Pivotal Politics and the Ideological Content of Landmark Laws. Thomas R. Gray Department of Politics University of Virginia

Transcription

1 Pivotal Politics and the Ideological Content of Landmark Laws Thomas R. Gray Department of Politics University of Virginia Jeffery A. Jenkins Department of Politics University of Virginia January 3, 2016 The pivotal politics model (Krehbiel 1998) has significantly influenced the study of American political institutions, but its core empirical prediction that the size of the gridlock interval is negatively related to legislative productivity has not found strong empirical support. We argue that previous research featured a disconnect between the exclusively ideological theory and tests that relied on outcome variables that were not purely ideological. We remedy this by dividing landmark laws (Mayhew 1991) into two counts those that invoke ideological preferences and those that do not and uncover results consistent with pivotal politics core prediction: the size of the gridlock interval is negatively related to the production of ideological legislation. We also find that the size of the gridlock zone is positively related to the production of non-ideological legislation. These results and the substitution effect they imply hold up in the face of both sensitivity analyses and robustness checks. An earlier version of this paper was presented at the 2015 annual meeting of the American Political Science Association, San Francisco, CA, and in the American Politics Workshop at Columbia University. We thank Scott Adler, Sarah Anderson, Joshua Clinton, Robert Erikson, Keith Krehbiel, Nathan Monroe, Sharyn O Halloran, Jesse Richman, Jason Roberts, Jonathan Woon, and various workshop participants for many helpful comments.

2 Introduction Since first appearing nearly two decades ago, Keith Krehbiel s pivotal politics model (see Krehbiel 1998) has been an influential and widely used analytical framework in the study of American political institutions. 1 Initially designed to study the U.S. lawmaking process, the pivotal politics model a one-dimensional spatial model predicated on the notion that outcomes are constrained by supermajority voting rules has proven remarkably flexible. In addition to inspiring work that has delved more deeply into legislative productivity (Wawro and Schickler 2004, 2006; Lapinski 2008, 2013), the pivotal politics model has also been adapted and extended to study unilateral presidential action (Moe and Howell 1999; Howell 2003), presidential agenda setting (Beckman 2010), congressional delegation to executive agencies (Epstein and O Halloran 1999), and advice and consent in treaty making (Auerswald and Maltzman 2003) and judicial selection (Johnson and Roberts 2005; Primo, Binder, and Maltzman 2008). Despite its influence and widespread usage, pivotal politics has received mixed results in empirical testing. Krehbiel (1998) found a significant, negative relationship between the size of the gridlock interval (the ideological space between the members who represent the cloture and veto-override pivots, respectively) and the number of important laws enacted in the post- World War II era, in keeping with the theoretical expectation. However, more recent tests of pivotal politics have not produced similarly supportive results (for a useful summary, see Woon and Cook 2015). This has been true not only in studies of legislative productivity, but also when various roll-call-based measures are used as dependent variables. 1 A common measure of influence is the number of Google Scholar citations. As of 1/3/16, Krehbiel (1998) has generated 1,709 citations. See EC6xzkAAAAJ:u-x6o8ySG0sC 1

3 In sum, the current state of the literature suggests that pivotal politics provides useful theoretical intuition (or perhaps serves as a useful starting point for theory building), but falls short in empirical verification. Going further, superior empirical results from various partybased models might imply that pivotal politics is too reductionist, and that too much real-world complexity has been stripped away in the model-building process. Our goal in this paper is to revisit pivotal politics, as applied to the production of landmark laws (a particular type of important laws, which we define in the empirical section of this paper), and extend the testing of the model through the present day. In addition, we argue that prior pivotal-politics-based studies of legislative productivity have been lax in fully connecting theory to testing, specifically in the construction of appropriate and consistent measures. That is, while the pivotal politics model is explicitly ideological (spatial), and the key independent variable (the size of the gridlock interval) is appropriately ideological (spatial) in nature, the dependent variable is typically just a count of important laws regardless of whether those laws actually invoke ideological preferences. Following Lee (2009), we code landmark laws by ideological content, and then examine the performance of the pivotal politics model in a fair test using landmark ideological laws as the dependent variable. We uncover results consistent with pivotal politics expectation: the gridlock interval is significant and negatively related to the number of landmark ideological laws. Moreover, we uncover the opposite effect (a significant, positive relationship) when looking at landmark non-ideological laws; while pivotal politics makes no predictions on such non-ideological laws, the results are instructive, and we offer some preliminary suggestions as to their broader meaning. 2

4 The paper proceeds as follows. We first provide a brief overview of the pivotal politics model and its history of empirical results. Next, we update the data through the 113th Congress and retest the standard pivotal politics model. Consistent with prior studies, we find no connection between the size of the gridlock interval and the number of landmark laws produced. We then separate our landmark-law measure into two groups ideological laws and nonideological laws and uncover results consistent with the theory on the former group and in the opposite direction on the latter group. We then conduct sensitivity analyses (regarding the distribution of status quos) and robustness checks (regarding the construction of the gridlock interval) and find that our initial results hold up. Finally, we conclude with some thoughts about the importance of matching testing to theory. Pivotal Politics: Theory and Evidence The pivotal politics model, as developed in Krehbiel (1998), is an elegantly simple approach to explaining legislative productivity that focuses on the pivotal actors in the lawmaking process: the legislators who decide (a) whether a filibuster on a bill will be broken and (b) whether a presidential veto on a passed bill will be overriden. The model assumes that a single dimension of preferences over policy outcomes (often thought of as a left-right ideological dimension) captures the entire lawmaking process, and that these policy preferences are the sole determinants of votes. More specifically, legislators are assumed to possess single-peaked, symmetric preferences over policy outcomes and, given two options, pick the one closer to their ideal point. There is no role in the theory for parties, elections, or any other concern, except so much as they endogenously alter members policy preferences. Pivotal politics thus functions as an extension of classic median-voter games, with the addition of two supermajority features of the federal lawmaking process: the filibuster in the 3

5 Senate and the presidential veto. These deviations from simple majority rule yield two new pivotal actors: a filibuster pivot and a veto pivot. The filibuster pivot is the senator who decides the success of a filibuster attempt (i.e., whether cloture is invoked or not). The veto pivot is the member who determines the success of an attempt to override a presidential veto. 2 The decisions of each pivot are clear from the model. Since all status quos, once considered, move in the direction of the median voter, any status quo that lies between the pivot and the median would move away from the pivot, making any passable change less desired than the status quo itself. Thus, the leftmost pivot on the line will not consider moving any status quo that lies between itself and the median member. The same relationship holds for the rightmost pivot. This effectively creates an interval of status quos between the leftmost and rightmost pivots that cannot be altered. Any of these status quos, which a majority could replace with a policy closer to the median, would either be filibustered or vetoed (with an insufficient number of votes to invoke cloture or override, respectively). This set of unmovable status quo policies is commonly called the gridlock interval. Figure 1 (adapted from Krehbiel 1998, p. 35) presents one possible set of pivotal actors: F represents the Filibuster Pivot; M, the Median Member; V, the Veto Pivot; and P, the President. No status quo between F and V can be altered in the equilibrium of the game. [Figure 1 about here] While many scholars and pundits have identified divided government as the cause of legislative stalemate (c.f. Mayhew 1991), pivotal politics points instead to the varying size of the gridlock interval as the chief explanation of changes in legislative productivity. If the 2 Generally, applications of the theory assume the President is more extreme than the veto pivot, though this is not a necessary assumption. Empirically, this has always been the case since, at least, 1947, according to Common Space DW-NOMINATE scores. 4

6 distribution of status quos is assumed to be uniform, 3 then the larger the gridlock interval, the fewer the policies that are available to be moved. Only status quos at the extremes of the distribution of members preferences can be altered. However, as the gridlock interval narrows, more moderate policies become available for change. This expanding and contracting gridlock interval could explain decreasing and increasing levels of legislative productivity. It could also explain why Congress passes many bills with greater-than-necessary majorities and often with bipartisan support. In short, the pivotal politics theory predicts a negative relationship between the size of the gridlock interval and the production of important laws. Krehbiel (1998) found initial empirical support for this negative relationship. 4 However, in the succeeding years, additional confirmatory evidence has been hard to come by whether studying legislative productivity or other more general phenomena. Chiou and Rothenberg (2003) found no significant relationship between the size of the gridlock interval and the proportion of significant bills on the agenda that were enacted in the post-wwii period. In a 2006 study, the same authors found similar results with regard to counts of significant laws over a period extending into the 19th century. Covington and Bargen (2004) and Stiglitz and Weingast (2010) found weak results for a purely preference-driven theory relative to a partisan gatekeeping model. Krehbiel, Meirowitz, and Woon (2005) uncovered ambiguous results for pivotal politics (in comparison to other, party-based models), while Clinton (2007) uncovered little support for a pivotal politics approach. Richman (2011) found mixed results for pivotal politics and stronger results for a partisan model when analyzing policy locations using NPAT surveys. Woon and Cook (2015) also presented mixed results for pivotal politics in a novel test 3 This is a common assumption, but see Woon and Cook (2015) for a recent innovation, which builds on earlier work by Krehbiel (2006a, 2006b). We relax this assumption in a later section of this paper. 4 This was true in terms of both important enactments (Sweep One and Two from Mayhew 1991) and landmark enactments (only Sweep One). 5

7 that departed from the common assumption of uniformly distributed status quos. Only Heitshusen and Young (2006) uncovered results consistent with pivotal politics, finding a negative relationship between policy production (based on the number of section changes to the U.S. Code) and the size of the gridlock interval for the era. Thus, in the nearly two decades since Krehbiel (1998), there have been more failures and ambiguous results than successes. Given the inherent logic of pivotal politics, these results might lead one to believe that the theory serves as a useful starting point, but that additional complexity (e.g., some role for parties) is needed to better capture the data generating process in lawmaking. We reject this supposition, which we document in due course over the next two sections, and in doing so, resuscitate the empirical bases of pivotal politics. Extending and Retesting Pivotal Politics We first retest the core legislative productivity prediction of pivotal politics, that gridlock interval increases lead to reduced legislative productivity. We begin with the 80th Congress ( ) and extend the data through the 113th Congress ( ), for a longer temporal range than has been used in previous research. Our primary dependent variable is the count of Landmark Laws as compiled by Mayhew (1991) and updated through The unit of analysis is each individual two-year Congress. Thus, each unit s dependent variable is the number of landmark enactments in that Congress. 6 Mayhew s dataset runs from 1947 to 2014, representing 34 Congresses. Note that Mayhew s list of important laws is based on two separate sweeps. Sweep One counts laws deemed significant by newspaper reporters at the time, while Sweep Two counts laws 5 Files for both the original dataset and the updates through 2014 are available at 6 We removed all treaties from Mayhew s counts, because treaties do not go through the entire lawmaking process (both House and Senate) that pivotal politics models. 6

8 considered significant by experts in retrospect. Sweep Two requires the passage of time and thus was only applied by Mayhew through While both the combined and Sweep One counts are imperfect, having a consistently measured dependent variable is essential. Even if both sweeps covered the same period, they are fundamentally different time series and combining them is not advisable (Howell et al. 2000). Therefore, we rely only on Sweep One counts in this paper, which Cameron (2000) and Howell et al. (2000) refer to as landmark laws. 7 Figure 2 shows Mayhew s Sweep One laws for the 80th through 113th Congresses. During this post-war period, just under 10 landmark laws were produced per Congress, on average, with a minimum of four (86th Congress; ) and a maximum of 19 (89th Congress; ). 8 The standard deviation is [Figure 2 about here] The key independent variable is a measure of the Gridlock Interval. Krehbiel s (1998) original measure relies on partisan information (shifting seat control between the two major parties) to determine a change in the gridlock interval despite the theory not actually including parties in any way. A better way of testing the theory would make use of data that are not explicitly partisan. One option is to use measures of revealed preference, like Common Space DW-NOMINATE scores (Carroll et al. 2015), which has become the standard approach of measuring the gridlock interval in the literature (see Chiou and Rothenberg 2003, 2006; Woon 2009; Richman 2011; Oh 2015; Woon and Cook 2015). 9 We follow this approach, as it allows 7 We will thus use Sweep One laws and landmark laws interchangeably throughout the rest of this paper. 8 Summary statistics for all variables used in this paper are presented in Appendix 1, Table A A measure using NOMINATE scores is well suited to testing the pivotal politics theory because it treats each legislator as an individual and does not rely on partisan information, which is exactly what the theory itself assumes. Krehbiel (1998, p. 74) explained his choice not to use NOMINATE with two arguments. First, he rejected the cardinality of NOMINATE scores and, second, he rejected comparisons over time using NOMINATE scores. The second complaint is easier to reject. With the introduction of Common Space DW-NOMINATE scores, members are comparable across time and chambers, making measuring a gridlock interval on a constant scale a plausible endeavor. The cardinality issue should be mitigated by the innovations of DW-NOMINATE scores. Even if his 7

9 us to estimate the gridlock interval on a single policy-preference dimension (line) for each Congress. 10 Common Space scores do not allow individual members to vary over time, but unlike dynamic DW-NOMINATE scores, they are comparable across chambers. Because the pivotal politics theory places senators and representatives on the same policy dimension, this is a necessity. While the constancy of member Common Space scores does diminish the variance between units, Poole (2007) has argued that members do not change significantly during their time in Congress, and the use of static scores is justified in this context. Figure 3 shows the size of the gridlock interval for the 80th through 113th Congresses, while Figure 4 shows the regions of the policy line gridlocked in each Congress over the same range. 11 The mean gridlock interval is 0.44 on the NOMINATE scale, which encompasses more than a fifth of the measure s theoretical range. The minimum is 0.24 (95th Congress; ) and the maximum is 0.65 (113th Congress; ). After the series low in the 95th Congress, the size of the gridlock interval has been increasing in a near continuous fashion (with just a few Congress-to-Congress declines), in keeping with the ever-rising polarization in Congress. [Figures 3 and 4 about here] suspicions carry some weight, the potential flaws must be weighed against his proposed alternative: a measure that assumes a one-to-one relationship between parties and ideologies, and is predicated on major assumptions about the relationships between presidents and congressional parties. A cardinality assumption of NOMINATE scores, in our minds, is ultimately less demanding than the assumptions his gridlock-change variable asks us to make. 10 In a later section, we replicate all models that use the Gridlock Interval variable with a measure derived from Adjusted ADA Scores. Our findings are robust to using ADA scores instead of NOMINATE-based measures. 11 Constructing a gridlock interval is not a straightforward process. Rarely, for example, does the membership of a Congress remain constant from the first day to the last. Thus, we must make choices about whom to count. To construct these intervals, we deleted from each chamber in each Congress the members who had cast the fewest votes until we arrived at the appropriate size of the chamber. From that, we created rank orderings of firstdimension scores within each chamber and within each Congress, ignoring party affiliation. We then took the firstdimension Common Space DW-NOMINATE score as the ideal point for each pivotal actor in the theory. For example, the 60th senator from each direction would decide the success or failure of a filibuster, while the 67th senator from the opposite direction as the president would decide the success of a veto override. (In each Congress, the number of the pivotal actor is appropriate for the number of members in the chamber at that point in time and the percentage of legislators required under the contemporary version of the relevant rule; for example, prior to the Cloture Rule change in 1975, a two-thirds majority was required to break a filibuster.) The gridlock interval is the distance between the leftmost and rightmost pivots. 8

10 We also include other factors that are potentially part of the data-generating process. This is important because lawmaking is inherently complex and responds not only to the preferences of legislators, but the influences exerted on them by their electorates, the context and time in which they serve, and exogenous shocks. Many macro trends influence the agenda and the immediate necessity of bills: international relations, economic cycles, and popular movements are three such examples. Any of these may trend in the same upwards direction as the gridlock interval and thus be subsumed into that variable should a similar important factor be omitted. While avoiding bias is crucial, it is also difficult since we begin with a maximum of 34 observations. Each included variable corrects for potential omitted variable bias, but also removes a crucial degree of freedom. Thus, we must be as minimal as possible while also incorporating the necessary variables to plausibly describe the legislative process. We include single proxy measures for the political, economic, electoral, and international relations context that existed during each Congress. For the political context, we use Unified Government, which is coded 1 when the President s party also controlled both chambers of Congress and 0 otherwise. Scholars like Mayhew (1991) and Binder (2003) have debated whether unified partisan control of government enables legislating. There are clear reasons to think that it would: if parties are useful and powerful institutions, then a unification of the levers of power should make it easier to enact a party s agenda. To capture the economic context, we include GDP Growth Rate, which is the average of the two annual growth rates of the United States Real Gross Domestic Product during a given Congress. 12 Economic concerns motivate many important laws, ranging from bailouts and jobs bills to industry subsidies and stimulus 12 These data come from the U.S. Department of Commerce s Bureau of Economic Analysis and are available at: 9

11 packages. Economic circumstances may change the immediacy of a bill s need and in extreme circumstances may represent a shock that temporarily alters preferences over certain policies. 13 For the electoral context, we include a measure of the national Policy Mood, as developed by Stimson (1991), which taps the public s desire for new federal policy. 14 When this measure is high, there should be greater electoral appetite (and demand) for new legislation. If members of Congress exhibit an electoral connection, then these changing pressures may influence their willingness to support legislation. 15 Finally, for the international relations context, we include an indicator of whether the United States was at War, which is coded 1 for all conflicts that lasted at least half of one Congress and 0 otherwise. 16 Wars often necessitate emergency funding legislation, as well as corresponding compensation systems for members of the armed forces (and their families). Though Congress technically declares wars, Presidents retained control of the military and war-making powers in this period to such a degree that endogeneity in this variable is not an overwhelming concern. In Table 1, we present results from OLS models estimating the relationship between the Gridlock Interval and Sweep One laws. 17 Model 1 is a simple bivariate model, while Model 2 includes controls. Directional expectations are included next to the variable names. [Table 1 about here] 13 As with almost all macroeconomic controls, there is risk for post-treatment bias by including this measure. Excluding it, however, risks omitted variable bias. Additionally, to use only a lagged, pre-election value likely increases measurement error. There is no perfect solution. We opt to include it. 14 The specific measures (biennial) used can be found at: These data cover the years We use the contemporaneous measure of Policy Mood rather than a lagged measure; however, our findings are also robust to a lagged specification. 16 These include the Wars in Korea, Vietnam, and the combined Afghanistan and Iraq conflict (or the War on Terror ). For the Wars in Afghanistan and Iraq, which are difficult to properly date, we included them through 2010, when combined American troop deployment in those conflicts dropped below 100,000 persons. 17 In this and all future tables, we incorporate OLS (with appropriate time-series corrections) for ease of interpretation. All of our results are robust to the use of models specifically designed for count data. Table 1 is reproduced in a Negative Binomial regression in Appendix 2, Table A

12 To reiterate, the pivotal politics model predicts a negative relationship between the size of the gridlock interval and legislative productivity. That is, as the gridlock interval expands, the number of status quos that can be altered (or legislated on) decreases; therefore, we should expect fewer landmark laws. With this in mind, the results in Table 1 are striking. In the bivariate model (column 1), the coefficient for Gridlock Interval, while signed correctly, is not statistically significant. This null result is also present when controls are added (column 2). Of the control variables, only the coefficient for War is significant in the expected direction. The coefficients on Policy Mood, Unified Government, and GDP Growth are all signed correctly, but are not significant at conventional levels. Rethinking Theory to Testing: Selecting Similar Measures The preceding analysis is not very surprising given the history of weak empirical results for a pure preference-based theory of pivotal actors. Yet, this type of testing suffers from deficiencies in accurately translating theories and hypotheses into data and analyses. Many data and testing choices made by Krehbiel (1998) have carried on for more than a decade afterwards. What once helped his cause has (seemingly) become a hindrance. One primary example is the use of Mayhew s important laws, which are a standard in the literature. On their own, Mayhew s laws are not incorrect. When just Sweep One counts are used, they represent one justifiable means of measuring importance: what journalists remarked upon at the end of a session. However, when Sweep One counts are incorporated as the dependent variable in a pivotal politics model, a disconnect between theory and testing occurs. 18 Pivotal politics describes the process of producing ideological laws. Legislators have preferences over policies based on their ideal points on a single dimension of ideological positions. If a status quo 18 The same argument holds for a dependent variable based on both Sweep One and Two laws. 11

13 cannot be placed on the line because it does not invoke political ideology, then the theory cannot make predictions about it. An ideal test of pivotal politics should assess whether the ideological distance between pivotal actors in Congress influences the production of ideological legislation. Yet, Mayhew s laws were chosen only for their importance, not for whether they represent any genuine conflict between liberal and conservative values. As Lee (2009) argues, not all issues generate conflict, and not even all issues for which there is partisan conflict can be coherently placed on an ideological line. This weakness provides an opportunity for a new dependent variable, but does not require departing from Mayhew. Instead, we create subsets of Mayhew s Sweep One laws, separating those that fit into a liberal-versus-conservative ideological conflict from those that do not. In this, we rely on Lee s (2009, pp ) definition of ideology and resulting coding scheme for classifying Senate roll-call votes as ideological or non-ideological. Lee s system focuses on four key conflicts in American ideological battles: economic issues, social issues, hawk-versus-dove debates, and multilateralism versus unilateralism. 19 Economic issues focus primarily on laws that change levels of economic regulation (such as environmental regulations for businesses) or redistribution (for example, changing the graduation of the tax schedule or expanding Medicaid funding) or impact the overall level of government spending and share of the economy (such as large economic stimulus spending). Social issues include civil rights legislation and criminal punitiveness, as well as all laws that push policy away from traditional gender, family, sex, and race norms (such as Don t Ask, Don t Tell, abortion rights, or school prayer). Hawk-versus-dove debates include authorizations for use of military force, weapons investment, and limitations on weapons testing. Finally, multilateralism 19 This definition of ideology is time-bound to the debates between mainstream liberals and conservatives in the post-war era. Lee provides more detail on each category and what should be included, as well as many examples. 12

14 versus unilateralism focuses on debates over the importance of international organizations to America s foreign policy (for example, policies that promote the United Nations). Laws outside of these four categories do not have a clear place in modern American ideological debates. In these cases, placing a policy alternative to the left or right of a status quo is not possible, which makes the logic of pivotal politics inapplicable. Many laws fall into this non-ideological category, including those that deal with good-governance and anticorruption efforts (such as lobbying reform in the 110th Congress), non-redistributive and nonregulatory programs (such as the anti-cancer efforts begun by the National Cancer Act of 1971), and the distribution of power between the branches and within the federal government (such as the War Powers Resolution). For each law that Mayhew identified in his Sweep One, we determine whether it fits into one of Lee s categories of ideological conflict. If so, the law is coded as ideological; if not, it is coded as non-ideological. 20 Figure 5 illustrates the resulting time series of ideological and nonideological Sweep One laws. The mean level of ideological laws (5.85) is higher than nonideological laws (4.09), but both series display meaningful variation (standard deviations of 2.83 and 2.19, respectively). The minimum and maximum for ideological laws are two (106th and 109th Congresses; and ) and 12 (111th Congress; ), while the minimum and maximum for non-ideological laws are one (86th, 95th, 98th Congresses; , , and ) and 12 (109th Congress; ). Overall, the two series exhibit no meaningful correlation (r =.03). [Figure 5 about here] 20 Most laws have many sections and components that are difficult to evaluate in such a dichotomous way. We focus on the core features of the law rather than any add-ons or unrelated provisions specifically, the aspect that Mayhew identified in his brief note on each law. When necessary, we used other historical descriptions of the laws to provide supplementary information. 13

15 Before moving to the tests, it is important to be clear about how these data fit the theory. Pivotal politics generates predictions only for ideological productivity, not for non-ideological productivity. Thus, the fairest test of the theory should focus only on how the size of the gridlock interval is related to the production of ideological laws. In Table 2, we present OLS regression results similar to those in Table 1, except divided by landmark law type. 21 In Model 1, the dependent variable is the count of Ideological Laws, whereas in Model 2, the dependent variable is the count of Non-Ideological Laws. We carry over all theoretical (directional) expectations from Table 1 to Model 1 in Table 2. Without clear theoretical expectations for Model 2, we present it descriptively, with two-tailed tests. [Table 2 about here] When used to explain the productivity of ideological laws (Model 1), Gridlock Interval has a negative, statistically significant coefficient exactly as the pivotal politics theory predicts. A one-standard-deviation increase in the size of the gridlock interval corresponds to a 2.14 decrease in the expected number of landmark ideological laws. We have argued that a proper test of the theory requires a data choice for the dependent variable that actually fits the concepts invoked in the theory: conflict over ideological preferences. Model 1 offers strong support for that argument. When tested with all Mayhew Sweep One laws, pivotal politics generated null results. Yet, when applied to a measure of the dependent variable that is more appropriate for the theory, pivotal politics achieves its predicted (negative) coefficient. The coefficient for Gridlock Interval in Model 2, however, is entirely different: positive and statistically significant. A one-standard-deviation increase in the size of the gridlock interval 21 This same model, estimated using a Negative Binomial regression, is presented in Appendix 2, Table A2-2. All results for the Gridlock Interval variable are consistent with those found using OLS. 14

16 corresponds to a 1.72 increase in the expected number of landmark non-ideological laws. 22 Taken together, the two models in Table 2 explain the model in Table 1. For one set of bills (ideological), the gridlock interval has a strong and significant negative relationship, and for a different set (non-ideological) it has a strong and significant positive relationship. When analyzed collectively, these work out to a small, negative, and insignificant relationship. Side by side, they show the dangers of data that poorly fit a theory: they can undermine a theory s empirical strength just as well as enhance it. 23 The two models feature opposite-signed effects of similar absolute sizes for the gridlock interval. And we see that overall productivity has been noisy, but fairly flat over time. These facts together imply that overall productivity itself is not being changed by the size of the gridlock interval, but rather that the proportion of laws that are ideological changes with the expanding and contracting gridlock interval. In models where the dependent variable is the fraction of landmark laws that are ideological, the Gridlock Interval variable has a strong negative (significant) coefficient, as we would expect given the results in Table Interestingly, Table 2 also sheds light on the debate over unified government. Mayhew (1991) and Krehbiel (1998) have argued that unified government does not, in practice, differ from divided government. This was a natural outgrowth of pivotal politics. If legislators are individual actors rather than partisan groups, then it is the distribution of preferences that matter, 22 These results are robust to secondary considerations of where in the policy space the gridlock interval is located. In Appendix 3, we present results for models that control for the amount of newly ungridlocked space (relative to the preceding Congress), which may contain a disproportionate number of newly movable status quos. 23 These results are not exclusive to studies of Mayhew s laws mixed with NOMINATE scores; they apply to the broader use of ideal-point measures to explain legislative productivity. For example, McCarty, Poole, and Rosenthal (2006) and McCarty (2007) have argued that partisan polarization (deeply related to the gridlock interval, which is a measure of the polarization of pivotal actors, almost always in separate parties) leads to reduced productivity. In Appendix 4 we replicate our preceding analysis using partisan polarization instead of the gridlock interval and find highly similar results. Increased polarization is associated with fewer ideological laws and more non-ideological laws. 24 These models are presented in Appendix 5. 15

17 not partisan control. Table 2 reveals a non-existent effect for unified government when explaining landmark ideological laws. But it also shows a significant, positive effect for unified government on landmark non-ideological laws. And this make some sense, since in the absence of ideological fault lines that members can use as individualists, partisanship may serve as an organizing principle to solve collective action problems within Congress (see Lee 2009). One question that remains is: why does an increase in the gridlock interval predict an increase in landmark non-ideological laws? One plausible explanation is that when the gridlock interval increases, making it harder for legislators to pass laws on which they have ideological differences, they simply switch their focus to laws that do not involve major disagreements. That is, they substitute non-ideological legislation for ideological legislation. More generally, facing a constant pressure to produce even as Congress becomes more polarized, legislators find points of common ground in the form of non-ideological legislation. Following the arguments of Lee (2009), these landmark non-ideological laws may solve electoral needs for members or may help augment the power of members or parties. Moreover, parties may be more cohesive than they are on ideological issues and perhaps even more cohesive across party lines when it comes to mutual electoral or power benefits. This may explain the relative stability of legislative productivity over time even as the gridlock interval has increased. Sensitivity to Assumptions about the Status Quo Distribution The preceding analyses rely on a core assumption common in the literature: that status quos are uniformly distributed across the theoretical range of the policy space. This assumption is highly convenient: it allows us to use spatial measures (the gridlock interval) as our main independent variable. The underlying logic of the negative relationship between gridlock and productivity is actually not spatial, but rather that when a larger percentage of status quos are 16

18 unmovable, then it should be harder to pass legislation. However, if we assume that status quos are uniformly distributed, then the length of the interval is perfectly correlated with the percentage of status quos in that space. Every length of 0.2 on the -1-to-1 policy line corresponds to 10% of all status quo points. An expansion of the gridlock interval by 0.2 anywhere on the line reduces the number of movable status quos by exactly the same amount: ten percentage points. Thus, the length of the line between the pivots is an accurate measure of the percentage of status quos gridlocked so long as those status quos are uniformly distributed. Despite being common, this assumption has not been immune to criticism (Krehbiel 2006a, 2006b; Richman 2011; Krehbiel and Peskowitz 2014; Woon and Cook 2015), and for good reason. On its face, it is a demanding assumption. To be true, it must be the case that at all points in time, 10% of status quo points are on the most liberal 0.2 segment of the policy line and on the most conservative 0.2 segment of the line and that these have the same number of status quos as the segment -0.1 to 0.1. If this were true, it would imply the complete irrelevance of history. Within the context of most theoretical models of lawmaking, the legislative process moves policies towards the median of the chamber. While the median moves, it is almost always around the middle of the line, not at the extremes. Thus, for there to be an equal number of status quos in the center as at the extremes, these processes must not have affected the overall distribution of status quos at all. That is a difficult assumption to accept. A more likely explanation is that since most status quos, when altered by new legislation, are moved in the direction of the center and, in Pivotal Politics, at least, get captured there status quos should be disproportionately found at the center of the line. Given this, a normal distribution is a more persuasive stylized distribution than a uniform one. This is significant because when the distribution is normal, an expansion by 0.2 in one section of the line does not 17

19 necessarily capture the same proportion of the overall status quos as an identical 0.2 expansion at a different point of the line. When the distribution is non-uniform, it is necessary to move away from spatial measures. In short, this becomes a case of measurement error: if the status quo distribution is non-uniform, spatial distances will not properly measure the actual percentage of status quos gridlocked. Consider the following stylized example: A fictional congress has a leftmost pivot (F) at -0.2 and a rightmost pivot (V) at 0.3 on a -1-to-1 policy line, creating a gridlock interval of 0.5. Under the uniform distribution assumption, this converts simply to 0.5/2 = 25% of status quos gridlocked, because the gridlock interval has length 0.5 and the total length of the space is two. However, if status quos are actually distributed normally around the center (0) of the ideological space, the gridlocked percentage might be considerably different. Figure 6 shows two hypothetical distributions one with a standard deviation of 0.2 and one with a standard deviation of 0.3. In the former case, a full 77% of status quos are gridlocked, while in the latter case, only 59% are. These are significantly greater than the 25% obtained using the uniform distribution. [Figure 6 about here] Another way the distribution affects the measurement of gridlocked status quos is through the mean of the distribution. In Figure 6, all distributions were centered at zero. However, if the policymaking process (the median voter) remains left- or right-of-center for any prolonged period of time, the overall distribution may shift in that direction as policies are moved from the periphery towards the median and then captured (as discussed by Krehbiel and Peskowitz 2014). Figure 7 shows how two distributions with equal standard deviations, but centered left- and right-of-zero impact the measurement of the gridlocked percentage. 18

20 [Figure 7 about here] Both distributions feature a standard deviation of Given the -0.2 and 0.3 pivots, the left-ofzero distribution (presumed to be after a period of liberal ascendancy) has 50% of its status quos gridlocked, while the right-of-zero distribution has 67% of its status quos gridlocked. Measuring accurately may thus require making defensible assumptions about the distribution of the data, its spread, and how history has impacted its current center. To test our arguments with these ideas in mind, we replicate the preceding analyses with a variety of possible distributions, and show that our results are robust and insensitive to these potential sources of measurement error. We include two primary categories of potential distributions. First, we analyze truncated normal distributions centered at zero, with a variety of standard deviations. To this, we add truncated normal distributions centered at points informed by the previous Congresses median voters. If policies have gravitated to these points, then these may be more reliable distribution means. We use three different values: the previous Congress s median voter, the average of the last three medians, and the average of the last five medians. These account for the possibility that this influence may be exerted over different lengths of time. Within each type of distribution, we consider different standard deviations. From each distribution, we calculate the percentage of status quos gridlocked with each Congress s pivot locations. We then use these values in place of the Gridlock Interval variable in regressions that replicate Model 1 in Table 2. The results of those regressions, which are reported in Table 3, indicate that our earlier results are not dependent on the uniform status quo distribution. In fact, they are maintained in all but one of the 16 regression models that we analyze. This is because while the magnitudes vary, the resulting measures are still highly correlated across all of the distributions we specify. This implies that the uniform-distribution 19

21 assumption may not create a serious measurement problem after all. In terms of overall model fit, measures based on a normal distribution centered at zero outperform all other measure types, while those that incorporate past median values perform worse than the model with the uniformdistribution measure. [Table 3 about here] Note that the uniform distribution s coefficient is about twice as large as that of any other model. This is striking at first, but has an intuitive explanation. A typical movement of the pivots around the center of the distribution gridlocks fewer status quos when the distribution is uniform than when it is normal, because status quos in the normal distribution are packed in the center. This is evident in the standard deviations of the measures used in Table 3. While the standard deviation of the percentage-gridlocked variable based on a uniform distribution is 0.05, the standard deviation of the percentage-gridlocked variable based on a truncated normal distribution with a mean of zero and a standard deviation of 0.2 is 0.14, almost three times as large. Thus, a typical change is considerably larger. This is true of all of the non-uniform measurements. Replication Using Adjusted ADA Scores to Measure the Gridlock Interval One limitation of our approach is that we must make strong assumptions about the nature of DW-NOMINATE scores. We have to assume that the NOMINATE methodology produces a first dimension of revealed ideological preferences from a set of all votes (votes on both ideological and non-ideological issues) and that the collection of issues that informs this dimension closely matches the set of issues that we use to define liberal-conservative ideology (from Lee). If non-ideological issues, or just a single subcategory of our definition of ideology (for example, position on the government s role in the economy), inform the first dimension 20

22 NOMINATE score, then our measure may be flawed. An ideal measure would estimate ideal points exclusively from votes that would qualify as ideological in our definition. For the range of years we examine, this would require a gargantuan and difficult coding effort. But the possibility of measurement error from the NOMINATE approach cannot be ignored. Thus, we consider a subset of votes that we have reason to believe are disproportionately ideological: votes selected for inclusion in the Americans for Democratic Action s Liberal Quotient Scores. Americans for Democratic Action (ADA) is a liberal interest group that describes itself as America s most experienced independent liberal lobbying organization. 25 The group s function most known to political scientists, however, is its scoring of members of Congress on key roll-call votes. The group typically selects a small number of important votes (20 votes per year in recent years) in each year (not Congress) and generates a score for each member based on the number of times he/she voted in the liberal direction out of the total number of votes. Because these scores are available for more than fifty years and have been constructed to be explicitly about ideology, they have served as measures of ideological preferences in Congress. Groseclose, Levitt, and Snyder (1999) introduced Adjusted ADA Scores to take into account the changing agenda from year to year: each year has different votes, so proposals in some years may be more extreme than in other years, making it easier to get a higher or lower score in certain years. Adjusted ADA Scores are a statistical attempt to correct for the variation in year-to-year (and chamber-to-chamber) averages and dispersions. We use these adjusted scores as calculated by Anderson and Habel (2009), and updated through 2012 rather than the raw (unadjusted) scores This quote comes from the About ADA page on the group s website, and is available at: 26 An alternative approach would be to disregard the scores themselves, even the adjusted versions, and simply use the ADA reports as a means of identifying a subset of ideologically salient votes and the liberal choice in those 21

23 We construct gridlock intervals in the same way we did with the Common Space DW- NOMINATE scores. For each Congress, we start with the career average Adjusted ADA Score of each member who received a score for that Congress. 27 As some members lack scores in individual years (because they served in only a portion of a Congress or missed many votes in a year), some Congresses have an incomplete number of members so we rescale the pivotal voter positions to fit the size of the chamber for which data are available. This generates a measure of the gridlock interval for each year from We then average the relevant two-year intervals to get a measure for each Congress. Figure 8 presents the changing ADA gridlock interval size over time. [Figure 8 here about here] Table 4 replicates Table 1 using these ADA gridlock interval sizes. These models analyze the relationship between the gridlock interval and the productivity of all Sweep One Laws. The results in Table 4 are similar to the results in Table 1. Both by itself (Model 1) and with controls (Model 2), the gridlock interval measured with ADA scores is not significantly associated with the production of Landmark Laws. Overall, then, models with ADA-based measures and NOMINATE-based measures produce results that are substantially the same. [Table 4 about here] Table 5 replicates Table 2, splitting Sweep One Laws into Ideological Laws and Non- Ideological Laws. Again, the results in Table 5 mirror the results in Table 2. In Model 1, the ADA gridlock interval has a statistically significant, negative relationship with the production of votes, and then run a NOMINATE or Bayesian Item Response Theory estimation procedure on them to generate new ideal points for each member. We are currently pursuing this and it will be featured in subsequent iterations of this paper. 27 We use career scores rather than a congress-specific score for two reasons. First, this is more directly comparable to the Common Space DW-NOMINATE scores, which feature one score for each member s entire career. Second, because each annual score is based on only votes, while careers often involve hundreds of ADA-scored votes, the career score is likely to be a more accurate measure of a member s preferences. 22