Redistricting and Polarization in California: Who Draws the Lines?

Transcription

1 Redistricting and Polarization in California: Who Draws the Lines? Corbett Grainger * Preliminary Draft: Do Not Cite Comments Welcome July 11, 2008 Abstract In the U.S., the process of drawing election districts is left to individual states, and critics of legislative redistricting often argue for independent panels to take control of the process. A common claim is that legislative redistricting has been a major contributor to polarization in the American political system; however, there have been few tests of this hypothesis. Previous attempts to test for a relationship between redistricting and polarization have generally relied on cross-state comparisons in redistricting methods and examining voting behavior in the House of Representatives. In this paper, I exploit the fact that the redistricting process in California has alternated between legislatively-drawn and panel-drawn districts for the past four decades. Using roll-call voting data, voter registration, and election returns at the state legisature level, I nd that evidence that legislative redistricting (compared with panel-drawn redistricting) is associated with increased polarization. *Economics Department, University of California, Santa Barbara. grainger@econ.ucsb.edu. The author gratefully acknowledges fellowship support from NSF Grant Number Introduction Throughout the U.S., there has been considerable debate surrounding legislative redistricting, or gerrymandering, the process whereby a state legislature draws election districts. 1 The redistricting process mandated by the Supreme Court's 1 Legislative redistricting is often termed Gerrymandering. Although ubiquitous, this term often has a negative connotation, so I refrain from using it. 1

2 one person, one vote ruling 2 requires that election districts be adjusted to changes in population following every decennial census so that population across districts within a state be approximately equal, but the process of redistricting itself is left to the individual states. In most states, the redistricting process is carried out by the state legislature, but in some cases districts are drawn by independent panels. 3 Some critics argue that the legislative redistricting has been a major cause of the much-documented polarization of the political system over the past several decades. According to this argument, allowing the state legislature to draw election districts leads to safe districts for some legislators, which favors incumbents and makes it dicult for seats to change parties. These safe districts allow legislators to shift to more extreme ideological positions in their roll call votes, which has led to increased polarization in congress and state legislatures. The state legislature in California oers a quasi-natural experiment to test this hypothesis; in the 1970s and 1990s, district lines were drawn by independent panels of judges, whereas the redistricting in the 1960s, 1980s, and 2000s was done legislatively. In this paper, I exploit this fact by using rollcall voting data, election returns and voter registration in California to test the polarization hypothesis. After every federal census, the California constitution requires that election districts be drawn by committees in the state Senate and Assembly to reect changes in population. This has been a contentious issue in California for years. 4 In recent years, four Propositions to change the state's redistricting policy have been put to a statewide vote, but all have failed to pass. 5 Alternative proposals would generally require legislative districts to be drawn by independent panels of judges. Interestingly, two of the past four redistricting plans in California were drawn by retired judges, so-called Special Masters, because in these decades the state legislature and governor reached an impasse on how the lines should be redrawn. This allows a unique opportunity to estimate the impact of dierent types of redistricting on legislative behavior. The eects of redistricting are dicult to quantify, and the end result depends on the goals of those drawing the lines as well as the composition and distribution of voters in a state (Cain, 1985). Political scientists have researched the eects of legislative redistricting, and much research has focused on the number of seats changing parties and the overall balance of power (either in the state or in the U.S. House of Representatives). 6 Partisan redistricting plans 2 See Baker v Carr, 1962, and Reynolds v Sims, There are currently 20 states in which panels draw district lines (Garrett, 2005). For a review of the history of redistricting and the major laws, see Mann and Cain (2005). The process in California will be discussed in more detail in section 2. 4 For example, in a recent issue of The Economist (The Penny Drops), the author's sentiment about the system in California was summed up as follows: California's districts are so shamelessly gerrymandered that hardly any races are competitive.... The result is a political system in which moderation is punished, both by voters and by other politicians. 5 Most recently Proposition 77 in the 2005 California referendum election was supported by 40.2% of voters. A new Proposition will be on the ballot in November, In addition, several authors have focused on racial gerrymandering; for example, see Tate (2003). 2

3 are generally associated with net gains in seats, but as Cain and Campagna (1987) discuss, other goals may be just as important to those drawing the lines. Retaining seats for inuential legislators, displacing incumbents from opposing parties, and maximizing the number of safe districts for one's own party can all be goals for the gerrymanderer. When districts are drawn legislatively, the majority party can aim for political goals. However, legal constraints on contiguity, compactness, and dilution of minority group inuences place limits on the types of districts that can be drawn. Friedman and Holden (2008) discuss optimal strategies for gerrymandering from an microeconomic perspective. Interestingly, Friedman and Holden's recent work calls into question conventional beliefs about redistricting. They show that the practice of packing the minority party into few concentrated districts and creating narrow majorities (cracking) for one's own party is not optimal in the face of uncertainty and other fairly general conditions. Consistent with their model, and important for this study, a risk averse gerrymanderer in the state legislature may instead opt for some safe seats for his own party as well as for the opposition. There has been some work done on the link between redistricting and polarization. Friedman and Holden (2007) argue that the high incumbent reelection rate in U.S. congress cannot be attributed to redistricting. They nd that, all else equal, the re-election rate has decreased over time, which they attribute to increased legal scrutiny. In addition, there have been two recent papers that explicitly test for links between redistricting and legislative polarization. Mc- Carty et al (2007) argue that much polarization in the US Congress cannot be attributed to redistricting, because Democrats and Republicans can both represent the same district between redistrictings in very dierent ways. This nding cannot be explained by changes in district composition. They conduct simulations of neutral districtings, which do not yield signicantly dierent polarization than what has actually been observed. Masket et al (2006) argue that the impact of redistricting on polarization is modest by using a cross-section of state legislatures with dierent redistricting methods. They nd that legislatures in states with partisan redistricting schemes are roughly as polarized as those in states with more neutral redistricting methods. However, because they do not control for state xed eects, their results may be biased. They also argue that state and federal legislatures have polarized between redistrictings as well as during redistricting, a nding that is consistent with the current study. In this paper, I exploit the fact that the redistricting process in California has alternated between legislatively- and panel-drawn districts for several decades. I rst show some anecdotal evidence that elections, on average, are less competitive under legislatively-drawn districts. Then, using panel data from , I estimate the relationship between who draws election districts and polarization in the state legislature. I, too, nd that there has been a time trend in polarization, much of which cannot be attributed to redistricting. However, I nd evidence that legislative behavior changes following each redistricting, and these changes are consistent with the polarization argument. Specically, drawing districts legislatively (compared to panels of retired judges) is associated 3

4 with increased polarization in the state legislature. The paper is organized as follows. Section 2 reviews the process of redistricting in California. Section 3 describes the data. Section 4 presents empirical evidence and tests implications of the polarization hypothesis. Section 5 provides a discussion and concludes. 2 Redistricting in California Following Baker v Carr (1962) and Reynolds v Sims (1964), California was forced to redraw district lines for an election in 1966 to equate (approximately) district populations to adhere to the one person, one vote ruling. Prior to this ruling, districts in California had populations that were anything but equal; urban districts were heavily populated while some rural districts, which were given equal representation in the state legislature and congress, had very small populations. 7 California's constitution calls on the state Assembly and Senate to draw district lines, so following the ruling districts needed to be adjusted legislatively in order to comply with the one person, one vote standard. The redistricting of 1966 has been historically referred to as a bipartisan gerrymander. A few years later, following the 1970 census, the Democratic majority and Republican governor, Ronald Reagan, reached an impasse regarding the proposed redistricting plan, and the task of drawing district lines was passed on to the so-called Special Masters, three retired judges appointed by the California State Supreme Court. The new districts were accepted in late 1973, and were in place for the elections of The story in the following decade was remarkably dierent. By the end of 1981, with a Democratic majority and governor, the Democrats produced election districts that were in place for the elections of To many observers, this redistricting heavily favored Democratic incumbents. A decade later, the 1991 redistricting plan was drawn by Special Masters following another dispute over the proposed redistricting plan, and in 2001 the district lines were drawn in a high prole bipartisan gerrymander. In the 2002 election, not one legislative seat in California changed party hands, which is often cited as an argument against the redistricting system in California. Figure 1 shows a timeline of the most important events in California redistricting in recent history. In this paper, I rely on the alternation of redistricting types to identify the relationship between redistricting type and polarization in the legislature. Districts were drawn by panels of Special Masters in the 1970s and 1990s because in both cases the Legislature (and its Democratic majority) came to an impasse with the Republican Governor about how districts should be drawn. Therefore there may be concern that the eects of panel-drawn districts on legislative behavior may be correlated with the party of the sitting Governor. While Republican Governors were in power when districts were drawn by panels, the 7 Los Angeles County, with a population of about 6 Million people, had one representative in the state senate, while some rural counties had equal representation for as few as 14,000 people. 4

5 Figure 1: Timeline of Redistricting in California relationship between the type of districts in place between districtings and the Governor's party is not systematic. A Democratic Governor was in power when the 1970s panel-drawn districts were in eect, and a Republican governor was in power for most of the 1990s. In fact, since 1967, there has been a Republican Governor in oce in California in all years except (Jerry Brown, Democrat) and (Gray Davis, Democrat). Pat Brown, a Democrat, was in oce during the legislative redistricting prior to the elections in 1966, but Ronald Reagan took his oath of oce in January, When districts are drawn legislatively, certain criteria must be met to satisfy legal requirements, placing constraints on the manipulations that can be done to district maps. A thorough discussion of these constraints is far beyond the scope of this paper, but the main guidelines include population equality across districts, contiguity and geographical compactness of districts. Redistricting plans that do not meet these specications, or those that disproportionately disadvantage minority groups, are subject to court challenges by opponents. Still, there remains much leeway in the gerrymanderer's ability to draw district maps, and research has been devoted to the optimal strategy (see Friedman and Holden, 2008). The Special Masters charged with drawing district lines in the 1970s and 1990s used criteria determined by the courts when drawing district lines. The objectives here are not partisan in nature; instead, a set of guidelines was developed to produce maps that adhere to a set of ideals determined by the courts in Legislature v. Reinecke. The general guidelines include avoidance of minority inuence dilution; population equality within 1 percent of ideal population; contiguity; geographical integrity; county and city integrity; nesting (i.e. combining two Assembly Districts to create each state Senate District); and maintaining communities of interest. The objectives of legislators and courts when creating district maps are strikingly dierent. The constraints that legislators face when drawing districts are non-negligible, but inuential legislators can exercise some choice in the composition of districts. In a sense, the panel-drawn districts of the 1970s and 5

6 1990s serve as a control to test the impact of legislatively-drawn districts on polarization. In theory, the polarization hypothesis could be consistent with the optimal gerrymandering model put forth in Friedman and Holden (2008). Suppose that the gerrymanderer (i.e. the Democratic leadership) in the state legislature during the 1960s, 1980s, or 2000s was risk averse. In that case, depending on the precision of their beliefs about the distribution of voters, they may wish to secure some safe seats for their own party, while packing some districts with constituents of the opposition. If Democratic legislators respond to more liberal districts by voting for more liberal issues, legislative redistricting could lead to increased polarization. On the other hand, in the case of panel-drawn districts, incumbent legislators must run for reelection in districts with constituencies determined outside their control, and once in oce, the composition of districts, on average, may be less favorable to members of both parties. This unique alternation between panel-drawn and legislatively-drawn districts in California allows me to test for dierences in behavior after new district lines are drawn. 3 Data Description In order to test for changes in polarization due to redistricting, I need to rst measure legislative behavior. I use interest group scores from the California Federation of Labor (CFL) and the California League of Conservation Voters (CLCV) for both chambers. 8 In the rest of this paper, I dene polarization as the distance between the average Democrat's and Republican's roll-call voting scores. 9 These interest groups rate members of California's state Senate and Assembly 10 based on their environmental roll call voting record for the previous year. The CFL scores are based on a variety of liberal issues including (but not limited to) labor, insurance, health, appropriations, and human services. A CFL score of 100 corresponds to a legislator voting pro-labor on every vote for which that legislator is present. Similarly, a legislator's CLCV score is simply the percentage of the time that he or she votes pro-conservation on the roll call votes selected by CLCV for that year. These scores have been published 8 To my knowledge, these are the only two scores available at the state legislature level for this time period. California is also unique to have a long history of interest group scores available, which makes it dicult to compare these results to redistrictings over this time period in other state legislatures. 9 An alternative denition that could be employed using interest group scores, which in this case is not too useful, would be to dene polarization as the amount of common space (i.e. overlap) between Democrats and Republicans. Figure 4 in the Appendix shows the number of Democrats with scores less than the most moderate 10% of Republicans, as well as the number of Republicans with scores greater than the most moderate 10% of Democrats. As shown in the gure, because of the time trend in polarization, there are very few such legislators in the past two decades. 10 There are 40 state Senators and 80 Assembly members in any given year. Senate terms are four years, with half of the legislators being up for re-election every two years. Assembly terms are two years. 6

7 since 1974, and were obtained directly from CLCV. 11 I restrict the sample to scores where the legislator votes at least half of the time in a given year. The divergence of Democratic and Republican roll-call vote scores can be seen in Figure 2 and Table 1. As measured by these roll-call scores, Democratic legislators have become steadily more liberal and Republicans more conservative, with some important uctuations. The question remains as to whether these changes in polarization are correlated with redistricting type. As preliminary evidence, in Table 1, Democrats' scores show a trend where they become more liberal when moving to legislatively-drawn districts, and they become more moderate when moving to panel-drawn districts. In examining the three-year average CFL scores of each party, moving to panel-drawn districts in the 1970s is associated with a drop in the average Democrats' score of 14 points. When moving back to legislatively-drawn districts in the 1980s, the average Democrat's score moved back toward the liberal end by 14 points. In the 1990s, scores decreased on average by 4.5 points. Similarly, the dierence between Democrats and Republicans shows a similar pattern, where polarization increases when moving to legislatively-drawn districts. Clearly there is also a time trend present, with the dierence growing over time. The time trend suggests that there are other forces at work driving polarization other than redistricting. Finally, critics of legislative redistricting argue that it leads to safe seats, which in turn leads to polarization. As a proxy for the competitiveness of districts, I use district-level voter registration to capture district constituency. The registration numbers are from October or November, and are the last measure of registration available from each year prior to the general elections before and after redistrictings. I also use election returns at the district level for state Senate and Assembly elections from biennial publications of the Statement of Vote from the Secretary of State in election years. On average, elections in districts drawn by Special Masters in California have been more competitive than those drawn legislatively. 12 Figure 3 shows the average margins of victory, by party, in the state legislature elections over this time period. 11 Often interest group scores are adjusted using a procedure by Groseclose, Levitt and Snyder (1999). This is because in each year and chamber of the legislature, the set of roll call votes, which scores are based on, is dierent. I choose to use nominal scores because the adjustment procedure assumes that any movement from a legislator's ideology is ideosyncratic and unrelated to districts; that is, the adjusted scores would not be valid for my purposes because the adjustment assumes away the behavior I am testing for. Furthermore, because I am using CLCV scores as the dependent variable in my regressions, any measurement error should not bias my results. An additional problem with these scores is that they are bound between 0 and 100. As a robustness check, I estimate all regressions here using symmetrically censored least squares (Chay and Powell, 2001), which does not change the results appreciably. 12 A more thorough analysis of the relationship between redistricting and election outcomes is beyond the scope of this paper. See Friedman and Holden (2007) for a more thorough examination of incumbent reelection rates. 7

8 4 Empirical Evidence My goal is to test for a relationship between how election districts are drawn and polarization. The argument generally involves two steps: 1) legislativelydrawn districts favor incumbents and create safe seats, and 2) safe legislators are then free to vote at their parties extreme, which is the root of polarization in the legislature. As descriptive evidence of the dierence between election competitiveness under panel-drawn and legislatively-drawn districts, as shown in Table 6 (in the Appendix), the margin of victory is larger for both parties under legislativelydrawn districts compared to panel-drawn districts. Figure 3 shows the average margins of victory for members of each party, by districting period. The overall average margin of victory for elections under panel-drawn districts in the California state legislature was about 5 percentage points smaller than in periods with legislatively-drawn districts. 4.1 Voting Patterns in Safe and Competitive Districts It appears that legislative redistricting is associated with an increased ease of reelection and that districts are safer, on average, when drawn legislatively. Legislators in very liberal or conservative districts may take more extreme positions in their roll call voting behavior. In order to test whether safe legislators behave dierently than others, I rst dene safe seats in terms of voter registration rather than election results because of concerns about endogeneity. 13 A safe seat for a Democrat is dened as having adjusted Democratic voter registration in the top quintile of all Democratic seats for that chamber and year. A competitive seat has adjusted Democratic voter registration in the bottom quintile of all Democratic seats for that chamber and year. I dene Republican seats analogously. These data are then merged with roll-call voting data for the year following that election year. To determine the impact of safe seats on legislative behavior, I estimate the following regression equation: 14 Score it = α 0 + α 1 Dem i + α 2 Safe i,t 1 + α 3 Comp i,t 1 (1) +α 3 Comp i,t 1 Dem i + α 3 Safe i,t 1 Dem i + ɛ it, 13 Using margins of victory yields similar results, but a legislator's margin of victory in any election could be based on his or her past voting behavior. I use voter registration for each district in the years 1972, 1974, 1980, 1982, 1990, 1992, 2000, and These cover the registration numbers for the last election under 'old' districts and the rst election following each redistricting in the dataset. 14 The safe legislators in 1983 did not vote signicantly dierent from other members of their own parties in 1981, which suggests that legislators were not given safe districts because of their past extreme voting behavior. However, safe legislators in 2003 had taken more polarizing positions than their average fellow party members. This suggests that the safe legislators of both parties in 2003 may have been given more comfortable districts in that bipartisan gerrymander. 8

9 where Score it is legislator i's roll-call voting score (either CFL or CLCV) in year t, Dem i is a dummy if legislator i is a Democrat, and Safe i,t 1 and Comp i,t 1 are dummy variable for how safe or competitive legislator i's seat was, in terms of voter registration, in the election the fall before year t. A safe Democratic seat is dened by having adjusted Democratic registration (in percentage terms) in the top quintile, whereas a competitive Democratic seat was in the bottom quintile of registration. Republican safe and competitive seats are dened analogously. Some specications also contain year xed eects, and I also estimate the regression separately for legislatively- and panel-drawn districts. The results are in Table 2. Strikingly, the results indicate that Republicans in safe seats take more conservative positions than Republicans in competitive seats. The dierence in voting behavior for Republicans in these regressions is as high as 18 points. 15 The results for Democrats are more noisy than for Republicans (it is signicant in 3 of the 8 regressions), but the signs on the coecients are always consistent with Democrats in safer districts taking more liberal positions in their roll-call votes compared to those Democrats with competitive seats. Interestingly, the dierence for Democrats is signicant only in CFL regressions, which may suggest that Democrats are more prone to change voting patterns due to district competitiveness on labor- or social-related issues than environmental issues. It appears that legislators of both parties behave dierently in safe districts, which is consistent with the polarization hypothesis. In the remainder of the paper I take a reduced-form approach, focusing on the impact of the type of redistricting on polarization in the legislature. 16 The redistricting regime itself can be viewed as the policy variable, as policymakers can change the laws governing redistricting but do not directly control the distribution of voters in the state. As discussed later in this paper, the results presented in this section are robust to the inclusion of additional control variables for partisan control of the legislature, the governor's party, and national trends in legislative polarization. 4.2 Polarization and Redistricting Regimes As a rst test of the hypothesis that legislative redistricting is associated with increased polarization, I begin by estimating the following simple regression: Score it = β 0 + β 1 P anel t + β 2 Dem i + β 3 P anel t Dem i + ɛ it (2) 15 Moreover, these results are robust to including legislator xed eects along with year xed eects in the regression (when estimating the regressions separately by party). This suggests that individual Republicans change behavior due to how safe or competitive the district is that they are representing. The dierence in behavior between safe and competitive seats in these regressions is about 11 points for both CFL and CLCV regressions. 16 To directly test the hypothesis, one could use two-stage approach, by instrumenting to estimate the impact of redistricting regimes on margins of victory and then the impact of margins of victory on legislative behavior. This would not be because of concerns of endogeneity, but instead because the argument is often formed in two stages. I argue that the equivalent of a reduced form regression is more interesting here, because this captures the impact of the policy variable on legislative behavior. 9

10 where Score it is legislator i's CFL or CLCV score in year t, P anel t is a dummy variable that equals 1 if districts in place in year t were drawn by Special Masters and 0 if legislatively-drawn, and Dem i is a dummy variable equal to 1 if legislator i is a Democrat (= 0 if Republican). These results are shown in columns (1) and (3) of Table 3. In column (1) in Table 3, Democrats in panel-drawn districts take more moderate positions as measured by the CFL scores. The CLCV results in (3) again suggest that Democrats behave signicantly dierent under panel-drawn districts than when districts are legislatively-drawn. They vote, on average, 6.8 points more pro-environmentally when districts are legislativelydrawn. To look for evidence of increased polarization due to legislatively-drawn districts, I test the null hypothesis that the dierence between the average scores of Democrats and Republicans is the same across redistricting regimes. From the regression in equation (2), one can test the hypothesis that the dierence in polarization between district types is signicant; in this simple regression, that amounts to simply testing H 0 : β 3 = 0. The dierence in polarization as measured by CFL scores is not signicant in this simple regression; however, when examining polarization on environmental issues, the dierence is about 6 points and is signicant at the 1% level. That is, polarization is higher under legislatively-drawn districts. The regressions from estimating equation (2) restrict the behavior to be the same in all legislatively-drawn districts (as well as in the panel districts of the 1970s and 1990s). I relax this assumption by extending equation (2) to allow for dierent levels for each set of districts. Specically, the equation is Score it = γ 0 + γ 1 Dem i + φ P + δ P Dem i + ɛ it (3) where P is a vector of period-specic dummy variables for the districts of the 1960s, 1970s, 1980s and 1990s redistrictings. The results are in Table 4. The results in columns (1) and (4) show the time trend in polarization, with some evidence that Democrats became more moderate with respect to their labor-related voting records in the 1970s compared to in the legislatively-drawn districts of the 1960s and 1980s. The dierence in behavior in later decades is not signicantly dierent here, which reects the fact that average scores were very close to the censoring points 0 and Controlling for Legislator Fixed Eects One test of the polarization hypothesis would be to examine how legislators change their behavior due to dierent types of redistricting. If some legislators changed to more polarizing positions as a result of gerrymandered districts, 17 it would be evidence in support of the polarization hypothesis. I proceed by modifying equation (3) to include legislator xed eects. Because party aliation is 17 Or, similarly, if legislators move to more moderate positions when changing from legislatively- to panel-drawn districts, this would be evidence in support of the polarization hypothesis. 10

11 time-invariant, the xed eects regressions were run separately for each party and interest group score. The results are shown in columns (2), (3), (5), and (6) of Table 4. Holding legislator-specic characteristics constant, Democrats take more extreme positions in both the CLCV and CFL regressions under the legislativelydrawn districts compared to the panel-drawn districts. Moreover, in the CFL regression, Democrats took more polarizing positions in the legislatively-drawn districts of the 1960s compared to the panel-drawn districts of the 1970s. The dierence between Democrats' behavior in the 1990s and 2000s is not signicant at the 5% level using either measure. This is not particularly surprising considering that the median CLCV and CFL scores for Democrats in the year 2000 were near 100, the censoring point. Republicans, on the other hand, show a constant tendency towards more conservative positions, even after including legislator xed eects. This trend persists using both the CLCV and CFL scores. These results suggest that Democratic legislators take more extreme voting positions when districts are drawn legislatively, while Republican behavior appears to be more of a time trend. Taken together, there is an increase in polarization when districts are drawn legislatively. These changes in polarization are largely driven by changes in Democratic behavior. 4.4 Time Trends and Discontinuities By examining the results so far, there is some evidence of polarization levels being correlated with redistricting type, but there is clearly a time trend for Democrats and Republicans that may be confounding the simple regression results presented so far. As pointed out by others (e.g. Masket, Winburn, and Wright, 2007), some polarization occurs between redistrictings, and this trend is also present in the California state legislature. Because the time trend is clearly important, I modify the regression in equation 2 to include exible quartic time trends for both parties. Doing so allows me to examine discrete jumps in rollcall voting behavior due to changes in district composition. This is essentially a regression discontinuity approach, with all years receiving equal weight. The results are shown in columns (2) and (4) of Table 2. After controlling for these time trends, the regression results for both roll-call voting measures suggest that legislatively-drawn districts are indeed more polarizing. On average, the dierence in polarization is more than 10 CLCV points and nearly 9 CFL points. The results including time trends in Table 2 were estimated constraining the impact to be the same for all legislatively-drawn districts (similarly for all paneldrawn districts). Relaxing this assumption, and allowing the discrete change to be dierent for each redistricting, I then estimate the following regression equation: Score it = γ 0 + γ 1 Dem i + φ P + δ P Dem i + f(t) + g(t) + ɛ it (4) where Score it is legislator i's CLCV or CFL score in year t; P is a vector of period-specic dummy variables for the districts of the 1960s, 1970s, 1980s and 11

12 1990s redistrictings.; f (t) is a quartic time trend for Democrats and the associated regression coecients; and g(t) is a quartic time trend for Republicans and the associated regression coecients. 18 These estimates are in Table 5. In column (1), the regression in equation (4) was estimated with CFL as the dependent variable. Again, Democrats alternate in their behavior, becoming relatively more liberal under legislatively-drawn districts compared to the paneldrawn districts. Republicans again become steadily more conservative, slowing down only slightly during the panel-drawn districts of the 1990s. Column (3) of Table 5 shows the estimates of equation (4) with CLCV scores as the dependent variable. The results show a striking pattern: after controlling for party trends with a quartic time trend, the average Democrat votes more pro-environmentally after moving to legislatively-drawn districts, and relatively anti-environmentally after moving to panel-drawn districts. The dierence in behavior for the average Democrat is signicant except moving from the 1990s to 2000s. A similar pattern emerges for Republicans as in the previous regressions, where it appears that a time trend dominates. To test for increased polarization during legislatively-drawn periods, I simply subtract the expected Republican score in any period from the expected Democratic score, and run pairwise tests comparing this dierence from decadeto-decade. In Panel B of Table 5, changes in polarization are consistent with the polarization hypothesis. Moving from the rst panel-drawn districts to the legislatively-drawn districts of the 1980s, polarization is signicantly higher in the 1980s. Then moving from the 1980s to the 1990s, polarization decreases after controlling for the quartic time trends. Finally, though the point estimates suggest that the change is in the right direction, we cannot reject the hypothesis that polarization in the 1990s was the same as in the 2000s. 19 For both the CLCV and CFL regressions, roughly one-quarter of the total observations lie at the censoring points (0 and 100), with the greatest share in later years in the data. As an attempt to correct for this censoring, I also estimate a symmetrically-censored least squares regression (Powell, 1986) in columns (2) and (4). The point estimates dier slightly, but the same general pattern emerges. The dierence in decade-to-decade polarization in the CFL censored regression is signicant in all cases, but only at the 10% level when comparing the 1990s to the 2000s. In the censored CLCV regression, the dierence between the 1990s and 2000s remains insignicant. The results in Table 5 reinforce what was seen by examining the three-year average interest group scores in Table 1. Polarization in the legislature increases 18 The quartic time trend is simply constructed by taking the number of years since the rst observation for that interest group score. Specically, for Democrats, f(t) = ρ 1 t + ρ 2 t 2 + ρ 3 t 3 + ρ 4 t 4. For the CFL regressions, t = 1 corresponds to 1968, t = 2 corresponds to 1969, and so on. For the CLCV regressions, t = 1 corresponds to As a robustness check, counterfactual periods were generated to create four periods that overlapped redistricting periods to the midpoints on either side. I then interacted these counterfactual periods with Democrat and Republican dummies and ran regressions analagous to those in Table 5. The results did not yield a systematic pattern, which provides support to the claim that the correlations in Table 5 are indeed capturing discrete jumps in behavior due to new districts. 12

13 when moving to legislatively-drawn districts, and it decreases when moving back to panel-drawn districts. After controlling for time trends, the changes in polarization were as high as 24 points between the 1980s and 1990s in the CLCV case, and as high as 16 points between the 1970s and 1980s in the CFL case. Democrats, who exercised considerable control during the redistrictings, change behavior signicantly depending on who drew district lines, and their behavior largely determined these changes in polarization. The results are rather robust, and although the dierence is not signicant across all specications, the point estimates suggest that the direction of change is indeed consistent with the claim that legislative redistricting causes increased polarization in the legislature. 5 Conclusion A common claim in the media and by some politicians is that legislative redistricting is a major cause of polarization in the United States, and California has been at the heart of this debate. Because the process of redistricting has alternated between courts and legislatively-drawn districts since the 1960s in California, focusing on the state Senate and Assembly provides a nice quasi-experiment to test for a relationship between polarization and redistricting schemes. Importantly, because the redistricting method has alternated every decade in the last four decades, the policy switch in this quasi-experiment goes both ways and is repeated. Using legislative voting data, I test the so-called polarization hypothesis and nd support for the claim that the type of redistricting is, in fact, associated with changes in polarization. These ndings are consistent with the hypothesis that legislative redistricting increases polarization in the legislature. Other authors have tested the polarization hypothesis, focusing mostly on the US House of Representatives. These tests, often using only one redistricting period, rely on cross-state comparisons of redistricting methods for identication. Other studies, such as that by Masket, Winburn, and Wright, show that the polarization trend persists between redistrictings. I, too, nd a strong time trend that cannot be explained solely by redistricting; this time trend is prevalent for both parties, but I also nd discrete changes in behavior induced by the types of redistricting in place. Importantly, the trends reverse themselves when moving from legislatively-drawn districts to panel-drawn ones. This paper shows that legislators in both parties vote dierently on both environmental and labor issues if their districts are safe, in terms of voter registration, compared to competitive ones. The dierence is especially pronounced for Republicans, with estimated dierences in interest group scores ranging from around 9 to 19 points. The dierence between CFL scores for Democrats under safe districts compared to competitive ones ranges from 11 to 16 points. Using roll-call voting scores (CLCV and CFL) in regressions on indicator variables for redistricting type and party, I nd evidence that the type of redistricting is correlated with polarization. There is clearly a time trend in polarization in California (and elsewhere in the United States), and determining all of the causes of this trend would be a dicult endeavor. This research shows, 13

14 however, that redistricting plans have an impact on polarization. Even after controlling for that trend with a exible (quartic) time trend for each party, there is evidence that the type of redistricting matters for the degree of polarization. Democrats tend to take more conservative stances on environmental and labor issues when they represent panel-drawn districts, and when districts are legislatively-drawn, they take signicantly stronger pro-environmental (or pro-labor) stances. The movement of the party, on average, as well as changes in individual legislator's voting patterns, suggests that legislative redistricting allows legislators to take more extreme positions in their roll-call votes, which leads to increased polarization. These results are rather robust across specications. Contrary to ndings elsewhere in the literature, it appears that it does matter who draws the lines. The approach taken in this paper is a reduced-form one, and because I am looking at political data over the past four decades, there are many possible confounding factors. However, because the policy changes both directions, and does so twice, any omitted variables that could be biasing these results would need to coincide with this policy experiment. As discussed earlier in the paper, partisan control of the legislature, the governor's party, and other national trends may be driving polarization. The results presented in this paper are rather robust to inclusion of controls for these eects. In Tables 8-11 in the Appendix I present regression results controlling for these factors. This study has policy implications for those interested in redistricting reform, and I believe that other states (as well as California) can learn from this policy experiment. Previous research has indicated that political elites, such as state level legislators, tend to take more extreme positions on issues than the general population (Fiorina and Levendusky). It appears that legislative redistricting serves as a catalyst to enable even more polarizing behavior in the state legislature. If policymakers and voters are concerned about polarization in the state legislature, perhaps other redistricting methods, such as independent panels, could be considered. The impact of polarization on welfare is beyond the scope of this study (see Coate and Knight, 2007 for a discussion of socially optimal districting), but further research should be conducted to determine the impact of legislative polarization on legislative outcomes and overall welfare. References [1] Berry, William D., and Thomas M. Carsey Revised Candidate-Level Legislative Returns in the United States with Adjusted Candidate Names, [2] Basehart, Harry, and John Comer Partisan and Incumbent Eects in State Legislative Redistricting. Legislative Studies Quarterly 16(1): [3] Cain, Bruce E Assessing the Partisan Eects of Redistricting. The American Political Science Review 79(2):

15 [4] Cain, Bruce E. and Janet C. Campagna Predicting Partisan Redistricting Disputes. Legislative Studies Quarterly 12(2): [5] Canes-Wrone, Brandice; David W. Brady; and John F. Cogan Out of Step, Out of Oce: Electoral Accountability and House Members' Voting. The American Political Science Review 96(1): [6] Chay, Kenneth, and Michael Powell. Semiparametric Censored Regression Models, Journal of Economic Perspectives, Fall 2001: [7] Coate, Stephen, and Brian Knight. Socially Optimal Districting: a Theoretical and Empirical Exploration. Forthcoming, The Quarterly Journal of Economics. [8] Fiorina, Morris P., and Matthew S. Levendusky Disconnected: The Political Class vs. the People. In Red and Blue Nation: Characteristics and Causes of America's Polarized Politics, ed. Pietro S. Nivola and David W. Brady. Washington, DC: Brookings Institution Press. [9] Friedman, John N. and Richard T. Holden. Optimal Gerrymandering: Sometimes Pack, But Never Crack. Forthcoming, American Economic Review (2008). [10] Friedman, John N. and Richard T. Holden. The Rising Incumbent Reelection Rate: What's Gerrymandering Got to Do With It? Working Paper [11] Garrett, E Redistricting: another California revolution?, IRI Report , February 2005 available at [12] Glazer, Amihai, and Marc Robbins Congressional Responsiveness to Constituency Change. American Journal of Political Science 29: [13] Groseclose, Tim "A Model of Candidate Location When One Candidate Has a Valence Advantage" American Journal of Political Science. 45 (October): [14] Groseclose, Tim; Levitt, Steven D.; and James M. Snyder, Jr. Comparing Interest Group Scores across Time and Chambers: Adjusted ADA Scores for the U S Congress." American Political Science Review, 1999, 93(1), pp [15] Huang, Allison. Does Environmental Legislation Matter for Elections? Evidence from the US Senate. Working Paper, [16] Kousser, J. Morgan Estimating the Partisan Consequences of Redistricting Plans-Simply. Legislative Studies Quarterly 21(4): [17] Lacayo, Hank. October 5, Redrawing the Political Boundaries: Reform or Power Grab? San Francisco Chronicle, Open Forum. 15

16 [18] Lee, David S.; Enrico Moretti; and Matthew J. Butler. August, Do Voters Aect or Elect Policies? Evidence from the U.S. House. Quarterly Journal of Economics [19] LeVeaux-Sharpe, Christine Congressional Responsiveness to Induced Constituency Change: An Extension to the 1990s. Legislative Studies Quarterly 26(2): [20] Mann, Thomas E., and Bruce Cain Party Lines: Competition, Partisanship, and Congressional Redistricting. Washington, D.C.: Brookings Institution Press. [21] Masket, Seth., Winburn, Jonathan. and Gerald Wright. The Limits of the Gerrymander: Examining the Impact of Redistricting on Electoral Competition and Legislative Polarization. Paper presented at the annual meeting of the The Midwest Political Science Association, Palmer House Hilton, Chicago, Illinois, Apr 20, [22] McCarty, Nolan; Poole, Keith T.; and Howard Rosenthal. Does Gerrymandering Cause Polarization? Working Paper. October, [23] Report and Recommendations of the Special Masters on Reapportionment, Special Masters Archives, IGS Library, UC Berkeley. Electronic version available at igs.berkeley.edu/library/reapp/reapp90- report/index.html. [24] Stratmann, Thomas Congressional Voting over Legislative Careers: Shifting Positions and Changing Constraints. The American Political Science Review 94(3): [25] Tate, Katherine Black Opinion on the Legitimacy of Racial Redistricting and Minority-Majority Districts. The American Political Science Review 97(1):

17 6 Figures and Tables Figure 2: Nominal CLCV Scores over Time Average (nominal) CLCV scores are shown by party. 100 corresponds to a proenvironmental roll-call voting record. Vertical lines denote times where new districts take eect. Periods are labeled by redistricting type. 17

18 Figure 3: Average Margins of Victory by Districting Margin of victory is dened as the percentage point dierence between the winner and the second-place candidate. Uncontested seats are included in this calculation (where margin=100%); the general shape remains without these observations. Vertical bars denote times where new districts take eect. The districts in the 1970s and 1990s were drawn by panels of Special Masters, whereas the districts in the 1960s, 1980s and 2000s were drawn legislatively. 18

19 Table 1: Average Polarization Pre- and Post-Redistricting: Three-Year Averages Roll-Call Measure: CFL CLCV Democrat Republican Di. Democrat Republican Di. Pre-1970s Districts (0.63) (4.40) (4.44) Post-1970s Districts (1.38) (2.90) (3.21) (1.44) (2.68) (3.04) Pre-1980s Districts (1.33) (1.66) (2.12) (1.95) (2.12) (2.88) Post-1980s Districts (0.87) (1.90) (2.09) (1.15) (2.30) (2.56) Pre-1990s Districts (0.85) (1.75) (1.94) (1.04) (2.75) (2.94) Post-1990s Districts (1.40) (1.16) (1.81) (2.18) (1.15) (2.46) Pre-2000s Districts (0.83) (1.13) (1.40) (1.32) (2.59) (2.90) Post-2000s Districts (1.01) (0.50) (1.12) (1.53) (1.08) (1.87) F-Statistics for Pre- and Post-Redistricting Dierences in Polarization, by Redistricting: 1970s (Panel) F(1, 502) 0.31 F( 1, 513) 1980s (Legislative) *** 23.30*** 1990s (Panel) s (Legislative) 11.62*** In the CLCV regressions, there is only one year of observations for the pre-1970s districts. Three-year averages are from a single regression on the respective dummy variables, with errors clustered by legislator. The results are broken up by party for readability. F-tests were conducted for dierences in polarization before and after each redistricting, with *** indicating signicance at the 1% level. 19

20 Table 2: Impact of Seat Competitiveness on Roll-Call Voting Behavior Dependent Variable: All District Types Panel-Drawn Legislatively-Drawn CFL CLCV CFL CLCV CFL CLCV CFL CLCV (1) (2) (3) (4) (5) (6) (7) (8) Democrat 75.17*** 56.75*** 75.06*** 56.43*** 74.57*** 49.76*** 76.06*** 66.26*** (1.96) (2.34) (1.82) (2.35) (1.62) (2.88) (3.16) (2.42) Safe * (3.11) (3.31) (3.10) (3.43) (3.38) (3.87) (4.02) (3.57) Competitive * (4.40) (3.95) (3.91) (3.92) (3.28) (4.93) (7.81) (5.66) Safe*Dem 6.76** ** * 6.11 (3.30) (4.34) (3.31) (4.46) (3.62) (5.13) (4.43) (4.84) Comp*Dem -9.00* ** ** (4.82) (4.78) (4.40) (4.75) (4.24) (5.87) (7.99) (6.49) Intercept 15.97*** 20.83*** 29.83*** 27.34*** 15.37*** 23.46*** 16.79*** 17.42*** (1.80) (1.80) (3.10) (2.53) (1.24) (2.37) (3.03) (1.91) Year Fixed Eects No No Yes Yes No No No No Observations F-Stats: SafeDem=CompDem 11.54*** *** *** SafeRep=CompRep 5.53** 5.01** 6.68*** 5.03** * 6.27** 2.02 A legislator's seat is dened as safe here if the Democratic registration in her district, as a fraction of Democratic and Republican registration, is in the top quintile for Democrats or bottom quintile for Republicans. Competitive seats are dened for the corresponding bottom or top quintiles for Democrats and Republicans, respectively. The dependent variable in the odd-numbered columns is a legislator's CFL score, and the dependent variable in even columns is a legislator's CLCV score. Robust standard errors are in parentheses; *, **, and *** correspond to the 10%, 5%, and 1% levels of signicance, respectively. 20

21 Table 3: Impact of Redistricting Type on Legislative Behavior Dependent Variable: CFL CLCV (1) (2) (3) (4) Panel -3.05** ** (1.15) (1.51) (2.03) (1.66) Democrat 72.03*** 45.88*** 62.79*** 55.43*** (1.44) (4.05) (2.17) (7.43) Panel*Democrat *** -5.96*** *** (1.39) (1.73) (2.27) (2.24) Intercept 20.23*** 31.72*** 23.87*** 44.80*** (1.36) (3.01) (1.97) (6.62) Party Year Trends No Quartic No Quartic Clusters N 3,910 3,910 3,708 3,708 The dependent variable in columns (1) and (2) is the CFL score for members of California's Assembly and state Senate for the years The dependent variable in columns (3) and (4) is the CLCV score for members of California's Assembly and state Senate for the years , excluding observations with scores based on fewer than 10 roll-call votes. All regressions are pooled with errors clustered by legislator. Standard errors are in parentheses. ***, **, and * correspond to the 1%, 5%, and 10% levels of signicance, respectively. 21

22 Table 4: Impact of Redistricting Type on Legislative Behavior Dependent Variable CFL CLCV (1) (2) (3) (4) (5) (6) 1960 (leg.) Dem *** 94.61*** (1.05) (0.93) 1970 (panel) Dem *** 84.36*** 72.08*** 78.63*** (1.17) (0.66) (1.51) (0.78) 1980 (leg.) Dem *** 93.37*** 85.81*** 90.28*** (0.67) (0.58) (1.11) (0.62) 1990 (panel) Dem *** 89.28*** 87.01*** 80.84*** (0.78) (0.66) (1.31) (0.67) 2000 (leg.) Dem *** 87.30*** 88.64*** 78.83*** (0.95) (1.09) (1.55) (1.21) 1960 (leg.) Rep *** 28.71*** (2.52) (1.58) 1970 (panel) Rep *** 20.62*** 38.81*** 32.04*** (1.82) (1.13) (1.83) (1.15) 1980 (leg.) Rep *** 18.69*** 31.65*** 30.32*** (1.76) (0.98) (2.25) (0.83) 1990 (panel) Rep *** 15.97*** 12.59*** 15.83*** (0.81) (1.10) (1.59) (0.84) 2000 (leg.) Rep. 5.93*** 13.61*** 4.76*** 11.43*** (0.45) (1.91) (0.96) (1.64) Legislator Fixed Eects No Yes Yes No Yes Yes Clusters/Groups Observations 4,027 2,418 1,609 3,708 2,251 1,457 The coecients and standard errors are calculated using the results from estimating equation 3. The dependent variable in columns (1)-(3) is the CFL score for individual legislators, and in (4)-(6) the dependent variable is the CLCV score. The regressions in (1) and (4) were run using pooled OLS with clustered standard errors; the regressions in (2), (3), (5), and (6) include legislator xed eects. Standard errors are in parentheses; in columns (1) and (4), errors are clustered by legislator. *, **, and *** correspond to the 10%, 5%, and 1% levels of signicance, respectively. 22

23 Table 5: Impact of Redistricting Type on Legislative Behavior: Controlling for Time Trends Panel A: Regression Discontinuity Estimates Panel B: Implied Polarization Dep. Variable: CFL CLCV District Type CFL CLCV (1) (2) (3) (4) (5) (6) (7) (8) Democrat 56.66*** 53.13*** 40.32*** 37.21*** 1960s 45.26*** 37.58*** (7.54) (7.09) (8.95) (7.58) (4.07) (2.84) Legis1960*Dem * ** 1970s (Panel) 40.26*** 31.63*** 45.98*** 49.91*** (6.17) (6.46) (6.56) (4.53) (6.89) (4.72) Panel1970*Dem *** *** s 54.88*** 48.31*** 53.41*** 53.80*** (5.22) (5.32) (5.39) (6.69) (6.67) (5.56) (7.82) (5.67) Legis1980*Dem ** 16.59** 1990s (Panel) 50.36*** 44.89*** 36.72*** 30.29*** (3.59) (4.14) (4.89) (5.39) (7.32) (6.40) (8.36) (6.56) Panel1990*Dem -6.30*** -8.25*** s 56.66*** 53.13*** 40.32*** 37.21*** (2.04) (3.15) (2.46) (4.95) (7.54) (7.09) (8.95) (7.58) Legis *** 18.24*** (5.07) (5.08) F-Stats for Dierences in Polarization: Panel * 0.59 Ho : (4.20) (4.18) (4.09) (5.58) 1960s=1970s ** Legis ** s=1980s 9.13*** 32.16*** 5.28** 1.54 (2.66) (3.26) (3.49) (4.53) 1980s=1990s *** 61.96*** Panel *** 5.25** 3.97** s=2000s 13.03*** 6.83*** (1.25) (2.50) (1.76) (4.37) Intercept 15.74*** 11.76** 34.61*** 37.56*** (5.95) (5.49) (7.41) (6.11) Estimator OLS SCLS OLS SCLS Clusters Observations 4,027 4,027 3,708 3,570 The dependent variable in columns (1) and (2) is the CFL score for individual legislators, and in (3) and (4) it is the legislators CLCV score. The excluded category is Legis2000. Panel A columns (1) and (3) were estimated using pooled OLS; columns (2) and (4) was estimated with symmetrically-censored least squares. Legis1960 and Legis1980 are dummy variables for periods under the districts drawn legislatively in the 1960s and 1980s, respectively. Panel1970 and Panel1990 are dummy variables corresponding to districts drawn by the Special Masters in the 1970s and 1990s, respectively. All regressions include quartic party time trends for each party. Standard errors are clustered by legislator, in parentheses, where *, **, and *** correspond to the 10%, 5%, and 1% levels of signicance, respectively. 23

24 7 Appendix Figure 4: Number of Moderate Party Members Over Time The number of Moderate Democrats (Republicans) is the number in that year and chamber whose CLCV scores are less (greater) than the most liberal (conservative) decile of the other party. 24

25 Table 6: Regression of Margin of Victory on Party-Decade Interactions Period / Party Margin of Victory 1970s Democrats 6.484** (2.83) 1970s Republicans (2.88) 1980s Democrats 10.35*** (2.95) 1980s Republicans 6.61** (2.92) 1990s Democrats 7.59** (2.84) 1990s Republicans -4.51** (2.23) 2000s Democrats 13.67*** (3.37) 2000s Republicans 29.38*** (2.073) The margins of victory for each period were calculated from a regression of margin of victory on a constant, a Democrat dummy variable, period-specic dummies, and interaction terms. The data span elections from , and observations include both Assembly and Senate for the California state legislature. There are 1,503 observations. The regression is pooled OLS with errors clustered by legislator. *, **, and *** correspond to the 10%, 5%, and 1% levels of signicance, respectively. 25

26 Figure 5: Kernel Density Estimates of CLCV Scores Pre- and Post-1980s Redistricting 26