Pith. sign in

REVIEW 4 major objections 5 minor 9 references

Counterfactual Situation Testing: From Single to Multidimensional Discrimination

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Counterfactual situation testing detects individual discrimination by comparing complainants to their counterfactual selves.

desk verdict CST is a real conceptual step forward, but the experiments overclaim detection performance without a ground-truth check. read the letter →

arxiv 2502.01267 v2 pith:47NRGD6F submitted 2025-02-03 cs.LG

Jose M. Alvarez , Salvatore Ruggieri This is my paper · ORCID
classification cs.LG
keywords counterfactualsituationtestingindividualdiscriminationfairnessgiventhedifferenceintersectionalmultiplestructuralcausalmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces counterfactual situation testing (CST), a causal data-mining method for finding individual cases of discrimination in a dataset of classifier decisions. Its central claim is that a discrimination test should compare a complainant not to similar people of another protected status holding everything else fixed, but to a counterfactual version of the complainant whose other attributes are allowed to change as they would under the protected-status manipulation. The paper shows this 'fairness given the difference' comparison finds more potential discrimination cases than standard situation testing in synthetic and law-school admission experiments, and can flag discrimination even when a classifier is counterfactually fair. It also formalizes multiple versus intersectional discrimination and reports evidence that the former misses the latter.

What carries the argument

The load-bearing object is the counterfactual dataset $D^{\mathrm{CF}}$, the mapping of every protected instance into the world where the protected attribute is set to its non-protected value through the structural causal model $M$. It matters because it supplies the test search center: the test group is chosen by similarity to the complainant's counterfactual self, not to the complainant's factual self, so group membership encodes the mutatis mutandis manipulation. The comparison engine is the rate difference $\Delta p$ with its Wald confidence interval, which turns the single factual-versus-counterfactual comparison of counterfactual fairness into a many-to-many, statistically testable claim. In the multidimensional setting the same machinery is reused either once per protected attribute for multiple discrimination or once on the intersectional attribute $A^*$ for intersectional discrimination.

What would settle it

Generate a synthetic dataset from a known causal model with a true discrimination effect, then run CST with a deliberately misspecified model—for example, one that omits a common cause of $A$ and $X$, or reverses the sign of the $A \to X$ coefficient. If the misspecified run still reports a comparable set of significant cases, the central dependence on correct causal knowledge is refuted; if the case set changes materially, the assumption is doing the work the paper assigns to it.

Watch

Extended reading notes

Core claim

Under CST, each complainant $c$ gets two search centers: her own profile $x_c$ for the control group, and her counterfactual profile $x_c^{\mathrm{CF}}$, generated by the abduction-action-prediction steps under the intervention $do(A:=0)$, for the test group. The control and test groups are the $k$ nearest protected and non-protected instances to those centers, and the test statistic is the difference in negative-decision rates $\Delta p = p_c - p_t$. A claim of individual discrimination is declared when $\Delta p > \tau$, and it is statistically significant when the one-sided Wald confidence interval excludes $\tau$. The paper argues this operationalizes fairness given the difference, because changing the protected attribute $A$ is allowed to change the seemingly neutral attributes $X$; it reports that CST detects more cases than situation testing and that counterfactual fairness neither implies nor is implied by CST discrimination, with experiments demonstrating discrimination even for counterfactually fair classifiers.

Load-bearing premise

The result depends on the structural causal model $M$ being a correct specification of how the protected attribute affects the other attributes, including no unobserved confounders; if $M$ is wrong, the counterfactual profiles and therefore the test groups are biased, and no discrimination claim under CST can be trusted.

Editorial extensions

If this is right

  • If CST is right, discrimination auditing can replace the idealized 'all else equal' comparison with a causal 'change what needs changing' comparison whenever stakeholders can agree on a structural causal model.
  • Counterfactual fairness alone is not enough to clear a classifier: a model can be counterfactually fair yet show statistically significant negative-decision differences around borderline complainants.
  • Counterfactual fairness claims can be reported with confidence intervals, giving courts and auditors a measure of certainty on individual claims.
  • Multidimensional audits should include an intersectional protected attribute, since multiple separate single-attribute tests can miss complainants who are discriminated at the intersection.
  • As the neighborhood size $k$ grows, the average detected rate difference converges to the demographic parity gap $P(\hat{Y}|A=1)-P(\hat{Y}|A=0)$, so CST's added value concentrates in small- and medium-sized neighborhoods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: a natural deployment is to run CST under a small set of plausible structural causal models and report the disagreement between the case sets, turning the single-model dependence into a sensitivity band.
  • Editorial inference: the counterfactual world is a normative target, so choosing an SCM is also choosing what society should look like; auditors should make that target explicit and contestable rather than presenting it as neutral data processing.
  • Editorial inference: the gap between intersectional and multiple case counts could serve as a quantitative 'intersectionality deficit' audit metric for deployed systems, beyond the two scenarios studied.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Counterfactual Situation Testing (CST), a framework for detecting individual-level discrimination in classifier decisions. For each complainant, CST builds a control group of similar protected instances and a test group of similar non-protected instances, but with the test group centered on the complainant's counterfactual generated from a structural causal model under the mutatis mutandis manipulation. Discrimination is flagged when the difference in negative-decision rates Δp exceeds a threshold τ, with a one-sided Wald confidence interval providing statistical significance. The framework is extended to multiple and intersectional discrimination, and experiments compare CST with situation testing (ST) and counterfactual fairness (CF) on synthetic loan data and law school admissions data. The paper reports that CST detects more cases than both ST and CF, including cases where the classifier is counterfactually fair, and argues that intersectional discrimination is not captured by multiple discrimination testing.

Significance. If validated, the paper makes a useful conceptual contribution: it operationalizes 'fairness given the difference' in a concrete way, links counterfactual fairness to confidence intervals, and provides an explicit treatment of multidimensional and intersectional discrimination. The framework is clearly defined, the k-NN implementation is simple and reproducible, and the code is available. The synthetic experiment is a strength because the data-generating SCM is known, which makes oracle counterfactual outcomes computable; however, the paper does not exploit this to validate detection accuracy. The real-data experiment is illustrative rather than confirmatory because the classifier is hand-chosen and the SCM is estimated on the same data. Overall, the central conceptual claim is defensible, but the empirical evidence for 'more discrimination cases' needs stronger validation against ground truth or at least a null model.

major comments (4)
  1. [§5.2, Table 1 and §5.2.1] The central empirical claim, that CST 'validates the use of CST w/o over ST,' is based solely on the number of cases where Δp > τ. Since no individual-level ground-truth labels of discrimination are defined or used, a higher count does not establish better detection. In the synthetic experiment the true SCM of Figure 1 is available, so an oracle counterfactual outcome for each complainant is computable; the paper should report precision, recall, and accuracy of CST, ST, and CF against this oracle label. Without such a comparison, the counts in Table 1 cannot distinguish genuine discrimination from artifacts of the procedure.
  2. [§5.2.1, Figure 2 and Table 3] The higher detection counts appear to be partly built into the construction of the test group: because the counterfactual center x_CF is shifted to systematically higher X1 and X2 (Figure 2), the test neighborhoods have lower rejection rates by construction, making Δp > τ more likely regardless of whether the flag corresponds to true discrimination. To support the claim that CST detects more real discrimination, the paper should include a null experiment where the SCM has no effect of A on X (or where b() does not use X) and report the false-positive rate of CST and ST under that null model. This would directly test whether the count difference is a mechanical consequence of the shifted center.
  3. [§3.1 and §6] The framework's validity depends on the correctness of the SCM M, but no sensitivity analysis is provided. Section 3.1 assumes causal sufficiency, acyclicity, and additive noise, and Section 6 acknowledges that a biased SCM biases DCF; however, the experiments always use a single fixed M. In the law school experiment (Section 5.3), the SCM is re-estimated on the same data used for evaluation, and b() is externally imposed. The paper should vary the SCM coefficients, introduce a hidden confounder, or compare against an alternative plausible SCM to show that the qualitative conclusions are not an artifact of a particular misspecified model.
  4. [Eq. (10) and Definition 4.5] The Wald confidence interval in Eq. (10) treats pc and pt as proportions from independent samples, but the control and test groups are selected as k-nearest neighborhoods from the same fixed dataset, and the neighborhoods of different complainants overlap. Moreover, thousands of complainants are tested simultaneously, and no multiple-testing correction is applied for the single-attribute runs (Bonferroni is used only for the q attributes in multiple discrimination). The statistical-significance counts in Tables 1, 10, 11, 12, and 13 may therefore be overstated. The authors should either justify the independence assumption, use a permutation or bootstrap procedure, or explicitly report the expected number of false positives under a null model.
minor comments (5)
  1. [Algorithm 1, Appendix B.1] Algorithm 1 uses x_CF_c in line 6 but lists no SCM or counterfactual dataset as input; the pseudo-code should either include DCF as an input or specify that x_CF_c is generated externally via the abduction-action-prediction steps described in Section 3.1.
  2. [Table 12, row ST, k=250] The entry '24 (5.1%)' is inconsistent with the stated percentage of non-white-female applicants (8.4% of 21,790, so 24 cases would be about 1.31%); the percentage appears to be a typographical error and should be corrected.
  3. [Table 8 and Table 9 captions] The captions contain the typo 'statical significance'; this should read 'statistical significance.'
  4. [§4.3, Eq. (10)] The text states that the confidence interval is one-sided but the formula writes the lower bound as [Δp − wα, +∞); it would be clearer to state explicitly that zα is the 1−α quantile and that this is a lower one-sided interval for Δp > τ.
  5. [Figure 4, center panel] The center plot shows average Δp for CST w/o and ST, but the text says 'both methods show similar trends' while the curves appear to diverge at small k; a brief explanation of why the trends are considered similar would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CST definitions and experiments are self-contained, with only non-load-bearing terminological self-citations.

full rationale

I find no circular step in the paper's derivation chain. Definitions 4.1-4.7 fix the control and test groups from factual and counterfactual search centers, and Definition 4.4 explicitly defines individual discrimination as Δp > τ; this is an operationalization, not a claim that a fitted constant predicts a held-out quantity. The self-citations to Álvarez and Ruggieri (2023, 2024) point to the authors' own terms 'fairness given the difference' and the CP/MM manipulation labels, but the concepts are developed in the text and anchored to Kohler-Hausmann (2018), so the citations are not load-bearing. In the synthetic experiment (Section 5.2), the SCM is the data-generating model by design, so the experiment is a controlled illustration of the framework's behavior; no parameter is fitted and then renamed as a prediction. The higher CST detection counts are analyzed by the paper as consequences of the mutatis mutandis test-center shift (Tables 2-3), not presented as externally validated discoveries. Section 6 explicitly acknowledges that indirect-discrimination testing 'sounds like a self-fulfilling prophecy' and explains why detection is not guaranteed; that is a limitation about validation strength, which the rules distinguish from circularity. Overall, the framework is self-contained and no prediction reduces to its inputs by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The CST framework itself contains no fitted parameters; the free parameters are user-chosen thresholds and the SCM coefficients. The core assumptions are the correct specification of the SCM and standard asymptotics for the statistical test.

free parameters (5)
  • Synthetic SCM coefficients (f1, f2) = chosen by authors
    In the synthetic loan experiment (Section 5.2), the authors choose the structural equations, so the counterfactual distribution is determined by their choices.
  • Law school SCM coefficients (β1, β2, λ1, λ2, bU, bL) = estimated from LSAC data
    In Section 5.3, the SCM from Kusner et al. is re-estimated on the same LSAC data used for discrimination testing.
  • Accepted deviation τ = 0.0
    Set by the authors for all experiments; defines the threshold for a discrimination claim.
  • Significance level α = 0.05
    Set by the authors; controls the confidence intervals in Equations 10 and 11.
  • Neighborhood size k = 15, 30, 50, 100, 250
    User-chosen; the number of detected cases varies with k.
assumptions (5)
  • domain assumption Causal sufficiency: no hidden common causes in the SCM
    Stated in Section 3.1; allows the noise terms to factor and is needed for the abduction step.
  • domain assumption Acyclicity of the causal graph G
    Stated in Section 3.1; restricts SCM to DAGs.
  • domain assumption Additive noise models for the structural equations
    Stated in Section 3.1; simplifies the abduction step, though the authors say it is not necessary.
  • domain assumption Protected attribute A has no causal parents and affects only X; Y = b(X) is deterministic
    Assumed in the DAGs of Figures 1 and 6; if A had parents or Y depended directly on A, the test group construction would differ.
  • standard math Asymptotic normality of the difference in sample proportions for the Wald confidence interval
    Used in Equations 10 and 11 for the significance tests.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Counterfactual Situation Testing: From Single to Multidimensional Discrimination." pith.science (2026). https://pith.science/paper/47NRGD6F

@misc{pith2026250201267,
  author       = {Pith},
  title        = {Pith review of: Counterfactual Situation Testing: From Single to Multidimensional Discrimination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/47NRGD6F}},
  note         = {Machine review of arXiv:2502.01267}
}
read the original abstract

We present counterfactual situation testing (CST), a causal data mining framework for detecting individual discrimination in a dataset of classifier decisions. CST answers the question ``what would have been the model outcome had the individual, or complainant, been of a different protected status?'' It extends the legally-grounded situation testing (ST) of Thanh et al. (2011) by operationalizing the notion of "fairness given the difference" via counterfactual reasoning. ST finds for each complainant similar protected and non-protected instances in the dataset; constructs, respectively, a control and test group; and compares the groups such that a difference in model outcomes implies a potential case of individual discrimination. CST, instead, avoids this idealized comparison by establishing the test group on the complainant's generated counterfactual, which reflects how the protected attribute when changed influences other seemingly neutral attributes of the complainant. Under CST we test for discrimination for each complainant by comparing similar individuals within the control and test group but dissimilar individuals across these groups. We consider single (e.g.,~gender) and multidimensional (e.g.,~gender and race) discrimination testing. For multidimensional discrimination we study multiple and intersectional discrimination and, as feared by legal scholars, find evidence that the former fails to account for the latter kind. Using a k-nearest neighbor implementation, we showcase CST on synthetic and real data. Experimental results show that CST uncovers a higher number of cases than ST, even when the model is counterfactually fair. CST, in fact, extends counterfactual fairness (CF) of Kusner et al. (2017) by equipping CF with confidence intervals, which we report for all experiments.

Figures

Figures reproduced from arXiv: 2502.01267 by the authors.

Figure 1
Figure 1. The auxiliary causal knowledge for Example 1.1 (and Section 5.2). Let b [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Distribution of the attributes used by the ADM [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 1
Figure 1. Marked by * are the number of statistically significant cases. [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figures from the paper (8 more)
Figure 3
Figure 3. Figure 3: The boxplots for a subset of cases detected by ST and CST w/o for [PITH_FULL_IMAGE:figures/full_fig_p020_3.png]
Figure 4
Figure 4. Figure 4: Left and center: Number of cases by CST w/o and ST and their avg. ∆ [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Number of cases by CST w/o and w/ and their average ∆ [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: The auxiliary causal knowledge for Section 5.3. Let [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 6
Figure 6. Figure 6: Marked by * are the statistically significant cases. [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Average ∆p and number of cases for race (R) and gender (G), respectively. We plot all and statistically significant (sig.) cases for each method. discriminate against female applicants? To answer this question using CST and CF, we generate the corresponding DCF G using…
Figure 8
Figure 8. Figure 8: Avg. ∆p and number of cases for multiple (left) and intersectional (right) dis￾crimination on R and G. We plot all and statistically significant (sig.) cases for all methods. tion testing procedures. Under multiple discrimination, we still rely on DCF R and DCF G for r…
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p040_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 7 canonical work pages

  1. [1]

    Adams-Prassl, J., Binns, R., & Kelly-Lyth, A. (2023). Directly discriminatory algorithms. The Modern Law Review , 86 (1), 144–175. Adler, J. S. (2019). Murder in New Orleans: The creation of Jim Crow policing . University of Chicago Press. Aggarwal, A., Lohia, P., Nagar, S., Dey, K., & Saha, D. (2018). Automated test generation to detect individual discri...

    work page Pith review arXiv 2023

  2. [31]

    Mutatis Mutandis: Revisiting the Comparator in Discrimination Testing

    ´Alvarez, J. M., & Ruggieri, S. (2023). Counterfactual situation testing: Uncovering discrim- ination under fairness given the difference. EAAMO, 2:1–2:11. 2318 Counterfactual Situation Testing ´Alvarez, J. M., & Ruggieri, S. (2024). Mutatis mutandis: Revisiting the comparator in discrimination testing. CoRR, abs/2405.13693. Angrist, J. D., & Pischke, J.-...

    work page Pith review arXiv 2023

  3. [50]

    Newcombe, R. G. (1998). Interval estimation for the difference between independent pro- portions: Comparison of eleven methods. Statistics in Medicine , 17 (8), 873–890. Pearl, J. (2009). Causality: Models, reasoning, and inference (2nd). Cambridge University Press. Pearl, J., Glymour, M., & Jewell, N. P. (2016). Causal inference in statistics: A primer ....

    work page 1998

  4. [58]

    2321 Alvarez & Ruggieri Rorive, I. (2009). Proving discrimination cases: The role of situation testing. Centre for Equal Rights and MPG . https : / / ec . europa . eu / migrant - integration / library - document/proving-discrimination-cases-role-situation-testing en Rose, E. K. (2022). A Constructivist Perspective on Empirical Discrimination Research. Jou...

    work page 2009

  5. [513]

    Kamiran, F., & Calders, T. (2009). Classifying without discriminating. Int. Conf. on Com- puter, Control and Communication , 1–6. Karimi, A., Sch¨ olkopf, B., & Valera, I. (2021). Algorithmic recourse: From counterfactual explanations to interventions. F AccT, 353–362. Karimi, A., von K¨ ugelgen, B. J., Sch¨ olkopf, B., & Valera, I. (2020). Algorithmic re...

    arXiv 2009

  6. [735]

    V., & Russakovsky, O

    Wang, A., Ramaswamy, V. V., & Russakovsky, O. (2022). Towards intersectionality in machine learning: Including more identities, handling underrepresentation, and per- forming evaluation. F AccT, 336–349. Weerts, H. J. P., Xenidis, R., Tarissan, F., Olsen, H. P., & Pechenizkiy, M. (2023). Algo- rithmic unfairness through the lens of EU non-discrimination l...

    work page 2022

  7. [841]

    Wachter, S., Mittelstadt, B., & Russell, C. (2020). Bias preservation in machine learning: The legality of fairness metrics under eu non-discrimination law. W. Va. L. Rev. , 123,

    work page 2020

  8. [1163]

    Kozodoi, N., Jacob, J., & Lessmann, S. (2022). Fairness in credit scoring: Assessment, implementation and profit implications. Eur. J. Oper. Res. , 297 (3), 1083–1094. Kusner, M. J., Loftus, J. R., Russell, C., & Silva, R. (2017). Counterfactual fairness. NIPS, 4066–4076. Lippert-Rasmussen, K. (2006). The badness of discrimination. Ethical Theory and Mora...

    work page 2022

Show all 9 references
  1. [3523]

    N., & Wu, X

    Carey, A. N., & Wu, X. (2022). The causal fairness field guide: Perspectives from social and formal sciences. Frontiers Big Data, 5, 892837. Cerda, P., & Varoquaux, G. (2022). Encoding high-cardinality string categorical variables. IEEE Trans. Knowl. Data Eng. , 34 (3), 1164–1...

    2022 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.