REVIEW 2 major objections 5 minor 30 references
Universal inference for variance components
T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read Split-data likelihood tests produce finite-sample valid confidence intervals for variance components at the boundary.
desk verdict A solid application of known universal inference machinery to the boundary/near-unity variance components problem, with real algorithmic contributions; the main caveat is that the approximate diagonalization in Sec. 3.4 is used without flagging that it falls outside the finite-sample coverage guarantee. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The split likelihood ratio test: partition Y into Y(0) and Y(1) independently of the data, compute estimates from Y(1), evaluate the conditional likelihood of Y(0) given Y(1) at those estimates under the null and the alternative, and reject when the ratio T_n exceeds U/alpha with U ~ Uniform(0,1). Its validity comes from the conditional expectation of the ratio equaling exactly one under the null, so Markov's inequality gives a finite-sample bound. Computation exploits the covariance structure tau^2 Psi(h^2): when the K_m share eigenvectors or are mutually orthogonal, Psi becomes diagonal, reducing likelihood and derivative evaluations from O(n^3) to O(n); crossed random effects are mapped o
What would settle it
Simulate Y from the stated Gaussian model with M=2, K_1 and K_2 fixed and known, true h^2_1 = 0 and h^2_2 = 0.99, n = 300, and construct the randomized split likelihood ratio confidence interval for h^2_1 over 10,000 replicates; if the empirical coverage falls below the nominal 95% level by more than simulation error, the finite-sample uniform validity claim is false.
Extended reading notes
Core claim
The paper establishes uniformly valid finite-sample confidence intervals for variance components and for proportions of variability such as heritability, including when the proportion is near one. For the Gaussian model Y ~ N(0, tau^2(sum_m h^2_m K_m + (1 - sum_m h^2_m) I)), the split likelihood ratio statistic compares the conditional likelihood of one data half at estimates from the other half under the null and alternative hypotheses. Markov's inequality bounds the rejection probability by alpha in any finite sample, and randomizing the threshold with a uniform variable preserves validity while increasing power. This yields, to the paper's knowledge, the first method that is uniformly val
Load-bearing premise
The finite-sample coverage guarantee assumes the Gaussian model is exactly right: the mean is zero or known, the K_m are known, and the response is multivariate normal; coverage is not claimed if the covariance structure is estimated from data, if the response is non-Gaussian, or if K_m is replaced by an approximation.
Editorial extensions
If this is right
- Tests and confidence intervals for one variance component remain valid when other variance components sit at zero or one, where score-based and asymptotic intervals under-cover.
- Heritability and other proportions of variability can receive finite-sample confidence intervals even when the estimate is near one, a case previously not handled.
- The same split-likelihood recipe applies to any correctly specified parametric model, so boundary problems in other mixed models inherit the finite-sample guarantee.
- The randomized threshold reduces conservativeness relative to the plain split likelihood ratio test without sacrificing the coverage bound.
- For models with shared eigenvectors, orthogonal K_m, or crossed random effects, the proposed algorithms make the method practical for n in the thousands.
Reading between the lines
- The finite-sample guarantee is purchased from the likelihood being exactly right; applying the same split-test recipe to non-Gaussian data or to K_m estimated from the same data would require a separate argument.
- Approximate diagonalization replaces the true covariance structure with a close but misspecified one; the simulation suggests the coverage loss is small, but the paper's theory does not formally cover that trade-off.
- The choice of how to split the data, uniformly at random versus balancing factor levels, likely affects power more than coverage; the paper leaves this as an open question, and comparing interval widths across split designs would test it directly.
- If the conservativeness can be reduced, the method could become a practical default replacement for Wald and profile intervals in mixed-model software, not just a safeguard for boundary cases.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper applies universal inference—specifically the split likelihood-ratio test and its randomized version—to Gaussian variance components models with parameterization θ=(h²,τ²). The authors show that tests and confidence intervals for variance components and their proportions are uniformly valid in finite samples, including settings where nuisance components are near their boundary or where the proportion of variability is near unity. They develop faster algorithms for models with diagonalizable covariance structure, including the case of crossed random effects, and they propose an approximate diagonalization scheme for near-block-diagonal settings. The main theorem (Theorem 1) is known from the universal inference literature, but the paper provides a self-contained proof and adapts the method to the variance components setting. Simulations and a crossed-random-effects data example illustrate the methods.
Significance. If the stated guarantees hold, the paper makes a useful contribution: it offers finite-sample valid inference for boundary and near-unity variance components, where standard asymptotic/score-based methods can be unreliable. The computational gains from diagonalization are concrete and potentially important for large n. The paper is honest about conservativeness, gives reproducible code, and includes a clear proof of the underlying split-LRT validity. Its main limitation is that the validity theorem requires the Gaussian model and the covariance matrices K_m to be correctly specified; the approximate diagonalization used in one simulation and the centering step in the data example are not covered by that theorem. These gaps are fixable by qualification or additional theory, but they currently make the abstract's and Section 6's blanket uniformity claims overbroad.
major comments (2)
- [Sec. 3.4 / Fig. 3 (middle)] The approximate diagonalization is not covered by Theorem 1. Replacing K_m by \tilde{K}_m in Eq. (4) means the numerator and denominator are densities from a misspecified model, so the key identity E_{θ*}[L_{\hatθ_1}/L_{θ*} | Y^(1)] = 1 in Appendix A fails and E[T_n] ≤ 1 is not guaranteed. The middle panel of Fig. 3 therefore demonstrates only empirical size/power, not finite-sample uniform validity. The abstract and Section 6 state uniform finite-sample validity without this qualification. Please either provide conditions/proofs for the approximate version or clearly label it as heuristic and restrict all validity claims to the correctly specified model.
- [Sec. 5 / model (11)] The data example centers the response by the sample mean and then analyzes the centered vector as if model (11) held with mean zero and full-rank covariance. Under the stated Gaussian model this creates a misspecified or singular likelihood, so the finite-sample guarantee of Theorem 1 cannot be directly invoked for the reported intervals. Please either use the known-X/unknown-β extension mentioned in the introduction or explicitly state that the data example is approximate and outside the formal guarantees.
minor comments (5)
- [Eq. (5)] The conditional log-likelihood appears to omit the n^(0) log τ² term. Later profiling formulas include the τ² dependence, so this is presumably a typographical omission, but it should be corrected for clarity.
- [Sec. 3.2] Equation (8) is stated for general joint diagonalizability, but Theorem 2 is proved only under the stronger condition K_m K_ℓ = 0. The crossed random effects construction in Sec. 3.3 yields Λ_m Λ_ℓ ≠ 0, so the reader cannot infer joint diagonalizability from Theorem 2 there. Since O is constructed explicitly, this is not an algorithmic gap, but the text should either state the standard commuting-symmetric-matrices fact or clarify that Theorem 2 is only a sufficient condition.
- [Sec. 5] The k-fold averaging of test statistics is used without formal justification. Please add a sentence explaining why the averaged statistic retains the e-value property, or cite the relevant result in Wasserman et al. (2020).
- [Fig. 3 caption] 'Mote Carlo' should be 'Monte Carlo'.
- [Abstract / Sec. 6] Since the formal theorem assumes a correctly specified Gaussian model with known K_m, I suggest adding an explicit qualifier such as 'under correct specification of the covariance model' when stating the finite-sample uniformity guarantee.
Circularity Check
No significant circularity: the validity proof is self-contained, and the approximate-diagonalization algorithm is presented as a heuristic rather than as inheriting Theorem 1.
full rationale
The central derivation is self-contained and does not reduce to its inputs. Theorem 1 is the standard universal-inference result, restated and proved in Appendix A using Markov's inequality and the conditional-density identity E_{θ*}[L(θ̂1)/L(θ*)] = 1. This identity holds because the likelihood is the true density of the correctly specified Gaussian model; no fitted constants or data-dependent parameters are inserted into the inequality. The application to variance components in Section 2.1 simply supplies the Gaussian log-likelihood (5) and defines θ̂0 and θ̂1 as maximizers on the split data, so the coverage guarantee is not definitionally tied to the quantity being inferred. The score-based comparisons and citations to prior work by the authors (Zhang et al. 2025; Ekvall and Bottai 2025) are background and benchmarks, not premises of the proof. The paper itself flags the main limitation: Section 5 states 'the proofs of validity of the proposed methods require a correctly specified likelihood,' and Section 3.4's approximate diagonalization replaces K_m by jointly diagonalizable K̃_m without any theorem stating that the resulting statistic remains an e-value. The middle panel of Fig. 3 is simulation evidence only. This is an overbreadth/correctness concern, not circularity, because the paper does not claim Theorem 1 covers the approximate implementation. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. Hence the circularity score is 0.
Assumptions & free parameters
assumptions (3)
- domain assumption The response Y follows a multivariate normal distribution with mean zero (or known fixed design) and covariance Sigma(theta) as in (1)-(3), with K_m known and nonnegative variance components.
- domain assumption Structural assumptions for the computational shortcuts: shared eigenvectors, Km*Kl = 0, or the specific Kronecker-product structure of crossed random effects.
- standard math Averaging e-values over multiple splits yields a valid combined e-value, as used for the k-fold procedure in Section 5.
Cite this review
Pith. "Pith review of Universal inference for variance components." pith.science (2026). https://pith.science/paper/U2NSM6KH
@misc{pith2026250900255,
author = {Pith},
title = {Pith review of: Universal inference for variance components},
year = {2026},
howpublished = {\url{https://pith.science/paper/U2NSM6KH}},
note = {Machine review of arXiv:2509.00255}
}
read the original abstract
We consider universal inference in variance components models, focusing on settings where the parameter is near or at the boundary of the parameter set. Two cases, which are not handled by existing state-of-the-art methods, are of particular interest: (i) inference on a variance component when other variance components are near or at the boundary, and (ii) inference on near-unity proportions of variability, that is, one variance component divided by the sum of all variance components. Case (i) is relevant, for example, for the construction of componentwise confidence intervals, as often used by practitioners. Case (ii) is particularly relevant when making inferences about heritability in modern genetics. For both cases, we show how to construct confidence intervals that are uniformly valid in finite samples. We propose algorithms which, by exploiting the structure of variance components models, lead to substantially faster computing than naive implementations of universal inference. The usefulness of the proposed methods is illustrated by simulations and a data example with crossed random effects, which are known to be complicated for conventional inference procedures.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[1]
Bates, D., M¨ achler, M., Bolker, B., and Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of statistical software, 67:1–48
work page 2015
-
[2]
Bloom, J. S., Kotenko, I., Sadhu, M. J., Treusch, S., Albert, F. W., and Kruglyak, L. (2015). 21 Genetic interactions contribute less than additive effects to quantitative trait variation in yeast. Nature communications, 6(1):8712
work page 2015
-
[3]
Crainiceanu, C. M. and Ruppert, D. (2004). Likelihood ratio tests in linear mixed models with one variance component. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 66(1):165–185
work page 2004
-
[4]
Da, Y., Wang, C., Wang, S., and Hu, G. (2014). Mixed model methods for genomic prediction and variance component estimation of additive and dominance effects using snp markers. PloS one, 9(1):e87666
work page 2014
-
[5]
Ekvall, K. O. and Bottai, M. (2025). Uniform inference in linear mixed models. arXiv, (2507.19633)
work page Pith review arXiv 2025
-
[6]
Ekvall, K. O. and Jones, G. L. (2020). Consistent maximum likelihood estimation using subsets with applications to multivariate mixed models. The Annals of Statistics, 48(2):932–952
work page 2020
-
[7]
Elkantassi, S., Bellio, R., Brazzale, A. R., and Davison, A. C. (2023). Improved inference for a boundary parameter. Canadian Journal of Statistics, 51(3):780–799
work page 2023
-
[8]
Ghosh, S., Hastie, T., and Owen, A. B. (2022). Scalable logistic regression with crossed random effects. Electronic Journal of Statistics, 16(2)
work page 2022
Show all 30 references
-
[9]
Goldstein, H. (2011). Multilevel statistical models. John Wiley & Sons
2011
-
[10]
N., Ssenyomo, G., Kamali, A., et al
Heckerman, D., Gurdasani, D., Kadie, C., Pomilla, C., Carstensen, T., Martin, H., Ekoru, K., Nsubuga, R. N., Ssenyomo, G., Kamali, A., et al. (2016). Linear mixed model for heritability estimation that explicitly addresses environmental variation. Proceedings of the National A...
2016
-
[11]
Hicks, C. R. and Turner, K. V. (1999). Fundamental concepts in the design of experiments. Oxford University Press, New York, NY, 5 edition
1999
-
[12]
Jiang, J. (2013). The subset argument and consistency of MLE in GLMM: Answer to an open problem and beyond. The Annals of Statistics, 41(1)
2013
-
[13]
Jiang, J. (2025). Asymptotic distribution of maximum likelihood estimator in generalized linear mixed models with crossed random effects. The Annals of Statistics, 53(3):1298–1318. 22
2025
-
[14]
P., and Ghosh, S
Jiang, J., Wand, M. P., and Ghosh, S. (2024). Precise Asymptotics for Linear Mixed Models with Crossed Random Effects
2024
-
[15]
Kreft, I. G. and De Leeuw, J. (1998). Introducing multilevel modeling. Sage
1998
-
[16]
and Nelder, J
Lee, Y. and Nelder, J. A. (1996). Hierarchical generalized linear models. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(4):619–656
1996
-
[17]
Lyu, Z., Sisson, S., and Welsh, A. (2024). Increasing dimension asymptotics for two-way crossed mixed effect models. The Annals of Statistics, 52(6):2956–2978
2024
-
[18]
and Segal, I
Milgrom, P. and Segal, I. (2002). Envelope theorems for arbitrary choice sets. Econometrica, 70(2):583–601
2002
-
[19]
O., and Zanella, G
Papaspiliopoulos, O., Roberts, G. O., and Zanella, G. (2020). Scalable inference for crossed random effects models. Biometrika, 107(1):25–40
2020
-
[20]
Pazokitoroudi, A., Liu, Z., Dahl, A., Zaitlen, N., Rosset, S., and Sankararaman, S. (2024). A scalable and robust variance components method reveals insights into the architecture of gene-environment interactions underlying complex traits. The American Journal of Human Genetic...
2024
-
[21]
and Manole, T
Ramdas, A. and Manole, T. (2023). Randomized and exchangeable improvements of markov’s, chebyshev’s and chernoff’s inequalities. arXiv preprint arXiv:2304.02611
2023 arXiv
-
[22]
and Goldstein, H
Rasbash, J. and Goldstein, H. (1994). Efficient analysis of mixed hierarchical and cross- classified random structures using a multilevel model.Journal of Educational and Behavioral statistics, 19(4):337–350
1994
-
[23]
Robinson, G. K. (1991). That blup is a good thing: the estimation of random effects. Statistical science, pages 15–32
1991
-
[24]
Runcie, D. E. and Crawford, L. (2019). Fast and flexible linear mixed models for genome-wide genetics. PLoS genetics, 15(2):e1007978
2019
-
[25]
Schweiger, R., Fisher, E., Rahmani, E., Shenhav, L., Rosset, S., and Halperin, E. (2018). Using stochastic approximation techniques to efficiently construct confidence intervals for heritability. Journal of Computational Biology, 25(7):794–808. 23
2018
-
[26]
E., M¨ arz, W., Eskin, E., Rosset, S., and Halperin, E
Schweiger, R., Kaufman, S., Laaksonen, R., Kleber, M. E., M¨ arz, W., Eskin, E., Rosset, S., and Halperin, E. (2016). Fast and accurate construction of confidence intervals for heritability. The American Journal of Human Genetics, 98(6):1181–1192
2016
-
[27]
G., Varona, L., and Legarra, A
Vitezica, Z. G., Varona, L., and Legarra, A. (2013). On the additive and dominant variance and covariance of individuals within the genomic selection scope. Genetics, 195(4):1223–1230
2013
-
[28]
Wasserman, L., Ramdas, A., and Balakrishnan, S. (2020). Universal inference. Proceedings of the National Academy of Sciences, 117(29):16880–16890
2020
-
[29]
A., Pasquale, L
Yang, J., Manolio, T. A., Pasquale, L. R., Boerwinkle, E., Caporaso, N., Cunningham, J. M., De Andrade, M., Feenstra, B., Feingold, E., Hayes, M. G., et al. (2011). Genome partitioning of genetic variation for complex traits using common SNPs. Nature genetics, 43(6):519–525
2011
-
[30]
O., and Molstad, A
Zhang, Y., Ekvall, K. O., and Molstad, A. J. (2025). Fast and reliable confidence intervals for a variance component. Biometrika, 112(2):asaf010. A Technical details Proof of Theorem 1.The arguments are available in the literature (Ramdas and Manole, 2023; Wasserman et al., 20...
2025
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.