Are Males' Incomes Influenced by the Income Mix of Their Male Neighbors? Explorations into Nonlinear and Threshold Effects in Stockholm

Abstract

We investigate the degree to which neighborhood income composition affects the subsequent income of individual male residents, and test the degree to which these effects are characterized by nonlinear, threshold-like relationships. We specify a fixed-effects model to reduce potential bias arising from unmeasured individual characteristics affecting neighborhood selection and income. We employ annual data on 124 000 working-age males residing in Stockholm over the 1991–2006 period to estimate parameters for innovative variables measuring the sequence, duration, and intensity of neighborhood exposures. We find that two thresholds—one above 20 per cent and the other above 40 per cent—best describe the strong inverse relationship between consistent exposure to higher percentages of low-income male neighbors and subsequent earnings of individual male residents. We draw implications for potential causal mechanisms behind this relationship and formulating public policy towards places of concentrated disadvantage.

Keywords::

• Neighborhood effects, social mixing, nonlinear effects, threshold effects, fixed effects models, neighborhood income mix

Acknowledgements

The authors wish to thank participants in the Glasgow University Urban Studies Seminar and the Inequality and Social Policy Seminar at Harvard University for their constructive feedback on an earlier version of this work, and the University of Amsterdam and Uppsala University – Institute for Housing and Urban Research (IBF) for their institutional financial support. A special debt of gratitude is owed to Prof. Paul Jargowsky, whose exceptionally careful and insightful critique of a preliminary draft of this paper proved immensely helpful. The research assistance of Dan Beard is gratefully acknowledged.

Notes

¹ Also see the reviews in Leventhal & Brooks-Gunn (2000) and Friedrichs et al. (2003).

² More modern sociological treatises closely related to collective socialization also suggest thresholds (Granovetter, 1978; Granovetter & Soong, 1983). An illustration is Wilson's (1987) contention that as a critical mass of middle-class families leave the inner city, low-income blacks left behind become isolated from the positive role models that the erstwhile dominant middle class offered. Economists also have developed several mathematical treatises involving collective socialization effects in which thresholds often emerge as solutions to complex decision problems under certain assumptions (Akerlof, 1980; Brock & Durlauf, 2001; Galster, 1987, chap. 3).

³ The empirical challenges associated with obtaining accurate estimates of the independent causal effect of neighborhoods on individual residents has been the subject of numerous scholarly reviews (see Dietz, 2002; Ellen & Turner, 2003; Friedrichs et al., 2003; Galster, 2005, 2008; Gephart, 1997; Leventhal & Brooks-Gunn, 2000; Sampson et al., 2002). Arguably the central methodological challenge that researchers confront in obtaining an unbiased estimate of the magnitude of neighborhood effects is geographic selection bias (Ginther et al., 2000). The issue is that certain types of individuals who have certain (unmeasured) characteristics will move from/to certain types of neighborhoods and these same unobservables may also affect the outcome in question. Any observed relationship between neighborhood conditions and outcomes for such individuals or their children may therefore be biased because of this systematic spatial selection process, even if all the observable characteristics are controlled (Duncan et al., 1997; Manski, 1993, 1995, 2000).

⁴ We define the metropolitan area as the municipalities of Stockholm, Solna, and Sundbyberg.

⁵ This restriction also means that we do not analyze international immigrants who entered Stockholm after 1991. In a companion paper (Andersson et al., 2014), we investigate neighborhood effects for this subpopulation and find that, while strong, they also have distinctive features related to the ethnic and employment composition of neighbors.

⁶ Formally, income from work is computed here as the sum of: cash salary payments, income from active businesses, and the value of tax-based benefits that employees accrue as terms of their employment (sick or parental leave, work-related injury or illness compensation, daily payments for temporary military service, or giving assistance to a handicapped relative).

⁷ The log-linear transformation not only is appropriate given the positive skew of the income distribution, but also has sound grounding in economic theory, implicitly suggesting that income is a multiplicative (not additive) function of personal, neighborhood, and labor market characteristics.

⁸ As a practical matter, the national distributions of males and females are remarkably similar in Sweden.

⁹ Multicollinearity resulted in only the squared term being retained in the models.

¹⁰ We recognize that this span is arbitrary and represents a compromise: longer periods place higher requirements on how many years we must compute [N] and thus the number of permutations of patterns possible; shorter periods reduce the length of duration for which we can test.

¹¹ The ranges of minimum percentages for both groups were selected to insure adequate sample sizes for each alternative.

¹² We include in this myopic benchmark model controls for the individual's year of birth and ethnic background (i.e., being Swedish or foreign born) because we do not use fixed effects.

¹³ The results for control variables did not vary appreciably with the alternative specifications of [N]; thus, for brevity we report only the results for the model using exposure to 50 per cent of more low-income neighbors as the [N%X] variables.

¹⁴ Here we discuss the subsets of income mix dummies showing variations in sustained exposures; other permutations of sequences of exposures also evince similar general patterns and their results are available upon request.

¹⁵ We experimented to ascertain more precisely where the threshold was, and determined it was approximately 40 per cent.

¹⁶ In a semi-log model it is proper to use the expression EXP(b) −  1 to calculate the percentage difference in income associated with a dummy variable changing from zero to one, where b is the coefficient and EXP is the exponentiation function.

¹⁷ Results available upon request to the first author.