Causal DAG Identification for Count Data via Poisson Thinning Structural Equation Models
Penggang Gao, Ming Cai, Hisayuki Hara
stat.ME · cs.LG · stat.ML
Quick verdict
Claims full DAG identifiability for count data via Poisson-thinning SEM; elegant sound theory, but tests cap at d≤10 and real-data i.i.d. fit is questionable.
Summary
The paper introduces the Poisson thinning structural equation model (PT-SEM), a count-valued SEM in which each unit of a parent variable generates a Poisson-distributed number of offspring, so thinning coefficients may exceed one and lie in [0,∞), and node-wise exogenous noises may come from arbitrary count families satisfying mild conditions (positive mass at zero, nonzero variance, finite MGF near zero). The authors prove full identifiability of the DAG, thinning coefficients, and exogenous distributions via a recursive sink characterization based on conditional cumulant generating functions, and show the same analysis makes the binomial-thinning PB-SCM identifiable whenever every nonsink has non-Poisson noise. For learning, they propose a plug-in BIC using method-of-moments estimates of thinning coefficients and family parameters, optimized exactly by subset dynamic programming, with consistency for DAG and family selection (Theorem 18). Simulations (d=4–10, N=100–10,000, 100 replications) show directed-edge F1 close to an oracle that knows the true exogenous families, clearly beating ODS, PC-RCIT, and PB-SCM baselines, though exact DP is exponential in d. An NBA play-by-play application (95,808 team-quarter observations, 10 seasons) recovers a prespecified five-node reference DAG perfectly in every season with stable coefficients (FOUL→FTA 1.187–1.297). Main limitations: simulation scale capped at d=10, i.i.d. assumption strained in the real data, and the reference DAG is author-constructed.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under (A1)–(A3), the observational distribution of a PT-SEM uniquely determines the causal DAG, the thinning coefficient matrix, and the node-wise exogenous distributions (Theorem 12).
- The binomial-thinning PB-SCM becomes fully identifiable whenever every nonsink node has non-Poisson exogenous noise (Theorem 14); Poisson noise is the only obstruction.
- The plug-in BIC with method-of-moments estimates, optimized by exact subset dynamic programming, is consistent for both DAG and exogenous-family selection (Theorem 18).
- In simulations up to d = 10, the method approaches oracle accuracy and outperforms ODS, PC-RCIT, PB-SCM, and PB-SCM-PGF on directed-edge F1.
- On NBA event counts, the method recovers the prespecified five-node reference DAG in all ten seasons with stable coefficients (FOUL→FTA between 1.187 and 1.297).
Strengths
Complete identifiability theory with full proofs
The sink characterization via conditional cumulant generating functions (Prop 10, Thm 12, App A.2) is modular and verifiable; the bivariate contradiction argument (Lemmas 5–7) reduces to a short polynomial-in-z computation forcing ω2 = δ2γ1β > 0 to vanish, and Lemma 11's positive-definiteness proof via Var(W) + a_r²Var(εr) is clean.
Substantive generalization of PB-SCM with a clarifying corollary
Coefficients move from (0,1] to [0,∞) and exogenous noises become node-wise heterogeneous; Theorem 14 shows Poisson noise is the sole obstruction to full identifiability under binomial thinning, sharpening Qiao et al. (2024).
Practical exact learner with end-to-end guarantees
Honest, ablation-rich simulation design
Reproducible real-data validation
Weaknesses
Scalability bounded by exact DP
i.i.d. assumption strained in the real application
Reference-DAG circularity
Baseline coverage gaps
Finite-sample instability of family selection
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets causal structure learning for count-valued data, generalizing the Poisson branching SCM (PB-SCM, Qiao et al. 2024) along two axes: Poisson offspring (thinning coefficients in [0,∞), interpreted as mean offspring per parent unit) and node-wise heterogeneous exogenous families. Three claims are made: (i) full identifiability of DAG, coefficients, and exogenous distributions (Theorem 12); (ii) the binomial-thinning corollary that non-Poisson noise at every nonsink suffices for PB-SCM identifiability (Theorem 14); (iii) a consistent plug-in BIC learner optimized by exact dynamic programming (Theorem 18). All three carry complete proofs in Appendix A. The scope is explicitly causally sufficient, acyclic, fixed-dimension settings (Section 7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is PT-SEM in simple terms?
PT-SEM (Poisson thinning structural equation model) is a statistical model for count (nonnegative integer) data in which each unit of a parent variable spawns a Poisson-distributed number of offspring that flow into a child variable. Unlike earlier binomial-thinning models (PB-SCM), edge coefficients can exceed 1, and each node's background noise may follow any count distribution such as Poisson, negative binomial, or zero-inflated Poisson.
What does the paper actually prove?
That under mild conditions (each noise variable has positive probability at zero, nonzero variance, and a finite moment generating function near zero), the full causal DAG, all edge coefficients, and each node's noise distribution are uniquely determined by the observational distribution. It also proves that a BIC-based algorithm using moment estimates and exact dynamic programming recovers the correct graph and noise families as sample size grows.
Is code or data available for this paper?
Yes. The authors provide code and reproducibility materials at github.com/KOUHOUKOU/PT-SEM, and the real-data analysis uses the public Kaggle dataset NBA WNBA play-by-play and shots data, version 8, aggregated to team-quarter counts over ten NBA seasons (95,808 rows).
What are the main limitations of this work?
The exact dynamic-programming search scales exponentially with the number of nodes (experiments stop at d = 10); the real-data analysis assumes independent observations, which pooled sports-event counts may violate; and the ground-truth NBA graph was specified by the authors themselves from league rules rather than by an independent source.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–4 and 6 are rendered as scattered axis/tick text by the PDF extraction, so plotted curve values could not be read directly; quantitative figure claims (e.g., F1 levels) rely on the surrounding prose and on Table 4, which extracted cleanly.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.