Segregation and individual employment: a longitudinal study of neighborhood effects

In order to describe the development of segregation in Sweden, we need to use some measure of the phenomenon. A common measure of segregation used in empirical studies is the dissimilarity index² (see, for example, Tauber and Tauber 1965; Cutler and Glaeser 1997; Cutler et al. 1999; Aldén et al. 2015), which dates back to Hoover (1941). This index measures how minority and majority group individuals are distributed across sub-areas, such as neighborhoods of a region. The more evenly the minority group is distributed relative to the majority group, the lower is the level of segregation. Following Duncan and Duncan (1955), the dissimilarity index, \(D\), is calculated as follows:where \(b_{k}\) is the number of minority group individuals in neighborhood \(k\), \(B\) the corresponding number in the respective region, \(a_{k}\) the number of majority group individuals in neighborhood \(k\), and \(A\) the corresponding number in the region. \(n\) is the number of neighborhoods in a region. \(D\) ranges from 0 to 1 and can be interpreted as the proportion of minority group individuals in the region that need to change neighborhood of residence in order to achieve the same geographical distribution as the majority group. In the present paper, the dissimilarity index is based on the distribution of minority and majority group individuals across parishes (which represent neighborhoods) within labor market regions. Each of the 2513 parishes in Sweden is located in one of the 93 labor market regions in Sweden.³ Minority group individuals are those with a foreign background, classified according to Statistics Sweden as either foreign born or native born with two foreign-born parents. Second generation immigrants are thus included in the minority group, because the cultural identities of parents are often upheld in the next generation (Waters 1994; Portes and Rumbaut 2005). Individuals born in Sweden with at least one native-born parent are categorized in the majority group.

To assess the overall change in segregation between 1990 and 2011 in Sweden, we calculate the yearly dissimilarity index for each of the 93 labor market regions and weight it with regard to population. This is summarized in Fig. 1.

The dissimilarity index for Sweden calculated for 1990 is approximately 0.1944, which implies that, on average, 19.44% of the individuals with foreign background needed to move to another neighborhood in order to reach the same geographical distribution as the majority population. Figure 1 shows a continuous increase in average segregation in Sweden, and in 2011, the weighted dissimilarity index reached approximately 0.2471, which implies an increase in average segregation by 5% units. The results of Aldén et al. (2015), who calculate the dissimilarity index for the 12 largest Swedish municipalities between 1990 and 2007, point in the same direction, particularly for the earlier part of the period. This development of residential segregation supports the conclusion from the bid-rent model with ethnic externalities described in Sect. 2.2, namely that segregation moves toward a complete spatial separation of different groups of the population.

Figure 1 gives an overall view of the change in segregation in Sweden between 1990 and 2011, and regional discrepancies are hence suppressed. To shed some light on differences among regions, Fig. 2 shows the level of segregation in 2011 as well as the change in the dissimilarity index between 1990 and 2011.

In 2011, the most segregated region in Sweden was Södertälje, which is located just south of the capital region of Stockholm. In 2011, the dissimilarity index for Södertälje was 0.3924, which implies that almost 40% of the population with a foreign background needed to relocate within the region to reach the same distribution as the population with native background. This figure can be compared to 28% in 1990. This increase implies that Södertälje is one of the regions where segregation increased the most between 1990 and 2011. The right map in Fig. 2 shows that a majority of regions in Sweden saw an increase in segregation between 1990 and 2011, albeit in varying degrees.

Figure 2 also shows that segregation is not merely an urban phenomenon. The three metropolitan regions in Sweden: Stockholm (Sthlm), Gothenburg (Gbg), and Malmö, do show up among the 25 most segregated regions, but considerably smaller regions and smaller cities also have relatively high degrees of segregation. In general, segregation and increases in segregation are greater in the more densely populated southern part of Sweden.

To understand the fundamental mechanisms in the housing market that influence the outcome in terms of ethnic segregation, we can refer to the neoclassical bid-rent theory, building on the seminal model presented by Alonso (1964), Mills (1967), and Muth (1969), often referred to as the Mills–Muth model. The original version of this model assumes a monocentric urban region with a central business district in its core. An extension of this model is to introduce prejudice as preference of an individual (or household) toward some other group of the population. This type of modification was presented by Yinger (1976), while Fujita (1989) presents a comprehensive overview of the application of the bid-rent model, including racial and ethnic perspectives, where neighborhood externalities are considered.

To understand the outcome of such bid-rent models, we follow Fujita (1989) and assume a situation with two groups of the population, \(A\) and \(B\). \(A\)-households have an aversion toward living close to \(B\)-households, while \(B\)-households are indifferent in this respect. The utility function for \(A\)-households can then be written as \(U_{A} (z,s,E(k))\), where \(E(k)\) is the externality \(A\)-households experience at location \(k\) (the level of experienced externality at \(k\) is negatively associated with closeness to \(B\)-households at \(k\)) given the spatial distribution of \(B\)-households in the city, \(z\) is the composite consumer good, and \(s\) is the consumption of land or the lot size of the house. Transportation costs increase with distance \(r\) from the CBD of a location \(k\). It is assumed that \(U_{A}\) increases when \(E\left( k \right)\) increases (Fujita 1989):Using the bid-rent function including externalities at a given location, several models have been formulated to explain spatial distributions of population groups in a city region. Bailey (1959) and Rose-Ackerman (1975, 1977) elaborate on the so-called border-model, which assumes that the two groups of households are completely segregated in a city, where one of the groups occupies the inner (outer) part of the city and \(E(k)\) increases as the distance from the border dividing the inner and outer parts of the city increases. The strength of the prejudices will influence the rate of speed as well as the magnitude for these processes as they develop over time. However, Schelling (1971) shows theoretically that even weak preferences can lead to fully segregated neighborhoods. For Sweden, these conclusions are supported empirically by Aldén et al. (2015), who show that increases in neighborhood concentration of immigrants are associated with decreases in the native population.

Kain (1968) argues that residential segregation isolates minority groups from job locations, which imposes spatial barriers. This spatial mismatch implies high costs due to physical distance, a situation that discourages individuals from seeking employment in other areas. In addition, the spatial barriers imply informational barriers about job opportunities as well as potential labor market discrimination against minority groups from employers in areas populated mainly by the majority group. Kain (1968) even argues that what is usually called employment discrimination is in fact a result of housing segregation. In addition, Kain (1968) presents an early empirical contribution, showing a negative relationship between housing segregation and the employment of minority group individuals. There is also a large body of studies on the connection between spatial mobility and social mobility (see Massey et al. 1987 for an overview).

As discussed already by Kain (1968), spatial separation reinforces social distance since segregation relegates already disadvantaged individuals to areas with fewer opportunities and less amenities (Logan 1978). Residential segregation also implies less interaction with natives and fewer incentives to acquire country-specific human capital and language skills, which might have long-term effects on labor market outcomes (cf. Chiswick and Miller 1995; Edin et al. 2003; Delander et al. 2005; Aldén and Hammarstedt 2014). These effects may be exacerbated by, for example, a lack of knowledge and information in segregated neighborhoods (cf. O’Regan and Quigley 1998) or lower educational aspirations among parents in these neighborhoods combined with school segregation (cf. Mookherjee et al. 2010; Böhlmark et al. 2016). Borjas (1995) argues that individuals brought up in more-advantageous environments have higher productivity due to exposure to beneficial social and economic factors, including role models (Ginther et al. 2000). Hence, observing and interacting with neighbors gives rise to externality effects that affect human capital accumulation (Glaeser et al. 2000) as well as behavior (cf. Debreu 1952; Schelling 1973; Becker 1974).

Earlier empirical studies mostly find a negative relationship between segregation and labor market outcomes (Leonard 1987: Cutler and Glaeser 1997; Collins and Margo 2000). However, there is a large self-organizing aspect to residential segregation, where individual choices and preferences play an important role in housing decisions (see, e.g., Schelling 1971). Some empirical studies that address this endogeneity find that the effect of segregation on labor market outcomes is still negative (Cutler and Glaeser 1997), while others find a nonnegative effect (Edin et al. 2003; Cutler et al. 2008; Andersson et al. 2014). This can be explained by the potential benefits of segregation, which are related to social interactions. The residential concentration of certain groups, in terms of, for example, age, education, race, and family status, enhances the formation of group-specific social networks, which increases access to job opportunities through, for example, job referrals (cf. Rees 1966; Bayer et al. 2008: Calvó-Armengol and Jackson 2004). In addition, residential concentration may be necessary to achieve critical mass to support firms and other institutions targeting group-specific needs (Waldfogel 2003). Andersson et al. (2014), who apply similar Swedish micro-data, as well as method, to the present study,⁴ find that the effect of immigrant concentration on immigrant income is positive. The authors argue that this could be due to the Swedish welfare system and redistributive policies.

Zenou (2007) shows that individuals who live in segregated areas rely on strong social ties, that is, family and friends in their immediate neighborhood, for information. If these strong ties are unemployed, the chances of finding a job through the social network are also limited (cf. Hensvik and Nordström Skans 2016). Indeed, Granovetter (1973, 1974, 1983, 1995) argues that strong ties are inferior to weak ties in providing employment opportunities. This argument is supported empirically by Alesina et al. (2015), who show that a background in strong family societies has a negative effect on employment. Additionally, Clark (2003) argues that the impact of unemployment on utility is reduced if a higher proportion of the neighbors are unemployed. This implies that segregated areas with lower socioeconomic status may have a social norm of unemployment, which decreases the incentives to actively seek a job. Social norms may thus be self-enforcing at both the individual and the societal level (cf. Dal Bó and Terviö 2013).

Individuals supply labor, and firms demand labor on the labor market. Both the demand side and the supply side of individual employment can be modeled using discrete choice theory. An individual will actively search for a job (supply labor) only if the expected utility from working exceeds the expected utility from not working. Based on neoclassical economic theory, individual \(i\)’s utility, \(U\), is a function of consumed goods, \(z\), and leisure time, \(F\), which are both subject to constraints. First, the individual cannot spend more than the total income, \(I\), which, if the individual is working, is equal to non-labor income, \(\bar{I}\), plus the hours spent at work times the wage, wh, minus the costs of getting to and from the place of work, \(C\). Second, the individual has a time constraint, where her total time, \(T\), is allocated between work, \(h\), and leisure. If individual \(i\) is not working, there is no time constraint and the total income is the sum of non-labor income and welfare allowances, \(S\).

Neighborhood externalities, \(E(k)\), enter the utility function as a separate factor. Following the discussion above of unemployment as a social norm, residing in a neighborhood with lower socioeconomic status decreases the relative utility from being employed, while it increases the relative utility from not being employed. As opposed to the bid-rent models above, \(E(k)\) is experienced by individuals from both groups of the population (\(A\) and \(B\)). Hence, with \(E(k)\) defined as above,⁵\(U^{\text{work}}\) is an increasing function of \(E(k)\), while \(U^{\text{non-work}}\) is a decreasing function of \(E(k)\). The supply side can thus be specified as follows (cf. Gravelle and Rees 2004):On the demand side, a firm will only employ an individual if the expected value of her marginal product, \(E[{\text{MP}}_{L} ]\), is at least as high as the wage the firm has to pay. An individual’s expected productivity, \(E[{\text{MP}}_{L} ]\), is largely a function of education, ED, and experience, EX (cf. Mincer 1958, 1974). However, working in a primarily Swedish context commonly requires specific skills, such as those pertaining to language, culture, and social norms. The accumulation of country-specific human capital, SE, is thus another factor that affects the expected productivity of an individual. Since the exposure to beneficial social and economic factors in more-prosperous neighborhoods increases the productivity of individuals (Borjas 1995), segregation, \(E(k)\), also enters the function for productivity. \(E[{\text{MP}}_{L} ]\) is thus an increasing function of \(E(k)\). The demand side of employment is specified below:The probability that an individual is employed, \(\Pr (Y = 1)\), is thus a function of a combination of factors from both the demand side and the supply side of the labor market. In addition, the probability of employment may depend on the type of region, R, the individual lives in, since employment opportunities are commonly greater in more-urban regions. This could be due to more-efficient matching on the labor market, which is identified by Duranton and Puga (2004) as a micro-foundation of agglomeration economies. Åslund et al. (2010) show that for refugees arriving to Sweden in 1990–1991, being placed in a location with good access to jobs had a positive effect on the probability of being employed in 1999.

This gives the following theoretical model:Not all of these variables are possible to capture empirically. This specifically concerns the supply side of employment, shown by Eqs. (3)–(5), except for the variable of main interest; segregation.

This study relies on full population geo-coded micro-data from Statistics Sweden. This type of data provides unique possibilities for conducting research with individuals as the units of observation. In addition, the geographical location for each individual is known, which allows for the inclusion of variables describing the surrounding milieu, such as the neighborhood. The neighborhoods defined as church parishes in Sweden are fixed during the whole time period. In general, parishes are more representative of actual neighborhoods in relatively denser areas. The data stretch from 1990 to 2011, which provides a relatively long panel. This is necessary to study neighborhood effects in an individual fixed effects setting, since neighborhood characteristics tend to change slowly over time. This implies that the identification comes mostly, but not solely, from individuals who move between neighborhoods. To test the robustness of the results, we also calculate the neighborhood variables at the scale of Small Areas for Market Statistics (SAMS).⁶ In Sweden, there are approximately 9000 SAMS, which imply that these are, on average, smaller geographical units than parishes.

The data availability allows us to construct a data set in panel form that follows the same individuals over 21 years, 1990–2011.⁷ Section 2.1 shows that on a regional level, there have been relatively large changes in segregation levels since 1990, which implies that there have been redistributions of individuals with native and foreign background on the neighborhood level. This is supported by Table 5, which shows that even though the between variation for foreign–native difference is greater, there is considerable within variation. Hence, even for individuals who do not move, there is variation over time in the neighborhood characteristics.

To include only those individuals who were potentially in the labor force for the whole period, we exclude those who in 1990 were younger than 20 or older than 40 years of age. We cannot know whether all included individuals were continuously in the labor force, since the data do not provide information on labor market participation of non-employed, that is, whether they are active job-seekers or not. This implies that we can measure the non-employment rate but not the unemployment rate. To follow the same individuals during the whole period, we also exclude those who for some reason disappear from the data set during the time period.⁸ In total, we follow 2,254,375⁹ individuals, of which approximately 16% have a foreign background. The data set does not allow for distinguishing among different types of foreign background, since it is not possible to trace individuals to their parents. Country of origin, grouped in regions, is thus only known for foreign-born individuals. Additionally, the data do not allow for distinguishing between labor migrants and, for example, refugee migrants.

In the estimations, we apply individual fixed effects to control for unobserved time-invariant individual characteristics, such as innate abilities, social skills, childhood experiences, and cultural background. These characteristics may influence both where a person chooses to locate and whether that person is employed, which implies a correlation between explanatory variables and the error term. If not controlled for, this endogeneity induces biased and inconsistent estimates. Applying individual fixed effects thus reduces issues of self-selection and omitted variable bias (cf. Weinberg et al. 2004; Andersson et al. 2014). This is crucial in studies of neighborhood effects due to the large degree of endogeneity in the location decisions of individuals. Another endogeneity issue arises from the collection of the data, namely that employment status is being reported in November, while location is based on place of residence by December 31st. This implies that an individual may have moved during the year, which infers a change in the neighborhood characteristics due to being employed. The reverse causality is addressed by measuring the neighborhood characteristics with a 1-year lag. However, the issue of reverse causality may not be fully eliminated since individuals can move due to knowing that they will start a new employment the next year. As robustness test, we introduce 2-year lags for the neighborhood variables.

Even though the measures described above may not completely eliminate endogeneity, they reduce it to a large extent. Both (time-invariant) unobserved individual heterogeneity and reverse causality are controlled and tested for in the estimations and by the robustness tests. One issue that does remain is the potential influence on employment status of time-variant unobserved individual characteristics, since these cannot be captured by the fixed effects estimation. This may affect, for example, the self-selection of movers, that is, that individuals with poor employment prospects are more likely to move to distressed neighborhoods that offer cheap housing, while employed high income individuals have greater possibilities to move to economically prosperous neighborhoods. This type of self-selection may induce biased results if individuals’ employment prospects are changing over time due to either unobservable learning or changes in their underlying abilities.

The dependent variable, \(y_{it}\), takes the value one if individual \(i\) is employed at time \(t\), and zero otherwise. The probability that \(y_{it} = 1\) is given by some function of the covariates \(x_{it}\), where \(x_{it}\) is theoretically defined in Sect. 2.3, while the empirically measured variables are presented in Sect. 3.3.

With a binary dependent variable, estimating logit or probit models may seem as the natural choice. However, as pointed out by Angrist and Pischke (2009), the nonlinearity of these models provides complications when working with panel data. Given that ordinary least squares (OLS) estimates and probit marginal effects have been shown to be empirically very similar (Angrist and Pischke 2009), we estimate linear probability models instead of the alternative nonlinear probit or logit models.¹⁰

The estimated linear probability model, in a panel setting using two-way fixed effects, can be specified as follows:where \(\alpha_{i}\) captures the individual fixed effects, which controls for time-invariant unobservable individual heterogeneity. In fixed effects estimation, the \(\alpha_{i}\)s are eliminated due to the within transformation. \(\gamma_{t}\) represents year fixed effects, which are introduced to control for business cycles, that is, that the probability of being employed is generally higher (lower) in times of expansion (recession). \(u_{it}\) is the usual error term.

To test for differences between males and females with native and foreign backgrounds, we introduce interaction terms between neighborhood segregation and gender/background. In addition, since the relationship between segregation and employment may vary across different types of regions, we run category-wise models for individuals residing in metropolitan, city, and rural regions, respectively.¹¹

The covariates \(x_{it}\) consist of variables at three different levels: individual, neighborhood, and regional. The variables of main interest concern neighborhood characteristics, while individual and regional characteristics control for factors that may affect the probability that an individual is employed. A complete list of variables is given in Table 1. The motivations for these variables are found in Sects. 2.2 and 2.3 as well as in the following subsections. All variables are constructed based on the full population micro-data from Statistics Sweden. Descriptive statistics are presented in Table 5, while Table 6 provides a correlation matrix.

As already mentioned, neighborhoods in Sweden, as in many other countries, that hold a large proportion of individuals with foreign background are characterized by lower socioeconomic status. Hence, we re-estimate the relationship between neighborhood segregation and individual employment, including the neighborhood non-employment rate as a measure of socioeconomic status. This shows whether it is the spatial separation of individuals with native and foreign background that is negatively associated with employment or whether it may be the distress of neighborhoods. The results for the neighborhood characteristics are presented in Table 4.¹⁶

In addition to the results shown in Table 4, we have tested whether the effect of socioeconomic status on individual employment is stronger for more-segregated neighborhoods, by use of an interaction term. The interaction between the neighborhood non-employment rate and foreign–native difference is, however, insignificant throughout the specifications in Table 4.

Table 4 shows that controlling for the socioeconomic status of the neighborhood changes the relationship between residential segregation per se and employment quite drastically. For females of foreign background as well as for natives, the relationship turns positive, or at least nonnegative,¹⁷ which is in line with the results by Andersson et al. (2014), who show a positive effect of immigrant concentration on incomes, using a similar fixed effects approach on Swedish micro-data. The positive effect may be a result of social networks, which can be argued to form more easily among individuals in segregated neighborhoods (cf. Zenou 2007). However, since the added estimate for males of foreign background points to an insignificant or even negative relationship between segregation and employment, the explanation provided by Andersson et al. (2014) may be more likely. Hence, the positive effect may reflect an added effort and effectiveness of Swedish labor market institutions in segregated neighborhoods, which individuals of native background are better equipped to take advantage of. The estimations employing SAMS as well as 2-year lags for the neighborhood variables strengthen this conclusion.

The negative association between segregation and employment found in Table 3 correlates with the lower non-employment rate that commonly prevails in distressed and segregated neighborhoods. However, since the interaction term between non-employment and foreign–native difference is insignificant, the negative effect of the non-employment rate on individual employment is not stronger in more ethnically segregated neighborhoods.

The results point to that it is not segregation in terms of spatial separation between individuals of native and foreign background that causes unemployment. It is rather the lower socioeconomic status, or distress of neighborhoods, here measured as the share of non-employed. In Sweden, one typical aspect of the ethnic segregation process is that individuals with foreign background are more and more agglomerated in distressed neighborhoods, characterized by inexpensive housing, low employment levels, low incomes, and low education levels. The bivariate correlation between the foreign–native difference and the non-employment rate at the neighborhood level has increased continuously between 1990 and 2011, from 0.13 in 1990 to 0.38 in 2011,¹⁸ which indicates that labor market effects tied to the process of segregation have been accelerating during the past decades. Hence, one significant challenge with segregation is perhaps not the spatial separation but rather that segregation relegates individuals who already face challenges in the labor market to areas with fewer opportunities for employment (cf. Kain 1968; Massey et al. 1987). This increases the distress and thus decreases the opportunities in these neighborhoods even further, which indicate that spatial and socioeconomic separation go hand in hand.

The lower employment rates in segregated areas may have significant implications for individual networks. Even though social networks may form more easily in segregated neighborhoods, due to individuals sharing similar backgrounds, these networks may not be very efficient from a labor market perspective. In addition, if the individuals in segregated neighborhoods interact mainly within strong ties networks, as found by Zenou (2007), the opportunities for finding a job are limited (cf. Granovetter 1973, 1974, 1983, 1995). This would mean that the potential benefits of segregation, such as group-specific networks, may not be as valid in the Swedish context. As already discussed, distressed neighborhoods in Sweden, as well as in other countries, are characterized by a relatively large immigrant population. However, segregated neighborhoods in Sweden are commonly composed of a wide variety of ethnic groups, with little or nothing in common besides being immigrants (Andersson and Bråmå 2004). This provides less breeding ground for the creation of group-specific social networks.

As opposed to the results on foreign–native difference, the main relationship between neighborhood non-employment and individual employment can now be found for individuals living in rural regions. The status of the neighborhood is thus less relevant for individual employment in metropolitan and city regions, which can be explained by the fact that larger labor market regions offer more opportunities for employment than smaller labor market regions (cf. Duranton and Puga 2004). The estimations using 2-year lags as well as the smaller SAMS to approximate neighborhoods show very similar results to Table 4, albeit with somewhat smaller relationships between the non-employment rate and individual employment.

The results for the control variables show that education is a key factor for employment. For example, having a master’s degree is related to a 0.5, or 50% points, higher level of the likelihood to be employed. However, the base category in the estimations is individuals with 9 years of compulsory schooling or less, which is a group with low employment rates that faces very large challenges in the Swedish labor market due to the absence of low-skilled jobs and high entry wages. Hence, the results confirm the importance of higher education for labor market outcomes (cf. Mincer 1958, 1974).

The results regarding civil status show that being married, and especially having children, is positively related to employment [except from only married in specification (1)]. As expected, experience, measured as age, has a nonlinear effect on employment. Table 2 shows that the likelihood of being employed increases with age but at a decreasing rate. Also years spent in Sweden is significantly related to employment of foreign-born individuals, which confirms previous research (Rooth 1999; Nekby 2003). Compared to native Swedes (base category), foreign born who has been in Sweden for up to 10 years is less likely to be employed, while foreign born with more than 15 years in the country is even more likely to be employed than natives.

Both control variables at the regional level are highly statistically significant. Population density is positively related to employment and an increase in the employment rate by 10% points is related to a six-percentage point increase in the likelihood to be employed. This reflects that employment opportunities are greater in regions with a high degree of urbanization as well as regions with a well-functioning labor market, which may be due to more-efficient matching when there is a greater variety in both employees and employers (Duranton and Puga 2004).