Evidence on the importance of spatial voting models in presidential nominations and elections

Transcription

1 Public Choice (2005) 123: DOI: /s C Springer 2005 Evidence on the importance of spatial voting models in presidential nominations and elections LAWRENCE W. KENNY 1, &BABAK LOTFINIA 2 1 Department of Economics, University of Florida, Gainesville, Florida , U.S.A.; 2 Bureau of Economic and Business Research, University of Florida, Gainesville, Florida , U.S.A. ( Author for correspondence: larry.kenny@cba.ufl.edu) Accepted 6 March 2004 Abstract. ADA scores and Nominate scores are used for the first time to examine the influence of spatial voting records on which candidate wins the party s presidential nomination and on which nominee wins the general election. We find that the most conservative Republican candidate and moderately liberal Democrats were most likely to win their party s nomination. For general elections we find that the candidate s spatial record has nearly as much impact on the outcome as economic growth, which has been the focus of most past empirical research. The nominee whose voting record is more moderate is more likely to be elected. Introduction In the spatial voting literature, the median voter model asserts that both candidates for an office select the same platform, that which is preferred by the median voter. 1 On the other hand, in the divergent platform or two constituencies model, candidates offer distinctly different platforms that straddle the median voter s position; thus the more conservative candidate receives votes from conservatives, and liberals support the more liberal candidate. 2 There is strong evidence that spatial voting plays an important role in determining the outcome of congressional elections. Legislators whose voting records are far from what would be predicted for their district are less likely to get reelected. 3 Research on U.S. Senate elections has found that the divergent platform (or two constituency) model outperforms the median voter model empirically. Democratic senators almost always have a more liberal voting record than Republican senators from the same state, who represent the same geographic constituency. 4 These differences have origins in the party nomination process. House members whose record is far from the party position in their state are less likely to run for the Senate and to be selected their party s standard bearer (Francis, Kenny, Morton, & Schmidt, 1994). Recognizing this, House members who later run for the Senate edge steadily toward the winning state party position (Francis & Kenny, 1996). Furthermore, the median voter s preferred position does not appear to be ideal for winning the general election. Senators who deviate from their party s position are less likely to get reelected, or even to seek reelection, than those whose records are close to the appropriate

2 440 state party platform. 5 More importantly, Schmidt, Kenny, and Morton (1996) found that distance to the state party position better explains reelection success than distance to the median voter s position. The relevance of spatial voting models often is illustrated in class with examples of voters punishing presidential candidates for taking the wrong platform. Goldwater lost in 1964 because he was too conservative. Mondale was trounced in 1984 for being too liberal. Thus it is surprising that the statistical literature examining the factors affecting the outcomes of multiple presidential elections has focused almost exclusively on whether the incumbent party is rewarded for economic performance or punished for unpopular wars. 6 Afew studies, however, have relied on perceptions about candidate spatial positions to explain the outcome of multiple presidential elections. 7 Zaller (1998) used a research assistant s ratings on a seven-point conservative liberal scale of presidential nominees from 1948 to 1996 and found that the more moderate nominee did better, ceteris paribus. Bartels and Zaller s (2001) estimates of the nominees spatial positions is based on expert ratings of candidates in elections between 1948 and and survey data on candidates between 1984 and 1996 (from better informed respondents in American National Election Studies surveys...). The research assistant s and experts estimates of candidate positions are, however, subjective and may be biased with the ex post rationalization that often accompanies hindsight. And although Francis et al. (1994) have examined the role of spatial voting records on success in becoming a nominee for the U.S. Senate, no study appears to have examined how candidates spatial voting records affect who wins the presidential nomination. Several ideological measures are available for evaluating the performance of these spatial models in explaining which candidate obtains the party s presidential nomination and which nominee wins the general election. Ratings produced by Americans for Democratic Action (ADA) place the voting records of members of Congress on the crucial conservative liberal dimension that seems to explain most political conflict, 8 and comparable scores for presidents in have been constructed by Zupan (1992). Groseclose, Levitt, and Snyder (1999) have transformed the raw House and Senate ADA scores to make them comparable across chambers and over time and claim that their indexed ADA scores should perform better than the raw ADA scores produced by the Americans for Democratic Action. Poole and Rosenthal (1997) tout their Nominate scores as preferable to interest group ratings such as the ADA scores for a variety of reasons. We use their common space coordinates, which also are comparable across chambers and over time, for members of Congress and for Eisenhower and subsequent presidents. This paper makes numerous contributions to the literature. We extend Zupan s presidential ADA score series of 11 years to Then, we use these three spatial measures of ideology to ascertain whether the candidates

3 441 actual or inferred spatial voting records matter in determining the winners of the presidential nominations and presidential elections, and, if so, whether spatial competition is better described by the median voter model or the divergent platform model. Ours is the first study to use the candidate s voting record in Congress to explain success in seeking the presidency. Our use of the voting record, as opposed to an expert s estimate of the candidate s spatial position, provides an important tie to the extensive literature as cited, which examines the effect of the voting record on subsequent success in getting reelected or in seeking higher office. 9 We find that the candidates spatial records play an important role in selecting their party s nominee. Democrats with a centrist or far-left voting record are less likely to get nominated than are candidates who are only moderately liberal. Similarly moderate Republicans are less likely to be nominated than candidates with a very conservative voting record. Thus the party nominees in the general election have spatially distinct platforms, which supports the divergent platform model. Spatial records also are very important in determining the winner of the general election. The median voter and divergent models fare equally well. In half the regressions, the median voter model provides the better fit; in these nominees who were closer to the median voter were more likely to win. In the other regressions, the divergent platform better explained the outcome. The nominee who was closer to the optimal party position was more successful. Note, however, that the optimal party positions for winning the general election were not that far apart. Our empirical analysis also provides evidence on the relative performance of the three spatial voting record variables in explaining candidate success. Measuring the Ideological Record of Presidents and Presidential Candidates Zupan (1992) developed an ingenious way to determine U.S. presidents spatial positions. The widely used Americans for Democratic Action rating equals the percentage of times the legislator supported a liberal position in key roll-call votes selected by the ADA. For about half the House or Senate votes selected by the ADA, the Congressional Quarterly Almanac has identified whether the president took a personal position on the vote. Using these votes, Zupan determined for each year from 1947 to 1989 the percentage of times the president s position on these House and Senate votes agreed with the ADA s liberal position. We have extended this series through Since Zupan s presidential ADA scores and the congressional ADA scores are based on roughly the same votes, these scores are sufficiently comparable to allow the use of both to examine the role that an ideological record

4 442 plays in determining mobility from Congress, and even from the governor s mansion, to the presidency and the president s subsequent success in getting reelected. 10 The lack of a presidential position on a bill is not counted in Zupan s index as a vote for or against the liberal position. In contrast, the raw ratings compiled by the ADA for Congress divide the votes favoring liberal positions by the total bills considered, thus counting an abstention as anti-liberal. Since Peltzman (1984), Francis et al. (1994), and others feel it more appropriate to treat abstentions as neither a vote for the ADA s position nor a vote against it, they created modified ADA scores, which equal the percentage of votes cast that were cast for liberal positions. 11 The modified ADA scores used in the congressional literature and in this paper deal with abstentions in the same fashion as Zupan s presidential ADA scores. The presidential modified ADA scores are reported in Table 1 and range from 0 to 100. Cumulative scores based on a simple average of each year s ADA scores are also given. It is clear that Eisenhower s career ADA score of 50 places him far from other presidents from either party. Excluding Eisenhower, the average score for Republicans was approximately 8. The average Democratic score was approximately 89. Groseclose, Levitt, and Snyder s conversion formulas were utilized on the modified ADA scores to create indexed ADA scores that are comparable over time and across chambers, with the 1980 House being the base year and chamber. Although in the 1980 House the indexed ADA scores lie between 0 and 100, the conversion formulas allow for other indexed scores to lie outside this range; this is due to some senators and representatives being more liberal or more conservative than the members of the 1980 House. Thus the indexed ADA scores between 1947 and 1999 range from 27 to 129 in the House and fall between 31 and 118 in the Senate. Since the president s ADA score is based on both House and Senate votes used by the ADA, the indexed ADA score for the president is the average of the indexed ADA score calculated using the House formula and the indexed ADA score based on the Senate formula. With two exceptions, the average ADA score over the first 3 years of the president s term is used to gauge the president s ideological record as he seeks reelection. In light of their limited record as president, Ford s votes over his 3 years in office and Johnson s ADA scores for were used for the 1976 and 1964 elections, respectively. 12 For presidential candidates who had served in Congress, we use their ADA scores averaged over the 6 years prior to their bid for president as an estimate of their expected ideological position as president. 13 Forexample, Bob Dole s average ADA score for was used for his 1996 presidential bid. Five presidents elected after Eisenhower had served in Congress. Their congressional indexed ADA scores are fairly close to their subsequent scores as

5 443 Table 1. Presidential modified ADA scores Republican ADA score Democrat ADA score Year President Year Cum. Mean Year President Year Cum. Mean 1953 Eisenhower Truman Eisenhower Truman Eisenhower Truman Eisenhower Truman Eisenhower Truman Eisenhower Truman Eisenhower Kennedy Eisenhower Kennedy Nixon Kennedy Nixon Johnson Nixon Johnson Nixon Johnson Nixon Johnson Nixon Johnson Ford Carter Ford Carter Ford Carter Reagan Carter Reagan Clinton Reagan Clinton Reagan Clinton Reagan Clinton Reagan Clinton Reagan Clinton Reagan Clinton Bush G.H Clinton Bush G.H Bush G.H Bush G.H All Presidents ave. weighted by years ave. weighted by years ave. weighted by Pres ave. weighted by Pres Not including Eisenhower ave. weighted by years 7.95 ave. weighted by Pres Source: : Zupan (1992) : compiled by Lotfinia.

6 444 president. The absolute differences are: Kennedy (2), Ford (6), Nixon (13), G.H. Bush (14), Johnson (25). Gauging the expected ideological record for a presidential candidate who has not served in Congress is more problematic. Since governors and senators run in statewide races, they should be constrained by the same state party position. Accordingly, with one exception we use the average (modified or indexed) ADA scores for a state s senators of the same party over the 11-year period spanning from 5 years before the election to 5 years after the election as the predicted ADA score for governors seeking to be president. 14 As expected, the prediction error is greater for the three governors who became president Reagan (7), Clinton (14), Carter (43) than for the presidents who had served in Congress. Poole and Rosenthal have produced various Nominate scores that measure the spatial location of members of Congress. The common space Nominate scores were used because these are comparable across chambers and over time. Nominate scores for presidents are based on the CQ Presidential Support Roll Calls. There is one lifetime score for each member of Congress and for each president, starting with Eisenhower. For governors running for president, average Nominate scores for senators in their party and state over the five (2 year) Congresses centered around the election are used to produce an expected spatial position. The first dimension of the Nominate scores is highly correlated with the ADA scores. To facilitate comparison with the ADA measures, this first dimension is used and has been rescaled so that the most liberal position in the congressional data ( 0.806) corresponds to 100 and the most conservative position (0.896) is mapped to 0. Selecting the Party s Nominee Now we seek to ascertain whether the candidate s voting record plays any role in whether he is ultimately selected as his party s nominee for president. This reduced form approach explaining the outcome of the nomination process does not delve into the determination of each s state s delegates, which is beyond the scope of this paper. Instead, we ask whether moderates on average over the years are less likely than liberals to become the Democratic party s nominee. Similarly, what is the optimal voting record for garnering the Republican nomination? And does deviation from the winning party position have much impact on the likelihood of being selected? As Cooper and Munger (2000: 338) note, theory does not tell us much about who enters and wins nomination races, which often have three or more candidates. To gain some insight into what might happen, they describe the outcomes of simulated races in which 10 candidates are drawn randomly from

7 445 a normal distribution centered at the party s median voter. In this setting, Cooper and Munger (2000: 346) find a high variance in the winner s position and that the distribution of the winner s position is if anything... multimodal. Because the ADA began evaluating legislators records in 1947, we start with the 13 presidential elections from 1952 to Our sample excludes eight nominations in which the incumbent president was seeking reelection (Eisenhower, 1956; Johnson, 1964; Nixon, 1972; Ford, 1976; Carter, 1980; Reagan, 1984; Bush, 1992; Clinton, 1996) and four nominations in which a sitting vice president ran for office (Nixon, 1960; Humphrey, 1968; Bush, 1988; Gore, 2000), since these candidates were always successful. The 1952 Republican race was tossed out because the major candidate and nominee, Eisenhower, had never held elective office and thus had no predicted ADA score; indeed both parties courted this war hero. Our list of primary candidates in the remaining 13 primaries includes all candidates who ran in at least two primaries, received at least 1% of the total vote across all primaries, and had been governor or a member of Congress. The first two restrictions limit the list of candidates to major candidates. 15 The third requirement limits our analysis to candidates whose ideological position is measurable; this is not exceedingly restrictive, since Eisenhower was the only major-party nominee in the elections not to have been a governor or member of Congress. Under these criteria, there were 17 Republican bids in five nomination races (1964, 1968, 1980, 1996, and 2000) and 38 bids by Democrats in eight nomination races (1952, 1956, 1960, 1972, 1976, 1984, 1988, and 1992). 16 The small number of presidential candidates, particularly Republicans, will limit the number of hypotheses that can be tested with these data. Tables 2 and 3 report estimates of the impact of the candidate s ideological record and whether the candidate was a governor (GOVERNOR = 1) in determining success in obtaining the party nomination. Linear probability regressions explaining whether the candidate won the nomination are found in Table 2; these regressions utilize Huber-White-sandwich robust standard errors to deal with the well-known heteroskedasticity problem in this setting. 17 Table 3 reports regressions, which also employ robust standard errors, explaining the candidate s share of the vote obtained by these candidates in the presidential primaries. Candidates who are farther from their party position are hypothesized to be less likely to win the primary. DISTANCE TO PARTY PRIMARY POSI- TION equals the distance of the candidate s ideological score (I c ) from the winning party primary position (I pp ), i.e. I c I pp.toascertain the winning party primary position, we searched over integer values for the party primary position I pp that produced the best fit, by minimizing the root mean squared

8 446 Table 2. Regression estimates of effect of distance from party position on winning presidential primaries (absolute t-statistic in parentheses, based on robust standard errors) Republican (n = 17) Democrat (n = 38) Modified Indexed Rescaled Modified Indexed Rescaled Modified Indexed Rescaled Modified Indexed Rescaled ADA ADA nominate ADA ADA nominate ADA ADA nominate ADA ADA nominate Measure of position (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) DISTANCE TO PARTY PRIMARY POSITION (2.78) (3.15) (2.94) (2.78) (2.65) (1.89) RELATIVE DISTANCE TO PARTY PRIMARY POS. (2.93) (3.37) (3.14) (3.03) (2.39) (1.80) GOVERNOR (0.10) (0.04) (0.38) (0.17) (0.13) (0.44) (3.00) (3.02) (3.00) (3.12) (2.94) (2.77) Intercept (2.81) (2.99) (3.60) (2.89) (3.02) (3.40) (2.54) (2.93) (2.35) (2.64) (2.42) (2.30) R Root mean sq. error Overall significance PARTY PRIMARY POSITION

9 447 Table 3. Regression estimates of effect of distance from party position on vote share in presidential primaries (absolute t-statistic in parentheses, based on robust standard errors) Republican (n = 17) Democrat (n = 38) Modified Indexed Rescaled Modified Indexed Rescaled Modified Indexed Rescaled Modified Indexed Rescaled ADA ADA nominate ADA ADA nominate ADA ADA nominate ADA ADA nominate Measure of position (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) DISTANCE TO PARTY PRIMARY POSITION (2.85) (3.06) (3.30) (2.61) (2.64) (1.80) RELATIVE DISTANCE TO PARTY PRIMARY POS. (2.87) (3.22) (2.97) (2.53) (2.28) (1.93) GOVERNOR (0.40) (0.28) (0.25) (0.47) (0.38) (0.17) (2.07) (2.19) (1.81) (2.06) (1.89) (1.94) Intercept (3.99) (3.97) (4.25) (4.21) (4.18) (4.54) (4.42) (4.27) (4.31) (4.40) (3.93) (4.18) R Root mean sq. error Overall significance PARTY PRIMARY POSITION

10 448 error in a regression. The estimates of I pp are labeled in these tables as PARTY PRIMARY POSITION. This was done for each of the three ideological measures: modified ADA score, indexed ADA score, and rescaled Nominate score. A candidate needs only to be closer than the other candidates in that primary to the winning party position to be the most desirable choice. RELATIVE DISTANCE TO PARTY POS reflects this realization. It equals the candidate s distance from the winning party primary position relative to that of the closest candidate in the primary. That is, it equals I c I pp I closest I pp. This variable equals 0 for the closest candidate in each primary. A search also produced an estimate of the optimal party primary position for this variable, as used with a particular ideological measure. Accordingly, each of the two tables has 12 regressions, representing estimates for two parties, three ideological measures, and two distance variables. Test statistics of overall significance indicate that the regressions in Tables 2 and 3 are significant at the 5% level in all but two instances. The only exceptions occur when rescaled Nominate scores are used to explain the share of votes garnered in Democratic primaries. The estimated optimal party positions are revealing. For Republicans, the most conservative position (among candidates) was almost always the ideal position. On the other hand, the Democratic party clearly is not choosing its most liberal candidate. 18 For each ideological measure there are several Democratic candidates with scores that are at least 5% higher than the typical estimate of the winning party primary position. Candidates need only be closer to the party position than their rivals for the nomination to be preferred over their opponents. Accordingly, relative distance (RELATIVE DISTANCE TO PARTY PRIMARY POS) should better explain who wins the nomination than distance (DISTANCE TO PARTY PRIMARY POSITION). This hypothesis is supported in every comparison for the Republican nominations and in half of the comparisons for the Democratic nominations. For each party, there is evidence that those who are farther from the optimal party primary position are less likely to win the party s nomination. All 24 DISTANCE TO PARTY PRIMARY POSITION or RELATIVE DIS- TANCE TO PARTY PRIMARY POS coefficients in Tables 2 and 3 are negative and thus consistent with this hypothesis. These results, on the other hand, are not consistent with the lack of pattern in the winning position in Cooper and Munger s (2000) simulations of 10 candidate races with random entry. For the Republican primaries, all the regression coefficients are significantly negative at the 0.8% level or better, under a one-tailed test. A one standard deviation rise in distance is estimated to lower the probability of winning

11 449 Table 4. Explaining who wins the democratic primaries: frequency analysis (using modified ADA scores, PARTY PRIMARY POSITION = 92) DISTANCE TO PARTY OUTCOME PRIMARY POSITION WIN LOSE % WIN Toward Liberal Positions Toward Moderate Positions Total 8 30 the nomination by and to reduce the candidate s share of the primary vote by The classification is perfect if the measure of relative distance from the winning Republican position of 3 is used in conjunction with the indexed ADA scores. All five Republican candidates with a relative position of 0 (i.e., the closest candidates to the party primary position of 3) won the nomination, and all 12 candidates who were 3 65 points farther away lost the nomination. A frequency distribution showing how success rates for Democratic candidates depend on the distance from the winning party primary position, based on modified ADA scores, is found in Table 4. The fraction winning the nomination is lower for those who are more moderate or more liberal than the winning Democratic primary position (92), but this relationship is not nearly as striking as that reported for Republicans. The regression results are consistent with this observation. In the Democratic primaries, under a one-tailed test eight coefficients are significant at the 1.5% level or better and four are significant at the % levels. The distance measures based on the Nominate scores are consistently the least significant. Deviating from the optimal party primary platform is less costly for Democrats than for Republicans. For Democrats, a one standard deviation rise in distance causes the probability of winning to fall by and results in a drop in the share of primary votes received. Governors may have an advantage over other presidential candidates because they have experience running a large government, which is valuable to being an effective president. Consistent with this reasoning, 20 of the 24 coefficients on GOVERNOR are positive. They are significant, however, only in the Democrat regressions. For Democrats, governors are estimated to have a higher probability of winning the nomination. Governors also obtain a larger share of the primary vote in Democratic primaries.

12 450 The General Election A variety of variables have been used to explain the success of the incumbent party in the November presidential elections. These measure economic performance, involvement in unpopular wars, candidate ideology, incumbent party fatigue, president s popularity, voter evaluations of the incumbent party s capabilities, primary performance of the incumbent party s nominee, voter support for the incumbent party s nominee (early in the campaign), and so on. 19 The president s popularity and the esteem with which the incumbent party is held undoubtedly reflect the incumbent party s success in bringing about economic prosperity and avoiding unpopular wars. The primary success and initial popularity as president of the incumbent party s nominee also may reflect the ideological position he has taken. Thus, we focus on the more fundamental factors behind the incumbent party s success in the general elections: economic performance, involvement in unpopular wars, and candidate ideology. Bartels and Zaller (2001) estimated 48 regressions using various measures of these three factors and incumbent party fatigue and found (2001: 11) that (Hibbs, 2000) Bread and Peace model is by far the best of the 48 models. Accordingly, we use Hibbs measures of economic performance ( bread ) and peace, supplemented by our novel variables capturing the parties relative spatial positions. Since we have no basis for assigning Eisenhower a predicted ideological score in 1952, the empirical analysis covers the 12 elections from 1956 to Let UNPOPULAR WAR be a dummy variable that equals 1 in 1968 and equals 0 otherwise. Past economic performance is discounted at a rate that is determined by the data. The economic record of the incumbent party for the 15 quarters from the second quarter in the first year in office to the last quarter in the election year is summarized by DISCOUNTED GROWTH t = g t + b g t 1 + b 2 g t 2 + +b 14 g t 14 (1) where g t equals the quarterly growth rate in real personal income per capita and b is the quarterly discount factor that is to be estimated. To this basic Bread and Peace specification, we add the relative closeness of the candidates ideological positions to a winning general election position. Our estimates of the nominees ideological positions are based on the nominees actual voting record, in contrast to earlier research that has relied on experts estimates of the nominees policy positions. Under the median voter model, the median voter s position is the winning general election position. That is, the candidate closer to the median voter s position is more likely to

13 451 win the election, other things equal. Let DIFFERENCE IN DISTANCES TO MEDIAN VOTER = incumbent party s distance from median voter s position challenger party s distance from median voter s position. If the voter s utility is symmetric and the median voter is indifferent between the two party positions, then the median voter s position is the midpoint of the two party positions. One estimate of the party position is the average voting record of the party s presidents, which is reported in panel A of Table The midpoint of the two presidential party averages is indicated Table 5. Party and median voter positions Modified Indexed Rescaled President ADA ADA nominate Panel A: presidents ideological ratings Republicans Eisenhower Nixon Ford Reagan Bush G.H Average Average without Ike Democrats Truman NA Kennedy Johnson Carter Clinton Average Panel B: estimating median voter based on presidents record Republicans a Median voter Democrats Panel C: estimating median voter based on median value of winning party primary positions Republicans Median voter Democrats a Based on mean without Eisenhower.

14 452 as Median Voter in panel B for each of the three ideological measures. The other estimate of the median voter s position is based on the winning party primary positions estimated in Section 3. Panel C lists the median of the winning party primary position estimates for each party and the resulting median voter s position. The estimates for the median voter s position in panels B and C are remarkably close for the modified ADA and Nominate measures. We use the average of the panel B and panel C median voter estimates as the median voter s position in the empirical analysis in this section: 47 for modified ADA, 38 for indexed ADA, 52 for rescaled Nominate. Does the nominee who is closer to the median voter fare better in the general election? Alternatively, is the nominee who is closer to a party position more likely to win the general election? We test this version of the two constituency hypothesis by requiring the parties to be equidistant from the median voter s position and searching for the distance to the median voter that produces the best fit. Table 6 reports regressions explaining the share of the two major party vote received by the incumbent party, which are similar to those found in the literature. Linear probability regressions explaining whether the incumbent party won the general election are found in Table 7. To be consistent with the vote share variable, Gore is coded as winning the 2000 election. Regressions based on the median voter model, in which the candidate closer to the median voter wins, are shown in the top panels of Tables 6 and 7. For half the specifications, we get a better fit allowing the parties to have different winning general election positions that straddle the median voter s position. These divergent platform regressions are reported in the bottom panels in Tables 6 and 7. Half the regressions employ UNPOPULAR WAR. UNPOPULAR WAR has significantly negative coefficients in all of the eight regressions in which it is used. As a result of Vietnam, the incumbent Democratic party is estimated to have lost.023 to.045 of the two-party vote and experienced a.686 to.765 lower probability of winning the election. The literature generally has focused on the role of economic growth in explaining the success of the incumbent party in presidential elections. We also find that the incumbent party is rewarded for rapid growth of the economy. All but two of the 18 coefficients for DISCOUNTED GROWTH are significantly positive. A one standard deviation rise in this variable is estimated to lead to a.033 to.043 increase in the incumbent s vote share and a.217 to.302 higher probability of winning the election. The quarterly discount factor (b) applied to prior economic growth was defined in Equation (1). The estimates of b, which are obtained through another search procedure, average 0.89 and range from 0.83 to To see what these discount factors imply, consider the average value for b. Economic growth four quarters (i.e., 1 year) before the election receives a weight equal

15 453 Table 6. Regression estimates of effect of distance from optimal position on vote share in presidential general elections ( ); dependent variable = incumbent vote share in two party vote (absolute t-statistic in parentheses, based on robust standard errors) Modified Indexed Rescaled Modified Indexed Rescaled Measure of position ADA ADA nominate ADA ADA nominate Median voter result DIFFERENCE IN DISTANCES TO MEDIAN VOTER (3.21) (2.56) (2.91) (3.80) (3.39) (4.59) UNPOPULAR WAR (2.78) (3.38) (1.96) DISCOUNTED GROWTH (5.13) (3.78) (6.73) (5.27) (3.35) (6.48) Intercept (19.9) (20.8) (24.6) (22.0) (24.2) (27.3) R Root mean sq. error Median voter position Qrtly discount factor Annual discount rate Divergent platform result (if better fit than median voter) DIFFERENCE IN DISTANCES TO PARTY POSITION UNPOPULAR WAR (1.88) DISCOUNTED GROWTH (3.02) (4.51) (4.50) (5.11) (5.31) (6.50) Intercept (20.3) (22.6) (27.3) R Root mean sq. error Republican position Democratic position Qrtly discount factor Annual discount rate to (= ). This is equivalent to voters applying an annual rate of discount of 59% to prior economic growth

16 454 Table 7. Linear probability regression estimates of effect of distance from optimal position on winning presidential general elections ( ); dependent variable = incumbent victory dummy (absolute t-statistic in parentheses, based on robust standard errors) Modified Indexed Rescaled Modified Indexed Rescaled Measure of position ADA ADA nominate ADA ADA nominate Median voter result DIFFERENCE IN DISTANCES TO MEDIAN VOTER (4.01) (3.74) (1.83) (3.94) (3.61) (3.14) UNPOPULAR WAR (4.89) (4.65) (2.92) DISCOUNTED GROWTH (3.26) (2.61) (2.99) (2.57) (0.88) (2.02) Intercept (0.13) (0.43) (0.04) (0.23) (0.95) (0.56) R Root mean sq. Error Median voter position Qrtly discount factor Annual discount rate Divergent platform results (if better fit than median voter) DIFFERENCE IN DISTANCES TO PARTY POSITION (3.53) (4.80) (3.86) UNPOPULAR WAR (4.29) DISCOUNTED GROWTH (2.38) (2.34) (0.72) Intercept (0.48) (0.13) (1.02) R Root mean sq. error Republican position Democratic position Qrtly discount factor Annual discount rate This is higher than both the 14% average return on the 5000 most active U.S. stocks for and Hibbs (2000) estimate of the annual discount rate of 23% on prior economic growth in presidential elections (based on b = 0.95) but is smaller than the 79% discount rate that Figlio (2000) estimated voters applied to prior indexed ADA scores for senators. Tables 6 and 7 report the annual rate of discount for each regression, which varies from 0 to 111%.

17 455 There is very little prior evidence on the role that nominee ideology plays in studies of multiple presidential elections, and there is no evidence on the effect of the nominee s voting record. In half the regressions in Tables 6 and 7, the median voter model fits the data best. That is, the nominee who is closer to the median voter s position is more likely to win the election. In the remaining half of the regressions, the nominee who is closer to his estimated winning party position is more likely to win. The winning general election party positions are 4 to 18 points apart, averaging an 11-point difference. These party differences are much smaller than the point spread that separates the average ideological records for Democrat and Republican presidents (see Table 5). Nevertheless, our estimates of the winning general election Democratic position (e.g., 45 in indexed ADA regression 11 in Table 7) are to the left of Carter in 1976 (e.g., 39.5), and our estimates of the winning general election Republican position (e.g., 31) are to the right of Eisenhower (e.g., 37.5); that is, both candidates were more moderate than was optimal in adivergent platform setting in the general election. The winning position equals the median voter s position under the median voter model and equals the winning party position under the divergent platform model. An increase in the incumbent party nominee s distance from the winning position, relative to the challenger party nominee s distance, is expected to cause the incumbent party to fare worse in the election. All 18 coefficients on the DIFFERENCE IN DISTANCES variables in regression Tables 6 and 7 have the predicted negative sign, and all but one are highly significant. 21 A one standard deviation rise in these variables is estimated to result in a decrease in the incumbent party s share of major party votes and a fall in the probability of an incumbent party victory. The importance of relative distance can be seen in Table 8, which shows that, based on the indexed ADA measures, the closer candidate won 10 of the 12 elections. The closer candidate lost only the two elections in which the differences between the candidates closeness to the median voter were the smallest (0.6 and 3.4 points, for the 1996 and 1960 elections, respectively). Note that the divergent platform model has the same success in classifying success with indexed ADA scores. The candidate who is closer to the winning position is favored to win. But it is important to recognize that this does not mean that those who are far from the median voter cannot win the general election. It can be seen in Table 8 that, based on the median voter model and indexed ADA data, three of the 12 elections were won by candidates who were 36 to 40 points from the median voter s position. Johnson defeated Goldwater, who was 3 points farther away, and Reagan trounced Mondale, who was 19 points more distant from the median voter. And Clinton beat Dole, even though Dole s record was less than a point closer to the median voter.

18 456 Table 8. Effect of distance from median voter position on winning presidential general elections ( ) indexed ADA scores (median voter position = 38) Closer candidate More distant candidate Closer candidate Indexed Dist. to Indexed Dist. to Diff in Name ADA Median Name ADA Median Dist. Outcome 1996 Dole Clinton Lose 1960 Nixon Kennedy Lose 1964 Johnson Goldwater Win 1968 Nixon Humphrey Win 1992 Clinton Bush, G.H Win 2000 Gore Bush, G.W Win 1980 Reagan Carter Win 1988 Bush, G.H Dukakis Win 1984 Reagan Mondale Win 1972 Nixon McGovern Win 1976 Carter Ford Win 1956 Eisenhower Stevenson Win Conclusion Although there is a wide range in our estimates of the annual discount rate that voters apply to presidential elections, the average estimated 59% annual discount rate for prior economic growth in presidential elections is quite high but somewhat lower than the 79% discount rate estimated by Figlio (2000) for prior ADA scores in Senate elections. Our study thus provides additional evidence on the but what have you done for me lately phenomenon, as well as confirming a large literature that has found that incumbent parties benefit from strong economic growth. We have extended Zupan s (1992) series on presidential ADA scores by 11 years to 2000 and have combined them with presidential candidates congressional and estimated gubernatorial ADA scores and Nominate scores to examine the influence of spatial voting records on which candidate wins the party s nomination for president and on which nominee wins the general election. The only other statistical analyses of multiple presidential general elections from a spatial perspective rely on subjective estimates of how liberal each nominee was, which may be biased by hindsight. No one appears to have studied the role that spatial position plays in how parties select their candidates for president over multiple elections. We examined the success of 17 Republican candidates in 5 elections and of 38 Democratic candidates in 8 elections. Although moderate candidates do better in the general elections, parties favor extremists in the nomination process.

19 457 Based on indexed ADA scores, the most conservative Republican candidate won each of the five nominations. For Democrats, the winning primary position is slightly more moderate (e.g., an average indexed ADA score of 78). Democratic candidates with more liberal or more moderate records than this are less likely to win the nomination. These results are inconsistent with the lack of pattern for winners in Cooper and Munger s simulations based on random entry in 10 candidate races; this inconsistency could be due to entry not being random or there being far fewer than 10 effective candidates. Our study of 12 general elections for president found that the presidential nominees spatial voting records have an important impact on the outcome of presidential elections. The evidence is mixed on whether a nominee fares better with a record that is close to the median voter s position or with one that is close to his winning party position. Half the regressions found that the median voter s position maximizes the nominee s success. In the other half of the regressions, the winning party positions are 11 points apart on average, which is much smaller than the typical distances that separates party voting records for presidents and members of Congress. Indeed only one of each party s nominees was between his party s winning general election position and the median voter s position. The 11 other nominees in each party were either Republicans who were more conservative than the winning Republican position or Democrats who were more liberal than the winning Democratic position. The relationship between the nominees ideological positions and electoral success is remarkable. Based on indexed ADA scores, the nominee who was closer to the median voter s position won 10 of the 12 elections. And the two misclassified elections were those in which the difference between nominee distances from the winning position was the smallest (1996 and 1960). The relative distance from the winning position is highly significant in explaining the incumbent party s vote share and whether the incumbent party won the election. Spatial position is an important factor in determining the outcome of presidential elections. It has nearly as much impact as economic growth, which has been the focus of much of the empirical research in this area. Our results differ sharply from those of Alvarez and Nagler (1995, 1998), who in their case studies of the 1992 and 1996 U.S. presidential elections found that perceptions about the economy overshadowed perceptions about policy positions in determining voting behavior. Thus, we have found that the nomination process selects candidates who in 11 of 12 elections were more extreme (i.e., more liberal for Democrats and more conservative for Republicans) than the winning party position in the general election. More moderate candidates would have had greater success in the general election. This outcome may be due to the two stage electoral process. In the theoretical literature, if there is some uncertainty about the general election outcome, the Democratic party s median primary voter, for

20 458 example, trades off having a more moderate candidate who is farther from her preferred ideology against that candidate s higher probability of winning the general election. As a result, she may maximize her expected utility by selecting a candidate who is more liberal than the median voter in the general election. 22 Therefore, our findings that party nominees diverge from the median voter s position and that more moderate nominees fare better in the general election are consistent with this set of models. Party primaries pull the nominees away from the winning general election party position. On the other hand, these two stage models are unable to explain Schmidt et al. (1996) finding for U.S. senators and our finding in half the regressions for presidential candidates that candidates who are too moderate are less likely to win. Adams and Merrill (2003a) suggest that this result could be due to a loss of votes from party loyalists caused by alienation as parties converge. This study also provides a comparison of the performance of three prominent ideological measures in explaining the outcomes of presidential primaries and general elections. Each outperforms the other two measures in some context. But the best overall measure is the indexed ADA scores (Groseclose et al., 1999), which are comparable across chambers and over time. This measure produces a better fit than the other two in explaining which candidate wins the Republican primary, which party wins the general election, and the candidate s vote share in the Democratic primary. The modified ADA scores are the best measure for determining who wins the Democratic primary and the vote share in the Republican primary (under one specification). Poole and Rosenthal s Nominate scores beat out the other two ideology variables in explaining the incumbent party s vote share in the general election and the vote share in the Republican primary (under another specification). Acknowledgments We have benefited from comments received at the 2001 Public Choice meetings and at the 2001 Southern Economic Association meetings. We are especially grateful to James F. Adams, Robert Emerson, Bernard Grofman, Rebecca Morton, Ben Russo, Donald Wittman and an anonymous referee for their help. Notes 1. See, for example, Enelow and Hinich (1984). 2. Huntington (1950), Fiorina (1974), Fenno (1977), and Peltzman (1984) were among the most vocal critics of the median voter model, arguing that divergent platforms were needed to adequately explain political behavior. 3. See Johannes and McAdams (1981), Glazer and Robbins (1985), Abramowitz (1988), Westlye (1991), Wright (1993), Goff and Grier (1993), and Figlio (2000).

21 See Bullock and Brady (1983), Peltzman (1984), Wright (1989), Grofman, Griffin, and Glazer (1990), Shapiro, Brady, Brody, and Ferejohn (1990), Poole and Rosenthal (1991), Jung, Kenny, and Lott (1994), and Francis and Kenny (2000). 5. See Whitby and Bledsoe (1986) and Schmidt et al. (1996). 6. Keech (1995) summarizes this literature. See also Peltzman (1990), Hibbs (2000), and Bartels and Zaller (2001). 7. There is a larger empirical spatial literature that uses survey data on perceived candidate positions and voter preferences for individual elections. Enelow and Hinich (1984), Poole and Rosenthal (1984), Erikson and Romero (1990), Alvarez and Nagler (1995, 1998), Adams and Merrill (1999, 2000, 2003b), Alvarez, Nagler, and Bowler (2000), and Alvarez, Nagler, and Willette (2000) describe candidate location, explain voting behavior, and explore the electoral implications of changes in the candidates policy positions. 8. See Poole and Rosenthal (1991). 9. In related research, Jenkins and Sala (1998) found that spatial positions were quite successful in explaining the House vote for presidential candidates in 1825 after no candidate won amajority in the Electoral College in Note that the set of votes used for the House and Senate are somewhat different, reflecting what votes were taken in these two chambers. Also, as indicated, the presidential ADA scores pertain only to those House and Senate ADA votes for which a presidential stance was noted in Congressional Quarterly Almanac. 11. This avoids classifying Ted Kennedy as a conservative in 1980 when he missed 12 of the 18 votes used by the ADA as he ran for President. The ADA gave him a 33 rating, but our modified ADA score for him is Johnson s career indexed ADA score as president (76) was much more liberal than his Senate record (50) but somewhat more conservative than his indexed ADA score in 1964 (82). 13. Earlier scores are used for candidates who had not been in Congress for several years, such as Nixon in the 1960, 1964, and 1968 elections. 14. During this period, as Georgia experienced a transition from a solidly Democratic state to atwo party state, its Democratic senators became much more liberal. Their average ADA scores for is used for Carter s 1976 presidential bid, but this average is still well below his subsequent ADA score as president. 15. Note that under these criteria Eugene McCarthy barely qualifies as a candidate for the Republican nomination in 1968, since he received 1.0% of votes cast over all the Republican primaries, obtaining a sizeable number of votes in six primaries. Write-in votes for him accounted for % of the Republican votes in Nebraska, New Jersey, and Oregon and for % of the Republican votes in Massachusetts, New Hampshire, and Pennsylvania. Source: Congressional Quarterly (1997). 16. The Republican candidates were 1964: Goldwater, Lodge, Nixon, Rockefeller, Scranton, Smith; 1968: Nixon, McCarthy (see footnote 15), Reagan, Rockefeller; 1980: Anderson, G.H. Bush, Reagan; 1996: Alexander, Dole; 2000: G.W. Bush, McCain. The Democratic candidates were 1952: Kefauver, Russell, Stevenson; 1956: Kefauver, Stevenson; 1960: Humphrey, Kennedy, Morse; 1972: Chisholm, Humphrey, Jackson, McCarthy, McGovern, Muskie, Sanford, Wallace; 1976: Brown, R.C. Byrd, Carter, Church, Harris, Jackson, Udall, Wallace; 1984: Glenn, Hart, McGovern, Mondale; 1988: Dukakis, Gephardt, Gore, Hart, Simon; 1992: Brown, Clinton, Harkin, Kerrey, Tsongas. 17. Similar probits explaining which candidate wins the nomination and later whether the incumbent party s nominee wins the general election are not reported because probit (or logit) in several specifications paradoxically is unable to produce an estimate of the impact of the candidate s voting record when the voting record perfectly or nearly perfectly