Pith. sign in

REVIEW 2 major objections 1 minor 23 references

Direct and Indirect Discrimination in Generalized Linear Models

T0 review · 2 major / 1 minor · reviewed 2026-06-26 · grok-4.3

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

Pith's one-line read A moment decomposition of GLM predictions isolates four linear disparity channels plus two curvature terms from the inverse link.

desk verdict The paper supplies explicit moment decompositions for GLM fairness that add curvature coupling and amplification terms, but the whole thing rests on an untested extension of the Wasserstein criterion past the linear-Gaussian case. read the letter →

arxiv 2606.23116 v1 pith:V7GFCY2P submitted 2026-06-22 stat.ME

Bertille Tierny , Arthur Charpentier , François Hu This is my paper · ORCID
classification stat.ME
keywords generalizedlinearmodelsgroupdisparitiesdecompositionfairnessdiagnosticsinverselinkfunctionactuarialmodelinglogisticregressionPoisson
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 develops a moment-based framework that diagnoses sources of group differences in generalized linear model predictions. It begins with a Wasserstein barycentric criterion that reduces to a two-moment decomposition with direct mean, indirect mean, interaction and structural components in the linear-Gaussian case. The leading term for GLMs retains those four channels while adding curvature coupling and curvature amplification caused by the nonlinear inverse link. Explicit formulas are supplied for logistic, Poisson and Tweedie specifications, and the decomposition is applied to medical-expenditure survey data. The resulting diagnostic separates explicit sensitive effects, proxy-mediated profiles, covariance differences and link-induced nonlinearity.

What carries the argument

The leading decomposition D1(f) of the empirical output-scale criterion U2(f), which extends the reduced Wasserstein two-moment criterion by preserving linear channels and adding inverse-link curvature terms.

What would settle it

A numerical check in which the sum of the six decomposed components deviates materially from the observed difference in average GLM predictions between two groups on the same data.

Watch

Extended reading notes

Core claim

In the exact linear-Gaussian benchmark the Wasserstein barycentric criterion reduces to a two-moment criterion and decomposes into direct mean, indirect mean, interaction and structural components. For GLMs the leading decomposition D1(f) preserves the four linear channels and adds two curvature components induced by the inverse link: curvature coupling and curvature amplification. Explicit formulas are derived for logistic, Poisson and Tweedie specifications.

Load-bearing premise

The moment-based reduction of the Wasserstein barycentric criterion remains a sufficient diagnostic for group disparities once the nonlinear inverse link is present.

Editorial extensions

If this is right

  • Disparities in GLM predictions can be attributed separately to explicit group effects, proxy covariates, covariance structure and nonlinear link effects.
  • Explicit formulas allow direct computation of curvature coupling and amplification for logistic, Poisson and Tweedie models.
  • The decomposition supplies a tractable actuarial diagnostic on real datasets such as medical-expenditure surveys.

Reading between the lines

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

  • The same moment decomposition could be applied to other GLM families or to models with multiple inverse links to test whether curvature terms scale with variance.
  • Comparing the leading D1(f) term against full distributional parity measures on the same data would quantify how much information is lost by the moment reduction.
  • The framework suggests auditing fitted models by recomputing the decomposition after removing each sensitive or proxy variable in turn.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper develops a moment-based decomposition framework for group disparities in fitted GLM predictions. In the linear-Gaussian benchmark the Wasserstein barycentric criterion reduces to a two-moment criterion that decomposes into direct mean, indirect mean, interaction, and structural components. For GLMs the authors distinguish the empirical output-scale criterion U_2(f), a within-group proxy ilde U_2(f), and a leading decomposition D_1(f) that retains the four linear channels while adding two curvature components (coupling and amplification) induced by the inverse link; explicit formulas are supplied for logistic, Poisson, and Tweedie families and the diagnostic is illustrated on medical-expenditure survey data. The framework is presented as a tractable actuarial tool rather than a full distributional-parity characterization or legal test.

Significance. If the explicit formulas are correct and D_1(f) remains informative, the work supplies actuaries with closed-form expressions that separate direct, indirect, structural, and nonlinear-link contributions to prediction disparities in common GLM specifications. The provision of exact formulas for three families and the explicit acknowledgment of scope limitations are concrete strengths that could support practical adoption.

major comments (2)
  1. [Abstract / derivation of D_1(f)] Abstract and derivation of D_1(f): the central claim that the leading decomposition preserves the four linear channels and adds explicit curvature terms rests on an unverified reduction from the Wasserstein barycentric criterion once the nonlinear inverse link is present; the manuscript supplies no numerical checks, simulation studies, or comparisons against the full distributional criterion to confirm that omitted higher-order link effects do not dominate in typical GLM regimes.
  2. [Illustration section] Illustration section: the medical-expenditure example is described only at a high level; without reported values of U_2(f), ilde U_2(f), and the separate curvature components it is impossible to assess whether the added curvature terms materially alter the linear-channel decomposition in a real GLM fit.
minor comments (1)
  1. [Methods] Notation for U_2(f) and ilde U_2(f) is introduced in the abstract but the precise definitions and their relationship to the Wasserstein criterion should be restated at the beginning of the methods section for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract / derivation of D_1(f)] Abstract and derivation of D_1(f): the central claim that the leading decomposition preserves the four linear channels and adds explicit curvature terms rests on an unverified reduction from the Wasserstein barycentric criterion once the nonlinear inverse link is present; the manuscript supplies no numerical checks, simulation studies, or comparisons against the full distributional criterion to confirm that omitted higher-order link effects do not dominate in typical GLM regimes.

    Authors: The derivation of D_1(f) is obtained via a first-order Taylor expansion of the inverse link around the group means; this expansion exactly retains the four linear channels and isolates the two leading curvature terms by construction. The manuscript positions D_1(f) explicitly as a leading-term diagnostic rather than an exact match to the full Wasserstein criterion outside the linear-Gaussian case. We agree that a numerical check would strengthen the claim and will add a short simulation study comparing D_1(f) to the full distributional criterion under representative GLM regimes in the revision. revision: yes

  2. Referee: [Illustration section] Illustration section: the medical-expenditure example is described only at a high level; without reported values of U_2(f), ilde U_2(f), and the separate curvature components it is impossible to assess whether the added curvature terms materially alter the linear-channel decomposition in a real GLM fit.

    Authors: We agree that the illustration section would be more informative with explicit numerical values. The revised manuscript will report the computed values of U_2(f), ilde U_2(f), and the individual components of D_1(f) for the medical-expenditure data, together with a short discussion of the relative size of the curvature contributions. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: decomposition constructed from GLM structure and Wasserstein benchmark

full rationale

The paper starts from the Wasserstein barycentric criterion (reducing to moments only in linear-Gaussian case), then defines empirical output-scale U2(f), within-group proxy, and leading D1(f) by adding explicit curvature-coupling and curvature-amplification terms induced by the inverse link. Closed-form expressions for logistic, Poisson and Tweedie follow directly from standard GLM mean-variance and link properties rather than re-expressing any fitted parameter as a 'prediction' of itself. No self-citation chain, uniqueness theorem, or ansatz-smuggling is invoked; the abstract explicitly disclaims full distributional parity outside the benchmark, confirming the construction does not collapse to its inputs by definition.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The framework rests on the choice of Wasserstein barycentric criterion as the disparity measure and on standard GLM assumptions about the inverse link; no new free parameters or invented entities are introduced.

assumptions (1)
  • domain assumption Wasserstein barycentric criterion measures distributional demographic-parity violation
    Invoked as the starting point for the linear-Gaussian benchmark reduction to two moments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Direct and Indirect Discrimination in Generalized Linear Models." pith.science (2026). https://pith.science/paper/V7GFCY2P

@misc{pith2026260623116,
  author       = {Pith},
  title        = {Pith review of: Direct and Indirect Discrimination in Generalized Linear Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V7GFCY2P}},
  note         = {Machine review of arXiv:2606.23116}
}
abstract

Generalized linear models are central to actuarial modelling of binary risk, claim frequency, utilization, and cost-related outcomes. Yet fairness diagnostics often rely on linear-model intuitions, although GLM predictions are obtained by transporting a latent score through a nonlinear inverse link. We develop a moment-based decomposition framework for diagnosing group disparities in fitted GLM predictions. In an exact linear-Gaussian benchmark, the Wasserstein barycentric criterion for distributional demographic-parity violation reduces to a two-moment criterion and decomposes into direct mean, indirect mean, interaction, and structural components. For GLMs, we distinguish the empirical output-scale criterion $U_2(f)$, a within-group proxy $\widetilde U_2(f)$, and a leading decomposition $D_1(f)$. This leading term preserves the four linear channels and adds two curvature components induced by the inverse link: curvature coupling and curvature amplification. We derive explicit formulas for logistic, Poisson, and Tweedie specifications and illustrate the diagnostic on medical-expenditure survey data. The framework is not a legal test of discrimination, nor a full characterization of distributional parity outside the linear-Gaussian case. It is a tractable actuarial diagnostic for identifying whether fitted prediction disparities arise from explicit sensitive effects, proxy-mediated covariate profiles, covariance-structure differences, or nonlinear link effects.

Figures

Figures reproduced from arXiv: 2606.23116 by the authors.

Figure 1
Figure 1. Schematic linear decomposition. The group-average score gap is read through an [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Linear model: coefficient shifts from the base model to the moment-aligned model. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Schematic extension to generalized linear models. On the latent scale, group [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Logistic model: coefficient shifts from the base model to the latent moment-aligned [PITH_FULL_IMAGE:figures/full_fig_p034_4.png]
Figure 5
Figure 5. Figure 5: Poisson model: coefficient shifts from the base model to the latent moment-aligned [PITH_FULL_IMAGE:figures/full_fig_p042_5.png]
Figure 6
Figure 6. Figure 6: Poisson model: conditional density of predicted office-based visits ( [PITH_FULL_IMAGE:figures/full_fig_p043_6.png]
Figure 7
Figure 7. Figure 7: Tweedie log-link model: coefficient shifts from the base model to the latent moment [PITH_FULL_IMAGE:figures/full_fig_p052_7.png]
Figure 8
Figure 8. Figure 8: Tweedie log-link model: conditional density of predicted office-based visits ( [PITH_FULL_IMAGE:figures/full_fig_p053_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 1 canonical work pages

  1. [1]

    Fairness and risk: An ethical argument for a group fairness definition insurers can use

    Baumann, J., Loi, M., 2023. Fairness and risk: An ethical argument for a group fairness definition insurers can use. Philosophy & Technology 36, 45

    2023

  2. [2]

    Wage discrimination: Reduced form and structural estimates

    Blinder, A.S., 1973. Wage discrimination: Reduced form and structural estimates. The Journal of Human Resources 8, 436–455

    1973

  3. [3]

    Insurance, Biases, Discrimination and Fairness

    Charpentier, A., 2024. Insurance, Biases, Discrimination and Fairness. Springer Actuarial, Springer

    2024

  4. [4]

    A fair price to pay: Exploiting causal graphs for fairness in insurance

    Côté, O., Côté, M.P., Charpentier, A., 2025. A fair price to pay: Exploiting causal graphs for fairness in insurance. Journal of Risk and Insurance doi:10.1111/jori.12503

    work page doi:10.1111/jori.12503 2025

  5. [5]

    Generalized linear models for insurance data

    De Jong, P., Heller, G.Z., 2008. Generalized linear models for insurance data. Cambridge University Press

    2008

  6. [6]

    Effective Statistical Learning Methods for Actuaries I: GLMs and Extensions

    Denuit, M., Hainaut, D., Trufin, J., 2019. Effective Statistical Learning Methods for Actuaries I: GLMs and Extensions. Springer

    2019

  7. [7]

    Generalized Linear Models with Examples in R

    Dunn, P.K., Smyth, G.K., 2018. Generalized Linear Models with Examples in R. 2 ed., Springer. 60

    2018

  8. [8]

    An extension of the blinder-oaxaca decomposition technique to logit and probit models

    Fairlie, R.W., 2005. An extension of the blinder-oaxaca decomposition technique to logit and probit models. Journal of economic and social measurement 30, 305–316

    2005

Show all 23 references
  1. [9]

    Decomposition methods in eco- nomics, in: Ashenfelter, O., Card, D

    Fortin, N., Lemieux, T., Firpo, S., 2011. Decomposition methods in eco- nomics, in: Ashenfelter, O., Card, D. (Eds.), Handbook of Labor Economics. Elsevier. volume 4A, pp. 1–102

    2011

  2. [10]

    Regression modeling with actuarial and financial appli- cations

    Frees, E.W., 2009. Regression modeling with actuarial and financial appli- cations. Cambridge University Press

    2009

  3. [11]

    Predictive modeling applica- tions in actuarial science

    Frees, E.W., Derrig, R.A., Meyers, G., 2014. Predictive modeling applica- tions in actuarial science. volume 1. Cambridge University Press

    2014

  4. [12]

    Predictive modeling applica- tions in actuarial science

    Frees, E.W., Derrig, R.A., Meyers, G., 2016. Predictive modeling applica- tions in actuarial science. volume 2. Cambridge University Press

    2016

  5. [13]

    Demographic parity constrained minimax optimal regression under linear model, in: Advances in Neural Information Processing Systems, pp

    Fukuchi, K., Sakuma, J., 2023. Demographic parity constrained minimax optimal regression under linear model, in: Advances in Neural Information Processing Systems, pp. 8653–8689

    2023

  6. [14]

    Generalized Additive Models

    Hastie, T.J., Tibshirani, R.J., 1990. Generalized Additive Models. Chapman and Hall

    1990

  7. [15]

    The blinder–oaxaca decomposition for linear regression models

    Jann, B., 2008. The blinder–oaxaca decomposition for linear regression models. The Stata Journal 8, 453–479

    2008

  8. [16]

    Modern Actuarial Risk Theory: Using R

    Kaas, R., Goovaerts, M., Dhaene, J., Denuit, M., 2008. Modern Actuarial Risk Theory: Using R. Springer

    2008

  9. [17]

    Components of a difference between two rates

    Kitagawa, E.M., 1955. Components of a difference between two rates. Journal of the American Statistical Association 50, 1168–1194

    1955

  10. [18]

    Discrimination-free insurance pricing

    Lindholm, M., Richman, R., Tsanakas, A., Wüthrich, M.V., 2022. Discrimination-free insurance pricing. ASTIN Bulletin 52, 55–89. doi:10. 1017/S0515036121000234. 61

    2022

  11. [19]

    Generalized Linear Models

    McCullagh, P., Nelder, J.A., 1989. Generalized Linear Models. 2 ed., Chapman and Hall

    1989

  12. [20]

    Male-female wage differentials in urban labor markets

    Oaxaca, R., 1973. Male-female wage differentials in urban labor markets. International Economic Review 14, 693–709

    1973

  13. [21]

    Non-life insurance pricing with generalized linear models

    Ohlsson, E., Johansson, B., 2010. Non-life insurance pricing with generalized linear models. volume 174. Springer

    2010

  14. [22]

    Decomposing direct and indirect biases in linear models under demographic parity constraint, in: Proceedings of the AAAI Conference on Artificial Intelligence

    Tierny, B., Charpentier, A., Hu, F., 2026. Decomposing direct and indirect biases in linear models under demographic parity constraint, in: Proceedings of the AAAI Conference on Artificial Intelligence

    2026

  15. [23]

    Generalized Additive Models: An Introduction with R

    Wood, S.N., 2017. Generalized Additive Models: An Introduction with R. 2 ed., Chapman and Hall/CRC. Appendix A. Empirical Design and Reproducibility This appendix describes the empirical design used in Sections 2.5, 4.4, 5.5, and 6.4. The numerical results can be reproduced by...

    2017

Pith tools

Reviewed June 26, 2026 · model on record in the stance chip above.