Graph-theoretic Inference for Random Effects in High-dimensional Studies

Lynna Chu; Yichuan Bai

REVIEW 3 major objections 5 minor 16 references

Graph-theoretic Inference for Random Effects in High-dimensional Studies

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a model-free, rank-based graph test for random effects that works with high-dimensional fixed effects, deriving its asymptotic null distribution and consistency.

desk verdict A clever Friedman–Rafsky adaptation that is a valid exchangeability test but not a test of Ψ=0; the central claim needs reframing before it can be published. read the letter →

arxiv 2506.07946 v1 pith:BZDCFK42 submitted 2025-06-09 stat.ME

Lynna Chu , Yichuan Bai This is my paper

classification stat.ME MSC 62G1062H1562J10

keywords randomeffectstestinghigh-dimensionalinferencegraph-basedtestpermutationnullrank-basedstatisticmixedmodelsmodel-free

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 is trying to establish a way to test whether random effects are present in a mixed model without needing to specify or fit the model, even when the number of fixed-effect covariates exceeds the sample size ($p>N$). The proposed test builds a similarity graph on the covariate vectors $X$, ranks observations by their distances in the response $Y$, and asks whether pairs that share a cluster label tend to be close in both spaces. If no random effect exists, the cluster labels should be exchangeable, and the paper proves the test statistic is asymptotically standard normal under that permutation null. If a random effect exists, the test is consistent under alternatives in which within-cluster ranks are bounded. A correct test of this kind would let practitioners check for clustering correlation in high-dimensional, possibly non-linear settings where conventional mixed-model inference is unreliable.

What carries the argument

The central object is the within-cluster edge-rank statistic $T_s = \sum_{(i,j)\in G_X} R_i(j)\mathbf{1}\{(X_i, X_j)\in C_s\}$, where $R_i(j)$ is the normalized rank of $Y_j$ among all observations ordered by distance from $Y_i$, and $G_X$ is a similarity graph (e.g., $k$-MST or $k$-NN) built on the covariate vectors. The test statistic $Z_I$ standardizes the summed $T_s$ under the permutation null; the machinery that carries the argument is the exact combinatorial computation of the moments of the $T_s$ over random permutations of cluster labels, which yields both the asymptotic normality and the consistency proof.

What would settle it

Generate data with zero random-effect variance but cluster-specific distributions for $X$ (for example, each cluster's covariates have a different mean) and a response $Y = X\beta + \varepsilon$; if the test rejects at the nominal $0.05$ level in substantially more than $5\%$ of trials, the exchangeability assumption underlying the null is violated. A weaker check is to verify whether conditions (2)--(4) of Theorem 3.1 hold for a $k$-NN graph on standard Gaussian data, since the proof depends on those constraints.

Watch

Extended reading notes

Core claim

The central claim is that the presence of a random effect is detectable through the within-cluster edge-rank sum $V_I = \sum_{s} T_s$, where $T_s$ sums normalized ranks $R_i(j)$ over edges of a graph $G_X$ whose two endpoints both lie in cluster $s$. Under the null of no random effect, the cluster labels are exchangeable given $X$ and $Y$, and the paper shows the standardized statistic $Z_I = (V_I - E(V_I))/\sqrt{\mathrm{var}(V_I)}$ converges in distribution to $N(0,1)$ under conditions on the graph density and rank concentration (Theorem 3.1). Under fixed alternatives for which within-cluster ranks are bounded, the test rejects with probability tending to one (Theorem 3.2). The paper also supplies exact combinatorial expressions for the expectation, variance, and covariance of the $T_s$ under permutation, and demonstrates through simulations that the asymptotic null distribution controls type I error while a recently proposed $F$-test does not.

Load-bearing premise

The test's null distribution assumes that, when there is no random effect, cluster labels can be permuted without changing the joint distribution of the data; that does not follow from the variance of the random effect being zero, so the test can reject for reasons other than a random effect if clusters differ in ways not captured by $X$.

Editorial extensions

If this is right

Researchers can test for a random block, batch, or subject effect in high-dimensional regression without estimating potentially many fixed effects.
The test remains usable when the relationship between $X$ and $Y$ is non-linear, because it uses only ranks and graph proximity.
The permutation-based version gives exact small-sample calibration for moderate sample sizes, while the asymptotic version provides a fast approximation.
The method offers a diagnostic for whether to add a random effect to a model, complementing model-selection procedures that require sparsity assumptions.

Reading between the lines

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

The same statistic could serve as a broader test of whether cluster labels carry information about $Y$ beyond what $X$ explains, which is not limited to variance-component testing; it may also detect cluster-specific mean shifts.
Because the test uses only pairwise distances and ranks, it could extend naturally to non-Euclidean similarity measures or incomplete data, though the paper does not develop this.
The graph-density condition $|G_X| = O(N^\alpha)$ with $\alpha < 1.5$ implies a tradeoff between power and theoretical validity; users may need to check that a chosen $k$ keeps the graph sparse enough.
A natural next step the paper leaves open is a power or sample-size calculation under a parametric alternative, which would make the test useful for study design.

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.

Desk Editor's Note

read the letter

Short version: this paper builds a Friedman–Rafsky-style within-cluster edge-rank statistic, derives exact permutation moments and a normal null, and shows in simulations that it controls size and has decent power in high-dimensional and nonlinear settings. That part is real and reasonably careful. The problem is the claim that this tests H0: Ψ=0. The permutation null it analyzes is exchangeability of cluster labels given X and Y. Ψ=0 does not give you that; fixed cluster means or cluster-dependent X distributions will produce exactly the within-cluster rank compression the test looks for, with zero random-effect variance. The paper itself states the exchangeability assumption in Section 2.1, so the gap is out in the open—it is not a hidden mathematical error, it is a mislabeled null hypothesis. Theorem 3.1 correctly gives the null distribution for the permutation procedure; it just does not deliver a test of the mixed-model null. Theorem 3.2 is weaker than it looks because bounded within-cluster ranks are effectively part of the alternative condition.

What is genuinely new: the adaptation of the graph-based association statistic to cluster-label permutation, the moment formulas in Lemma 2.1, and the asymptotic justification under conditions. The simulations are honest and include a useful comparison showing Law–Ritov's F-test can badly over-reject when β is not sparse. The real-data application is illustrative rather than decisive.

The soft spots beyond the null gap: Conditions (2)–(4) of Theorem 3.1 are unverified for k-MST/k-NN, so the asymptotic null is conditional on unproven graph behavior; the proofs are in appendices that aren't in the supplied text. The consistency theorem assumes the signal it is trying to detect. The choice of k is left open, which is a practical limitation the authors acknowledge.

My bottom line: as written, the central scientific claim is not supported, so a strict reject is defensible. But the underlying procedure is a valid, useful exchangeability test, and the authors are one reframing away from a solid paper. I would send this to a serious referee, with the expectation of heavy revision: change the stated null to cluster-label exchangeability (or explicitly add conditions under which Ψ=0 implies it), verify or weaken the graph conditions, and provide the appendix proofs. The paper deserves referee time because the construction is new and the simulations are informative.

Referee Report

3 major / 5 minor

Summary. The paper proposes a graph-theoretic, rank-based test for the null hypothesis H0: Ψ=0, where Ψ is the covariance matrix of random effects in a mixed model with potentially high-dimensional fixed effects. The test statistic V_I sums, over all clusters, the normalized ranks of within-cluster edges in a similarity graph G_X constructed on the covariate vectors; small values of the standardized statistic Z_I are taken as evidence of a random effect. Lemma 2.1 gives combinatorial expressions for the mean and variance of V_I under a permutation null in which cluster labels are randomly shuffled. Theorem 3.1 claims asymptotic standard normality of Z_I under this permutation null, and Theorem 3.2 claims consistency under fixed alternatives with bounded within-cluster ranks. The paper reports simulation studies under linear and non-linear mixed models and applies the procedure to a sorghum data set. The central claim is that the procedure tests whether the random-effect variance is zero without estimating the high-dimensional fixed effects or specifying the model correctly.

Significance. If the central claim were correct, the paper would provide a genuinely model-free test for random effects in high-dimensional settings, with the attractive feature that no estimation of the fixed-effects coefficients is required. The combinatorial moment derivation in Lemma 2.1 is a useful contribution, and the analytic null distribution in Theorem 3.1, if applicable, would avoid expensive permutation computations. The null simulations with iid covariate vectors across clusters show calibration close to the nominal level. However, the paper's central identification claim is not supported: the test is calibrated for cluster-label exchangeability, which is strictly stronger than Ψ=0, and the consistency theorem assumes essentially the signal the test is meant to detect. These issues are load-bearing because they concern what hypothesis the procedure actually tests. The paper also leaves the sufficient conditions of Theorem 3.1 unverified for standard similarity graphs. As a result, the contribution as a test for random effects is not established.

major comments (3)

[Section 2.1] The paper states: 'Under H0, there is no random effect and the cluster labels are exchangeable.' This implication is false. Ψ=0 does not imply that cluster labels are exchangeable given X and Y. For example, under Y_i = X_iβ + μ_{s(i)} + ε_i with fixed cluster-specific means μ_s and Ψ=0, the cluster labels carry information about Y and the within-cluster edge ranks will be systematically small, so Z_I will reject even though the random-effect variance is zero. Since Theorem 3.1 derives the null distribution under the permutation null, the rejection region {Z_I ≤ z_α} is calibrated for exchangeability, not for H0: Ψ=0. This is a load-bearing identification error. The authors should either state and justify an additional condition under which H0: Ψ=0 implies exchangeability, or reframe the test as a test of cluster exchangeability rather than of zero random-effect variance.
[Theorem 3.2] The consistency result assumes that, under the alternative, the within-cluster ranks are bounded above by constants γ_s. This is essentially the property the test statistic measures and is not derived from Ψ≠0; fixed cluster means also produce bounded within-cluster ranks. Consequently, Theorem 3.2 does not establish power specifically against random-effect alternatives, and it is consistent with the identification problem described above. Please provide verifiable conditions on the data-generating process (for example, on Ψ and the design) under which the bounded-rank condition holds, or weaken the consistency claim accordingly.
[Theorem 3.1, conditions (2)–(4)] The conditions on the similarity graph G_X are stated as sufficient for asymptotic normality, but the paper does not show that any standard graph construction, such as k-MST or k-NN, satisfies them for typical data distributions. Condition (1) holds for a k-MST with α=1, but conditions (2)–(4) constrain the behavior of rank sums over graph neighborhoods and are neither verified analytically nor checked numerically. Without such verification, the asymptotic null distribution has no demonstrated domain of applicability beyond the specific simulation settings. Please provide examples, a verification lemma for common graphs, or a characterization of data-generating regimes under which these conditions hold.

minor comments (5)

[Section 2.1, equation for T_s] The definition of T_s contains a double sum over i twice; it should be a sum over i and j, i.e., T_s = Σ_i Σ_j a_{ij} R_i(j) c_{ij}(s).
[Table 2] The SNR=4 rows in Table 2 are identical to the SNR=4 rows in Table 1, even though Table 2 is intended to report results under t_3 errors. This appears to be a copy-paste error and should be corrected, as it affects the credibility of the non-normal calibration results.
[Section 5] The word 'sorgum' should be 'sorghum'.
[Theorem 3.1] The theorem statement writes 'ZI' without a subscript; for consistency with the rest of the paper, it should be Z_I.
[Data application] The paper does not provide code or data for the sorghum application, so the empirical results are not reproducible from the manuscript alone.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the statistic, permutation moments, and asymptotic null distribution are externally defined; no parameter is fitted to the outcome and no load-bearing self-citation appears.

full rationale

The claimed derivation is self-contained. The test statistic Z_I is built from the observed graph G_X on X and normalized Y-ranks; no parameter is fitted to the outcome, so there is no fitted-input-called-prediction step. Lemma 2.1 derives E(T_s), var(T_s), and cov(T_s,T_t) combinatorially under the permutation null, and Theorem 3.1 gives asymptotic normality under explicit graph/rank conditions (1)-(4) stated independently of the test decision; simulations and the Law-Ritov comparison provide external checks. The citation to Friedman and Rafsky (1983) supplies the base statistic, not a load-bearing self-citation, and the authors have no prior-work self-citations in the derivation chain. Theorem 3.2's assumption of bounded within-cluster ranks is a sufficient-condition hypothesis on the alternative, not an equivalent restatement of the conclusion; it is the usual kind of alternative-direction condition in rank-test power proofs and does not define the test or its null distribution. The more serious issue—that H0: Psi = 0 does not in general imply exchangeability of cluster labels—is a scientific identification and validity concern, not a circularity of the derivation chain, so it does not raise the circularity score. Overall, no significant circularity is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted to the outcome; the tuning constant k is selected by hand. The main burden is the exchangeability assumption and the unverified technical graph conditions.

free parameters (1)

k (number of edges in k-MST similarity graph) = 20 in power simulations; 30 in data application
Graph density is chosen by the analyst, and the paper notes the choice is application-dependent and that an optimal choice remains open; power and validity depend on it.

assumptions (3)

domain assumption Cluster-label exchangeability under H0: permuting cluster labels preserves the joint distribution of (X, Y, labels).
Section 2.1 states under H0 'cluster labels are exchangeable'. This is not implied by Psi = 0; fixed cluster-level means or cluster-specific X distributions break it, so the test can reject under the null.
ad hoc to paper Theorem 3.1 conditions (1)-(4) hold for the chosen similarity graph and rank configuration.
The paper calls them sufficient and discusses (1) for k-MST, but does not verify (2)-(4) theoretically or empirically before applying the normal approximation.
ad hoc to paper Within-cluster ranks are bounded by constants gamma_s under the alternative.
Theorem 3.2's consistency is stated only under this alternative condition, which is close to the signal the test statistic itself measures and is not derived from Psi != 0.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Graph-theoretic Inference for Random Effects in High-dimensional Studies." pith.science (2026). https://pith.science/paper/BZDCFK42

@misc{pith2026250607946,
  author       = {Pith},
  title        = {Pith review of: Graph-theoretic Inference for Random Effects in High-dimensional Studies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZDCFK42}},
  note         = {Machine review of arXiv:2506.07946}
}

abstract

We study the problem of testing for the presence of random effects in mixed models with high-dimensional fixed effects. To this end, we propose a rank-based graph-theoretic approach to test whether a collection of random effects is zero. Our approach is non-parametric and model-free in the sense that we not require correct specification of the mixed model nor estimation of unknown parameters. Instead, the test statistic evaluates whether incorporating group-level correlation meaningfully improves the ability of a potentially high-dimensional covariate vector $X$ to predict a response variable $Y$. We establish the consistency of the proposed test and derive its asymptotic null distribution. Through simulation studies and a real data application, we demonstrate the practical effectiveness of the proposed test.

Figures

Figures reproduced from arXiv: 2506.07946 by the authors.

**Figure 1.** GX constructed on the X ∈ R N×p space, such that N = 60, s = 1, . . . , 4, and p = 2 . Here GX is an MST using Euclidean distance; each Xi is a node in the graph and its color represents its cluster label. The solid lines represent edges connecting within clusters. Blue solid edges correspond to C1, purple solid edges to C2, red solid edges to C3, and green solid edges to C4. The dashed black lines represent edges c… view at source ↗

**Figure 2.** Boxplots of Ts and PI s=1 Ts for four clusters under the null (τ 2 = 0) and alternative (τ 2 ̸= 0) hypothesis over 100 trials. Under H0, there is no random effect and the cluster labels are exchangeable, with equal probability given to all possible N n1,n2,...,nI shuffling of the cluster labels. We refer to this as the permutation null distribution. If the cluster labels are indeed exchangeable, then our observed … view at source ↗

**Figure 3.** Heatmap of distances between observations, subsetted by tissue type (leaf or root) and treatment type (drought or well-watered). Each black square represents the pairwise distances within the same block. The X and Y axis labels indicate the block to which the observations belongs to. We also test for both block and genotype as random effects. The resulting p-value is less than 0.001, indicating strong evidence that … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 14 canonical work pages

[1]

N., Qi, Y., Goren, E., Chiniquy, D., Sheflin, A

Chai, Y. N., Qi, Y., Goren, E., Chiniquy, D., Sheflin, A. M., Tringe, S. G., Prenni, J. E., Liu, P., and Schachtman, D. P. (2024). Root-associated bacterial communities and root metabolite composition are linked to nitrogen use efficiency in sorghum. Msystems , 9(1):e01190--23

work page 2024
[2]

Chen, F., Li, Z., Shi, L., and Zhu, L. (2015). Inference for mixed models of ANOVA type with high-dimensional data. Journal of Multivariate Analysis , 133:382--401

work page 2015
[3]

Cui, S., Ji, T., Li, J., Cheng, J., and Qiu, J. (2016). What if we ignore the random effects when analyzing RNA -seq data in a multifactor experiment. Statistical Applications in Genetics and Molecular Biology , 15(2):87--105

work page 2016
[4]

Ekvall, K. O. and Bottai, M. (2022). Confidence regions near singular information and boundary points with applications to mixed models. The Annals of Statistics , 50(3):1806--1832. Publisher: Institute of Mathematical Statistics

work page 2022
[5]

and Li, R

Fan, Y. and Li, R. (2012). Variable selection in linear mixed effects models. Annals of statistics , 40(4):2043

work page 2012
[6]

Friedman, J. H. and Rafsky, L. C. (1983). Graph- Theoretic Measures of Multivariate Association and Prediction . The Annals of Statistics , 11(2):377--391. Publisher: Institute of Mathematical Statistics

work page 1983
[7]

and Tutz, G

Groll, A. and Tutz, G. (2014). Variable selection for generalized linear mixed models by l 1-penalized estimation. Statistics and Computing , 24:137--154

work page 2014
[8]

and Nguyen, T

Jiang, J. and Nguyen, T. (2021). Linear and Generalized Linear Mixed Models and Their Applications . Springer Series in Statistics . Springer New York, New York, NY

work page 2021

Show all 16 references

[9]

and Ritov, Y

Law, M. and Ritov, Y. (2021). Inference without compatibility: Using exponential weighting for inference on a parameter of a linear model. Bernoulli , 27(3):1467--1495

2021
[10]

and Ritov, Y

Law, M. and Ritov, Y. (2023). Inference and Estimation for Random Effects in High - Dimensional Linear Mixed Models . Journal of the American Statistical Association , 118(543):1682--1691. Publisher: Taylor & Francis \_eprint: https://doi.org/10.1080/01621459.2021.2004896

2023
[11]

Lee, O. E. and Braun, T. M. (2012). Permutation tests for random effects in linear mixed models. Biometrics , 68(2):486--493

2012
[12]

T., and Li, H

Li, S., Cai, T. T., and Li, H. (2022). Inference for high-dimensional linear mixed-effects models: A quasi-likelihood approach. Journal of the American Statistical Association , 117(540):1835--1846

2022
[13]

M \"u ller, S., Scealy, J., and Welsh, A. (2013). Model selection in linear mixed models. Statistical Science , 28(2):135--167

2013
[14]

Schelldorfer, J., B \"u hlmann, P., and DE GEER, S. V. (2011). Estimation for high-dimensional linear mixed-effects models using l1 penalization. Scandinavian Journal of Statistics , 38(2):197--214

2011
[15]

Wang, L., Zhou, J., and Qu, A. (2012). Penalized generalized estimating equations for high-dimensional longitudinal data analysis. Biometrics , 68(2):353--360

2012
[16]

and Wu, T

Yang, L. and Wu, T. T. (2023). Model-based clustering of high-dimensional longitudinal data via regularization. Biometrics , 79(2):761--774

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

N., Qi, Y., Goren, E., Chiniquy, D., Sheflin, A

Chai, Y. N., Qi, Y., Goren, E., Chiniquy, D., Sheflin, A. M., Tringe, S. G., Prenni, J. E., Liu, P., and Schachtman, D. P. (2024). Root-associated bacterial communities and root metabolite composition are linked to nitrogen use efficiency in sorghum. Msystems , 9(1):e01190--23

work page 2024

[2] [2]

Chen, F., Li, Z., Shi, L., and Zhu, L. (2015). Inference for mixed models of ANOVA type with high-dimensional data. Journal of Multivariate Analysis , 133:382--401

work page 2015

[3] [3]

Cui, S., Ji, T., Li, J., Cheng, J., and Qiu, J. (2016). What if we ignore the random effects when analyzing RNA -seq data in a multifactor experiment. Statistical Applications in Genetics and Molecular Biology , 15(2):87--105

work page 2016

[4] [4]

Ekvall, K. O. and Bottai, M. (2022). Confidence regions near singular information and boundary points with applications to mixed models. The Annals of Statistics , 50(3):1806--1832. Publisher: Institute of Mathematical Statistics

work page 2022

[5] [5]

and Li, R

Fan, Y. and Li, R. (2012). Variable selection in linear mixed effects models. Annals of statistics , 40(4):2043

work page 2012

[6] [6]

Friedman, J. H. and Rafsky, L. C. (1983). Graph- Theoretic Measures of Multivariate Association and Prediction . The Annals of Statistics , 11(2):377--391. Publisher: Institute of Mathematical Statistics

work page 1983

[7] [7]

and Tutz, G

Groll, A. and Tutz, G. (2014). Variable selection for generalized linear mixed models by l 1-penalized estimation. Statistics and Computing , 24:137--154

work page 2014

[8] [8]

and Nguyen, T

Jiang, J. and Nguyen, T. (2021). Linear and Generalized Linear Mixed Models and Their Applications . Springer Series in Statistics . Springer New York, New York, NY

work page 2021