Affective Polarization in the Canadian Party System, 1988–2021

Abstract This article brings three decades of broadly consistent survey data on survey respondents’ feelings about the parties as evidence of affective polarization. It also presents evidence about policy differences among the parties and makes an explicit link between elite and mass data with multilevel modelling. The article shows that affective polarization is real and also demonstrates its connection to the ideological landscape. But it also shows that conceptual categories originating in the United States must be adapted to Canada's multiparty system and to the continuing contrasts between Quebec and the rest of Canada. It suggests that accounts of Canada's twentieth-century party system may not apply to the twenty-first century.


Introduction
Canadian parties and their supporters are going down a road of affective polarization, or so it seems. This claim is a regular media trope, and supportive evidence is accumulating in academic research. This includes comparative evidence (for example, Boxell et al., 2020;Gidron et al., 2020), which places Canadian polarization near the centre of the global pack. Canadian trends (Cochrane, 2015) seem roughly parallel to those in the United States (Iyengar et al., 2012). But conceptual guideposts from the United States presuppose a nationalized two-party system and a separatedpowers framework, which do not travel well to a multiparty parliamentary system with an electorate segmented between Quebec and the rest of the country. Moreover, US research typically ignores the large body of nonpartisans. The comparative literature addresses some of these gaps but at the cost of suppressing depth and context. And that literature makes contradictory claims about trends.
This article addresses these shortfalls by mobilizing three decades of broadly consistent Canadian survey data on feelings about the parties. It sets up comparisons across each Canadian party and its supporters by presenting evidence for each entity separately but in juxtaposition. It does so in parallel for Quebec and the rest of Canada. It links trends in affect to changes in parties' ideological positioning.
The article confirms that, in the main, affective polarization has increased, both inside and outside Quebec. This is true even though some parts of the system have depolarized. Shifts in affect correspond to shifts in the policy gaps among the contestants for power. The article provides suggestive evidence that the link between policy and affect is causal: policy drives affect. It concludes with a discussion of further implications, including for our models of the Canadian party system.

Conflicting evidence on trends
Scholarly evidence tilts toward increased polarization in Canada, but the record is mixed. Certainly, negativity toward rival parties is a theme in recent work-for example, by Caruana et al. (2015) and McGregor et al. (2015). This negativity radiates into social relations and perceptions: partisans of one party overestimate their policy differences with other-party supporters and allow such considerations to affect even the choice of romantic partners (Merkley, 2021b).
These patterns seem to correspond with those in the United States (Iyengar, et al. 2012). In contrast to feelings toward ethnic and religious groups, which are constrained by considerations of social desirability, disdain toward political parties is unrestrained: Unlike race, gender, and other social divides where group-related attitudes and behaviors are constrained by social norms . . . there are no corresponding pressures to temper disapproval of political opponents. If anything, the rhetoric and actions of political leaders demonstrate that hostility directed at the opposition is acceptable, even appropriate. Partisans therefore feel free to express animus and engage in discriminatory behavior toward opposing partisans. (Iyengar and Westwood, 2015: 690) Cochrane (2015: Figure 8.4) argues that these trends also appear in Canada. But comparative work that includes Canada is less clear. Supporting Cochrane's view is Boxell et al. (2020: Figures 1 and 2). In their nine-country dataset, the only other systems with increasing polarization are the United States and Switzerland. The same study (Boxell et al., 2020: Table 1) places Canada among the most polarized-more polarized than the United States. 1 But Gidron et al. (2020: Figure 1) place Canada among the less polarized and find that Canadians have become less polarized, not more.
This confusion reflects variation in method and in data sources. Multicountry studies rely on index numbers for comparison, but the indices differ in how many parties enter the calculation, whether parties are weighted, how they are weighted, and whether nonpartisans are included. They also differ in the timescale for imputing trends. Cochrane's work spans 43 years, Boxell et al. cover 30 years,and Gidron et al. cover about 20 years. Comparative datasets can include studies that differ by fieldwork mode and, by implication, sample composition. Some datasets focus on preelection data, others on post-election data and some combine the two.

Affect and ideology
It is natural to ask whether affective polarization reflects substantive policy differences among parties. Decades of research on belief systems in mass publics might make us hesitate. Converse (1964) painted a bleak picture for the United States, and there is no reason to think that his conclusions do not apply to Canada. His core finding is that survey respondents struggle with even minimal notions of ideology. Such structure as appears is a by-product of group sentiment. In the United States, the widening affective gap between partisans finds no parallel in the gaps between self-described liberals and conservatives (Iyengar et al., 2012: Figure 2;Kinder and Kalmoe, 2017: Figure 5.3). Similarly, mutual antipathy masks underlying similarities between the rival parties' mass bases (Mason, 2015). Thus far, the active ingredient would seem to be partisanship as such.
Evidence on the other side is piling up, however. Wagner (2021), Ward and Tavits (2019) and Medeiros and Noël (2014) all demonstrate links between ideology and affect. Less clear is what causes what. For Medeiros and Noël, the path leads from ideology to affect. For Ward and Tavits, the opposite is true. Wagner models the relationship as if ideology follows affect, but he does not insist on a causal interpretation. And to the extent that all this work relies on respondent self-reports for both sides of the equation, the findings are inherently ambiguous.
The ambiguity is clearing up, however. Druckman and Levendusky (2019) show that social distance between partisan groups is only incidentally about citizens en masse; the essential focus is on elites. A Canadian example of the elite focus turns on the country's multiparty context: McGregor et al. (2015) show that the driving force of partisan negativity is primarily ideological distance, defined not by self-reports but by the publicly verifiable left-right ordering of parties. Gidron et al. (2020: Rogowski and Sutherland (2016) and Webster and Abramowitz (2017) zero in on the causal story by manipulating policy distances experimentally. Lelkes (2021) extends the argument by excluding mere party identification as the key driver. For many citizens, the distance between the parties has rendered perceptual metrics superfluous. The consistency of party policy differences enables candidates to gain a "reputational premium" (Sniderman and Stiglitz, 2012): if the candidate is on the correct side (as defined by the party ordering) relative to the other party's candidate-even by just a little bit-he or she gains the premium. Voters are not required to know much, just to get the order right-and this is now easy to do. The pattern is such that "ideology is now a 'natural' part of [US voters'] psychological functioning and will always be present in one form or another" (Jost, 2006: 667).
The elite pattern for Canada has striking similarities to that in the United States. Starting in the 1970s, the major parties diverged on a left-right axis (Cochrane, 2010: Figure 1;Cochrane, 2015: Figures 8.1 and 8.2;Johnston, 2017: Figure 4.1), and voters, especially "near elites," noticed the divergence (Cochrane, 2015: Figure 8.3). This has produced yet another parallel to the United States: voters are "sorting." In the United States, the issue and ideological overlap between Democratic and Republican identifiers has shrunk (Levendusky, 2009). In Canada, this seems to be happening for left-right self-placement (Johnston, 2014;Merkley, 2022) and extends to substantive policy domains (Kevins and Soroka, 2018;Merkley, 2022).
Either way, increased consistency between party and ideology ups the affective stakes: If political identity is a substantial driver of ingroup bias, activism, and anger, why would these things increase over time? . . . the answer has largely to do with changes in the alignment of partisan and ideological identities over time. Sorting has brought our ideological and partisan identities into agreement, and this new alignment has increased the strength of those identities. (Mason, 2018: 130-131) This is an argument about the net reduction of "cross-pressure," 3 as memberships or identities pull less and less in opposite directions (Brader et al., 2014).

Where things stand
All this said, the Canadian evidence to date is scattered and yet is over-integrated. Evidence has accumulated on affective gaps, and also on ideological ones. But the observation of trends has not attended closely to the actual time path. And talk of widening gaps obscures the fact that some gaps have narrowed. The focus on gaps has neglected the more general patterns of overall variance in party evaluation and the relative balance of positivity and negativity. As for affect and ideology, only one account (Cochrane, 2015) juxtaposes them and is clear on timing. But the actual connection between the two is left to the reader's eyeballs. Meanwhile, Canada's place in the comparative literature is confused by differing treatments of its multiparty character and by varying timescales. If that literature -some of it, at least-is properly attentive to the country's multipartism, the impulse to compare has suppressed critical interpretive detail. And it is weak on causal inference.
All this neglects the fact that the electorate is segmented between Quebec and the rest of Canada. Historically, polarization in Quebec has not been over left and right but over the "national question" (Beauchemin, 2015;Nadeau, 1992): How should Quebec be situated in the federation or should it belong to the federation at all? On this the Liberals have been "the party of Canada" (Johnston, 2017: chap. 4). If any big party took the other side, it was Conservatives, notwithstanding resistance from the party's supporters outside Quebec (Johnston, 2017: chap. 4). In 1993, that position was assumed by the Bloc Québécois.
This puts the following questions on the table. The first four refer to the structure of affect itself: 1. Over all, has affect for the parties become more polarized? 2. Over all, has the balance between positivity and negativity shifted toward negativity? 3. Have gaps between party identification groups expanded? 4. Have nonpartisans become more generally more negative to parties?
Two questions pertain to the link between ideology and affect: 5. Is there a longitudinal relationship between party position-taking and affect toward each party expressed by party identification groups? 6. Is this relationship the result of elite induction?
For each question, it is natural to ask if the pattern is the same for all parties. Similarly, how much do patterns differ between Quebec and the rest, and are the Quebec/rest differences widening or narrowing?
Questions 1 through 4, about the landscape of affect, can be addressed with survey-based indicators. Questions 5 and 6, about the descriptive link between affect and ideology, require additional data on elites. To smooth the exposition, I juxtapose discussion of indicators and methods to accounts of results, stage by stage.

Survey sources and indicators
To get as stable a platform as possible, the focus is Canadian Election Study (CES) data from 1988 to 2019 inclusive. Before 1988 the CES mode was basically face-to-face (FTF), which dictated that fieldwork took place after the election. 4 The shift to the telephone mode (RDD) in 1988 enabled a pre-post design. RDD was supplemented in 2015 with an online component, and in 2021 fieldwork was online only. All analyses in this article use data from the campaign-period wave, to minimize projection from the election's outcome. The 1988 CES also brought a critical change in the measurement of party identification, the key moderator for this article.
The indicator of partisan affect is the so-called feeling thermometer, which ranges from 0 to 100, with 50 as the point of indifference. 5 I take the thermometer rating as given, which is to say I live with the fact that the indicator is vulnerable to differential item functioning (Wilcox et al., 1989). I make no attempt to control for the general warmth of a respondent's ratings or for their variance across objects. It is common to overcome some of these problems by taking differences across objects-for example, the gap in rating between one's own party and that of the main alternative. As my interest lies in relations among all the parts of the system, I do not want to mask any of those parts. The indicator is deployed in three ways: as a gauge of overall polarization, to assess the relative balance of positivity and negativity, and as the dependent variable in estimations where party identification is a moderator. 6 The direction of party identification is captured by the Canadian version of the classic indicator. Respondents are asked if they "think" of themselves in partisan terms, with the parties explicitly named and non-identification validated as an alternative. Without this validation, the indicator would be vulnerable to short-term forces (Johnston, 1992). This is interpretively critical because, as we shall see shortly, period-specific forces are clearly visible on the landscape.

Overall polarization
I address overall polarization in feeling with the standard deviation of the feeling thermometer, which gives a quick summary and facilitates visual comparison over time. In a summary sense, according to Figure 1, feelings about all parties in both Quebec and the rest of Canada have become more polarized. The 1993-1997 interval seems to be a break point for every party. After 2000, the Conservatives led the way. Dispersion in Conservative Party ratings jumped in 2004 and continued to grow almost year by year, such that by 2015 the standard deviation was half again larger than in 1988. For the Liberals and New Democratic Party (NDP), further increases are modest between 1997 and 2015. Between 2015 and 2019, however, the Liberals recorded a big jump.
The Conservative trend needs to be seen in tandem with the short history of Reform. The mere emergence of Reform increased the overall polarization outside Quebec. As with the other parties, Reform's indicator jumps in 1997. The merger of Reform and the Progressive Conservatives in 2003 incorporated this polarization into feelings about the new Conservative Party.
Trends in Quebec and the rest of Canada are broadly similar, but with the presence of the Bloc as a key qualifier. At every point, feeling is more polarized about the Bloc than about any other party in Quebec, or in the rest of Canada for that matter.
Alongside the general trend is an unanticipated subtheme: being in government ratchets up the dispersion. The first Liberal surge spans the first four years of the Chrétien government, and the same was true for the Trudeau years. For the Conservatives, the gains occasioned by the merger with Reform plateaued between 2004 and 2006. But thereafter, the now-governing Conservatives saw dispersion increase, even as the Liberals stalled. After 2015, polarization gains for the Conservative Party, now out of power, slackened. Instead, as we have seen, the biggest gain was for the Liberals, now back in government. Note: Dispersion indicator is the standard deviation. Underlying data include all respondents in each region.

Positivity/negativity
For the relative positivity or negativity of the charge, the main story is not so much a net shift in either direction as a hollowing out of the middle. Figure 2 presents the evidence in kernel density plots with vertical lines at scale values of 25, 50 and 75. These visual aids show relative concentrations of affect around benchmarks. Another visual aid is a horizontal line at 0.008, positioned to help interpretation of each distribution's tails.
Visually most striking is the massing at extreme negativity. In 1988, both major parties drew quite symmetrically distributed response, with neither tail markedly fatter than the other. The shift toward asymmetry began in 1993 and was mostly about the insurgents-Reform and the Bloc. The pattern now characterizes all the parties. Some shifted more quickly than others, and no year stands out as a break point. But by 2015 or 2019, a second mode, below 25 degrees, appeared for all the parties. Over the same period, the height of the main mode dropped. Clearly, the electorate harbours a lot of ill feeling.
But if the main mode for most parties has become less densely populated, it has come to represent a warmer value. This is not just an illusory by-product of reduced numbers around 50 degrees. In recent years, the share at or above 75 degrees is typically larger than in 1988.

Moderation by party identification
Unsurprisingly, much of the story in the preceding sections is driven by-or at least, is conditional upon-party identification. Evidence is in Figure 3. The impact of partisanship is shown by plots of conditional means and 95 per cent confidence intervals over consecutive years for each party identification group and for nonpartisans. To simplify the display, the nonpartisan group includes supporters of niche parties, most of which do not last beyond a single election. Also treated as nonpartisans, despite their staying power, are the Greens. Their numbers are always small, and the CES did not consistently capture their presence.

Canada outside Quebec
Both Liberal and Conservative identifiers warmed to their own party. To the naked eye, Conservative gains seem especially dramatic, but this is partly an illusion created by the dip in the 1990s. The net gain from 1988 is about five degrees. The Liberals actually have the greater cumulative 1988-2019 gain, about 10 points.
Trends in feeling toward parties other than the respondent's own depend on the pairing. Broadly speaking, all other partisan groups became more negative about the Conservatives. The sharpest drop was for Liberal identifiers, a drop of 10 degrees. Conservative supporters reciprocate the feeling, but especially toward the Liberal Party, a drop of about 15 degrees. Critically, Liberal/Conservative gaps have grown relative to the rest of the system. In 1988, Conservatives disliked the Liberals no more than did New Democrats or nonpartisans. In 2019, Conservatives dislike the Liberal Party far more than any other group does. And they dislike the Liberals more than they do any other party except the NDP. The feeling is mutual: where in 1988 Liberals felt about the Conservative roughly as nonpartisans did, now they dislike Conservatives as much as New Democrats once did. Reform identifiers made a double contribution to partisan negativity.  Note: Entries are mean thermometer ratings and associated 95 per cent confidence intervals. Colours and labels indicate identification group. Cells are keyed to the party being evaluated.
They were even more negative toward the Liberal Party and the NDP than Conservatives were. But they were also negative about the Conservatives, slightly less so than NDP supporters and more negative than Liberal identifiers.
The Liberal/NDP dynamic stands in marked contrast to all this. Gaps between the NDP and the Liberals have narrowed. Where Liberals used to dislike the NDP as much as they did the Conservatives, this is no longer true; indeed, their feelings are now net positive. Similarly, New Democrats have warmed to the Liberals, although their movement is less dramatic.
In general, nonpartisans seem mildly disapproving of all parties and have been so throughout. Only for the NDP is there a trend, a positive one.

Quebec
Patterns in Quebec are broadly similar to those for the rest of Canada but with twists related to the national question. Own-party ratings have become more positive, but only slightly so and mainly in comparison with dips in the middle of the series. Other-party ratings are also moving in the same directions as in the rest of Canada, if from different starting points. Every partisan group likes the Conservative Party less-a small shift for Liberal identifiers and big ones for New Democrats and Bloquistes. Every group (except Conservative supporters) came to like the NDP more. Gaps between the Liberals and Conservatives widened. The sympathy between Conservatives and Bloc that was evident in 1993 no longer holds, even as Bloquistes and New Democrats have become more sympathetic to each other. In sum, the national question is yielding ground by something like left-right ideology.
But not completely, thanks to the Bloc's continued presence. The party is quite distinctively disliked. The mutual antipathy of Liberals and the Bloc is essentially unchanged, and this is still the Quebec electorate's strongest polarity.
Quebec nonpartisans are clearly more negative toward the parties than those in the rest of the country, and they exhibit more trends: down for the Liberals, Conservatives, and the Bloc; upward only for the NDP.

Ideology and Affect
This section draws the connection between shifts in affect and shifts in parties' leftright positioning. This requires discussion of how to measure positioning, of how to capture an elite/mass link, and of estimation strategies. The ideological landscape is mapped with data from the Manifesto Project (Volkens et al., 2018;hereafter MARPOR), as used by Cochrane (2015) and Johnston (2017). The measure is a summary left-right scale (RILE), which aggregates party platform commitments across the broad range of cultural and economic policies following the classification scheme in Laver and Budge (1992). This particular indicator is a subject of controversy and requires some justification. It represents an attempt by the MARPOR group to extract policy positions out of data originally coded for emphasis. Although emphasis often signals position, it does not always do so. Critics of the scale argue that analysts should go further back into the data to rebuild the index from scratch, either by applying a pre-set coding scheme to individual words through text analysis (Laver and Garry, 2000) or by a purely inductive strategy implemented by exploratory factor analysis (Gabel and Huber, 2000). These strategies produce closer correspondence than RILE to a commonsense reading of political dynamics or to placements in expert surveys. Cochrane pushes even further into the question by querying whether ideational relations are symmetric between left and right. He finds that they are not-not for Canada (Cochrane, 2010: 592-99, especially Figure 3) and not for party systems in general (Cochrane, 2015). He proposes a coding scheme that accommodates this asymmetry (Cochrane, 2015). But none of these exercises produces an off-the-shelf dataset comparable to the institutionalized and ongoing MARPOR archive. Even the critics concede that in the aggregate, the traditional classification travels well with the alternatives (see, for example, Laver and Garry, 2000: Table 3; Gabel and Huber, 2000: 102). To the extent that RILE fails to capture the underlying concept, it works against my thesis. The MARPOR data are deployed as originally coded, then, as a proof-of-concept indicator. Party positions are specific to each year; no pooling across years is attempted, as there seems to be no principled basis for doing so.
Evidence for the full postwar period appears in Figure 4. 7 Before the 1970s the big parties essentially did not differ, as both sat slightly right of centre. The divide at the core of the old system was not over left and right but over the ethnic definition of the nationality (Johnston, 2017: chap. 5). To the extent that the system had ideological energy, its source was on the left, with the NDP and its predecessor, the Co-operative Commonwealth Federation (CCF). In the 1970s, the energy shifted to the other side, as the Conservative Party pulled to the right.
As a visual aid, 1988, the first year for the survey data, is marked. This, of course, is also the brink of a major disruption. By 1993, the year of the breakthrough by Reform, the Conservatives had moved far to the right. Not only was this insufficient  The pattern on the centre and left was the opposite. In most elections from 1979 on, the Liberals have been closer to the NDP than to the Conservatives, usually a lot closer. The convergence between these parties was not because the Liberals shifted left but because the NDP gradually drifted rightward, to the centre. The Bloc, meanwhile, was ideologically centrist all along. This crowding at the centre explains the finding in Gidron et al. (2020) that the Canadian system has depolarized.
Relationally speaking, however, the Canadian centre has emptied out. In most elections, the Liberal vote usually outweighs the NDP and Bloc votes, such that when distances are weighted, the gap between the two most plausible parties of government is much greater than before (Cochrane, 2015: 157). The temporary displacement of the Liberals by the NDP in 2011 did not alter this fact.

Model and estimation
In each estimation, the dependent variable is the thermometer rating for one party and the primary independent variable is that same party's left-right positioning. The impact of positioning is moderated by the same party identification groups as in Figure 3. Specifically, the party's left-right position in a given election is interacted with dummy variables for the party identification groups in that year's CES sample. The regression slope describing the impact of party on affect is thus allowed to vary across identification groups. My expectations for how identification conditions that slope are mapped in Table 1. As values on the left-right scale increase as a party moves to the right, the table couches predictions in terms of a rightward shift. All non-NDP identifiers should like the NDP more as it moves to the right (that is, toward the centre). All non-Conservative identifiers should like the Conservative Party less as it moves to the right (that is, away from the centre). I posit that nonpartisans' response to rightward shifts should be non-negative for the NDP and non-positive for the Conservatives. By implication, the response could simply be null. Supporters of the Bloc and the Liberals, as centrists, should be collectively ambiguous about rightward Cell entries indicate predicted impact on feelings of a rightward ideological shift by the party being evaluated, contingent on the respondent's party identification. For example, the column for the NDP shows that an NDP identifier is predicted to like that party less, the further to the right it is located, while Liberal and Conservative identifiers are predicted to like the party more. movement by their own party. I posit that supporters of the flanking parties, the NDP and the Conservatives, are more extreme than the parties themselves. On this assumption, as their respective parties move to the right, NDP supporters should like their party less and Conservative supporters should like their party more. As party positions are common for all respondents for a given year and because parties take positions in advance of elections, it seems intuitively correct to characterize the relationship as hierarchical and recursive. But if multilevel modelling seems conceptually appropriate, there are only nine observations per party at the election-year level (eight for the Bloc). What this implies for estimation is a matter of dispute. Stegmueller (2013) argues that conventional multilevel techniques yield both biased estimates and underestimated standard errors. He advocates Bayesian estimation. Elff et al. (2021) question Stegmueller's claim and advocate fitting models by restricted maximum likelihood. 8 I adopt the latter, as it is less computationally intensive. In the estimations, intercepts are allowed to vary as random effects, but variation in slopes is precisely what we are modelling.
All estimations are partitioned between Quebec and the rest of Canada. To the extent that the national question is more salient and the left-right axis less salient in Quebec than elsewhere (Johnston, 2008), left-right patterns should be sharper in the rest of Canada than in Quebec.
Some parties are excluded from presentations. The Bloc appears only in the Quebec estimations, obviously. Outside Quebec, Reform was a critical player in pushing the system to the right, but it contested only three elections overall and none since 2000. In Quebec, the NDP ran candidates over the entire period, but its band of identifiers was small. Hence, for Reform in all regions and for the NDP in Quebec, marginal effects estimates are not reported. 9

Results
All of this section's exhibits, in Figure 5, are organized by region and by the party being evaluated. Within a panel, the figure shows how a unit rightward shift in party positioning affects feeling toward that party, contingent on respondents' party identification.

The rest of Canada.
Patterns are broadly but not perfectly consistent with expectations. Consistency is greatest for the Conservative Party: its rightward movement makes its supporters feel warmer toward it and makes all others cooler. A unit shift to the right makes Conservatives' feeling toward their own party 0.27 degrees warmer. But if we compare on the extremes in Figure 4-the low in 2000, when the Conservative platform was more centrist than it had been for decades, and 2015, the party's rightmost positioning to date-the shift is 41 points, which translates into predicted gain for Conservative identifiers of 11 degrees: greater than the observed gain. For Liberals and New Democrats, in contrast, a unit rightward shift induces decrements of 0.34 and 0.38, respectively, in their feelings for the Conservative Party. The party's total 2000-2015 shift is estimated to have cooled their rivals' feelings about 11 degrees. This is the entire observed shift among Liberal identifiers and about two-thirds of that among New Democrats. For the NDP, the overall pattern is mixed but the message is clear. For Liberal identifiers, a one-unit rightward shift in NDP positioning induces a 0.42 increase in affect toward the NDP. The maximum policy shift, which happens to be between the 1988 and 2015 end points, is 31 units. On this count, the NDP shift to the centre predicts a 13-degree gain in Liberal warmth of feeling. This is larger than the observed gain. Conservative identifiers seem indifferent to the NDP's location. A possible explanation is that Conservative supporters have also moved to the right (Merkley, 2021a), such that the gap just did not close. Contrary to my assumption, NDP identifiers may not sit to the left of their party: a rightward shift makes them like it more (although the confidence interval for the coefficient overlaps zero).
For feelings about the Liberal Party, the main story is about Conservative identifiers. A unit rightward shift of the Liberal platform would induce a 0.37 gain in their affect for the Liberals. In fact, according to Figure 4, the Liberals have moved more to the left than the right. The biggest gap is between 1997, when the Liberals prioritized fighting the deficit, and 2015, when they outflanked the NDP. Figure 4 indicates that this is also the span over which Conservatives' feelings about the Liberal Party plummeted. The leftward shift was 27 points, implying an affective drop among Conservative identifiers of about 10 degrees. This is about half the total movement captured in Figure 4. Contrary to expectation, NDP identifiers did not respond to Liberal policy shifts. Neither did Liberal identifiers, although here there is no expectation.

Quebec
Where Quebeckers remain distinct, unsurprisingly, is in relation to the Bloc. The evidence is on the bottom row of Figure 5. We have no expectation of response by Bloquistes to Liberal left-right movement, and none appears. For Bloc movement, there are no expectations for response by Liberal or Bloc identifiers, and none appears. This confirms that affective polarization between Liberals and the Bloc, revealed in Figure 3, is clearly not based on left-right ideology. Conservative identifiers, seemingly, like the Bloc the further it is to the left. We can make no sense of this. But then, as Figure 1 showed, the Bloc exhibited no net shift on the left-right spectrum; its policy movement was trendless.
Otherwise, parallels with the rest of Canada are evident, as is convergence. The clearest parallels are where we should expect them, for dynamics between the Canada-wide parties. For feelings toward the Liberal Party, Conservative identifiers' policy link is exactly the same as outside Quebec, a coefficient of 0.37. Also as in the rest of Canada, Conservatives like their own party more as it moves to the right, Liberals are unresponsive to their own party's movement, and nonpartisans are unresponsive to all parties. The one difference with the rest of Canada is that Quebec Liberals' increasing negativity to the Conservative Party does not seem tied to left-right policy shifts.

Indicator and estimation strategy
Arguments that party movement induces affective shifts-rather than the other way around-would be more credible if a communication link could be established. In particular, we need a plausible indicator of individual differences in receptivity to politically relevant messages. Unfortunately, the CES has no indicator of knowledge or interest that is comparable across the decades. 10 What I can work with is the intensity of party identification, which is consistently measured throughout and is a powerful motivator of attentiveness to the media. We should expect, then, that the more intense a person's partisanship, the more responsive he or she should be to signals about policy. But for many, the point of such attentiveness is to seek reinforcement of predispositions. Compounding this is the fact that partisan intensity can itself be related to ideological positioning. This is clearly true for New Democrats and Conservatives, as identification with each party has intensified since 2000. But for Liberal identifiers, there is scarcely any relationship between intensity and ideological self-placement, and in the aggregate, Liberal partisanship intensified only modestly over the years. Finally, Liberal identifiers as a group are all but indistinguishable ideologically from nonpartisans. 11 The estimation setup involves a different interaction from that outlined in the preceding section. As before, the slope that describes the impact on feeling from policy shifts is allowed to vary across identification groups. Here, however, the groups are not defined by direction but by intensity, specifically variations in intensity of Liberal identification with the lowest intensity defined by non-identification. Of course, only Liberals and nonpartisans enter the estimation. The only targets worth evaluating are the NDP and Conservatives-the flanking parties-as only they made (mostly) consistent and significant ideological moves over the full period. As the patterns are clearest outside Quebec, the estimation focuses on the rest of Canada. Multilevel estimation is also the setup of choice here. Figure 6 displays the evidence. For both parties, the patterns are consistent with expectation, and the scale of moderation is quite similar: amplification across the range from non-identifier to strong identifier is 0.3 or more. 12 Although there is a hint that even nonpartisans view the NDP's centrist drift positively, moving up the Liberal intensity ladder amplifies this considerably: a strong partisan is about 2.5 times as sensitive as a nonpartisan and almost twice as sensitive as a weak partisan. For the Conservative Party, nonpartisans are simply unresponsive to the policy shift, while strong Liberal identifiers are nearly three times as responsive as weak ones.

Recapitulation
In answer to question 1 above, Canadians' feelings about political parties have become more polarized. This is true in the minimal sense that dispersion has increased for all parties both in and outside Quebec. For most of the period, this was more true of the Conservatives than of the Liberals, but the latter caught up in 2019. The mere presence of Reform and the Bloc raised the affective stakes.
For question 2, sentiment has not become, on balance, more negative. One exception is for the Conservative Party, to which Quebeckers have clearly cooled. The NDP, conversely, has become generally more popular. All that said, the most dramatic shift has been the massing of feeling for the big parties at the negative extreme. By 2019 (2015 for the Conservatives), the negative end had effectively ceased to be a tail of the distribution; it had become a second mode. But even as the negative tail fattened for most parties, the main mode shifted toward more positive values. Although movement in this direction is less striking than shifts the other way, the major parties now attract more positive sentiment than before -especially when they win.
For all parties, in answer to question 3, the source of gains at the high end is each party's own supporters. Toward parties other than one's own, the trend is mainly negative. Specifically, Conservatives like the Liberal Party and the NDP less than ever, and the feeling is reciprocated. And from the moment of their appearance, Reform and the Bloc evoked sharply differentiated response. To all of this, the exceptional pair is the Liberal Party and the NDP. Supporters of each party have warmed to the other, even to the point of outright positivity.
The basic answer to question 4 is that nonpartisans have not become systematically more negative overall. The rise of negativity is a partisan phenomenon, and nonpartisans remain mostly indifferent. The patterns in Quebec are subtly different, however. Quebec nonpartisans seem to have soured on the Conservatives and the Bloc but have warmed to the Liberals and the NDP.
The answer to question 5 seems clear: affect follows policy-or at least they track each other. The rightward shift of the Conservative Party made its supporters like the party more and made Liberal and NDP supporters like it less. The rightward drift of the NDP made Liberals like that party more. Contrary to expectation, this did not have a cooling effect on the NDP's own supporters. To the extent that the Liberals shifted left, they alienated Conservative supporters.
Question 6 asks if the policy-affect link is causal, as opposed to merely correlational. Addressing this question required some contortions, but the effort points to a causal story. For respondents in the ideological middle, the greater their interest, the stronger their affective response to the repositioning of the flanking parties, the Conservatives and the NDP. Party identification cannot be treated simply as the "unmoved mover" . Certainly, partisanship affects the perceptual landscape; in that sense, it is a "mover." 13 But it is not always "unmoved." The partisan flux between 1988 and 2004 was existential and hard to square with the notion that individuals' psychological identifications unfailingly serve as a bulwark against forces in society and economy. To be sure, change in distributions of party identification is sluggish. But right from the start, the numbers claiming to identify with new parties were not small. Some of the flux in the moderating power of identification, then, must reflect not so much evolution of affect within identification groups as shifts in membership between ideologically adjacent ones. The Reform Party is the clearest example. Reform drew from the disgruntled right flank of the old Conservative Party (Johnston, 2017: Table 8.4). Much of the post-2000 drop in Conservative identifiers' affect for the Liberal Party reflects the return of Reformers to the fold, even as some centrists in the old Progressive Conservative base declined to identify with the new, unhyphenated party. In sum, part of apparent flux in partisan impact is compositional in origin.

Data and estimation
Three methodological issues must be acknowledged. First, although the CES platform has been conceptually quite stable since 1988, trends common to all survey research may account for some of the patterns to which I have given a substantive interpretation. Response rates to the RDD component, although high by comparison with commercial polls, have steadily declined. Plausibly, the disappearing respondents sit at the low end of political interest and engagement, such that increasing unit nonresponse probably accounts for some of the polarization trend (Cavari and Freedman, 2023). 14 Second, the addition of online samples in 2015 and 2019 and the elimination of the RDD component in 2021 changed the landscape independently of anything in the electorate. I decided to include the web mode so as to extend the series to 2021 and because by 2015 it was not obvious that an RDD frame carried any presumption relative to an online one. But in both 2015 and 2019, polarization was greater among online respondents than among RDD ones, such that part of the apparent change between 2011 and 2015 reflects a shift in mode. 15 Third, the multilevel analyses of policy and affect push the methods boundary, as the number of upper-tier groups is small.

Further implications
We need to rethink our basic models of the Canadian party system. Its dominant parties no longer engage in brokerage, and its dynamics are no longer driven by polarized pluralism. Policy gaps widened over succeeding decades and were accompanied by a polarization of feeling between the big parties. Neither major party seems like a big tent anymore, not even the Liberals (Carty, 2015). For polarized pluralism as a defining feature of the twentieth-century system, the key was the system's domination by a party of the centre, the Liberals. From this domination flowed both the multiplicity of parties and a peculiarly Canadian pattern of cyclical volatility. As the policy and affective gaps between the Liberals and the NDP shrank, the formerly centrist vehicle now seems more like the major partner in a centre-left cartel. And the Liberal Party arguably ceased to be dominant in 1984.
And yet the electorate may be even less than before a "god of vengeance and reward" (Key, 1964: 568), able to deliver retrospective judgments on the performance of incumbents. The old system did reward or punish the incumbent for the performance of the economy (Bélanger and Godbout, 2010). Sometimes the punishment was dire, as in 1935, 1958, 1962, 1984 and 1993. Now it takes the form of exchanges within sides. The system can be conceived in continental European terms as comprising party families or blocks. Volatility should be separated into within-and between-block components (Bartolini and Mair, 1990). Victory by the major player in one block is increasingly unacceptable to voters in the other block. Crossing the boundary between blocks no longer seems available as a performance judgment. By implication, judgments on a governing party may increasingly be confined within the block. As a rough gauge of this shift, consider the Conservatives and any party to its right as the "right" block and the Liberals and all others as the "left." That this division is arbitrary for earlier years is precisely the point: those years have passed. A clear implication of this article is that the division is no longer so arbitrary, and blocks should have become more stable. Facts correspond to expectations: the flux-indicated by the standard deviation-in block shares since 1990 is only half that for the years between 1945 and 1990. Thinking in terms of blocks highlights asymmetry between the sides. Cochrane (2010Cochrane ( , 2015 has already identified a qualitative ideational asymmetry: where the left is rather seamless, the right comprises indigestible chunks. This article captures a corresponding asymmetry in partisan feeling and behaviour. On the left, the parties exhibit high existential continuity, even as they adjust their policy positioning and converge on the field of sentiment. They appeal to each other's supporters but hesitate to take each other hostage. On the right, the relations are existential-and unstable. Voters on the right flank occasionally threaten to bring the whole house down. Exactly this happened in 1993, and the threat continues to be credible. Is there a link between Cochrane's findings and these findings? Finally, the article has implications for multicountry causal inference, as in Gidron et al. (2020). Keeping the parties separate, rather than subsuming them in an overarching index, delivers more than aesthetic satisfaction. It also gives leverage on causal linkage between policy and affect. It is striking, of course, that the Conservatives pulled to the right and that affective differences followed. But no less important is that the NDP moved to the centre and converged on the Liberal Party. Liberal and NDP identifiers responded with affective depolarization. This is not just nuance; it is interpretively critical. This convergence in policy and affect underscores the role of policy in driving the divergence on the system's other flank. It suggests that the focus for future analyses should not be on trends in culture or in mass communication and social media. Rather it should be on party organization: what drove the Conservatives to the right and the NDP to the centre? Is there similar heterogeneity in other countries? effects (Kish, 1965). A minor exception to the FTF rule was the 1980 study, carried out with emergency funding and by telephone, and comprising respondents who had been interviewed in 1979. It was like the other pre-1988 studies in being post-election with a clustered sample. 5 Exact wordings for thermometer items can be found in appendix Table A1. 6 Polarization over parties is not the by-product of some more general shift in item functioning. Evidence can be found in section 5 of the appendix, where I compare distributions of affect toward various places and social groups. For these "placebos" there is simply no increase in the dispersion of sentiment. 7 Certain omissions require comment. For visual clarity, and as it is not relevant to this article, the Social Credit Party, last coded for 1974, does not appear. No Canadian Manifesto data have been catalogued since 2015. 8 Elff et al. (2021) argue that Stegmueller's findings of bias reflect a fixed seed for his Monte Carlo simulations and too few simulations overall: 2,000 rather than 10,000. For my part, I tested Bayesian estimation on a handful of examples on the lines Elff et al. suggest. In each case, the correspondences with restricted maximum likelihood (REML) are close, including between REML confidence intervals and Bayesian credible distributions. A comparison of REML and Bayesian estimations for the most consequential relationship, the left-right position of the Conservatives, can be found in Table A13 in the online appendix. An additional possibility is just to estimate a nonhierarchical model but with a Huber-White correction for clustering of errors by year. This is more conservative than any variant of hierarchical modelling, as evidenced by the comparison (again for Conservative positioning) in appendix Table A15. These estimates do not alter the general picture, but they widen confidence intervals, such that more span the zero boundary than in the multilevel estimations. 9 NDP and Reform identifiers are included in the underlying estimations for feelings toward the other parties, however, as their presence adds precision. See appendix Tables A10 and A11. 10 For more detail on CES knowledge and interest indicators, see appendix section 3.1. 11 Evidence on these propositions can be found in appendix section 3.2. Analyses are based on the leftright battery of the Comparative Study of Electoral Systems (CSES), which the CES has included in its postelection mailback survey since 1997. Because mailback samples are relatively small, are not representative cross-sections and do not extend back to 1988, ideological self-placements cannot be incorporated directly into this article's estimations. 12 Although Figure 6's confidence intervals for each level of partisan intensity overlap those for each other level in each estimation, the underlying interactions are clearly real. On each side, the coefficient on the interaction term is about five times greater than its standard error. See appendix Table A12. 13 As, for example, in Merkley (2021b). 14 Discussion and evidence on the CES response rate and its implications can be found in appendix section 4. 15 Detailed comparison between the RDD and web modes can be found in appendix section 1.