Fairness interventions often targets disparities associated with a single sensitive attribute in binary classification tasks. However, when multiple sensitive attributes are associated with a data subject, improving fairness with respect to one attribute does not necessarily improve it across all attributes: it may exacerbate disparities associated with other sensitive attributes, including those that are unavailable to the practitioner. We refer to these unintended effects as cross-attribute externalities of fairness. We study how fairness with respect to one sensitive attribute varies when an initially unfair classification or regression model is corrected with respect to another attribute. We characterize these externalities using Wasserstein barycentric disparities and develop a distributional framework that captures changes in both marginal and joint fairness measures. This highlights a limitation of conventional fairness evaluation, in which the success of an intervention is assessed mainly for the attribute it targets. In a jointly Gaussian model, we derive explicit trajectories for marginal and joint disparities and identify conditions under which an intervention eliminates disparities with respect to its target attribute while increasing disparities associated with another attribute. We also derive sharp partial-identification bounds on the resulting externalities when a sensitive attribute is unavailable and extend our results to categorical sensitive attributes. We illustrate these externalities using classification and regression real-data: in the classification experiments, interventions increase unfairness with respect to the hidden attribute in the majority of scenarios across more than half of the datasets considered; in the regression experiments on the ACSIncome dataset, correcting for race reduces sex disparities, whereas subsequently correcting for age increases them.
Figures & tables
Figure 1. A before-and-after audit of full correction for D1 . Gaussian sensitive attributes with β1=1 and β2=−0.6 . The three disparity indices (the targeted disparity U1 , the cross-attribute externality U2 , and the joint unfairness Ujoint ) are plotted as functions of ρ . Solid curves: original score; dashed: full correction for D1 . The indices do not depend on the distribution of RD .
Figure 2. Partial correction need not have the same effects on different audits. For ρ∈{0.4,0.7} , β1=1 , β2=−0.6 , the targeted ( U1 ) and joint ( Ujoint ) disparities decrease, while U2 immediately increases ( ρ=0.4 ) or ends up increasing after an initial improvement ( ρ=0.7 ). The exact trajectories follow ( 9 ).
ρ
U1(0)
U2(0)
U2(1)
Ujoint(0)
Ujoint(1)
0.2
0.7744
0.1600
0.3318
1.1200
0.3456
0.4
0.5776
0.0400
0.2540
0.8800
0.3024
0.6
0.4096
0.0000
0.1475
0.6400
0.2304
0.9
0.2116
0.0900
0.0130
0.2800
0.0684
Table 1. Cross-attribute externalities under full correction for D1 . Marginal ( U1,U2 ) and joint ( Ujoint ) disparity indices before and after full correction for D1 for different values of ρ∈[0,0.9] in a jointly Gaussian model with (β1,β2)=(1,−0.6) .
Figure 3. Partial identification of a hidden externality. We fix a=Corr(D1,Z)=0.5 and vary the correlation ρ=Corr(D1,D2)∈[0,1] along the horizontal axis. The left panel shows the sharp bounds on the change in disparity associated with D2 . The central blue curve is their arithmetic midpoint and corresponds to conditional independence ( η=0 ); the exterme curves correspond to η=±1 . The right panel shows the sharp identified interval for residual disparity after full correction of D1 , [0,(1−a2)(1−ρ2)] .
Dataset
Mitigated
Hidden
ρ1h
Mean ΔUh (sd)
Min ΔUh
Max ΔUh
% ΔUh > 0
Adult
Race
Sex
0.101
-0.019 (0.013)
-0.060
0.005
6.7
Sex
Race
0.101
-0.059 (0.020)
-0.111
0.003
0.8
Compas
Race
Sex
0.066
-0.015 (0.011)
-0.051
0.018
5.0
Sex
Race
0.066
-0.011 (0.007)
-0.033
0.004
5.8
Default
Race
Sex
-0.068
0.002 (0.007)
-0.012
0.037
56.7
Sex
Race
-0.068
0.002 (0.007)
-0.025
0.022
67.5
Table 2. Changes in hidden-attribute unfairness, ΔUh , under Wasserstein barycenter mitigation (EquiPy). For each dataset and mitigated–hidden attribute pair, reports include the mean change (standard deviation in parentheses), minimum, maximum, and % of runs with ΔUh>0 (indicating a deterioration in fairness), across 30 random seeds and four classification models (LR, RF, SVM, and DNN). ρ1h is the correlation between the mitigated and hidden attributes.
Figure 4. Intermediate and joint mitigation effects on hidden-attribute unfairness with EquiPy. Mitigating both age and sex, we report race-related unfairness for the baseline model, after age or sex mitigation, and after joint age–sex mitigation, for Mep15 (left panel) and Mep16 (right panel). Individual mitigation steps can reduce hidden-attribute unfairness, whereas the subsequent joint correction can increase it (below the baseline level). Error bars and markers indicate variability across seeds and models.
Dataset
Mitigated
Hidden
ρ1h
Mean ΔUh (sd)
Min ΔUh
Max ΔUh
% ΔUh > 0
Law School
Race
Sex
0.077
-0.035 (0.059)
-0.184
0.133
23.3
Insurance
Sex
Age
0.003
-0.138 (0.240)
-0.802
0.396
27.8
Communities and Crime
Race
Age
0.085
0.002 (0.043)
-0.089
0.169
50.0
Age
Poverty
0.138
-0.019 (0.034)
-0.105
0.149
22.2
Poverty
Race
-0.336
-0.164 (0.026)
-0.235
-0.118
0.0
ACSIncome (PA)
Race
Sex
0.022
-0.027 (0.072)
-0.358
0.121
34.4
Table 3. Changes in hidden-attribute unfairness, ΔUh , under Wasserstein barycenter mitigation (EquiPy) . For each dataset and mitigated–hidden attribute pair, ΔUh is summarized across 30 random seeds and the regression models (GLM, RF, and NN) considered. The table reports the mean change (standard deviation in parentheses), minimum, maximum, and percentage of runs with ΔUh>0 , and ρ1h denotes the correlation between the mitigated and hidden attributes. Positive values indicate an increase in hidden-attribute unfairness. The reverse mitigation directions are reported in Table 9 in Appendix.
Dataset
Mitigated
Hidden
ρ1h
ρ2h
Mean ΔUh
Min ΔUh
Max ΔUh
% ΔUh > 0
Communities and Crime
Race/Age
Poverty
-0.3358
0.1383
-0.193 (0.054)
-0.316
-0.057
0.0
Race/Poverty
Age
0.0853
0.1383
-0.014 (0.036)
-0.097
0.089
28.9
Age/Poverty
Race
0.0853
-0.3358
-0.175 (0.025)
-0.240
-0.115
0.0
ACSIncome (PA)
Race/Age
Sex
0.0219
0.0123
0.024 (0.112)
-0.260
0.336
64.4
Sex/Age
Race
0.0219
0.0896
-0.547 (0.368)
-1.340
0.120
3.3
Sex/Race
Age
0.0123
0.0896
-0.049 (0.082)
-0.275
0.2325
22.75
Table 4. Changes in hidden-attribute unfairness, ΔUh , under Wasserstein barycenter mitigation (EquiPy) . For each dataset, the table considers pairs of mitigated attributes (D1,D2) and a hidden attribute h . Results are summarized across 30 random seeds and the regression models (GLM, RF, and NN). The table reports the mean change (standard deviation in parentheses), minimum, maximum, and percentage of runs with ΔUh>0 , and ρ1h , ρ2h denote the correlation between the hidden attribute h and the first and second mitigated attribute, respectively. Positive values of ΔUh indicate an increase in hidden-attribute unfairness.
Figure 5. Intermediate and joint mitigation effects on hidden-attribute unfairness. The left panel reports age-related unfairness for the baseline model, after race or poverty mitigation, and after joint race–poverty mitigation in the Communities and Crime dataset. The right panel reports sex-related unfairness for the baseline model, after race or age mitigation, and after joint race-age mitigation in the ACS Income dataset. The plots illustrate that an individual mitigation step can increase unfairness associated with a hidden attribute, whereas the subsequent joint correction can reduce unfairness below its baseline level. Error bars and markers indicate the variability or range across the evaluated results (across seeds and models).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Assignments 1–18
Assignments 19–36
No
D2,1:4
D2,5:8
U2(Z)
U2(Z′)
ΔU2
No
D2,1:4
D2,5:8
U2(Z)
U2(Z′)
ΔU2
1
0110
1001
0.250
0.000
-0.250
19
0011
0011
0.250
0.250
+0.000
2
1001
0110
0.250
0.000
-0.250
20
0011
0101
0.250
0.250
+0.000
3
0011
0110
0.250
0.125
-0.125
21
0011
1001
0.125
0.125
+0.000
4
0011
1010
0.125
0.000
-0.125
22
0101
0011
0.250
0.250
+0.000
5
0011
1100
0.125
0.000
-0.125
23
0101
0101
0.250
0.250
+0.000
Appendix
Table 5. The 36 admissible assignments h∈A of the unobserved attribute D2 across two observed groups of four observations, ordered by ΔU2 . D2,1:4 denotes the values assigned to D2 in the group D1=−1 , while D2,5:8 denotes the values assigned to D2 in the group D1=+1 . Lower and upper bounds on ΔU2 are shown in bold.
ΔU2
Number of assignments
Proportion
−0.250
2
5.6%
−0.125
16
44.4%
−0.000
16
44.4%
+0.125
2
5.6%
Total
36
100%
Appendix
Table 6. Distribution of externalities compatible with the demographic constraints, across the 36 admissible assignments of the unobserved attribute D2 from Table 5 .
Mitigation
Dataset
Mitigated
Hidden
ρ1h
Mean ΔUh (sd)
Min ΔUh
Max ΔUh
% ΔUh>0
DIR
Adult
Race
Sex
0.101
-0.026 (0.052)
-0.253
0.238
20.0
Sex
Race
0.101
-0.026 (0.046)
-0.211
0.150
15.0
Compas
Race
Sex
0.066
-0.002 (0.017)
-0.069
0.050
34.2
Sex
Race
0.066
-0.006 (0.013)
-0.046
0.034
25.8
Default
Age
Sex
-0.068
-0.012 (0.024)
-0.064
0.105
20.8
Sex
Age
-0.068
-0.002 (0.040)
-0.177
0.217
40.8
Appendix
Table 7. Changes in hidden-attribute unfairness, ΔUh , under pre-processing mitigation DIR and RW. For each dataset and mitigated–hidden attribute pair, reports include the mean change (standard deviation in parentheses), minimum, maximum, and % of runs with ΔUh>0 (indicating a deterioration in fairness), across 30 random seeds and four classification models (LR, RF, SVM, and DNN). ρ1h is the correlation between the mitigated and hidden attributes.
Figure 6. Model-dependent mitigation externalities. The figure compares hidden-attribute unfairness across the baseline model, the intermediate mitigation step, and the subsequent joint correction for the LR, SVM, RF, and DNN classification models. The direction of the change in unfairness following each correction depends on the model. For Mep15, an individual mitigation step may either increase (LR, RF, SVM) or decrease (DNN) hidden-attribute unfairness, whereas for Mep16, it consistently increases it. For both datasets, the subsequent joint correction reduces hidden-attribute unfairness, although it sometimes remains above the baseline level (Mep15 with LR). These results highlight the model- and dataset- dependent nature of fairness externalities under sequential mitigation.
Figure 7. Unfairness as a function of the mitigation parameter λ . Each panel reports the mean observed, hidden-attribute, and joint unfairness across random seeds and classification models (LR, SVM, RF, and DNN) for the indicated dataset and attribute configuration. Solid, dashed, and dotted lines distinguish observed, hidden-attribute, and joint unfairness, respectively; shaded bands represent one standard deviation around the mean. For Mep15, when mitigating sex (bottom-right panel), race-related unfairness increases as λ grows, while joint unfairness decreases markedly. This illustrates that improvements in joint fairness can coexist with adverse externalities on a hidden attribute, highlighting the importance of examining both marginal and joint indices.
Dataset
Mitigated
Hidden
ρ1h
ρ2h
Mean ΔUh
Min ΔUh
Max ΔUh
% ΔUh > 0
Mep15
Age/Sex
Race
-0.013
-0.016
−0.004 (0.014)
-0.053
0.036
37.1
Mep16
Age/Sex
Race
-0.046
-0.017
−0.018 (0.014)
-0.056
0.017
8.3
Appendix
Table 8. Changes in hidden-attribute unfairness under Wasserstein barycenter mitigation (EquiPy) with two observed attributes. For each dataset with three sensitive attributes (Mep15 and Mep16), we consider pairs of mitigated attributes and a hidden attribute h . Results are aggregated over 30 random seeds and four classification models (LR, SVM, RF, and DNN) out-of-sample. Reports include the mean change in unfairness (standard deviation in parentheses), minimum, maximum, and % of runs with ΔUh>0 (indicating increased hidden-attribute unfairness). Here, ρ1h and ρ2h denote the correlations between the first and second mitigated attributes, respectively, and the hidden attribute h .
Dataset
Mitigated
Hidden
ρ1h
Mean ΔUh (sd)
Min ΔUh
Max ΔUh
% ΔUh > 0
Law School
Sex
Race
0.077
-0.016 (0.013)
-0.053
0.007
7.8
Age
Sex
0.003
-0.033 (0.0132)
-0.650
0.300
31.1
Communities and Crime
Age
Race
0.085
-0.004 (0.009)
-0.024
0.036
22.2
Poverty
Age
0.138
0.002 (0.031)
-0.070
0.087
54.4
Race
Poverty
-0.336
-0.181 (0.050)
-0.314
-0.041
0.0
ACSIncome (PA)
Sex
Race
0.022
-0.006 (0.043)
-0.188
0.176
41.1
Appendix
Table 9. Changes in hidden-attribute unfairness, ΔUh , under Wasserstein barycenter mitigation (EquiPy) . For each dataset and mitigated–hidden attribute pair, ΔUh is summarized across 30 random seeds and the regression models (GLM, RF, and NN) considered. The table reports the mean change (standard deviation in parentheses), minimum, maximum, and percentage of runs with ΔUh>0 , and ρ1h denotes the correlation between the mitigated and hidden attributes. Positive values indicate an increase in hidden-attribute unfairness.
Figure 8. Unfairness as a function of the mitigation parameter λ . Each panel reports the mean observed, hidden-attribute, and joint unfairness across random seeds and regression models (GLM, RF, and NN) for the indicated dataset and attribute configuration out-of-sample. Solid, dashed, and dotted lines distinguish observed, hidden-attribute, and joint unfairness, respectively; shaded bands represent one standard deviation around the mean.
Figure 9. Model-dependent mitigation externalities. The figure compares hidden-attribute unfairness across the baseline model, the intermediate mitigation step, and the subsequent joint correction for the GLM, RF, and NN predictive models. The direction of the change in unfairness following each correction depends on the model: an individual mitigation step may increase or decrease hidden-attribute unfairness, and the subsequent joint correction may further increase it or reverse the intermediate improvement. These results highlight the model dependence of fairness externalities under sequential mitigation.
Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.