acceptodds
Under review as a conference paper at ICLR 2027

PoET: Test-Integrated Nonparametric Poisson DAG Learning via Equidispersion Testing

Abstract

Poisson DAG models enable causal discovery for count data through a mean–variance asymmetry: conditioning on all parents yields equidispersion, whereas omitted parents induce overdispersion. Existing ranking-based methods exploit this asymmetry to construct a causal ordering by repeatedly selecting the node with the smallest overdispersion, without testing whether the selected node is statistically compatible with equidispersion. These methods also face difficulties in nonparametric conditional-mean estimation or avoid them by prespecifying nodewise link functions, which limits modeling flexibility. We propose PoET, a test-integrated framework for Poisson DAG learning that replaces score ranking with test-based layer eligibility. Its new cross-fitted tests for conditional uncorrelatedness and equidispersion use Conway–Maxwell–Poisson (CMP)-kernel smoothing to estimate conditional means nonparametrically, without prespecified nodewise link functions. To mitigate the conditioning burden of nonparametric smoothing, conditional-uncorrelatedness tests screen parent candidates to reduce conditioning dimension, while equidispersion tests guide layer assignment and parent recovery. If no candidate is eligible for layer assignment, PoET stops rather than forcing further graph decisions. We establish graph-recovery consistency under correct specification and consistent violation detection under local misspecification. Simulations show competitive recovery under heterogeneous nonlinear mechanisms and reliable detection of dispersion violations. NBA and Lending Club analyses illustrate, respectively, test-based incompatibility detection and complete layerwise learning without prescribed links.

Primary area:
causal reasoning
Keywords:
causal discovery, Poisson DAG, count data, graph recovery, equidispersion, CMP kernel

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_odihOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces PoET, a novel approach for graph estimation in SEMs with Poisson likelihoods. PoET uses a test-based approach, starting from an empty graph and successively adding edges if they approximately satisfy a dispersion-based criterion (which is satisfied for actual Poisson DAGs). The test is carried out using a nonparametric approach that allows the likelihood to depend nonlinearly on previous nodes. Empirical results are given showing that the PoET approach tends to outperform competitors at very large numbers of samples

Critical strengths
  • The paper is clear and presents the PoET method in an easy-to-understand and clear way. The general argument, that the method is a good approach for structure learning in Poisson DAGs, is generally well-made and seems correct. The overall approach is a nice combination of different existing methods, and the story is convincing.
  • The empirical arguments are convincing (at least for high numbers of data points, see weaknesses) and the theoretical consistency arguments are correct, at least as far as I can tell.
Critical weaknesses
  • The empirical results could be more compelling. The results for the synthetic experiments don't have any results for n <= 1000, and the results at n=1000 generally show a pretty severe drop compared to 3k. As far as I understand, in these SEM cases we are often quite concerned with relatively small number of data points (e.g. the standard Sachs dataset has 853 observational samples). Similarly, the real-world-data examples are the NBA example (rejected for not being Poisson) and the lending club data, which is only 5-dimensional and could presumably be studied with direct enumeration.
  • As far as I understand, the advance is somewhat incremental. The technique consists of combining several previously-known ideas, and it's not obvious there is a huge novelty. However, the careful treatment of the approach including proofs and empirical results means it's, as far as I can tell, about in-line with the novelty of previous DAG papers at ICLR in previous years, so this is not a blocker for acceptance. There is perhaps some over-selling in the paper, e.g. the repeated claims about testing being better than scoring. I'm not entirely sure I understand why we can't just threshold our score to give a test
Questions for authors
  • Can you either give results showing PoET's performance at lower n, or explain why the high-n regime is important? Compared to previous work, this n seems quite high (e.g. NOTEARS gives results for n=20, n=1000).
  • Can you comment on the distinction between scoring and testing, given that we can presumably construct a test by thresholding the score? Is there a question around the calibration/consistency of such a score-driven test?
4 comments
AuthorsOpenReview ↗

Response to Reviewer odih - W1 & Q1

(W1/Q1) Small-sample performance and the role of larger samples.

We thank the reviewer for pointing out that the original manuscript did not adequately evaluate small-sample performance. We have added experiments in Appendix A.3.8 at n=500n=500 and p=5p=5, close to the real-data settings, comparing PoET with all eight baselines over 100 replications. PoET's mean F1 ranges from 0.558 to 0.689, while ODS achieves the highest mean F1 in all six settings.

Average graph-recovery F1 scores (standard errors) at n=500n=500 and p=5p=5. Each method uses 100 replications.

Graph density: g=2g=2

MethodHubBAER
PoET0.689 (0.032)0.604 (0.030)0.616 (0.032)
ODS0.715 (0.016)0.693 (0.019)0.769 (0.020)
LearnDAG0.669 (0.018)0.606 (0.024)0.738 (0.022)
DirectLiNGAM0.152 (0.017)0.130 (0.017)0.186 (0.019)
MRS0.539 (0.021)0.555 (0.021)0.558 (0.025)
NOTEARS0.641 (0.019)0.584 (0.023)0.685 (0.026)
TLDAG0.690 (0.017)0.615 (0.023)0.639 (0.025)
PC0.366 (0.028)0.244 (0.017)0.402 (0.031)
ZICO0.463 (0.021)0.493 (0.029)0.391 (0.028)

Graph density: g=4g=4

MethodHubBAER
PoET0.558 (0.035)0.628 (0.031)0.558 (0.035)
ODS0.757 (0.014)0.755 (0.017)0.757 (0.014)
LearnDAG0.617 (0.018)0.689 (0.018)0.617 (0.018)
DirectLiNGAM0.305 (0.019)0.230 (0.017)0.305 (0.019)
MRS0.497 (0.019)0.541 (0.019)0.497 (0.019)
NOTEARS0.575 (0.016)0.627 (0.015)0.575 (0.016)
TLDAG0.653 (0.017)0.683 (0.017)0.653 (0.017)
PC0.421 (0.019)0.381 (0.023)0.421 (0.019)
ZICO0.208 (0.017)0.336 (0.022)0.208 (0.017)

Larger samples are important for PoET for two reasons.

First, PoET checks model assumptions while learning the graph. Unlike score-based methods, which always return a best-ranked graph decision, PoET estimates local null distributions and withholds further decisions when assumption violation is detected. Its local tests require sufficiently accurate conditional-mean estimates [1]. At small sample sizes, estimation error and limited test power can affect both graph recovery and model diagnostics. Larger samples help support these testing decisions throughout learning.

Second, broader functional flexibility carries a nonparametric estimation cost. PoET estimates unspecified multivariate conditional means, allowing nonlinear effects and interactions beyond prespecified parametric forms. This flexibility makes estimation demanding, especially as the number of conditioning variables increases [2]. The low-dimensional p=5p=5 setting may also favor ODS's exact-cell estimation.

These results illustrate the trade-off between flexible, test-integrated learning and small-sample performance. We have clarified this trade-off and PoET's small-sample limitations in the revised manuscript.

[1] Shah, R. D., and Peters, J. The hardness of conditional independence testing and the generalised covariance measure. The Annals of Statistics, 48(3):1514–1538, 2020.

[2] Bühlmann, P., Peters, J., and Ernest, J. CAM: Causal additive models, high-dimensional order search and penalized regression. The Annals of Statistics, 42(6):2526–2556, 2014.

AuthorsOpenReview ↗

Response to Reviewer odih - W2 & Q2

(W2/Q2) Distinction from score-based methods.

We thank the reviewer for raising this point. We agree that PoET's distinction from score-based methods was not sufficiently clear in the original manuscript. We have added how PoET is related to score thresholding in Section 3.1.

To construct a test by thresholding the overdispersion score S^i(Z)\widehat{S}_i(\boldsymbol{Z}) using a normal approximation, we would need to establish the asymptotic distribution of n(S^i(Z)−1)\sqrt{n}(\widehat{S}_i(\boldsymbol{Z})-1). For a naive plug-in score, this expansion can involve a first-order conditional-mean estimation-error term of the form 1n∑t(m^i(k)(Zt)−mi(Zt))\frac{1}{\sqrt{n}}\sum_t \left(\widehat{m}_i^{(k)}(\boldsymbol{Z}_t)-m_i(\boldsymbol{Z}_t)\right). With nonparametric estimation, which converges at a rate slower than n\sqrt{n}, this term need not be asymptotically negligible and can even diverge, so consistency of the estimated score alone does not justify a normal approximation. We therefore use QnQ_n, whose residual-based construction and cross-fitting control the first-order estimation error, leaving a quadratic bias that is negligible under the stated conditions.

PoET can nevertheless be viewed as a score-thresholding method in a broader sense by interpreting QnQ_n as a new score. Its contribution is the statistical justification of the threshold. For the equidispersion test, we use

TQ,n=n,Qnσ^Qand reject when∣TQ,n∣>z1−α/2.T_{Q,n}=\frac{\sqrt{n}\\,Q_n}{\widehat{\sigma}_Q} \qquad\text{and reject when}\qquad |T_{Q,n}|>z_{1-\alpha/2}.

Under the stated conditions, Lemma 2 establishes TQ,n⇒N(0,1)T_{Q,n}\Rightarrow N(0,1) for a fixed equidispersion null, giving an asymptotic rejection probability of α\alpha. This statistical justification is the distinction we would like to emphasize.

Thanks for your replies. Although I expected PoET to have worse performance at lower nn, it's good to see that the performance doesn't crater massively. For the comment about the score vs thresholding, your response makes sense--essentially arbitrarily thresholding a score isn't statistically justified while the approach in PoET is well-specified.

I don't think the paper is necessarily at the 'clear accept' level but your replies have increased my confidence in recommending a weak acceptance.

Thank you for taking the time to review our paper and follow up on our responses. We are glad you found the lower nn results reassuring and that our explanation clarified how PoET’s statistically justified thresholds differ from arbitrary score thresholding. We appreciate your continued support.

Reviewer_F88YOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+2)
Read review
Paper summary

The paper proposes PoET, a causal discovery method for count data under a nonparametric Poisson DAG model. Each node is Poisson given its parents, while leaving out a parent pushes the variance up. Earlier dispersion-based methods (ODS, MRS, TLDAG) use this asymmetry as a score: they keep picking the "best" node even when none fits the model. PoET instead tests equidispersion directly to decide which nodes join the next layer, and it stops when no node passes. Nodes left over are labelled as violations. Conditional means are estimated with cross fitted Conway Maxwell Poisson (CMP) kernel smoothing, so no link functions need to be specified. A conditional-uncorrelatedness test screens parent candidates to keep conditioning sets small, and equidispersion tests then drive parent expansion and pruning. The theory gives asymptotic validity for both local tests, consistent graph recovery under correct specification, and consistent violation detection under local misspecification. Experiments cover simulations (p = 30, n from 1,000 to 10,000, hub/BA/ER graphs, nonlinear links) and two real datasets (NBA and Lending Club).

Critical strengths
  1. It is better for a method to admit the model doesn't fit than to force a graph, and the stopping rule does exactly that.
  2. The appendix includes false-violation rates under correct specification, tuning sensitivity, a screening ablation and larger-graph experiments.
Critical weaknesses
  1. Binomial violations are detected much better than negative binomial violations or confounding. The generating mechanisms explain why: binomial is underdispersed, while negative binomial and confounding push the variance up, just like a missing parent. So the hard violations look like the very signal the method uses to build the graph. This matters because real count data are often overdispersed, and the paper barely discusses it.
  2. Income is binned into an index, which is not a count. Changing the binning or relaxing the sample filters breaks the graph. The result looks highly dependent on preprocessing, and there is no ground truth to check it against.
  3. No runtimes are reported, even though the method smooths repeatedly. The larger-graph experiment is also limited.
  4. PoET loses to the baselines at small n and only pulls ahead at larger n. The main text should say this clearly instead of leaving it mostly to a figure. Both real datasets are also small, at around 450 rows each.
  5. The main ingredients (existing tests, kernels and cross-fitting) are already known.
Questions for authors
  1. What are the runtimes compared with the baselines, and how do they scale with the number of nodes and the sample size?
  2. Negative binomial nodes and latent confounding both raise the variance, just like a missing parent. How does PoET tell these apart from a genuinely missing parent? Does expansion tend to add spurious parents until the extra variance is absorbed? Could you report how often misspecified nodes end up placed in the graph with wrong parents, rather than flagged?
  3. Both real datasets are much smaller than the smallest simulated sample size. Are the tests calibrated at that size? Could you report false-violation rates and graph recovery at around the real-data sample size?
10 comments
AuthorsOpenReview ↗

Response to Reviewer F88Y - W1

(W1) Clarification on strength and direction of dispersion.

We thank the reviewer for pointing out that the paper did not clearly explain the strength of the two types of departures. PoET uses a two-sided equidispersion test that targets both underdispersion and overdispersion without prioritizing either direction. The original settings have average variance-to-mean ratios of about 0.32 for binomial nodes and 1.61 for negative binomial nodes, so the two settings differ in the magnitude of their departures from equidispersion. Detection performance depends on the strength of the departures, not their direction alone.

We vary the dispersion parameters in Appendix A.3.3 while retaining the conditional-mean functions and all other simulation settings. For binomial nodes, we vary the number of trials M∈20,60,120M\in\\{20,60,120\\}, where M=20M=20 is the original setting, adjusting the success probabilities to preserve the conditional means. Increasing MM moves the variance-to-mean ratio toward one and weakens underdispersion. For negative binomial nodes, we vary the size parameter r∈2,1,0.5r\in\\{2,1,0.5\\}, where r=2r=2 is the original setting. Decreasing rr increases the variance-to-mean ratio and strengthens overdispersion. These comparisons examine how detection performance changes with departure strength within each family.

Average violation-detection F1 scores (standard errors) with p=30p=30, n=1000n=1000, g=2g=2, and 100 replications. The variance-to-mean column reports the range across graph families of the average conditional ratio at misspecified nodes.

MechanismVariance-to-mean ratioHubBAER
Binomial, M=20M=200.316–0.3210.885 (0.016)0.770 (0.017)0.785 (0.012)
Binomial, M=60M=600.772–0.7740.597 (0.028)0.561 (0.026)0.608 (0.020)
Binomial, M=120M=1200.886–0.8870.141 (0.020)0.105 (0.013)0.280 (0.020)
Negative binomial, r=2r=21.604–1.6330.365 (0.032)0.229 (0.028)0.238 (0.020)
Negative binomial, r=1r=12.209–2.2630.755 (0.028)0.554 (0.026)0.609 (0.021)
Negative binomial, r=0.5r=0.53.414–3.5220.784 (0.027)0.619 (0.025)0.671 (0.020)

Weakening binomial underdispersion lowers violation-detection F1, while strengthening negative binomial overdispersion improves it. Negative binomial with r=0.5r = 0.5 exceeds binomial with M=120M=120 in all three graph families. These results suggest that detection depends strongly on how far the conditional distribution departs from equidispersion, rather than being determined by the overdispersion or underdispersion alone.

AuthorsOpenReview ↗

Response to Reviewer F88Y - W2

(W2) Concerns on preprocessing and the lack of ground truth.

We thank the reviewer for highlighting the need to clarify our preprocessing choices and the interpretation of the Lending Club results. We present the Lending Club analysis as an exploratory illustration of PoET on real data. The variables, sample filters, and baseline income binning were selected based on domain considerations and data characteristics.

Sensitivity to preprocessing. We acknowledge that the discretized income index is not a count variable and treat it as an empirical approximation. We retained it to avoid omitting an important variable that may affect both mortgage- and installment-account holdings. The table below summarizes the 15-, 20-, and 25-bin analyses in Appendix A.4 and adds a 30-bin sensitivity check. Across the alternative specifications, three or four of the five baseline edges are retained, although the 15- and 30-bin specifications leave one node unassigned. We have revised and added the results and relevant discussion in Section 6 and Appendix A.4.

Sensitivity to the number of equal-width income bins. All other settings are held fixed at the baseline values (n=455n=455, seed =100=100, α=0.05\alpha=0.05, and five-fold cross-fitting). The 20-bin specification is the baseline.

Income binsAssigned nodesLearned edgesBaseline edges retainedResult
154/533/5Incomplete
205/555/5Complete
255/554/5Complete
304/533/5Incomplete

Sensitivity to sample filtering. The learned graph is sensitive to sample filtering. However, because PoET does not directly model categorical variables as nodes, we used them as sample filters to reduce heterogeneity across their categories. The table below compares groups for each filter while holding the other three restrictions fixed.

Distributional differences across the categorical groups used for sample filtering. The other three restrictions are held fixed. The pp-values are from separate two-sided Mann–Whitney UU tests; percentages and medians are descriptive summaries.

FilterComparisonpp-valueDescriptive summary
HomeownershipRENT vs. MORTGAGE1.51×10−881.51\times10^{-88}Mortgage accounts: zero accounts, 83% vs. 9%
GradeD vs. A6.55×10−86.55\times10^{-8}Delinquencies: at least one, 24.5% vs. 9.9%
Application typeJoint vs. individual1.46×10−81.46\times10^{-8}Annual income: median $\65{,}000\ vs. $\87{,}812\
Loan purposeNon-debt-consolidation vs. debt consolidation0.003350.00335Current installment accounts: zero, 14.2% vs. 9.5%; at least three, 34.7% vs. 42.9%

Together with domain knowledge, these category-specific differences suggest that pooling may leave relevant borrower characteristics unmodeled as direct causes of individual variables or as common causes of multiple variables. Filtering is therefore a reasonable way to reduce such heterogeneity. This analysis shows how PoET can be applied to count variables in a dataset that also contains categorical variables: the categorical variables are used for sample filtering rather than included as DAG nodes.

Interpretation without ground truth. Without a ground-truth DAG, we cannot directly evaluate causal recovery accuracy. Nevertheless, several recovered relationships are consistent with domain knowledge and provide some plausibility check. For example, PoET recovers edges from annual income to both mortgage accounts and current installment accounts, consistent with the expectation that borrowers' financial resources influence their access to and holdings of credit accounts. Although this does not replace ground-truth validation, it provides a limited domain-based assessment of the estimated structure.

AuthorsOpenReview ↗

Response to Reviewer F88Y - W3

(W3) Limited larger-graph experiments.

We thank the reviewer for raising this point. As the reviewer also noted among the strengths, Table 5 in Appendix A.3.6 evaluates larger graphs with p=100p=100 and simple sizes ranging from n=1000n = 1000 to 1000010000.

We acknowledge that p=100p=100 is not large by the standards of all DAG-learning methods. It is nevertheless a meaningful setting for PoET, which repeats multivariate nonparametric smoothing and testing during graph learning. For context, Bühlmann et al. ([1]) report that RESIT, a DAG-learning method combining nonparametric regression with independence tests, was computationally infeasible at p=100p=100. PoET accommodates nonlinear and nonadditive conditional means and checks model assumptions during learning. These add computational cost, and we do not claim scalability to arbitrarily large graphs.

[1] Bühlmann, P., Peters, J., and Ernest, J. CAM: Causal additive models, high-dimensional order search and penalized regression. The Annals of Statistics, 42(6):2526–2556, 2014.

Reviewer_pCK1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PoET, a test-integrated framework for nonparametric Poisson DAG learning from count-valued data. The method combines cross-fitted Conway–Maxwell–Poisson (CMP) kernel smoothing with conditional-uncorrelatedness tests for parent-candidate screening and equidispersion tests for layer assignment, parent expansion, and pruning. The authors establish asymptotic guarantees for the local tests, graph-recovery consistency under correct specification, and consistent violation detection under local misspecification. Experiments evaluate graph recovery and violation detection on simulated data and two real datasets.

Critical strengths
  • The paper addresses a limitation of existing ranking-based dispersion methods, which may select the best-ranked candidate even when none is statistically compatible with equidispersion, by proposing the test-integrated PoET framework.

  • The framework estimates conditional means nonparametrically without prespecified nodewise link functions and uses equidispersion-guided expansion to recover parents missed by screening.

  • The paper establishes asymptotic guarantees for the local tests and graph-recovery consistency under correct specification, together with consistent identification of the region affected by local misspecification under the stated assumptions.

Critical weaknesses
  • PoET builds on existing Poisson DAG identification principles and residual-covariance testing. Its methodological novelty primarily lies in integrating these ideas with nonparametric estimation, parent search, and diagnostic stopping, rather than introducing a new identification principle.

  • Table 8 compares full PoET with a variant that jointly removes screening and expansion while retaining equidispersion testing. The incremental benefit of test-based layer selection over score-based ordering is therefore not established through a controlled comparison.

  • Theorem 5 assumes Lipschitz conditional means, whereas the negative-binomial mechanism in Appendix A.3.3 can produce non-Lipschitz conditional means for nodes with multiple parents. Consequently, the reported negative-binomial experiments are not necessarily covered by the stated consistency guarantee.

Questions for authors
  • Could you clarify the main technical advances in the proposed CMP-kernel tests, particularly which aspects of their construction and theoretical analysis go beyond existing equidispersion and residual-covariance methods?

  • The current ablation study only compares PoET (screening + expansion) with All previous nodes. Could you provide additional ablation studies to evaluate the individual contributions of PoET’s components to the performance improvements?

4 comments
AuthorsOpenReview ↗

Response to Reviewer pCK1 - W1

(W1) Contribution beyond identification.

We thank the reviewer for the opportunity to clearly separate the existing identification principle from PoET's learning and inference contributions. We agree that PoET builds on the former, and we have revised Introduction to state more precisely that our contributions concern the learning framework, modeling flexibility, and its statistical theory.

Beyond forced graph completion. To our knowledge, PoET is the first count-data DAG learning framework to integrate model-assumption testing, diagnostic stopping, and partial-graph output into a single learning procedure. When no remaining node satisfies the local equidispersion assumption, PoET stops, preserves the graph learned so far, and explicitly flags the unresolved region, rather than forcing further graph decisions.

PoET broadens the learning scope, not the identifiability scope. Building on established Poisson-DAG identifiability results, PoET accommodates unspecified nonlinear conditional-mean functions through CMP-kernel smoothing. This allows flexible nonlinear effects and interactions without the prespecified link structures used in [1–3], while mitigating the sparse-cell instability of exact-cell approaches such as ODS [4] by borrowing information across nearby count configurations.

PoET develops new count-kernel inference theory supporting graph-learning guarantees. To the best of our knowledge, PoET is the first DAG-learning framework based on CMP-kernel smoothing. Beyond applying the smoother, we derive CMP-kernel concentration bounds, establish uniform convergence of conditional-mean estimators (Lemma 7), and prove asymptotic validity and power (Lemma 2). These results underpin consistent graph recovery under correct specification (Theorem 4) and violation detection under misspecification conditions (Theorem 5).

[1] Gunwoong Park and Sion Park. High-Dimensional Poisson Structural Equation Model Learning via ℓ1\ell_1-Regularized Regression. Journal of Machine Learning Research, 20(95):1–41, 2019.

[2] Thi Kim Hue Nguyen, Monica Chiogna, and Davide Risso. Unguided structure learning of DAGs for count data. arXiv:2406.04994, 2024.

[3] Wei Zhou, Xin He, Wei Zhong, and Junhui Wang. Efficient learning of quadratic variance function directed acyclic graphs via topological layers. Journal of Computational and Graphical Statistics, 31(4):1269–1279, 2022.

[4] Gunwoong Park and Garvesh Raskutti. Learning Large-Scale Poisson DAG Models Based on OverDispersion Scoring. Advances in Neural Information Processing Systems 28, 631–639, 2015.

AuthorsOpenReview ↗

Response to Reviewer pCK1 - W2 & Q2

(W2/Q2) Individual contribution of PoET's components.

We thank the reviewer for requesting a component-wise evaluation. We have added ablation of PoET's each component in Appendix A.3.11. The suggestion led us to substantially extend the original ablation study by separately removing screening, expansion, and test-based layer selection. The data, seeds, and other tuning choices are matched within these comparisons. The table below reports graph recovery and violation detection separately.

Ablation comparisons on Hub graphs with n=1000n=1000, p=30p=30, and g=2g=2, following the setting in Appendix 3.10. Entries report average F1 scores (standard errors) over 30 replications. A dash indicates that the variant does not produce violation labels.

SettingCorrect model / Graph-recovery F1Binomial / Violation-detection F1
PoET (screening + expansion + testing)0.516 (0.025)0.850 (0.034)
Without screening0.395 (0.036)0.610 (0.044)
Without expansion0.493 (0.025)0.806 (0.022)
Without test-based layer selection0.427 (0.011)—

We consider three variants of PoET. Without screening uses all previously assigned nodes as initial parent candidates. Without expansion retains the screening procedure but does not add further parent candidates. Without test-based layer selection retains the parent-search procedure but selects the node with the lowest overdispersion score at each step instead of applying test-based layer selection. This variant therefore does not produce violation labels.

Full PoET achieves graph-recovery F1 of 0.516, compared with 0.395 without screening, 0.493 without expansion, and 0.427 without testing. Removing screening or expansion also reduces violation-detection F1 from 0.850 to 0.610 or 0.806, respectively. These results support the use of full PoET framework.

AuthorsOpenReview ↗

Response to Reviewer pCK1 - W3

(W3) Lipschitz condition in the negative binomial experiments.

We thank the reviewer for identifying this mismatch between the original negative-binomial experiment and the sufficient conditions of Theorem 5. We agree that the original mechanism need not satisfy the stated Lipschitz condition. Although performance outside these sufficient conditions does not contradict the theorem, the experiment was not necessarily covered by its guarantee. We therefore revised the negative binomial mechanism data generation process and the corresponding results in Section 5.2 and Appendix A.3.4 to satisfy this condition and reran the experiments, retaining the graph generation, coefficient distributions, misspecified-node selection, and PoET tuning from Appendix A.3.3.

For each misspecified node j∈Wj\in W, we draw signed coefficients with magnitudes from Unif⁡(0.15,0.45)\operatorname{Unif}(0.15,0.45) and define

ηj=∣∑k∈Pa⁡(j)βjklog⁡(1+Xk)∣,μjNB=min⁡105,,1+ηj.\eta_j = \left| \sum_{k\in\operatorname{Pa}(j)} \beta_{jk}\log(1+X_k) \right|, \qquad \mu_j^{\mathrm{NB}} = \min\\{10^5,\\,1+\eta_j\\}.

We generate

Xj∣XPa⁡(j)∼NB⁡(2,μjNB),X_j\mid X_{\operatorname{Pa}(j)} \sim \operatorname{NB}(2,\mu_j^{\mathrm{NB}}),

retaining the original size parameter of 2.

For each remaining node j∉Wj\notin W, we bound the original Poisson conditional mean μj0(XPa⁡(j))\mu_j^0(X_{\operatorname{Pa}(j)}) from below by 10−810^{-8} and from above by 10510^5:

\mu_j^{\mathrm{P}} = \min\left\\{10^5,\\, \max\left\\{10^{-8},\mu_j^0(X_{\operatorname{Pa}(j)})\right\\}\right\\}, \qquad X_j\mid X_{\operatorname{Pa}(j)} \sim \operatorname{Poisson}(\mu_j^{\mathrm{P}}).

Relative to the original setting, we replace exp⁡(ηj)\exp(\eta_j) with 1+ηj1+\eta_j in the negative binomial mean and impose fixed bounds on the conditional means. The upper bounds ensure global Lipschitz continuity of the structural and relevant conditional means on the count support, as well as finite fourth moments. The positive lower bound keeps the Poisson means bounded away from zero.

The table below reports the completed results from 100 replications. Across graph families and densities, the average violation-detection F1 scores range from 0.184 to 0.433 at n=1000n=1000 and from 0.887 to 0.941 at n=10000n=10000. Thus, the improvement with sample size is also observed under this mechanism satisfying the Lipschitz condition.

Average violation-detection F1 scores (standard errors) of PoET under the bounded-mean negative binomial mechanism with p=30p=30 and g∈2,4g\in\\{2,4\\}. Each entry is based on 100 replications.

Graph densitynnHubBAER
g=2g=210000.326 (0.031)0.197 (0.026)0.184 (0.018)
g=2g=230000.808 (0.026)0.608 (0.027)0.653 (0.023)
g=2g=250000.899 (0.021)0.798 (0.023)0.806 (0.017)
g=2g=270000.929 (0.019)0.868 (0.019)0.853 (0.015)
g=2g=2100000.940 (0.017)0.887 (0.017)0.903 (0.011)
g=4g=410000.427 (0.035)0.433 (0.038)0.373 (0.019)
g=4g=430000.809 (0.022)0.743 (0.028)0.742 (0.021)
g=4g=450000.879 (0.020)0.837 (0.022)0.858 (0.014)
g=4g=470000.918 (0.016)0.867 (0.020)0.899 (0.012)
g=4g=4100000.941 (0.014)0.915 (0.015)0.925 (0.009)

In addition, thanks to your comment, we noticed that on the integer lattice Zd\mathbb{Z}^d (or N0d\mathbb{N}_0^d), the Lipschitz assumption is not restrictive in practice. Indeed, for any 0<m<M0 < m < M, any function f:Zd→[m,M]f : \mathbb{Z}^d \to [m, M] automatically satisfies the global Lipschitz condition: since ∥z−z′∥≥1\Vert{}\boldsymbol{z} - \boldsymbol{z}'\Vert{} \geq 1 whenever z≠z′\boldsymbol{z} \neq \boldsymbol{z}', we have

∣f(z)−f(z′)∣≤M−m≤(M−m)∥z−z′∥for all z≠z′∈Zd\vert{}f(\boldsymbol{z}) - f(\boldsymbol{z}')\vert{} \leq M - m \leq (M - m)\Vert{}\boldsymbol{z} - \boldsymbol{z}'\Vert{} \quad \text{for all } \boldsymbol{z} \neq \boldsymbol{z}' \in \mathbb{Z}^d

(and trivially for z=z′\boldsymbol{z} = \boldsymbol{z}'). This implies that any function on the integer lattice can be made Lipschitz simply by clipping it from below at a small m>0m > 0 and from above at a large M<∞M < \infty.

AuthorsOpenReview ↗

Response to Reviewer pCK1 - Q1

(Q1) Technical advances in the CMP-kernel tests.

We thank the reviewer for raising this point. We have added a discussion of CMP-kernel related statistical advances that we have established in Appendix A.1.

Compared with classical dispersion tests for parametric Poisson regression ([1]), PoET estimates the relevant conditional means nonparametrically using CMP kernels. Its conditional-uncorrelatedness test shares the residual-product construction used in the generalised covariance measure (GCM; [2]). Our technical contribution is to develop and theoretically justify these CMP-based tests for multivariate count data. This requires the following two results that we have established in our paper.

First, we establish a geometric rate of CMP-kernel concentration. Huang et al. ([3]) showed that the variance of the mean-parameterized CMP distribution converges to zero as the dispersion parameter increases, for each fixed integer mean, without specifying a convergence rate. We establish an explicit geometric rate and use the resulting concentration bounds to control the kernel smoothing bias.

Second, we establish uniform convergence of CMP-based conditional-mean estimators. Uniform convergence is well studied for continuous kernel regression ([4]), while related discrete associated-kernel studies provide pointwise bias and variance results ([5]) or semiparametric asymptotic results ([6,7]). In our count-DAG setting, Lemma 7 establishes a uniform op(n−1/4)o_p(n^{-1/4}) rate for the CMP-based Nadaraya-Watson estimators over an expanding truncated count support.

These results provide the regression-error control needed to establish the asymptotic validity of both local tests in Lemma 2. Combined with cross-fitting and the first-order insensitivity of the residual moments, they ensure that conditional-mean estimation errors are negligible at the n\sqrt{n} scale. Thus, we provide explicit sufficient conditions for calibrated testing with CMP-based nonparametric regression.

[1] Cameron, A. C., and Trivedi, P. K. Regression-based tests for overdispersion in the Poisson model. Journal of Econometrics, 46(3):347–364, 1990.

[2] Shah, R. D., and Peters, J. The hardness of conditional independence testing and the generalised covariance measure. The Annals of Statistics, 48(3):1514–1538, 2020.

[3] Huang, A., Sippel, L., and Fung, T. Consistent second-order discrete kernel smoothing using dispersed Conway–Maxwell–Poisson kernels. Computational Statistics, 37:551–563, 2022.

[4] Hansen, B. E. Uniform convergence rates for kernel estimation with dependent data. Econometric Theory, 24(3):726–748, 2008.

[5] Somé, S. M., and Kokonendji, C. C. Effects of associated kernels in nonparametric multiple regressions. Journal of Statistical Theory and Practice, 10(2):456–471, 2016.

[6] Abdous, B., Kokonendji, C. C., and Senga Kiessé, T. On semiparametric regression for count explanatory variables. Journal of Statistical Planning and Inference, 142(6):1537–1548, 2012.

[7] Senga Kiessé, T., and Rivoire, M. Discrete semiparametric regression models with associated kernel and applications. Journal of Nonparametric Statistics, 23(4):927–941, 2011.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.