstrbuton. Yet, our emprcal and theoretcal understandng remans lmted. At the ndvdual level basc economc varables, such as ncome and educaton, are only weakly related to polcy preferences or to votng (Margalt 2013; Lynch and Myrskyla 2009), and at the macro-level hgh and rsng nequalty, contrary to expectatons, has not ncreased support for redstrbuton and the poltcal left (Georgads and Mannng 2012; Allan and Scruggs 2004). In ths paper we propose that an mportant element n solvng these puzzles s the (changng) ncentves that dfferent groups have to be nformed about poltcs, and the logc of votng choce for those wthout sgnfcant nformaton. 1 We present a model of votng, consstent wth ratonal choce, wth two new features: endogenous ncentves to acqure relevant nformaton and the logc of votng when nformaton s lmted. Jontly these features sgnfcantly mprove our ablty to explan ndvdual votng behavor, and they have strkng mplcatons for partsan and dstrbutve poltcs across countres, ncludng the poltcal shft to the rght. Specfcally, our model mples that the unnformed tend to vote centrst, and that the ncentves to acqure nformaton (ncreasngly) favors the poltcal rght. Our model of class votng flls what we see as a major lacuna n comparatve poltcal economy. The exstng lterature s largely focused on testng models of economc nterests the mportance of ncome (e.g., Moene and Wallersten 2001; Cusack et al. 2006), rsk and job securty (e.g., Rueda 2005; Rehm 2011; Margalt 2013), educaton and sklls (e.g., Iversen and 1 Assumng they decde to vote. The paper does not examne the logc of partcpaton, although we suggest below why endogenzng ths choce s lkely to renforce our conclusons.

4 2 Soskce 2001; Busemeyer 2012), property and debt (e.g., Ansell 2012), sector (e.g., Rehm and Wren 2012), and so on and how these nterests shape preferences for polces and poltcal partes. But ths work does not address the queston of how people come to understand ther nterests (f they do) and the effect of publc polces on these nterests. Ths s not only an error of omsson but of commsson because Downs long ago argued that voters have an ncentve to be ratonally gnorant (Downs 1957). Moreover, we show that those who are nformed vote dfferently than those who are not, and that the dstrbuton of ncentves to be nformed s nonrandom. How people come to understand ther nterests s therefore consequental for explanng votng and publc polces. Smply put, where most current research n the feld seeks to pn down the nature of peoples nterests under the assumpton that these are common knowledge, we take nterests as gven and ask how people come to know and vote on these nterests. Our answer suggests a major research fronter at the ntersecton of comparatve poltcal economy and socologcal network analyss, whch complements comparatve work on economc class and votng (see Evans and de Graaf 2013 for the most comprehensve recent treatment) by brngng n the role of nformal networks and dscusson. In modelng the process of nformaton acquston we post two ratonal motvatons. The frst s the prvate economc ncentve to seek nformaton about publc polces and regulatons, and how governments may change these, because such nformaton facltates better fnancal and management decsons when the outcomes depend on future publc polces. As orgnally argued by Larcnese (2001), prvate ncentves tend to be concentrated among those wth hgh ncomes and wealth, or among those who are n manageral or supervsory postons where famlarty wth publc polces and how they may change s mportant for makng sound economc decsons. Prvate ncentves are not exclusvely reserved for the rch and powerful because ordnary people are also makng longer term nvestments n schoolng, pensons and

5 other areas where poltcal nformaton about the future may be relevant to current decsons. But we suggest that prvate ncentves are ncreasng n economc varables that are conducve to rght votng. The second, and equally general, motvaton s socal. Mountng evdence n socal psychology shows that human behavor s strongly motvated by the desre to gan the recognton and respect of others (Baumester and Leary, 1995), and ths nsght has ganed nfluence n both poltcal scence (e.g., Abrams et al. 2011) and n economcs (e.g., Brennan and Pettt 2004). The desre to belong to groups and to receve the approval of peers s now wdely thought to have an evolutonary bass. As Baumester and Leary (1995) suggest, these socal ncentves are as fundamental as the desre for personal securty and materal welfare, and they cannot be reduced to the economc ones (even f they surely complement each other n many cases). The socal desre for approval, we argue, has two effects: Frst, t motvates some people to acqure costly knowledge about poltcs when poltcs s a recurrent topc of dscusson n the groups and nformal networks whch they belong to. Second, t motvates other people to assmlate more passvely the opnons of those around them when group members can reasonably be assumed to have well-algned nterests and when conformty s valued. Formal groups lke unons make use of socal ncentves, snce n effect the nformal socal networks of many workers are partally wthn unons; and we wll argue that ther declne snce the 1980s has been a sgnfcant cost to the poltcal left. The role of dscusson of poltcs has a long pedgree n socology, begnnng wth the poneerng work of Lazersfeld and Berelson and ther assocates (Lazersfeld et al 1944; Berelson et al 1954) and echoed more recently n the hghly nfluental study by Huckfeld and Sprague (1991, 1995) and an nsghtful recent book by Walsh (2004). We buld on ths work and connect t to poltcal economy and ratonal choce n a manner that algns t wth manstream models n 3

6 comparatve poltcs. What we seek s a cross-dscplnary ntegraton of these nsghts that sheds lght on comparatve patterns of votng and partsan poltcs from a ratonal choce perspectve. Our econometrc tests show the hghly sgnfcant role of poltcal dscusson n nformal poltcal networks n correctly dentfyng one s poltcal nterest. But why exactly? Here the composton of networks s of great mportance. The crtcal glue between poltcal scence and socologcal analyses s what socologsts call homophly: Informal socal networks tend to be structured on an economc bass, so that people from smlar economc strata are much more lkely to assocate wth each other than those from dfferent strata. People from the same class marry each other at hgher rates; they tend to lve n the same neghborhoods, they work n the same offce or factory, and they send ther chldren to the same schools: Brds of a feather flock together s the ttle of most wdely cted revew of the lterature (McPherson et al 2001). Homophly s a crtcally mportant fact about socal realty because t means that what people learn from poltcal dscusson n ther networks tends to be close to ther own nterests. Overrdng the Downsan ncentve to be gnorant, poltcal dscusson n socal networks gves poltcal economy models explanatory power. Organzed groups understand the power of socal ncentves and they try to use socal networks to ther advantage. Buldng on a new study by Ahlqust and Lev (2012), we fnd strong ndcatons n our data of such organzatonal effects n the case of unons. As powerfully argued by Przeworsk and Sprague (1982), organzed uses of socal ncentves have hstorcally served as counterweghts to the concentraton of prvate ncentves among rght consttuences; and we show that the declne of unons snce 1980, addng to the effect of the declne of mass partes, has played a major role n the declne of the left. We present our argument n the form of a formal model that can be drectly estmated usng the data we have. We wll show that when nformaton s lmted there s a strkng centrst 4

7 5 bas n votng, whch we beleve s a completely new result. Because the ncentves to be nformed are stronger on ths centrst bas s more consequental on the left, wth mplcatons for the left-rght balance of support. The presence of partsan bases s strongly confrmed n publc opnon data from 16 advanced democraces, whch shows that among those who rarely dscuss poltcs and are not n unons, the support for the center and rght s much hgher. We further show that left support s strongest when poltcal dscusson s nteracted wth unon membershp. The rest of the paper s dvded nto three sectons. The frst presents the model, the second tests the model on comparatve survey data on votng, and the thrd summarzes the key results and dscusses broader mplcatons of the argument. 2. The Model Our pont of departure s a very smple spatal model of votng, where voters are ncompletely nformed about ther own poston relatve to those of partes that they can vote for. Wth lmted nformaton the model shows that people are much more prone to vote centrst. We then ntroduce prvate and socal ncentves and consder the consequences for partsan poltcs A smple spatal baselne model Take a very smple case of lmted nformaton from whch we wll derve some key ntutons. In ths case there are three partes, left (L), center (C), and rght (R), wth polcy postons unformly dstrbuted on a left-rght scale wth the range[ aa, ]: C = 0, L= br, = ba ; > b. We assume that voter chooses L, C or R to mnmze expected losses, where the ex-post loss to from votng for party P s the dstance between the party poston and the voter poston, y: (1) L = P y. Wth complete nformaton about y and P, ths reduces to a smple spatal votng model where always chooses the party closest to her (Adams 2001).

8 6 The proxmty result s smple and ntutve, and t s wdely used n the comparatve votng lterature. Yet, surprsngly, t s fundamentally altered f voters have ncomplete nformaton. 2 Voters may be uncertan about both ther own polcy preferences and those of poltcal partes. The latter arses smply from lmted nformaton about polcy platforms, and the credblty and capacty of partes to mplement these platforms. Voters may also be uncertan about ther own polcy preferences. Even f preferences were smply a functon of current ncome, as n the canoncal Meltzer-Rchard model (Meltzer and Rchard 1981), what s optmal depends on the progressvty of taxaton, the effcency costs of such taxaton, and the composton of spendng. The nformaton problem s greater stll when preferences are dependent on multple factors n addton to ncome, as mpled by most poltcal economy models,. Even f voters knew the polcy postons of partes, and could trust partes to mplement these, they would not necessarly know whch polces would best advance ther nterests. So f there are costs to knowledge acquston, t s not hard to see why people often make decsons that devate from standard model predctons. To begn wth the most extreme possblty, magne now that a voter s completely unnformed about poltcs ( ratonally gnorant ). If she votes, who should she vote for? 3 If the 2 In the spatal votng tradton t s sometmes recognzed, gong back to Downs, that voters are nformaton msers who partes can only capture by assmlatng a large number of ssues nto a sngle deologcal left-rght dmenson (Munger and Hnch 1994). The spatal analyss s then appled to ths smpler world. But, of course, ths does not explan how voters decde where they belong on the deologcal left-rght scale. 3 Snce the ncentves to acqure nformaton are almost certanly related to the ncentves to vote, those wthout nformaton are also less lkely to vote (see Abrams et al, 2011). Yet, voter non-

9 7 partes are unformly dstrbuted, the temptaton would be to say that she should vote for each party wth equal probablty. After all, each party s equally lkely to best represent her preferences. But that ntuton s wrong. In fact, wth reasonable assumptons the voter would always vote for party C. The reason s that voters are tryng to mnmze losses (measured as spatal dstances), and votng for C mnmzes the expected loss. Note agan that ths concluson s an mplcaton of the standard spatal votng model under uncertanty. It does not requre any specalzed assumptons or non-lneartes n the utlty functon. 4 To llustrate the logc, magne that voter s has no nformaton and shares the postons of L, C, or R wth equal probablty (we wll subsequently generalze ths smple case to any dstrbuton of voters). If votes L the expected loss, L, λ, s ( C L) ( R L) , where the frst term s the (zero) loss f L turns out to be the correct party (e., the most representatve), and the other two terms are the losses f ether C or R s the correct party. Smlarly, f votes C the expected loss, λ, s ( C L) ( R C) C, It s easy to see that the only dfference between the two expected losses s the last term, whch s greater f votes L than f votes C. So λl, > λc,. The same holds f we compare R to C snce λl, = λr, f we assume symmetrcal spacng. Votng C s therefore always the best choce for. Ths logc wll prove mportant to our results, and we show n appendx A that t can be generalzed to any dstrbuton of subjectve probabltes across the three partes. Specfcally, turnout that s related to the lack of ncentves to be nformed only renforces our conclusons, so we assume that everyone votes for smplcty. 4 Some spatal votng models assume that voters mnmze squared dstances. Ths makes no dfference to our substantve results.

10 8 we show that for any contnuous dstrbuton f ylc when the porton of the dstrbuton to the rght of ylc (2) F( y LC ) > 0.5 s the mdpont between L and C, votes left s less than 0.5, or: An analogous result holds for votng rght. The mplcatons s that for to vote left or rght there must be a greater than one half probablty that L or R s the best party. If the choce between the three partes was symmetrcal π, = π, = π, = 1/3, so ths condton would not hold. Agan, L C R there s a centrst bas n votng under uncertanty, and ths bas has a very precse meanng. An ntuton behnd ths result, whch to our knowledge has not been derved before, s that people who are uncertan about ther polcy preferences relatve to those offered by partes try to nsure themselves by votng for the centrst party rather than non-centrst partes. If the center party s close to the status quo, ths wll appear emprcally as a status quo bas, but note that such a bas s a consequence of loss averson n our model; t s not an assumpton. Note also that the centrst bas does not mean that voters tend to choose partes that are dfferent from ther own stated polcy postons. Polcy-preferences wll mrror party preferences n the sense that uncertan voters who vote centrst are also lkely to express centrst polcy opnons. Our model s therefore fully consstent wth mountng evdence that voters tend to poston themselves at the center of the poltcal space compared to most poltcal partes (see Macdonald, Lsthaug and Rabnovtz 1991; Iversen 1994; Kedar 2005). Indeed, nsofar as partes represent the preferences of nformed voters (ncludng, presumably, party eltes themselves), our model explans ths long-standng puzzle. Up to ths pont the left and rght are equalty dsadvantaged by the centrst bas, but ths depends on the level of nformaton beng equal across the poltcal spectrum. We therefore need to be explct about the process by whch people acqure nformaton. Such nformaton could be

11 about both the postons of poltcal partes or voters own postons, but for smplcty we can fx ether and vary the other wthout loss of generalty. We assume here that party postons are gven but that voters are uncertan about ther own preferences relatve to those of partes Endogenzng the acquston of knowledge Our startng pont s a ratonally gnorant voter who only knows that there s some probablty dstrbuton of hs or her nterests across the polcy space. We assume that the dstrbuton s symmetrc and centred at the mean of the space so that we do not prejudce the results n one partsan drecton or another. Apart from ths requrement, the dstrbuton could have any concevable shape, ncludng mult-peaked, polarzed, and normal. As we showed above that the centrst bas s ndependent of the dstrbuton. In the followng we assume a normal dstrbuton because t allows us to easly translate the model nto probt, whch can be estmated drectly. 5 We can also justfy ths assumpton on emprcal grounds snce the dstrbuton of voter preferences n most countres s approxmately normal, and t would be natural to thnk that the overall dstrbuton s the startng pont for most voters. But, agan, none of our substantve analytc results depend on assumng a normal dstrbuton. More precsely we assume that (unnformed) voter s pror s a normal dstrbuton wth 2 mean y and a varance ofσ : 5 Ths s strctly speakng an approxmaton because the polcy space s constraned to the range [ aa, ], but snce we can choose a to be arbtrarly large there are no practcal consequences of dongs ths. The unform dstrbuton we consdered above s now approxmated by a normal dstrbuton whereσ 2 for large enough a.

12 10 y 2 (, ) N y σ It s natural to thnk that y s close to, or dentcal to, the poston of C, but we can allow the data to determne the exact locaton. All people start out wth ths pror dstrbuton, but they are lkely over tme to acqure at least some nformaton about ther nterests from newspapers and other meda, from the formal groups they belong to (such as unons and churches), and from the dscusson of poltcs wth others n ther nformal networks of famly, frends, neghbors, and co-workers, whch we treat as exogenous. 6 We assume that voters use ths knd of nformaton to update ther pror and arrve at a poston that s closer to ther true preferences. 7 Specfcally, the sgnal receves through s network or organzatonal afflatons s called y*, and we assume that the sgnal s drawn from a normal dstrbuton centered on s own nterests, y: y N( y, σ ). * 2 Informaton s thus unbased, and ths assumpton s lkely to be a good approxmaton to realty as long as the network or group s small and homogeneous, consstng of people who are smlar to n terms of relevant socoeconomc attrbutes -- what socologsts call homophly (see McPherson et al. 2001). We wll dscuss below what happens f the group s large and 6 It s possble that some people choose ther work, neghborhood, and so on because people n these have poltcal vews that are well-algned wth ther own, although we thnk ths s rare. Note however that n ths case socal nteracton nsde networks should have no ndependent effect on atttudes snce they are decded pror to network membershp. 7 We do not have n mnd objectve nterests n the Marxan sense, but smply the preferences we would observe f people were fully nformed about poltcs. We are agnostc n ths paper about the content of these preferences or what determnes them.

13 11 heterogeneous, such as natonal unons, because then some members wll have economc postons that are dstnct from the center of the group-defned poston. In ths case our model mples that groups can shape ndvdual preferences away from the predcton based on economc poston. In lne wth standard Bayesan updatng, the new nformaton receves through s network s * 2 y and σ, not y tself. In other words, gets a nosy (whle unbased) sgnal about her own nterests. The pror dstrbuton (wth center y ) s then combned wth the new, updatng dstrbuton to yeld a posteror dstrbuton of y gven the sgnal y y * y y y* N σ + σ, σ σ σ σ The posteror dstrbuton s also a normal dstrbuton, and the mean s a weghted average of y* and y wth weghts α and 1 α: 1 2 σ where α σ σ y * y σ σ Ey y* = = αy* + (1 α) y, σ σ 2 2 s the relatve precson of the sgnal * y, and determned by the varance of the pror and updatng dstrbutons. In other words, usng Bayesan updatng, how much weght places on the new nformaton depends on the uncertanty assocated wth ths nformaton compared to the uncertanty assocated wth the orgnal nformaton. Usng the loss functon n (1) wll vote left (L) ff (3) Ey y y y y * * = α + (1 α) < LC,

14 12 where y LC s the mdpont between L and C. Ths s because y y s normally (hence * symmetrcally) dstrbuted, so that Ey y < y mples F( y ) > 0.5. As explaned n * LC LC Apprendx A, votes L when more than half the dstrbuton s to left of the mdpont between L and C. Compared to the example of complete uncertanty, the posteror dstrbuton for a voter to the left of ylc may satsfy ths condton because the nformaton receved from s network s centered on y. If there s complete nformaton, a voter wth y < y wll always vote left snce LC the entre dstrbuton wll be to the left of y LC complete nformaton.. Ths s equvalent to standard spatal votng wth 2.3. Predctng the vote Votng n our model s determnstc, even f there s ncomplete nformaton. Ths s because voter knows the sgnal he or she receves and hence αy expected value of y gven the sgnal * + (1 α) y exactly (ths s Ey y, the * y ). Snce y s also known, whether the best choce s to LC vote L follows from Equaton (3). An analogous argument apples to C and R. But as analysts we do not see the sgnal, * y, that each voter receves. Instead we have to rely on our knowledge of y, whch the voter does not know, to predct s vote probablstcally. Appendx B shows exactly how ths s done, but the ntuton s smple. Essentally we use nformaton about y to predct the probablty that receved a partcular sgnal, whch then determnes the vote. Snce there s a dstrbuton of sgnals we can ask what the probablty of us (as analysts) observng a partcular y s when the sgnal s to the left of the mdpont between L and C, to the rght of the mdpont of C and R, or between the two mdponts. The probablty of votng left s now: *

15 13 (4) 1 σ P( v = L) =Φ σ ( ylc y ) + ( y ) 2 LC y σ Analogous results hold for votng center and rght (shown n Appendx B). The centrst bas when has low nformaton s clear from (4) f yrc > y > y snce when LC σ (e., when the sgnal s very nosy) the probablty of votng center goes to 1 (shown formally n Appendx B). If voters who have no pror nformaton learn nothng from ther network they wll vote centrst. Ths corresponds to the smple example of complete uncertanty that we started out wth. The fnal step t to endogenze the prvate and socal ncentves to acqure nformaton. We do so explctly n Appendx C, but the ntuton s very straghtforward. Socal ncentves to be nformed arse from engagement n poltcal dscusson, D, wth people who can be assumed to share your nterests. If poltcs s a recurrent topc of dscusson n the groups or networks to whch people belong, there s a socal ncentve to acqure nformaton about the group nterest n order to beneft from the socal approval that comes from beng knowledgeable. Prvate ncentves, k p,, arse when nformaton about poltcs enables people wth substantal wealth to make better nvestment decsons, or people n manageral or hgh admnstratve postons to make better busness and organzatonal decsons. Any choces that have consequences beyond the next electoral perod and that are affected by publc polces can be mproved wth knowledge about how these polces may change. 8 8 Note that the prvate ncentve for to acqure poltcal knowledge arses when needs to know, n takng a decson now the return on whch depends on future polces, how lkely those polces are to change. Those on low ncomes may be well-nformed about rules governng, say, welfare and tax credts that affect them, but they only have an ncentve to know about poltcs f

16 14 We combne the two types of knowledge nto a smple addtve utlty functon wth standard concave propertes. Snce nformaton s costly, there are levels of nformaton acqured for prvate and socal purposes that wll be optmal to the ndvdual. The acqured nformaton n turn helps voters choose partes that are well-algned wth ther nterests. As explaned n Appendx C, we assume that when makng the votng decson nformaton acqured for prvate and publc purposes can substtute for each other. Based on these assumptons the predctons of the model can now llustrated as n Fgure 1. The fgure assumes that L, C, and R are located at -1, 0, and 1 n a space that ranges from -2 to +2 (approxmately 95 percent of the fctonal electorate), and t smulates the probablty of votng for each party f the drectons of effects are as predcted by the model. 9 Panel a) shows the probablty of votng for each party when we set k P, = 0, so that prvate knowledge plays no role. Note that when dscusson s low that s, when voters receve only very mprecse sgnals about ther own nterests and therefore lean heavly on the pror dstrbuton the support for the center party tends to be hgh across the left-rght scale. The relatonshp between economc poston and left and rght votng s correspondngly weak. When dscusson and nformaton they are takng decsons (for example nvestng n tranng) the return on whch depends on future polces; by contrast to many professonals or hgh-ncome earners takng fnancal decsons, we assume (and show some evdence) that ths s a relatvely unusual stuaton. 9 The predctons are based on the votng equaton at the end of Appendx C. Snce the model only predcts the drecton of effects, ther magntudes are arbtrary. Snce the model only mples the drecton of effects, ther magntudes are arbtrary. We have chosen these magntudes to roughly correspond to the range of emprcally observed propenstes to vote for dfferent partes, gven our set of varables.

17 15 rse, the support for the center declnes among those who are natural consttuences for the left and rght. The strength of the relatonshp between left-rght poston and left or rght votng correspondngly ncreases. [Fgure 1 about here] Panel b) shows what happens when we ntroduce prvate nformaton (now gnorng the center). When prvate ncentves to acqure nformaton are rsng n y say, a smple lnear functon such as k, P = α + λy -- we fnd that the effect of poltcal dscusson n the support for the rght s smaller than for the left. Ths s because hgh-y ndvduals, who are predsposed to vote for the rght, have greater prvate ncentves to be nformed. They consequently have less reason to acqure addtonal nformaton as dscusson rses. Ths logc has mportant mplcatons for the partsan balance of power because at low levels of poltcal dscusson the rght has a poltcal advantage. Ths advantage dsappears as dscusson rses because low-y voters now have a greater ncentve to be nformed and therefore catch up to the level of nformaton n rght consttuences. In the emprcal analyss we test ths effect and measure ts magntude n terms of vote shares Group versus ndvdual preferences The model we have presented assumes that people are n groups wth others who share ther economc stuaton, and hence ther underlyng nterests. Ths s reasonable because homophly s well-documented n the socologcal lterature. At the same tme, the extent to whch ths s true does vary over tme, across countres, and across segments of the populaton. So what happens when people are members of groups wth heterogenety n terms of socoeconomc stuaton? An ntrgung possblty rased by our argument s that f people assume that those around them have congruent nterests, they wll adopt the mean of the dstrbuton of vews n the

18 group. It s also sensble to assume that what yelds socal approval n groups s to help dentfy the group nterest, not ther own ndvdual nterest. When all do the same, the sgnal people n heterogeneous groups receve s 16 y N( y, σ ) * 2 where y s the center of the dstrbuton of preferences n s group, or the group nterest. Untl now we have assumed that the dstrbuton s centered on y ; the ndvdual nterest. Beng a member of heterogeneous groups where poltcal dscusson and socal ncentves matter wll nduce people to adopt the group nterest as ther own, even f t s away from ther ndvdual nterest as defned by the poston they would adopt wth complete nformaton. Ths means that dscusson could affect not only the level of nformaton, but also preferences. Emprcally, ths wll show up as an ndependent effect of group membershp, shftng the probablty of votng for partcular partes to the left or to the rght of the predcton usng our best estmate of y. As we wll dscuss there s strong evdence of such effects n the case of unon membershp. 3. An emprcal test The emprcal test of our model conssts of drectly estmatng Equaton (6) usng nonlnear ordered probt. Because we do not have nformaton on people s prvate nformaton we only consder the effects of poltcal dscusson, but wth the mplct hypothess that dscusson wll dsproportonately affect the support for the left because of the concentraton of prvate ncentves among rght consttuents (as llustrated n Fgure 1b). To smplfy the estmaton we use a lnear frst-order approxmaton for the coeffcent on y wth a postve slope, LC

19 17 ( α + βd ) y, whle that on y s lnear n D wth a negatve slope, ( η ρd ) y. These LC approxmatons mply the followng estmatng equatons: (5) = =Φ ( ) α + β LC γ + δ η ρ [ αylc βdy LC γ y δdy ηy ρdy ] = =Φ [ α RC + β RC γ δ η + ρ ] [ αylc βdy LC γ y δdy ηy ρdy ] = = Φ [ α + β γ δ η + ρ ] P( v L) D y ( D ) y ( D ) y =Φ + + Pv ( C) y Dy y Dy y Dy Φ + + Pv ( R) 1 y Dy y Dy y Dy RC RC Equaton (7) s techncally a nonlnear ordered probt, and snce ths does not exst as as a canned procedure, we wrote a program for Stata. The basc code for ths program s provded n Appendx D. The model s estmated usng maxmum lkelhood The data We use data from the 2004 Internatonal Socal Survey Program (ISSP) survey of ctzenshp, whch s unque n askng respondents how often they engage n poltcal dscusson. Our analyss covers 16 establshed democraces for whch we have complete data on all the key varables. 10 In total we have nearly 23,000 ndvdual observatons. The survey queston about poltcal dscusson reads: 10 The countres are: Australa, Austra, Brtan, Canada, Denmark, Fnland, France, Germany, the Netherlands, New Zealand, Portugal, Norway, Span, Sweden, Swtzerland, and the Unted States. We had to exclude Ireland and Japan because more than half of the respondents dd not ndcate ther preferred party (64 and 56 percent, respectvely). In addton these two party systems are also dffcult to classfy on a left-rght dmenson because of clentlsm n the case of japan and the hstorcal dvson of Irsh partes on the ndependence ssue.

20 18 When you get together wth your frends, relatves or fellow workers how often do you dscuss poltcs? Note that the queston specfcally refers to people who we would usually consder part of an nformal network, whch s precsely what our model calls for. The respondent could answer never (1), rarely (2), sometmes (3), or often (4). In our pooled sample the dstrbuton of answers across the four categores s 14, 30, 43, and 13 percent, respectvely. The dependent varable s based on a queston about party preferences. In most cases the respondent was smply asked to ndcate whch party he or she voted for n the prevous electon, or would vote for n the next. In some cases respondents were asked more broadly about the party they supported or felt close to. In a sngle case, the US, respondents were gven a non-party choce and asked whether they thnk of themselves as Republcan, Democrat, or Independent, and how strongly they felt about ther allegance. We treated those wth a clear Democratc or Republcan preference as left and rght, respectvely, and the rest (36 percent) as centrst. In all other cases we classfed partes nto left, center, and rght, as detaled n Appendx E. 11 Some small regonalst and sngle-ssue partes were excluded from the analyss because they do not lend themselves easly to a left-rght nterpretaton. 12 On average the dstrbuton of the 11 Although our model assumes (at least) three partes, t can stll be used to predct votng n two-party systems. In ths case unnformed people wll place themselves n the center of the poltcal space and be equally lkely to vote left and rght. As n the three-party case, nformaton wll lead to more nterest-based votng. 12 A few large partes are also dffcult to classfy because they are really cross-class polces. The most notable example s the German Chrstan Democratc Unon (CDU/CSU), whch s classfed as a center party (n lne wth Manow and Van Kersbergen 2009).

21 19 electorate nto the three groups s farly even: 31 percent for the left, 39 percent for the center, and 30 percent for the rght. Our classfcaton devates somewhat from one devsed by the ISSP s survey, wth a correlaton of.91. As a robustness test we ran our regresson on the ISSP party choce varable, and the substantve results are nearly dentcal (the results for the ISSP codes are reported n Appendx F). 13 The results dscussed n the body of the paper are based on the party classfcaton n Appendx E. The estmate of ndvduals left-rght poston, y, s arrved at through the followng procedure. We frst regressed party choce on a large set of economc predctors for each country separately, usng standard ordered probt. 14 The full set of economc explanatory varables s: famly ncome (standardzed nto nne natonally-specfc quantles), educaton, supervsory poston, self-employment, unemployment, occupatonal unemployment, part-tme employment, publc sector employment, skll-specfcty, retrement status, gender, rural resdence, and an ndcator varable for not beng n the labor market. These varables reflect a varety of standard poltcal economy arguments that need not be recounted here. Agan, we are treatng economc nterests as gven and seek nstead to understand the effects of ncentves to be nformed about them. 13 We collapsed the fve categores on the ISSP scale nto three, whch yelds fewer centrst votes and more left votes compared to our classfcaton (25 and 38 percent, respectvely). ISSP do not use party choce to classfy Span and New Zealand, where we retaned our classfcaton. 14 Ths procedure s n effect the frst step n a 2-Step MLE procedure whch mples consstency (Wooldrdge 2002, 414).

22 In addton to these economc varables we follow Rueda (2012) and nclude a measure of altrusm, whch s based on a survey queston about the respondent s vews on what t takes to be a good ctzen. One opton s to help people [n the respondent s country] who are worse off than yourself, wth respondents beng able to ndcate one of seven degrees of mportance, rangng from not at all mportant to very mportant. As one mght expect, the measure s heavly left-skewed wth about one thrd sayng very mportant, 25 percent ndcatng the next hghest level of mportance, and only 1.3 percent sayng not at all mportant. Answers undoubtedly reflect a healthy dose of cheap talk, but lke Rueda we fnd that they are n fact assocated wth dstnct votng patterns (controllng for the economc varables above), so they are not meanngless. If we allow people to be altrustc, then poltcal dscusson should have exactly the same effect as shftng people to the left when they have low ncome, hgh rsk, and so on. Each natonal y-varable s calculated as a lnear combnaton of altrusm and the economc varables wth weghts determned by the regresson coeffcents. 15 Fnally, we standardzed each natonal measure by dvdng the populaton nto decles where a value of 1 means that an ndvdual s n the bottom decle of the y-dstrbuton and a value of 10 means that he or she s n the top decle. Ths s to ensure that the left-rght measure has the same nterpretaton across countres. As we wll see, t also makes t easy to translate predcted probabltes nto shares of the populaton who support dfferent types of partes Because there are mssng observatons on every varable, we dropped regressors wth p-values less than 0.2, startng wth the least sgnfcant, then the next, untl all retaned varables were sgnfcant at a 0.2 level.

23 Results Table 1 shows the regresson results. The frst column s the predcted probablty of supportng the left, center and rght, solely as a functon of people s left-rght poston. The results from ths regresson are llustrated n Fgure 2. The frst thng to note s that standard poltcal economy varables do a reasonably good job at predctng left and rght votng, mrrorng the fndngs n Rehm (2011) and n Cusack et al (2006). For those n the lowest decle of the y dstrbuton the probablty of supportng the left s almost 50 percent, whle the probablty of supportng the rght s less than 20 percent. For those n the hghest decle the pattern s reversed and of smlar magntude. Gven that the support for the left and rght s about even n our sample (31 versus 30 percent), ths symmetry s unsurprsng. Keep n mnd that the overall support for dfferent partes s dependent on the exact strategc postonng of partes, and we only know ther rankng. [Table 1 and Fgure 2 about here] A notable aspect of the pattern n Fgure 2 s that votng for center partes s hgh throughout the left-rght space, fallng only slghtly below 35 percent on the extremes. In our model the explanaton of ths puzzle s that even voters wth extreme nferred left-rght postons frequently vote center because they vew centrst partes as safe havens whenever they are uncertan about how ther welfare wll be affected by dfferent partes. Loss-averse ndvduals do not lke to take chances wth extreme partes unless they are confdent that the polces are rght for them. The mplcaton of ths logc s that a large number of voters wll shft ther support away from centrst partes as ther poltcal nformaton rses. Ths s not because nformed voters are less loss-averse than unnformed voters, but because they have less reason to worry about noncentrst partes leadng to faled polces. Ths expectaton s powerfully confrmed n Fgure 3. It

24 22 shows the estmated probabltes of supportng dfferent partes contngent on the level of poltcal dscusson (detaled results are n Appendx E). The drop n support for the center s remarkable, declnng from nearly 50 percent across the y-scale when dscusson s low to less than 30 percent when dscusson s hgh. [Fgure 3 about here] Conversely, the relatonshp between y and left and rght votng gets progressvely stronger, resemblng a par of openng scssors. Ths polarzng effect of poltcal dscusson stands n sharp contrast to Habermas-nspred arguments about delberaton as a source of poltcal moderaton. In the real world, those who dscuss poltcs tend to be strong partsans. 16 An equally clear result s that the effect of dscusson s much hgher on the left than on the rght. For someone n the lowest decle of the dstrbuton, gong from never dscussng poltcs to often dscussng poltcs nearly doubles the probablty of votng left from 32 to 61 percent, whereas the probablty of votng rght for someone n the top decle only ncreases from 41 to 51 percent. Ths dfferental effect of dscusson s exactly as predcted by the model, f 16 It s concevable that people frst become partsan and then begn to dscuss poltcs. But ths cannot explan why dscusson has dfferent effects on the left and rght; a pattern we document next. Also, f dscusson s caused by partsanshp, party choce must be the result of somethng other than left-rght poston snce we are comparng people wth the same l-r poston (but wth dfferent levels of dscusson). Ths then has the strange mplcaton that whle people become partsan for reasons that are unrelated to ther left-rght poston, dscusson makes left-rght poston an ncreasngly strong predctor of ther party choce. The much smpler and plausble story s that dscusson yelds useful nformaton about who to vote for.

25 23 prvate nformaton s ndeed rsng n y. At low levels of dscusson many on the rght have prvate ncentves to be nformed. Another notable aspect of Fgure 3 s the rght-shftng locaton of the pont of ntersecton between the support curves for the left and rght. To the left of ths pont the probablty of supportng the left exceeds the probablty of supportng the rght, whle the opposte s true to the rght of ths pont. Because the y-varable s measured n decles the projecton of the pont of ntersecton onto the x-axs (ndcated wth a vertcal lne) convenently shows the share of the electorate whose probablty of votng left s greater than the probablty of votng rght. At the lowest level of dscusson ths share s 40 percent. As dscusson rses, the pont moves to the rght, and at the hghest level of dscusson the left dsadvantage has turned nto an advantage wth more than 60 percent supportng the left. Poltcal dscusson thus unambguously benefts the left more than the rght. Ths reflects the lack of prvate ncentves for those on lower ncomes. But t also partly reflects that those wth a hgh sense of altrusm are more lkely to vote left when poltcal dscusson ncreases (as we should expect). On average the two blocks end up wth s about the same share of the total vote. Agan, the aggregate shares of the vote depend on relatve party postons, whch we do not have data on. But the relatonshp between dscusson and left and rght advantage s precsely as expected from the model. At low levels of dscusson the left s at a dsadvantage because, unlke the rght, few n ther consttuences have prvate ncentves to be nformed. Agan, we do not have drect evdence that hgh-y voters have greater prvate ncentves to be nformed, but ndrect evdence can be gleaned from ILO s standard classfcaton of occupatons (ISCO-88). For some occupatons we have strong prors that publc polces would affect the optmal decsons people make about the future. A manager makng hrng decsons, or nvestng n new equpment, would clearly beneft from nformaton about potental changes

26 n job protecton legslaton, rules about tax deprecaton, mnmum wage regulatons, and so on. Ths s true for a range of occupatons and ndustres, even though the exact reasons dffer. A company that reles on publc procurement, for example, would want to know whether the polcy n the affected area s lkely to contnue. Surely legslators have prvate ncentves to be poltcally nformed! We thnk t s uncontroversal to conjecture that they are partcularly strong n the followng range of occupatons: 24 ISCO-88 class: Descrpton Legslators and senor offcals Corporate managers General managers Busness professonals (accountants, personnel and career professonals, others) Legal professonals (lawyers, judges, others) Safety and qualty nspectors Insurance representatves Estate agents Apprasers Busness servces agents and trade brokers Customs, tax and related government assocate professonals If we assgn these occupatons a value of 1, and all others a value of 0, t s hardly surprsng that the varable exhbts a strong assocaton wth our left-rght ndcator (although ths does not nclude drect nformaton about occupaton). Thus, whereas someone n the bottom decle of the y-dstrbuton has less than fve percent probablty of beng n one of these occupatons (based on a smple probt regresson), someone the top decle has a greater than 30 percent probablty. The lkelhood of havng a hgh level of educaton also rses notably from low to hgh y- values, and whle educaton tself s not a measure of prvate ncentves, t almost certanly reduces the costs of actng on prvate ncentves to acqure nformaton. For those at the low end of the y-dstrbuton Martn Glens provded us wth some tellng, f ultmately nconclusve,

27 evdence from the 2004 Amercan Natonal Electon Study. Respondents were asked to place the two man poltcal partes and ther presdental canddates (Bush and Kerry) on a left-rght scale. Assumng the Republcan Party and Bush were to the rght of the Democratc Party and Kerry, only slghtly over half of those n the bottom quartle of ncome were able to rank them correctly (53 percent for partes and 56 percent for canddates). In the top quartle of ncome about three quarter of respondents were able to do so (77 percent for partes and 71 percent for canddates). Agan, ths does not mply that low ncome people are ll-nformed about publc polces that drectly affect them, such as food stamps and housng subsdes, but they have no prvate ncentves to lnk these polces to partes because such knowledge yelds no prvate benefts. The evdence for the mportance of poltcal dscusson s to our knowledge novel. Yet, t hghlghts an age-old understandng of the dfferent mportance attached by left and rght partes to the moblzaton of voters. As already argued by Duverger (1951), the left was hstorcally dfferent than the rght because the former depended more on actvatng the masses and breakng ther tradtonal allegance to economc and relgous patrons. Wth the event of democracy such moblzaton requred mass organzatons that could reach consttuences and foster a sense of common cause. In Przework and Sprague s (1988) nfluental formulaton, buldng a strong left necesstated the creaton of a workng class dentty, dstnct from the bland mddle-of-the-road alternatve of the center. And from the work of Huckfeld and Sprague (1991, 1995), and more recently Ahlqust and Lev (2012), we know that such denttes are constructed n part through poltcal dscusson and network nvolvement. Incentves for poltcal moblzaton are of course also present on the rght, but to a lesser extent. As descrbed by Duverger, the rse of class-mass partes produced a contagon from the left as rght partes began to mtate the left and buld stronger mass organzatons to compete outsde ther orgnal upper and upper mddle class consttuences. But class-mass partes 25

28 26 remaned a predomnantly left phenomenon wth membershp rates far outpacng those on the rght (Mar 1998). From the perspectve of our model, the greater mportance of class moblzaton to the left s explaned by the dfferent ncentves to acqure poltcal nformaton The role of unons In our data we can suggest the mportance of mass moblzaton wth reference to the changng role of unon. Over the past four decades party membershp has collapsed everywhere, and whle new forms of mass-meda communcaton and targeted campagnng have substtuted for tradtonal membershp organzatons (see Epsten 1967; Dalton et al. 1984; Norrs 2006), the declne of organzatonal lfe s a real concern for the left because t lkely reduces engagement n socal nteracton nvolvng poltcal dscusson. We cannot here examne the role of party membershp, n part because t s so low n our sample (less than four percent n average) and n part because membershp s tself a functon of partsanshp. But we can explore the role of unons. Most workers jon unons for economc reasons (more on ths below), and membershp n unons s stll farly common n our sample (22 percent n average). Followng the logc of the model, unons can exert an nfluence on votng n two man ways. Frst, they can nfluence the level and dstrbuton of poltcal dscusson by engagng members n poltcal dscusson. Second, they can nfluence the content of poltcal dscusson by presentng focal ponts for collectve nterpretatons of economc nterests (the group nterest as defned n secton 2.5). Both logcs are succnctly captured n a new study on unons by Ahlqust and Lev (2012). Reportng results from ntervews wth unon leaders and members, they descrbe how unons can create a socal foundaton for the poltcal belefs of members. Referrng to the Longshore and Warehouse Unon they observe that [The hrng hall] provded the rare place where [workers] could get warm, drnk coffee, read the newspaper, and nteract off the job wth other workers and unon offcers. They

29 could talk poltcs and rase questons about the unon and ther offcers. [The] clear sense of an occupatonal communty was enhanced by the mportance of famly and neghborhood connectons among workers, common socal actvtes, debate and dscusson... To estmate the effect of unons on the level and dstrbuton of poltcal dscusson we regressed poltcal dscusson on unon membershp (n addton to y, the prvate nformaton dummy varable, and fxed effects). It turns out that the probablty of frequent dscusson ( sometmes or often ) s about 10 percent hgher among unon members (about 60 percent n average) than among non-unon members (about 50 percent n average). Ths s a much stronger effect than that of y on dscusson, so unonzaton appears to be an effectve way to overcome lack of poltcal engagement among relatvely low-ncome groups. But an even more mportant role of unons s to cultvate a common understandng of the collectve nterest (also emphaszed n Ahlqust and Lev). In our model, f ndvdual members assume that there s a correspondence between the group nterest and the ndvdual nterest, and f the group nterest s observable wthout engagng n dscusson, unons can exert an effect even among those who do not actvely engage n dscusson. As long as sharng the common dentty wth others s a source of socal approval, followng the poltcal preferences of the unon leadershp s optmal (see secton 2.5). Hence, unons can have a sgnfcant effect on the preferences of members whose underlyng y-poston s away from the medan. We test ths possblty by ncludng unon membershp n our nonlnear model and explorng how t shapes votng preferences jontly wth left-rght poston at dfferent levels of dscusson. Fgure 4 shows the estmated votng probabltes among members compared to nonmembers. Note that unon membershp notably ncreases the proporton of the electorate where the left s preferred to the rght. Among those who do not dscuss poltcs, the share of the 27