Ethnic Preferences, Social Distance Dynamics, and Residential Segregation: Theoretical Explorations Using Simulation Analysis∗

In this paper I consider theories of residential segregation that emphasize social distance and ethnic preference dynamics. I argue that these theories are more compelling than many critics have supposed, and I conclude that they deserve to be considered more carefully. I then use simulation methodology to assess the potential impact of social distance and ethnic preference dynamics on ethnic segregation under certain theoretically interesting conditions. Based on the results from the simulation analyses, I offer three conclusions: (1) status preferences and status dynamics have the capacity to produce high levels of status segregation but do not produce high levels of ethnic segregation under the specified simulation conditions; (2) ethnic preferences can, under certain theoretically interesting conditions specified in these simulations, produce high levels of ethnic segregation in the absence of housing discrimination; and (3) ethnic preferences and social distance dynamics can, when combined with status preferences, status dynamics, and demographic and urban-structural settings common in American cities, produce highly stable patterns of multi-group segregation and hyper-segregation (i.e., high levels of ethnic segregation on multiple dimensions) of minority populations. Based on these model-based theoretical explorations I speculate that the persistence of segregation in recent decades may have been overdetermined, that is, it may have been sustained by multiple sufficient causes including not only discrimination, but also social distance and preference dynamics. This raises the possibility that reductions in housing discrimination may not necessarily lead to large declines in ethnic segregation in the short run because social distance and preference dynamics may be able to sustain ethnic segregation at surprisingly high levels in the absence of housing discrimination.

∗An earlier version of this paper was presented at the Annual Meetings of the American Sociological Association Chicago, Illinois, August 1999. I thank two anonymous reviewers for helpful suggestions for improvement of an earlier version of this paper. The research reported here has was facilitated in part by funding support from NIH Grants R43HD38199 (Simulating Residential Segregation Dynamics: Phase I) and R44HD038199 (Simulating Residential Segregation Dynamics: Phase II). In addition, the research reported here was made possible in part by faculty development leave support provided by Texas A&M University.

Keywords:

• agent-based model
• simulation
• ethnic segregation
• urban ecology
• prejudice
• ethnic preferences
• social distance

Notes

∗An earlier version of this paper was presented at the Annual Meetings of the American Sociological Association Chicago, Illinois, August 1999. I thank two anonymous reviewers for helpful suggestions for improvement of an earlier version of this paper. The research reported here has was facilitated in part by funding support from NIH Grants R43HD38199 (Simulating Residential Segregation Dynamics: Phase I) and R44HD038199 (Simulating Residential Segregation Dynamics: Phase II). In addition, the research reported here was made possible in part by faculty development leave support provided by Texas A&M University.

¹The predictions that low income groups will be centralized are grounded in the assumption that city neighborhoods are organized in a zonal pattern with new, high-quality construction being concentrated in the periphery of the city in low density, high-status residential areas.

²Peter Blau (1977) has emphasized the relationship between social distance segregation in a structural theory that while distinct from the ecological tradition, is compatible with it.

³Since Hawley (1944a 1944b) ecological theory recognizes that coordination and institutional practice can play an important supporting role in creating and maintaining segregation.

⁴See Massey and Denton (1993), Yinger (1995 1998), Galster (1991 1992), and Farley, Danziger, and Holzer (2000) for discussions focusing on the impact of discrimination on residential segregation.

⁵The individual preferences tradition in economics does not try to explain the existence of “ethnic preferences.” It merely assesses the implications of such preferences when they exist. In this regard, the “social distance” tradition in human ecology is more ambitious. It seeks to explain preferences as emerging from the competition between groups with disparate interests based on socioeconomic position and ethnic culture.

⁶The individual preferences tradition in economics does not dismiss housing discrimination as a possible cause of residential segregation, but does not itself address the dynamics of housing discrimination. In contrast, ecological theory is more general. It identifies mechanisms – categoric group formations (e.g., ethnic groups), intergroup competition, and dominance relations – that can produce segregation via discrimination and exclusion. Thus, while the focus in this paper is on strains of ecological theory that emphasize the role of voluntary choice and social distance dynamics, the ecological tradition is not restricted to explaining segregation solely in these terms.

⁷Recent articles by Massey (2001) and Yinger (1998) affirm their commitment to this position.

⁸The documentary film This is Our Home, It is Not for Sale (Schwarz, 1987) provides detailed ethnographic evidence of this dynamic. The film documents ethnic transition in an upper middle-class neighborhood in central Houston during the 1960s. It shows how discrimination (including symbolic and actual violence) impeded the initial integration of this neighborhood. However, after the “pioneering” phase of integration, minorities gained access to housing in the neighborhood through conventional means (e.g., realtors). A few sections of the neighborhood underwent rapid white-to-black succession based on “white” flight in response to block-busting activities. In the main, however, the area's transition from predominantly white to predominantly black came about gradually over two decades as white households exited the neighborhood for “normal” (i.e., life cycle-related) reasons and were disproportionately replaced by black rather than white households.

⁹One recent line of research focusing on white “flight” and “avoidance” behavior in residential decisions (e.g., Quillian, 2002; Crowder, 2000; Emerson, Yancey, and Chai, 2001) has begun to accumulate evidence that the researchers in this area interpret as consistent with key elements of Schelling's models.

¹⁰Statements by Farley and his associates have changed over time. Earlier statements, (at least up to Farley, Fielding, and Krysan, 1997) optimistically argued that white and black preferences could permit significant movement toward integration if discrimination were reduced. Later statements by Farley and associations are more cautious about the possibilities for integration under prevailing preferences. A very recent statement (Krysan and Farley, 2002) recognizes that preferences may be important factors in segregation but dismisses certain interpretations offered by preference theorists. I give the Krysan and Farley article extended attention below.

¹¹Schelling's work gives careful consideration to a broad range of issues including heterogeneous preferences on the part of whites for different neighborhood ethnic mixes. His theoretical models do not hinge on assumptions that whites seek all-white neighborhoods.

¹²The research literature consistently reports that, while status dynamics are not a central cause of ethnic segregation, status inequality and status dynamics does produce non-negligible levels of ethnic segregation.

¹³By definition, literal first arrivers cannot be evenly distributed across all neighborhoods. Later arrivers tend to locate nearby because: they have similar status characteristics; they depend on earlier migrants for information about residential options; and they often desire to locate near earlier migrants.

¹⁴As later discussion shows, minorities need only seek disproportionate co-ethnic presence for their preferences to be segregation-promoting.

¹⁵The most common first choice is for a 50/50 mix. The next most common first choice is for a mix that is substantially greater than 50% black (e.g., 70%).

¹⁶Proportionate or even distribution means that, for two groups being compared, the proportion of each group's population in each area is the same. Thus, for example, if an area contains a certain fraction (e.g., 0.02) of the city's white population, it also will contain that fraction of the city's black population.

¹⁷While not specifically drawing on urban ecological theory, Thernstrom and Thernstrom make a similar point (1997: 225–230).

¹⁸For convenience, I use the terms “city” and metropolitan area synonymously, focusing on the notion of the greater urban area that includes both central city and suburban ring.

¹⁹This value can be obtained from the standard formula D = 1/2Σ|w_i − b_i|, where i indexes neighborhoods and w and b measure each neighborhood's percentage share of the city-wide total of whites and blacks, respectively. The city's all-white neighborhoods can be treated as a single area for which |w_i − b_i| is 82.35 (since w_i is 82.35 and b_i is 0.0). Similarly, the remaining integrated neighborhoods can be treated as a single area for which |w_i − b_i| also is 82.35 (since b_i is 100.0 and w_i is 17.65). Thus, 1/2Σ|w_i − b_i| is 82.35.

²⁰Few whites (one third or less) indicate that they would be willing to move into a 50/50 white/black neighborhood and the number is undoubtedly much lower for majority black neighborhoods. Very few whites identify this neighborhood mix as a first choice or second choice. The vast majority of whites have a first choice for a majority white neighborhood.

²¹The city's population is 85% white. They will accept a proportion of 0.10 black. The expression 0.10 = X/(85 + X) represents X as the maximum proportion of the city's population that could be black and co-reside with whites without violating whites' preferences. Solving for X yields a figure of 9.44. The city's population is 15% black. Thus, 62.96% of blacks (62.96 = 100 · 9.44/15) would live in integrated neighborhoods and 37.04% of blacks would live in all-black neighborhoods. The city's all-black neighborhoods can be treated as a single area for which |w_i − b_i| is 37.04 (since w_i is 0.0 and b_i is 37.04). The integrated neighborhoods can be treated as a single area for which |w_i − b_i| also is 82.35 (since b_i is 62.96 and w_i is 100.0). Thus, D = 1/2Σ|w_i − b_i| = 37.04.

²²Schelling points out that the least tolerant whites leave an integrated area first. As they leave the ethnic mix changes making some of the remaining, less prejudiced whites uncomfortable and they leave and so on.

²³Even people who understand that neighborhood change might be driven by self-fulfilling prophecies may nevertheless act in ways that reinforce the dynamic because they rationally fear being a victim of the process (i.e., they fear being the last to sell their home in a neighborhood that is “turning”).

²⁴The fact that predominantly black suburbs tend to be older and lower status than predominantly white suburbs does not undercut this point. Black suburbs typically have lower crime and better services and amenities than predominantly black inner-city neighborhoods.

²⁵I prepared simple tabulations from the Multi-City Study of Urban Inequality to explore this issue and could not find evidence that preferences for black representation in neighborhoods were lower for upper- and middle-status black households; they also prefer 50/50 “integrated” neighborhoods as a first choice. Krysan and Farley (2002: 960) also report data that is consistent with this conclusion. All else equal, then, greater congregation is required to satisfy this preference.

²⁶Status homogeneity reduces the possibility that minorities can enter high-status areas at the “low end”. In light of this, it is interesting to note that area stratification within metropolitan areas appears to have increased in recent decades (Jargowsky, 1997). Gated communities and other class-homogeneous “up scale” development strategies effectively insure that high-status whites will have limited residential contact with blacks of any status and almost no contact with working-class blacks.

²⁷Socioeconomic assimilation provided greater means for successful competition in housing markets. But, the attenuation of socioeconomic differences also reduced the social distance between groups.

²⁸The one exception is when aversion is based on the expectation of extreme discrimination and hostile treatment. This issue is discussed later.

²⁹As is discussed in more detail below, their coding procedures limit “cultural” reasons. In addition, they identify reasons for desiring disproportionate co-ethnic presence such as “want diversity” or “would not feel comfortable” without considering whether “comfort” is linked to the presence of minority culture and institutions or whether “diversity” would be prized if it involved proportionate minority representation.

³⁰See for example, Park's (1926) discussion of spatial variation in groups that differ on life-style groupings differentiated on age, sex, and stage in the life-cycle.

³¹Alternatively, these sentiments could be negatively expressed as “doesn't feel right,” “doesn't feel comfortable,” “wouldn't fit in,” or “lacks a sense of community.”

³²My conversations with an admittedly non-random sample of middle-class blacks (mostly academic professionals and college students) on this topic provide no basis for contradicting Patterson's conclusion. For example, I have never encountered the view that it would be difficult for a middle-class black household to move to a home in a predominantly white neighborhood in the city I live in; yet the segregation level in this city is near the average for all metropolitan areas nationwide.

³³That is, they report that 89% of blacks are willing to move to a neighborhood that is 86% white before the move and 80% white after the move and that 73% of whites are willing to move to a neighborhood that is 78.6% white before the move and 80% white after the move.

³⁴In the interest of space, I excised a discussion on this subject. I review this in more detail in a forthcoming book (Fossett, forthcoming).

³⁵If households consider options in rank order of preference, blacks' second- and third-ranked choices are for areas that are 71–73% and 100% black, respectively.

³⁶My examination of census tract data for the MCSUI cities shows that intermediate mix neighborhoods, while not common, are hardly rare. But, for theoretical analysis it is important to note that even if they did not exist, “willingness to mix” preferences would readily permit their creation.

³⁷If all “willing” pioneers disperse evenly in feasible, partially integrated neighborhoods, black representation in these neighborhoods would be 100 · A/(A + B) where A is the number of willing black pioneers (0.35 · 20 = 7.0) and B is the number of willing whites (0.73 · 80 = 58.4). Thus, 100 · 7.0/(58.4 + 7.0) = 10.7.

³⁸All 89% of blacks who indicate they are willing to move to a 14.3–20% black neighborhood (before-after entry) could reside with whites in neighborhoods that are 18.2% black (20 · 0.89 = 17.8; 17.8/[80 + 17.8] = 18.2).

³⁹Even more integration is clearly possible since Krysan and Farley report that an additional 16% of whites would be willing to move to neighborhoods that are 6.7–7.1% black and could be paired with available blacks who would be willing to accept this neighborhood mix. However, more detailed preference data are needed to make exact calculations of the minimum D that could be achieved without violating any individual's willingness to mix preferences.

⁴⁰Similarly, while the vast majority of black households report being “willing” to reside in neighborhoods that are 14–20% black (a fact which permits high levels of integration under the model of “optimal mixing”), about 9 out of 10 of these households prefer areas that are 50–73% black over areas that are 14–20% black.

⁴¹There is some variation by neighborhood type; 15% of those who ranked all-black neighborhoods as their top choice mentioned concerns about discrimination and white hostility, but these concerns were mentioned by only 4% of the much larger group of black respondents who ranked 50–73% black areas as their top choice.

⁴²In a National Academy of Sciences report on segregation, Pettigrew highlighted the simulation work of Freeman and Sunshine (1970) focusing on discrimination and white prejudice as an “important new beginning…demonstrating the complex interface of prejudice attitudes and social and market factors” (1973: 77). He also noted that some of the implications of their work are “interesting and not necessarily obvious.” For whatever reason, this work did not stimulate future studies building on their initial efforts.

⁴³There are a variety of simulations that implement variations and extensions of the models Schelling, explored (e.g., Schelling, 1969a; Epstein and Axtell, 1996; Gaylord and D'Andria, 1998; Laurie and Jaggi, 2003). By comparison with SimSeg, these models are much simpler in terms of their representations of ethnic demography, socioeconomic inequality, urban structure, ethnic preferences, and behavioral rules. The Freeman and Sunshine (1970) effort stands as one of the more ambitious efforts of its time. But, it too implements a model that is much simpler than the SimSeg model.

⁴⁴The program is implemented in Delphi, an object-oriented variant of the Pascal programming language. The first versions of the program were developed to run under the DOS operating system. The current version runs under the Windows operating systems.

⁴⁵In the simulations reported here, search and movement is not constrained beyond means-testing, but the SimSeg also permits the activation of algorithms that also constrain search and movement by simulating housing discrimination. These dynamics are not activated in the simulations reported here.

⁴⁶The web address is http://vlab-resi.tamu.edu/simseg/simseg.htm. The program and documentation files are distributed in compressed “archives” (i.e., “zip” files). The documents are distributed as “PDF” files created using the Adobe Acrobat program.

⁴⁷The available documentation is for a DOS version of the SimSeg program (version 1.0). The simulation results presented here are generated by the Windows version of the program (SimSeg 2.00a). Separate documentation is not presently available for the Windows version of SimSeg, but all of its essential algorithms are identical to those used in the DOS version (the key differences are in the program's user interface).

⁴⁸The SimSeg Lite program can be found at http://vlab-resi.tamu.edu/simseg/simseg.htm. This program cannot be used to reproduce the specific simulations reported in this paper. But it is much easier to use and can illustrate all of the major findings reported here.

⁴⁹The fact that this is the case indicates that SimSeg's results are driven by the basic settings of its key variables. Complex refinements and enhancements allow for more precise representations of various theories, but generally do not alter simulation outcomes in substantively important ways.

⁵⁰All results reported here have been replicated using virtual cities with larger numbers of neighborhoods and households. The neighborhood dimensions used here are selected to yield a graphical representation that is visually intelligible (i.e., in large cities, households become visually indistinct).

⁵¹The overall status distribution has a median of 35, an interquartile range of 30.0, and an interdecile range of 53.2. The distribution has considerable intra-group status inequality with a gini index of concentration of 30.2 and an interdecile range for status scores extending from 12.9 to 66.1.

⁵²Housing and status are measured on the same scale. Housing is means-tested against household status.

⁵³Cycles are the basic measure of time in the simulation. Under the settings used in these simulation scenarios, they roughly correspond to time periods of 6–12 months (based on comparisons with observed patterns of residential mobility and turnover).

⁵⁴This simulates a wide range of factors that generate household movement; for example, job transfers, changes in household form, the “death” and “birth” of households, and so on.

⁵⁵The SimSeg program provides for the possibility of activating algorithms that simulate institutional constraints (i.e., various types of housing discrimination). Not surprisingly, they can produce ethnic segregation when they are activated. However, since these algorithms are not relevant to the issues under review here, none were active in the simulation experiments performed for this paper.

⁵⁶This norming procedure is particularly relevant in the present case because neighborhood size is small. Non-normed values of D yield much higher segregation levels reflecting the fact (discussed earlier in the text) that random assignment produces considerable variation in neighborhood ethnic composition. See Fossett (2005) for a detailed discussion of measuring segregation in computer simulations.

⁵⁷Segregation between socioeconomic groups is not reported in the tables for space considerations.

^a Norming yields lower numerical scores. The procedure is discussed in the text.

^b Measured by in-group representation in bounded neighborhood. Expected values under random assignment are approximately 60, 20, and 20 for whites, blacks, and Hispanics, respectively.

^c Measured using an index which yields − 100 when all group members are on the periphery, 0 when group members are randomly distributed, and 100 when group members are at the city center. Negative values are observed for all groups in Experiment 0 because vacancies are concentrated in central neighborhoods.

^d Measured by in-group representation in adjacent bounded neighborhoods. Expected values under random assignment are approximately 60, 20, and 20 for whites, blacks, and Hispanics, respectively.

^a Norming is discussed in the text. Normed measures yield lower numerical scores.

^b Measured by in-group representation in bounded neighborhood. Expected values under random assignment are approximately 60, 20, and 20 for whites, blacks, and Hispanics, respectively.

^c Measured using an index which yields –100 when all group members are on the periphery, 0 when group members are randomly distributed, and 100 when group members are at the city center. Negative values are observed for all groups in Experiment 0 because vacancies are concentrated in central neighborhoods.

^d Measured by in-group representation in adjacent bounded neighborhoods. Expected values under random assignment are approximately 60, 20, and 20 for whites, blacks, and Hispanics, respectively.

⁵⁸To do this, I first calculated the absolute difference between the normed dissimilarity score between whites and blacks and the mean for this same score over 1,000 replications; I then repeated this calculation for the normed dissimilarity score for whites and Hispanics and the normed dissimilarity score between Hispanics and blacks; next I summed the three figures; and then selected the experiment where the sum was closest to 0.0. In all cases, the sum for the expected experiment was very close to 0.0.

^a Norming is discussed in the text. Normed measures yield lower numerical scores.

^b Measured by in-group representation in bounded neighborhood. Expected values under random assignment are approximately 60, 20, and 20 for whites, blacks, and Hispanics, respectively.

^c Measured using an index which yields –100 when all group members are on the periphery, 0 when group members are randomly distributed, and 100 when group members are at the city center. Negative values are observed for all groups in Experiment 0 because vacancies are concentrated in central neighborhoods.

^d Measured by in-group representation in adjacent bounded neighborhoods. Expected values under random assignment are approximately 60, 20, and 20 for whites, blacks, and Hispanics, respectively.

⁵⁹The variability in segregation outcomes depends on the outcome considered and the particular experiment. With 1,000 replications, however, it is easy to distinguish systematic differences. For example, the largest 95% confidence interval around a mean reported in Appendix Table 2 has a width of only 0.775 and most are well under 0.500.

⁶⁰Given the relatively low sample-to-sample variability in the outcomes and the sample size of 1,000, differences in means of 1.0 or more are significant at 0.001 using a two sample difference of means test. Accordingly, I focus on substantive significance (rather than statistical) significance when discussing results.

⁶¹That is, the r-squared statistic from the non-linear regression of housing value on distance from the city center is 0.60. This figure falls within the range of between neighborhood variation in housing values seen in “real” cities (based on data for block groups).

^a Gray shading indicates that the model parameter value is changed relative to the previous scenario.

^b Not applicable because processes drawing on this model parameter are not active.

^c SES distributions are determined by five model parameters (the last five items shown in this section of the table). Other SES outcomes are reported for descriptive purposes, but, strictly speaking, are not model parameters and thus are not shaded.

⁶²I report median status because it is easy for readers to interpret. For technical reasons, the SimSeg program uses Lieberson's index of net difference (ND) to quantify inter-group inequality. ND is 40 for the white-black comparison, 25 for the white-Hispanic comparison, and 15 for the Hispanic-black comparison.

⁶³Jaret's (1995) discussion of the status dynamics hypothesis suggests that a proponent of this hypothesis might object that the level of ethnic inequality implemented here is not extreme enough. The level implemented is within the ranges of inequality observed for education, occupation, and income in real cities. However, group inequality in wealth and credit worthiness, which might be seen as important in housing dynamics, may be even greater (Oliver and Shapiro, 1995).

⁶⁴As noted earlier, segregation between socioeconomic groups is not reported in the tables.

⁶⁵The dispersion factor is implemented as a random normal deviate applied to a log-odds (logit) version of the preference target. This is theoretically appropriate for working with a percentage score bounded at 0 and 100. The logit version of the preference targets are always normally distributed. The “raw” score preference targets are approximately normally distributed when the median raw score target is 50 percent. The raw-score distributions are skewed left when the median raw score target is above 50 and skewed right when the median raw score target is below 50. These patterns are consistent with observed preference distributions.

⁶⁶Ethnic mix in adjoining areas is given half as much weight as ethnic mix in the immediate area. Interestingly, consideration of ethnic composition in adjoining areas tends to diminish the effect of ethnic preferences on “evenness of distribution” (the dimension of segregation measured by the index of dissimilarity) – it creates a competing goal. However, it tends to accentuates the effect of ethnic preferences on the “clustering” dimension of segregation and drives overall residential patterns from “checkerboarding” and toward “ghettoization.”

⁶⁷The specific methods of weighting the various considerations are discussed at length in the program documentation (Fossett, 1998a).

⁶⁸This is controlled by activating the model parameter for considering ethnic mix among “nearby neighbors” and the associated model parameter that sets the “nearby-neighbor range” to 5. Activating these parameters greatly increases the computational burden of the simulation experiment. For this reason, and since the results are otherwise little different, these nearby neighbor calculations are not used in other simulations.

⁶⁹The 95% center spread (i.e., approximately±1.96 standard deviations) in the distributions of in-group targets is 26.5 for whites (70.6–97.1), and 57.8 for blacks and Hispanics (21.1–78.9). These percentage targets and spread ranges are obtained by transforming relevant deviates from the normal distributions of logit-transformed preference targets.

⁷⁰For example, in Experiment 3 most of the “suburban” fringe is higher-status white, but a higher-status black area coalesced around area D11 and higher-status Hispanic areas coalesced around areas K5.

⁷¹Concentration, the fifth of the five dimensions of segregation Massey and Denton (1988) identify, is not relevant here because the density of housing units does not vary by neighborhood. However, it might be assumed that, following standard urban patterns, average lot size increases with distance from the city center. Under that assumption, central neighborhoods will be more densely settled and concentration effects will follow centralization effects.

⁷²The following points are established by experiments not reported in this article.

⁷³Whites do not seek out-group contact, so this preference comes into play when whites are choosing among neighborhoods that have some nonwhite presence.

⁷⁴Again, this is another example of how simulation analysis produces results with interesting theoretical implications. Concerns for ethnic representation in the broader area increase clustering, but dampen dissimilarity and isolation. One key qualification, in this scenario the macro area is defined in terms of immediate adjacency. If it were instead defined in terms of a larger area with a fixed boundary (e.g., a school district), the effects could be quite different.

⁷⁵Analyses not reported here indicate that about half of that can be attributed to the introduction of out-group targets and about half can be attributed to the specific configuration of out-group preferences.

⁷⁶Interestingly, the mean for normed black-Hispanic dissimilarity declines by over 8 points.

⁷⁷The notion of viewing the results of simulation models as “empirical evidence” regarding the behavior of a theoretical model is new in sociological research on segregation. But it is widely practiced in many fields where theoretical models are too complex to be investigated using deductive or analytic methods and instead must be explored using computational methods to generate data about model behavior.

⁷⁸I report simulation findings illustrating this possibility in Fossett (2001).