Data and Measures: The Mapping Local Communities Canada Study

We use online survey responses from 7,811 English-speaking Canadians who answered the Mapping Local Communities Canada Study (MLCC) in April–July 2012. A non-partisan electoral education initiative sponsored by the Canadian Broadcasting Corporation called Vote Compass provided the sampling frame (see Appendix for more details). The respondents in our convenience sample were better educated, wealthier, more likely to be informed about politics, and more comfortable using technology than the average Canadian. Our aim is to provide a focused assessment of the relationship between perceptions of place and attitudes about social cohesion, and our results are likely conservative given that misperceptions would be greater for the Canadian population as a whole; better educated, wealthier, and more politically informed survey respondents are likely to differ less from one another in regard to their perceptions than survey respondents representing all of English-speaking Canadian society (Delli Carpini and Keeter Reference Delli Carpini and Keeter1996; Nadeau et al. Reference Nadeau, Niemi and Levine1993).

Canada suits our study because of its immigrant history, rapidity of demographic change, and ethnic heterogeneity across the nation. ‘Visible minorities’, the Canadian Census term for non-white and non-Aboriginal Canadians, grew from less than 1 per cent of the population in 1971 to about 16 per cent in the 2006 Census. The effect of this diversity is noticeable in its biggest cities. For example, by the 2016 Census, visible minorities made up more than one in five Canadians and both Toronto and Vancouver were majority–minority. Furthermore, Canada has an official policy of multiculturalism, and diversity is a salient political issue (Bouchard Reference Bouchard2008). Because we care about theories of threat posed by minorities and social cohesion in the context of ethnic diversity, our main analyses in this article focus on majority group members (non-visible minorities and non-Aboriginals).

What Do the Census and People See?

Measures of objective context

For our objective-context measures, we use data from the 2006 Canadian Census for levels of visible minority population and used both the 2006 and the 2016 Canadian Census to create a measure of change in visible minority population.Footnote ⁵ We created an index of the percentage of visible minorities for Census dissemination areas (DAs); they are the smallest Census unit for which all information is disseminated.Footnote ⁶ We followed the Statistics Canada definition of ‘visible minority’ (which includes ‘persons who are non-Caucasian in race or non-white in colour and who do not report being Aboriginal’).Footnote ⁷ In 2006, 50 per cent of Canada’s roughly 55,000 DAs contained less than 6 per cent visible minorities.Footnote ⁸

Measures of subjective context

To create a measure of personally relevant places that operationalizes context as pseudoenvironments, respondents see an online map centered on the postal code where they live and then draw their ‘local community’. (For details about the map-drawing measure – including its validity and reliability – see Wong et al. (Reference Wong, Bowers, Rubenson, Fredrickson and Rundlett2018).) Figure 1 shows an example of fifty such maps drawn by people living in the Greater Toronto Area (GTA) overlaid on each other and on a Google Map of the GTA. Each map was unique.

Figure 1. A random sample of fifty ‘local community’ maps drawn by residents of Toronto in the MLCC.

After respondents drew their ‘local community’ on the map, they answered a battery of questions about their perceptions of the relative size of ethnic/racial groups captured in their drawing: ‘Just your best guess – what percentage of the population in your local community is …’ The list of groups included the following: black, Canadian Aboriginal, white, Chinese, Latin American, South Asian (East Indian, Pakistani, Sri Lankan, etc.), and Other Asian (Korean, Japanese, Filipino, etc.). The percentage perceived visible minority in a context was an index adding together responses following the government’s definition of ‘visible minority’.Footnote ⁹

We also showed respondents, at random with equal probability, a map of one of six administrative geographic areas highlighted, centered on their address or postal code: the area ranged from the respondent’s smallest Census unit (DA) to Canada as a whole. Respondents were told the name of the administrative unit the map represented, and were shown its boundaries overlaid on a Google Map. They then were given the same battery of questions about the demographic make-up of this fixed geography. We use both perceptions of their own hand-drawn map and perceptions of fixed Census neighborhoods (DAs) as our measures of perceptions in this paper. Since more than one respondent might have been shown the same fixed Census neighborhood, this will allow us (in Appendices C.2, C.3, and C.4) to show that the effects of differences in perceptions are not driven by differences in maps drawn.

The respondents overestimate the percentages of visible minorities in all their contexts: the median response was 28 per cent visible minority for those viewing a map of their DA, while the median Census measurement for these ‘neighborhoods’ was 7 per cent; the median perception of proportion of visible minorities in their hand-drawn ‘local community’ was also 29 per cent, while our median estimate of the proportion of visible minorities in the boundaries of their maps using Census data was 10 per cent (weighting all DAs touching the hand-drawn map equally in their contribution to objective context within the subjective boundaries) or 8 per cent (weighting each DA in proportion to the land area it accounted for within the hand-drawn map).Footnote ¹⁰ These overestimates are consistent with previous research (see Guay et al. (Reference Guay, Marghetis, Wong and Landy2025) for a recent example).

Design: Isolating Perceptions from Objective Context

Our research design uses 3,772 respondents living in 3,098 DAs and creates pairs of people with different perceptions, living in very similar (or even identical) DAs in terms of proportion of visible minorities, the amount of change in the proportion of visible minorities between 2016 and 2006 as measured by the Census, and within the 2006 Canadian Census categories for urban versus rural.Footnote ¹¹ Because DAs do not vary much in total population – most contain between 400 and 700 people by design – our matched respondents live in environments with very similar total populations.Footnote ¹²

This paired research design removed differences between people due to objective context without trying to hold constant any other covariates. Without pairing, we would have compared white people living in areas with no visible minorities to white people living in places with nearly 100 per cent visible minorities. In the design we present here, 99 per cent of the pairs differed by less than 0.036 percentage points in proportion of visible minority, the maximum difference in percent of visible minority within a pair was less than 0.04 percentage points, and 70 per cent of the matches were identical in regard to objective context. Also, no two people within a pair differed by more than 3 percentage points in change in visible minority population between 2006 and 2016, with 50 per cent of pairs differing by less than 0.8 percentage points of change.

Within pairs, holding nearly constant the objective characteristics of where people live, do people still differ in their perceptions of their surroundings? Our primary research design requires that each pair differ in perceptions – after all, a difference in social-cohesion attitudes between two people who differ in perceptions is undefined if those two people do not in fact differ in perceptions, and those pairs would drop from the analysis. To answer the preliminary question of whether two people living in the same objective context are likely to differ in their perceptions, we recreated the matched design described above without any requirements on the amount of difference in perceptions. If objective context is a good proxy for perceptions, we should expect nearly no variation within pairs in perceptions.Footnote ¹³ The pairs in this unrestricted perceptions design (where people were matched on the same objective-context numbers as the main design but with no mention of perceptions) were successfully matched on objective context: 80 per cent of the pairs were identical in proportion of visible minority, with the largest difference being less than 1 percentage point, and 90 per cent of the pairs differed by less than 2 percentage points in change in visible minority between 2016 and 2006. Yet, the majority of pairs differed in perceptions: only 45 out of 1,638 pairs (roughly 3 per cent) had two people with the same perceptions, the mean difference in perceptions is 21 percentage points, and 50 per cent differed by between 7 and 30 percentage points.

Returning to our main design, which requires non-zero differences in perceptions in the same way that an experiment would require both treated and control units within a pair, we can report that those pairs differ in perceptions on average by 25 percentage points, with a minimum difference of 0.01 percentage points by construction, and 50 per cent of the differences fall between 7 and 33 percentage points. That is, differences in perceptions between people living in the same kinds of context are the norm in this study.

This design and the versions described below hold objective context (nearly) constant so that we may assess the differences in social cohesion attributable to perceptions isolated from objective context. The description of our design also reveals that perceptions differ within a pair even when the objective context is the same (whether by construction or without restriction on perception differences within a pair). The Appendix contains the results of four other designs, including a design that compares people who live in exactly the same DA who are asked to evaluate the visible minority proportion of that DA: even those people differ in their perceptions of diversity of their DA. Notice that this approach to research design allows us to describe the relationship between perceptions and outcomes while partialling out the influence of objective context in a non-parametric manner. This relationship is exactly as would be implied by the two-path theory, and we show it below. We are not estimating the causal effect of perceptions of diversity on perceptions of social cohesion under a selection on observable assumptions (after all, we are only matching on one or two covariates); rather, we are describing how perceptions of diversity relate to perceptions of social cohesion after holding constant objective context. If objective context and perceptions stand in for one another, the two-path theory is not useful. If perceptions vary within pairs, and those within-pair differences in perception are systematically related to within-pair differences in social cohesion, we will have discovered that context matters via the path of perceptions. Notice also that we are not exploring the interaction of perceptions and objective context in this paper: many other papers have established that objective context has a relationship with outcomes such as social cohesion. By adding perceptions, we help explain why those relationships have varied – some of them arise from the operation of direct experience with the context and others, our evidence suggests below, should arise from the operation of perceptions.

Analysis: How Do Pseudoenvironments Influence Social Cohesion?

Our empirical strategy in this paper comes in two general stages. In the first stage (which we called ‘Design’ in the previous section), we created pairs of survey respondents who live in objectively similar places in regard to ethnic homogeneity in their home Census DAs. The second stage (which we call ‘Analysis’) describes how people who perceive more ethnic diversity than their paired respondent also differ with regard to their responses to survey questions about social cohesion and collective efficacy.

If we could have implanted two different perceptions in the mind of a person and compared their responses to the outcome questions, we would not have needed to create pairs in order to isolate perceptions from objective context: we would have had a single person living in an identical place with two different perceptions of that place. Alternatively, if we had been able to randomly assign perceptions to the minds of many people, we could also have isolated the two mechanisms of the bundle of drivers that we are calling ‘perceptions’ and the bundle of drivers that we call objective context as they influence the outcomes. These two idealized designs suffer from problems of feasibility and, more importantly, connection to the theory. The effect of being told to perceive one’s environment in a particular way is different from the effect of the natural perceptions that are generated from experience upon facing a survey question. That said, the paired designs that we created should allow us to isolate the way perceptions of place matter for the outcome in a way that is separate from the way objective context matters: if the person who perceives more diversity in their environment within the pair tends to be the person who also reports less social cohesion, then we can claim that this difference is not due to differences within pairs arising from differences in the objective context. Our analysis of the data makes exactly such calculations, and Appendix C shows a total of five different paired designs demonstrating that the substantive results hold regardless of the details of how we create the pairs.

How do we do this analysis? The simplest analysis would subtract outcome values within pairs and regress those differences on within-pair differences in perception.Footnote ¹⁴ The analyses that generate the results that we present here build on and extend this pair-differences approach because (1) common calculations of standard errors using this approach will tend to produce overly small standard errors that do not account for the fact that people are clustered in dissemination areas such that two or more people might share a DA even if they are spread across different pairs,Footnote ¹⁵ and (2) we would like to reduce the influence of any small within-pair objective-context differences left over from the design stage.Footnote ¹⁶ Multilevel models allow us to describe the relationship between within-pair differences while calculating standard errors that account for cross-pair clustering by DA and removing any remaining average linear relationship between pair-members in objective context. We use that modeling strategy for the results presented here. The disadvantage of the multilevel model is that it requires an assumption about the probability process generating the outcomes (a likelihood function), which would not be necessary were we to pursue the simpler but perhaps overly optimistic approach of fixed effects. To address these trade-offs, Appendix D shows that the simple, fixed effects approach and two multilevel modeling based approaches (one fully Bayesian and the likelihood based approach presented here) all give substantively the same answers, and nearly the same answers numerically.Footnote ¹⁷

Equations (1) and (2) show the multilevel model describing how mean social cohesion (μ) for respondents i = 1, …, n in pairs matched on objective context s = 1, …, S and dissemination areas d = 1, …, D varies as a function of perceptions and remaining objective-context differences within pairs. The outcome and two random effects are modeled as arising from three different normal distributions in the creation of the likelihood function that we use for estimation and statistical inference.Footnote ¹⁸ If we had only written μ _is = α _s + β ₁perceptions _is we would have the equivalent of a fixed effects or within-pair model where β ₁ is the relationship between within-pair perceptions differences and the within-pair differences in social cohesion.Footnote ¹⁹ The random effect term α _d models the dissemination area clustering across pairs, and the term β ₂objective _id adjusts the calculation of β ₁ for any remaining linear relationship between perceptions and objective contexts within pairs: although most pairs are nearly identical in objective context, we would like to ensure that if there is any overall relationship such that the higher perceiver tends to live in the more diverse place than the lower perceiver within the pair, we remove that relationship.

(1) $$\eqalign{ & {y_{isd}} \sim N({{\rm\mu} _{isd}},{{\rm\sigma} _{isd}}) \cr & {\alpha _s}\sim N({{\rm\gamma} _{s,0}},{{\rm\sigma} _s}) \cr & {\alpha _d}\sim N({{\rm\gamma} _{d,0}},{{\rm\sigma} _d}) \cr} $$

(2) $${{\rm\mu} _{isd}} = \underbrace {{\alpha _d}}_{{\rm{DA}}\;{\rm{random}}\;{\rm{effect}}} + \underbrace {{\alpha _s}}_{{\rm{Pair}}\;{\rm{random}}\;{\rm{effect}}} + \underbrace {{\beta _1}{\rm{perception}}{{\rm{s}}_{is}}}_{{\rm{Effect}}\;{\rm{of}}\;{\rm{perception}}\;{\rm{differences}}} + \underbrace {{\beta _2}{\rm{objectiv}}{{\rm{e}}_{id}}}_{{\rm{Remaining}}\;{\rm{objective}}\;{\rm{DA}}\;{\rm{diffs}}}$$

Since we have two random effects that are not nested, we present profile confidence intervals below rather than Wald intervals. The Wald interval uses the second-derivative of the log-likelihood function evaluated at its maximum as an estimator of a standard error and uses a normal approximation and assumption of symmetry of the likelihood function around the maximum to create confidence intervals. However, when the likelihood surface is complicated (such as when we have two random effects), the Wald approach can yield inaccurate estimates of standard errors. The profiling approach calculates an interval of possible parameter estimates around the maximum of the likelihood function for a given parameter instead of calculating a standard error and then relying on assumptions about symmetry and normality to calculate an interval. Pinheiro and Bates (Reference Pinheiro and Bates2000) and Raudenbush and Bryk (Reference Raudenbush and Bryk2002) suggest the use of profile-based intervals for complex multilevel models, and we follow their advice below.

Results

Figure 2 shows the relationship between perceiving more minorities and responses to the social cohesion and collective efficacy scales. For two respondents who live in almost identical contexts, the one who perceives more minorities in her local community is more likely to think people who live in that community do not share the same values, do not get along, and would not help each other.

Figure 2. The influence of perceptions of visible minorities on social cohesion and community efficacy holding constant objective context. Points show the average difference in perceptions of visible minorities in the hand-drawn ‘local community’ on outcomes (listed on the y-axis) conditional on objective-context matched pairs (3,772 pairs, 3,098 DAs) and using the model shown in eq. (2). The segments show 95 per cent profile-likelihood confidence intervals (D Bates et al. Reference Bates, Mächler, Bolker and Walker2015).

The coefficient from the model for social cohesion is −0.065 – representing the average difference in social-cohesion score between a person perceiving no visible minorities in their local community (0 per cent) and a person who perceives that 100 per cent of their local community is a visible minority. We present such an unrealistic coefficient both to connect it with the model itself and to the multiple auxiliary analyses that we present later in the paper. For now, however, let us put this estimate in perspective. Few pairs differ by 100 percentage points in perceptions: 50 per cent differ by 16 percentage points or less and 90 per cent of pairs differ by less than 54 percentage points. Given a typical difference in perceptions within pairs, for example, of 20 percentage points, our results predict a difference of −0.065 * 0.20 ≈−0.013 on the social-cohesion scale. Is this a large difference in regard to the outcome? Pairs differ in social cohesion by about 0.17 and community efficacy by about 0.13 (the median differences) on 0 to 1 coded scales. So, a roughly median difference in perceptions would be associated with about a tenth of a median difference in social cohesion or community efficacy: a moderate effect if we are thinking about typical differences in surveys. We could also say that an extreme difference in perceptions within pairs would lead to a moderate difference in the outcomes within pairs (0.065 being close to 1/2 of the median difference within pairs in the outcomes). Another way to see this relationship is in standardized scales. The standard deviation of pairwise differences in perceptions is 0.20 and the standard deviations of social cohesion and community efficacy are 0.13 and 0.17 respectively. So, the effect of 1 sd difference in perceptions is about $ - .1 = - 0.065 \times \left( {{{.20} \over {.13}}} \right)$ standard deviations on the social-cohesion scale or about $ - .08 = - 0.065 \times \left( {{{.20} \over {.17}}} \right)$ standard deviations on the community efficacy scale. These are all moderate to low effect sizes estimated relatively precisely by our research design.

Another description of the relationship between perceptions and social cohesion within pairs is simply the proportion of pairs in which the person who reports perceiving more diversity is also the person with the lower social-cohesion score. In this case, 56 per cent of pairs have this pattern versus 44 per cent of pairs either tied in social cohesion or where the person perceiving more diversity has the higher social-cohesion score. For collective efficacy, the numbers are similar: 59 per cent of pairs have the higher diversity perceiver with the lower collective efficacy score.

These results help explain why diversity diminishes support for social goods provision: senses of reciprocity and shared preferences are weaker when majority group members believe they live among more outgroup members, even if they actually live among relatively few outgroup members according to the Census. The social cohesion and collective efficacy indices capture attitudes about what respondents think other people in their community would do.

Alternative Explanations

One might wonder whether antecedent conditions uncorrelated with objective context could explain both perceptions and social cohesion. Thus, our results would be explained not by differences in perceptions within pairs, but differences in ethnocentrism or socio-economic status that are uncorrelated with the ethnic diversity of a place, for example. We address this topic briefly to assuage worries about alternative explanations based on prejudice and socio-economic status.Footnote ²⁰

Could Ethnocentrism Drive Our Results?

What if prejudice explains our description of differences more than perceptions? For example, prejudiced individuals may react more strongly to outgroup members and thus perceive more of them (Nadeau et al. Reference Nadeau, Niemi and Levine1993). In other words, more prejudiced white individuals living in nearly identically diverse neighborhoods may both perceive more visible minorities living near them and be more likely to believe their neighbors do not share their values. In that case, while (mis)perceptions may be part of the story, the real engine driving the relationship would be the underlying ethnocentrism that exists as a stable personality trait unaffected by objective context (since this is held constant) or subjective perceptions. Since it is beyond the scope of this paper to explore the influence of all of the ways that people form perceptions – including their stereotypes, direct experience, or education – we engage with the question about ethnocentrism briefly here to show that, in fact, the results are not explainable by ethnocentrism.

If ethnocentrism drives our results, we should see a positive relationship between perceptions and expressions of prejudice. To assess this possibility we created a racial resentment measure from three items in the survey, and using the pairs from our first design we examined whether individuals who were more prejudiced within pairs were more likely to perceive more visible minorities in their local communities.Footnote ²¹ Figure 3 shows the opposite: the person within each pair who expresses more racial resentment on average is also likely to see fewer minorities in her community than her counterpart. As expected from past research, the figure also shows that more racial resentment is related to (slightly) lower levels of social cohesion and (appreciably) lower collective efficacy. It is clear that differences in ethnocentrism are not the same as differences in perceptions; nor do they act the same way in regard to our outcomes.Footnote ²²

Figure 3. Does ethnocentrism predict both perceptions and social cohesion? Within pairs, greater ethnocentrism is associated with lower perceived ethnic diversity, social cohesion, and collective efficacy. Pairs matched on Census-based per cent visible minority in the respondents’ DA, Census Subdivision (CSD) urbanicity, and change in per cent visible minority in the DA between 2006 and 2016 as described above.