Leave-one-out testing for node-level differences in Gaussian graphical models

arXiv:2601.15896 · stat.ME · Submitted 2026-01-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Leave-one-out testing for node-level differences in Gaussian graphical models".

Jane: Here's the gist of the paper. We study two-sample equality testing in Gaussian graphical models by proposing a node-level inference procedure via a leave-one-out Bartlettadjusted test on a fully connected graph <ref:2601.15896#pg2>.

Tom: This method yields increments with standard chi-square null limits, enabling calibrated significance for single nodes and fixed-size subsets <ref:2601.15896#pg5>.

Jane: The proposed procedure extends the two-sample LRT of Banzato et al.

Tom: (2023) to a leave-one-out framework on a fully connected graph <ref:2601.15896#pg3>.

Paper summary: Tom: So we're wrapping up this deep dive on "Leave-one-out testing for node-level differences in Gaussian graphical models" by Banzato et al., and what it means for us right now.

Jane: Basically, they took a standard likelihood ratio test and turned it into a leave-one-out procedure to check if two samples have the same structure, one node at a time.

Tom: Yeah, the big point is that this gives us calibrated significance for single nodes and even small groups of nodes instead of just getting one big score for the whole system.

Lu: They use a Bartlett adjustment to keep things accurate when the sample size is small, and they show simulations match what's expected under certain conditions.

Meng: So, if I'm deploying this to production, what's the biggest limitation they admit? What can you rely on?

Tom: They note that a major limitation is relying on a fully connected working graph when doing this analysis. They say this assumption is fine if the true structure is unknown or when you assume a maximal clique idealization.

Jane: It sounds like they're not just giving us a black box test, but actually proposing how to fit this into a larger workflow for finding where the difference actually lies in the data structure.

Lu: The authors suggest a two stage strategy: first find the actual structure of that clique with another method, and then use this leave-one-out analysis inside that specific group.

Tom: It’s about getting meaningful evidence for specific parts of the model structure without needing perfect knowledge of every single connection beforehand.

Jane: This opens up new ways to look at data where you want to know if a tiny piece is behaving differently than the rest.

Lu: We'll talk more about how this approach compares to other methods in our next segment when we look at the simulation results and see the power of this node-level inference.

Conclusion: Tom: So we're wrapping up this deep dive on "Leave-one-out testing for node-level differences in Gaussian graphical models" by Banzato and his team, and what it means for us right now.

Jane: Basically, they took a standard likelihood ratio test and turned it into a leave-one-out procedure to check if two samples have the same structure, one node at a time.

Tom: Yeah, the big point is that this gives us calibrated significance for single nodes and even small groups of nodes instead of just getting one big score for the whole system.

Lu: They use a Bartlett adjustment to keep things accurate when the sample size is small, and they show simulations match what's expected under certain conditions.

Meng: So, if I'm deploying this to production, what's the biggest limitation they admit? What can you rely on?

Tom: They note that a major limitation is relying on a fully connected working graph when doing this analysis. They say this assumption is fine if the true structure is unknown or when you assume a maximal clique idealization.

Jane: It sounds like they're not just giving us a black box test, but actually proposing how to fit this into a larger workflow for finding where the difference actually lies in the data structure.

Lu: The authors suggest a two stage strategy: first find the actual structure of that clique with another method, and then use this leave-one-out analysis inside that specific group.

Tom: It’s about getting meaningful evidence for specific parts of the model structure without needing perfect knowledge of every single connection beforehand.

Jane: This opens up new ways to look at data where you want to know if a tiny piece is behaving differently than the rest.

Lu: We'll talk more about how this approach compares to other methods in our next segment when we look at the simulation results.

Davide Benussi, Ester Alongi, Erika Banzato

Department of Statistical Sciences, University of Padova

stat.ME

Submitted: 2026-01-22

Updated: 2026-11-15

Journal ref: Journal of Statistical Planning and Inference, 248 (2027), 106486

DOI: 10.1016/j.jspi.2026.106486

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 86/100

The gist: We study two-sample equality testing in Gaussian graphical models by proposing a node-level inference procedure via a leave-one-out Bartlettadjusted test on a fully connected graph…

Key concepts

Leave-one-out framework
This procedure tests a hypothesis by comparing the test statistic calculated on the entire graph against the statistic calculated after removing one node at a time. This allows researchers to assess how much each specific node contributes to the overall difference being tested.
Bartlett-adjusted test
The original likelihood ratio test is corrected using Bartlett's correction. This adjustment improves the accuracy of finite-sample results, ensuring that the significance levels are more reliable when dealing with smaller datasets compared to standard tests.
Null limiting distribution
This describes the theoretical distribution of the test statistic when there is no true difference between groups (the null hypothesis). The paper shows this distribution converges to a chi-square value, which allows for precise calibration of the significance levels.

Terminology

Summary

We study two-sample equality testing in Gaussian graphical models by proposing a node-level inference procedure via a leave-one-out Bartlettadjusted test on a fully connected graph <ref:2601. p. 0

Where the paper says it

Classical likelihood ratio tests on decomposable graphs admit clique-wise factorizations, offering limited localization and unstable finite-sample behaviour. We propose node-level inference via a leave-one-out Bartlettadjusted test on a fully connected graph.

15896#pg2>. This method yields increments with standard chi-square null limits, enabling calibrated significance for single nodes and fixed-size subsets <ref:2601. p. 0
Where the paper says it

We propose node-level inference via a leave-one-out Bartlettadjusted test on a fully connected graph. The resulting increments have standard chi-square null limits, enabling calibrated significance for single nodes and fixed-size subsets.

15896#pg5>.

Methodology of Node Inference

The proposed procedure extends the two-sample LRT of Banzato et al. (2023) to a leave-one-out framework on a fully connected graph <ref:2601. p. 1

Where the paper says it

We propose a node-level inference procedure extending the two-sample LRT of Banzato et al. (2023) to a leave-one-out framework on a fully connected graph, either treated as a maximal clique or as a simplifying assumption when the structure is unknown.

15896#pg3>. For any candidate node, the procedure contrasts the Bartlett-adjusted LRT on the full graph with the statistic recomputed after removing that node <ref:2601. p. 1
Where the paper says it

(2023) to a leave-one-out framework on a fully connected graph, either treated as a maximal clique or as a simplifying assumption when the structure is unknown. For any candidate node, the procedure contrasts the Bartlett-adjusted LRT on the full graph with the statistic recomputed after removing that node.

15896#pg4>. The resulting increment, under the null hypothesis of equality of distribution, has an asymptotic chi-square distribution with degrees of freedom depending only on the number of nodes removed and the total number of nodes <ref:2601. p. 1
Where the paper says it

For any candidate node, the procedure contrasts the Bartlett-adjusted LRT on the full graph with the statistic recomputed after removing that node. The resulting increment, under the null hypothesis of equality of distribution, has an asymptotic chi-square distribution with degrees of freedom depending only on the number of nodes removed and the total number of nodes.

15896#pg5>.

Null Limiting Distribution Results

Theorem 1 establishes that under H0, as n = n1 + n2 → ∞, the increment ∆M = T V n - T V M n d−→ χ2h(l,p), where h(l, p) = l(2p − l + 3)2 <ref:2601.15896#pg5>. Corollary 1 specifies that for a singleton removal (l = 1), the null limit is ∆j d−→ χ2p+1 <ref:2601. p. 3

Where the paper says it

Note that h(1, p) = p + 1, h(2, p) = 2p + 1, and h(3, p) = 3p, corresponding to singletons, pairs, and triplets, which are our main focus, though Theorem 1 holds for any 1 ≤ l ≤ p − 1. Corollary 1 (Singleton removal).

15896#pg5>. Furthermore, Conjecture 1 suggests that under H1, the increments behave as ∆j d−→ χ2p+1(λS) for altered nodes and ∆j' d−→ χ2p+1(λSc) for unaltered nodes <ref:2601.15896#pg5>.

Finite Sample Calibration and Error Control

The Bartlett-type correction improves finite-sample accuracy, and the adjusted test statistic T V n is defined as T V n = δ V n W V n, with δ V n = f V - 2µ V n <ref:2601.15896#pg4>. Numerical experiments in Section 3 indicate that Bartlett-adjusted increments provide better marginal type I error control than their Wnbased counterparts, particularly at small sample sizes <ref:2601. p. 5

Where the paper says it

05, for each subset M. Bartlett-adjusted increments provide better marginal type I error control than their Wnbased counterparts, particularly at small sample sizes; see the leftmost columns for n1 = n2 = 10.

15896#pg8>.

Practical Application and Caveats

The methodology is computationally light and inherits improved finite sample calibration from the Bartlett adjustment <ref:2601.15896#pg5>. A limitation noted is the reliance on a fully connected working graph, which is defensible when the true structure is unknown or when a maximal clique idealization is common <ref:2601. p. 8

Where the paper says it

A limitation is the reliance on a fully connected working graph. This is defensible when the true structure is unknown and a maximal clique idealization is common.

15896#pg10>. The authors suggest a practical two-stage strategy: first localize the responsible clique with a valid clique-wise method, and then apply this leave-one-out analysis within that clique <ref:2601. p. 8
Where the paper says it

This is defensible when the true structure is unknown and a maximal clique idealization is common. This also suggests a practical two-stage strategy: first localize the responsible clique with a valid clique-wise method (Salviato et al.

15896#pg10>.

Simulation Validation

The simulation study confirms the validity of the null limiting distribution stated in Theorem 1 through empirical distributions matching the expected chi-square limits <ref:2601.15896#pg8>. The results demonstrate that Bartlett-adjusted increments track the nominal level more closely than Wn-based tests, and power results show improved complete recovery when shifts affect few nodes <ref:2601. p. 3

Where the paper says it

(ii) Using Wn in place of Tn preserves the form of the results but degrades the finite-sample chi-square approximation. (iii) Numerical experiments in Section 3 indicate Wn-based tests are anti-conservative, whereas Tn-based tests track the nominal level more closely.

15896#pg8>.

Conclusions

We developed node and subset level two-sample inference for Gaussian graphical models by defining leave-one-out counterparts of a Bartlett-adjusted global LRT <ref:2601.15896#pg5>. The procedure is computationally light, inherits improved finite sample calibration from the Bartlett adjustment, and performs well in both simulations and data analyses <ref:2601. p. 7

Where the paper says it

When the number of altered nodes is unknown, classical multiple testing adjustments yield valid selection of the set of altered nodes. The procedure is computationally light, inherits improved finite sample calibration from the Bartlett adjustment, and performs well in both simulations and 8

15896#pg5>. The authors conclude that this approach provides calibrated significance for individual nodes and fixed-size subsets <ref:2601. p. 7
Where the paper says it

4 Conclusions We developed node and subset level two-sample inference for Gaussian graphical models by defining leave-one-out counterparts of a Bartlett-adjusted global LRT. The associated increments have standard null limits, which provide calibrated significance for individual nodes and fixed-size subsets.

15896#pg5>.

Acknowledgements

The authors thank Prof. Davide Risso (Department of Statistical Sciences, University of Padova) for his valuable insights during the preparation of this manuscript <ref:2601. p. 8

Where the paper says it

Acknowledgements The authors thank Prof. Davide Risso (Department of Statistical Sciences, University of Padova) for his valuable insights during the preparation of this manuscript.

15896#pg9>.

References

Banzato, E., Chiogna, M. p. 1

Where the paper says it

Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al. , 2019; Djordjilovi´c and Chiogna, 2022), but do not directly address node-level inference and rely on decomposability.

p. 1
Where the paper says it

Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al. , 2019; Djordjilovi´c and Chiogna, 2022), but do not directly address node-level inference and rely on decomposability.

, Djordjilovi´c, V. p. 1
Where the paper says it

Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al. , 2019; Djordjilovi´c and Chiogna, 2022), but do not directly address node-level inference and rely on decomposability.

p. 1
Where the paper says it

Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al. , 2019; Djordjilovi´c and Chiogna, 2022), but do not directly address node-level inference and rely on decomposability.

p. 1
Where the paper says it

Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al. , 2019; Djordjilovi´c and Chiogna, 2022), but do not directly address node-level inference and rely on decomposability.

, and Risso, D. p. 8
Where the paper says it

Acknowledgements The authors thank Prof. Davide Risso (Department of Statistical Sciences, University of Padova) for his valuable insights during the preparation of this manuscript.

(2023). A Bartlett-type correction for likelihood ratio tests with application to testing equality of gaussian graphical models. p. 8
Where the paper says it

, and Risso, D. A Bartlett-type correction for likelihood ratio tests with application to testing equality of gaussian graphical models.

Stat. Probab. Lett., 193:109732 <ref:2601. p. 8
Where the paper says it

A Bartlett-type correction for likelihood ratio tests with application to testing equality of gaussian graphical models. , 193:109732.

15896#pg9>.

Bartlett, M. p. 1

Where the paper says it

In decomposable graphs, the likelihood ratio statistic (LRT; Wilks, 1938) admits convenient clique-separator factorizations and asymptotic chi-square limits, with finite-sample accuracy improved by Bartlett corrections (Bartlett, 1937; Banzato et al. Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al.

S. (1937). Properties of sufficiency and statistical tests. p. 8
Where the paper says it

Bartlett, M. Properties of sufficiency and statistical tests.

Proc. R. Soc. Lond. A, 160:268–282 <ref:2601. p. 8
Where the paper says it

Properties of sufficiency and statistical tests. A, 160:268–282.

15896#pg5>.

Benjamini, Y. p. 8

Where the paper says it

, 2019; Djordjilovi´c and Chiogna, 2022; Banzato et al. , 2023); then apply our leave-one-out analysis within that clique, controlling the overall false discovery rate by an acrossstage adjustment (Benjamini and Heller, 2008; Van den Berge et al.

and Heller, R. p. 8
Where the paper says it

, 2019; Djordjilovi´c and Chiogna, 2022; Banzato et al. , 2023); then apply our leave-one-out analysis within that clique, controlling the overall false discovery rate by an acrossstage adjustment (Benjamini and Heller, 2008; Van den Berge et al.

(2008). Screening for partial conjunction hypotheses <ref:2601. p. 8
Where the paper says it

and Heller, R. Screening for partial conjunction hypotheses.

15896#pg10>.

Bonferroni, C. p. 4

Where the paper says it

Because |M(l)| grows combinatorially, testing all {∆M : M ∈ M(l)} without adjustment inflates the probability of spurious findings, especially since the tests are dependent. We therefore use multiplicity-adjusted p-values (Bonferroni, 1936; Holm, 1979) to control the family-wise error rate (FWER) under arbitrary dependence.

E. (1936). Teoria statistica delle classi e calcolo delle probabilit'a <ref:2601.15896#pg5>.

Chiaretti, S. p. 6

Where the paper says it

dataset (Chiaretti et al. , 2005; Li, 2025).

, Li, X., Gentleman, R. p. 9
Where the paper says it

Chiaretti, S. , Gentleman, R.

, Vitale, A. p. 9
Where the paper says it

, Gentleman, R. , Vitale, A.

, Wang, K., Mandelli, F. p. 9
Where the paper says it

, Vitale, A. , Mandelli, F.

, Fo'a, R., and Ritz, J. p. 9
Where the paper says it

, Mandelli, F. , and Ritz, J.

(2005). Gene expression profiles of b-lineage adult acute lymphocytic leukemia reveal genetic patterns that identify lineage derivation and distinct mechanisms of transformation. p. 9
Where the paper says it

, and Ritz, J. Gene expression profiles of b-lineage adult acute lymphocytic leukemia reveal genetic patterns that identify lineage derivation and distinct mechanisms of transformation.

Djordjilovi´c, V. and Chiogna, M. p. 1

Where the paper says it

Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al. , 2019; Djordjilovi´c and Chiogna, 2022), but do not directly address node-level inference and rely on decomposability.

(2022). Searching for a source of difference in graphical models <ref:2601. p. 9
Where the paper says it

and Chiogna, M. Searching for a source of difference in graphical models.

15896#pg4>.

Holm, S. (1979). A simple sequentially rejective multiple test procedure <ref:2601. p. 9

Where the paper says it

, 190:104973. A simple sequentially rejective multiple test procedure.

15896#pg5>.

Lauritzen, S. p. 0

Where the paper says it

Simulations confirm validity, and a case study shows practical utility. Keywords: Likelihood ratio test; Gaussian graphical models; Hypothesis testing; Node-level inference 1 Introduction Gaussian graphical models (GGMs) parametrize multivariate normal distributions by a mean vector and a precision matrix constrained by an undirected graph, where zeros in the precision encode conditional independences (Lauritzen, 1996; Whittaker, 2009).

L. (1996). Graphical Models <ref:2601. p. 0
Where the paper says it

Leave-one-out testing for node-level differences in Gaussian graphical models Davide Benussi∗1 , Ester Alongi†2 , and Erika Banzato‡1 1Department of Statistical Sciences, University of Padova, 35121, Padova, Italy January 23, 2026 Abstract We study two-sample equality testing in Gaussian graphical models. Classical likelihood ratio tests on decomposable graphs admit clique-wise factorizations, offering limited localization and unstable finite-sample behaviour.

15896#pg4>.

Li, X. (2025). ALL: A data package <ref:2601. p. 9

Where the paper says it

Clarendon Press. ALL: A data package.

15896#pg4>.

Salviato, E. p. 1

Where the paper says it

In decomposable graphs, the likelihood ratio statistic (LRT; Wilks, 1938) admits convenient clique-separator factorizations and asymptotic chi-square limits, with finite-sample accuracy improved by Bartlett corrections (Bartlett, 1937; Banzato et al. Recent contributions exploit decomposable structure to localize changes through clique-wise testing and set operations (Salviato et al.

, Djordjilovi´c, V., Chiogna, M., and Romualdi, C. p. 9
Where the paper says it

, Chiogna, M. , and Romualdi, C.

(2019). SourceSet: A graphical model approach to identify primary genes in perturbed biological pathways <ref:2601. p. 9
Where the paper says it

, and Romualdi, C. SourceSet: A graphical model approach to identify primary genes in perturbed biological pathways.

15896#pg4>.

Van den Berge, K. p. 8

Where the paper says it

, 2019; Djordjilovi´c and Chiogna, 2022; Banzato et al. , 2023); then apply our leave-one-out analysis within that clique, controlling the overall false discovery rate by an acrossstage adjustment (Benjamini and Heller, 2008; Van den Berge et al.

, Soneson, C. p. 9
Where the paper says it

Van den Berge, K. , Soneson, C.

, Robinson, M. p. 9
Where the paper says it

, Soneson, C. , Robinson, M.

D., and Clement, L. p. 9
Where the paper says it

, Robinson, M. , and Clement, L.

(2017). stageR: a general stage-wise method for controlling the gene-level false discovery rate in differential expression and differential transcript usage <ref:2601. p. 9
Where the paper says it

, and Clement, L. stageR: a general stage-wise method for controlling the gene-level false discovery rate in differential expression and differential transcript usage.

15896#pg10>.

Whittaker, J. p. 0

Where the paper says it

Simulations confirm validity, and a case study shows practical utility. Keywords: Likelihood ratio test; Gaussian graphical models; Hypothesis testing; Node-level inference 1 Introduction Gaussian graphical models (GGMs) parametrize multivariate normal distributions by a mean vector and a precision matrix constrained by an undirected graph, where zeros in the precision encode conditional independences (Lauritzen, 1996; Whittaker, 2009).

(2009). Graphical Models in Applied Multivariate Statistics <ref:2601. p. 9
Where the paper says it

Whittaker, J. Graphical Models in Applied Multivariate Statistics.

15896#pg4>.

Wilks, S. S. (1938). The large-sample distribution of the likelihood ratio for testing composite hypotheses <ref:2601. p. 9

Where the paper says it

Graphical Models in Applied Multivariate Statistics. The large-sample distribution of the likelihood ratio for testing composite hypotheses.

15896#pg3>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Leave-one-out inference for Gaussian graphical models allows for node-level inference by contrasting a Bartlett adjusted likelihood ratio test on a fully connected graph with one excluding a single node, resulting in valid node-level p-values while retaining the finite-sample accuracy of the Bartlett correction.

  2. The system can perform valid subset level testing by calculating increments like Delta M = T V n - T V M n, which is stochastically large when M intersect S != empty and asymptotically negligible when M subset Sc.

  3. When the true structure is unknown, the system can adopt a two-stage strategy: first localize the responsible clique with a valid clique-wise method, then apply leave-one-out analysis within that clique for node inference while controlling false discovery rate by an acrossstage adjustment.

Abstract

We study two-sample equality testing in Gaussian graphical models with the aim of localizing distributional differences at the node level. The proposed procedure targets departures in the joint distribution, which may arise from mean shifts, variance changes, or covariance differences. We extend Bartlett-adjusted likelihood ratio tests to a leave-one-out framework on a fully connected working graph. The resulting increments have standard chi-square null limits, enabling calibrated inference for single nodes and fixed-size subsets. We also discuss how decomposable graphical structure can be exploited to apply the approach locally on larger graphs. Simulation results indicate good calibration and competitive power, and a real data application highlights the practical applicability of the method.

Related papers