# Ideology and Interests in American Politics

Table of Contents Dedication ...................................................................................................................... ii

Acknowledgements ....................................................................................................... iii

Preface ........................................................................................................................... iv List of Figures ............................................................................................................... xi

List of Tables .............................................................................................................. xiii List of Appendices .......................................................................................................xv

Chapter 1: Ideology and Intersts in the Political Market Place ............................... 1

1. Introduction .............................................................................................. 2

2. Motivation ................................................................................................ 5

3. Recovering Ideological Positions from Campaign Finance Records .... 7

4. Statistical Model of Contribution Behavior......................................... 10

4.1. Estimation ...................................................................................... 15

5. Data ......................................................................................................... 16

6. Results from a Joint Scaling of the 1980-2008 Election Cycles ......... 19

6.1. Measuring the Ideological Consistency of PAC Contributions . 23

7. Assessing Model Fit ............................................................................... 26

7.1. External Validation ........................................................................ 27

7.2. Classification of Roll Call Vote Outcomes ................................... 32

7.3. Predictive Accuracy ....................................................................... 35

8. What Motivates PACs to Contribute? ............................................... 36 9. Conclusion .............................................................................................. 43 10. Appendices ............................................................................................ 44

ix

Chapter 2: Operationalizing the Spatial Model with an Interactive Survey on Public

Spending Preferences ............................................................................... 48

1. Introduction ............................................................................................ 49

2. Fiscal Preferences, Public Opinion and Representation ..................... 52

3. Recovering Ideal Points from Survey Data on Fiscal Preferences .... 55

4. Survey Implementation ......................................................................... 60

5. The Linkage between Self-reported Ideology and Fiscal Preferences 62

6. Scaling and Dimensional Analysis ........................................................ 68

7. Understanding the Rival/Non-Rival Spending Dimension ................ 72

8. The Contours of Fiscal Preferences ...................................................... 73 9. Conclusion .............................................................................................. 78 10. Appendices ............................................................................................ 80

Chapter 3: A Day by Day Measure of Legislative Ideology ................................... 82

1. Introduction ............................................................................................ 83 2. Smoothing Legislator Time Trend with Localized Kernel Estimates86

3. Classification Results from the U.S. Senate......................................... 94

4. Individual Legislator Trends ................................................................. 97 4.1. President Trends .......................................................................... 101

5. Is Senate Polarization a Result of Replacement or Adaptation ...... 104

x

6. Conclusion ............................................................................................ 112 7. Appendices ............................................................................................ 116

Bibliography .............................................................................................................. 121

xi

List of Figures Chapter 1 Figure 1.1: Frequency plot of contribution amounts during all primary or general elections from 1980 to 2008 ............................................................................. 9 Figure 1.2: Density plots of ideal point estimates for 3572 PACs and 3314 candidates from a joint scaling of the 1980-2008 election cycles. ............................ 20 Figure 1.3: Money-weighted standard deviations plotted against ideal point location ............................................................................................................... 24 Figure 1.4: Candidate CFscores plotted against DW-NOMINATE scores from a scaling of the 1980-2008 Election Cycles. ..................................................... 29 Figure 1.5: Correct classification rate in each Congress (1980-2008) ......... 34 Figure 1.6 Box and whisker plot of the estimated conditional means on the observed data for the 1980-2008 Election Cycles .............................................. 35 Chapter 2 Figure 2.1 Mean preferences for spending dimensions on self-reported ideology ........................................................................................................................ 64 Figure 2.2 Loadings for spending categories on the first and second principal components and respondent ideal points projected onto the recovered space ........................................................................................................................... 70 Chapter 3 Figure 3.1: Four legislator time trends from a scaling of the 1 st -110 th US Senates, with corresponding DW-NOMINATE trends (gray)... ............................ 99 Figure 3.2: Presidential Time trends from Eisenhower to G.W. Bush from a joint two-dimensional scaling of the 80 th -110 th US House and Senate. .............. 103

xii

Figure 3.3: Political polarization measured by the distance between party means ......................................................................................................................... 106 Figure 3.4: Cumulative increase in polarization from member replacement .................................................................................................................................... 109 Figure 3.5: Cumulative increase in polarization from ideological adaptation .................................................................................................................. 110 Figure 3.6: APRE from held out votes from the training sets plotted against specified window sizes ................................................................................. 116

xiii

List of Tables

Chapter 1 Table 1.1: PACs affiliated interest group that issue ratings ...................... 22 Table 1.2: Comparison of correlations with DW-NOMINATE Scores between PAC-NOMINATE, IMWA, and CFscores for the 1980-2008 election cycles............................................................................................................................. 30 Table 1.3: Correct classification rates for Congressional roll call votes for four competing models ................................................................................................ 33 Table 1.4: Testing for significance of competing determinants of PAC contributions ................................................................................................................ 38 Table 1.5: Comparisons of deviance residual measures of explained variance across PACs grouped by industry coding.................................................. 40 Table 1.6 Correct Classification of Candidate Pairs for ideological and non-ideological motives across PACs grouped by industry coding ........................ 42

Chapter 2 Table 2.1: Mean interpersonal spending preference distances ................. 62 Table 2.2: Summary statistics for spending categories .............................. 67 Table 2.3: Correlations across measures of ideology and partisanship ...... 68 Table 2.4: Mean positions for the first and second PCA dimensions and self-reported ideology across selected groups ............................................................ 68 Table 2.5: Regressing Measures of relative demand for rival versus non- rival spending on ideology, partisanship, gender and religion ................................. 76

xiv

Chapter 3 Table 3.1: Classification for 1st through 110th US Senates .................... 95 Table 3.2: Correlations of legislators estimates from two-dimensional scalings ......................................................................................................................... 97 Table 3.3: Comparison of error rates for presidents in static and dynamic scalings ....................................................................................................................... 104 Table 3.4: Dates of change points and corresponding events ................. 111

xv

List of Appendices Chapter 1 Appendix 1.A: Iterated Money-Weighted Averaging ................................. 44 Appendix 1.B: Web Supplemental Data Sources ........................................ 46

Chapter 2 Appendix 2.A: Allocation Section of Questionnaire .................................... 80 Appendix 2.B: Questionnaire Wording ........................................................ 81

Chapter 3 Appendix 3.A: Cross-validation .................................................................. 116 Appendix 3.B: Legislatures Outside the U.S. ............................................ 117

1

Estimating Ideological Positions of Candidates and Contributors from Campaign Finance Records

Abstract: I develop a statistical method to measure the ideology of candidates and contributors using campaign finance records. The method recovers ideo- logical positions for incumbents that strongly correlate with ideological meas- ures recovered from voting records, while simultaneously recovering positions for political action committees (PACs), unsuccessful challengers and open-seat candidates. The method shows promise as a platform for testing hypotheses about contribution behavior. I illustrate by examining which motivations best explain PAC contributions. The results reveal that ideology features promi- nently in the contribution behavior of PACs, but its influence varies consider- ably across categories of PACs.

2

1. Introduction

Many applications in political science require reliable ideological measures of individuals and groups. Methods for scaling roll call votes recover precise ideo- logical estimates but are limited to legislators with voting records. In contrast, me- thods for scaling political texts show great promise in extending estimation outside the confines of legislative bodies but are not yet able to locate individuals with much precision. Scaling campaign contributions offers an attractive middle ground between these scaling methods. Ideological measures recovered from contribution data rival the reliability and precision of those recovered from roll call data. At the same time, scaling contribution data shares lexical analysis’s promise of extending ideological estimation to a more comprehensive set of political actors. A core tenet of spatial models of politics is that political actors prefer ideo- logically proximate outcomes to those that are more distant. The proximity as- sumption predicts slightly different outcomes depending on structure and rules of the observed behavior. In the context of contribution behavior, each contributor must decide how to distribute funds across thousands of eligible candidates with- out violating campaign finance laws. Thus, the proximity assumption predicts that a contributor will rank order candidates by their proximity in ideological

3

space and move down the list giving the maximum legal amount to each candi- date until its budget is exhausted. 1

The problem with creating an ideological map based on a parsimonious spatial theory is that ideological proximity is but one of many considerations that influence contributors. Contributors may give to campaigns in order to purchase legislator favors such as votes or other constituency services (Grossman & Help- man, 1994, 2001a, 2001b; Snyder, 1990), in order to gain access to the candidate to communicate their concerns directly (Hall & Wayman, 1990), or in order to influence regulatory agencies (Baron, 1989; Gordon & Hafer, 2005). Electoral con- siderations can weigh in as well (Mueller, 2003). Most contributors presumably find benefit in supporting candidates with a realistic chance of winning, even if it means contributors with extreme ideologies must overlook many of their most fa- vored candidates.

Despite the complexity present in the data-generating process, the data have a manageable structure. Contribution data can be organized as a contingen- cy matrix, where the rows are contributors, the columns are candidates and each cell represents a contribution amount. This makes scaling techniques already fa- miliar to political scientists suitable for contribution data. I present two of these here.

1 In actuality a contributor’s budget may be endogenous to the selection of candidates and campaign finance laws, but for the sake of exposition I leave this discussion for later.

4

The first scaling method is a variant of the reciprocal averaging procedure common to the correspondence analysis literature (Hill, 1973). I term this method Iterated Money-Weighted Averaging (IMWA). This simple algorithmic method is easy to understand, straightforward to implement and requires no distributional or functional assumptions. It simply models candidate ideal points as the money- weighted averages of their contributors’ ideal points, and in turn, models the con- tributor ideal points as the money-weighted averages of their recipient candidates’ ideal points. This process is iterated until convergence. 2 Despite operating under the incorrect assumption that spatial proximity alone determines contribution outcomes, the IWMA technique successfully recovers ideological estimates that align to standard liberal-moderate-conservative dimension recovered from roll call data (Clinton, Jackman, & Rivers, 2004; Poole & Rosenthal, 1997). The IMWA scaling confirms that ideological proximity influences contri- bution behavior, but it fails to account for non-ideological motives that, in con- trast to Congressional roll call voting, are thought to be important components of the data-generating process. For this reason, I develop a statistical method that builds upon item response theory (IRT) count models developed for educational testing and later adapted to scale political text. I find including both ideological and non-ideological covariates in the model improves the quality of the ideological estimates and provides a powerful framework for the analysis of contributor beha-

2 The IMWA technique is detailed in the appendix.

5

vior. I refer to ideological estimates recovered from the IRT count model as Cam- paign Finance Scores (CFscores). 3

This paper unfolds as follows. Section 2 motivates why scaling contribution data advances the methodology of ideological measurement and marks an impor- tant contribution to the study of contribution behavior. Section 3 discusses me- thodological issues associated with scaling contribution data along with the pro- posed solutions. Section 4 introduces an IRT count model for scaling contribution data. Sections 5 and 6 discuss the data and present results from a joint scaling of the 1980-2008 election cycles, followed by an assessment of model fit in Section 7. The final section focuses on data from the 2003-2004 election cycle to test hypo- theses about PAC contribution behavior.

2. Motivation

Campaign finance is an expansive arena of competing interests that brings together the masses, elites, special interest groups and politicians. Roll call votes are confined to legislatures, but the vast, interconnected flows of political money pervade nearly every level of American politics. The contributors that give to can- didates from different institutions are ideal bridge actors that provide the “glue” (Gerber & Lewis, 2004; Poole & Rosenthal, 1997) needed to construct a common-

3 The term “CFscores” is not interchangeable with “IRT count model results” since the CFscores are only a component of the overall IRT count model.

6

space scaling. Consequently, scaling federal contribution records automatically puts contributors and House, Senate and Presidential candidates on a common scale. 4 Moreover, a large and growing number of states make their campaign finance databases available to the public. As such, the methods I develop easily extend to candidates for state legislative, judicial, and gubernatorial office as well as ballot measure campaigns. We stand to learn much from such an extension. For one, empirical tests of spatial models of electoral politics require data on candidate positioning, includ- ing those who never serve in Congress. The proposed methods can recover esti- mates for most of these unsuccessful candidates. In addition, PAC ideal points are quantities of interest in their own right and have the potential to cast light on a wide range of political phenomena. Among other things, such measures make it possible to examine whether PACs from a given industry have adopted similar ideologies, suggesting a coherent lobbying platform, or have more diffuse policy preferences, suggesting diverging or opposed lobbying platforms. Moreover, if indi- vidual contributors are included in scaling, it then become possible to explore how ideology differs across professions, or to examine the extent to which the ideologi- cal distribution of individuals differs from that of special interest groups. Recovering ideological estimates for a more expansive set of political ac- tors is not the only advantage. The empirical traction provided by the IRT count

4 The ideological positions of candidates are not ideal points, per se, because they are not directly derived from choices made by the candidates. Rather, they are representations of candidate ideol- ogy as perceived by contributors. The distinction is subtle but merits acknowledgement.

7

model has the potential to aid researchers in addressing a number of important questions in the campaign finance literature. A question central to the literature asks whether PAC contributions are position-induced (i.e. ideological) or service- induced (i.e. quid pro quo transactions) (Ansolabahere, de Figueiredo, & Snyder Jr., 2003; C. Cameron & Morton, 1992; McCarty & Rothenberg, 1996; Poole, Romer, & Rosenthal, 1987). The analysis in Section 8 uses the IRT count model to assess how well these hypothesized motives account for contribution behavior of PACs.

3. Recovering Ideological Positions from Campaign Finance Records

Contribution data lends itself less readily than roll call data to ideal point estimation. With respect to theory, the spatial model of voting is an elegant repre- sentation of a voter’s choice function. It simply predicts that a voter chooses the outcome nearest his ideal point. With respect to empirics, roll call votes naturally present themselves as binary outcomes. Assuming abstentions are missing data; a legislator can vote either “yea” or “nay”. This greatly reduces the complexity of possible outcomes. A spatial model of campaign contributions is less forthcoming. This is in part owing to restrictions placed on contribution amounts, but it largely stems from an increase in the number of possible outcomes from which to choose. A PAC can give to everyone, no one or to any one of the possible candidate combi-

8

nations. In other words, a multinomial choice problem best characterizes the deci- sion to contribute, which complicates the relationship between spatial theory and estimation. One conceptually sound method of reducing this complexity is to simplify the choice problem by treating incumbent-challenger pairs as the unit of observa- tion. By assuming that PACs consider electoral races on a per case basis, one can restructure the dependent variable in one of two ways. The first is to code contri- butions to incumbents as positive values and contributions to challengers as nega- tive values (Poole, et al., 1987; Wand, 2009). The second is to treat contributions as voting outcomes, where a contribution to the incumbent is coded as a vote for and a contribution to the challenger as a vote against the incumbent (McCarty & Poole, 1996). This approach does much to reduce the complexity of potential outcomes but comes at a cost. Using candidate pairs as the unit of analysis is an intuitive way to simplify empirics, but it restricts analysis to congressional races with a via- ble challenger. In the 2008 Election, only 139 House incumbents faced challengers who raised more than $100,000; the remaining House incumbents ran essentially unopposed. Yet, on average, unchallenged House incumbents raised 83 percent as many funds as incumbents who faced competitive challengers. In fact, PACs di- rect a sizable majority of their funds at non-competitive races. Consequently, ana- lyzing candidate pairs has the undesirable effect of leaving the majority of contri- bution behavior unexplained.

9

The alternative is using contribution amounts for contributor-candidate pairs as the unit of observation. Although contribution amounts are naturally measured in monetary values, the data closely approximate count values. Figure 1 shows that in practice PACs tend to contribute in multiples of $500. As a result, con- verting the contribution amounts into count values by rounding up to $500 inter- vals results in a negligible loss of information. Fortunately, models for scaling count data are common in the IRT literature. They include the Rasch Poisson Counts Model (RPCM) (1980), an extension of the RPCM by Van Duijn and Jansen (1995), and negative binomial factor analysis (Ogasawara, 1999). A similar group of models estimate ideological positions of legislators and parties from text (Laver, Benoit, & Garry, 2003; Monroe, Colaresi, & Quinn, 2008; Monroe & Maeda, 2004; Slapin & Proksch, 2008).

Figure 1: Frequency plot of contribution amounts during all primary or general elections from 198 0 to 2008. Each contribution amount represents a contributor-candidate pair.

10

The model presented in the following section extends IRT count models to contribution data. A notable addition is the incorporation of item-specific cova- riates for non-spatial candidate characteristics widely believed to influence contri- bution decisions. The section begins with a discussion of the contributor utility function and choice problem. A presentation of the IRT count model is next, fol- lowed by a description of the estimation routine.

4. Statistical Model of Contribution Behavior

There are two standard perspectives on why PACs contribute to candi- dates. The first is known as the selection hypothesis. It conjectures that contribu- tions are position-induced. Interest groups identify candidates with ideological po- sitions proximate to their own and contribute in order to bolster these candidates’ election efforts (McCarty & Poole, 1996; Poole & Romer, 1985). In this account, interests groups seek to influence the ideological composition of Congress without expecting legislative favors in return. According to the second perspective, interest groups exchange contributions quid pro quo for either legislative services or access (Kroszner & Stratmann, 1998; Snyder Jr., 1990). Legislative services may take the form of purchasing legislative votes or promises from legislators to pressure regula- tory agencies (Baron, 1989). Access entails the privilege to communicate directly to legislators the contributor’s interests, insights and concerns (Hall & Wayman, 1990).

11

Some candidate characteristics are associated with different motives to contribute. The literature categorizes these characteristics into three general types: ideological, structural, and electoral. Ideological characteristics are consistent with position-induced giving and are a function of spatial proximity. Structural charac- teristics are consistent with service-induced giving and include whether an incum- bent is a member of the party leadership, a committee chair, and the incumbent’s committee assignment. Electoral characteristics include status as a candidate for the House or Senate, incumbency status, and electoral competitiveness. Contribut- ing with respect to electoral considerations concerns picking winners and can be consistent with either service-induced or position-induced giving. I assume PACs derive utility from contributing to candidate as a func- tion of ideological, structural, and electoral motives. PACs experience quadratic utility loss with respect to the ideological distance. I further assume that the non- ideological portions of the utility function are additively separable. 5 The determi- nistic portion of the utility function summing over all candidates is displayed be- low.

5 It is possible to implement alternative functional forms for the utility function. The estimation software was designed with this in mind and allows considerable flexibility in specifying the contri- butor utility function. Adjusting the model to include interactions or weighting terms is a relative- ly straightforward process. Nonetheless, I have found that the additively separable utility function specified above works well and is a good baseline since it requires the fewest additional assump- tions.

,

,,

,

,

,

,

,

,

= − (

−

)

+

+

+

+

4.1

12

and

represent the ideal points for candidate and contributor .

and

weight the importance of ideological proximity for candidate and contributor . 6

represents the matrix of covariates associated with electoral characteristics and

represents the matrix of covariates associated with structural characteristics.

and

are the direction and magnitudes of electoral and structural characteristics.

can be loosely interpreted as measuring a candidate’s latent fundraising ability not registered by the other candidate characteristics – for example, a candidate may be charismatic fundraiser or exert more time and effort than others soliciting contributions. Lastly,

captures any motives to contribute not picked up by the other covariates. Such motives might include signaling strength to Congress or regulatory agencies (Gordon & Hafer, 2005), consumption benefits (Ansolabahere, et al., 2003), or contributing broadly seeking access to large group of candidates.

The choice problem for contributor is how best to distribute its budget across candidates. I assume that a PAC will give until it exhausts its budget. Al- though PACs occasionally raise more funds than they contribute during a cycle, the typical PAC spends until its budget is exhausted. Every term in the utility function is exogenously determined; thus, the contributor’s problem is simply to identify the set of candidates that maximize his utility.

6 By assumption,

> 0 and

> 0.

13

Error is introduced into the model via a right-censored negative-binomial distribution. The model assigns each contributor a likelihood of giving to each candidate as a function of contributor utility. The negative-binomial distribution is preferred to the more commonly used Poisson distribution because it adjusts for the severe over-dispersion present in the data—the variance is routinely ten times that of the mean. 7 The likelihood equation is right-censored at ≥ 10 to account for the $5,000 contribution limit (A. Cameron & Trivedi, 1998). The data is organized as a matrix of counts, where cell (,) is the contri- bution amount of PAC to candidate . Let

be a vector of length that represents PAC ’s contribution profile. The model takes the following form:

=

(

)

4.2

(

| λ

,σ

) = ℎ(

| λ

,σ

) i f < 10 1 − ℎ(| λ

,σ

)

i f = 10

4.3

7 In robustness checks of distributional assumptions, I have found that the negative-binomial model greatly outperforms the Possion model in terms of measures of model fit but that the recovered ideal point estimates are nearly identical. I also ran tests using zero-inflated (Greene, 1994; Lambert, 1992; Long, 1997) and hurdle models (King, 1989; Mullahy, 1997), which are designed to account for the mass point at zero. Although zero-adjusted models have a number of desirable qualities for this type of application, the large increase in the number of parameters associated with these models is unwarranted. The zero-adjusted models nearly double the number of parameters but only marginally affects the quantites of interest. The correlations of ideal point estimates from the negative binomial and ZINB models are 0.996 for candidates and 0.991 for contributors. Consequently, I find that the negative binomial model offers the best compromise between computational efficiency and model fit.