Introduction

The divisive primary hypothesis, first suggested by Key (1953), posits that when a party’s primary is competitive or the eventual nominee does poorly in the primary, the party suffers in the general election. However, in presidential elections, measuring the impact of divisiveness is complicated by the fact that campaigns are waged both at the state and national level. As a consequence, the relationship between divisive state primaries and general election outcomes is substantially different in presidential campaigns than in subnational campaigns.

Substantial research exists on the impact of divisive state primaries, however this research generally ignores the important distinction between national and subnational elections. Presidential elections, unlike state-level elections, directly involve the national parties. In any state, a divided national party could have a negative impact on the performance of its presidential candidate even though that state’s presidential primary was not divisive. Thus we do not know if national-level or state-level divisiveness exerts greater influence on state-level outcomes in presidential elections because existing models do not account for national party division (NPD).

In a single-state primary, the winner of the popular vote becomes the party nominee. Presidential primaries are part of a larger, more complex environment. In presidential campaigns, individual state primaries do not determine the identity of the nominee. Rather, they select (or apportion) delegates to the national convention who then select the party nominee. It is common in the literature to use the term “divisive presidential primary” either to describe the divisiveness of an individual state primary or to describe the divisiveness of the national party during the nomination process.

There has long been concern that the presidential nomination process undermines party cohesion and encourages intraparty factionalism. When one national party is divided and the other party united, the divided party usually loses the election. The relative divisiveness of the national parties is a critical component of the national campaign, yet it is included neither in models of state primary divisiveness nor in models of aggregate presidential election outcomes. Excluding NPD from models of presidential election outcomes has the potential to bias the estimate of the impact of divisive state primaries or other variables. To measure the full impact of the divisiveness in presidential campaigns, it is necessary to measure both NPD as well as divisive primaries in individual states.

NPD is not simply an aggregation of divisive state primaries. NPDs are deeper and larger than state party divisions. A set of divisive state primaries does not necessarily indicate a divided national party in the general election or vice versa.Footnote 1 Absent the influence of NPD, measuring the impact of divisive state primaries might seem relatively straightforward. However, studies of subnational divisive primaries have reached a confusing variety of conclusions (Lengle and Owen 1996).Footnote 2 Measuring the impact of divisive state primaries in presidential campaigns is further complicated by the impact of NPD. In this research, we establish and measure the impact of NPD and that of divisive state primaries (DSP). To these ends, we first develop a comprehensive model of state outcomes in presidential campaigns that incorporates both state and national-level controls.

As we will explain, there are several ways to define the appropriate timeframe and to specify the model. We test several possible measures of NPD, and several measures of divisive state primaries found in previous research. We show that our findings are not artifacts of statistical techniques, timeframes or operational definitions. The results are consistent; varying only slightly across a wide range of methods and specifications.

Divisive State Primaries in Presidential and Subnational Campaigns

The causes and consequences of divisive presidential primaries are somewhat different than those of divisive subnational primaries.Footnote 3 The literature on divisive congressional and gubernatorial primaries posits a link between a divisive state primary and the general election outcome in that state. Because presidential nomination campaigns are sequential and national in scope, some of what occurs in individual state primaries spills over to other state contests.Footnote 4

Previous studies (Hacker 1965; Kenney and Rice 1987; Atkeson 1998; Lazarus 2005; Southwell 1986) have suggested that a divisive subnational primary decreases that party’s vote because (a) supporters of the losing candidate are alienated or discouraged, (b) the primary battle provides rhetorical “ammunition” for the opposing party, or (c) the state party’s resources are depleted. Each of these effects manifests differently in presidential campaigns than in subnational campaigns. In a congressional or gubernatorial campaign, a competitive primary may divide the state party and deplete its resources, hurting its ability to compete in the general election. In a presidential campaign, a few divisive state primaries would neither divide the national party nor deplete its resources. Presidential candidates allocate resources to states based on their strategic importance; if state party resources are depleted in a battleground state, the national campaign will pump money into that state.Footnote 5

In a presidential campaign, because of national media coverage, supporters of losing candidates may be alienated even if there was not a divisive primary in their state. Rhetorical attacks made by intraparty rivals in a few primaries may be co-opted and disseminated nationally, influencing subsequent national media coverage of that candidate. More generally, the negative image of an internally divided national party may be a potent cue to general election voters. In presidential campaigns, some of the influences on general election outcomes derive not from divisive primaries in specific states but from a divided national party.

An incumbent president may be challenged within his/her own party during the primaries, but the challenge is national, not restricted to a specific state. Both incumbent and challenger choose the state primaries in which they will compete vigorously. That decision is based on national-level factors as well as state factors. Similarly, both frontrunners and challengers in non-incumbent nomination campaigns run in particular states to bolster their chances of winning the national nomination. Differences in the context and causes of divisiveness in presidential and subnational primaries help explain why various studies have come to such different conclusions—analyzing presidential and subnational primaries separately leads to more clear and meaningful results. In this research, we focus exclusively on presidential primaries and their impact on state-level presidential general election results.

Developing a Model of State Vote Outcomes

We develop a comprehensive model of a state vote outcomes to estimate the impact of NPD, and to test a variety of operationalizations of divisive state primaries to determine the extent to which they influence the estimated effects. Examining the literature on divisive presidential primaries, we find that virtually all models are under-specified. Typically, they include few state-level controls and few, if any, national-level controls. This allows the possibility of excluded variable bias—that the estimated impact of divisive state primaries includes some of the impact of excluded variables, typically inflating and biasing that estimate. Our model of state general election outcomes takes into account a wide set of national-level and state-level variables that generally correspond to the factors that influence individual voting behavior. The use of state-level data, including partisanship and ideology, should provide strong controls to measure the impact of NPD and to re-examine the divisive primary hypothesis.

The Key Variables

The dependent variable used here is the proportion of the major party state vote won by the Democratic Party in the general election.Footnote 6 To measure NPD, we use the proportion of delegate votes received by the Democratic nominee (on the first ballot) at the convention minus the corresponding proportion for the Republican nominee as a measure of relative NPD. Below, we will show that this measure, though not ideal, leads to results that are substantively the same as results obtained using very different measures of NPD such as the difference in aggregate poplar vote. To measure divisive state primaries we use the proportion of the state primary vote received by the eventual Democratic nominee minus the corresponding proportion received by the eventual Republican nominee as the measure of divisive state primaries (Kenney and Rice 1987). Below, we test a variety of possible measures of divisive state primaries to determine the degree to which the operationalization of this key variable influences the results.

State-Level Variables

Some previous studies of divisive primaries have controlled for state-level effects by including one or more previous presidential election results (Mayer 1996; Atkeson 1998). In this study, state-level effects are accounted for by controlling for state partisanship, state ideology, and the home states of the presidential and vice presidential nominees.

Rabinowitz et al. (1984) analyzed the vote outcomes of the states and found that presidential elections are structured by party and ideology (Jackson and Carsey 1999; Erikson et al. 1993). Research at the individual and state levels shows that partisanship exerts substantial influence in presidential elections. State partisanship is measured here as the average of the most recent statewide votes for Governor, Senator and U.S. House. Previous presidential vote is not included since that might reflect national factors involving previous presidential politics rather than underlying state-level partisanship. In this model, state partisanship is not fixed; rather, its values often change to some degree from election to election.

One might argue that previous voting in congressional and gubernatorial elections is not a good measure of partisanship for the southern states until recent decades. In their analysis, Rabinowitz et al. (1984) found that conservative Democratic states such as Alabama and South Carolina tended to cluster in a different part of the factor space than either liberal Democratic states or conservative Republican states. Thus, we include two measures of state ideology, a general left–right scale and a scale that involves civil rights and social issues.Footnote 7 These controls should capture changing state-level effects.Footnote 8

Although the general ideology measure fails to capture some salient issues, it does reflect many issue-oriented differences across state populations. Civil rights issues (integration, voting rights, affirmative action, etc.) and social issues (abortion, gay rights, gun control, etc.) have been powerful over many elections and, critical to this study, their impact has been regional, affecting states differently (McCarty et al. 2006; Zaller 1992). Both ideology variables are measured using mean DW-NOMINATE roll call data scores for the U.S. House delegation in each state in the term prior to the presidential election (Poole and Rosenthal 1997). The expectation is that the civil rights/social issues variable will have a negative effect on voting for the Democratic Party in elections since the 1970s.

Candidate evaluation is difficult to measure at the state level. Polls, which could provide such information, are rarely consistent in format and rarely available for every state. Thus we are limited to controls for the home state of the presidential and vice presidential candidates, and the presidential candidate’s home region. Presidential candidates tend to do better in their home region than elsewhere and both presidential and vice presidential candidates tend to do well in their home states. Each of these variables can take on values of −1, 0 or +1 (e.g., the Republican candidate’s home state has a value of −1), although 0 is by far the most common. Home region is measured as all states adjacent to a candidate’s home state (Holbrook 1991).

National-Level Variables

In the literature on forecasting presidential elections (e.g., Campbell 2001; Bartels and Zaller 2001) there is general agreement that the national economy has a powerful impact on the national popular vote. In the current research, national economic conditions are operationalized as the annual change in RDI. Since the dependent variable is Democratic vote share, this variable is multiplied by −1 when the incumbent president is Republican. Thus, credit or blame for the economy is directed at the incumbent party.Footnote 9

Several studies suggest that the apparent effect of divisiveness may be spurious (Jacobson and Kernell 1981; Kenney 1988; Atkeson 1998). It is quite possible that an unpopular incumbent would attract more or stronger challengers; similarly, a popular incumbent might “scare off” strong challengers. Thus the apparent relationship between divisiveness and vote outcomes may be an artifact of spuriousness—both divisiveness and vote outcomes are strongly influenced by the strength of the incumbent. This argument is supported, directly or indirectly, by numerous studies of subnational primaries (e.g., Hacker 1965; Partin 2002; Lazarus 2005).

We address this concern in two ways. At the presidential level, an unpopular incumbent may encourage intraparty challenges, which may exacerbate existing regional or ideological divisions within the party. However, unlike most subnational election campaigns, the out-party typically begins with 5–9 legitimate candidates (e.g., senators and governors) for the nomination, whether the incumbent is popular or not. Historically the number and “quality” of nomination candidates in the out-party (or open seat presidential campaigns) appears unrelated to the quality of the opposing party’s candidate.Footnote 10 Ford (in 1976) and Carter (in 1980) did attract strong intraparty opponents, but such cases are rare. Popular incumbents like Nixon and Reagan were challenged by relatively strong fields of opponents.

We also address this concern statistically. If the strength of the incumbent president is causing a spurious relationship to appear causal, including such a variable in the model would cause the parameter estimates of divisiveness to diminish or lose statistical significance. It is difficult to measure candidate quality directly, but we can do so by controlling for national economic conditions (which, except for 2008, rarely change much during the election year) and presidential approval, measured as the Gallup approval rating in January of the election year, before the primaries begin (Atkeson 1998).Footnote 11. These variables reflect the perceived quality of the incumbent party candidate. (Indeed, these are the two most common factors in the presidential election forecasts.)

Each party is a coalition of diverse elements. The longer a party holds power the more likely party fissures will develop (Campbell 2000). Thus, like the forecast models, we include a variable (“terms”) that controls for the length of time a party has held the White House. If the incumbent party has been in office for only one term, this variable has a value of 0; if it has been in office for two or more consecutive terms, this variable has a value of 1. We include a separate dummy variable indicating whether or not the incumbent president is running.

Typically, the major parties nominate relatively centrist candidates, but occasionally one party nominates a relative extremist. In general, candidates who are perceived as more ideologically extreme are disadvantaged in presidential elections. Bartels and Zaller (2001) combined expert ratings of candidates 1948–1980 (Rosenstone 1983) with NES data 1984–1996. We extend this measure through 2012. Higher absolute values indicate greater relative extremism. In addition, we include a control variable for the impact of war. Wars have the capacity to divide parties and affect the decisions of candidates and voters. This variable is measured as the number of combat fatalities as a proportion of the national population. Increased fatalities are expected to disadvantage the incumbent party. War is measured as a national, not a state, variable because voters in all states receive national news of foreign affairs and because variations across elections are greater than those across states (Table 1).Footnote 12

Table 1 Descriptive statistics 1948–2012
Full size table

Data and Methods

Presidential elections are not singular national elections, as they are often treated, but are 51 separate contests. The influence of state factors is not included in most national studies. The model developed here measures the effects of NPD and divisive state primaries on presidential general election outcomes across space (states) and over time (elections). A pooled time-series allows both state-level and national-level effects to be tested concurrently. The model is applied to the set of presidential elections from 1948 though 2012.

This study melds several research streams including those of divisive primaries and general election forecasting. The model used in this research differs from the national forecasting models in several ways. First, the purpose is explanation not prediction; we do not use previous presidential vote outcomes or national trial heat polls since they do not seem to add to the explanatory power of the model. Second, variables representing “special circumstances” such as Watergate or a Catholic candidate are not included. Instead, the model includes only factors that occur regularly in presidential elections. Third, the unit of analysis is the state rather than the nation. This study is not the first to develop a model of state-level presidential voting; Gelman and King (1993), Campbell (1992), Holbrook (1991), and Rosenstone (1983) have provided useful guidance.

A time-series cross-sectional design should have enough power to generalize about the relationship between nomination campaigns and general elections. Because multiple units in time are observed, we need to control for the election year context (Stimson 1985). The substantive relevance of the election year is highlighted by Atkeson (1998). It is preferable to explicitly model national effects with actual national-level variables rather than leaving those effects in the “black box” of an election year dummy. Because election year dummies would be perfectly collinear with the national variables that change over time but do not vary across states, random effects for time are employed. However, fixed-effects are used to capture state-level heterogeneity. More formally, the standard two-way error component panel data model (Baltagi 2005) is given by

$$\it Y_{it} \, = {\text{a}} + \, X_{it} \, {\beta} \, + \, u_{it}$$

where \(u_{it} = \, \mu_{i} + \lambda_{t} + \, v_{it}\) and i indexes units and t indexes time.

Thus u it is a compound error term with a unit specific error term µ i , an election specific error term λ t , and an observation specific error term v it .

OLS estimates of this model fail to account for both unit-specific and time-specific unobserved effects, which leads to incorrect standard errors (and potentially biased estimates if either unit- or time-specific effects are correlated with the independent variables). Thus this model is estimated via maximum likelihood using fixed effects for states (to explicitly model unit-specific effects) and random effects for time (since fixed effects for time prohibit the estimation of coefficients of variables, like NPD, that vary over time but not across states). A Wald test shows that the state-specific effects are jointly significant (p < .001), and a likelihood ratio test against a model without random effects for time shows that the unrestricted model fits the data better than the restricted model (p < .001).Footnote 13

1948 was chosen as the starting date because, as the first post-WW II election, it in some ways represents the beginning of the modern era of electoral politics and because national economic data (in particular RDI) are not available prior to this date. Although some state primaries played a role in some earlier nomination campaigns (1912, 1928–1944), they played a substantial role in the 1948 Republican campaign and in subsequent campaigns.Footnote 14 The 1968 election has been excluded because of the anomalous character of both the primary campaign and the general election makes it inappropriate for this research.Footnote 15 With only 16 national elections, one of them might be an influential outlier that biases the coefficients. As a test, the model was estimated repeatedly, each time omitting one election year. The effects on the key variables were minimal.

There are 781 valid cases in the dataset.Footnote 16 Both primaries and caucuses are included. Lengle et al. (1995) analyzed the impact of caucuses, divisive primaries and non-divisive primaries on presidential election outcomes. Their results show very similar patterns for caucuses and non-divisive primaries, in sharp contrast to divisive primaries. They concluded that a caucus is a “non-divisive mechanism” that is virtually identical to a non-divisive primary in its impact on the general election. Accordingly, caucuses were assigned a value of 0 in terms of primary divisiveness (i.e., no advantage to either party). This substantially reduces the number of missing cases, facilitating more reliable parameter estimates. As a test, the model was estimated without any caucus states (see online Table A-1.)

Measuring National Party Division and Divisive State Primaries

NPD is generally driven by elites: activists, office holders, and opinion leaders (see Steger 2008). The actions of elites, especially candidates, can either exacerbate or mitigate NPDs. Unified, cohesive parties help the nominee, but parties that are internally divided and lacking cohesion hurt their candidates’ chances by diverting resources and tarnishing the nominee’s image (Campbell 2004). When party officials work together, with little ideological conflict, their electoral and policy goals are relatively clear (see Herrera 1993).

The greater the fragmentation, the less likely that one candidate will be seen as the dominant frontrunner in the “invisible primary”. Lack of consensus during the invisible primary does not necessarily indicate party division during the general election campaign. Contested nominations are the norm in presidential campaigns; competition for the nomination does not indicate a divided party. A strong diverse field of candidates can exacerbate existing divisions; however, the dynamics of the system are such that one candidate could quickly capture the nomination.Footnote 17

When a nomination campaign is divisive, the nominee and party elites attempt to reunite the party. They will not be able to erase years of ideological, regional or demographic differences, but they may be able to persuade disparate factions to work together temporarily to help the party win the presidency. The appearance of unity or division at the convention can influence undecided voters who are just beginning to focus on the campaign (see Holbrook 1996).Footnote 18

Measuring National Party Division

Measuring NPD presents multiple difficulties. Because this variable is central to our research, several measures were tested. Through the 1970s delegate votes at the national conventions provided a rough measure of divisiveness. Thus NPD could be measured as the proportion of convention delegate votes received by the Democratic nominee (on the first ballot) minus the corresponding proportion for the Republican nominee. This is a reasonable, though not ideal, measure of NPD, at least through the 1970s. However, beginning in the 1980s party elites have “stage-managed” convention votes (CV), perhaps in part to present the national television audience with the appearance of party unity.

A delegate-based measure mainly taps the behavior of party activists, chosen during the presidential campaign by fellow partisan voters. The national conventions are typically the most influential events of the entire campaign (Holbrook 1996). They occur when many voters, especially independents and weak partisans are beginning to focus on the two parties and their nominees. Most nominees receive overwhelming support on the convention ballot; the exceptions occur when the national party is severely divided.

As an alternative measure one could compare the proportion of the national primary vote won by the two nominees (aggregate primary vote, or APV). This measure is more appropriate to the post-reform period than to earlier elections.Footnote 19 Now, voting in a primary is consequential; delegates to the convention are allocated based on votes in primaries. Before 1972, there were few primaries and voting in primaries bore little if any relationship to the choice of the parties’ nominees. Neither CVs nor APV is ideal. Thus we considered, tested, but eventually rejected, several other possible measures, including the nominee’s New Hampshire primary vote (Norpoth 2001), the proportion of early-deciding partisans, and several variations of CV and APV (See online appendix). In the analyses to follow we estimate the impact of NPD using both CV and APV to demonstrate that our substantive conclusions about the impact of NPD do not depend on how the variable is measured.

Measuring Divisive State Primaries

There is no consensus on how best to operationalize state-level primary divisiveness. Different ways of operationalizing the concept might account for the differing results seen in previous studies. In presidential campaigns, a divisive state primary can be thought of as one in which the state primary electorate generally prefers a candidate(s) other than the eventual nominee. This implies that there is a large pool of voters who may be inclined to abstain or defect. This concept of divisiveness is measured by the proportion of the vote for candidates other than the eventual nominee (Born 1981). Kenney and Rice (1987) argue that the proportion of the state primary vote received by the eventual Democratic nominee minus the corresponding proportion received by the eventual Republican nominee is the best measure of state primary divisiveness (also see Atkeson 1998). This approach seems advantageous since it accounts for the relative divisiveness of the two parties’ state primaries.

The major alternative approach focuses on the competitiveness of the primary. A close, hard-fought primary may lead some voters to harbor intense negative feelings about the eventual nominee. This concept of divisiveness is measured by the vote margin between the two leading candidates in the primary (Lengle et al. 1995), sometimes operationalized as a dummy variable (e.g., less than 20 % vote margin).

Although both approaches measure aspects of divisiveness that could influence the general election outcome, they relate to substantively different phenomena. Consider a state primary in which the eventual nominee comes in second with only 30 % of the vote while the winner of that primary receives 35 % (the remaining votes distributed among other candidates). Such a primary would be considered highly divisive by the former measure (support for other candidates) but relatively non-divisive by the latter (margin of victory). Several measures used in previous research were tested. The analyses will indicate the kind of “divisiveness” that leads to diminished performance in the general election.Footnote 20

Results

The main results of the analyses are shown in Table 2. They indicate that both NPD and divisive state primaries are statistically significant at the .01 level (whether NPD is measured using CV or APV) and exert a potentially meaningful (i.e., non-trivial) impact on election results. The parameter estimates of the control variables vary, largely because of the different time-periods involved. As expected, the parameter estimates of the national economy, state partisanship and state ideology are all statistically significant and in the expected direction.

Table 2 Impact of national party division and divisive state primaries on state vote outcomes
Full size table

The results indicate that the impact of divisive state primaries is limited, while the impact of NPD can be substantial. For example (using CV as the measure of NPD), if in a certain state one party’s primary is divisive, with its eventual nominee receiving only 50 % of the state primary vote, and the other party’s primary is non-divisive with its nominee receiving 90 %, the former would lose only 1.12 % in that state in the general election. In comparison, smaller differences in NPD lead to greater differences in the national outcome. If the nominee of one party receives 70 % of the vote at his/her national convention while the nominee of the other party receives 90 %, then the former would lose 2.43 % in the national popular vote. During the 1948–2012 period, NPD ranged from –46.2 to +35.3 (negative values indicate greater division in the Democratic Party). Considering this range, the coefficient of .121 indicates that the effect of relative NPD on the Democratic popular vote varied from −5.6 % to +4.3 %.

The mean absolute value of DSP for all states is 19.18 %. The mean absolute shift in state vote outcome caused by DSP is .52 %, and the maximum is 2.7 %. The mean absolute shift in state vote outcomes caused by NPD is 2.9 %, and the maximum shift is 5.59 %. Assuming all the DSP scores in an election have the same sign (which is plausible though unlikely) and setting both DSP and NPD at their (absolute) means, the mean national shift caused by NPD is more than five times the mean national shift caused by DSP.Footnote 21

The Dependent Variable

A critical concern is the operationalization of NPD. We argue above that relative CV (per cent of first ballot votes for the Democratic nominee minus the corresponding percentage for the Republican nominee) is a suitable measure of NPD. We note however that since the 1980s, conventions and convention voting have become increasingly stage-managed. As indicated above, APV, is well suited to the post-reform era (from 1972 on). Thus we employ the basic model, as described above, in two ways: one using CV for the entire post WW II period (1948–2012), and the other using APV for the post-reform period (1972–2012).

Table 2 shows the results of the two models, identical except for the measure of NPD: CV in one, APV in the other. Because these variables are measured on different scales, we do not expect the coefficients of CV and APV to be similar. We do however expect, if both are reasonable measures of NPD, that the coefficient of divisive state primaries will not be affected much by which measure of NPD is used. Indeed, this is what we find: the coefficients of NPD differ according to their scale of measurement (.121 using CV 1948–2012, .237 using APV 1972–2012) but both are statistically significant at the .01 level; the coefficients of divisive state primaries are strikingly similar (.0279 using CV, .0258 using APV; both significant at the .01 level).

Measured as the proportion of CVs received by the Democratic nominee minus the corresponding number for the Republican nominee (CV), NPD indicates that, for example, if one nominee receives 90 % of the delegates, while the other receives 65 %, the latter would lose 3.03 % in the general election. The impact of NPD was at least 3.29 % in 9 of the 16 elections. The coefficient for NPD measured as relative APV (taking into account the difference in the unit of measure) is larger than when measured as relative CV, suggesting that the impact of NPD may be greater than estimated using CV. The similarity between the results using CV and APV provides evidence that the results are not very sensitive to the way that NPD is operationalized. Except as noted, the analyses to follow will use CV as the measure of NPD.

Tests for Robustness

A number of sensitivity tests were performed. To test for possible realignment effects and/or the McGovern-Frasier reforms, the model was estimated with dummies for 1972 or later and for 1980 or later, as well as interactions with these variables. Both January and July presidential approval ratings were tested. The war variable was tested both as a state-level and as a national-level effect. Third party votes were incorporated in several ways. Changes in coefficients (using either CVs or APV) are negligible when these variables are included. To make sure that no one election was driving or distorting the results, the model was estimated repeatedly, each time excluding one election. The results are very consistent, varying slightly across a wide range of statistical methods, model specifications, and measurements.

Neither possible realignment effects, the timing of presidential approval ratings, the way war was incorporated into the model, nor the exclusion of any one election led to results that failed to confirm the hypotheses. As shown in Table A-1 (online appendix), the parameter estimate for divisive state primary is similar if caucuses are excluded, though the parameter estimate for NPD is about 20 % lower; the substantive conclusions are unchanged. The parameter estimates are very similar whether or not 1968 is included. Nonetheless, for the reasons discussed above, in footnote 15, and the online appendix, we decided that the 1968 data are not appropriate to this research.

The model was also estimated with a two-way random effects model, a GLS estimator for panel data with AR(1) serial correlation (Baltagi and Wu 1999), Panel Corrected Standard Errors (both with and without a lagged dependent variable), and a mixed model that adds a random effect on a third party variable, effectively allowing it to vary across elections (while maintaining fixed-effects for states). The results are robust to the use of these alternative statistical techniques. The size of the coefficients varies to some extent, but the substantive message is the same: NPD (using either the CV measure or the APV measure) has large and statistically significant effects on election outcomes and divisive state primaries have small but significant effects (Table A-1, online).

With few exceptions, the results are very similar across varying specifications. The coefficient of NPD was very stable, and significant at the .01 level. (More precisely, the coefficients using CV as the measure of NPD are very similar to one another; the same is true of the coefficients using APV as the measure of NPD). The coefficient of divisive state primaries, measured in terms of support for the eventual nominees, was extremely stable and significant at the .05 level, two-tailed. (As discussed below, when measured by margin of victory, the divisive state primaries variable in not statistically significant.) The results of these tests indicate that that our findings are robust: both NPD and state primary divisiveness significantly influence state-level presidential outcomes.Footnote 22

Sensitivity to the Operationalization of DSP

As discussed above, several measures have been used in previous research to represent divisive state primaries. The two main approaches are to define a divisive state primary as (a) one in which the eventual nominee does poorly, and (b) one in which the victory margin is small. Each can be represented either as two variables (one for each party) or as one variable (the difference between the two parties in that state). Furthermore, margin of victory can be represented either as a continuous or as a dummy variable (e.g., using a 20 % cutoff).

As shown in Table 3, the results indicate that operationalizing divisive state primaries in terms of the eventual nominee’s performance is more reliable than operationalizing it in terms of victory margin.Footnote 23 Indeed, both of the former are statistically significant while none of the latter are. This suggests that diminished performance in the general election occurs because there is a large pool of voters who did not support the eventual nominee in the state primary rather than because some voters evaluate the nominee negatively because the state primary was close and competitive (Note that the coefficient of NPD is very stable and highly significant regardless of the way that divisive state primaries is measured).

Table 3 Results of tests of alternative measures of divisive state primaries (DSP)
Full size table

Although the impact of divisive state primary (measured in terms of the nominees’ relative performance) is statistically significant, its national impact is relatively minor (State party divisions are nonetheless quite important to state party leaders, who generally have little or no effect on NPD). Even in the most extreme possible case, where one party’s nominee received 100 % of the state primary vote while his opponent received 0 %, the impact on the general election vote in that state is only 2.79 % (using CV; 2.58 % using APV). State partisanship and state ideology have a greater impact than divisive state primaries (using standardized coefficients for comparative purposes). This confirms the expectation that divisive state primaries have a small and usually inconsequential effect on electoral outcomes. On the other hand, NPD is one of the more influential variables. These results support the hypothesis that NPD potentially has a substantial negative effect on electoral outcomes.Footnote 24

A divisive state primary leads to a maximum possible decrease in a state’s general election vote of 2.8 %, while a divided national party more often than not leads to a decrease of more than 3.2 %. The impact of divisive state primaries is limited to a subset of states while the impact of NPD is not. Taken together, these results are consistent with the thesis that the overall negative impact of NPD is greater than that of divisive state primaries. Even in a close election, it is unlikely that divisive state primaries would make the difference in terms of who wins the Electoral College, though NPD may well have such an effect.

Substantive Impact of Divisive State Primaries and NPD

To provide an overview of the potential substantive effects of DSP and NPD, we calculated the estimated vote in each state in each election, absent the effects of divisive primaries, and absent the effects of NPD. Table 4 shows the number of states (and electoral votes) that likely would have been won by the other party; below we describe the potential effects on each national election. These estimates are intended to be illustrative. There is no way to know which states actually would have switched. Candidate, media and voter behavior would have been different in various ways. The number of states that would switch is partly a function of the closeness of the election. A close national election combined with state or national divisiveness can lead to a number of states “switching”. The table reflects in substantive terms the results of the statistical analysis: the impact of divisive primaries is small and limited while the impact of NPD is greater and more widespread.

Table 4 Substantive impact of divisive state primaries and national party division (states and electoral votes expected to switch absent divisiveness effects)
Full size table

Based on the results shown in Table 4, among all 16 elections, only 16 states likely would have switched because of divisive state primaries alone; 95 states would have switched because of NPD alone. In none of the 16 elections do the results indicate that more than two states, or more than 53 electoral votes, would have switched because of divisive state primaries. In comparison, there were six elections in which NPD would likely have switched at least eight states with more than 112 electoral votes. (This illustration uses CVs to measure NPD; these estimates are more modest than those using the APV measure).

In six of the 16 elections, the impact of divisiveness would have substantially changed the results. Without the effects of state primary divisiveness, we estimate that Kerry in 2004 would have won Ohio and thus the presidency and in 1948 neither candidate would have won a majority of electoral votes. Without the effects of NPD, we estimate that in 2000 Gore would have won Florida and thus the presidency; in 1980, had it not been for the national division between Carter and Kennedy, Carter would have won 17 additional states, putting him within striking distance of Ronald Reagan. Similarly, in 1976 (absent the effects of NPD) Ford would have won New York, Texas, Pennsylvania and five other states leading to a substantial victory over Carter; in 1960, Kennedy would have won 18 additional states leading to a landslide victory over Nixon; and in 1948, Dewey would have won an additional 10 states, verifying the Chicago Tribune headline “Dewy defeats Truman”.

Conclusion

The relationship between divisive state primaries and general election outcomes is substantially different in presidential campaigns than in subnational campaigns. To appropriately estimate the impact of divisiveness in presidential campaigns, one must measure both the impact of NPD and the impact of divisive primaries in individual states. To this end, we developed a comprehensive model of state outcomes in presidential campaigns and tested several measures of NPD and several measures of divisive state primaries. We find that, in presidential campaigns, both NPD and divisive state primaries significantly influence state-level general election outcomes, with the former having a greater and more widespread impact.

In addition, we have demonstrated that the impact of state primary divisiveness is sensitive to how the concept is operationalized. One can conceptualize a divisive state primary as one in which the state primary electorate only weakly supports the eventual nominee, implying that there are many partisans who may abstain or defect. This is measured by the proportion of the vote for candidates other than the eventual nominee. Alternatively one can conceptualize a divisive state primary as close and competitive, causing some partisans to harbor negative feelings about the eventual nominee. This is measured by the vote margin between the two leading candidates in the primary. These two approaches relate to substantively different phenomena. The analyses indicates that the former leads to diminished performance in the general election (the latter is not statistically significant).

The results indicate that the impact of divisive state primaries is limited, while the impact of NPD can be substantial. A divisive state primary leads to no more than a 2.8 % decrease in the general election in that state. In comparison, NPD more often than not leads to decreases of at least 3.2 % across states. The impact of divisive state primaries is limited to a subset of states while the impact of NPD is not. Taken together, these results confirm the general thesis that the overall negative impact of NPD is greater than that of divisive state primaries.

This research demonstrates that NPD is a critical component of divisiveness in presidential campaigns, but one that generally is not included in previous research. By incorporating a comprehensive set of appropriate controls, we have reliably estimated the impact of NPD and of divisive presidential primaries. We show that the national component is potentially powerful; the state-level component pales in comparison. Previous analyses of divisive state presidential primaries have measured a minor effect while ignoring the greater effect.

Although this research has focused on the relative impact of NPD and divisive state primaries, the analysis also sheds light on the behavior of states in presidential elections. Among other findings, it indicates that election-specific national factors are critical to understanding general election outcomes and that long-term state-level factors such as partisanship and ideology play a major role in state-level electoral behavior.Footnote 25

It is hoped that this study contributes substantially to resolving the controversy over the impact of divisive presidential primaries. It has been shown that a divisive state presidential primary does have a negative effect on the vote outcomes in the general election, although the magnitude of the effect is relatively small, unlikely to change the winner in more than one or two states in a presidential election. Because of the control variables included in the model, it is unlikely that this relationship is an artifact of unpopular incumbents, weak economies, or the partisan or ideological predispositions of state electorates. Furthermore, the results are consistent—varying little across a wide range of methods and model specifications.

Having established this, it would be useful and interesting to differentiate between the relative impact of early and late primaries, those with high versus low turnout, and those with few or many active candidates. These and other state-specific factors could cause some divisive state primaries to have greater or lesser impact on general election results than others. That would represent a potentially important avenue for future research. Such research would present a number of measurement problems. Primary turnout is difficult to gauge because the denominator is generally unknown. The three factors are inter-related: greater turnout tends to occur when there are more candidates which tends to happen early in the primary season. Similarly, once the race has been called, both turnout and the number of candidates diminishes.

One implication for the parties is clear: in terms of winning the presidency, a divided national party does more damage than a set of divisive state primaries. Competition among candidates does not necessarily hurt a party’s general election chances, but schisms within the party’s base can be truly harmful. The analysis indicates that a divisive state primary will decrease the party’s general election vote in that state, but usually by less than 2 %. In comparison, a divided national party decreases the party’s vote across states, usually by more than 3 %. Party leaders no doubt understand that NPD leads to negative consequences in the general election. However, in most cases, they need not be concerned about the effects of divisive primaries in individual states. Except in the most pivotal states, such as Florida and Ohio, in a close national election, a decrease in the range of 1–2 % in the popular vote will not influence which party wins the Electoral College.

Much can be gained by investigating the causes and consequences of divided national parties. What causes the underlying long-term divisions within the parties? Under which circumstances do presidential nomination campaigns exacerbate such divisions? What can candidates do before and after the convention to unite their party? And what can the parties do between elections to diminish the chances that divisions will intensify during the next campaign cycle?