Financing Direct Democracy: Revisiting the Research on Campaign Spending and Citizen Initiatives

Transcription

1 Duke Law Duke Law Scholarshp Repostory Workng Papers 2010 Fnancng Drect Democracy: Revstng the Research on Campagn Spendng and Ctzen Intatves John de Fgueredo Duke Law School, Chang Ho J Thad Kousser Repostory Ctaton de Fgueredo, John; J, Chang Ho; and Kousser, Thad, "Fnancng Drect Democracy: Revstng the Research on Campagn Spendng and Ctzen Intatves" (2010). Duke Law Workng Papers. Paper Ths Artcle s brought to you for free and open access by Duke Law Scholarshp Repostory. It has been accepted for ncluson n Workng Papers by an authorzed admnstrator of Duke Law Scholarshp Repostory. For more nformaton, please contact dunshee@law.duke.edu.

2 Fnancng Drect Democracy: Revstng the Research on Campagn Spendng and Ctzen Intatves John M. de Fgueredo UCLA Chang Ho J La Serra Unversty Thad Kousser UCSD 08 June 2010 Abstract The conventonal vew n the drect democracy lterature s that spendng aganst a measure s more effectve than spendng n favor of a measure, but the emprcal results underlyng ths concluson have been questoned by recent research. We argue that the conventonal fndng s drven by the endogenous nature of campagn spendng: ntatve proponents spend more when ther ballot measure s lkely to fal. We address ths endogenety by usng an nstrumental varables approach to analyze a comprehensve dataset of ballot propostons n Calforna from 1976 to We fnd that both support and opposton spendng on ctzen ntatves have strong, statstcally sgnfcant, and countervalng effects. We confrm ths fndng by lookng at tme seres data from early pollng on a subset of these measures. Both analyses show that spendng n favor of ctzen ntatves substantally ncreases ther chances of passage, just as opposton spendng decreases ths lkelhood. We would lke to thank semnar partcpants at Stanford Unversty, Duke Unversty, and UCLA for helpful comments. Please send correspondence to de Fgueredo, UCLA Anderson School of Management, 110 Westwood Plaza #D508, Los Angeles, CA , jdefg@anderson.ucla.edu.

3 To what extent does campagn spendng shape the electoral fate of ctzen ntatves? Ths emprcal queston has mportant mplcatons for the normatve debate over the role of money n drect democracy electons (Gerber 1999, Broder 2000, Matsusaka 2004). If groups wth a concentrated nterest n publc polcy can use ther organzatonal advantages to rase large campagn war chests that translate nto votes for the ntatves that they propose, drect democracy may not smply amplfy the voce of the people as ts early advocates promsed (see Grantham, 1983, Ells 2002, Goebel 2002, Pott 2003). By contrast, f spendng n favor of ntatves does not ncrease ther chances of passage, powerful nterests wll not gan an advantage through the money that they can rase. Whle ths mportant queston has generated much scholarshp from poltcal scentsts, economsts, and legal scholars, two dstnct streams of the lterature have provded dfferent answers to t. A long seres of works that treat campagn spendng levels as exogenous created a conventonal wsdom that expendtures have asymmetrc effects: Money spent n opposton to ballot propostons (ncludng both ctzen ntatves and measures placed on the ballot by legslatures) lowers ther expected vote shares or chances of passage, whle spendng n favor of them s drastcally less effectve (Allen 1979, Lydenberg 1979, Mastro et al. 1980, Lowensten 1982, Magleby 1984, Schockley 1980, 1985, Schmdt 1989, J 1998, and Bowler and Donovan 1998). A recent revew of the lterature reports that these studes found that heavy spendng aganst a measure tended to lead to the measure s defeat, whereas heavy spendng n favor of a measure had a mnmal mpact (Lupa and Matsusaka 2004, p. 470). Though only a few studes 1

4 have probed why spendng mght have ths asymmetrc effect 1 or why ballot proponents spend so much even f ther money appears to be wasted, 2 many have ponted to the reassurng normatve mplcatons of ths emprcal fndng. Such studes mply a role for money n drect democracy that cuts aganst popular stereotypes. [T]he deep pockets of busness groups do not allow them to buy favorable polcy (Lupa and Matsusaka 2004, pp ). At the same tme, the asymmetrc effect of campagn spendng may enable such groups to defend a status quo polcy aganst an energetc challenge. Another stream of scholarshp (Zsk 1987, Gerber 1999, Garrett and Gerber 2001) has found, often by explctly consderng that spendng levels are endogenous to the nature of a proposton and ts backers (Banducc 1998, Stratmann 2005, 2006), that the effects of pro and con spendng can be symmetrc under certan poltcal condtons. 3 These works have found that both proponents and opponents of a ballot measure can wn more votes when they spend more money suggestng a dfferent vew of money and drect democracy. If a well-organzed nterest or busness group opens ts deep pockets for a ballot campagn, t may ndeed be able to buy support for a favorable outcome. Balancng ths worry s the assurance that groups seekng to defend exstng polcy have no predetermned advantage n a campagn aganst those who wsh to overturn the status quo. 1 Lowensten (1982, pp ) examnes possble ratonales for ths asymmetrc effect, rejectng explanatons based on voter confuson or a predsposton to vote aganst propostons, and nstead postng that the knd of ballot measures that attract hgh levels of support spendng are lkely to be less popular wth voters from the outset. Ths lne of reasonng serves as the bass for our two-stage model, whch confrms Lowensten s conjecture. 2 Gerber s (1999, p. 82) survey of 156 nterest groups actve n proposton campagns provdes some potental solutons to ths puzzle: Group representatves often contended that they backed ballot measures more to send sgnals to legslatures, draw attenton to ther agendas, or respond to ther members than to make polcy drectly. Note that Gerber s nvestgaton focused specfcally on the nablty of busness groups to buy support for propostons that promsed a narrow economc beneft, whle expendtures by ctzen groups appeared to be more nfluental. 3 Garrett and Gerber (2001) fnd that support spendng can be effectve n the absence of opposton spendng, Gerber (1999) shows that spendng by ctzen groups rather than by busness can boost a proposton s electoral fortunes, and Zsk (1987) fnds symmetrc effects for support and opposton spendng n approxmately half of all cases. 2

5 We seek to address ths consequental controversy n the lterature by analyzng a comprehensve dataset of ballot propostons n Calforna 4 from 1976 through 2004 and applyng a novel approach to dealng wth the endogenety of spendng levels, solatng ctzen ntatves from other types of propostons, and addressng other estmaton ssues. We argue that the much of the exstng lterature underestmates the effect of spendng n favor of ctzen ntatves because narrow nterests often spend large sums to pass measures that begn campagns wth slm majortes, whle proponents of measures wth a wde appeal can spend less and stll wn. Borrowng nsghts and technques from earler studes as well as from the lterature on the effects of spendng n canddate electons, we hghlght the crtcal mportance of estmaton choces. We begn by takng the tradtonal approach to estmatng the effect of spendng on ballot measures, whch produces the asymmetrc fndng. When we treat campagn spendng as exogenous and assume that t has constant returns, t appears that opposton campagns hurt ballot measures whle spendng n favor of them s wasted. Ths puzzlng asymmetry dsappears, however, when we use more flexble econometrc technques that allow for dmnshng returns and consder how the concentraton of costs and benefts can determne fnancal support for a polcy proposal (Low 1964, Stgler 1971, Peltzman 1976, Wlson 1980). Just as expendtures n opposton to ctzen ntatves dmnsh ther support, campagn spendng n favor of ntatves appears to have a postve and sgnfcant effect on ther chances at the ballot box. Accordng to our fnal model, spendng an addtonal $100,000 n support of an ntatve ncreases ts chances of wnnng by a predcted 1.43 percent, whch s almost as large as the 1.90 percent predcted decrease n the lkelhood of passage that would 4 We focus only on Calforna ballot measures because they serve as the dataset for nearly all of the exstng lterature (see Lee 1978, Lowensten 1982, Magleby 1984, Myers 1988, Prce 1988, Bowler and Donovan 1998, Gerber and Garrett 2001, Stratmann 2006) and because a hstorcal archve descrbng ther polcy content s avalable (Hastngs Law Lbrary, 2005). See Smth (2001), Lupa and Matsusaka (2004), and Smth and Tolbert (2007) for revews of the lterature on money and propostons across all states. 3

6 result from an addtonal $100,000 of opposton spendng. We supplement our nstrumental varables approach wth a tme seres analyss of a more lmted set of ntatves, showng that both pro and con expendtures appear to shft the level of popular support for ballot measures over the course of a campagn. In the end, our analyss provdes more evdence n favor of newer works challengng the conventonal wsdom n ths area by showng that nterest groups, for better or for worse, can boost the chances of passng an ntatve that they favor just as effectvely as opponents can spend to stop them. I. Reasons to Doubt the Conventonal Wsdom Whle earler studes compared the effect of spendng patterns on the success rates of ballot measures through bvarate studes (Lee 1978, Allen 1979, Lydenberg 1979, Lowensten 1982, Owens and Wade 1986), recent works use multvarate models to predct the fate of propostons by the levels of spendng for and aganst them, along wth other control factors (J 1998, Bowler and Donovan 1998, Gerber 1999). Table 1 presents the most basc multvarate approach wth our data. It reports the results of a probt model explanng the success or falure of 329 propostons n Calforna from 1976 to The coeffcents on these varables can be gven a causal nterpretaton only f pro and con spendng levels have been exogenously and randomly assgned to propostons. We also nclude controls for the type of electon, the type of proposton, the presence of a competng ntatve, the number of other measures on the ballot, and economc condtons. We coded all dchotomous varables that could be posed as questons wth a 1 ndcatng an answer of yes and a 0 ndcatng no. We estmate robust standard errors, and nclude year fxed effects to account for the possblty that fscal or electoral condtons make one year a good or bad tme for all ntatves on the ballot. 4

7 [Table 1 About Here] The results n Table 1 parallel the conventonal fndng n the lterature. Money spent aganst a proposton has a negatve and strongly sgnfcant effect on ts chances of passage. An addtonal $1 mllon n opposton spendng leads to an estmated 4.1% decrease n the probablty that t wll pass. By contrast, the mpact of spendng n favor of a ballot measure s substantvely mnuscule and falls far short of statstcal sgnfcance. Ths basc model, whch follows the approach of many prevous studes, reproduces the puzzlng conventonal wsdom that support and opposton spendng have asymmetrc effects on the outcomes of drect democracy electons. 5 Instead of acceptng ths counterntutve result, we explore the substantve methodologcal assumptons upon whch t rests. Relaxng each assumpton has a clear econometrc mplcaton, and we explore the consequences of these estmaton choces later n our emprcal fndngs secton. In ths secton, we focus on our conceptual objectons to the basc model reported n Table 1, provdng data and referencng other works that support our crtques. A frst troublng feature of the basc model s that t treats every proposton that appears on a ballot as the same type of phenomenon, gnorng key dfferences n the path that a proposal took to reach the ballot. We wsh to draw a dstncton between measures put on the ballot by legslatures (the three 6 types of measures that are commonly referred to as legslatve propostons ) and drect ctzen ntatves. Legslatve propostons are less controversal, much less lkely to nspre spendng on ether sde, and more lkely to wn than ntatves qualfed by voters and nterest groups. The reasons for ths le n the rules that govern access to the ballot 5 In addton to the works that have nformed the conventonal wsdom, another stream of research fnds that nether sde can spend effectvely n proposton campagns: Lee (1978), Owens and Wade (1986), Myers (1988), and Prce (1988) all conclude that nether opposton nor support spendng plays a central role n determnng the fate of propostons. 6 Legslatve propostons nclude consttutonal amendments proposed by the legslature, legslatve bonds, and legslatve proposals to alter statutes that have been put n place by the ctzens through ntatves (Legslatve Analyst s Offce 2009). 5

8 and n the realtes of Calforna poltcs. For an dea to reach the ballot through the legslatve process, t must wn two-thrds votes n both the Senate and the Assembly. Nether party ever held such a supermajorty durng our perod of study, meanng that some level of bpartsan agreement was necessary to send every one of these legslatve propostons to the voters (Natonal Conference of State Legslatures 2002, State Net 2005). Because of ths, the measures were often deas wth broad support that dd not generate ntense controversy or massve campagn spendng. Intatves, by contrast, need not reflect the vews of a wde cross-secton of Calforna offceholders or voters. What they requre s the ntense support of a group that s suffcently well organzed and well funded to collect a large number of voter sgnatures n a short amount of tme. To qualfy for the ballot, statutory ntatves must be sgned by the number of regstered voters that equals 5% of the votes cast n the last gubernatoral race, whle ntatve consttutonal amendments must be backed by 8% of that total. Wth only 150 days to collect these names, nearly all ntatve authors pay sgnature gatherng frms $1-3 per name to collect the requste number of vald names (Gerber et al. 2001, p. 10). If a group s wllng to pay more than a mllon dollars to put ts dea on the ballot, rather than spendng a smaller sum lobbyng the legslature to pass t, we can expect that ts proposal s opposed by a majorty of legslators n Sacramento (Lowensten 1982, Kousser and McCubbns 2005). We can also expect that these ctzen-backed ntatves are less lkely to draw wde support n the electorate than ther legslatvely-proposed counterparts. A look at the ballot measures n our sample demonstrates the dfferences between legslatve propostons and ntatves. The 192 legslatve propostons that reached the ballot from 1976 to 2004 attracted an average of $478,406 spent n support of them and $220,273 spent n opposton, n constant dollars. Much more money flowed nto ntatve campagns. 6

9 Mean spendng on the 137 ctzen-qualfed measures n our sample was $3.6 mllon for pro campagns and $2.4 mllon aganst them (agan adjusted for nflaton). Intatves also dd much worse at the polls than legslatve propostons. In our sample, 43% of ctzen-authored measures passed, whle 70% of the legslature s proposals were accepted by voters. Because they are so clearly dfferent from each other, we argue that these two types of propostons should not be combned nto the same analyss. Dong so would bas the estmated effects of spendng n favor of ballot measures toward zero, snce t ncludes a class of measures (legslatve propostons) that generate lttle support spendng and have hgh chances of passage because of the polcy proposals that they contan. Separatng them from the analyss of ctzen ntatves s a frst step toward breakng the endogenous lnk between the nature of a proposton and the spendng that t attracts, and thus solatng the treatment effect of campagn expendtures. Some prevous works have made the same dvson of ballot measures that we propose. Whle the more comprehensve multvarate analyses by J (1998), Gerber (1999), and Bowler and Donovan (1998) nclude both ntatves and legslatve propostons, many of the earler bvarate analyses such as Lowensten (1982), Shockley (1985), Prce (1988), Meyers (1988), and Schmdt (1989) looked only at ctzen-authored ntatves. Garret and Gerber (2001) study both types of ballot measures but nclude a dummy varable ndcatng the type of proposton n ther models. Because t s the ctzen ntatve that most commentators are referrng to when dscussng the mpact of money on drect democracy, analyses of ths class of propostons wll be most relevant to the normatve debate. We focus only on ntatves n the rest of our analyss. 7 7 Because they are also ntated by ctzens, are often controversal, and frequently generate expensve campagns, we nclude the eght referendums on the ballot durng our perod of study nto the ntatve category. These are attempts to overturn laws passed by the legslature whch qualfy for the ballot n the same manner as ntatves. 7

10 Our second objecton to the conventonal approach s that t assumes that spendng wll have a constant, lnear effect on electon outcomes. Under ths reasonng, f the Indan trbes that spent $66 mllon n support of Proposton 5 n 1998 (Calforna Secretary of State, 1999) had spent a $67 th mllon on t, ths would have gven ther ntatve the same boost that the second mllon dollar would have provded to a less lavshly funded campagn. Ths strans common sense, and also fals to ft wth the broader lterature on the role of money n Amercan poltcs. Studes of the effects of campagn spendng n Congressonal canddate races (Green and Krasno, 1988), Senate contests (Gerber, 1998) and state legslatve races (Kousser and LaRaja, 2002) have found that hgh spendng brngs dmnshng margnal returns. Ths makes sense, because havng enough money to attan a basc level of recognton and support for a cause would seem to be more mportant than saturatng the arwaves wth more and more commercals. Rather than assumng lnear effects, these works test alternatve specfcatons and often fnd that nonlnear transformatons of expendtures provde a better ft of the data. In our emprcal secton, we follow a smlar strategy, examnng whether the square root of spendng provdes a better predcton of electoral outcomes than raw expendtures, to allow for a dmnshng margnal value of money. (In Secton III, we consder log transformatons as well.) At the same tme that t provdes a better ft for what we know about money and campagns, ths approach also gves less weght to outlyng ntatves that generated enormous spendng levels. Whle 88% of the ntatve campagns n our sample spent less than $5 mllon ether for or aganst an ntatve, 2% of them spent $20 mllon or more (n dollars). A lnear specfcaton would gve equal weght to the money spent n all ntatves, allowng the outlers to drve much of the results. Snce the very hghest expendtures n our dataset are often made by 8

11 support campagns, ths could dsproportonately bas the estmated effect of pro spendng toward zero. Our transformaton of spendng levels wll avod ths ptfall. Our fnal 8 reason to doubt the fndngs from our basc model and much of the exstng lterature s perhaps the most serous one: Campagn spendng on an ntatve may be endogenous to ts chances of passage. Most exstng works treat spendng for or aganst a measure as f t were randomly assgned to the ntatve and uncorrelated wth characterstcs such as ts prospects early n the campagn. The lterature on the role of campagn spendng n canddate races hghlghts the problems created by such an assumpton, and suggests one possble soluton. Ever snce Jacobson s (1980) nfluental Money n Congressonal Electons, scholars have recognzed that contrbutons to US Congressonal nomnees are lkely to be endogenous to ther chances of wnnng. Strong canddates, the reasonng goes, do well n electons and attract lots of money, creatng a correlaton between fundrasng prowess and electoral performance when there may be no causal lnk. Donors may smply be heapng funds on those who wll wn anyway, perhaps because they wsh to nfluence the members once n offce. We post that campagn spendng, especally money that goes to support ntatves, s also endogenous to ther electoral chances but n just the opposte way. Money may flow to ntatves that are n danger of falng. Snce contrbutors cannot hope to buy nfluence over the unchangng language of a ballot measure, there s no reason to ple money on one that s already headed for passage. Yet f an ntatve looks vulnerable, ts supporters especally those who have already 8 Another potental mprovement to the conventonal model would separate out money spent gatherng sgnatures to qualfy ntatves for the ballot from the money spent on the campagns supportng them. Unfortunately, the Calforna Secretary of State s offce reports qualfcaton and campagn spendng separately for only 75 of the 137 ntatves n our sample. In an earler draft of ths paper, whch we wll make avalable to nterested readers, we used ths lmted data to estmate qualfcaton spendng for all 137 ntatves and analyzed the effects of only campagn spendng on ntatve passage. Ths dd not change the sgn or sgnfcance of any of our estmated spendng effects, and the models were generally robust to ths change. Consequently, we use combned qualfcaton and campagn spendng n all of the analyss here snce t s reported for all of the ntatves n our sample. Results are avalable from the authors (see also SSRN verson of paper). 9

12 nvested substantal sums to qualfy t for the ballot may contrbute large amounts to campagn for ts survval. If ths story s correct, consder ts mplcatons for the conventonal econometrc approach. The model n Table 1 found that spendng n favor of propostons has no effect on ther chances of wnnng. But ths could be an artfact of the fact that a proposton s unmeasured ntal chances of passage are omtted from ths model, negatvely correlated wth support spendng, and postvely correlated wth ts performance on electon day. If ths plausble chan of relatonshps holds true, t wll bas the estmated effect of support spendng downward toward zero (as long as t s uncorrelated wth other regressors). Money spent n favor of propostons may ndeed ncrease ther chances of passage, but the bas created by the endogenous nature of campagn expendtures may be hdng ts true effect. In order to uncover our estmate of the mpact of spendng on ctzen ntatves, we follow the approach of Jacobson and others by fndng nstruments that capture the unmeasured characterstcs of an ntatve that determne ts campagn fnances. Jacobson (1980, p. 140) used a canddate qualty measure based on poltcal bographes, along wth other exogenous varables, to predct fnancal support n the frst stage of a two-stage least squares model. Two prevous authors have employed nstrumental varables estmaton technques n ballot measure fnancng papers. In an often overlooked and rarely cted paper, Banducc (1998) s the frst author to recognze ths problem and attempts an nstrumental varables estmaton on only proponent spendng. Her two key nstruments are opponent spendng and whether there s busness regulaton. Whle the former s clearly endogenous tself and thus requres ts own nstruments, the latter seems a somewhat promsng start to the analyzng the problem the degree to whch busness nterests are affected. Even wth these weak nstruments, Banducc s able to 10

13 get results whch show that proponent spendng may have an effect n some cases on ballot measure outcomes. Stratmann (2005, 2006) also recognzes the endogenety problem n spendng. Stratmann examnes spendng on televson advertsng n fve major meda markets n all statewde ballot measures n Calforna from and ts effect on county level votng outcomes usng a fxed effects estmaton, and fnds that televson advertsng for and aganst affect vote outcomes. Ths novel and effectve research desgn may, by ts nature, face lmts to ts external valdty: The data set covers only 18 ntatves and fve meda markets. Moreover, televson advertsng accounts for only 38% of total group expendtures, leavng 62% of expendtures outsde the analyss (Stratmann 2006: 792). The man method that Stratmann (2005, 2006) uses to address endogenety the ncluson of county- and ballot-level fxed effects has also faced crtcsm. 9 Taken together, though, Banducc (1998) and Stratmann s (2005, 2006) works effectvely challenge the conventonal assumpton that spendng on propostons s exogenous, suggest the value of the nstrumental varables approach as an alternatve, and fnd that support and opposton spendng may have symmetrc effects. To buld on ther work, we analyze all Calforna ballot measures from 1976 to 2004 (329 n all), we examne all spendng (rather than only televson advertsng), and we use multple nstruments that vary across each observaton n the dataset to obtan better dentfcaton of the econometrcs. We gather new measures of how concentrated the 9 de Fgueredo (2005) provdes a lengthy and detaled explanaton of why fxed effects does not solve the endogenety problem for the ballot measures dataset. To summarze that crtque, f countes have dfferent underlyng preferences for dfferent ballot measures, then ballot fxed effects (whch control for dfferences n ballot measures) and separate county fxed effects (whch control for dfferences n countes) do not solve the endogenety problem. Rather, one needs ballot-county fxed effects. (See Mundlak 1963 for a further dscusson of ths ssue.) Unfortunately, the Stratmann research desgn does not allow for these controls. As a separate robustness check, Stratmann (2006) does report the use of (but few of the results from) an nstrumental varables estmaton wth three nstruments: the number of measures on the ballot, the number of undecded voters, and the prce of advertsng per voter. However, these nstruments are plagued wth precsely the same problems as the fxed effects measures they vary by county or by ballot measure (on the state level) but do not vary by county-ballot measure as would be requred of standard nstruments n a research desgn of ths type. 11

14 costs and benefts of an ntatve are as our major nstruments. Our codng system s descrbed more fully n the secton on data sources below, but s based on the logc set forth by Low (1964), Stgler (1971), Peltzman (1976), and Wlson (1980) and summarzed n Baron (1996, pp ). Prevous work by Donovan et al. (1998) apples ths same typology to Calforna ballot propostons. When a polcy brngs concentrated benefts to one group rather than spreadng ts advantages around to the entre populaton, ths group wll be able to solve ts collectve acton problem and organze a strong effort to push for the polcy. The same logc apples for the polcy s opponents, and the proposal s success wll often be determned by the way that the concentraton of ts effects structures the nature of the poltcal competton over t. Under ths theory, ntatves that would yeld large but narrow payoffs, such as expansons of trbal gamng or relaxatons of ndustry regulatons, can be expected to rase a lot of money. But we hypothesze that these are precsely the types of measures whch are lkely to begn ther campagns wth a low probablty of passage. They do not promse a broad beneft, and n some cases may mpose costs on a broad cross-secton of voters. Examples of ths nclude Proposton 31 n 2000, whch mposed lmts on nsurance clams, and Proposton 64 n 2004, whch mposed lmts on unfar labor practce suts. Voters were ntally wary of these lmts, wth only 26% and 21% of voters supportng them, respectvely, when the Feld Poll gauged publc opnon at the outset of ther campagns. Yet these types of ntatves promse a tangble beneft to concentrated ndustry groups. Ther poltcal consultants wll have more money at ther dsposal, but a tougher job to do. Our measures, then, wll capture the way that the ntal prospects of a proposton help to determne ts fnances. The scope and magntude of the costs and benefts of ntatves may help to predct spendng levels n a frst stage model, and should not be gnored. 12

15 In summary, our three reasons to doubt the conventonal fndng that ntatve supporter waste ther money are substantve objectons that have methodologcal mplcatons. We beleve that analyses of propostons should be based on a sample of comparable phenomena such as ctzen ntatves, rather than mxtures of legslatvely- and ctzen-proposed measures. Instead of assumng lnear effects, we should test whether spendng on ntatve campagns yelds constant or dmnshng returns. Fnally, the endogenous nature of ntatve spendng should be taken as serously as the endogenous nature of canddate spendng. In an onlne appendx, we present a smple model that formalzes our conceptual approach and motvates both our assumpton of dmnshng margnal returns to spendng and our use of nstrumental varables. II. Data Sources In order to estmate the effects that spendng for and aganst Calforna ballot propostons has had on ther chances of passage, we analyze data from offcal state sources reportng ther campagn fnances, the subject areas whch they addressed, ther potental polcy mpacts, and ther fnal votes. We cover the 1976 to 2004 perod, wth excepton of two ntatves and 37 legslatve propostons for whch no campagn fnance documents were avalable. 10 Ths gves us a total of 329 propostons, ncludng 137 ntatves qualfed by ctzen groups and 192 legslatve propostons. We used the 1976 through 2004 edtons of the Calforna Secretary of State s The Statement of the Vote and Supplement to the Statement of Vote to dstngush between the two routes to the ballot. We used these same sources to collect nformaton on the proposton s vote 10 In the Unversty of Calforna s lbrary system, we could not fnd spendng data on Propostons 1-4 n the 1979 Specal Electon, Propostons from the 1984 General Electon, Propostons and 52 from the 1986 Prmary Electon, Propostons from the 1986 General Electon, Propostons from the 1992 Prmary Electon, and Propostons 168, 169, 171, 173, and 174 from the 1993 Specal Electon. Snce only two of these were ctzen ntatves, ths randomly mssng data should not bas our man analyss, whch looks only at ntatves. 13

16 share, whether or not t was a counter ntatve to another measure on the same ballot, whether t was a bond, and whch type of electon, prmary or general, was beng contested. To dentfy the polcy area covered by the proposton, we used A Study of Calforna Ballot Measures (Calforna Secretary of State, 2002) and updated t for the 2003 and 2004 electons ourselves. Our spendng fgures come from standardzed reports publshed by Calforna s Secretary of State and Far Poltcal Practces Commsson. 11 All expendtures were deflated nto constant dollars usng the Calforna CPI (Calforna Department of Fnance, 2005). Our most complcated task was to gather new varables assessng whether ntatves would brng dffuse or concentrated costs and benefts, and how large they mght be. All three of the authors made these judgments ndependently after readng summares of the measures produced by the Calforna Legslatve Analyst and collected n the Hastngs Law Lbrary Calforna Ballot Measures Databases (Hastngs Law Lbrary, 2005). We dd not converse about any of our codng choces. For each proposton, we recorded two objectve measures whether ths was a ctzenproposed bond and whether t changed tax levels and sx subjectve, nterval varables. As noted below, four of these varables wll become the omtted nstruments used n the frst stage equaton and two wll be used n both the frst stage and second stage equatons. These sx came from measures of the magntude and the scope of the economc benefts and costs that the ntatve could provde to some sets of wnners and losers, compared wth the status quo. We defne wnners for an ntatve as those who would beneft from changng polcy from the status quo to the ntatve's proposal. The losers are those who would pay the costs of changng polcy from the 11 Dfferent reports cover dfferent years, so our sources vary from Calforna Far Poltcal Practces Commsson (1986) for all electons from 1976 to 1986, Calforna Far Poltcal Practces Commsson (1988, 1989) for the 1988 electons, Calforna Far Poltcal Practces Commsson (1991a,b) for the 1990 electons, Calforna Secretary of State (1993) for the 1992 General Electon, Calforna Secretary of State (1995) for the 1994 electons, Calforna Secretary of State (1997) for the 1996 electons, Calforna Secretary of State (1998, 1999) for the 1998 electons, and Calforna Secretary of State (2005) for the electons. 14

17 status quo to the ntatve's proposal. We estmated expected benefts for the wnner and costs for the losng group, over a fve year perod. The best way to descrbe our new measures s to gve examples of recent Calforna ntatves that fall nto each category of dfferent varables. In fact, we began wth the followng set of examples to gude our subjectve codng of the benefts that would go to the backers of an ntatve f t was successful, and used a parallel set of measures to code the costs that would be pad by those who would lose f the ntatve passed. Magntude of Economc Benefts (four ascendng levels): 0 f there s no economc effect on anyone or f the effects are neglgble. For nstance, the ban on exportaton on horses for slaughter contaned n 1998 s Prop. 6 had lttle economc effect on any human, except for perhaps a neglgble boost for those sellng glue made from a dfferent technology. 1 for a gan of thousands or hundreds of thousands of dollars. Ths s the total that a small number of lawyers mght expect to gan after 2004 s Prop. 59 changed publc records law to make t easer to brng legal cases. 2 for a gan of mllons or tens of mllons of dollars. Calforna busnesses probably expected to save about ths much by avodng unjustfed lawsuts under the lmt on unfar busness practce suts mposed by 2004 s Prop for a gan n the hundreds of mllons or bllons. Those who proft from Calforna s trbal gamng casnos and those who wanted to end the trbes monopoly on Las Vegas-style gamng may have expected to gan benefts on ths massve scale from 2004 s unsuccessful Propostons 68 and 70. The county and cty governments whch wll beneft from the protecton of ther revenues guaranteed by 2004 s Prop. 1A probably expect to gan about ths much over a fve year perod, snce state budgets have occasonally taken back more than $1 bllon normally allotted to local governments. These varables wll be two of the nstruments omtted from the frst stage estmaton. 15

18 Scope of Effect on Industres (three unordered categores). We code an ntatve as brngng a Concentrated Industry Beneft only f an ntatve prmarly benefts a concentrated group of 1-3 ndustres, smlar to Prop. 70 s concentrated effect on the gamng ndustry. An ntatve brngs a Dffuse Industry Beneft f an ntatve a broad range of ndustres, smlar to 2004 s Prop. 64 s lmt on lawsuts. An ntatve brngs No Industry Benefts when t s a measure lke 2004 s Prop. 62, whch contaned a change n prmary electon law that promsed no clear effect on any ndustry. Note that ths extends the deas of Banducc s (1998) busness regulaton nstrument. These varables wll be the two addtonal nstruments omtted from the frst stage. Scope of Benefts for Ctzens (four ascendng levels): 0 f t benefts a concentrated group of a thousand or fewer resdents. 1 f t benefts a group of Calfornans n the tens of thousands, such as the 29,000 or so prsoners who would get early release under 2004 s Prop. 66, whch altered the state s Three Strkes sentencng law. 2 f t benefts a class of Calfornans n the hundreds of thousands or mllons, such as the one mllon unnsured resdents whose employers would have been forced to contrbute to ther health nsurance under 2004 s Prop f t benefts all Calfornans equally, as Prop. 59 s expanson of open records dd. Because benefts (costs) to ctzens are lkely to affect both the ablty of proposton proponents (opponents) to rase money as well as the ballot outcome drectly through ctzen votng, these varables are ncluded n both the frst stage and second stage equatons. Whle many publshed papers rely on a sngle coder for more subjectve varables, we ncluded three ndependent codngs for each varable for each ntatve. After each of us coded each ntatve, we created a combned measure by takng the medan of the three scores. Consequently, f our judgment was unanmous or f two out of three of us agreed on a score, the 16

19 consensus score became the fnal measure. Our subjectve codng correlates qute hghly for fve of these sx measures. We conduct a Cronbach s alpha test for the nter-rater relablty of the codngs. Each of the sx varables, except one, has a scale relablty coeffcent of at least.62 and as hgh as.86. One outler (Scope of Costs for Ctzens) has a relablty of.41. Ths means that t s a nosy measure of ctzen cost scope. We thnk t prudent to keep ths varable, despte the nose, because t s lkely that a broad scope of a measure wll have an mpact on vote outcomes. That sad, when we have ndependent coders assess ths same measure (as descrbed n the robustness secton), the alpha test has a relablty of Thus, there does seem to be some mert to ncludng ths varable. A separate measure of Spearman correlaton coeffcents, whch s sometmes preferred n the lterature, reflects the relablty found n the Cronbach s alpha test. Whle there s not a consstent standard for levels on statstcal congruence of nter-rater relablty measures, the.60 to.80 range s seems to be the wdely accepted range n the lterature for acceptable relablty of codng of ths type (Nunnally 1978; Murphy and Davdshofer 2001). Thus for at least fve of the sx varables coded, there seems to be relatvely good consstency across coders and our combned scores worked well as nstruments. As a fnal method for the reader to gan confdence n our codng, we present a sample of ntatves n Appendx Table A1 wth the results of our codng system for nspecton. III. Emprcal Fndngs A. Results Usng ths new data, we can reevaluate the conventonal fndng reached by many prevous studes and n the analyss reported n Table 1. Table 2 presents the full model whch reflects all 17

20 our substantve suggestons, yelds results stand n clear contrast to the conventonal wsdom about the nfluence of money n ntatve campagns. 12 [Insert Table 2 About Here] Model 2 n Table 2 ntegrates all of our objectons to the conventonal approach by takng serously the possble endogenety of spendng levels. Because contrbuton levels may be affected by an unmeasured ntal assessment of an ntatve s lkelhood of passage, spendng could be correlated wth the errors n our models through ths omtted varable. To purge spendng levels of ther lnks wth the error term, we follow the approach that models of canddate spendng take and estmate a two-stage model. The frst stage uses a set of nstruments our measures of the sze and concentraton of the benefts and costs brought by each ntatve, along wth other controls to predct spendng by both yes and no campagns. We analyze the square root of spendng levels n all of the followng models, ncorporatng our fndng of dmnshng returns. The results of the frst stage model, reported n Appendx Table A2, generally ft wth our expectatons. More money can be rased to support ntatves when they promse larger economc benefts or smaller economc costs, when the scope of ther effects on ctzens are concentrated, when they qualfy for general electon ballots, and when they would lock n polcy gans through a consttutonal amendment. Opposton forces spend more when the economc stakes are hghest, when ndustry costs are sgnfcant and concentrated, and when counter ntatves are on the ballot. One concern that arses s the qualty of the nstruments. We examne ths from both a conceptual and statstcal approach. Conceptually, the nstruments whch we use n the frst stage 12 In de Fgueredo, J, and Kousser (2007), we alter the conventonal model to ncorporate the econometrc mplcatons of our three objectons to the tradtonal approach, one by one. We report results at each step, makng the emprcal consequences of each of our alteratons clear. 18

21 and are omtted n the second stage are those nstruments whch relate to the costs, benefts, and structure of ndustry. The logc underlyng ths s that these ndustry-specfc factors are lkely to affect the amount of money a group can rase n support of or opposton to a gven ballot measure, but are unlkely to affect the vote outcome drectly. The effect operates through the money the ndustry rases. Contrast ths to the effect of the costs and benefts to ctzens. These mght affect both the amount of money rased from ctzens and the way n whch ctzens vote, so we nclude these varables n both stages of the analyss. Industres, unlke ctzens, cannot vote. Statstcally, we examne the qualty of the nstruments through a number of dfferent measures. In ths frst stage, the coeffcents have the expected sgn, and the t-statstcs on the coeffcents of the key nstruments are large, We subject the nstruments to further testng n conductng a Hausman specfcaton test for nstrumental varables. The Hausman test rejects consstency of the non-nstrumented model at the 95% level of confdence, and ponts to the twostage model as a better model for estmatng the effects of spendng on outcomes wth vald nstruments. That sad, t s also lkely the case that the collectve acton logc set forth by Low (1964), Stgler (1971), Peltzman (1976), and Wlson (1980) does not govern the fundrasng fate of every ntatve. Some grassroots groups have managed to rase large sums to back envronmental, poltcal reform, and anmal rghts ntatves, even when these measures would not brng a large fnancal beneft to a concentrated group. Many ntatve campagns n Calforna are funded by a wealthy donor or a poltcal patron usng the measure to help hs or her campagn, and the fundrasng success of these types of measures would also be poorly predcted n our model. Thus, though our predctons are not perfect n every case, the nstruments n the frst stage do help 19

22 us to more relably estmate the effects of money on ntatve campagns n our second stage model. Ths model, the nstrumental varables probt reported n Table 2, confrms our ntuton that both predcted support and predcted opposton spendng affect an ntatve s chances of passage. Note that the coeffcents are nearly dentcal n magntude, though opposte n sgn, and both of these measures are statstcally sgnfcant n ther expected drectons, ndcatng that the postve effects of support spendng are just as large as the negatve effects of opposton money. We translated these coeffcents nto predcted changes n an ntatve s probablty of passage brought by a shft n spendng, holdng all other varables n the model constant at ther mean values. These are nonlnear approxmatons for spendng and outcomes because we use the square root of spendng. Frst, we examned the potental effect of ncreasng support spendng from ts mean value of $3.6 mllon (n dollars). Boostng support spendng by $100,000 from ths level would brng an estmated 1.45 percent ncrease n an ntatve s chances of passage. Increasng the amount of opposton spendng from ts mean of $2.4 mllon (agan deflated to dollars) would have almost exactly the opposte effect. An addtonal $100,000 n spendng by the opposton campagn would lead to an estmated 1.93 percent decrease n an ntatve s chances. Both amounts are statstcally dfferent from zero, but not statstcally dfferent from each other. Ths fts wth our clear ntuton that expendtures on ether sde of the ssue should have equal and offsettng effects. It also stands n strkng opposton to our basc model, whch uses many conventonal technques to analyze exactly the same data. When we apply the lessons of the broader lterature on money and electons, we come down on the sde of the new stream of the lterature challengng the conventonal wsdom that the supporters of ntatves cannot spend ther way to vctory. 20

23 B. Robustness To explore the robustness of our regresson results, we examned a number of alternatve statstcal specfcatons. Frst f we replcate the nstrumental varables specfcaton ncludng qualfcaton spendng, the results are largely the same. The margnal value of pro-money decreases, whch s to be expected gven that the pro-money numbers are nflated by the amount of qualfcaton spendng. Although the scholarshp n ths area looks prmarly at a proposton s chances of passng, we also conducted parallel analyses of a measure s vote share. The results of these least squares regressons pont n the same drecton as our probt models, but do not acheve statstcal sgnfcance. When we estmate the conventonal model, support spendng appears to have a negatve, statstcally nsgnfcant effect on vote share, as n our prevous Model #1. In our fnal two-stage least squares model, ts coeffcent s postve (0.004) but only slghtly larger than ts standard error (0.0039). The dsjuncture between the probt and least squares results ndcates that support spendng has a greater mpact on shftng vote shares from 45% to 55% than t does on ncreasng support from 70% to 80%, as one would expect. In ths paper we focus on the frst boost gettng the ntatve from the loss column to the wn column, because ths s what most poltcans care about and scholars wrte about. Second, we pursue a number of other robustness checks. The results of these analyses are presented n Table Because many of the tests nvolve sub-samples of our 137 observatons, t s reasonable to expect the statstcal sgnfcance of the coeffcents to declne because of the lower power of the tests. In general, we avod tests nvolvng sub-samples of less than 50 observatons. 13 We present only the coeffcents for support and opposton spendng. However, all models are estmated usng the nstrumental varables procedures and all the varables (unless otherwse noted) found n Table 2. 21

24 The frst set of robustness tests break the sample by tme. Model 3 presents results for ntatves on the ballot through 1989, whle Model 4 presents results for the subsample of ntatves on the ballot from 1990 and later. The coeffcents on support and opposton spendng reman roughly symmetrc, and n three of the four cases, the coeffcents are statstcally sgnfcant at the 95% level of confdence. The second set of robustness tests slce the data by electon type general, off-year (no presdental electon), and on-year (presdental electon). 14 The general and the mdterm electons exhbt postve (negatve) coeffcents for support (opposton) spendng; n both Models 5 and 6, the coeffcents are statstcally sgnfcant at the 90% or 95% level. However, Model 7 suggests that nether support nor opposton spendng exert a statstcally sgnfcant effect n the natonal electon sub-sample.. We do not have a good answer for why ths s the case but speculate below. The thrd set of robustness analyss examnes the vote share (as a contnuous varable on [0, 100]) rather than vote outcome (a dchotomous varable [0,1]) as the dependent varable. Model 8, usng all 137 ctzen ballot propostons, generates the same result found n Table 1: opposton spendng s hghly effectve whle the effect of support spendng s not statstcally dfferent from zero. Ths result s not surprsng because there are many ntatves for whch the campagn s a slam dunk to begn wth, and others that have lttle chance of passng. To solve ths problem, we pursue two approaches. Frst, n the next secton of the paper, we examne pollng data to dscover how the ntal characterstcs of the ntatve affect money rased and ballot queston outcomes. Second, n the current dataset, we elmnate ntatves from the analyss where the outcome was greater or less than (35-65). It s these ntatves where money s lkely to have lower effect. In Model 9 we present the results for ths sub-sample of potentally 14 There are two few observatons to examne prmary electons. 22

25 compettve ntatves. Table 3 shows that when we lmt our analyss to these closer ntatves, there are statstcally sgnfcant, substantvely sgnfcant, and almost symmetrc effects of money on vote share. An addtonal $100,000 spent for (aganst) moves the vote share up (down) by about 0.9%. We conduct our fourth and ffth set of robustness analyses n Models 10 and 11. In Model 10, we nclude n our regressons changes n Calforna gross state product (GSP) as an addtonal economc ndcator varable to complement the Calforna unemployment varable. In Model 11, we remove qualfcaton spendng from the dependent varable. As Stratmann (2005, 2006) has argued, money spent by ntatve backers to gather the sgnatures necessary to qualfy ther measures for the ballot may have very lttle effect on ther eventual success at the ballot box. In both Models 10 and 11, the coeffcents on proponent spendng and opponent spendng are almost equal and opposte n sgn as well as statstcally sgnfcant. A sxth robustness check examnes the concern that because the authors coded the ntatves, there may be bas n the codng. To address ths, we also employed a graduate student and an undergraduate poltcal scence student to code the ballot ntatves usng the orgnal codng sheet wth three addtonal questons. Nether coder was aware of the ams of the project, nor had they read any draft of ths paper. Because of omssons n codng, four observatons were dropped leavng us wth 133 observatons. The results are presented n Model 12 of the paper. It shows that when we use these coders, the coeffcent on yes money s statstcally sgnfcant at the 90% level and the coeffcent on no money s statstcally sgnfcant at the 95% level. The effects of money on vote share are agan substantvely sgnfcant and almost symmetrc. A seventh concern that has been artculated s that the Economc Benefts varable has a somewhat odd scalng. To address ths concern, we used an alternatve scalng of Economc 23

26 Benefts that contaned fve, rather than four, categores and altered the cut-offs between them. 15 Because ths concern was artculated only late n the revson process and because we were afrad of nfusng author bas nto the codng, we had the ndependent coders code ths varable and then re-analyzed the data. The results are presented n Model 13 and are extremely smlar to Model 12, the model wth the ndependent coders and the orgnal codng of the varable. A thrd measure of Economc Benefts, whch smply asked coders to wrte down the best estmated n dollars of the economc beneft agan yelds smlar results (not reported here). Overall, the results of ths secton are smlar to the results n the Model 2. The exceptons are Models 7 and 8. We have already dscussed Model 8. Wth respect to Model 7, we speculate that natonal poltcal campagns overwhelm the proposton spendng; however, ths s merely speculaton. Overall, the results of these robustness checks suggest a result smlar to the man result of the paper n the prevous subsecton the effects of opposton and support spendng on ntatve outcomes are roughly symmetrc, wth the effect of opposton spendng somewhat larger than that of proponent spendng. IV. Supportng Evdence from Polls To ncrease our confdence n ths fndng, we gathered data from publc opnon polls taken at the begnnng of ntatve campagns and tested two addtonal mplcatons of our clams. Frst, f we are correct that spendng on ntatves s endogenous to ther ntal chances of passage, then we should observe a strong lnk between these early pollng fgures and subsequent spendng. Second, we can deal wth ths endogenety by takng advantage of the tme seres 15 In ths alternatve codng scheme, coders were asked to categorze an ntatve as a 0 f t promsed no or neglgble economc benefts, a 1 for an estmated gan between $1 and $500,000, a 2 for an estmated gan between $500,001 and $5,000,000, a 3 for an estmated gan between $5,000,001 and $100,000,000, and a 4 for an estmated gan greater than $100 mllon. 24