Daily Summary for 2026-11-11

daily

Transcript

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

Tom: It's the eleventh of November, twenty twenty-six, and this is the day's research.

Jane: 26 new papers came out today.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We'll take the day in one pass, then pull out the papers we're staying with.

Jane: Start us off with Graph Canonical Coherence Analysis.

The day's research: Tom: We need to handle conflicting answers from different judges or models. Existing methods just flatten the disagreement into one simple label and hide uncertainty.

Jane: That is tricky in pairwise comparisons between human judgments and large language models where item ambiguity obscures things.

Lu: BayesJudge introduces an online Bayesian meta-evaluation layer to handle those conflicting streams of human and LLM judgments by treating latent preference as a variable.

Meng: This approach separates response quality from evaluator bias by jointly modeling rater-specific confusion matrices for human annotation bias and an explicit order parameter for LLM judges.

Lalam: It results in posterior distributions over winning responses instead of single hard labels. BayesJudge shows this accurately recovers biases online on synthetic data.

Tom: That contrasts with other techniques like Drift-Aware Spectral Conformal Prediction for Non-Exchangeable Streaming Data which narrows inference intervals in streaming data based on tuning budget.

Jane: gBottomup’s graph-based bottom-up framework is useful when structural breaks in high-dimensional sequences are densely clustered offering a more adaptive segmentation than traditional top-down procedures.

Lu: Calibration of Time-Varying SEIR Models provides insights into how direct forward solver least absolute deviations performs in fitting epidemic models, favoring LAD for isolated spikes.

Meng: Causal Network Discovery from Interventional Count Data with Latent Linear DAGs tackles identifying latent causal structures in count data using a latent linear Gaussian directed acyclic graph model.

Lalam: A Fully Specified Bayes Factor Range for Hypothesis Testing and Sensitivity Analysis in Process Tracing provides a rigorous way to compute Bayes factors by fully specifying the generative model.

Tom: The most significant finding involves a rank-based debiased machine learning framework for partially linear accelerated failure time models providing valid inference even under right censoring.

Jane: This method achieves near-nominal coverage rates up to sixty percent in simulation studies contrasting sharply with biased results from the nonorthogonal plug-in estimator.

Lu: This framework applies to electronic health record data from adults with diabetes and advanced chronic kidney disease estimating links between serum albumin and systolic blood pressure.

Meng: Another piece addresses measuring disease burden through recurrent events by developing a novel non-parametric estimator combining cohort and population data.

Lalam: This estimator shows convergence to a mean zero Gaussian process providing a consistent way to estimate the covariance function for excess recurrent events.

Tom: The papers cover BayesJudge, gBottomup, Calibration of Time-Varying SEIR Models, Drift-Aware Spectral Conformal Prediction for Non-Exchangeable Streaming Data, Graph Canonical Coherence Analysis, Causal Network Discovery from Interventional Count Data with Latent Linear DAGs.

Jane: We also have A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials and A non-parametric estimator for excess recurrent events.

Lu: A Fully Specified Bayes Factor Range for Hypothesis Testing and Sensitivity Analysis in Process Tracing is another key finding we reviewed.

Meng: Debiased Machine Learning for Partially Linear Accelerated Failure Time Models provides a rank-based framework to estimate exposure effects under right censoring.

Lalam: Comparative e-backtesting develops a parametric anytime-valid framework for comparing risk measures using e-values and e-processes in financial backtesting.

Tom: A non-parametric estimator for excess recurrent events extends relative survival to recurrent events by combining cohort and population data.

Jane: No Universal Winner in Model Selection compares regression versus representational similarity analysis using Monte Carlo simulations and empirical data.

Lu: Evidence on Slowing Progress in Longevity -- Is it Misleading? uses a Bayesian change-point framework to show flattening is due to low variance rather than mean pace decline.

Meng: Exploring complex dependence structures using Bayesian bi-clustering and log-linear graphical modelling uses Bayesian partitioning to reveal complex dependence structures between subjects and categorical variables.

Lalam: Streaming robust covariance estimation and sequential outlier detection proposes an online method for simultaneously estimating the geometric median and variance in streaming data.

Tom: The price of a weak shadow variable investigates inference costs when the relevance of a shadow variable approaches zero in nonignorable missing data settings.

Jane: Estimation and inference in penalized function-on-function regression with within-curve residual dependence analyzes how within-curve dependence affects point estimates and confidence intervals.

Lu: Quantifying Long-Range Dependence in Object-Valued Time Series develops an intrinsic framework for defining long memory in object-valued time series using the trace of an isometric embedding's lag-covariance operator.

Meng: Causal Inference for All introduces new estimands to characterize potential outcome trajectories in longitudinal studies where outcomes are truncated by death.

Lalam: Invariant Image Reparameterisation bridges symbolic and numerical methods for parameter non-identifiability analysis proposing invariant image reparameterization.

Tom: A Bayesian Approach for Earthquake Impact Modelling develops a generalized Bayesian approach within a likelihood-free framework to generate earthquake impact estimates across mortality and displacement.

Jane: Spatial fusion of DHS and MICS surveys with jittered and geomasked locations proposes a spatial fusion model to account for positional uncertainty in survey data from different sources.

Lu: Rank-Transformed Dissimilarity Profiles for High-Dimensional Classification proposes a dissimilarity-profiling classification framework that transforms high-dimensional features into class-wise rank profiles.

Meng: MML Probabilistic Principal Component Analysis describes a new approach to automatically selecting the number of principal components in probabilistic PCA using the Bayesian minimum message length method.

Lalam: A weighted edge-count two-sample test for multivariate and object data proposes a modified edge-count test that uses within-sample edges with different weights to improve power.

Tom: The list includes BayesJudge, gBottomup, Calibration of Time-Varying SEIR Models, Drift-Aware Spectral Conformal Prediction for Non-Exchangeable Streaming Data, Graph Canonical Coherence Analysis, Causal Network Discovery from Interventional Count Data with Latent Linear DAGs.

Jane: We also have A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials and Debiased Machine Learning for Partially Linear Accelerated Failure Time Models.

Lu: A non-parametric estimator for excess recurrent events extends the notion of relative survival to recurrent events by developing a novel non-parametric estimator combining cohort and population data.

Meng: Comparative e-backtesting develops a parametric anytime-valid framework for comparing risk measures using e-values and e-processes in financial backtesting.

Lalam: A Fully Specified Bayes Factor Range for Hypothesis Testing and Sensitivity Analysis in Process Tracing shows how to compute fully specified Bayes factors for process tracing evidence.

Tom: The paper on Debiased Machine Learning provides a rank-based framework to estimate exposure effects in partially linear accelerated failure time models under right censoring.

Jane: A non-parametric estimator for excess recurrent events extends relative survival to recurrent events by developing a novel non-parametric estimator combining cohort and population data.

Lu: Comparative e-backtesting develops a parametric anytime-valid framework for comparing risk measures using e-values and e-processes in financial backtesting.

Meng: A Fully Specified Bayes Factor Range for Hypothesis Testing and Sensitivity Analysis in Process Tracing shows how to compute fully specified Bayes factors for process tracing evidence.

Lalam: The list includes A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials, Debiased Machine Learning for Partially Linear Accelerated Failure Time Models, and A non-parametric estimator for excess recurrent events.

Tom: The paper on Debiased Machine Learning provides a rank-based framework to estimate exposure effects in partially linear accelerated failure time models under right censoring.

Jane: A non-parametric estimator for excess recurrent events extends relative survival to recurrent events by developing a novel non-parametric estimator combining cohort and population data.

Lu: Comparative e-backtesting develops a parametric anytime-valid framework for comparing risk measures using e-values and e-processes in financial backtesting.

Meng: A Fully Specified Bayes Factor Range for Hypothesis Testing and Sensitivity Analysis in Process Tracing shows how to compute fully specified Bayes factors for process tracing evidence.

Lalam: The list includes Comparative e-backtesting, A non-parametric estimator for excess recurrent events, and a few others.

Lucky paper: 2601.09038: Jane: So we're looking at Graph Canonical Coherence Analysis, which extends canonical correlation analysis to multivariate graph signals in the graph frequency domain.

Tom: It aims to discover relationships between two multivariate graph signals defined on a common underlying graph by maximizing their coherence in the spectral domain.

Lu: The motivation here is addressing challenges with graphs like discreteness and irregularity, using the Brillinger frequency-domain approach to connect time-domain and graph-domain analyses.

Meng: How does this framework actually handle the estimation steps? What are we looking at when we talk about Graph Cross-Spectral Densities?

Jane: Step one involves estimating those Graph Cross-Spectral Densities using methods like the graph crossperiodogram. Then, step two estimates canonical graph signals by estimating those filters H hat lambda l and F hat lambda l based on eigenvectors of spectral matrices.

Tom: And the final step is constructing the canonical graph signals Z hat i and W hat i using those filters along with the original signals X one through X p and Y one through Y q.

Lu: The interpretation of these canonical graph signals happens in the graph frequency domain, representing how much variance in the graph Fourier coefficients is linearly shared with those canonical graph signals.

Meng: The authors introduce "graph canonical communality" as the sum of graph canonical loadings for each signal, and "graph canonical adequacy" as the average of all those loadings for a given set relative to a single signal.

Jane: Empirical validation was done on economic and energy datasets from G20 countries and the USPS handwritten digit dataset. The results showed that canonical coherences decrease from Z hat one W hat one to Z hat five W hat five across all graph frequencies.

Tom: That suggests gCChA can be effectively applied in image classification tasks, which is interesting since they used the USPS data for comparison.

Lu: Promising future research involves extending this framework to accommodate more than two graph datasets for canonical dependence analysis or exploring application areas like brain functional networks and environmental sensing networks.

Meng: I'm curious about the comparison between gCCA and gCChA, since the authors mention it's based on training accuracy due to differences in how the graphs were constructed between training and test sets.

Jane: That is a fair point; those differences in graph construction can certainly impact how accurate those comparative results are.

Tom: It seems like a solid extension of canonical correlation analysis into a frequency domain that accounts for the specific structure of graphs.

Lu: We also see other work touching on related ideas, like Graph Canonical Coherence Analysis which proposes extending canonical correlation analysis to multivariate graph signals in the graph frequency domain.

Meng: Another paper we saw is Causal Network Discovery from Interventional Count Data with Latent Linear DAGs, which uses a latent linear Gaussian directed acyclic graph model for causal discovery from interventional count data with Poisson measurement error.

Jane: That tackles identifying latent causal structures in count data using that specific model structure.

Tom: And we also have Calibration of Time-Varying SEIR Models, which develops a direct forward solver least absolute deviations calibration method to fit time-varying epidemic models to reporting interval incidence data.

Lu: It's notable that LAD is favored in many comparisons when dealing with isolated spikes in those epidemic models.

Meng: How does the complexity of the graph structure affect the estimation of those Graph Cross-Spectral Densities? That seems like a major bottleneck for any real deployment.

Jane: Well, the authors note that enhanced estimation strategies for GCSDs remain an area of interest for future work. It sounds like they're still pushing on improving how fast and accurate we can get those initial density estimates.

Tom: It’s good to see that while the framework is sound, there’s a clear path forward in improving the estimation techniques themselves, rather than just applying it more broadly right now.

Lu: I agree, that suggests the theoretical foundation is strong enough for further refinement on the practical estimation side.

Lucky paper: 2505.21447: Tom: Alright, we're looking at A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials. The core problem here is handling patients who don't survive to get those non-terminal outcome measurements in CRXO trials.

Jane: So what the authors are tackling is how to get a causal effect estimate when you have truncation by death, and they define this as the survivor average causal effect or SACE, which is a subgroup treatment effect for always-survivors.

Lu: They set up six specific structural and modeling assumptions to identify this SACE, including one that says potential outcomes are only governed by the treatment assigned to their cluster period.

Meng: Their Bayesian estimation methodology augments the posterior with latent principal stratum membership so conditional posteriors become more tractable for Markov chain Monte Carlo sampling.

Lalam: They complete the observed-data likelihood by including each individual’s principal stratum membership, which lets them use the EM algorithm or define an augmented posterior distribution that is suited to simpler Markov chain Monte Carlo methods.

Tom: In simulation studies, Model one shows low bias and achieves close to nominal coverage for both the difference and ratio of means SACE estimands across all scenarios considered.

Jane: That’s good because it contrasts with other techniques where performance depends heavily on a tuning budget, like Drift-Aware Spectral Conformal Prediction.

Lu: They apply their methods to the PEPTIC trial, getting estimates for the SACE in difference in means of zero point zero six eight and ratio of means in days of one point zero six three with ninety-five percent HPD credible intervals.

Meng: Those credible intervals suggest a statistically significant increase in hospitalization length of stay for those assigned to PPIs versus H2RBs among always-survivors in the PEPTIC trial.

Tom: But they also mentioned some limitations, specifically assuming survival monotonicity, which is untestable and could be challenged if we look at comparative effectiveness studies.

Jane: And they noted that the proximity of these SACE estimates might just be because the classified always-survivors are similar to the observed survivors in those studies.

Lu: They also showed that Model one performs best overall relative to simpler comparator models in terms of bias, root mean square error, and coverage in aggregate.

Meng: The authors cautioned that the SACE effects described by this paper can only be interpreted from an intention-to-treat perspective because of that assumption.

Lalam: They also pointed out that Model one and Model two may still give rise to some undercoverage, suggesting a remedy might be employing different families of priors like half-Cauchy priors.

Tom: What about the algorithmic framework itself? How does the Gibbs sampler work in practice for drawing from that posterior distribution?

Jane: The framework involves ten steps, including initializing components of theta ext and augmentation variable g, followed by updates for nonmortal regression and SACE parameters.

Lu: For the posterior update for LDiff, which is the difference in means on the log-time scale, it's approximated by E(muLDiffdata) which involves terms like alpha one one one plus beta T one one one times Xij k dF.

Meng: And for ROM, the ratio of means on the time scale, it's given by an expression involving exp alpha two X plus delta T two kappa plus beta T two j i k o exp beta T two X plus delta T two kappa.

Lalam: For the PEPTIC trial specifically, they obtained estimates for LDiff as zero point zero six eight with a credible interval of zero point zero one six to zero point one one eight in log-days and ROM as one point zero six three with an interval of one point zero zero eight to one point one one eight in days, right?

Tom: Those results show a statistically significant increase in hospitalization length of stay for those assigned to PPIs versus H2RBs among always-survivors, which is what we were looking for.

Jane: It's important to remember that the authors are using data augmentation with P´olya-Gamma random variables to represent the contributions of individuals to each principal stratum.

Lu: They demonstrated that including cluster-period random effects in the outcome model was necessary to maintain coverage across all datagenerating processes.

Meng: They also showed that Model one performs just as well if not better than Model A, which explicitly includes those cluster-period random effects.

Lalam: The paper suggests a remedy for undercoverage might be employing half-Cauchy priors instead of the diffuse inverse gamma priors used in their setup.

Lucky paper: 2608.07841: Tom: Okay, we’re looking at Debiased Machine Learning for Partially Linear Accelerated Failure Time Models today. It’s about rank-based U-statistic inference in partially linear accelerated failure time models to give valid inference under flexible nuisance estimation.

Jane: So, the core idea here is they are taking a semiparametric partial linear accelerated failure time model and letting machine learning handle the flexible nuisance component while keeping the main exposure part linear.

Lu: They build on the double/debiased machine learning framework and use a partialling-out principle. For right-censored outcomes, they use a rank-based U-statistic estimator built on pairwise comparisons to extend guarantees to those censored results.

Meng: What’s the main gain from this orthogonalized rank based U-statistic with the censoring corrected influence function? I need to know if it actually improves sample efficiency over standard cross fitting.

Tom: The key contribution is that they developed an orthogonalized rank based U-statistic with a censoring corrected influence function, which they present as the first Neyman orthogonal estimator of this type. They also introduced a block pairwise cross fitting scheme to decouple pairwise moments while keeping the sample efficiency of standard cross fitting.

Jane: That decoupling sounds important for managing the complexity when you have these multiple streams of judgment we talked about earlier.

Lu: They establish n over n consistency and valid asymptotic inference on the primary exposure effect under ML nuisance estimation, which they prove using Neyman orthogonality eliminating the first order nuisance effect and product rate conditions controlling the second order remainder.

Meng: So if it eliminates that first order effect, does that mean we don't have to worry about bias from that specific part of the model anymore?

Lalam: Proposition two proves n over n (beta hat minus beta zero) converges to a normal distribution with variance V over A squared, where A and V are defined by the Jacobian and Hoeffding projection variance.

Tom: Simulations confirm finite sample performance under various error distributions, showing that both the plug in and the proposed Orth variants achieve near nominal coverage across all settings, with the Leurgans weighting closely tracking the oracle benchmark.

Jane: That's good to hear about coverage rates being near nominal, but what about how this translates to real-world data? I’ve heard they applied it to electronic health record data from adults with diabetes in the All of Us Research Program cohort.

Lu: They focused on patients with chronic kidney disease stage 3b or worse before diabetes diagnosis, using pre index serum albumin as the exposure, and found that the orthogonal correction further attenuates estimates to beta hat equals zero point two zero and beta hat equals zero point one nine at the five percent level.

Meng: So those specific results are significant for their application in chronic kidney disease patients with diabetes, but what about heavy censoring? The discussion section mentions future work might improve performance under heavy censoring because of the nuisance estimation under censoring.

Tom: The discussion section interprets these results as estimates of association unless a causal design is in place, which is a necessary caveat we need to keep in mind for deployment.

Jane: So, putting it all together, this paper introduces a rank based debiased machine learning framework that uses a censoring corrected influence function and cross fitting to achieve valid inference even when the nuisance functions are estimated via machine learning.

Lu: The practical performance really hinges on that nuisance estimation, specifically for the term l zero equals E of y sub T Z under censoring, which is what they flagged as a future work area.

Lucky paper: 2609.26328: Tom: So we are looking at A non-parametric estimator for excess recurrent events today. This paper tackles how to measure disease burden by extending relative survival to recurrent events using a novel estimator that combines cohort data with population level data.

Jane: It seems like the main motivation is bypassing the problem where recorded reasons don't accurately capture the true disease burden, which is a real issue in public health research.

Lu: The paper formulates an additive excess rate model where individual recurrent events rates are decomposed into an excess rate and a general population rate, which gives us the expected number of excess recurrent events up to time t as mu E of t equals Z t zero dLambda R E of u S of u.

Meng: That formula looks complex; what exactly is that mu E of t supposed to represent in practical terms for someone deploying a model?

Lalam: The marginal non-parametric estimates for survival S hat and excess recurrent events rate dLambda hat R E are defined using product limit estimators, and the final estimate mu hat E of t is obtained by plugging those into the equation.

Tom: And theoretically, they have three theorems detailing the weak convergence of this estimator to a zero mean Gaussian process. That sounds like a lot of rigorous proof to get there.

Jane: Theorem one shows that the right continuous version of a specific process converges weakly as n goes to infinity to a multivariate zero mean Gaussian process. What does that convergence tell us about the reliability of this estimate?

Lu: Theorem two proves that the process involving the estimator itself converges weakly to a zero mean Gaussian process W equals W of t with covariance function v(s, t).

Meng: So, in simulation, they tested this on sample sizes n equals five hundred and n equals two thousand. Did those results show that the standard error performed well in finite samples?

Lalam: Yes, the simulations showed that the expected number of excess recurrent events estimator mu hat E of t and its standard error perform well in finite samples.

Tom: They applied this to Swedish data on patients with rectal cancer and found an estimated number of excess hospitalisations at the end of study was four point eight seven with a ninety-five percent confidence interval of four point zero two, five point seven three.

Jane: Forty-eight point eight seven excess hospitalisations over the study period; that's a substantial burden on both patients and the healthcare system, right?

Lu: The paper notes this estimator is based on a marginal interpretation averaging over both disease related and non-disease related patient characteristics to provide an estimate of the total burden of disease.

Meng: But they also admit it doesn't produce a causal estimate because overall survival is part of the estimand; that’s a key limitation we need to keep in mind when deploying this kind of model.

Lalam: The paper concludes that earlier detection would be of immense value both to society and patients, suggesting stratification might be preferable for individual patient decision making.

Lucky paper: 2608.22375: Tom: Alright, we're looking at Exploring complex dependence structures using Bayesian bi-clustering and log-linear graphical modelling today.

Jane: The paper is from the University of St Andrews and it talks about how Bayesian partitioning simultaneously on subjects and categorical variables reveals complex dependence structures through views.

Lu: It uses variable selection to highlight which variables drive the clustering of subjects within each view, which relates to marginal independence and conditional independence.

Tom: I'm curious about that connection between mixture modeling, which benefits from many variables, and log-linear graphical modeling that explicitly describes dependence structure.

Jane: The paper shows how bi-clustering results inform a summary matrix T which translates the clustering output into information relevant to selecting a graphical log-linear model.

Meng: If we look at the empirical validation, simulation studies show the mean squared differences between the underlying true T matrices and those from single-view clustering are roughly twice as large for three simulations.

Tom: That's a significant gap if that much difference exists between the methods being compared in terms of model search efficiency.

Jane: In real data analyses for coronary heart disease, the matrices T CHD and T CHD,SV are quite similar because the first five variables are placed in the same view by the bi-clustering.

Lu: That similarity suggests it captures better edges like 'AD' and 'DE' in that specific context.

Tom: So, when using this approach for GRACE data, utilizing bi-clustering information reduces the average number of iterations to the best model compared to single-view clustering or a completely random search.

Jane: That’s an efficiency gain when you’re looking at those high-dimensional sequences.

Meng: The authors note that while bi-clustering informs the model search algorithm, it is not used to form the log-linear modelling prior distributions.

Lu: They suggest investigating if those two priors match or how they should be formed to effect prior matching in future work.

More episodes

← Home