Is Cultural Diversity Good for the Economy?

Is Cultural Diversity Good for the Economy? Wesley Sze Honours Undergraduate Thesis Written under the supervision of: Dr. Nicole Fortin Dr. Florian Hoffmann University of British Columbia Abstract. This paper describes the relationship between cultural diversity and the mean earnings and rents of native-born residents of Canada s metropolitan areas from 1981-2006. A highly robust positive correlation is found, where native-born workers living in more diverse cities received higher earnings and paid higher housing rents, even after controlling for relevant explanatory variables and city and time fixed effects. The relationship is not dependent on any particular city and is remarkably stable to robustness checks and time trend controls. Using instrumental variable estimation and economic reasoning, I argue that the correlation is at least partly driven by a causal effect stemming from diversity. Using a simple equilibrium model of earnings and rents, this causal effect is consistent with a dominant positive productivity effect of diversity on the native-born. An important new finding is that this diversity effect is almost entirely driven by the highly educated class diversity of the low educated does not contribute to earnings and rents. At the same time, everyone experiences the benefits of a diverse high education group, suggesting human capital type spillovers.

1 Introduction With the increase in immigration rates during the latter half of the twentieth century, many cities in North America, Europe, and Australia have experienced dramatic changes in the ethnic landscape of its populations. These cities have rapidly evolved into multiethnic and diverse urban communities within a relatively short period of time. With the undergoing of such profound demographic change, it is no wonder that immigration is of great importance for many today. One central issue that remains controversial is the effect immigrant inflows have on the well being of the native-born. While some see great value and benefit from increased cultural diversity, others fear that increased immigrant presence comes only at the expense of the native-born (Borjas 1994a). In addition to the social and cultural implications of immigration, its economic impacts form a critical part of the public discourse, connecting issues of ethnic diversity and immigration policy with labour market outcomes and economic growth. As such, many economists have taken an interest in better understanding the consequences of the growing cultural diversity brought about by immigrant inflows (for example, see Borjas [1990, 1992, 1994a], Card [2001], Peri [2007] among others). This paper examines the economic impacts of cultural diversity in a Canadian context, identifying the dominant utility or productivity effect that diversity brings to the native-born. It has been suggested, both in the economics and broader social science literature, that diversity has both consumption and production value (Glaeser et. al 2001). In the case of consumption, cultural diversity can lead to a wider availability of consumption goods (e.g. ethnic restaurants, arts and entertainment events), positively affecting individuals with love of variety preferences. 1 I call this the consumption variety!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 1 This term is borrowed from Dixit and Stiglitz (1977), who incorporate a taste for greater product variety through constant elasticity of substitution (CES) preferences. 2

hypothesis. Diversity can also have a productivity value. Workers from different cultural backgrounds may bring skills, abilities, and creativity that are complementary to each other in the production process (Lazear 1999). By complementing each other s skills, workers from different backgrounds may have a positive externality effect on one another that increases productivity. I refer to this as the complementary skillset hypothesis. While these two positive effects of diversity have been identified, a priori, one cannot conclude that a positive diversity effect certainly exists. Negative effects of diversity on utility and productivity may offset these benefits, either partially or entirely. In the case of utility, some individuals may have distaste for living in diverse environments. This can arise if one s cultural values feel threatened by the presence of foreign born. This view may be reinforced by the popular media and prevailing cultural attitudes, which may or may not be well founded. In addition, intergroup conflict may become problematic and further exacerbate the issue. Productivity may also be negatively affected by diversity if the skillsets of diverse workers are actually non-complementary or conflictual (O Reilly et. al 1997). For instance, differences in language, communication, or work ethic may offset any positive benefit from diversity in the production process. As such, I conjecture that cultural diversity may be a relevant determinant in consumption and production, though the direction of this effect is not immediately obvious. While some studies have addressed this question in the case of the United States (Ottaviano and Peri 2006; Peri 2007) and Europe (Bellini et. al 2008), little work has been done on identifying overall effects of cultural diversity in Canada. One of the goals of this paper is to quantify this effect in a Canadian context. There are several reasons why examining immigration from a Canadian perspective is worthwhile and meaningful. First, the composition of immigrants is significantly different in Canada than in the United States. 3

Whereas Mexico and Central America account for a large proportion of immigrants in the U.S., Canada s immigrants tend to come from a broader range of countries (though concentrated from China, Hong Kong, Southeast Asia, and India). Immigrants from different countries of birth presumably bring a different set of values and skills ( ethnic capital ) that play a role in the overall diversity effect. Second, undocumented immigration is a much greater concern in the United States than in Canada. It has been suggested that up to 40% of new arrivals in the United States enter through illegitimate means, often with little education and no English proficiency (Passel 2005). Not only does this create an unfavourable attitude toward immigrants as a whole in the U.S., but it also brings into question the accuracy of data since illegal immigrants may be underreported in data gathering processes. Lastly, a comparison of the results between the U.S. and Canada can help us better understand the role of institutional framework and governmental policy in determining diversity effects. One other contribution of this paper is in examining the role education plays in the diversity effect. The way diversity affects utility and production could depend on education level. This is especially true for the complementary skillset hypothesis, where certain industries and education groups may be better suited to benefit from the varied skills and creativity in diverse populations (Florida 2002). For example, industries with highly educated workers that emphasize creativity and originality (e.g. software development and IT) may be better suited to benefit from diversity than sectors that do not emphasize originality or independent thought. As to whether or not education is relevant in diversity s utility effect (e.g. instilling tolerance and preference for ethnic diversity) is less obvious 4

and perhaps more controversial. 2 Regardless, education seems to be a significant factor when looking at the impacts of cultural diversity. In this paper, I explicitly incorporate education into my analysis by segmenting diversity effects into both within and out-of group interactions. This method for controlling for education is an innovative feature of this paper and is found to be extremely relevant to our study of diversity. The empirical analysis focuses on identifying the relationship between cultural diversity and average earnings and rents of the native-born in metropolitan areas. Building upon a simple model of mobile workers and competitive firms originally developed by Roback (1982), earnings and rents are estimated together to identify the dominant utility or productivity effect of diversity. Using an index of fractionalization as a measure for diversity, I control for relevant explanatory variables and city and time fixed effects to determine any residual correlation between diversity and the earnings and rents of the native-born. Relationships within the whole population are examined first, followed by a decomposition of diversity, earnings, and rents according to education group. The main result I show is as follows: cultural diversity is positively related with native-born earnings and rents, though this relationship is driven only by the diversity of the highly educated. This second qualification is, to my knowledge, a new contribution to the understanding of how diversity affects economic outcomes. This diversity correlation is at least partially causal in nature and consistent with a dominant positive productivity effect stemming from diversity. The remainder of the paper is organized as follows: section 2 provides a summary of the existing economics literature that addresses the relationship between immigration,!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 2 The relationship between education and attitudes towards racial diversity has been more extensively studied by scholars in the fields of sociology and psychology of education. See Troyna (1990) and Smolicz (1996) for a more detailed discussion. 5

cultural diversity, and economic outcomes. Section 3 briefly introduces Roback s theoretical model that is used to develop a consistent estimation procedure for the diversity effect on mean earnings and rents. Section 4 describes the data sources, key summary statistics, and stylized facts of my research. I present and discuss my key estimation results in section 5, including variations on the basic specifications, robustness checks, and addressing causality and endogeneity. A conclusion of my findings is summarized in section 6, along with a discussion of the limitations of this study and directions for future research. 2 Literature Review Economists have been interested in understanding the economic implications of cultural diversity for as long as immigration has been occurring. Its relevance has only increased as technological advancements and political reform have ushered in an era of unprecedented migration previously not possible. Initially, the economics literature focused on examining the short run effects of immigrant worker inflows on the labour market outcomes of the native-born. Much of this early work was likely motivated by a concern that rising immigration would bring down the employment and earnings of the native-born. However, little economic evidence supports this hypothesis. In a survey of the economic literature on immigration, Borjas (1994) finds no evidence for any significant negative effect of immigration on native outcomes. The methodology employed by early empirical work focused on correlating changes in immigration rates with changes in key labour market indicators (earnings, employment rate), studying short run changes within localized geographic regions. This so-called area-analysis approach found only very modest effects, with estimates for the elasticity of native wages with respect to the number of immigrants converging around -.02 (Grossman 1982, Borjas 1990). Estimates for the elasticity of native 6

employment with respect to immigration have been even less significant (Borjas 1990, Altonji and Card 1991). A similar result was confirmed by Card (1990), who exploited an exogenous change in the immigrant labour supply in Miami caused by Cuban immigrants from the Mariel Boatlift of 1980. Using this natural experiment, Card found only a negligible effect on the employment outcomes of native-born low-skilled workers, consistent with the results of the area-analysis methodology previously employed. Overall, there appears to be minimal empirical support for a substantial adverse effect of immigration on the labour market performance of natives. This area-analysis approach, though originally helpful, is limited in scope. One weakness is that it fails to take into account worker movement in response to immigration. For example, it assumes closed geographic regions, whereas in reality workers are mobile across regions. Migration flows are also rarely exogenous and workers and immigrants presumably self-select into areas with the best opportunities. In addition, local workers may out-migrate from immigrant-receiving cities, such that the actual effects of immigration are understated (Frey 1996). Card (2001) attempts to address this issue of worker mobility by looking only at how much immigration drives out similarly skilled workers in U.S. cities. He finds little evidence for out-migration of workers in response to immigrant inflows, and in fact finds weak support for the opposite: that increases in immigration actually encourage a net inflow of similarly skilled workers. These results, along with previous area-analysis studies, seem to confirm that the arrival of immigrants has a negligible effect on the labour market outcomes of natives. While there have been numerous economic studies on immigration, there has been far less study of cultural diversity, an important consequence of immigration. The two, though closely related, are distinct concepts. Diversity emphasizes the cultural richness and 7

variety that immigrants bring to society, and not just the presence of foreign-born or their measurable skill level or wealth. For example, an area with many immigrants from the same place of origin may have a large immigrant share yet low diversity. Most economic studies have focused on immigrants as a single, homogenous group. As such, they have little to say about how diversity affects individuals living in multiethnic communities. Social scientists have long hypothesized that cultural diversity may play a significant role in affecting productivity and consumption values. Much of the study of diversity has been confined to cities, where the effects of diversity are most discernible. Quigley (1998) identifies diverse populations within cities as one of the driving forces behind their economic growth and success. Through the increased variety in the availability of goods and services, Quigley argues diverse cities are able to draw upon shared inputs in production and consumption, taking advantage of economies of scale. This, he argues, results in more innovative and productive firms and workers in diverse cities. He also finds that heterogeneous labour markets have greater availability of workers with different skills, resulting in more efficient matching of workers and jobs (cf. the complementary skillset hypothesis). In a similar vein, sociologist Florida (2002) identifies diversity as one of the key features of cities that attract skilled and creative individuals. This concentration of the so-called creative class in cities creates environments conducive to innovation, productivity, and economic growth. Economists and psychologists have also long conjectured a love-of-variety in preferences that yields positive utility effects when more choices are available (cf. the consumption variety hypothesis). Applications of tastes for variety have been widely used in the international trade literature, where some of the gains from trade arise from the increased variety of consumption goods (for example, see Krugman 1980 and Broda and Weinstein 2006). 8

Research in human capital externalities and peer effects are also relevant to our study of diversity. Although the native-born do not contribute to cultural diversity, they may still experience the effects of immigrant diversity. Furthermore, I hypothesize that diversity effects are closely related with education. If so, a question to consider is how different education groups contribute and respond to changes in diversity. Just as education is thought to have a positive externality that exceeds private return (Moretti 2003, 2004), diversity within certain groups may very well have spillover qualities that benefit society as a whole. Peer effects have also been identified in ethnic capital, the unique social and cultural experiences and upbringings that ethnic groups possess (Borjas 1994b). Borjas finds that a person s ethnicity plays an important role in the formation of human capital, where one s productivity and skill is related with their ethnic group s average skill level. This is not to say that the effects of increased cultural diversity are entirely beneficial. Easterly and Levine (1997) find societal fragmentation to be a key factor explaining cross-country differences in growth rates. The authors examine the case of countries in Sub-Saharan Africa, using an index of fractionalization based on ethnolinguistic groups to proxy for ethnic diversity. Using cross-country panel regressions over three decades, they find ethnic diversity to be negatively correlated with per capita GDP growth. In addition, they find that regions of higher ethnic diversity are associated with low schooling, political instability, and lack of public infrastructure. The paper argues that more diverse societies promote rent-seeking and predatory behavior that impedes economic growth. Similarly, Alesina and La Ferrara (2005) look into the effects of ethnic diversity on economic performance and policies at the country, county, and small village levels. Indeed, their findings corroborate those of Easterly and Levine: they find that ethnic diversity is negatively correlated with growth rates, even after controlling for region variables. They 9

point to difficulties in communication and incompatibility of skills and preferences across cultural groups as factors that outweigh the benefits of diversity. It should be noted that these studies have disproportionately focused on developing countries with inherently weak and underdeveloped political institutions. Despite these findings, political scientist Paul Collier (2001) suggests that ethnic diversity, when framed within well functioning political institutions and processes, is not necessarily harmful to economic growth. Collier argues that the drawbacks of societal fractionalization can be overcome with adequate democratic institutions, even allowing the private sector to flourish in diverse societies. This suggests that strong institutions are necessary for diversity to benefit society and resolves seemingly inconsistent findings about diversity in different areas. Most similar to this paper s study of diversity is the work of Ottaviano and Peri (2006), who model cultural diversity as a city amenity affecting utility and production functions. They build upon a simple model to identify long run equilibrium effects of diversity on the average wages and rents of U.S. born residents after allowing for free mobility of residents and firms. Using Census data from 1970 and 1990, the authors calculate an index of diversity based on country of birth as a key explanatory variable to explain differences in mean wages and rents across cities over the two time periods. Their findings suggest a dominant positive productivity effect of diversity, where wages and rents are positively and significantly correlated with diversity. Their results are robust to the inclusion of additional control variables and instrumental variable estimation is used to confirm a causal effect. The work of Ottaviano and Peri is thought to be one of the first to look at the long run equilibrium effects of diversity. Bellini et al. (2008) have replicated this study for areas 10

in the European Union, finding results that are similar to the United States. Peri (2007) also repeated the original study for the case of California, finding a similar positive effect of immigrant diversity on the native-born in California. This paper builds upon this same approach with an extension to Canadian cities. A modification to the methodology is made to address education as a mechanism through which diversity affects the economy, a question previously left unanswered. 3 Underlying Economic Framework This section presents a model to formalize how diversity affects earnings and rents, resulting in a strategy to estimate the effect of diversity on equilibrium earnings and rents. The model is an adaptation of a well-cited equilibrium framework proposed by Roback (1982), which focuses on the role of city amenities in allocating workers across cities based on wages and rents. This model has been extensively used in the social sciences to measure the value of unique city qualities. Closely following the work of Ottaviano and Peri (2006), I modify Roback s general framework to model cultural diversity as a city-specific feature that enters into the utility and production functions. If diversity is an amenity (disamenity), then workers in equilibrium would be willing to pay higher (lower) rents in cities with a high level of diversity. Similarly, if diversity has a positive (negative) productivity value, workers in equilibrium would receive higher (lower) earnings to reflect their marginal productivity. By estimating earnings and rents together in relation to diversity, a dominant utility or productivity effect of diversity can be determined. 11

A concrete example confirms this intuition and provides a specific empirical identification strategy. 3 Suppose there are a given number of distinct cities with freely mobile households and competitive firms. Cities are distinct in the sense that they do no overlap and do not contain residents working in another city (or have firms employing workers residing in a different city). Free mobility assumes that in the long run individuals and firms are able to freely choose their location. Preferences of individuals are Cobb Douglas in the consumption of land! and a homogenous composite good!, with diversity affecting a shift factor!. The utility function for individuals living city! is defined as!!!!!"#!!!!!!!!!! (1) where!!!!!. If diversity is an amenity, then!"!"#$ is positive. Utility maximization yields an indirect utility function based on the rent!!, wage!!, and diversity!"#! of city!. The free mobility assumption implies that in the long run, all workers in each city receive the same level of utility, such that no worker desires to change cities. That is,!!!!!!!!!"#!!!!"#!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! (2) where!!!!! is the indirect utility function,!! is the expenditure of an individual, and! is the equilibrium level of utility. Wages and rents adjust to equalize utility in all locations. The free mobility condition implies that an exogenous increase in wages or rents is matched with a similar increase the in the price of the other factor. A graphical representation of equation (2) is shown in figure 1, denoted by an upward sloping curve with diversity acting as a shift factor in the wage-rent graph space. The case for firms is parallel. The production function for the composite good! is Cobb Douglas with two factors of production, land! and labour!. Diversity affects!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 3 The example I use is analogous to the model introduced by Ottaviano and Peri (2006). For brevity, only the key elements are presented in this paper. For a more detailed exposition and discussion of the model, see Roback (1982) and Ottaviano and Peri (2006). 12

production through a productivity shift factor!. A firm in city! operates according to the production function!!!!!!"#!!!!!!!!!!! (3) where!!!!!. Again, if diversity is productive,!"!"#$ is positive. In perfect competition, firms profit maximize and set a price equal to marginal cost. When the good is freely traded between cities the zero-profit equilibrium condition implies!!!!!!!!!!!!!!!!!!!!!!!!!!"#!!!!!!!!!!!!!!!!!! (4) Again, wages and rents adjust to equalize the price of the good across cities, depending on diversity levels in each city. In equilibrium, the condition requires that an increase in the cost of one input to be offset by a decrease in the other, such that all firms earn zero profit. The downward sloping curve in figure 1 is a graphical representation of equation (4) in the wage-rent space. Again, diversity acts as a shift factor if diversity is productive, it shifts the curve upwards to reflect higher factor prices. Figure 1 shows the spatial equilibrium of our model. A pair of wages and rents exists such that the free mobility and zero profit conditions are met. As diversity changes, the two curves shift to form a new equilibrium. Observing how the equilibrium wages and rents shift in response to changes in diversity forms the basis for the identification procedure. Solving equations (2) and (4) give explicit expressions for equilibrium rents and wages in each city:!"!!!!!!!"!"!!!!!!!!!!!"!!!!!!!!!!!!!"!!"#!!!!"#! (5)!"!!!!!!!!!"!!!!!!!!!!!!!"!!!!!!!!!!!!!"!!!"!!"#!!!!!"#!!!! (6) These two equations make it clear that observing only wages or rents in isolation cannot identify the diversity effect. For example, a positive correlation between rents and diversity is consistent with both a positive utility or productivity effect. At the same time, a positive correlation between wages and diversity is consistent with either a negative utility effect or 13

a positive productivity effect. However, jointly estimating equations (5) and (6) together yields the following identification procedure:!"!"#$!"!"#$ Dominant Effect Positive Positive Positive productivity effect Positive Negative Positive utility effect Negative Negative Negative productivity effect Negative Positive Negative utility effect Thus by observing the signs of the diversity coefficient in each of the earnings and rent regressions a dominant effect of diversity can be identified. Most relevant to this study is the dominant positive productivity effect, which is associated with a positive correlation between diversity, earnings, and rents. 4 Data Description and Summary Statistics 4.1 Data Source and Diversity Index The primary unit of observation used for this study is at the Census Metropolitan Area (CMA) level, defined by Statistics Canada as closely integrated urban areas with populations greater than 100,000. The use of CMA level data is useful for several reasons. Theoretically, the economics literature suggests that the hypothesized utility and productivity effects from diversity are most prevalent in the context of urban agglomerations. CMAs also tend to exhibit higher levels of variation in cultural diversity than in non-urban areas, which tend to have lower levels of diversity with little change over time. On a more practical note, the primary data source used to calculate city level variables is the Statistics Canada Public Use Microdata File (PUMF), which restricts geographic information to the CMA level. For these reasons, using aggregated CMA level data is both theoretically appropriate and practically feasible. 14

This study covers the time period 1981-2006, with data collected in five-year intervals for a total of six time period observations. Over this time span, Statistics Canada has added new CMAs as more urban areas reach the 100,000-population threshold, thereby increasing the number of unique CMA identifiers from 13 in 1981 to 23 in 2006. Use of an unbalanced panel dataset is not problematic in this case, as the population criteria is consistently applied to all urban areas and should not be correlated with the errors. 4 In total, there are 99 unique city-year observations included in the time series panel data set. One peculiarity worth noting is the merging of certain urban areas in the PUMF as a single CMA. This is done to ensure anonymity of results; however, it also presents a data quality issue. Although several city combinations are geographically and culturally justified (e.g. Ottawa-Gatineau 5, St. Catherine s-niagara), many others share little commonality (e.g. Regina-Saskatoon, Sudbury-Thunder Bay, Sherbrooke-Trois Rivières). The merging of these cities data violates the assumption of non-overlapping and distinct cities and could impact the way diversity affects earnings and rents. However as a practical consideration, in order to maximize the number of observations all merged CMAs have been treated as single urban areas. Robustness checks are performed to examine the sensitivity of the estimation results to the inclusion of these mismatched CMAs. The key variable of interest is the measure of cultural diversity. I use an index of fractionalization, an index commonly used in demographic research to quantify levels of variation in categorical data (see Ottaviano and Peri [2006] Easterly and Levine [1999] for previous economic studies using this measure). As we are interested in measuring cultural!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 4 That is, the only factor deciding whether or not a city is included is whether it meets the population threshold. Since the scope of this study is already limited to urban areas, this restriction does not pose any systematic selection bias. 5 For the purposes of this study, the Ottawa-Gatineau CMA is considered to be in the province of Ontario, even though Gatineau is a part of Quebec. 15

diversity, the place of birth response variable on the PUMF is used as a proxy for cultural background. Place of birth is likely to capture differences in language, educational background, social values and attitudes, religion, and other aspects that contribute to overall cultural diversity. An advantage of using this variable is that it is easily identified and consistently measured in each time period. Calculated for every CMA in each year, the diversity index is defined as!!"#!!!!!!!!"#!!!!!!!! (7) where!!"#!!!!! is the square of the share of people from place of birth! amongst residents of city! in year!. Individual countries have been aggregated into groups such that the diversity index is always calculated with the same categories of origin (see the data appendix for a detailed list of country groupings). The index quantifies diversity by approximating the probability that two randomly chosen individuals were born in the same country. A perfectly homogenous city would have an index of zero, whereas a perfectly heterogeneous city where no individual is born in the same place would have an index approaching a value of one. 6 Table 1 reports summary statistics for the mean diversity indices by education group for each CMA in descending order, from most to least diverse. There is remarkable amount of variation in diversity across the cities. The most diversified cities, Toronto and Vancouver, have indices of.632 and.540, whereas the Sherbrooke-Trois Rivières and Québec CMAs show a much higher degree of homogeneity with values of.049 and.046. A further decomposition of diversity has been made according to education group. Diversity indices for the high and low education groups have been calculated using equation (7), except restricting the sample population to those with at least some college education!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 6 For a given number of places of birth N, the maximum value of the index is 1-N -1. This value approaches one as N becomes large. 16

(high education group) and those with at most high school completion (low education group). Breaking down diversity this way reveals several facts. Naturally, there is a strong correlation between the diversity of the entire population and the diversity of each education group (i.e. the more diverse the entire population, the more diverse each education group). This correlation holds stronger for the low education group (correlation coefficient of.993) than for the high education group (correlation coefficient of.892). Diversity is also found to be significantly different between the two education groups, even within the same city. Low education groups tend to be significantly more diverse than their highly educated counterparts. In some cities, the difference is as large as 50%, suggesting that immigrants tend to be less educated than those of the native-born, thus contributing to the higher diversity of the low education group. 4.2 Mean Earnings and Rents Table 2 presents descriptive statistics for the key dependent variables, mean earnings and rents of the native-born. Again, each variable has been calculated for the entire population, as well as high and low education groups. Because we are interested in utility and productivity effects stemming from diversity itself (and not just changes to the population due to immigrants), the sample for calculating mean earnings and rents is limited to nativeborn non-immigrants. I also control the sample for other demographic variables that could affect earnings and rents that are unrelated to diversity. Mean earnings have been calculated from the wages and salary income response on the PUMF for Canadian-born, full-time employed males aged 30-50 for each CMA in each time period. To account for inflationary growth in earnings, nominal values have been deflated into 2002 real dollars using the Canadian Consumer Price Index (2001 basket content). A similar calculation was 17

used for obtaining the mean monthly rent per room (gross rent divided by number of rooms response on the PUMF) for each CMA. For the rent calculations, the sample population was controlled for native-born males aged 16-65 for each CMA in each time period, again expressed in 2002 dollars. Using the same order of descending diversity in table 1, table 2 summarizes the average mean earnings and rents for each city during the time period, with further distinctions made according to education group. The summary statistics confirm our intuition on the relationship between earnings and rents. Again, there is a positive correlation between mean earnings and rents (correlation coefficient.6346): non-immigrants living in cities with higher average earnings also tend to pay higher rents. As expected, the highly educated receive higher earnings than those with less education. The earnings premium for the highly educated ranges from 40 to 60 percent. Lastly I note that in addition to receiving higher earnings, those with more education also tend to pay higher rents. Table 3 presents summary statistics for the main control variables used in the regression framework. The mean, standard deviation, minimum, and maximum values for each variable are reported. Variables for population and employment rate were obtained from the Statistics Canada CANSIM database. For combined CMAs, population and employment rate were calculated as the weighted average of each city s specific value. The shares of high and low education groups are defined as earlier and are used as a control for a city s education level. 18

4.3 Stylized Facts Economic theory suggests there is the possibility of a causal relationship stemming from diversity to mean earnings and rents. Before using formal statistical methods to test this hypothesis, a preliminary graphical representation helps to reinforce our finding of a positive relationship. Figures 2a and 3a show the scatterplots of correlations between change in diversity (entire population) and percentage change in mean earnings and rents of the native-born from 1991 to 2006. The positively sloped best-fit line suggests a positive relationship between diversity and mean earnings and rents. The slope coefficients estimate a.1 increase in diversity to be associated with an 11 and 17 percent growth in earnings and rents, respectively. Moreover, the positive relationship does not seem to be driven by any particular city. A notable outlier in figure 2a is Calgary, which experienced very high growth in earnings that exceeds the trend. This is perhaps explained by an exogenous boom to Alberta s energy sector, raising wages above normal. All in all, however, these initial figures offer evidence for an underlying positive diversity effect on earnings and rents. A further distinction of changes in diversity within education groups further confirms our hypothesis that education matters in determining the diversity effect. Figures 2b and 3b show the scatterplots for changes in diversity amongst the highly educated with mean earnings and rents of the native-born. The positive relationship is clearly strengthened with a better fit and lower dispersion around the regression line. This suggests that the diversity effect is stronger for the high education group. On the contrary, when controlling for diversity amongst the low education group in figures 2c and 3c, the relationship between diversity, earnings, and rents is lost. The regression line has poor fit, a low R-squared value, and is no longer significant. This indicates that the low educated group may contribute very little to the overall diversity effect. 19

Although these simple scatterplots do not take into account relevant controls or fixed effects estimation, they do provide visual evidence that reinforces this paper s key findings. Indeed, they support the proposition that cities with the highest growth in diversity also experienced highest growth in earnings and rents for the native-born. Furthermore, the stronger fit when controlling for diversity amongst the high education group emphasizes the strength of high education diversity in driving the overall correlation. 5 Estimation Results 5.1 Basic Earnings and Rent Regressions The theoretical model requires the simultaneous estimation of both earnings and rent regressions for each city to identify dominant utility or productivity effects of diversity. To estimate mean earnings for each CMA in each time period, the following OLS regression is used as an implementation of equation (6):!"!"#$%$&'!!!!!!!!!!"#!!!!!!!!!!!!!!!!! (7) The dependent variable,!"!!!"#$%$&'!!!! is the natural logarithm of the mean real annual earnings for native-born, full time employed males aged 30-50 in CMA! at time!. The variable of interest,!"#!!!!! is the diversity index calculated amongst the population of city! in education group! at time!. Education groups are divided into two broad categories: low education (high school or less) and high education (some college or greater). An overall diversity index that is not specific to a particular education group is also used to measure overall diversity of the entire population.! is a row vector of coefficients for!!, a column vector of control regressors. A measure to control for a city s education level is included in every regression. City fixed effects!! and year fixed effects!! are included, as is an 20

unobserved residual! with zero conditional mean. 7 In our discussion!! is the relevant slope that captures the equilibrium effect of diversity on earnings. The identification procedure also requires the simultaneous estimation of a parallel regression for mean rents. Analogous to the earnings regression, the following OLS regression is used as an application of equation (5):!"!"#$!!!!!!!!!!"#!!!!!!!!!!!!!!!!! (8) The dependent variable,!"!!!"#$!!!!, is the natural logarithm of mean real monthly rent per room for native-born males aged 16-65 in CMA! at time!. The measure of diversity,!"#!!!!! is the same as used in the earnings regression, and! is also a row vector of coefficients for!!, a column vector of control regressors. In addition, the regression includes city fixed effects!! and time fixed effects!!. Again,! is an unobserved residual with zero conditional mean.!! is the coefficient of interest that captures the equilibrium effect of diversity on rents. The OLS regression estimates for the basic earnings and rent equations (7) and (8) are reported in tables 4 and 5. Specifications I and II in table 4 show that diversity amongst the entire population has a positive, but not significant, correlation with earnings. The point estimate suggests that a.1 increase in diversity is associated with a 4% increase in earnings of the native-born. The positive correlation I find is consistent with Ottaviano and Peri (2006), though at a smaller magnitude. Specifications III-VI break down diversity according to education group. This decomposition confirms that education does matter in the diversity-earnings correlation: diversity amongst the highly educated is strongly significant, whereas diversity amongst the low educated group is a much weaker and insignificant!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 7 This zero conditional mean assumption is required for the OLS estimation to be unbiased. In section 5.4, I test for robustness to omitted variables to check if the basic regression gives a consistent estimate for the real parameter value. 21

predictor of earnings. The point estimate for the coefficient for diversity amongst the highly educated is approximately 1.22 (more than eight times greater than the estimate for the low educated diversity), suggesting a.1 increase in diversity to predict a 12% increase in average earnings. All specifications have been tested both with and without a control for employment rate, which is partially endogenously determined in our model, but could also have effects on productivity (Ciccone and Hall 1996). Whether or not a control for employment is included, however, does not significantly change the estimates for diversity. One important control for the earnings regressions is education. While some similar studies have used average years of schooling as an education control, there is not enough variation is average years of schooling across cities in our sample to obtain a reliable and statistically meaningful control variable. 8 Instead, the share of highly educated within a city is used as a measure for education level. Not only does this control for the educational composition of a city s workforce, but also takes into account human capital spillover effects of the high education class that have been hypothesized by several economists (notable examples are Moretti [2004a, 2004b], Shapiro [2005]). The results of the basic earnings specifications show a positive, though insignificant, estimate for the share highly educated. Here the insignificance of these estimates may be due to the relatively low variation in this variable over time. Therefore, this measure still may not accurately control for education s actual effects on earnings. However, because education is an important factor to include, I return to this question of education in section 5.2 where earnings, rents, and diversity are decomposed by education level. The parallel rent regressions are reported in table 5. The results are akin to those of the wage regressions, although the standard errors are generally larger than those in table 4.!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 8 This control for education is used as a robustness check in specification XIII of table 8. Using this alternate measure does not meaningfully change the results. 22

The initial estimate in specification I for the coefficient on diversity is.391 which, like the case of the earnings regression, is positive but insignificant. Breaking diversity down into the two education groups yields very similar results. Specifications III and IV show that diversity amongst the highly educated is, again, a highly significant predictor of rents with an even larger magnitude of 1.6 to 1.7. This suggests that a.1 increase in diversity is matched with a 16 to 17% increase in rents, which is slightly larger in magnitude than for earnings. In specification V, the coefficient for diversity is virtually zero, suggesting that diversity amongst the low education group is uncorrelated with rents. All specifications have been tested with and without a control for population, which is partly endogenously determined in the model through the free migration condition. Although the inclusion of this partially endogenous variable may bias the estimates, in this case it does not qualitatively change the results. This set of rent regressions shows that the diversity of the highly educated is a consistently strong predictor of rents, whereas the diversity amongst the low educated group remains insignificant. As a whole, the basic earnings and rent regressions suggest that diversity of the entire population is weakly positively related with wages and rents (!!!!!!!!! in specifications I of tables 4 and 5). This is consistent with previous studies, which have shown a positive correlation between diversity, earnings, and rents (Card 2009). The results, however, are not statistically significant for the diversity variable of the overall population. Nonetheless, we find that diversity amongst the highly educated is a consistently strong predictor of earnings and rents with a very low p-value in every case. As the highly educated group is a subset of the entire population, the decomposed regressions indicate that the high education group drives the overall positive diversity effect. Diversity of the 23

low education group is never significant and seems to contribute only very weakly (or even negatively) to the overall diversity effect. 5.2 Role of Education In this section I consider a refinement to the estimation strategy of the previous section to better understand the role of education in determining the diversity effect. In addition to decomposing diversity by education, mean earnings and rents are also calculated for each education group. This shows whether diversity effects are localized within specific groups or spillover to other groups. It allows us to answer the questions: do those with low education benefit from a more diverse highly educated workforce? Are diversity effects confined to particular education groups, or are its benefits experienced by all? Previous studies have shown that these spillovers exist for human capital (Moretti 2004); this section will examine if such spillover effects also apply to diversity. The estimates for the coefficients on the diversity indices for the mean earnings and rents for each education group are reported in table 6. 9 Each entry represents a separate regression, each controlled for city and year fixed effects. The diagonals represent within group effects, whereas the off-diagonal entries represent spillover effects into other groups. The most obvious result is that diversity of the highly educated is positively and significantly correlated with the earnings and rents of every education group. This reinforces the finding of section 5.1, which showed that the diversity effect is almost entirely driven by the high education group. Diversity effects should also be strongest within group and weaker in others. Indeed, the within-group estimates (1.444 and 1.807 for earnings and rents, respectively) are larger in magnitude than the out-of-group estimates (.817 and 1.626,!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 9 The last columns for earnings and rents correspond to the estimates found in tables 4 and 5, and are included as a point of reference. 24

respectively). A diverse high education group benefits the highly educated the most, though also benefits others to a lesser extent. The coefficients are always positive and statistically significant. Using the model s identification strategy, this is consistent with a dominant positive productivity effect of a diverse high education group (!!!!!!!!!!. On the other hand, diversity of the low education group is never significant, not even affecting own-group earnings or rents. Although the standard errors are comparable to the estimates for the high education diversity, the point estimates are considerably lower. As such, there is little evidence for any diversity effect coming from the low education group. This is in contrast to the diversity of the high education group, which is consistently significant. One possible explanation for the difference between the two groups is nature of certain industries, where some sectors benefit more from diversity than others. For instance, the high-tech sector may be more adept to exploit diversity productivity effects, whereas a labour intensive industry (e.g. construction, manufacturing) may be less suited to experiencing the benefits of a diverse workforce. Whatever the reason, diversity seems to affect outcomes only through the highly educated class, whether it is from the nature of highly educated industries or ethnic capital spillovers (Borjas 1992, 1994b). 5.3 Long Run Equilibrium Effects Lastly, employment and population are considered as endogenous variables in the model. The theoretical model suggests that in long run equilibrium, workers are mobile and sort themselves based on the distribution of wages and rents across cities. Earlier, employment and population were described as partly endogenously determined in the model. Table 7 explicitly tests for this by regressing employment and population against diversity. 25

Specifications I-III show that diversity and employment are uncorrelated. This confirms the hypothesis that in the long run, under free migration conditions, workers sort themselves across cities to an equilibrium such that there is no incentive to move. Employment rate, being related to job opportunities and prospects for employment, may act as a draw factor for prospective workers. An unusually high employment rate would attract an inflow of workers until the rate returns to normal, at which point a new equilibrium is reached. The insignificant relationship between diversity and employment rate verify this equilibrium effect. Specifications IV-VI show that diversity is positively correlated with population. This reinforces the expectation that more diverse cities also tend to be the most populous, as immigrants tend to settle in larger metropolitan areas with larger preexisting immigrant communities (Edin et. al 2003, Stark 1991). It also provides counter-evidence for the outmigration hypothesis. The positive correlation between population size and diversity indicates that immigrant-receiving cities have net growth in population. Thus, it is unlikely that the outflow of natives in response to immigration is significantly large. In addition, cities seem to be able to absorb the increase in population with new job opportunities, ultimately leaving the employment rate unchanged. If immigrant inflows really led to substantial internal migration of natives, we would expect to see fluctuations in employment rates as workers adjusted themselves across cities. These additional tests on the endogenous variables of the model reinforce our understanding of the spatial equilibrium and lend support for the dominant positive productivity effect. 26