acceptodds
Under review as a conference paper at ICLR 2027

Multi-Method Inference for Critical Thresholds from Sparse Observational Data

Abstract

Critical thresholds are often latent functionals rather than directly observed breakpoints. Under sparse observations, structurally different estimators can return similar numerical locations while relying on incompatible inverse mappings or targeting different quantities. We introduce a qualification-first framework that fixes the scientific target before model choice, profiles each method's native inferential object and failure modes, and uses known-truth validation to assign one of four roles: target-responsive, calibrated within a stated validation range, native diagnostic, or abstention. In a compact mechanistically aligned but observation-model-mismatched experiment at truths 0.90, 0.94, and 0.97, a hierarchical Bayesian model (HBM) responds monotonically but compresses near the upper end and is positively biased at the two lower truths (RMSE 0.032, 0.026, and 0.006); change-point detection (CPD) does not track the mechanistic target and remains native-only. Applied to 10,153 U.S. county-years from 2,212 counties during 2018–2023, the HBM posterior median unconstrained coverage crossing is 0.991, with 31.8% of posterior draws implying no feasible crossing at coverage under . CPD returns a nonsignificant native breakpoint of 0.992 (), while a biologically informed neural network (BINN) abstains because profile support spans 0.88–0.99. The framework therefore assigns different inferential roles and performs no fusion: numerical proximity alone is not evidence that heterogeneous methods recover the same latent threshold.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
critical threshold inference, scientific machine learning, learning from sparse data, inductive bias, model misspecification, uncertainty quantification, statistical inverse problems, identifiability, hierarchical Bayesian modeling, change-point detection, biologically informed neural networks, computational epidemiology

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yjCGOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies estimating a critical threshold, the value of a control variable at which a latent system changes state, from sparse data. The case study is the measles herd-immunity threshold: the MMR coverage at which the population-average effective reproduction number falls to one.

The authors propose a “qualification-first” framework: fix the target quantity, profile each method’s native estimand, and use a known-truth simulation to assign each method a role (target-responsive, calibrated, native diagnostic, or abstaining). Methods are fused only if at least two are calibrated for the same target. Three methods are tested: a hierarchical Bayesian model (HBM), change-point detection (CPD), and a biologically informed neural network (BINN).

In a simulation with three true thresholds and three replicates each, the HBM tracks the truth with positive bias, while CPD’s breakpoint moves the opposite way. On 10,153 US county-years (128 with cases), the HBM gives R0=25.87R_0 = 25.87 and a threshold of 0.991, CPD returns a non-significant breakpoint at 0.992, and the BINN abstains. No fusion is performed.

Critical strengths
  1. Critical-threshold estimation is an important problem, especially when simulating and calibrating epidemic models against observations. The core caution is correct and well illustrated. Similar numbers from different methods are not evidence they estimate the same quantity.
  2. Honest reporting. The paper refuses to fuse, lets the BINN abstain, and flags its own estimate as extrapolative.
Critical weaknesses

The problem is never made concrete, and it is smaller than the framing suggests. The measles threshold is the closed-form herd-immunity relation τ=(1−1/Rˉ0)/η\tau = (1-1/\bar R_0)/\eta with η\eta fixed at 0.97, so the task reduces to estimating a mean R0R_0 from county-year counts. The “known-truth” generator (App. B) is a static regression with no transmission dynamics. The main text never states this, so a reader cannot tell how the threshold enters the model or what CPD and the BINN add. The framework itself (fix the target, state each estimand, validate by simulation, abstain when unidentified) is standard identifiability and simulation-based calibration practice. The contribution is a vocabulary for it plus one case study that yields no estimate.

Questions for authors

Clarifications

  1. Please state in the main text that the threshold is (1−1/Rˉ0)/η(1-1/\bar R_0)/\eta and the task is estimating Rˉ0\bar R_0. What does the framework add beyond standard identifiability checks?
  2. Why does the validation grid stop at 0.97 when the estimate is 0.991?
  3. Why is this paper suited to ICLR? The only learning component abstains.

Suggestions

  1. Replace Fig. 1 with a diagram showing the data, the latent ReR_e, and where each method reads the threshold.
  2. Validate against a dynamic transmission simulator with importation.
Reviewer_aEV4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a qualification-first framework for estimating latent critical thresholds from sparse observations. It evaluates whether different methods target the same quantity before comparing or combining their estimates. A measles case study applies hierarchical Bayesian modeling, change-point detection, and a biologically informed neural network, assigning them different inferential roles rather than pooling their outputs. The central message is useful and clearly presented, but the validation is too limited to establish the framework's reliability.

Critical strengths
  1. The paper clearly distinguishes numerical agreement from agreement about the underlying estimand. Explicitly allowing abstention is valuable when sparse data cannot reliably identify a threshold.

  2. The authors transparently report unsuccessful validation, model dependence, and uncertainty. They appropriately avoid presenting the three methods as independent confirmation of a single threshold.

Critical weaknesses
  1. The qualification procedure is supported by only three replicates at each of three threshold values, without a comparable controlled evaluation of the final BINN profiling procedure. This is insufficient to establish reliable qualification or abstention decisions. The historical estimate of 0.991 also exceeds the largest tested truth of 0.97. Broader simulations should assess both successful recovery and failure detection, including sensitivity to the chosen qualification tolerances.

  2. The main empirical result depends on assumptions whose sensitivity is not evaluated. Transmission enters only the positive-burden component, leaving 128 positive county-years to inform it directly. The protection parameter is fixed at 0.97, and prior, zero-policy, and reference-weighting sensitivity analyses are absent. Although acknowledged, these omissions make it difficult to determine how much the estimated crossing and feasibility probability are driven by data versus modeling choices.

Questions for authors
  1. How reliably does the framework distinguish recoverable from nonidentifiable thresholds across larger simulations, including the final BINN procedure?

  2. How sensitive are the HBM crossing and feasibility probability to the transmission prior, protection parameter, and treatment of unreported county-years?

Reviewer_SR1uOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a qualification framework for estimating latent critical thresholds from sparse observations. To be frank, I don’t see anything particularly innovative in what was actually done.

Critical strengths

N/A

Critical weaknesses
  1. This article is a completely nonsensical piece of AI-generated content, riddled with a vast number of issues. It consists of a multitude of meaningless sections and charts.

  2. The author should carefully verify the validity of the article. The main text is divided into nine sections; however, the subsections and subsubsections within each section are so brief that it is impossible to extract any meaningful information from them. Furthermore, the list of references is extremely sparse, making it unclear what the specific significance of this work is. I do not believe that this article establishes a new field—a scenario that might otherwise explain the lack of relevant references.

  3. Section 4 defines an interpretable decision procedure, but the experiments do not establish how reliably it assigns inferential roles. The paper demonstrates rejection of unsupported claims in one application without evaluating false qualification or unnecessary abstention. The practical justification for the 0.03 recovery tolerance and 0.04 profile-width criterion is also limited.

Questions for authors

Please refer to weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.