unobservables. In this case, the estimated effect of living in a deprived neighborhood on un-
employment is reduced by 4% compared with the simple probit estimate.21 Then sorting on
observable characteristics is accounted for by the estimation of the simultaneous model of two
probits. The naive effect of living in a deprived neighborhood declines further to 1.94 (column
4). Finally, the comparison of neighborhood effects in columns 4 and 5 assesses the added
value from explaining location in a deprived neighborhood by the public housing variable. This
specification produces the strongest decrease in the estimated effect, that looses 40% as soon
as the public housing variable is included among explanatory variables in the neighborhood
equation (comparison of columns 4 and 5 or columns 2 and 3). In fact, when introducing the
public housing variable in the neighborhood equation, we better account for the concentration
of disadvantaged individuals in deprived neighborhoods. Therefore, we better control for self-
selection effects and we obtain a much more reliable estimate of the neighborhood effect. This
result shows that the particular situation of public housing renters in France provides a valuable
opportunity to estimate the impact of neighborhood on socioeconomic outcomes.
As to public housing accommodation, the predicted effect on unemployment probability is
3.15 points in the baseline specification (Table 6, column 5). As we already explained, this effect
is entirely due to the influence of public housing on neighborhood choice, and its intensity is due
to the large impact of public housing accommodation on the probability to live in a deprived
neighborhood.
As highlighted by Ginther et al. (2000), another potential concern in the estimation of
neighborhood effects is an inadequate correction for unobserved heterogeneity. Although the
estimation of the simultaneous probit system ensures that the correlation between unobserv-
able characteristics is taken into account, it is worth performing an informal exercise in order to
roughly evaluate the potential biases generated by unobservable traits. Therefore, we reestimate
the model with two different specifications in which some known characteristics are assumed to
be unobservables. The first specification consists in dropping the individual’s occupational sta-
tus in both equations, meaning that we neglect a characteristic which is quite important in
determining the individual’s behavior on the labor-market and on the housing market. The
second specification eliminates the spouse nationality, a feature that has a weaker impact on un-
employment (see Table 4). As expected, the correlation of residuals and the predicted marginal
effect of neighborhood increase in both cases (Table 6, columns 6 and 7) compared with the
21 The correlation between residuals is positive, because the fact that neighborhood type is explained by ob-
servable traits is not taken into account.
18