Residential Segregation and Unemployment: The Case of Brussels

Transcription

1 45(1) , January 2008 Residential Segregation and Unemployment: The Case of Brussels Claire Dujardin, Harris Selod and Isabelle Thomas [Paper first received, February 2003; in final form, March 2007] Abstract This paper investigates the causal effects of the spatial organisation of Brussels on unemployment propensities. Using census data at the individual level, the unemployment probability of young adults is estimated while taking into account personal, household and neighbourhood characteristics. The endogeneity of residential locations is solved by restricting the sample to young adults residing with their parents; the potential remaining bias is evaluated by conducting a sensitivity analysis. The results suggest that the neighbourhood of residence significantly increases a youngster s probability of being unemployed, a result which is quite robust to the presence of both observed and unobserved parental covariates. 1. Introduction For decades, sociologists, economists and geographers have written extensively on how the spatial structure of cities reflects socioeconomic differences in the population (see the seminal contributions of Burgess, 1925; Hoyt, 1939; and Harris and Ullman, 1945; and their respective descriptions or models of urban stratification). Strikingly, most cities today are characterised by stark disparities opposing city centres and peripheries. In the US for instance, inner cities are usually poor and the catalyst of many social problems, whereas suburbs are more well-off. Brussels exhibits a similar spatial structure since its inner city concentrates unemployed workers and disadvantaged communities, including many unskilled workers and ethnic minorities (see for example, Vandermotten et al., 1999; Thomas and Zenou, 1999; Goffette- Nagot et al., 2000; Kesteloot et al., 2001). Claire Dujardin and Isabelle Thomas are in the Department of Geography and CORE (Center for Operations Research and Econometrics) and at FNRS. CORE, Université catholique de Louvain, 34 Voie du Roman Pays, 1348 Louvain-la-Neuve, Belgium. Fax: s: claire.dujardin@uclouvain.be and isabelle.thomas@uclouvain.be. Harris Selod is in the Paris School of Economics (Laboratoire d Economie Appliquée UR 1043, INRA F Paris) and at CREST and CEPR, London. Paris School of Economics, 48 boulevard Jourdan, Paris, France. Fax: selod@ens.fr Print/ X Online 2008 Urban Studies Journal Limited DOI: /

2 90 CLAIRE DUJARDIN ET AL. To explain the general pattern of residential segregation in urban areas, the economics literature has often stressed the role of location choices through which individuals spontaneously sort themselves into a city according to their different socioeconomic characteristics. In this respect, residential segregation is a standard result in both public and urban economics (see Tiebout, 1956, for whom individuals vote with their feet for the provision of local public goods; or Fujita, 1989, for a formalisation of the standard monocentric model with heterogeneous agents). However, urban stratification should not just be considered as a process whereby households with different socioeconomic backgrounds sort themselves in the city. In fact, an abundant literature has investigated the reverse causality, arguing that the spatial organisation of cities could explain differences in social and economic outcomes. In the US context, it has been argued that the high level of unemployment among inner-city minorities could be explained by residential segregation and/or disconnection from job opportunities (see Kain, 1968; Kasarda, 1989; Jencks and Mayer, 1990; Wilson, 1987 and 1996; Cutler and Glaeser, 1997). Even though the theories which link labour market outcomes to the spatial organisation of cities have mainly inspired empirical papers on US metropolitan areas, the mechanisms put forward clearly have a general validity. Yet only a few works have focused on European cities (see for instance Fieldhouse, 1999, on London; and Gobillon and Selod, 2007, on Paris). The objective of the present paper is to test whether city structure can be a source of unemployment in the Brussels metropolitan area. Despite the huge interest on the topic, the bulk of empirical studies still has not reached a consensus regarding the role of spatial factors in explaining individual labour market outcomes. As argued by Ginther et al. (2000) and Dietz (2002), this lack of consensus can probably be explained by the great diversity regarding the methods used to test for the existence of spatial factors. In particular, because individuals sort themselves into different parts of urban space on the basis of their personal characteristics (including their labour market outcomes), studies encounter an endogeneity bias, as an individual s choice of residential location both influences and is influenced by its labour market outcomes. Moreover, it is likely that some individual and household characteristics which are unobserved to the researcher influence both labour market outcomes and residential location choice. Therefore, one might wrongly attribute to residential location the effect of these unobserved characteristics on labour market outcomes. Unfortunately, no perfect solution for this problem exists at present (Glaeser, 1996). Among the few empirical studies that try to deal with the endogeneity issue, most choose to restrict their sample to young adults residing with their parents, arguing that the location choice of parents can be thought of as fairly exogenous to the employment status of young adults (see for example, O Regan and Quigley, 1996, 1998). However, it has been argued that this solution does not completely eliminate endogeneity as there may exist unobserved parental characteristics that influence both the residential choice and the employment status of young adults (Glaeser, 1996). In the present paper, we nevertheless choose to focus on young adults living with their parents. However, in order to get better confidence in our results, we resort to a strategy inspired by Ginther et al. (2000) and Harding (2003) which enables us to analyse the sensitivity of our estimated spatial effects to the presence of both observed and unobserved parental covariates. The paper is structured as follows. Section 2 presents a brief synthesis of the economic literature that links the formation of unem-

3 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 91 ployment to city structure. Section 3 describes the database and presents the studied area. Section 4 describes our methodological approach. Section 5 presents some stylised facts about the Brussels metropolitan area. Section 6 presents and discusses the main results. Section 7 concludes. 2. Urban Unemployment and City Structure: A Brief Review of the Literature Over the past 30 years, unemployment has become a mass phenomenon in most industrialised countries and the source of poverty in urban areas. Economists have put forward a variety of mechanisms to explain and analyse the problem (Snower, 1994). A first line of analysis argues that unemployment is caused by insufficient growth and the weakness of demand, while a second school of thought considers that unemployment is caused by institutional distortions such as the existence of a minimum wage that prevents labour markets from clearing and causes unemployment among low-skilled workers. A third line of analysis blames the frictions that prevent the reallocation of the labour force from declining traditional industries towards growing ones. This argument is strengthened in a context of economic globalisation (which accelerates the decline of non-competitive industries and fosters delocalisation) and because of skill-biased technological change detrimental to low qualifications. Finally, a series of arguments point at a variety of labour market imperfections ranging from job-search inefficiency to non-competitive behaviours such as labour market discrimination against particular population groups (women, immigrants, senior workers, etc.). Interestingly, these analyses are often carried out in a non-spatial framework: the spatial concentration of unemployment is usually viewed only as a consequence of harmful industrial specialisation in a region or of the spatial sorting of unemployed workers themselves. The issue of how space or location might itself cause unemployment has often been over-looked by labour economists. An abundant literature in sociology and urban economics, however, suggests that the spatial organisation of cities can exacerbate unemployment among disadvantaged communities. In this perspective, labour market outcomes should depend on individual characteristics (age, education, ethnicity, etc.) and also on the exact location within the city. Two types of factor have been put forward to explain the existence of such spatial effects, focusing either on the role played by the physical disconnection from jobs or on the harmful effects of residential segregation and the social composition of neighbourhoods. The first studies purporting to demonstrate the influence of space on individual labour market outcomes were based on the spatial mismatch hypothesis put forward by Kain in 1968 (see Gobillon et al., 2007, for a mainly theoretical survey of the literature; and Ihlanfeldt and Sjoquist, 1998, for an empirical one). In theory, there are several mechanisms according to which distance to job opportunities can be problematic. One important mechanism is that job-seekers residing in areas disconnected from job opportunities are likely to reject job offers if commuting costs are too high in view of the offered wages (Brueckner and Martin, 1997). Another mechanism is that workers jobsearch efficiency may decrease with distance to jobs since it is obviously more difficult to search far away from one s place of residence (as modelled in Wasmer and Zenou, 2002). The intensity of the search effort may also decrease with distance from job opportunities for instance, if workers residing far away from job centres face lower rents and thus feel less pressured to find a job quickly (as in Smith and Zenou, 2003). Similarly, high search costs may also deter workers from searching

4 92 CLAIRE DUJARDIN ET AL. far away (as in Ortega, 2000). Finally, firms could discriminate against distant workers for instance, if distance makes them less productive because of long and tiring commutes (as in Zenou, 2002). General empirical tests confirm that the disconnection between places of residence and job locations exacerbates unemployment (see Weinberg, 2004, and Martin, 2004), but few of the detailed mechanisms have been specifically tested (see Gobillon et al., 2007, for more details on this issue). Other works have focused on the role of residential segregation and more generally on the quality of the social environment on individual socioeconomic outcomes. In this respect, several mechanisms can account for an adverse effect of residential segregation, either directly on unemployment, or indirectly through low employability. One mechanism is that residential segregation can be a hindrance to human capital acquisition: in neighbourhoods which concentrate lowability students, human capital externalities can lower school achievements and employability (Bénabou, 1993). Social problems which reduce the employability of workers can also spread through neighbourhood interactions. For instance, Crane (1991) develops an epidemic theory of ghettos in which the propensity of youngsters to adopt a socially deviant behaviour depends on the proportion of same-behaviour individuals in the neighbourhood. This contagion is all the more prevalent as adults are themselves unemployed and do not provide a role model with which youngsters can identify (Wilson, 1987). Another mechanism whereby residential segregation can exacerbate unemployment is that it can weaken social networks in disadvantaged communities. This is a crucial point since a significant proportion of jobs is usually found through personal contacts (Mortensen and Vishwanath, 1994) and since low-skilled workers, young adults and ethnic minorities often resort to such informal search methods (Holzer, 1987 and 1988). In particular, in neighbourhoods where local unemployment rates are higher than average, local residents know fewer employed workers who could refer them to their own employer or provide them with professional contacts. In this respect, Reingold (1999) concludes that the poor quality of social networks explains a significant portion of unemployment problems in disadvantaged urban areas in the US (see Selod and Zenou, 2001 and 2006, for formalisations). Another mechanism which links labour market outcomes to segregation involves the reluctance of employers to hire workers residing in disadvantaged neighbourhoods. The stigmatisation of these neighbourhoods is at the root of redlining, a practice in which employers draw an imaginary red line around a stigmatised neighbourhood and beyond which they discriminate against residents (see Zenou and Boccard, 2000, for a model). General tests show that segregation has an adverse effect on the labour market (see Cutler and Glaeser, 1997), but distinguishing which particular mechanisms explain this adverse effect remains on the research agenda. 3. Data and Studied Area 3.1 Studied Area This paper focuses on Brussels which, like Flanders and Wallonia, forms one of the three institutional regions of Belgium: the Brussels Capital Region, which consists of 19 municipalities (called communes in French) and has around 1 million inhabitants in an area of 163 square km. However, as with most cities, the Brussels functional metropolitan area extends far beyond its institutional limits. Several studies have tried to measure the spatial extent of its urban area, applying various methods (see Thomas et al., 2000, for a synthesis). In the present paper, we use the so-called Extended Urban Area (EUA, see Figure 1), which perfectly reflects the social

5 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 93 dualism between the city centre of Brussels and its close suburbs. 1 The EUA has 1.4 million inhabitants and extends over an area of 723 square km. The smallest spatial unit for which census data are usually available is the statistical ward, a sub-division of the municipality defined in 1971 according to social, economic and architectural similarities (Brulard and Van der Hagen, 1972). Statistical wards that present a common functional or structural character (for example, a common attraction pole like a school or a church) can further be grouped into larger entities which we will refer to as neighbourhoods; these constitute an intermediate level between the statistical ward and the municipality. It is the spatial level of analysis chosen in this paper. There Figure 1. The Brussels Extended Urban Area

6 94 CLAIRE DUJARDIN ET AL. are 328 such neighbourhoods in the EUA, grouping on average 4250 inhabitants. For statistical reasons, neighbourhoods of less than 200 inhabitants were not considered in the subsequent analyses. 3.2 The Data The empirical analysis in this paper is based on two datasets extracted from the 1991 Census of Population carried out by the Belgian National Institute of Statistics (in French, Institut National de Statistiques, INS hereafter). The two datasets differ with respect to the level of aggregation. In the first dataset, the basic statistical unit is the individual. For all individuals aged and residing in the Extended Urban Area, the dataset provides the main personal characteristics, including age, gender, education, citizenship, employment status, statistical ward of residence and kinship with household s head. The dataset also contains several household characteristics (for instance, whether the household owns a car) as well as an identification number, which makes it possible to identify individuals who belong to the same household. In the second dataset, the basic statistical unit is the statistical ward. The dataset includes various indicators of the socioeconomic composition and average housing characteristics of the statistical wards. This dataset was complemented with the average income of households in the statistical wards computed by INS from 1993 fiscal sources. These ward-level data were further aggregated to form neighbourhood-level variables. 4. Methodological Approach The objective of the present paper is to investigate whether the spatial structure of Brussels may have an effect on unemployment, controlling for the standard individual determinants of unemployment. 2 To this end, we estimate unemployment probabilities at the individual level, taking into account personal, household and neighbourhood characteristics, using the following logistic model P i Log = α + βii + γhi + δ Ni 1 P (1) i where, P i is the unemployment probability of individual i; I i is a vector of personal characteristics; H i is a vector of household characteristics; N i is a vector of neighbourhood characteristics (social composition and physical access to jobs); and α, β, γ and δ are vectors of parameters that will be estimated using maximum likelihood estimation (MLE). In particular, δ, when significantly different from zero, identifies an impact of spatial factors on unemployment, which we will refer to as neighbourhood effects in the rest of the paper. Using (1), the individual probability of unemployment P i is given by ( α + βii + γhi + δni) e Pi = ( α + βii + γhi + δni) 1+ e 4.1 Definition of Neighbourhood Characteristics (2) Two types of indicator are used to account for the neighbourhood characteristics which may potentially influence individual unemployment probabilities. The first type relates to the spatial mismatch theory and characterises the disconnection of neighbourhoods from jobs, while the second type relates to the social composition of neighbourhoods. Measuring spatial access to job opportunities is a difficult task as data on job offers are generally not available. In the absence of data on job openings, we use the individual declarations of workplaces in the census to compute the number of occupied jobs in each neighbourhood. 3 Two indicators are

7 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 95 then computed for each neighbourhood n: distance to jobs and job density. First, distance to jobs D n is defined as the average distance from neighbourhood n to each neighbourhood m in the EUA (d nm ) weighted by the number of jobs located in each one of these neighbourhoods (E m ) D n = d m m nm E E m m (3) with the interneighbourhood distance d nm defined as the Euclidian distance between the neighbourhoods centroids and the intraneighbourhood distance d nn being equal to two-thirds of the radius of a disc of an area equivalent to the area of neighbourhood n (which comes down to assuming that the population in each neighbourhood is uniformly distributed around a central point which concentrates all jobs). Secondly, we consider that the relevant job density for residents of neighbourhood n is the ratio of the number of jobs located in that neighbourhood and in the adjacent neighbourhoods to the overall labour force residing in the same areas. This definition has the advantage of smoothing job density over space and attenuating extreme values. These two indicators were computed both for all jobs and for low-skilled jobs (i.e. jobs occupied by workers having at most a diploma from the junior secondary education segment equivalent to O level in the UK system). Regarding the variables which characterise the social composition of neighbourhoods, researchers generally resort to one or several quantitative measures of the aggregate characteristics of residents. However, even though it is likely that individual outcomes are determined by a wide variety of neighbourhood characteristics (employment, education, racial composition, etc.), considering all these characteristics together in a single regression may cause collinearity problems (making parameter values and significance levels unstable) as many indicators of neighbourhood composition are highly intercorrelated (O Regan and Quigley, 1996; Johnston et al., 2004). To circumvent this problem, we use standard factorial ecology methods (see for example, Johnston, 1978) to summarise these multiple characteristics into different types of social environment within Brussels (see section 5.2 for a detailed presentation of the typology). 4.2 The Endogeneity of Residential Location Linking individual labour market outcomes to residential location raises the issue of the endogeneity of location choices (Glaeser, 1996; Dietz, 2002). It is well known that individuals with similar socioeconomic characteristics notably, similar labour market outcomes tend to sort themselves into certain areas of the urban space. For instance, individuals with well-paid jobs will choose to reside in neighbourhoods with a good social environment. There is thus a two-way causality: on the one hand, residential location influences individual labour market outcomes; and, on the other hand, individual outcomes influence the choice of a residential location. Stated differently, it is possible that individual characteristics that influence labour market outcomes may also influence residential choices. Of course, standard models like equation (1) make it possible to control for some individual and household characteristics which might influence both neighbourhood choice and individual outcomes. However, it is likely that some individual and household characteristics which are unobserved to the researcher (and therefore not included in I i or H i ) influence both the outcome of interest and neighbourhood choice. For example, individuals with a low labour market attachment (which directly influences the probability of unemployment)

8 96 CLAIRE DUJARDIN ET AL. may choose to reside in poor neighbourhoods for some economic or social reason. As a consequence, what the researcher perceives as a neighbourhood effect through the estimated parameter δ may simply stem from a correlated effect reflecting common residential choice. Various strategies have been developed to correct for the endogeneity of neighbourhood choice. For example, the existence of quasiexperimental situations (such as governmentsubsidised relocation programmes in the US like the Gautreaux or the Moving To Opportunity programmes) should make it possible to obtain more reliable estimates of neighbourhood effects (to the extent that households are moved from one neighbourhood to another through an exogenous intervention; see Oreopoulos, 2003, for a review). However, due to the scarcity of such experiments, researchers are often constrained to resort to more questionable strategies. Most studies restrict their sample to young adults residing with their parents, arguing that location choices were previously made by the parents and can thus be thought of as fairly exogenous to the employment status of young adults (see for example, O Regan and Quigley, 1996, 1998). This is also the strategy adopted in this paper, as our studied sample is restricted to young labour force participants aged and residing with their parents. However, this approach does not completely eliminate the endogeneity bias. Indeed, there may still exist parental unobserved characteristics which determine their residential choice and also influence the employment outcomes of their adult children (Glaeser, 1996). For example, lack of commitment to work may induce parents to locate in highpoverty neighbourhoods but may also influence the motivation of their children to search for a job as well as the intensity of their job-search effort. In this context, one cannot distinguish neighbourhood effects from the effect of those unobserved parental characteristics when estimating the unemployment probability of young adults living with their parents. 4.3 Sensitivity Analysis Following a suggestion made by Glaeser (1996, p. 62), we evaluate the potential remaining endogeneity bias by conducting a sensitivity analysis in order to assess the robustness of our estimated neighbourhood effects. We resort to a two-step strategy. In a first step, we follow the approach used by Ginther et al. (2000) and estimate several models of individual unemployment probability which incorporate various sets of household characteristics H i, moving away from a model with no household variable towards a model including an extensive set of parental controls (parental employment and professional status, parental education, male or female household head, possession of an automobile). The comparison of the estimated neighbourhood effects in these models enables us to test the robustness of our results to the presence of observed parental covariates. In a second step, we test the sensitivity of our results to the endogeneity bias which results from the omission of an unobserved covariate which is correlated to both the probability of unemployment among young adults and parental residential choice, following the approach developed by Rosenbaum and Rubin (1983) and recently applied by Harding (2003) in the context of neighbourhood effects. In order to carry out the sensitivity analysis, we build a dummy variable which equals one if the individual resides in a deprived neighbourhood and zero otherwise (see section 6.3 for the exact definition of deprived neighbourhoods). The goal of our analysis is to assess how an unobserved binary covariate which affects both the probability of unemployment among young adults and the choice of parents to reside in a deprived or

9 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 97 a non-deprived neighbourhood would alter our conclusions about the magnitude and significance of neighbourhood effects. This is done by generating a series of unobserved binary variables U that vary according to their degree of association with the neighbourhood dummy variable X and with the binary outcome measure Y. The degrees of association between U and X and between U and Y are classically measured by odds ratios. In practice, we use the Sensuc function in the Design S library written by Harrell (2003) to generate a binary variable U sampled according to the following logistic model Ui Log 1 U i = α + log( by ) i + log( c) X (4) i where, Y i is a binary variable indicating whether individual i is unemployed or not; X i is a binary variable indicating whether i resides in a deprived neighbourhood or not; and b and c are chosen odds ratios measuring the strength of the association that we impose between U and Y and U and X respectively. 4 The value of α is determined so that the overall prevalence of U = 1 is 0.5 (or any other given value). More precisely, the previous equation and the three imposed constraints (the values of the two sensitivity parameters b and c and the overall prevalence of U = 1) are used to determine the proportion of U = 1 in each of the four sub-groups defined by X and Y, i.e. P(U = 1 X = 1, Y = 1), P(U = 1 X = 1, Y = 0), P(U = 1 X = 0, Y = 1), and P(U = 1 X = 0, Y = 0). Then, a random variable U is generated in each sub-group defined by X and Y following a Bernoulli distribution of parameter P(U = 1 X = 1, Y = 1), etc. Once each individual has received a value for U, new estimates of neighbourhood effects are obtained by including the U variable in the unemployment probability model (1). This is repeated for increasing values of the sensitivity parameters b and c in order to investigate what level of endogeneity bias (i.e. the strength of the association between U and Y and between U and X) would be needed to invalidate the results and render our estimated neighbourhood effects not significant. 5. The Spatial Structure of Brussels 5.1 Stylised Facts The Brussels Extended Urban Area presents a well-marked spatial structure characterised by important disparities opposing its city centre to the periphery. Figure 2 shows the percentage of unemployed workers among labour force participants, highlighting a zone of very high unemployment rates (above 20 per cent) in the central part of the urban area, along the former industrial corridor (the Charleroi Willebroek canal). By contrast, unemployment is much lower (below 10 per cent or even 7 per cent) in the suburbs (notably in the municipalities of Tervuren, Overijse, Grimbergen, Dilbeek and Sint-Peeters-Leeuw). Figure 3 shows the percentage of North Africans (Moroccans, Algerians and Tunisians) and Turks among the whole population. These nationalities correspond to the latest wave of labour immigration in the second half of the 1960s (Kesteloot and Cortie, 1998). We consider them as a single group since they usually face the same type of problems in the labour market. 5 Figure 3 shows a high level of ethnic residential segregation: North Africans and Turks are mainly located around the centre of Brussels, with a concentration above 10 per cent or even 25 per cent in some neighbourhoods. Moreover, as can be seen from the computation of dissimilarity indices (Duncan and Duncan, 1955), 6 North Africans and Turks form the most segregated

10 98 CLAIRE DUJARDIN ET AL. Figure 2. Percentage of unemployed workers among labour force participants Source: INS (1991). group of foreigners in Brussels: 64 per cent of them would have to be relocated to another neighbourhood in order to obtain a uniform mix with the Belgian population. For non- Belgian citizens of the European Economic Community (consisting of 12 countries in 1991), the dissimilarity index goes down to only 32 per cent. It is widely acknowledged that the concentration of North Africans and Turks in the central neighbourhoods of Brussels is mainly due to the functioning of the local housing market. Indeed, the promotion of homeownership an important housing policy goal in Belgium led to the quasiabsence of a social housing sector and to the massive suburbanisation of high- and middleincome households towards the periphery. In this context, suburbanisation accelerated the degradation of central zones as housing units were rented to low-income households and maintenance was neglected by owners mainly concerned with the additional profit they could derive from their investments. This gave rise to a residual rental sector which concentrates the oldest, most poorly equipped and cheapest housing units of the city, which are the only dwellings poor foreigners can

11 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 99 Figure 3. Percentage of North Africans and Turks in the total population Source: INS (1991). afford (Kesteloot and Van der Haegen, 1997; Kesteloot and Cortie, 1998). By comparing Figures 2 and 3, it can easily be seen that neighbourhoods with a high proportion of North Africans and Turks also exhibit high unemployment rates, the correlation between the two variables being equal to This high value is partly explained by neighbourhood composition (foreigners are usually more likely to be unemployed because they have a lower education or because they are discriminated against in the labour market). It could also be explained by the exacerbating effect of residential segregation on labour market outcomes, as we hypothesise. Figure 4 shows the density of jobs, for all jobs (on the left-hand side) and for low-skilled jobs (on the right-hand side). It shows a zone of very high job densities in the centre of the EUA, where some neighbourhoods have more than five jobs per labour force participant. These neighbourhoods constitute what is usually called The Pentagon, where one can find the Administrative City as well as many financial institutions and headquarters of national and international firms. This zone of very high job densities is surrounded by a zone

12 100 CLAIRE DUJARDIN ET AL. Figure 4. Density of total jobs and low-skilled jobs Source: INS (1991). of high, although lower, job densities (more than two jobs for one resident), comprising the European institutions and extending towards the north-eastern periphery where the national airport as well as some industries are located. Note that the spatial distribution of job densities is quite similar for all jobs and for low-skilled jobs. However, the very high density zone in the centre (more than five jobs per labour force participant) and the high density zone in the north-eastern periphery are spatially less extended when it comes to low-skilled jobs. By comparing Figures 2 and 4, it can be seen that central zones with high unemployment rates also have relatively high job densities, which seems in contradiction with the spatial mismatch theory. Indeed, the correlation between unemployment rate and total job density is significant and positive (0.35 for all jobs and 0.31 for lowskilled jobs) and the correlation between unemployment rate and average distance to jobs is significant and negative (-0.52 for both all and low-skilled jobs). In other words, in Brussels, individuals living in high job density neighbourhoods close to the city centre are more likely to be unemployed. 5.2 Typology of Neighbourhoods Standard factorial ecology methods are used to identify socially homogeneous areas within Brussels, which will be subsequently used in the regression analyses in section 6. We first run a principal component analysis which

13 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 101 defines a limited number of non-correlated factors summarising the information carried by a set of neighbourhood variables (see Table A1 in the Appendix for the list of variables and their contribution to the retained factors). Then, neighbourhoods are grouped according to their co-ordinates on the factorial axes, using a hierarchical ascending classification (with the Ward method which minimises intragroup variance). We obtained five neighbourhood types 7 which are presented in Figure 5 (see Table A2 in the Appendix for their mean characteristics). They approximately correspond to the fieldwork knowledge that urban planners and geographers have of Brussels (for example, Grimmeau et al., 1994; Vandermotten and Vermoesen, 1995). The first neighbourhood type comprises very deprived areas in the centre of Brussels and corresponds to neighbourhoods with high proportions of North Africans, Turks and female-headed households as well as neighbourhoods with very high unemployment rates. These neighbourhoods are characterised by low educational levels and have the lowest income levels of the whole agglomeration. They are surrounded by a group of deprived neighbourhoods presenting Figure 5. Typology of neighbourhoods Source: authors calculations based on data from INS (1991).

14 102 CLAIRE DUJARDIN ET AL. similar characteristics but with a smaller proportion of North Africans and Turks and a less severe situation in terms of education and unemployment. The third group of neighbourhoods extends to the south-west and north-east of the inner city, along the former industrial axis. It groups neighbourhoods that also have a lower socioeconomic status, in particular an overrepresentation of bluecollar workers and individuals with a lower education. However, the unemployment rates and income levels are closer to the city s average. The two remaining groups are characterised on average by higher levels of education and professional status. The first, which we have labelled well-off, corresponds to the periphery of the Brussels EUA, while the second, which we have labelled very well-off, occupies the south-east part of the Brussels Capital Region as well as its continuation in the periphery (notably the municipalities of Waterloo and Lasne). The latter differs from the former due to its high proportion of executives and its very high average income. 6. The Effect of Spatial Structure on Unemployment: Main Results The previous section has shown that Brussels exhibits a high level of residential segregation associated with high local unemployment rates, but that disadvantaged neighbourhoods are usually close to job locations. However, a statistical analysis is required before one can determine the role of spatial factors on unemployment in Brussels. The aims of the present section are: to investigate the role played by the spatial structure of Brussels in individual unemployment probabilities; and, to test for the robustness of our results. As explained in section 4.2, in order to limit the endogeneity problems associated with residential sorting, the sample is restricted to young labour force participants living with their parents i.e. a sample of individuals. We first consider a model including only individual characteristics, which we comment in section 6.1, and then, in section 6.2, we progressively add neighbourhood characteristics to the baseline specification. Section 6.3 tests the robustness of the results to the presence of both observed and unobserved parental characteristics. 6.1 The Role of Individual Characteristics Table 1 presents the odds ratios for five different models explaining the individual unemployment probability. Model Ia considers only the role of individual characteristics (gender, age, education and citizenship), while the four other models add various combinations of neighbourhood characteristics to the specification. Three levels of education are distinguished: lower for individuals with at most a diploma of junior secondary education (normally corresponding to an age of 15); intermediate for those with a diploma of senior secondary education (normally aged 18); and higher for those with a higher diploma. Concerning citizenship, four main groups are defined: Belgians, foreigners from the European Economic Community (EEC hereafter; 12 countries in 1991), North African and Turkish foreigners, and other foreigners. Furthermore, the nationality of parents is used to approximate the concept of ethnicity, by distinguishing between Belgians of Belgian parents, Belgians of EEC parents, Belgians of North African or Turkish parents, and Belgians of parents of other nationality. Note that individuals for whom important characteristics were missing had to be set aside. Individuals residing in neighbourhoods of less than 200 inhabitants were also omitted from the sample. All in all, this leads to a sample of individuals covering 95 per cent of all labour force participants aged and residing with their parents. In Model Ia and in all other regressions, men or educated workers are less likely to

15 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 103 Table 1. A logistic regression of the unemployment probability: comparison of different neighbourhood variables (N = ) Model Ia Model Ib Model Ic Model Id Model Ie Likelihood ratio Akaike information criterion Neighbourhood variables Unemployment rate 1.027*** Neighbourhood type Very deprived 1.873*** 1.737*** 1.741*** Deprived 1.740*** 1.638*** 1.632*** Industrial axis 0.746*** 0.742*** 0.737*** Well-off Ref. Ref. Ref. Very well-off 1.990*** 1.970*** 1.962*** Distance to jobs NS Job density 1.041*** Distance to low-qualified jobs NS Low-qualified job density 1.071*** Individual characteristics Male 0.914*** 0.921*** 0.925** 0.926** 0.926*** Age 0.911*** 0.912*** 0.902*** 0.901*** 0.901*** Education Lower Ref. Ref. Ref. Ref. Ref. Intermediate 0.600*** 0.618*** 0.635*** 0.637*** 0.637*** Higher 0.523*** 0.554*** 0.529*** *0.530*** 0.530*** Citizenship Belgian (of Belgian parents) Ref. Ref. Ref. Ref. Ref. Belgian (of EEC parents) 1.460*** 1.349*** 1.311*** 1.305*** 1.303*** Belgian (of North-African or 2.263*** 1.665*** 1.810*** 1.785*** 1.783*** Turkish parents) Belgian (of parents of other 3.272*** 2.735*** 2.628*** 2.600*** 2.600*** citizenship) EEC 1.553*** 1.247*** 1.282*** 1.262*** 1.259*** North African and Turkish 2.897*** 2.039*** 2.271*** 2.238*** 2.234*** Other 3.204*** 2.680*** 2.656*** 2.634*** 2.631*** Notes: Figures give the odds ratios. ***significant at the 1 per cent level; **significant at the 5 per cent level; *significant at the 10 per cent level; NS not significant at the 10 per cent level. be unemployed than women or workers with a lower education. The probability of unemployment also decreases with the age of the individual and when the individual is a foreigner. North Africans and Turks are more disadvantaged than EEC citizens. Interestingly, young Belgian adults born of foreign parents are also more likely to be unemployed than young Belgian adults born of Belgian parents. This result suggests that, besides citizenship, the name and/or visible characteristics associated with foreign origin are a handicap in the labour market. This is consistent with the existence of both labour market discrimination and social networks of lower quality for individuals of foreign origin.

16 104 CLAIRE DUJARDIN ET AL. 6.2 The Role of Spatial Factors in Explaining Unemployment Probabilities As can be seen from the goodness-of-fit statistics (the likelihood ratio and the Akaike information criterion) in Table 1, introducing neighbourhood characteristics significantly increases the fit of the regression. Model Ib uses only the unemployment rate to represent neighbourhood influences. It shows that the local unemployment rate significantly increases the unemployment likelihood of young adults. As suggested by the theory, this may hinge upon lower-quality local social networks: the higher the proportion of unemployed neighbours, the more difficult the insertion of a young adult in the labour market. Model Ic substitutes the type of neighbourhood for the unemployment rate, which increases the fit of the model, thus confirming that the unemployment probability is affected by a wide variety of neighbourhood characteristics, instead of only by the unemployment rate. All else being equal, the unemployment probability is the lowest for young adults residing in the industrial axis and in well-off areas. In accordance with the theory, living in a central area with a socioeconomic environment of lower quality (deprived or very deprived) significantly increases the unemployment probability of young adults (with odds ratios of and respectively) in comparison with living in well-off areas. Surprisingly, however, living in a very well-off area also significantly increases the probability of unemployment. However, we do not wish to conclude that living in a wealthy area is detrimental to finding a job. A possible interpretation for this unexpected result could be that unemployed youth in very well-off areas do not feel under pressure to search intensively for a job if they get enough financial support from their parents. Another possible explanation could be the existence of a selection bias if living in a wealthy area enhances the chances of finding a well-paid job, so that young adults originating from these areas are likely to move out of their parents dwelling quickly. This would leave an overrepresentation of unemployed workers with adverse unobserved characteristics living with their parents in very well-off areas. Investigating the veracity of this would require modelling unemployment probabilities jointly with the process of leaving parental home through a selection model. However, as the choice to leave home is probably influenced by parental characteristics (such as household financial resources) as well as neighbourhood characteristics, this would require more specific data. In particular, we would need to be able to identify the household and neighbourhood of origin of young adults who do not live with their parents. This is unfortunately not feasible using the Belgian census data. 8 Model Id complements the specification by adding indicators of the physical disconnection from jobs. As might be suspected from the statistical description of Brussels (section 5.1), the effect of distance to jobs is not significant. As for job density, its impact is significant but works in the wrong direction: controlling for all other variables, young adults who reside in areas with the highest job densities are less likely to hold a job. The same obtains when one considers distance to lowskilled jobs and low-skilled job density. This last result is not consistent with the spatial mismatch hypothesis and suggests that spatial mismatch is not a problem in Brussels, a city in which the unemployed reside close to the jobs they could occupy. 6.3 Sensitivity Analysis In order to test the robustness of our estimated neighbourhood effects, in a first step, we conduct a sensitivity analysis to the presence of parental observed characteristics, estimating different models which incorporate various sets of parental and household characteristics. Parental characteristics were built assigning

17 RESIDENTIAL SEGREGATION AND UNEMPLOYMENT 105 to each young adult the characteristics of the household head when present in the dataset (i.e. aged 64 or younger). If the household head was not in the dataset or if values were missing, we used the characteristics of the household head s spouse. A total of 4786 cases had to be excluded, either because values were missing for both the household head and his spouse or because none of them was present in the dataset. Table 2 presents the results of our sensitivity analysis, moving away from a model with no household characteristics (Model I) towards models including an increasingly comprehensive set of parental controls (Models II VI). Note that distance to jobs and job density were not included as they played little role in the previous models. Models II VI show that the unemployment probability of a young adult is higher when the household head (or spouse) is not participating in the labour force or is unemployed, than when he is employed. This effect is highly significant and is consistent with social network theories (at the household level, unemployed parents being unable to help their job-seeking children) and socialisation considerations (unemployed parents failing to provide their children with an image of social success with which they could identify). Living in a female-headed household (a proxy for single-mother households) also significantly increases the likelihood that a young adult is unemployed, suggesting that these households are more frequently exposed to social problems detrimental to finding a job. Living in a household which does not own a car (an indirect measure of lack of financial resources and low mobility) also significantly increases the unemployment probability. The effects of parental professional status and educational level are not always significant and, when significant, seem counter-intuitive. Indeed, all other things equal, having a parent with an executive profession or a higher level of education increases the unemployment probability of young adults. This effect mirrors our counter-intuitive finding on residence in very well-off areas. As mentioned previously, this could be explained if rich children are under less pressure to search intensively for a job (because of their parents financial support) or by the presence of a selection bias. With the present dataset, we are unfortunately not able to test nor distinguish between these two explanations. By comparing the parameters and significance levels of the neighbourhood types across the different models, one can assess the sensitivity of our estimated neighbourhood effects to the inclusion of a more comprehensive set of parental controls. Table 2 shows that, although the inclusion of parental and household characteristics significantly increases the fit of the model (see the likelihood ratio and the Akaike information criterion), the estimated neighbourhood effects change very little (for example, the odds ratio associated with residence in very deprived areas varies from to between Model I and Model VI) and all parameters remain significant at the 1 per cent level. In a second step, we complement the sensitivity analysis by testing the robustness of our estimated neighbourhood effects to the presence of unobserved characteristics, using the methodology developed by Rosenbaum and Rubin (1983). Since this methodology requires the use of a binary neighbourhood variable, we redefined our five-modality variable by grouping together the two least-favoured neighbourhood types (very deprived and deprived), while considering all other neighbourhood types as another category. Doing so yields a dummy variable equal to one if the individual resides in a deprived neighbourhood and equal to zero otherwise. Table 3 presents the estimated odds ratios associated with the effect of living in such a deprived neighbourhood on the unemployment probability, for three combinations of parental and household characteristics: no parental characteristics (as in Table 2, Model I);

18 106 CLAIRE DUJARDIN ET AL. Table 2. Sensitivity analysis to the presence of parental and household characteristics (N = ) Model I Model II Model III Model IV Model V Model VI Likelihood ratio Akaike information criterion Neighbourhood variables Neighbourhood type Very deprived 1.903*** 1.817*** 1.844*** 1.863*** 1.819*** 1.707*** Deprived 1.724*** 1.689*** 1.699*** 1.709*** 1.649*** 1.574*** Industrial axis 0.741*** 0.732*** 0.746*** 0.756*** 0.760*** 0.760*** Well-off Ref. Ref. Ref. Ref. Ref. Ref. Very well-off 2.026*** 2.026*** 1.976*** 1.927*** 1.869*** 1.853*** Individual characteristics Male 0.936** 0.941* 0.941* 0.940* NS NS Age 0.898*** 0.893*** 0.891*** 0.889*** 0.890*** 0.893*** Education Lower Ref. Ref. Ref. Ref. Ref. Ref. Intermediate 0.641*** 0.652*** 0.648*** 0.641*** 0.650*** 0.667*** Higher 0.545*** 0.562*** 0.542*** 0.517*** 0.531*** 0.548*** Citizenship Belgian (of Belgian parents) Ref. Ref. Ref. Ref. Ref. Ref. Belgian (of EEC parents) 1.391*** 1.376*** 1.414*** 1.439*** 1.540*** 1.520*** Belgian (of North-African or 1.617*** 1.472*** 1.519*** 1.563*** 1.623*** 1.617*** Turkish par.) Belgian (of parents of other 2.464*** 2.330*** 2.326*** 2.393*** 2.516*** 2.487*** citizenship) EEC 1.297*** 1.227*** 1.268*** 1.305*** 1.367*** 1.361*** North African and Turkish 2.350*** 2.110*** 2.161*** 2.215*** 2.386*** 2.403*** Other 2.493*** 2.248*** 2.293*** 2.393*** 2.426*** 2.348*** Parental and household characteristics Employment status and professional status Not participating in labour force 1.281*** 1.314*** 1.315*** 1.226*** 1.191*** Unemployed 1.545*** 1.570*** 1.577*** 1.521*** 1.459*** Employed Ref. Executive 1.222*** NS 1.118* 1.128* Office worker NS NS NS NS Intermediate profession Ref. Ref. Ref. Ref. Farmer, trader NS NS NS NS Blue-collar 0.893* NS NS NS Education Lower Ref. Ref. Ref. Intermediate 1.087* 1.090* 1.116** Higher 1.330*** 1.343*** 1.390*** Male 0.670*** 0.758*** Possession of an automobile 0.670*** Notes: Figures give the odds ratios. ***significant at the 1 per cent level; **significant at the 5 per cent level; *significant at the 10 per cent level; NS not significant at the 10 per cent level.