Party Identifier Party Elite Ideological Polarization

H_o: No relationship between party identifier ideology and party elite ideology.

H_a1: Simple: Party identifier ideology determines party elite ideology.

H_a2: ME Aftershock: Lagged party identifier ideology determines party elite ideology.

H_o: No relationship between party elite ideology and party identifier ideology.

H_a1: Simple: Party elite ideology determines party identifier ideology.

H_a2: EM Aftershock: Lagged party elite ideology determines party identifier ideology.

There are two primary sources of data for this analysis. One is the series of American National Election Studies from the Cumulative Data File. The cumulative data file consists of variables derived from the 1948-2004 series of biennial ("time-series") SRC/CPS National Election Studies⁷⁸. My analysis uses data ranging from 1954-2004. The second source of data for this analysis is the series of congressional ideology (D-W Nominate) scores developed by Poole and Rosenthal. Use of the D-W Nominate scores was necessary for cross-Congress comparisons.⁷⁹ Analysis was conducted using SAS version 9.0. The dataset employed in this analysis was constructed using a multi-step process. First, the Poole-Rosenthal D-W Nominate scores (hereafter referred to as nominate scores) were merged with the cumulative file by year, state, and district. This created a data set where each respondent in the NES cumulative file had corresponding nominate scores for the congressional representatives from his or her district and senators from his or her state. The second stage in constructing the dataset for analysis involved calculating means and standard deviations for all relevant variables in the merged dataset. Then, using those means and standard deviations, I created a set of variables which translated the means and standard deviations for the respondent ideology and legislator nominate scores into z-scores for the partisan groupings and the full sample.

Ż = Equation 12.1: Z-Scores

Where:

The third stage involved estimating averages for each of the relevant computed and merged variables for each year of the biennial NES time series. Those values were then entered into a separate dataset for the purpose of statistical analysis. The ultimate product is a data set of average respondent ideology and average legislator nominate scores by year (1954-2004) for the means, standard deviations, and z-scores as calculated from the second data set. The total sample size for the final dataset is 27 observations corresponding to each NES biennial survey.

The means and standard deviations for selected variables included in the analysis can be found in Table 12.1. The Party ID variable used for this analysis collapses the traditional 7-point party identification scale into a 3-point party identification scale with Republican and Democratic leaners collapsed into the Republican and Democrat categories with only non-leaning independents included in the Independent category. The ideology variable is the traditional 7-point scaling of ideology that the NES incorporated in 1972.

In addition, the thermometer for conservatives is included as a proxy for ideology. The Pearson’s correlation between ideology and the conservative thermometer from the cumulative NES cross-section is just .637 (P<.0001), however the conservative thermometer was asked as early as 1964, and thus it allows for several more degrees of freedom in any statistical model using it to assess political polarization. The conservative thermometer was not asked of respondents to the 1978 NES. The conservative thermometer for 1978 is set at the mean of the two closest years (1976 & 1980) to retain biannual continuity. However, given the distance between the conservative thermometer and ideology, the models assessing elite-mass and partisan polarization use the 7-point ideology measure. Except where the scales were the same, difference variables use average z-scores in their calculations. Method

Modeling the Relationship between Mass Ideology and Objective Measures of Elite Ideology

The polarization analyses consist of five distinct types of models utilizing OLS regression and GLS regression where appropriate. The first set of models asses simple regressions of year on the mass, elite, and partisan ideologies. Given the limited number of observations in the data set, more robust multivariate regression models simply are not possible, as they quickly expend the available degrees of freedom. For full sample ideological means, positively sloped parameter estimates indicate a shift towards the upper pole (conservative) of the ideological distribution. For full sample ideological standard deviations, a positive slope indicates an increase in political polarization, as the dispersion of ideology will have increased over the time series. For the partisan variables, the direction of the slope and its relationship to polarization is determined by which party it is. For example, for the Republican ideological mean, a positive slope would indicate polarization, as the average ideology of Republicans would be shifting in a more conservative direction. However, for the Democratic ideological mean, a

positive slope would indicate moderation, as the average ideology of Democrats would be shifting the center. This distinction is best illustrated using the Z-scores. The Republican mean Z-Score for ideology is 0.487. A positive slope would indicate that average Republican ideology is moving away from the mean. The Democratic mean Z-score for ideology is -0.408. Here a positive slope would indicate that average Democratic ideology is moving towards the mean. These models are reported in Table 12.2.

Equation 12.2

The second set of models (Table 12.4) assess the simple regression of year on the difference between Republican and Democratic ideology as well as the difference between Republican and Democratic feeling thermometers on conservatives as a group. The table reports the unstandardized means and the standardized z-score for both ideology and the ideology proxy.

Equation 12.3

The third and the fourth set of models include both simple OLS regressions and generalized least squared (GLS) regressions for the models assessing ideological polarization, partisan polarization, and elite versus mass causation of that polarization. As mentioned earlier, whether or not elites lead and the masses follow is a significant source of contention among scholars, though the emergent consensus suggests that elites condition changes in mass attitudes. The models reported in Tables 12.4 and 12.5 assess polarization and elite-mass causation empirically.

For each group under consideration, there are eight regression models to test this relationship. Four models test a mass  elite causal relationship using the average ideology of respondents to predict the average ideology (nominate scores) of the elites (legislators). Whereas the other four modes test an elite  mass inferential model where the average nominate scores are used to predict average respondent ideology. For each group, a simple OLS regression is tested for both the mass  elite and the elite  mass models. However, given that this analysis is a cross-sectional time-series, accounting for the potentially serious problem of serial correlation necessitates employing a GLS model using autoregression techniques. For each of the three sets of independent variables, a GLS model is estimated. The Durbin-Watson statistic testing for first order autocorrelation and the probability of positive autocorrelation are reported in Tables 12.4 and 12.5 for the appropriate models.

In order to assess the theoretical problem of mass vs. elite causation (i.e. which is the chicken and which is the egg) I use two lagged independent variables (2 year lag & 4 year lag) of mass and elite ideology. Thus three possible causal relationships are employed testing both mass  elite and elite  mass polarization. The first model sans lagged variables tests whether there is a simultaneous relationship between mass ideology and elite ideology. If mass and elite ideology are both predictive in the same year, then I conclude this is strong evidence supporting the recursive model of mass and elite polarization. Indeed, the recursive model is supported, even if we find lagged effects, when there is significant within-year causation between mass and elite ideology. If lagged mass ideology (2 or 4 years) predicts elite ideology, then the mass  elite model is supported. If, however, lagged elite ideology (2 or 4 years) predicts mass ideology, then the elite  mass model finds support. In other words, if the mass or elite ideology from 1972 predicts the mass or elite ideology for 1974 or 1976, then we have a strong temporal basis for pointing the causal arrow in one direction or the other. The last set of models in Table 12.5 assess whether the differences between Republican and Democratic partisan identifiers is driving the observed differences between Republican and Democratic legislators, and vice versa.

Equations 12.4 – 12.9: Models for Mass  Elite & Elite  Mass Causation

Equation 12.4 Equation 12.5

Equation 12.6

Equation 12.7

Equation 12.8

Equation 12.9

The fifth set of models assess the squared differences between normalized legislator nominate scores from the two parties and normalized constituent ideologies scores by year for both the House and the Senate. The squared difference is calculated in order to retain the ideological differences between the constituent and his or her representative while permitting an assessment of the absolute difference between a legislator and his or her constituents. Using constituent ideologies means that a legislator’s ideological distance from respondents, Republican and Democrat, from his district or, if he is a senator, from his state. Furthermore, there is an important difference between this set of models and those previous. For the squared difference variables, the differences were first calculated in the second data set. This was necessary as the constituent model assesses distances between a legislator and that legislator’s constituents. The aggregation makes this kind of analysis impossible in the third data set. The problem with this analysis from the outset is that the ANES sample is not appropriate for state and district level aggregations. While the constituent models reported in Table 12.6 are at the party and house levels of aggregation, the differences between any one included legislator and his constituent may rest on just one respondent…or it could be twenty respondents. While these problems may average out over the long run, they do present a peculiar problem for this analysis. The assessment of constituent effects may further necessitate examining only the constituents from the party of the legislator (Fenno’s re-election constituency), however that would only exacerbate the previously mentioned problem. While this set of models is presented here tentatively, we hesitate to rely on them without further diagnostics.

Equation 12.4

FINDINGS & ANALYSIS

Ideological Polarization over Time

First, I will address the primary theoretical and empirical problem that students of political polarization grapple with: namely, well, polarization…is there any? The short answer is, yes. A great

** significant at .05 level

***significant at .01 level

deal! Table 12.2 reports the linear trends in political ideology for the mass public, elite politicians, partisan identifiers, and party elites (legislators). The first model regressing year on party identification demonstrates a slight Republican trend over the course of the time series. It is thus not particularly surprising that the trend in mean ideology indicates a shift towards the conservative end of the ideological spectrum for the mass public in the aggregate. Both represent evidence for mass ideological polarization, as both means are trending towards the poles of the distribution rather than the center. There was no apparent trend in average respondent feelings towards conservatives, though this finding does underscore the fact one must be cautious in using it as a stand in for ideology. Then again, perhaps the additional years in the time series accounts for the conservative FT’s poor showing. Of most interest in the aggregate models is the significant positive trend in the standard deviation for respondent average ideology. An increase in dispersion for ideology is confirmatory evidence in favor of the centrifugal change hypotheses for mass public ideological polarization. Ideologically speaking, there is growing dispersion in the aggregate electorate as well as a shift in the average ideological position towards the extremes of the ideological continuum and not to the center as Fiorina and his fellow

Culture Wars skeptics have suggested. Note in particular the standard deviation model reported in Table 12.2. I find a statistically significant increase in the variance in ideology over the time series. This runs directly counter to Fiorina’s and the other Culture Wars skeptics who suggest that there has been a centripetal trend among the mass electorate over the last 40 years.

Partisan Polarization over Time

While one must sift through the aggregate findings for subtle bits of evidence on our theoretical questions, there is no such problem when it comes to the partisan models and partisan polarization. What is of interest is that both the partisan identifiers in the mass public model s and the partisan elite models evidence a strong trend in ideological polarization. The later has been well documented by Poole and Rosenthal (Poole and Rosenthal 1984; McCarty, Poole, and Rosenthal 2006), but the former, as our previous discussion aptly demonstrated, is a major bone of contention. Fiorina argues that the party elites have become beholden to activists and interest groups while ignoring the mass electorate. They polarize to service the interested and passionate few at the expense of a largely centrist but unorganized public. But the results of this analysis indicate that the mass public is polarizing to some degree, and that partisan constituencies in the mass electorate are polarizing to a much larger degree. This is not a small point. An isolated party elite beholden to the organized (i.e. rich) and the passionate (I.e. the poles) could be a significant problem for American democracy. Certainly Fiorina thought so when he sounded the warning. But a party elite that is responsive to its adherents in the electorate—that may not be a problem at all. It sounds a lot like what parties are for. Indeed, it wasn’t so long ago that the APSA was explicitly calling for “responsible party government” where there were clear partisan differences, party discipline, and thus party responsibility. Is this a case of be careful what you wish for…or is the “hijacking of American democracy” really just responsible party government with sinister music playing in the background? Table 12.2 finds strong evidence that the Democrat and Republican identifiers in the mass public have polarized along the ideological dimension. Both in terms of the direction of the mean (towards the poles) and in the relative placement of the two parties on the ideological scale (the Z-score model), there is strong statistical evidence that both parties have moved away from the other. There is increased ideological disparity between Republican and Democratic identifiers since 1970.

Table 12.3: Models Regressing Differences b/w Republican & Democrat Identifiers by Year

MODEL	N	Intercept	Parameter Estimate		Standard Error	R2
IDEOR-D(m) = B0 + B1 (year) + e	17	-43.515	0.023	***	0.004	0.681
IDEOR-D(z) = B0 + B1 (year) + e	17	-25.946	0.014	***	0.002	0.668
CONFTR-D(m) = B0 + B1 (year) + e	21	-395.978	0.206	***	0.025	0.787
CONFTR-D(z) = B0 + B1 (year) + e	21	-23.575	0.012	***	0.001	0.816

* significant at .10 level

** significant at .05 level

***significant at .01 level

Looking to Table 12.3, we can see that looking at the question of polarization in terms of the difference between Republican and Democratic identifiers places the issue in stark relief. There is strong evidence of an ever widening divide between Republican and Democratic identifiers, even in our previously disappointing conservative feeling thermometer. Republican and Democratic identifiers in the American electorate are moving away from each other and taking up further distance positions on the ideological spectrum.

The part of this analysis most fraught with pitfalls is the assessment of temporal causation in mass and elite polarization. Is it a) mass-driven polarization or is it b) elite driven polarization? Or is it, as Fiorina argues, in fact, c) elite polarization irrespective of mass behavior (polarized or no). Table 12.4 provides some powerful evidence on this point. It is evident from this table that there is no support for the recursive model of ideological polarization between the mass public and political elites. Average ideological scores for the mass public are not correlated with the nominate scores for legislators of the same year in the House, Senate, nor the two combined. Furthermore, there’s little support for the Elite  Mass hypothesis, as none of the models that use lagged nominate scores to predict contemporaneous average mass public ideology scores is significant. While this specific finding doesn’t speak to Fiorina’s argument, it does run counter to a significant sub-literature which suggests that elites drive mass political opinion (Hetherington 2001; Conover, Gray, and Coombs 1982; Converse 1964; Sullivan, Piereson, and Marcus 1978; Mutz 2006; Hill and Hurley 1979). As can be seen in Table 12.4, I find no evidence that elite polarization is driving trends in the ideological disposition of the American public.

There is evidence in support of the Mass  Elite hypothesis. In all three sets of models, the lagged ideology of the mass public was a significant predictor of elite ideology. In the House and Senate models, the ideological polarization of the public is a significant predictor of the polarization in our elite legislative bodies on a two year lag (Table 12.4). In the full congressional model, this relationship is significant at the .01 level, giving strong evidence that the apparent relationship is not due to chance. For every single unit of change in the ideology of the masses there is a near full point (.749) change in the ideology of Congress. About half of the variation in the model is explained using the 2 year lagged ideology of the masses (0.549).

This suggests that elite polarization was, in fact, preceded by mass polarization. This result is consistent with my alternative formulation which argues that elites are sensitive to the opinions of the mass public and responsive to shifts in the aggregate views of the electorate. The causal arrow points from the masses to the elites and not, as the consensus in the literature suggests, from the elites towards the masses. And this elite-responsive-to-mass citizenry relationship is apparent in both the House of Representatives and the Senate models, arguing against redistricting or district characteristics as the prime mover mass-elite ideology. This evidence alos runs directly counter to the Fiorina conjecture of a runaway political elite dashing to the poles while ignoring the largely centrist public. Rather than elites diverging completely independent of mass centrism, I find that the mass public has polarized on ideology, and that elite polarization has lagged behind the polarization in the mass public. It is the mass public that has departed for the poles first, while elite legislators have followed—polarizing in response to mass ideological polarization. This suggests that Fiorina’s expressed concern over the democratic process in light of political polarization is somewhat overblown. While political polarization in the legislature reduces the probability of compromise and incremental politics, it is not in and of itself undemocratic. When elites polarize in response to mass polarization, this is an indicator of representative democracy, not a departure from it. This responsiveness is evident at the level of the general mass public and the elites in the congressional institutions, but the most powerful evidence of elite/mass responsiveness is among partisan elites and mass partisans.

** significant at .05 level

***significant at .01 level

While the models assessing the ideological trends of the mass public and political elites failed to support the recursive hypotheses, there is ample evidence of a recursive relationship between party identifiers and partisan elites in Table 12.5. All of the models that regress partisan identifier ideology on party elite ideology and models that do the opposite are significant with strong reduction in error. The best performing models for the House, as measured by R-Square, are the elite  mass models, with the contemporaneous model performing the best, explaining just short of half the variation in mass ideology (.499). It outperforms the lagged models, suggesting that increasing polarization among party identifiers and party elites is a simultaneous event. And while Republican identifiers are strongly responsive to Republican elites, the Democrat models perform markedly better than the Republican models. The R-Square for the elite  mass models for the Democrats is nearly 15% higher than in the Republican models (.631). The coefficients for the Democratic models are also, on average, larger than the counterpart Republican models. The model coefficient for the contemporaneous elite  mass model is 2.499, while the corresponding coefficient for the Republican elite  mass model is 1.290, a full point difference on the ideological scale.

The difference models examining the trend in ideological differences between partisan identifiers and the ideological differences in Congress are particularly strong, indicating that partisan identifiers have responded to the ideological polarization of Congress and that congressional legislators have responded to their partisan constituents becoming less ideologically diverse and more ideologically extreme. Note, here we are looking at the differences between the mass identifier ideological scores and the elite legislative nominate scores. Again, the coefficients for the elite  mass models are significantly larger than those for the mass  elite models, though all meet typical standards of statistical significance (0.428 vs. 2.050 on the high end). This is highly significant partisan polarization along ideological lines both in the mass public and among elite legislators. While Fiorina might be tempted to dismiss this as sorting, this result runs counter to Fiorina’s argument that little mass sorting

Table 12.5: Mass Party Identifiers & Party Elite (Legislators) Simple & Autoregressive Models of Partisan Ideology

** significant at .05 level

***significant at .01 level

has occurred and only moderate sorting has occurred at the elite level. I find strong partisan polarization at the elite and the mass level, and that the trends in mass and elite polarization over time are recursively related.

Table 12.6 reports models regressing the squared difference between the z-standardized ideology of respondents in the states/districts of each Republican and Democrat House representative or senator for the full set of time series data. Thus the model looks at the squared difference between the standardized ideology of the constituent minus the standardized ideology of the elite legislator who represents those constituents. The model thus doesn’t assess whether constituents are further polarized relative to their representatives (or vice versa), but whether there has been a linear increase in the distance between the constituents of representatives and the representatives themselves. The constituent models are intended to assess the difference between a mass public -elite responsiveness based on party identifier positions and aggregate partisan elite ideology versus elite partisan legislators responding to changes in their geographical constituents. The overall trend in these models is of increasing distance between Republican legislators and their constituents while, at the same time, there is decreasing distance between Democratic legislators and their states and districts. While the Republican Senate model is insignificant, the coefficient is in the positive direction, and the House Republican model indicates significant constituent-representative polarization (0.013). Indeed, the House model explains half of the variation in the squared differences between Republican elite representative ideology and Republican mass constituent ideology. On the Democratic side of things, there is a significant decline in the polarization of representatives and constituents in the House (-0.008) and the Senate (-0.001). The Democratic models range between 20% and 35% of the variance explained in the distance between constituents and representatives.

Table 12.6: Models Regressing Squared Difference b/w Normalized Constituent Ideology and Normalized Legislator Nominate Score by Year

MODEL: DV = B0 + B1(YEAR) + e	N	Intercept	Parameter Estimate		Standard Error	R2
DV: (HR ConIdeo Z– HR Nom Z)2	17	-24.74	0.013	***	0.003	0.507
DV: (HD ConIdeo Z – HD Nom Z)2	17	14.581	-0.008	**	0.004	0.227
DV: (SR ConIdeo Z – SR Nom Z)2	17	-8.526	0.005		0.003	0.135
DV: (SD ConIdeo Z – SD Nom Z)2	17	2.440	-0.001	**	0.867	0.341

* significant at .10 level

** significant at .05 level

***significant at .01 level

Why do the House models outperform the Senate models for the Republicans? Indeed, the coefficients are larger for the House models over the Senate models for both political parties. The effect of redistricting (gerrymandering) is a likely suspect. Over the time series under consideration we have a significant change from more Democrats and fewer Republicans in the earlier periods to, as a consequence of the Republican revolution, more Republicans and fewer Democrats in the later periods for the U.S. Congress. Thus there were almost certainly more Democrats representing marginal districts and states in the early period of the time series, while more Republicans represented marginal districts and states in the later periods. More marginal districts and states would have the consequence of more Democratic-leaning constituents for Republicans in the later period and vice versa in the earlier period. This would explain the significant positive coefficients in the Republican models and the negative coefficients in Democratic models. A more nuanced constituent-based argument for the partisan and ideological polarization would require looking at only same-party identifiers among constituents. This more thorough cut of the data, along with an analysis that takes into account district characteristics, will be necessary before any definitive conclusion could be reached.