Pith. sign in

REVIEW 3 major objections 6 minor 3 references

Falsifying Discriminant Validity of Predictive Algorithms

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read The paper claims that a statistical comparison of calibrated prediction losses can falsify whether an algorithm predicts an impermissible proxy like race at least as well as its intended outcomes; in the LSAC admissions model it flags race,

desk verdict The framing is useful but the calibrated-loss test is confounded by base rates; the LSAC race finding is a base-rate artifact rather than evidence of predicting race. read the letter →

arxiv 2601.17146 v2 pith:TKQYEAVU submitted 2026-01-23 stat.ME cs.CYcs.LGstat.ML

classification stat.MEcs.CYcs.LGstat.ML MSC 62F0362G10
keywords discriminantvalidityfalsificationframeworkproxyoutcomesPlattscalingnonparametrichypothesistestingconstructalgorithmicevaluationfairness
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

The paper argues that a key pre-deployment question — "is this algorithm predicting an impermissible quantity such as race, gender, or age instead of what we intended?" — can be answered with a principled statistical test. The proposed falsification framework formalizes discriminant validity: an algorithm passes when its calibrated prediction loss on intended (permissible) proxy outcomes is significantly lower than its loss on the impermissible proxy. Platt scaling puts losses for different outcomes on a common scale, and nonparametric tests (a paired t-test or Wilcoxon signed-rank test for one proxy; a conditional rank test for several) decide the comparison. In a law-school admissions model the test establishes discriminant validity for gender but fails for race, where the model predicts race better than GPA or bar passage; in the COMPAS recidivism setting it fails to establish validity with respect to age. The intended use is as an early, necessary-but-not-sufficient validity check that prompts rethinking of problem formulation when it fails.

What carries the argument

The central machinery is the pair of definitions loss discriminance (the algorithm's loss on the impermissible proxy exceeds its loss on all permissible proxies) and loss indiscriminance (loss on the impermissible proxy is no larger for at least one permissible proxy). The procedure rests on separate Platt scaling of the algorithm's predictions for each outcome — the paper argues this calibration is crucial so that losses are compared on a common calibrated scale — followed by nonparametric tests: a one-sided t-test or Wilcoxon signed-rank test on paired loss differences for a single permissible proxy, and, for multiple proxies, a conditional rank test whose statistic is the mean within-samp

What would settle it

Take the LSAC model, randomly permute its predictions (or replace them with a constant), and run Algorithm 1 with race as the impermissible proxy and GPA as permissible. If the test returns INDISCRIMINANT with race ranked first for roughly 93% of observations, the flag is driven by the white/93% base rate rather than by the model encoding race. Equivalently, compute the calibrated log loss of the constant majority-class predictor for each outcome: if the test declares race 'better predicted' than GPA solely from these floor losses, the comparison scale is base-rate-determined.

Watch

Extended reading notes

Core claim

The paper claims that discriminant validity of a predictive algorithm can be operationalized as a falsification procedure: compare per-outcome Platt-calibrated prediction losses and test whether the impermissible proxy's loss is higher than every permissible proxy's loss (loss discriminance). When the test fails to reject, the result is labelled INDISCRIMINANT — an inconclusive safety flag that should prompt re-examination of problem formulation, not model tuning. In the LSAC admissions analysis the model passes with respect to gender (p≈0) and fails with respect to race (p≈1, with race best-predicted for 91% of applicants); in COMPAS the decile scores fail with respect to age (p≈1) and give

Load-bearing premise

The test assumes that after Platt scaling, losses for outcomes with very different base rates are directly comparable, so that a lower calibrated loss genuinely means better prediction; for heavily skewed outcomes the loss floor itself is much lower, which can make an uninformative model appear to 'predict' the skewed impermissible proxy well.

Editorial extensions

If this is right

  • An INDISCRIMINANT result is an actionable safety flag: developers should revisit problem formulation, proxy choice, and task suitability rather than tune the model further.
  • The method works without access to the true target outcome, using only permissible proxy outcomes and the measured impermissible variable, and makes minimal distributional assumptions.
  • When several impermissible proxies are evaluated, standard multiple-testing corrections (Bonferroni, Holm) apply; in COMPAS the race result is borderline once corrected.
  • Standard performance metrics like AUC can be ambiguous about discriminant validity (in COMPAS, AUC for re-arrest and age are nearly equal); the loss-comparison test resolves that ambiguity statistically.
  • Passing the test is necessary but not sufficient for validity: convergent and external validity still require complementary analysis.

Reading between the lines

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

  • The calibration step may make the test sensitive to outcome base rates: for a Bernoulli outcome the achievable loss floor is its entropy, so a proxy with a 93% majority class (like 'white' in LSAC) can show artificially low calibrated loss even for an uninformative predictor. The race finding may partly reflect this base-rate asymmetry rather than the model genuinely encoding race; the appendix's
  • A concrete testable extension would be to run the same procedure on a null model (e.g., predictions permuted or constant majority-class predictions). If the null model also fails discriminant validity for skewed impermissible proxies, a base-rate-normalized loss, or a calibration that maps scores to a common reference distribution, would be needed before the test can support causal claims about wh
  • The framework maps naturally onto negative controls in causal inference, with the impermissible proxy playing the role of a negative control outcome; combining it with selection-bias adjustments (as needed in COMPAS, where only released defendants are observed) would strengthen its external-validity claims.
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

3 major / 6 minor

Summary. The paper proposes a falsification framework for assessing the discriminant validity of predictive algorithms. The claim is that by comparing Platt-calibrated prediction losses on permissible proxies (e.g., GPA) against losses on an impermissible proxy (e.g., race, gender), a nonparametric hypothesis test can determine whether the algorithm predicts the intended outcome better than the impermissible one. A single-proxy version uses a one-sided t-test or Wilcoxon signed-rank test on paired loss differences, while a multiple-proxy version uses a conditional rank/permutation test. The method is illustrated on LSAC admissions data (race fails the test, gender passes) and on COMPAS recidivism data (age fails), with additional negative-control and ablation analyses in the appendix.

Significance. If the proposed test were valid, it would provide a concrete, operational link between psychometric discriminant validity and algorithmic evaluation, which is a valuable and timely contribution. The paper usefully synthesizes placebo tests, negative controls, and discriminant-validity practices from other fields, and its explicit discussion of the inverse fallacy and the inconclusiveness of failing to reject is a strength. The negative control in Appendix B is good empirical practice. However, the central methodological claim — that separate Platt scaling makes calibrated losses comparable across outcomes with different marginal distributions — is incorrect, and the headline race finding is an artifact of base-rate entropy. This invalidates the core contribution as presented.

major comments (3)
  1. [§5.1, Algorithm 1, Table 3/A.1] The claim that separate Platt scaling 'ensures that the loss comparisons between different proxies are made on a common, calibrated scale' is incorrect. For a binary outcome with base rate p, the minimum achievable expected log loss is the entropy H(p), attained by a perfectly calibrated constant predictor. After Platt scaling, a model with zero information about race outputs ~0.93 probabilities, giving log loss ≈ H(0.93) ≈ 0.25 nats, whereas a balanced GPA proxy has a floor of H(0.5) ≈ 0.69 nats. The race loss is therefore lower for nearly every observation regardless of whether the model encodes any race information. This is exactly what Table 3 shows: the race result flips from mean difference −0.03 (DISCRIMINANT) to +0.50 (INDISCRIMINANT) solely because calibration is applied. The appendix dismisses the uncalibrated comparison as an artifact, but the calibrated comparison is equally
  2. [§6.1.2, Fig. 1 (right)] The headline finding — race is best-predicted (rank 1) for 91% of observations — is reproduced by a null model with no race information. Under Platt scaling, the calibrated constant predictor for race achieves loss H(0.93) for every row, while the GPA proxies have floors around H(0.5); thus race attains rank 1 for roughly 93% of rows. The p≈1 result is therefore not evidence that the algorithm 'predicts race better than permissible outcomes'; it is a base-rate artifact. The test as implemented cannot distinguish 'impermissible predicted better' from 'impermissible outcome has a lower entropy floor,' and the paper's interpretation of this p-value is not supported.
  3. [§5.2 and Definition 4.2] The conditional rank test's null hypothesis of loss exchangeability is not aligned with the paper's notion of loss indiscriminance. Definition 4.2 declares indiscriminance when l(Ŷ, Ỹ) ≤ l(Ŷ, Yj) for some j. But with log loss, that inequality is satisfied by a constant predictor whenever H(p̃) ≤ H(p_j), independent of any property of the algorithm. The exchangeability assumption used in Algorithm 2 therefore does not correspond to 'the algorithm fails to predict the impermissible outcome'; it corresponds to equality of the complete loss distributions, which is false at baseline whenever outcome marginals differ. The permutation p-values are not calibrated for the intended null.
minor comments (6)
  1. [Terminology] Inconsistent terminology: Definition 4.2 uses INDISCRIMINANT, while Algorithm 1's Ensure and step 17 use INDISCRIMINATE (inconclusive). Please harmonize.
  2. [Appendix A.1] Appendix A.1 states the race base rate is 94%, while the main text §6.1 states 93%. Please reconcile the discrepancy.
  3. [Abstract] The phrase 'the algorithm is predicting another outcome than what was intended' is ungrammatical; suggest 'an outcome other than the intended one.'
  4. [Figure 1] The caption says the dashed line shows the expected 25% under the null, but the null expected rank for M=3 is 2.5, not a 25% line. Please clarify what the dashed line represents.
  5. [Title] The full-text title is 'Falsifying Predictive Algorithms' while the arXiv title is 'Falsifying Discriminant Validity of Predictive Algorithms'; ensure consistency.
  6. [Tables 3 and 4] P-values are reported as ≈1 and ≈0. Reporting exact values or confidence intervals would be more informative.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: the falsification test is held-out and self-contained; the base-rate critique is a validity concern, not a circular reduction.

full rationale

The paper's derivation chain is a standard statistical testing construction: Def. 4.1 defines loss discriminance via loss comparisons; Def. 4.3 defines a falsification procedure as control of the probability of returning INDISCRIMINANT under indiscriminance; Alg. 1/2 instantiate this with paired t/Wilcoxon and conditional-rank/permutation tests. The only near-tautological element is Def. 4.3, which restates Type I error control; this is a definition rather than a claim derived from the target, so it does not constitute a circular step. The empirical results in §6 are held-out applications: the model is trained on a training split, Platt scaling is fit on a calibration split, and losses are computed on an evaluation split (§6.1), so the race/gender findings are not fitted parameters renamed as predictions. The self-citations (Coston et al. 2023; Guerdan et al. 2023; Coston et al. 2020; Rambachan et al. 2022) motivate framing and limitations but are not load-bearing: no uniqueness theorem or ansatz is imported from them. The serious base-rate/entropy problem with comparing calibrated losses across outcomes with different marginal distributions is a substantive validity threat, not circularity; the paper itself notes in Appendix A.1 that the raw result without calibration is 'an artifact of comparing losses computed from uncalibrated predictions against labels with different base rates (race: 94%, GPA: 50%)', and it flags that INDISCRIMINANT is inconclusive (§5.1, §6.2.1) and that the COMPAS sample has external-validity limits (§6.2). Those acknowledgments support an honest non-finding on circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new entities; its machinery is standard calibration and hypothesis testing. The central dependency is the (unjustified) comparability of losses across outcomes with different base rates.

free parameters (3)
  • Platt scaling coefficients A_y, B_y per outcome = not reported
    Fit by logistic regression on a held-out calibration set for each outcome (Algorithms 1-2); the test statistic is computed from these calibrated losses, so the test result depends on these fitted parameters.
  • Proxy binarization thresholds = race = white vs non-white; age < 25; high income = category 5; GPA above median
    Chosen by hand in §6 and Appendix C.2; results depend on these definitions, especially for family income (92% of observations show lower loss than GPA after binarization).
  • Loss function = log loss vs Brier score
    Two options presented; the paper reports qualitative robustness (Appendix A.2), but the choice affects the test statistic.
assumptions (4)
  • domain assumption Permissible proxies are valid measures of the intended construct
    The framework's interpretation of DISCRIMINANT/INDISCRIMINANT presupposes that outcomes like first-year GPA or re-arrest are acceptable proxies for the target. The paper itself flags this for COMPAS (§6.2).
  • domain assumption Platt scaling produces well-calibrated probabilities for each outcome
    Assumes the logit of P(y=1|f(x)) is linear in f(x); used in Algorithms 1-2, step 2.
  • ad hoc to paper Losses are comparable across outcomes after calibration
    The paper asserts this in §5.1 but does not justify it for outcomes with different base rates; the review analysis shows the assertion is false (a null predictor achieves lower loss for skewed outcomes).
  • standard math Standard paired-test and permutation-test validity conditions
    i.i.d. samples, exchangeability under the null for the permutation test, etc.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Falsifying Discriminant Validity of Predictive Algorithms." pith.science (2026). https://pith.science/paper/TKQYEAVU

@misc{pith2026260117146,
  author       = {Pith},
  title        = {Pith review of: Falsifying Discriminant Validity of Predictive Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKQYEAVU}},
  note         = {Machine review of arXiv:2601.17146}
}
read the original abstract

Empirical investigations into unintended model behavior often show that the algorithm is predicting another outcome than what was intended. These expos\'es highlight the need to identify when algorithms predict unintended quantities - ideally before deploying them into consequential settings. We propose a falsification framework that provides a principled statistical test for discriminant validity: the requirement that an algorithm predict intended outcomes better than impermissible ones. Drawing on falsification practices from causal inference, econometrics, and psychometrics, our framework compares calibrated prediction losses across outcomes to assess whether the algorithm exhibits discriminant validity with respect to a specified impermissible proxy. In settings where the target outcome is difficult to observe, multiple permissible proxy outcomes may be available; our framework accommodates both this setting and the case with a single permissible proxy. Throughout we use nonparametric hypothesis testing methods that make minimal assumptions on the data-generating process. We illustrate the method in an admissions setting, where the framework establishes discriminant validity with respect to gender but fails to establish discriminant validity with respect to race. This demonstrates how falsification can serve as an early validity check. We also provide analysis in a criminal justice setting, where we highlight the limitations of our framework and emphasize the need for complementary approaches to assess other aspects of construct validity and external validity.

Figures

Figures reproduced from arXiv: 2601.17146 by the authors.

Figure 1
Figure 1. Rank distribution of impermissible proxies (gender and race) for Alg. 2 on the LSAC [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Paired difference distributions for Alg. 1 on COMPAS with age (left) and race (right) as [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

3 extracted references · 1 linked inside Pith

  1. [1991]

    Sally Ho and Garance Burke

    doi: 10.1007/BF00960446. Sally Ho and Garance Burke. Oregon dropping AI tool used in child abuse cases. Associated Press, June 2 2022. URLhttps://apnews.com/article/ politics-technology-pennsylvania-child-abuse-1ea160dc5c2c203fdab456e3c2d97930. Accessed: 2026-01-08. Sally Ho and Garance Burke. Justice department scrutinizes Pittsburgh child welfare AI too...

  2. [2020]

    impermissible

    doi: 10.1056/NEJMms2004740. Angelina Wang, Sayash Kapoor, Solon Barocas, and Arvind Narayanan. Against predictive op- timization: On the legitimacy of decision-making algorithms that optimize predictive accuracy. ACM Journal on Responsible Computing, 1(1):1–45, 2024. Jamelle Watson-Daniels, Solon Barocas, Jake M Hofman, and Alexandra Chouldechova. Multi- ...

  3. [2021]

    Vic Barnett, Toby Lewis, et al.Outliers in statistical data, volume 3

    URLhttps://arxiv.org/abs/2106.05498. Vic Barnett, Toby Lewis, et al.Outliers in statistical data, volume 3. Wiley New York, 1994. Solon Barocas and Andrew D Selbst. Big data’s disparate impact.Calif. L. Rev., 104:671, 2016. Solon Barocas, Moritz Hardt, and Arvind Narayanan.Fairness and Machine Learning: Limitations and Opportunities. MIT Press, Cambridge,...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.