2.1 Municipal ideological representation

Local public opinion estimates have been especially valuable for scholars of urban politics and representation, helping to resolve a long-standing debate about the presence and role of ideological disagreement in local elections. For years, many urban political scientists held that local elections are distinctively non-ideological, especially outside big cities, animated instead by unpredictable local controversies and assessments of the managerial competence of incumbents (Oliver, Reference Oliver2012). This absence of ideological disagreement in municipal elections originated, it was argued, in some combination of local socio-demographic homogeneity, the technical and non-redistributive policy domains in which municipalities legislate, and the disciplining forces of inter-municipal competition and mobile capital (Peterson, Reference Peterson1981).

While most scholars continue to acknowledge that municipal jurisdictional limits and inter-municipal competition do constrain municipal attempts at redistributive policy making, a large body of evidence suggests not only that ideology plays an important role in municipal vote choice (Holman and Lay, Reference Holman and Celeste Lay2021; Sances, Reference Sances2018) but also that there is a clear relationship between municipal residents’ ideological positions and the policies that their municipalities enact (Tausanovitch and Warshaw, Reference Tausanovitch and Warshaw2014; Einstein and Kogan, Reference Einstein and Kogan2016). Moreover, this evidence consistently shows that ideological voting and policy responsiveness are equally apparent across municipalities of all population sizes (Sances, Reference Sances2018; Schaffner et al., Reference Schaffner, Rhodes and La Raja2020). Canadian findings have closely resembled those in the United States, with clear evidence of ideological voting in Canadian city elections (Cutler and Matthews, Reference Cutler and Scott Matthews2005; McGregor, Reference McGregor2017; Lucas and McGregor, Reference Lucas, Michael McGregor, Lucas and Michael McGregor2020) and aggregate evidence of partisan and ideological representation in municipalities of all population sizes and institutional types (Lucas, Reference Lucas2020).

Having found little evidence to support the persistent argument that municipal politics is uniquely non-ideological, urban political scientists have now turned from the question of if to the question of how—how, and in what ways, ideological disagreement shapes local elections and representation (Anzia, Reference Anzia2021; Warshaw, Reference Warshaw2019). One important branch of this new research has focused on responsiveness: the relationship between the ideological complexion of local municipal publics and the policies that their governments produce (Einstein and Kogan, Reference Einstein and Kogan2016; Sances, Reference Sances2021; Tausanovitch and Warshaw, Reference Tausanovitch and Warshaw2014). A second, more recent branch of research has explored what Schaffner et al. (Reference Schaffner, Rhodes and La Raja2020) call “ideological congruence representation”—the relationship between constituents’ attitudes and those of the politicians who represent them. This research, which grows out of a larger body of research on the ways politicians resemble (or fail to resemble) their constituents (Sheffer et al., Reference Sheffer, Loewen, Soroka, Walgrave and Sheafer2018) and understand (or misunderstand) local public opinion (Broockman and Skovron, Reference Broockman and Skovron2018), has recently been undertaken in the United States using voter file data (Schaffner et al., Reference Schaffner, Rhodes and La Raja2020) and in Canada using surveys of municipal political elites (Lucas, Reference Lucas2020).

Our analysis in this article extends this recent research by providing Canada's first direct comparison of mass and elite ideology in municipal government (in fact, it is, to our knowledge, the first direct comparison of mass and elite ideology at any level of government in Canada). We combine our MRP estimates of Canadian municipal ideology with novel survey data on the ideological positions of more than 900 Canadian municipal politicians to understand the size and character of the relationship between politicians’ ideological positions and those of their constituents. Our findings support a growing consensus on ideological representation in the urban political science literature, suggest new paths for further research on local representation and provide a practical model for the study of ideological representation at other levels of government in Canada.

2.2 A note on terminology

Before we turn to our data and methods, it is worth pausing to clarify a few key terms. Traditionally, most political scientists outside the Marxian tradition have understood ideology to refer to an interconnected system of beliefs and preferences about the scope, actions and priorities of government (Converse, Reference Converse2006). More recently, however, scholars have emphasized that an individual's ideological self-understanding—conservative, libertarian, progressive, and so on—is also a form of social identity, one that is related to, but not necessarily determined by, their policy preferences (Kinder and Kalmoe, Reference Kinder and Kalmoe2017; Mason, Reference Mason2018; Stimson, Reference Stimson2004). As a result, most political scientists now distinguish between symbolic ideology, which captures the social identity component of ideological self-placement, and operational or policy ideology, which captures the structure of an individual's actual policy attitudes (Jessee, Reference Jessee2012). This latter concept is the quantity we seek to estimate in what follows, and in keeping with other recent work (Caughey et al., Reference Caughey, O'Grady and Warshaw2019), we refer to it as policy ideology. This term usefully distinguishes policy ideology from symbolic ideology and also serves to remind readers that our measure is fundamentally about the structure and character of Canadians’ policy preferences.

3.1 Measuring policy ideology

As we discussed above, we have two main desiderata for a measure of individual policy ideology. First, following Cochrane (Reference Cochrane2015), we would like a measure that avoids defining specific policy positions as the core elements of an individual's ideological position. Ideally, an ideology measure will capture the family resemblance character of ideological agreement, such that two individuals may disagree on any particular issue but still hold a larger set of policy attitudes that is recognizable to their allies and opponents as having a distinctly left or right character. To this end, our model relies on publicly available data from the 2019 CES (Stephenson et al., Reference Stephenson, Harell, Rubenson and Loewen2020), a multiwave survey of nearly 38,000 Canadians undertaken during and after the 2019 Canadian federal election. We began by identifying every policy-relevant question in the survey—defined as a question about an individual's attitude on an issue for which a government could or should take action—producing an overall list of 56 survey items. Because of the modular structure of the CES survey, the number of responses to these questions ranges from less than 1,000 to the full sample; however, our measurement model allows us to incorporate information about individual ideological positions even from questions that were answered by a small subset of the overall sample. We provide additional detail on the survey items in our measurement model, along with descriptive statistics, in the supplementary material (SM1).

Second, we would like a measure that helps us both reduce measurement error and accept the inevitable uncertainty involved in complex political-psychological phenomena—reducing measurement error by aggregating data across many survey questions (Ansolabehere et al., Reference Ansolabehere, Rodden and Snyder2008) while also accepting measurement uncertainty by including some estimate of precision alongside the measured ideology values. To construct this measure, we fit the following Bayesian latent measurement model:

$$y_{ik} = \beta _k\xi _i + e_{ik}$$

where y _i,k are the values for observation i on each of the k survey questions. The β _k are the coefficients relating the latent variable to the observed indicators. ξ _i is each individual's policy ideology. The indicators (detailed in the supplementary materials) are mostly (44/56) five-point Likert-type items. Of the remaining 12 items, 9 are three-point ordinal measures and 3 are binary. While some of these variables would not ordinarily be amenable to linear modelling, Kamata and Bauer (Reference Kamata and Bauer2008) have demonstrated an equivalence between traditional factor analysis and the item response theory model; moreover, our goal is to recover the latent variable rather than make inferences about the nature of the underlying relationship. For these reasons, we are comfortable using a Bayesian analog to the common factor model to estimate the latent variable of interest on standardized versions of y _i,k.

We note that our measurement model takes a unidimensional approach to policy ideology. We do so for two reasons. First, while the structure of citizens’ policy attitudes is complex, past research in the United States has demonstrated that unidimensional measures can meaningfully summarize opinions even for large batteries of policy questions (Tausanovitch and Warshaw, Reference Tausanovitch and Warshaw2014; Fowler et al., Reference Fowler, Hill, Lewis, Tausanovitch, Vavreck and Warshaw2021). Recent studies in Canada have also demonstrated that elite discourse is clearly structured by a single left-right dimension with the sole exception of questions regarding Quebec sovereignty (Rheault and Cochrane, Reference Rheault and Cochrane2020).Footnote ⁴ Second, a unidimensional approach is empirically supported by CES responses themselves, which are related to one another across economic, social and immigration policy attitudes (an inter-item correlation plot and further analysis is available in supplementary material SM5); our preliminary analysis suggested that higher-dimensional solutions, including those that pre-organize survey questions into “economic” and “social” items (along the lines of Caughey and Warshaw [Reference Caughey and Warshaw2018] and Caughey et al. [Reference Caughey, O'Grady and Warshaw2019]), added complexity without meaningfully changing our findings. Thus, without denying that we have much to learn from higher-dimensional analyses, our unidimensional measure provides a useful and valid summary of the structure of Canadians’ issue attitudes.

We implement our model using automatic differentiation variational inference (ADVI) (Blei et al., Reference Blei, Kucukelbir and McAuliffe2017) to find the posterior distribution of the model parameters. There are some important caveats with VI, in general, and with ADVI, in particular—namely, that strong independence assumptions are made. This makes measurement models (particularly common factor models) good candidates, as they are models that assume conditional independence. Standardizing variables removes the model intercept, leaving, at least in one dimension, a single coefficient for each indicator. These coefficients are unlikely to be highly correlated across indicator variables. We note that variational inference offers exciting possibilities for measurement of very large datasets and multidimensional latent measurement problems that would be very computationally intensive using Markov chain Monte Carlo (MCMC) sampling (Imai et al., Reference Imai, Lo and Olmstead2016).

We assume the latent variable to have a standard normal prior and impose zero mean and unit variance in the posterior by standardizing each draw from the latent variable. This identifies the centre and scale of the latent variable distribution. This identifies the model up to a reflection, but since ADVI finds the most likely posterior mode, it will identify one side of the reflection and allow us to sample from that posterior. This could produce a result such that the coefficients and latent variable all need to be multiplied by −1 to make it substantively meaningful; for example, we could have individuals with recognizably left policy attitudes on the right side of the scale, and vice versa. In this circumstance, multiplying everything by −1 will put the observations on their appropriate side of zero. We also assume the coefficients relating the latent variable to the observed variable to have standard normal priors. The residual variances for each of the 56 indicators have half-Cauchy priors with β = 1. We ran the ADVI optimizer for 30,000 iterations, at which point the trend in the loss function evened out; we then sampled 1,500 observations from the joint posterior of all of the model parameters—latent variables, coefficients and residual variances. The results below are a function of these values.

3.2 Estimating local ideology

We use MRP to translate our individual ideological measures, described above, into aggregate local estimates. MRP can itself be divided into two substeps. First, we model individual ideology as a function of an individual's demographic characteristics (age, sex and education level), their local residence (in a municipality or a federal electoral district) and their region. We fit this model in a Bayesian multilevel setup, with varying intercepts for demographics, local community (municipality or electoral district) and region.Footnote ⁵ This approach allows us to pool information about demographic predictors of an individual's ideology from the full dataset while also drawing information from respondents’ local contexts when sufficiently informative data are available. We thus model the ideological position of individual i as follows:

$$Y_i = \theta _0 + \alpha _{\,j[ i ] }^{age.sex.edu} + \alpha _{k[ i ] }^{local} + \alpha _{l[ i ] }^{region} $$

In this model, we assume that age-gender-education intercepts are drawn from a normal distribution with mean zero:

$$\alpha _j^{age.sex.edu} \sim {\cal N}( {0, \;\sigma_{age.sex.edu}^2 } ) $$

We model intercepts for local communities—municipalities or federal electoral districts—as predicted by regional intercepts as well as a set of aggregate predictors for each municipality or district: population density, 2019 Conservative party vote share,Footnote ⁶ median income, proportion with university education, and proportion visible minority:

$$\eqalign{& \alpha _k^{local} \sim {\rm {\cal N}}\left( {\mu _k^{local}, \sigma _k^2 } \right) \cr & \mu _k^{local} = \alpha _{l\left[ i \right]}^{region} + \gamma _1{\rm densit}{\rm y}_k + \gamma _2{\rm vs}{\rm .co}{\rm n}_k + \gamma _3{\rm incom}{\rm e}_k \cr & \quad \quad \quad + \gamma _4{\rm edu}{\rm c}_k + \gamma _5{\rm rac}{\rm e}_k}$$

Finally, we assume that region intercepts (British Columbia, Prairies, Ontario, Quebec and Atlantic Canada) are drawn from a normal distribution with mean zero:

$$\alpha _l^{region} \sim {\cal N}( {0, \;\sigma_{region}^2 } ) $$

Using the results of this model, we predict ideology values for every possible demographic, municipal/district and regional combination, and we weight the resulting predictions on the basis of each subgroup's proportion of the overall municipal/district population. We use data from the Canadian census to construct these poststratification proportions. The result is an estimate of the average ideological position of each electoral district and municipality.Footnote ⁷

We then use the model parameters to estimate municipal ideology. For each municipality (which is itself nested in a region), there will be 12 different age-sex-education intercepts—one for each combination of age group (1 = 18–34; 2 = 35–54; 3 = 55 +), sex (1 = male; 2 = female) and education (1 = no university degree; 2 = holds university degree). From the Canadian census, we use the proportion of observations within municipality j that falls in each of the m = {1, …, 12} combinations mentioned above; call this p _j,m. For each posterior draw, we calculate ${\rm ideology}_j^{( t ) } = \mathop \sum \limits_{m = 1}^{12} \alpha _m^{age.sex.edu} p_{j, m} + \alpha _{l[ j ] }^{region} + \alpha _j^{local}$ where j is an index for municipality and l[j] indicates the region into which municipality j falls. The superscript (t) indicates that this is a single draw from the posterior distribution of the ideology measure. We then summarize these posterior draws, calculating the posterior mean, standard deviation and 95 per cent credible intervals.

3.3 Municipal ideological representation

Our analysis of ideological representation in Canadian municipalities not only requires that we have an estimate of each municipality's ideological position—which we construct using the procedure described above—but also requires information on the politicians who are elected in those municipalities. We draw the latter information from the Canadian Municipal Barometer (CMB), an annual survey of mayors and councillors in every Canadian municipality above 9,000 population. We use two quantities from this dataset. First, we compare municipal ideology to politicians’ symbolic ideology using ideological self-placement values from politicians who responded either to the CMB annual survey or two subsequent surveys of municipal politicians related to COVID-19, all of which contained a standard left-right self-placement question.Footnote ⁸ In total, we have 908 responses to this question across more than 400 municipalities.

We also compare municipal ideological values to politicians’ policy ideology using a suite of policy issue questions from the 2020 CMB annual survey. We measure each politician's ideological position using a Bayesian latent variable model identical in character to the CES model described above. In the supplementary material (SM5), we provide additional detail on this model, including question wording and validation tests. To our knowledge, the results of this model represent the first latent measure of the policy ideology of Canadian politicians ever constructed in Canada and allow us to more directly compare citizens’ ideological positions to local politicians’ attitudes on important municipal policy issues.

4.1 Validation tests

To provide some sense of the face validity of our measure of policy ideology, we can also check that the measure is related in expected ways to other indicators. For instance, while the relationship between symbolic and operational ideology is complex, we would expect to see that policy ideology is strongly related to ideological self-placement: in the aggregate, those who place themselves on the ideological left should also be more likely to hold recognizably left-leaning policy attitudes, and vice versa. Similarly, while individual partisanship in Canada is shaped by a variety of factors, we also have some general expectations about the relationship between partisanship and policy ideology: we would expect Conservative partisans to tend rightward, and we would expect Liberal and New Democratic party (NDP) partisans to tend leftward.

Figure 2 visualizes these tests. In the left plot, we use a non-parametric smoothed line to summarize the relationship between an individual's policy ideology and their ideological self- placement; the relationship is strong, clear and in the expected direction. In the right plot, we summarize the distribution of individual policy ideology for those who identify with seven political parties, as well as non-partisans.Footnote ⁹ Here, too, the results confirm our expectations: Liberal and NDP partisans tend leftward, and Conservative partisans tend rightward—as do partisans of the People's party of Canada, just as we would expect. The plot also suggests interesting possibilities for further analysis; consider, for instance, the internal challenges faced by the NDP, whose partisans are divided among those on the extreme left, those on the moderate left, and a not-insignificant number of individuals on the moderate right. The Green party, with roots in conservationist conservatism but a recent history of firmly left policy positions, appears to face similar internal heterogeneity.

Figure 2. Policy Ideology Measure, Self-Placement (left) and Partisanship (right)

4.2 Geographic variation

Turning from overall distributions to our local ideological estimates, Figure 3 provides an overview of the average policy ideology of all non-northern federal electoral districts, with right-leaning districts in blue and left-leaning districts in red. Small urban ridings are difficult to see in the overall map (though readers who are viewing the map on a computer can zoom in to see specific regions more clearly), so we also include smaller maps of Canada's four largest metropolitan regions. The maps provide a striking summary of the overall variation in policy ideology across Canada: moderate ideological positions through much of the Maritimes and Quebec, slightly more conservatism in Ontario and British Columbia, and a marked band of deep blue stretching from southern Manitoba through to Alberta's northern boundary. Below the large map, the metropolitan maps reveal what the larger map conceals: nearly all of Canada's most strongly left-leaning districts are nestled in the central core of the country's largest urban regions. These geographic patterns align well with general expectations about Canadian political geography; in the supplementary material (SM4), we show that our MRP measure of policy ideology is strongly correlated with Conservative party vote share in both federal electoral districts and municipalities.

Figure 3. Geographic Variation in Policy Ideology, by Electoral District

A second way to summarize geographic variation in district and municipal ideology is to simply plot each local community's estimated ideological position. In Figure 4, we do so for the 20 most left-leaning and right-leaning municipalities and districts. Among the municipalities in our dataset, Alberta clearly dominates the rightward end of the distribution, with places such as Leduc, Grande Prairie and Red Deer among the most conservative. At the other end of the spectrum are big cities like Montreal and Vancouver, along with medium-sized Maritime cities such as Saint John and Fredericton and the left-leaning municipalities of southern Vancouver Island. The same general patterns are visible among federal electoral districts, with rural Alberta at the rightmost end of the spectrum, and urban districts in Montreal, Vancouver and Toronto dominating the leftmost end of the spectrum.

Figure 4. Highest and Lowest Ideology Values, Municipalities and Electoral Districts