Pith. sign in

REVIEW 2 major objections 5 minor 37 references

A predictor's expected share of relative importance equals its reallocation row sum, and enforcing row sums of one removes a systematic bias.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Enforcing doubly stochastic reallocation (via Sinkhorn-Knopp) fixes the a priori bias in the GCD relative-importance measure, and the corrected GCD-SK approximates General Dominance better than GCD and often than Relative Weights.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection The row-sum characterization of the a priori bias is the real contribution; the doubly stochastic correction is sensible and effective in simulation, but its practical value depends on a prior the authors explicitly choose, and that should be debated in review. the 2 major comments →

arxiv 2607.13850 v1 pith:HG44LMFI submitted 2026-07-15 stat.ME

Bias Correction for Relative Importance Measures via Doubly Stochastic Reallocation

classification stat.ME MSC 62J0562H2062-08
keywords relative importance analysisgeneral dominanceorthogonalization-reallocation measuresa priori biasdoubly stochastic matricesSinkhorn-Knopp algorithmmulticollinearityvariance inflation factor
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tries to prove that a structural bias in a class of fast relative-importance measures is fully captured by the row sums of their reallocation matrices, and that erasing that row-sum imbalance removes the bias. It shows that, averaged over all possible responses (with coefficients uniform on the unit sphere), General Dominance gives every predictor an expected share of 1/p, while an orthogonalization-reallocation measure gives predictor i the row-sum_i/p. For the Green–Carroll–DeSarbo measure, those row sums are weighted averages of variance-inflation-factor ratios, so predictors with high VIFs are favored before any response is seen. The paper then maps the regression-based reallocation matrix to a doubly stochastic matrix, producing GCD-SK, and reports simulations in which this correction substantially improves on GCD and often beats Relative Weights when the first principal component dominates. A symmetric-correlation analysis is used to separate the bias from multicollinearity itself and to characterize Relative Weights' leveling problem as excess shrinkage.

Core claim

The central claim is that the a priori bias of an orthogonalization-reallocation measure is exactly the deviation of its reallocation matrix row sums from 1. Under the paper's response model, Proposition 1 shows the expected normalized importance of predictor i is row-sum_i / p, while General Dominance gives every predictor 1/p. Proposition 2 shows that for GCD, whose reallocation matrix is regression-based, each row sum is a weighted average of VIF ratios, so a predictor with the maximum VIF is over-credited on average and, because the matrix is column-stochastic, other predictors are under-credited. The paper corrects this by projecting the regression reallocation matrix onto the set of do

What carries the argument

The reallocation matrix A, whose entries a_ij assign importance from orthogonal predictor j back to original predictor i, and the identity that the expected normalized importance of predictor i is (1/p) times the i-th row sum. Under the minimal orthogonalization, the GCD reallocation matrix has entries gamma_ij^2 / VIF_j, making each row sum a weighted average of VIF ratios. Because this matrix is column-stochastic but not row-stochastic, the correction applies Sinkhorn–Knopp alternating normalization (or alternating Euclidean projections) to reach the nearest doubly stochastic matrix, thereby forcing all row sums to 1 while preserving the complete allocation of each orthogonal predictor's i

Load-bearing premise

The response coefficients are assumed to be uniformly distributed over all directions on the unit sphere, so 'a priori' means averaged over that uniform sphere; if a user's responses concentrate in a few directions, enforcing equal row sums may not correct the bias that actually matters.

What would settle it

Simulate or collect real responses where the coefficient vector is concentrated along the first principal component of the predictors, so the uniform-sphere assumption is violated; if GCD-SK's expected importance or rankings are no closer to General Dominance than uncorrected GCD's, then the uniform-direction prior does not capture the responses that matter.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A row-sum deviation from 1 is a complete diagnosis: any ORM whose reallocation matrix has unit row sums is free of the a priori bias under the paper's response model.
  • GCD-SK is a computationally cheap alternative to GD that the simulations show recovers much of the gap between uncorrected GCD and the GD benchmark, especially under severe multicollinearity.
  • In correlation structures with a dominant first principal component, where RW suffers from leveling, GCD-SK is the recommended bias-free choice; under mild multicollinearity with a weak first principal component, RW remains preferred.
  • Compound symmetry shows that homogeneous multicollinearity alone does not bias GCD, so practitioners should attribute the bias to heterogeneous VIFs rather than to multicollinearity per se.
  • RW's leveling problem is not an a priori bias: even a doubly stochastic matrix can shrink importance too strongly toward equality, so the two defects are structurally distinct.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same row-sum diagnostic could be applied to any column-stochastic importance attribution scheme, including Shapley-value approximations in explainable AI, to detect structural favoritism before data arrive.
  • One could test whether the same correction helps in high-dimensional settings where the number of predictors approaches or exceeds the sample size; the paper's simulations stop at p = 10.
  • The uniform-sphere prior is the crux: under response distributions concentrated on a few directions, a doubly stochastic matrix may not be the right target, and enforcing it could distort importance for the responses a user actually cares about.
  • Because the Sinkhorn–Knopp correction is multiplicative, it preserves relative structure better than Euclidean projection; a similar multiplicative correction might improve other proportional allocation rules.
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 / 5 minor

Summary. This paper studies a class of relative importance measures based on orthogonalization and reallocation (ORMs). It formalizes the 'a priori bias' under a response-generating model in which the systematic part of the response is y0 = z0'u0 + eps0 with u0 uniform on the unit sphere and x0 = Sigma^{1/2}z0, so beta0 = Sigma^{-1/2}u0. Proposition 1 shows that the expected normalized importance is 1/p for GD and row-sum_i/p for an ORM with reallocation matrix A. Proposition 2 shows that GCD's RegPA row sums are weighted averages of VIF ratios. Under compound symmetry, Proposition 3 gives closed forms for CorPA, GDA, and RegPA, all of which are doubly stochastic; Corollary 2 formalizes RW's leveling problem as shrinkage beyond the GDA benchmark. The paper corrects GCD by mapping RegPA to a doubly stochastic matrix using alternating projections (GCD-MAP) and the Sinkhorn-Knopp algorithm (GCD-SK). Simulations for p=3,...,10 with random correlation matrices and responses drawn from the same prior show that GCD-SK improves closeness to GD and often outperforms RW when the first principal component is dominant. The paper closes with empirical selection guidelines.

Significance. The theoretical contribution is substantial if the framework is accepted: a clean row-sum characterization of expected ORM behavior, a VIF-ratio explanation of GCD's bias, and a closed-form compound-symmetry analysis that separates a priori bias from multicollinearity and identifies RW leveling as shrinkage beyond the GD benchmark. The proofs in Appendix A are self-contained and, in my reading, correct. The paper also provides reproducible code via OSF, and no fitted parameters enter the central propositions. The main weakness is that the normative force of 'a priori bias' rests on the uniform-on-sphere prior in Eqs. (12)-(13), and the simulation evaluation draws responses from exactly that prior. Consequently, the external validity of the practical claims—especially that GCD-SK 'removes the structural row-sum bias'—is not established for other response-uncertainty models. The correction is a theorem under the stated prior but a heuristic under alternative priors.

major comments (2)
  1. [§3.1 (Eqs. 12-13) and §4.1 (Eq. 30)] The a priori bias is defined with respect to F induced by u0 ~ U(S^{p-1}), and the simulation evaluates ORMs under beta(b) drawn from the same F. Proposition 1 is correct conditional on this model, but the practical claim that GCD-SK 'removes the structural row-sum bias' and 'substantially improves' is only established under this prior. If a user's response uncertainty is not isotropic in z-space (e.g., beta has independent entries with unequal variances or is supported on a low-dimensional subspace), row-sum balance is no longer the expected-importance criterion, and the doubly stochastic mapping may distort importance for the responses actually encountered. Please add a sensitivity analysis over alternative beta priors, or explicitly restate the conclusions as conditional on the isotropic-uniform prior.
  2. [§4.3 (Eq. 30), Figs. 3 and 5] The performance evaluation uses closeness to GD as the metric, not row-sum balance. The theoretical result in Proposition 1 does not imply that a doubly stochastic RegPA is closer to GD in RMSE or Kendall's tau. The reported improvements of GCD-SK over GCD are therefore empirical, and the mechanism ('bias correction') is inferred rather than demonstrated. I recommend reporting how the improvement varies with the initial row-sum imbalance (e.g., max_i |r_i - 1| or sum_i (r_i - 1)^2) and comparing GCD-SK with other doubly stochastic projections, so that the causal interpretation is supported rather than assumed.
minor comments (5)
  1. [§2.2, Eq. (8)] Calling GDA 'theoretically ideal' overstates what is proven; it is a natural benchmark reallocation, but no optimality theorem is given.
  2. [§3.3] The statement that T=100 iterations is 'empirically sufficient' for MAP and SK would be stronger with a convergence criterion or diagnostic plot, especially since RegPA is not guaranteed to be strictly positive in all cases.
  3. [§3.4] The predictor correlation matrix Sigma used in the numerical illustration (Figure 1 and Table 2) is not specified in the text. Please include it or provide a pointer to the code so the illustration is reproducible.
  4. [Figures 3 and 5] The numeric labels in the stacked bars appear to be percentages, but the y-axis is labeled 'Proportion'. Please clarify the units and define the reported quantity precisely.
  5. [References] The reference to 'V on Neumann' should be 'von Neumann'.

Circularity Check

0 steps flagged

No significant circularity; the core Propositions are self-contained derivations, with only background self-citations and a prior-specific normative assumption.

full rationale

The central derivation is not circular. Proposition 1 (Appendix A.1) proves under the stated response model (Eqs. 12-13) that E[normalized GD] = 1/p and E[normalized ORM] = row-sum_i/p, via rotational invariance of the uniform-on-sphere prior. Proposition 2 (Appendix A.2) derives the VIF-ratio expression for RegPA row sums, and Proposition 3 with Corollaries 1-2 are closed-form compound-symmetry calculations. The proposed corrections (Section 3.3) are deterministic matrix-balancing algorithms (MAP/SK) applied to RegPA; they introduce no fitted parameters and are not optimized against the simulation metric, so the reported RMSE/Kendall improvements are not forced by the bias-removal objective. The simulation draws beta from the same prior F used to define the a priori bias, which is a legitimate consistency check but also a limitation: the normative force of row-sum balancing depends on uniform-on-sphere response uncertainty. Self-citations to Chang and Chen (2025) supply the ORM framework, the GDA benchmark, and simulation scaffolding (thresholds, win-loss design), but the load-bearing propositions and correction construction are proven or specified in this paper; no uniqueness theorem or ansatz is imported as an external fact. Therefore no load-bearing step reduces to its own input.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 0 invented entities

The proof apparatus uses only standard results; the only genuinely new postulate is the uniform-in-z response prior, which is the basis for the bias notion. No free parameters are fitted; simulation thresholds (lambda1/sqrt(p)=1.5, VIFmax/p=4) are descriptive diagnostics, not parameters of the proposed methods.

axioms (6)
  • ad hoc to paper Response model (12)-(13): y0 = z0'u0 + eps0, u0 ~ U(S^{p-1}), independent of noise; beta0 = Sigma^{-1/2}u0
    Defines the probability space in which 'a priori bias' is evaluated; no independent evidence that responses are uniform in this sense.
  • domain assumption Predictors and response are centered and scaled to unit L2 norm; X has full column rank (Section 2)
    Standard normalization in RI analysis; needed for covariance identities in the proofs.
  • domain assumption Johnson's minimal transformation Z = X(X'X)^{-1/2} is used as the default orthogonalization (Eq. 4)
    Adopted from the prior framework; results may differ under other orthogonalizations.
  • standard math Sinkhorn-Knopp convergence for strictly positive (or mildly nonnegative) square matrices (Section 3.3)
    Justifies existence and convergence of the SK correction; RegPA is generally strictly positive.
  • standard math Von Neumann alternating projections and Bregman projection properties (Section 3.3)
    Convergence of MAP relies on convex analysis results.
  • standard math Shapley value/GD equivalence (Eq. 1)
    GD is the Shapley value of the R^2 value function; used in Prop 1.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias Correction for Relative Importance Measures via Doubly Stochastic Reallocation." pith.science (2026). https://pith.science/paper/HG44LMFI

@misc{pith2026260713850,
  author       = {Pith},
  title        = {Pith review of: Bias Correction for Relative Importance Measures via Doubly Stochastic Reallocation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HG44LMFI}},
  note         = {Machine review of arXiv:2607.13850}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Relative importance (RI) analysis quantifies each predictor's contribution to the explained variance of a linear model. General Dominance (GD), a widely used benchmark, requires evaluating $2^p-1$ sub-models and becomes computationally intensive as the number of predictors $p$ grows. Orthogonalization-Reallocation Measures (ORMs), including Relative Weights (RW) and the Green--Carroll--DeSarbo measure (GCD), provide efficient alternatives by assigning importance to orthogonalized predictors and reallocating it to the original predictors. Each, however, has a structural limitation: RW exhibits a leveling problem that compresses differences among predictor importance values, whereas GCD exhibits an a priori bias that systematically favors certain predictors before a response is observed. We show that this bias is governed by the row-sums of the reallocation matrix. A closed-form analysis under compound symmetry relates the reallocations underlying GCD and RW to a GD-based benchmark, showing that homogeneous multicollinearity alone does not induce an a priori bias and formalizing RW's leveling problem as excess shrinkage relative to the benchmark. We correct GCD's bias by mapping its reallocation matrix to a doubly stochastic matrix using the Method of Alternating Projections (MAP) and the Sinkhorn--Knopp (SK) algorithm, yielding GCD-MAP and GCD-SK. Comprehensive simulations show that GCD-SK removes the structural row-sum bias, substantially improves upon GCD, and often outperforms RW when the first principal component is dominant. We conclude with empirical guidelines for selecting among the measures.

Figures

Figures reproduced from arXiv: 2607.13850 by Argon Chen, Tien-En Chang.

Figure 1
Figure 1. Figure 1: A p = 10 numerical illustration of the a priori bias in RegPA (GCD) and its correction via MAP and SK algorithms. 4 Simulations To evaluate the effectiveness of our proposed corrections for the a priori bias, we adopt the comprehensive Monte Carlo simulation framework developed by Chang and Chen [2025]. Their design is specifically aimed at assessing how well ORMs approximate the gold standard GD across a … view at source ↗
Figure 2
Figure 2. Figure 2: Performance of ORMs across predictor structures with mild multicollinearity and mild first principal [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Win–loss analysis between RW and GCD-SK under mild multicollinearity (Scenarios 1.1 and 1.2), based on [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance of ORMs across predictor structures with severe multicollinearity and mild first principal [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Win–loss analysis between RW and GCD-SK under severe multicollinearity (Scenarios 2.1 and 2.2), based on [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 1 linked inside Pith

  1. [1]

    Organizational research methods , year=

    History and use of relative importance indices in organizational research , author=. Organizational research methods , year=

  2. [2]

    Budescu, D. V. , title =. Psychological Bulletin , year =

  3. [3]

    , author=

    The dominance analysis approach for comparing predictors in multiple regression. , author=. Psychological methods , year=

  4. [4]

    Journal of Marketing Research , year=

    A new measure of predictor variable importance in multiple regression , author=. Journal of Marketing Research , year=

  5. [5]

    Psychometrika , year=

    The minimal transformation to orthonormality , author=. Psychometrika , year=

  6. [6]

    Multivariate behavioral research , year=

    A heuristic method for estimating the relative weight of predictor variables in multiple regression , author=. Multivariate behavioral research , year=

  7. [7]

    LeBreton, J. M. and Ployhart, R. E. and Ladd, R. T. , title =. Organizational Research Methods , year =

  8. [8]

    Chao, Y. C. and Zhao, Y. and Kupper, L. L. and Nylander-French, L. A. , title =. Journal of occupational and environmental hygiene , year =

  9. [9]

    Contributions to the Theory of Games (AM-28), Volume II , pages=

    A Value for n-Person Games , author=. Contributions to the Theory of Games (AM-28), Volume II , pages=. 1953 , publisher=

  10. [10]

    2025 , journal=

    Understanding and Using the Relative Importance Measures Based on Orthogonalization and Reallocation , author=. 2025 , journal=

  11. [11]

    Pattern Recognition , pages=

    Variable selection using relative importance rankings , author=. Pattern Recognition , pages=. 2026 , publisher=

  12. [12]

    , author=

    The accuracy of dominance analysis as a metric to assess relative importance: The joint impact of sampling error variance and measurement unreliability. , author=. Journal of Applied Psychology , year=

  13. [13]

    Multivariate behavioral research , year=

    On Johnson's (2000) relative weights method for assessing variable importance: A reanalysis , author=. Multivariate behavioral research , year=

  14. [14]

    Statistica Sinica , year=

    Decomposition of R 2 in multiple regression with correlated regressors , author=. Statistica Sinica , year=

  15. [15]

    Statistical Applications in Genetics and Molecular Biology , year=

    High-dimensional regression and variable selection using CAR scores , author=. Statistical Applications in Genetics and Molecular Biology , year=

  16. [16]

    Reduction theory , author=

    On rings of operators. Reduction theory , author=. Annals of Mathematics , year=

  17. [17]

    arXiv preprint arXiv:1309.1541 , year=

    Projection onto the probability simplex: An efficient algorithm with a simple proof, and an application , author=. arXiv preprint arXiv:1309.1541 , year=

  18. [18]

    Pacific Journal of Mathematics , year=

    Concerning nonnegative matrices and doubly stochastic matrices , author=. Pacific Journal of Mathematics , year=

  19. [19]

    2010 IEEE International Conference on Data Mining , year=

    Learning a bi-stochastic data similarity matrix , author=. 2010 IEEE International Conference on Data Mining , year=

  20. [20]

    The annals of mathematical statistics , year=

    Generalized iterative scaling for log-linear models , author=. The annals of mathematical statistics , year=

  21. [21]

    SIAM Journal on Matrix Analysis and Applications , year=

    The Sinkhorn--Knopp algorithm: convergence and applications , author=. SIAM Journal on Matrix Analysis and Applications , year=

  22. [22]

    Set-Valued Analysis , year=

    On the convergence of von Neumann's alternating projection algorithm for two sets , author=. Set-Valued Analysis , year=

  23. [23]

    SIAM review , year=

    On projection algorithms for solving convex feasibility problems , author=. SIAM review , year=

  24. [24]

    Wiley interdisciplinary reviews: Computational statistics , volume=

    Variable importance in regression models , author=. Wiley interdisciplinary reviews: Computational statistics , volume=

  25. [25]

    Organizational Research Methods , volume=

    Determining the relative importance of predictors in logistic regression: an extension of relative weight analysis , author=. Organizational Research Methods , volume=

  26. [26]

    , author=

    Multivariate relative importance: extending relative weight analysis to multivariate criterion spaces. , author=. Journal of Applied Psychology , volume=

  27. [27]

    Proceedings of the 2012 Winter Simulation Conference (WSC) , pages=

    Dominance index for many-to-many correlation and its applications to semiconductor yield analysis , author=. Proceedings of the 2012 Winter Simulation Conference (WSC) , pages=. 2012 , organization=

  28. [28]

    , author=

    Determining the statistical significance of relative weights. , author=. Psychological methods , volume=

  29. [29]

    Organizational Research Methods , volume=

    Residualized relative importance analysis: A technique for the comprehensive decomposition of variance in higher order regression models , author=. Organizational Research Methods , volume=

  30. [30]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Regression shrinkage and selection via the lasso , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

  31. [31]

    Knowledge-Based Systems , volume=

    Comprehensive relative importance analysis and its applications to high dimensional gene expression data analysis , author=. Knowledge-Based Systems , volume=

  32. [32]

    Advanced Engineering Informatics , volume=

    Many-to-many comprehensive relative importance analysis and its applications to analysis of semiconductor electrical testing parameters , author=. Advanced Engineering Informatics , volume=

  33. [33]

    Biometrika , volume=

    A new measure of rank correlation , author=. Biometrika , volume=

  34. [34]

    Johns Hopkins University, Tech

    Sampling uniformly from the unit simplex , author=. Johns Hopkins University, Tech. Rep , volume=

  35. [35]

    The American Statistician , year=

    Generating correlation matrices with specified eigenvalues using the method of alternating projections , author=. The American Statistician , year=

  36. [36]

    , author=

    The relative importance of task and contextual performance dimensions to supervisor judgments of overall performance. , author=. Journal of applied psychology , volume=

  37. [37]

    International Journal of Market Research , volume=

    Utilizing relative weight analysis in customer satisfaction research , author=. International Journal of Market Research , volume=

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.