Macropartisanship Revisited

Abstract

Canonical work argues that macropartisanship—the aggregate distribution of Democrats and Republicans in the country at a given time—is responsive to the economic and political environment. In other words, if times are good when Democrats are in charge (or bad when Republicans are in charge), more Americans will identify with the Democratic Party. We extend the pioneering work of MacKuen, Erikson, and Stimson (1989), who analyzed macropartisanship from 1953 through 1987, to 2021, assessing whether consumer sentiment and presidential approval still influence macropartisanship in an era of nationalized elections and affective polarization. We find that change has occurred. The effect of consumer sentiment on macropartisanship is no longer statistically distinguishable from zero, and we find evidence of “structural breaks” in the macropartisanship time series. Macropartisanship appears to have become less responsive to economic swings; approval-induced changes in macropartisanship have become more fleeting over time.

Prominent theories of partisanship suggest that party identification is responsive to the broader political and economic environment. At the aggregate level, scholars have shown that the share of Democrats and Republicans in the electorate is a function of economic perceptions and presidential approval such that when times are good, people increasingly affiliate with the president’s party (MacKuen, Erikson, and Stimson 1989 [hereafter MES]).¹ These ideas about aggregate-level partisanship, dubbed “macropartisanship,” have profoundly influenced the study of political opinion; to date, they represent the most important empirical challenge to the long-standing characterization of party identification as a stable attachment that is largely unresponsive to short-term forces (Campbell et al. 1960).

However, contemporary politics may differ in fundamental ways from the politics of the 1970s and 1980s, when these theories were developed and tested. For example, the mass public today is affectively polarized, expressing increasingly negative views toward the opposition party, its leaders, and members (e.g., Iyengar, Sood, and Lelkes 2012; Iyengar and Westwood 2015). Perhaps as a consequence, Americans are increasingly voting for candidates of only one political party (Hopkins 2018; Jacobson 2015). Indeed, even those too young to vote seem to have more clearly defined attitudes toward the parties than a generation ago (Tyler and Iyengar 2023).

Such findings raise questions about macropartisan responsiveness in the modern political era. Put simply, it is unclear whether there are enough “floaters” who may be induced to change their party identification on the basis of short-term evaluations of economic and political conditions (Converse 1962). Instead, Americans today may simply be too resolute in their commitment to their preferred party to attend to performance indicators, such as the state of the economy (Ellis and Ura 2021).

We extend data and analysis to the “modern” era, assessing the relationship between the economy and macropartisanship through early 2021. Our analyses search for a change in the extent to which consumer sentiment and presidential approval influence macropartisanship. Overall, we find somewhat weaker effects of consumer sentiment—our indicator of economic conditions—on macropartisanship in the modern era, relative to the period studied by MES (the early 1950s to late 1980s). The estimated effect of consumer sentiment on macropartisanship post-1987 is no longer statistically significant. Further, presidential approval remains predictive of macropartisanship, but approval-driven changes to macropartisanship seem to dissipate more quickly in the modern period.

Another approach, which searches for evidence of structural change in the macropartisanship series both during and after the period studied by MES, reveals evidence of structural breaks that occurred roughly in 1966, 1984, 1995, and 2006. Still, different tests for structural breaks reveal somewhat different breakpoints. In sum, we find suggestive evidence, however equivocal, that the short-term influences of aggregate partisanship in the country have changed since MES’s analysis.

Previous Evidence of Macropartisan Change

MES were among the first to apply time-series methods to aggregate partisanship survey data, leveraging the fact that the Gallup Poll had conducted at least quarterly measurements of party identification dating back to the beginning of the Eisenhower administration. Operationalizing macropartisanship as the ratio of self-identified Democrats over the sum of Democrats and Republicans, they pointed out that aggregate partisanship moved faster and further than would be expected if population replacement were the sole factor perturbing lifelong party attachments.

Drawing on the extensive literature on aggregate pocketbook voting (Kramer 1971; Tufte 1978), MES sought to test whether changing economic evaluations affected macropartisanship—in other words, whether macropartisanship is buffeted by changing performance evaluations. Specifically, they predicted that a strong economy should burnish the image of the party in control of the presidency; conversely, poor economic conditions should reduce the share of the public that identifies with the president’s party. A second hypothesis is that presidential approval, which reflects both economic and noneconomic performance evaluations, influences macropartisanship. All else being equal, the share of Democrats should rise when a Democratic president becomes more popular or when a Republican president becomes unpopular. ² Their analysis showed support for both of these claims using data from 1953 through 1987.

What Has Changed?

To our knowledge, there has been no comprehensive assessment of whether the original findings—that the aggregate distribution of partisanship in the country depends on economic evaluations and presidential approval more generally—still hold true today. There are reasons to suspect that macropartisanship may have become less responsive to short-term forces.

First, affective polarization has increased in the American public, particularly post-2000 (Iyengar, Sood, and Lelkes 2012; Iyengar and Westwood 2015). Specifically, partisans are increasingly viewing the opposition party more negatively than their own, and political behavior has become increasingly driven by negative attitudes toward the opposition (Abramowitz and Webster 2016). At the same time, election outcomes have become nationalized, such that citizens tend to vote for candidates of one party across all elective offices (Hopkins 2018; Jacobson 2015). As a consequence, there is much less ticket splitting, and more party loyalty, than ever before. For macropartisanship, the rise of distaste for the other party may decrease the likelihood that partisans switch sides in response to economic or political events.

Second, the increasing availability of partisan media may limit some citizens’ exposure to objective information about the opposition party or the state of the country (Prior 2007; Stroud 2011). For instance, recent work finds that exposure to CNN as opposed to Fox News during the Trump administration changed viewers’ perceptions of facts and also their evaluations of elected officials (Broockman and Kalla 2022). While perceptions of the state of the economy have surely always been influenced by party identification (Bartels 2002; Campbell et al. 1960; Gerber and Huber 2009), the opportunities for selective exposure in today’s information environment may dampen the connection between economic swings and macropartisan attachments. We note too that scholars have made progress in identifying when these changes have occurred—in other words, when political attitudes and behavior became so tethered to party. Some of this work identifies the rise of Newt Gingrich and the Republican takeover of the House in 1994, coupled with the introduction of Fox News, as an inflection point when elite polarization and conflict accelerated (e.g., Lee 2016; Theriault 2013), and in turn changed the behavior of the electorate. For example, Bonica and Cox (2018) show that since 1994, voters no longer punish candidates for ideological extremity and rather vote exclusively on the basis of national party preferences. Likewise, Hamel and Miller (2019) show that after 1994 politicians engulfed in scandal have experienced a much smaller decline in their vote share than they did previously. They also show that partisan campaign donors give more to scandal-tainted politicians today, suggesting that partisan donors today defend rather than punish bad behavior. Our goal in this essay is to extend previous work and reassess how consumer sentiment in particular shapes macropartisanship. That is, what has changed (and when) since the ink dried on MES’s analysis? We thus not only update the findings of a canonical study, but build on newer work on the malleability of partisanship in the modern era.

Data

The three key time-series variables in MES’s model are macropartisanship, consumer sentiment, and presidential approval. Each series raises some subtle measurement issues. Macropartisanship has provoked a certain amount of controversy because of the manner in which the party-identification question has been asked by the Gallup Poll. Unlike the American National Election Study, which asks “In general when it comes to politics, do you think of yourself as a Republican, Democrat, Independent, or what?” Gallup begins its version of the party-identification question with the phrase “In politics as of today.” By calling respondents’ attention to the present moment, the Gallup wording arguably accentuates the effects of recent events and short-term forces, a point demonstrated by Abramson and Ostrom (1991). Another complication is the change in survey mode over time, as Gallup moved from face-to-face interviews to phone interviews in the 1980s. ³

Approval presents a challenge insofar as it is measured using the same Gallup surveys used to measure macropartisanship. The repurposing of the same polls to create both a left-hand side and right-hand side measure runs the risk of exaggerating the apparent effect of approval, since a poll that draws an extra dollop of Democrats will also draw higher approval of a Democratic president. One way around this issue is to use approval measured by a different poll, such as CBS/New York Times (Green, Palmquist, and Schickler 1998), but this approach will not work here because no other poll runs concurrently with Gallup from 1953 through 1987, or from 1988 to 2021. ⁴

We therefore follow MES in employing quarterly data from the Gallup time series for both our measures of macropartisanship and presidential approval, recognizing that this approach may tend to overstate the causal influence of approval. We begin with data from Smidt (2018), extending that series from its end point of 2012 forward to 2021 using all available Gallup polls conducted in those subsequent years, placing them within quarters, and calculating quarterly averages for our variables of interest. We adjust for Gallup’s shift from face-to-face polling to phone polling by using concurrent surveys that employ both modes to calibrate the effects of this mode change. ⁵

To measure economic perceptions, MES use the Michigan Index of Consumer Sentiment (ICS), which conveniently spans the entire time period from 1953 onward. This measure combines survey responses to questions about the financial condition of respondents and their perceptions about current and future economic conditions. Enns, Kellstedt, and McAvoy (2012) show that this index tracks closely with subjective evaluations of the economy as measured by Gallup, but the latter measures are only available from 1985 onward.

Although a case could be made for augmenting or substituting measures, for present purposes our priority is to compare time-series relationships over time, which is best done with a consistent set of measures spanning the longest possible time period. Thus, our analysis leaves the basic variables used by MES largely intact (apart from the necessary mode adjustment to macropartisanship) and simply extends them forward in time. Descriptive statistics can be seen in table 1.⁶

Table 1 Descriptive Statistics

Note: “party-adjusted” variables are multiplied by -1 for quarters during Republican administrations.

Visualizing Over-Time Patterns

Before delving into statistical analyses of these time series, we begin with a graphical depiction of the macropartisanship time series covering the period from Eisenhower’s inauguration to Biden’s. Figure 1 displays the full series, with a vertical line demarcating the dates covered by the original MES analysis, which runs through the fourth quarter of 1987. To the right of the vertical line is the “Modern” period (from the first quarter of 1988 through the first quarter of 2021).

Figure 1 Macropartisanship, 1953–2021

Note: macropartisanship is the proportion of party identifiers who are Democrats, based on data from Gallup. This quarterly time series has been adjusted to account for Gallup’s shift from in-person to phone surveys.

With the exception of the marked shift in favor of Republicans during the 1980s, the two periods are broadly similar in terms of amplitude. Although we see short-term fluctuations in both the 1953–87 and 1988–2021 periods, they are relatively small in magnitude. The median quarterly change (the absolute value of current minus lagged macropartisanship) is 1.07 percentage points in the MES period and 1.08 points in the Modern period. The median yearly change (the absolute value of current minus four-lagged macropartisanship) is 1.87 points for the MES period and 1.71 points for the Modern period.

Figure 2 arrays all three time series for each presidency from Eisenhower through Trump. As expected, approval and consumer sentiment move reliably together for most presidents. For example, the decline in consumer sentiment during the first three years of the Carter administration coincided with erosion of Carter’s approval ratings. By contrast, the connection between short-term forces and macropartisanship is subtle. Declines in the popularity of Democratic presidents, such as Johnson or Carter, diminished the share of Democrats. Conversely, declines in the popularity of Republican presidents, such as George H. W. Bush and George W. Bush, coincided with an uptick in Democratic identifiers. On the other hand, the pattern of improving economic conditions under Clinton and Obama seems to have contributed little to the share of Democratic identifiers. The same holds for deteriorating economic conditions under Ford or Trump. The next section attempts to quantify the extent to which short-term forces had different effects on macropartisanship during the two periods.

Figure 2 Macropartisanship, ICS, and Presidential Approval: Eisenhower to Trump

Note: macropartisanship (the proportion of Democrats among respondents who identify with a major party) and presidential approval are measured by the Gallup Poll. Consumer sentiment is measured by the Michigan Index of Consumer Sentiment, divided by 100.

Statistical Model

Our workhorse model will be a minor variant on the model originally presented by MES. ⁷ Using the subscript t to denote each quarterly observation, the regression of macropartisanship (M_t) includes a lagged dependent variable (M _t−1), a party indicator (P_t), an interaction between the Index of Consumer Sentiment (I_t) and the party indicator, and a disturbance term: ⁸

(1) $$ {M}_t={\alpha}_0+{\alpha}_1{M}_{t-1}+{\alpha}_2{I}_t{P}_t+{\alpha}_3{P}_t+{u}_t $$

The party indicator is scored 1 if the president is a Democrat and -1 otherwise. This indicator is interacted with consumer sentiment so that good economic times under Democratic presidents increase macropartisanship, while good economic times under Republican presidents decrease it.

The basic structure of the regression model remains the same when we instead regress macropartisanship on presidential approval (A_t):

(2) $$ {M}_t={\beta}_0+{\beta}_1{M}_{t-1}+{\beta}_2{A}_t{P}_t+{\beta}_3{P}_t{+}{u}_t^{\prime } $$

Again, the party indicator is interacted with presidential approval so that popularity of Democratic presidents increases macropartisanship, while popularity of Republican presidents decreases it. ⁹

Results

We begin by estimating the parameters in equation 1, which describes the relationship between consumer sentiment and macropartisanship. The first column of table 2 shows the estimates for the entire quarterly time series, from 1953:1 through 2021:1. The lagged dependent variable is strongly predictive of next-period outcomes, consistent with an extensive literature showing that partisanship in the United States re-equilibrates slowly in the wake of a short-term shock. The coefficient (0.9180, SE = 0.0235) implies that a one-unit shock to macropartisanship has an enduring effect of 0.9180¹⁶ = 0.2544 units four years later. Put another way, any perturbation that moves macropartisanship by four points at time t will continue to move it by one point at time t + 16. Given the extent to which effects add up if sustained over time, care must be taken not to dismiss the effects of short-term forces as inconsequential based solely on their regression coefficients.

Table 2 Predicting Macropartisanship: Models with Consumer Sentiment or Presidential Approval, by Era

^∗∗∗p < 0.001; ^∗∗p < 0.01; ^∗p < 0.05

Note: consumer sentiment and presidential approval are multiplied by -1 during Republican administrations. MES refers to 1953:1–1987:4; Modern refers to 1988:1–2021:1.

The estimated slope for consumer sentiment is 0.0248 (SE = 0.0088), which implies that a one standard-deviation shift in this index (0.116 index units) has an immediate effect of 0.0248 ∗ 0.116 = 0.0029 on macropartisanship, which has a standard deviation of 0.045. Although such an effect would easily escape detection in a tracking poll, the effects of a sustained shift in consumer sentiment are more substantial. For example, when Carter took office in 1977, ICS was close to its mean (0.87) but fell by one standard deviation in 1978 and by another standard deviation in 1979, where it remained roughly through the end of his presidency. According to the regression model, the cumulative effect of this sustained period of economic pessimism over the course of Carter’s presidency was a decline in macropartisanship of 0.036, which would have been detectable by the surveys of the time. Columns 2 and 3 of this table suggest that the effects of consumer sentiment may have diminished over time. During the MES period, the immediate effect of a one-unit change in consumer sentiment appears to be 0.0383 (SE = 0.0131). By comparison, the apparent effects of consumer sentiment seem to be weaker during the Modern period. The immediate effect of a one-unit change in consumer sentiment appears to be just 0.0121 (SE = 0.0115).

Not only is the instantaneous effect of consumer sentiment weaker in the Modern period, the effects of lagged macropartisanship are weaker as well, which implies that the effects of economic shocks dissipate more quickly. To return to the example from the Carter era, how much would macropartisanship have changed had this economic contraction occurred in the post-MES period? According to the regression model in column 3, the cumulative decline in macropartisanship would have been only 0.009.

One way to test whether the coefficients in columns 2 and 3 differ by more than chance is to perform a Chow test, which in essence compares the sum of squared residuals from the pooled regression in column 1 to the sum of squared residuals of columns 2 and 3 combined. An F-test assesses whether the four additional parameters used to characterize macropartisanship in the two periods significantly improve the fit of the model. The answer appears to be yes: F(4, 264) = 4.17, p = 0.003. Although the apparent effect of ICS is not significantly different in the two periods (two-tailed p = 0.134), the overall pattern of coefficients changes significantly.

Turning our attention to the effects of presidential approval, we again see evidence of significant effects when analyzing the entire time series. The immediate effect of a one-point increase in A_t (equivalent to a 100 percentage-point change in raw approval ratings) is 0.0362 (SE = 0.0086). Because the lagged effects of macropartisanship are strong ( $ \hat{\alpha} $ ₁ = 0.8988), a sustained increase in approval can have substantial cumulative effects. For example, George W. Bush started his second term in office close to the grand mean for approval (53.5) but gradually slid into the high 20s by the end of 2008. According to the pooled model in column 4, the cumulative impact of this four-year run of subpar approval ratings was a 0.0567 gain in Democratic macropartisanship.

When we partition the data into two periods, MES and Modern, the coefficients change markedly. The instantaneous effect of approval rises from 0.0452 in the MES period to 0.0779 in the Modern period. That increase seems to suggest that macropartisanship grew more responsive to approval, but the lagged effects of macropartisanship complicate the interpretation. The estimated effects of lagged macropartisanship fall from 0.8239 (SE = 0.0454) to 0.4983 (SE = 0.0663), implying that approval’s effects now dissipate much more quickly. Returning to George W. Bush’s second term, the cumulative effects are predicted to be 0.0517 under the MES-era regression but 0.0380 under the Modern-era regression. The Modern period seems to feature quicker initial adjustment but also faster decay. Overall, a Chow test of structural change is highly significant: F(4, 264) = 10.68, p < 0.0001.

Estimating Structural Breaks

The previous section looked for signs of structural change by comparing the MES period to the Modern period. An alternative approach is to pool the full time series and use statistical tools to detect changes in model fit at any point in time. The strucchange package in R (Zeileis et al. 2002) supports this approach. We test for structural breaks with equation 1 over the entire time series using a moving window with a bandwidth that examines 20% of the data at a time. We set the number of breaks to investigate at four—the maximum number the software allows us to search for. This approach allows us to see search results had we looked for one, two, three, or four structural breaks.

We cross-validate this search with the framework developed by Bai and Perron (1998; 2003) for detecting unknown structural breaks and then forming 95% confidence intervals around when they occur. This approach, which is implemented in the Stata package xtbreak, outputs the apparent number of breaks, when they are said to occur, and the confidence intervals around these estimated occurrences. ¹⁰ Xtbreak allows up to five breaks. Typically, the more breaks one investigates, the more uncertainty there will be about their timing. Recognizing that identifying the timing of multiple structural breaks may overtax the available data, we focus primarily on whether the two approaches tell a similar story about structural change.

The results from these methods are summarized in table 3. Applying the xtbreak framework to equation 1 (consumer sentiment) detects five structural breaks over the entire time series: 1963, quarter 2 (prior to Johnson’s legislative juggernaut); 1974, quarter 2 (Watergate); 1984, quarter 2 (as Reagan expanded the Republican Party’s voter base en route to a landslide reelection victory); 1996, quarter 1 (a period of conflict between Clinton and House Republicans); and 2006, quarter 1 (George W. Bush’s slide in popularity). As we search for fewer breaks—moving from five to one—we see the breaks from the early 1960s, mid-1990s, early 2000s, and early 1970s disappear in turn.

Table 3 Location of Structural Breaks

Note: “xtb” denotes results from the xtbreak package in Stata. “sc” denotes results from the strucchange package in R. Strucchange does not support searches for more than four breaks.

The strucchange package supports a search for up to four breaks. As table 3 indicates, the strucchange search reveals a substantial structural break in late 1966, around the time Johnson had secured major legislative victories, including the Civil Rights Act of 1964, the Medicare and Medicaid Act of 1965, and the Voting Rights Act of 1965. Regardless of the number of breaks we search for (including one), the strucchange package finds a structural break occurring in either the third or fourth quarter of that year. If we search for two breaks, we find one in 1966 and another in 1995, quarter 4 (the Clinton–Republican conflict). A search for three breaks finds them in 1966, quarter 3; the third quarter of 1983 (the Reagan realignment); and the fourth quarter of 2005, when George W. Bush’s popularity began to slide. Finally, a search for four breaks finds them in 1966, 1980, 1994, and 2007.

Both the xtbreak and strucchange packages reveal a similar pattern when we use presidential approval as a predictor of macropartisanship instead of consumer sentiment. Indeed, in all cases but one the break points identified for presidential approval are within two years of those for consumer sentiment, and in most cases they are exactly the same. The one exception is the strucchange search for one break: the strucchange package finds that break in 1974 for presidential approval, and in 1966 for consumer sentiment.

The two methods for detecting structural breaks do not agree on the precise timing of post-MES changes, but they both point to the first Clinton administration and the second Bush administration as turning points. The two statistical packages also turn up ample evidence of structural change during the period that MES studied. Both approaches find evidence of structural change during the early to mid-1960s, and both suggest that change was afoot sometime between 1980 and 1984. The two approaches diverge mainly on the question of whether a structural break occurred during Watergate. Interestingly, neither method detects a structural break during Trump’s rise to the presidency or during his term in office.

Conclusion

MES offered an innovative approach to the study of partisanship. While most previous efforts focused on explaining party identification at the individual level (e.g., Campbell et al. 1960), their approach used pooled cross-sectional surveys and time-series methods to assess whether and how American partisanship in the aggregate adjusts in response to short-term shocks such as changes in consumer outlook and approval of the sitting president. While work at the individual level tends to characterize party identification as highly stable because respondents in panel surveys tend to maintain their relative positions along the partisan spectrum even if the overall mean changes from one period to the next (Green and Platzman 2022; Green and Yoon 2002), this approach described the aggregate distribution of partisans as responsive to the economic and political environment.

Our goal in this essay has been to reexamine the evidence for short-term influences on macropartisanship in the contemporary period. Our curiosity about structural change stems from the sense among political observers that the character of American politics is changing. Partisanship today seems more contentious than before, and attitudes toward opposition party elites and their partisan supporters are increasingly negative and hostile. At the same time, the parties themselves are in flux. The rise of Donald Trump brought about a new rhetorical style and a new set of policy positions; challenges from the progressive wing of the Democratic Party have come close to upending the party’s leadership and policy priorities. It is unclear whether this blend of internecine struggle and cross-party animus has made the public’s partisan attachments more or less responsive to economic events or other drivers of presidential popularity. Given that we now have roughly twice as many time-series observations as were available to MES, it seems an appropriate moment for an empirical reassessment.

Our evidence is suggestive but equivocal. We see some indications that the time-series parameters of the canonical macropartisanship model are changing, but the statistical evidence is less than decisive. Inspection of the estimates over time suggests the declining influence of consumer sentiment. We notice some structural changes in the relationship between approval and macropartisanship as well, but because the autoregressive character of macropartisanship is weaker in the recent period, the overall pattern is complex. We tentatively conclude that over time macropartisanship has become less responsive to changes in consumer sentiment. Changes in presidential approval continue to influence macropartisanship strongly, but these effects subside more quickly than in the past. One interpretation is that the contemporary news cycle moves from topic to topic more rapidly than in the past; another is that presidential evaluations, when moved, snap back to their equilibrium more quickly in the current polarized environment (Donovan et al. 2020). ¹¹

Omnibus tests for structural breaks uncover some intriguing evidence that the macropartisanship series has undergone a number of changes at crucial historical junctions. Given the evident change in party identification during the Reagan era, it is not surprising that tests of structural change point to the mid-1980s. But there are also hints that the structure of macropartisanship changed sometime between 1994 and 1996. Since this break was detected empirically, we can only speculate about its meaning, though as noted previously, earlier research has also pointed to this time period as an inflection point in political behavior (Bonica and Cox 2018; Hamel and Miller 2019). Indeed, the most obvious interpretation attributes the break to Newt Gingrich’s historic victory in 1994, ending Democrats’ long-standing control of the House of Representatives and ushering in a new era during which Republicans competed vigorously for unified control of government. Perhaps this was the harbinger of the bitter struggles to come, as both parties focused less on valence issues such as economic stewardship and turned toward appealing to core supporters.

More generally, our essay demonstrates the importance of revisiting canonical findings as new data emerge. While updates of this kind are perhaps not as enticing as altogether new questions and avenues of inquiry, our understanding of politics depends on how well our theories and evidence keep up with the political times. Given that “the times” have changed markedly since many of the authoritative findings in our field were first uncovered, it is critically important for their claims to be continually assessed.