Daily Summary for 2026-11-12
daily
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: It's the twelfth of November, twenty twenty-six, and this is the day's research.
Jane: 29 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 FDR-Controlled Variable Selection for Generalized Linear Models and Cox Regression with Virtual Dummies.
The day's research: Tom: We are focusing on tools for complex data structures and uncertainty across many fields.
Jane: The most pressing area is evaluating retrieval augmented generation systems within the R ecosystem.
Lu: ragR provides a unified workflow in R for document ingestion, embedding, and retrieval evaluation.
Meng: This moves evaluation from Python tools into a reproducible R environment for practical experimentation.
Lalam: Is there research on improving statistical inference methods for causal inference?
Tom: Yes, the Modified Egger Intercept Test proposes a new test for horizontal pleiotropy in Mendelian randomization.
Jane: That test corrects bias using a rerandomized IVW estimator, which is better than the standard EI test.
Lu: Another piece addresses variable selection with T-Rex, an FDR-controlled method for generalized linear models.
Meng: T-Rex extends virtual dummy construction to select predictors while controlling the false discovery rate.
Lalam: Are there efforts to formalize uncertainty visualization using a mathematical framework in R?
Tom: Yes, a framework allows users to replace data with random variables while preserving convergence properties.
Jane: That work integrates uncertainty directly into the grammar of graphics for sounder exploratory data analysis.
Lu: There is a finding about high-resolution ensemble members improving forecast skill in operational weather forecasting.
Meng: Augmenting low-resolution forecasts with high-resolution members yields substantial gains when using many high-resolution members.
Lalam: Is there a development on reliably assessing calibration in probabilistic predictions?
Tom: Yes, the Fair Box ordinate transform introduces a fair version of the Box ordinate transform for Gaussian law forecasts.
Jane: This new transformation correctly identifies miscalibration even with moderate sample sizes for operational forecasts.
Lu: There are methodological advances in handling complex experimental designs and model averaging.
Meng: The Sphericity Filter selects the best design by minimizing geometric loss divided by sphericity in noisy experiments.
Lalam: That filtering technique is useful when dealing with constrained mixtures where standard methods emphasize extrapolation errors.
Tom: Renewable model averaging proposes ReMAQ, a novel method approximating model averaging using quadratic optimization for streaming data.
Jane: That allows sequential updates in streaming data analysis without needing access to all historical data.
Lucky paper: 2609.29335: Tom: Okay, so we're diving into FDR-Controlled Variable Selection for Generalized Linear Models and Cox Regression with Virtual Dummies.
Jane: The paper tackles a huge problem in genomics and clinical studies where you have thousands of candidate predictors but only a few are actually associated with the outcome.
Lu: It proposes the Terminating-Random Experiments selector, which controls the false discovery rate by letting synthetic null variables compete with the real predictors.
Meng: The key mechanism here is that instead of storing an n-vector for each dummy, they store just its projections onto an adaptively growing subspace, which is memory efficient.
Tom: That's a specific detail—how it handles the memory bottleneck when you have tens of thousands of variables.
Jane: They state that this virtual dummy construction ensures the virtual and augmented selection paths are equal in distribution, so the FDR control carries over from the T-Rex selector.
Lalam: I see how that connection matters for people who are trying to build scalable variable selection tools right now.
Tom: So, they extend this concept to score-based forward selection for Bernoulli, Poisson, multinomial, and Cox responses by adapting the virtual dummy construction.
Jane: That means it's not just theoretical; it works across different types of response data like counts or survival times too.
Lu: The empirical validation shows that the virtual dummy framework is equivalent to explicit augmentation, and in genomic applications, VD-Multinomial recovers more than twice as many causal SNPs within the target FDR.
Meng: That's a big win for power in disease classification tasks when you have high dimensions to deal with.
Tom: And resource-wise, they claim the memory footprint adds only around two hundred five MB for Poisson, Bernoulli, and Cox instantiations compared to the AD-GLM.
Jane: So it’s not just theoretically sound; it’s actually faster than a competitor in terms of runtime for high-dimensional settings.
Lu: They show absolute VD-GLM runtimes ranging from about twenty-five seconds to thirty-seven seconds, while AD-GLM takes between fifty-eight and eighty-nine seconds.
Meng: That kind of speed difference is what makes it deployable in real-world scenarios where you're running models on production data.
Tom: The conclusion says the selection path is equal in distribution to explicit augmentation, inheriting the FDR control from T-Rex.
Jane: They also note that the loss in true positive rate is most pronounced for multinomial response, but it's negligible for Bernoulli responses.
Lu: Looking ahead, they suggest future work could involve robust forward selectors like the LAD or Huber, or other path algorithms to push this further.
Lucky paper: 2506.22601: Tom: So, we're looking at the Fair Box ordinate transform for forecasts following a multivariate Gaussian law again. Jane, what is the core problem this paper is trying to solve with this approach?
Jane: The core problem is reliably assessing calibration when you have a moderate sample size and you are dealing with multivariate predictions. Other versions of the box transform fail in those conditions.
Lu: It derives the fair version, uF n m S; xzero by looking at the squared Mahalanobis distance D squared which is calculated as xzero minus m T S negative one xzero minus m.
Tom: That calculation looks complicated. How does that specific formula ensure that the distribution of u is standard uniform for any ensemble size n?
Jane: It ensures that for calibrated forecasts, the distribution of u is standard uniform regardless of the ensemble size n, and they confirm this with p-values like zero point three six zero two when compared to theoretical BOT results.
Meng: When they test under different scenarios, like Type one miscalibration where variance is underforecast, how do those D values behave?
Lu: For Type one miscalibration, the D values for the fair sample BOT are around zero point two for the underforecast variance and around zero point one three for the overforecast variance case.
Tom: So even in severe misspecification, like a very low size to dimension ratio, does this method still perform well compared to other versions?
Jane: Yes, it performs well even when other versions fail, and they suggest combining the fair BOT histogram with univariate rank histograms and lower-dimensional subspace fair BOT histograms to reduce the chance of falsely concluding reliability.
Meng: Does this mean we can use this for verifying post-processed multivariate forecasts from things like ECMWF?
Lu: The paper verifies it using operational ECMWF ensemble forecasts for various weather quantities and lead times, showing that when verified against analyses, the fair sample BOT histograms correctly reflect the situation when the forecast is nearly calibrated and samples a Gaussian law.
Tom: That's interesting because it moves us away from just looking at one big number. What does this suggest about how we should interpret uncertainty in these forecasts?
Jane: It suggests that a reliable forecast of the multivariate normal distribution will result in a flat histogram of fair BOT values regardless of ensemble size.
Lu: The shape of the fair sample BOT histograms is qualitatively similar to that of the theoretical BOT histograms for several unreliable scenarios, which isn't true for classical sample BOT versions.
Meng: So, if we are building production systems, does this mean we can trust these calibration checks more with moderate samples?
Tom: It suggests combining the fair BOT histogram with univariate rank histograms to reduce the likelihood of falsely concluding reliability when things get tricky.
Jane: And they also mentioned that you can utilize this fair BOT to verify post-processed multivariate forecasts, which is a practical step for operational use.
Lucky paper: 2512.00312: Tom: Alright, we're looking at Kicking for Goal or Touch? An Expected Points Framework for Penalty Decisions in Rugby Union.
Jane: So, the authors are developing an Expected Points framework to quantify the value of choosing between attempting a shot at goal or kicking to touch during a penalty.
Lu: The core involves constructing two context-aware surfaces: one for the expected points of a possession starting with a lineout based on field location, and another for the expected points of a kick at goal.
Meng: They use multiple linear regression to estimate the Expected Points of a Lineout, noting that the coefficient on meter line is negative, reflecting intuition about score value decreasing as you move away from the try line.
Tom: And they handle double counting across consecutive possessions with the same state by randomly retaining just one representative observation from each group of n consecutive lineouts.
Jane: Then for the penalty kick at goal, they calculate expected points using a probability of success based on distance and lateral angle, which is modeled as a logit function.
Lu: The continuation value of a miss is approximated by looking at restarts that don't follow a change in score and computing their next expected score using the same Expected Points framework.
Meng: The decision quantity they use to map out strategy is the difference between the expected points of a lineout and the expected points of a kick at goal.
Tom: That difference is what shows where kicking for goal or kicking to touch maximizes your return, and positive values favor the lineout.
Jane: They also look at scenario analyses showing how yellow cards and team quality shift that indifference frontier, meaning stronger teams realize larger expected points from lineouts across most of the pitch.
Lu: A case study shows that sub-optimal decisions were frequent but usually near the decision boundary, leading to a total regret in expected points that was modest relative to the final score margin.
Meng: I'm a little concerned about combining phase-level data from club competitions with penalty-kicking data from international matches; that introduces potential inconsistencies in player quality.
Tom: That’s a fair point, but they also acknowledge that selection bias might influence the results because stronger teams are more likely to generate attacking lineouts in the opposition half.
Jane: What about the limitations on what this framework covers? They only restrict team choices to kicking for goal or kicking to touch, which suggests extending it to scrums would be a natural next step.
Lu: The authors also flag that incorporating player- and team-specific models for both goal-kicking accuracy and meters gained to touch could produce more personalized decision maps.
Lucky paper: 2511.08088: Tom: Alright, we're moving on to paper number four of five today, this one is called Who's Afraid of the Wallenius Distribution.
Jane: It looks like this paper deals with analyzing contingency tables with two or more groups and fixed margins using the Wallenius distribution.
Lu: The authors are showing that for two-by-two tables, it’s straightforward, but for tables with more than two groups, it gets much more nuanced.
Meng: It introduces the Wallenius noncentral hypergeometric distribution for analyzing contingency tables with fixed row margins and a sample size.
Tom: So we have weights which are positive and sum to one defining the distribution on a regular simplex.
Jane: The univariate setting looks like an urn model where the probability of selecting a ball is equal to its weight divided by the total weight in all balls in the urn.
Lu: I'm interested in how they approach this from different statistical lenses, like frequentist or Bayesian inference.
Tom: The frequentist lens involves constructing a confidence distribution and attendant confidence density by setting the function C of w equal to Prob of X greater than or equal to 41w.
Jane: And the maximum likelihood estimate for that weight w is given as zero point six thousand two hundred eighty-seven.
Meng: The Bayesian lens involves a posterior distribution proportional to the likelihood times the probability density function of the beta distribution for w.
Lu: For the multivariate setting, they introduce a novel Markov chain Monte Carlo algorithm called sphere walk Metropolis, or SWM.
Tom: SWM is implemented in Cartesian coordinates on the reference simplex and uses a Metropolis filter in barycentric coordinates on the regular simplex.
Jane: They also perform a weight transformation to yield an unconstrained optimization problem by transforming weights into working parameters.
Meng: They use scaling constants to scale very small probabilities without compromising the elements of inference because they are most generally based on ratios of the likelihood.
Lu: The study uses datasets from literature, including a genetic study about rabbit neonates, to illustrate the application of this distribution framework.
Tom: The analysis is delineated in Section six for multiple three by two contingency tables, where the maximum likelihood estimate for the weights was obtained as wb equals zero point three hundred forty-six comma zero point two hundred twenty-eight comma zero point
Lucky paper: 2610.36012: Tom: So we're looking at this paper, A Sphericity Filter for Space-Filling Designs in Noisy Experiments, and it’s about how to select the best experimental design when you have noisy multiresponse experiments with constrained mixtures.
Jane: The core idea is this sphericity filter which selects a complete design by minimizing geometric loss divided by sphericity without changing the points themselves.
Lu: It’s interesting because it mirrors the advantage of I-over A-optimality for highly constrained mixture designs, and it develops a low-order working model to protect the estimation precision we are seeking.
Meng: So this filter selects one complete design from a candidate pool by minimizing that score Q equals g of Q divided by S minus one of CB. What is the secondary control in this process?
Tom: The secondary control is the gI selector, which minimizes geometry times the complete I nu, and it uses coordinates matched to a prediction region.
Jane: That sounds like they are trying to balance uncertainty in the working model coefficients using the raw A criterion's shape factor.
Lu: The gains reported are about three point three percent reduction in root risk under OLS and GP fitting, and one point four percent relative improvement under nonlinear GP fitting at a median MaxPro increase of zero point six one percent.
Tom: So it improved both summaries over best geometry in five out of six bounded settings, which is promising for practical applications.
Jane: But the paper also notes that model-based designs can trade this off, reducing quadratic OLS root risk by twenty-three percent but increasing nonlinear GP root risk by fifty-seven percent.
Meng: That sounds like a trade-off we have to consider when deploying models at scale, especially if we are dealing with noisy data streams.
Lu: The evidence for the gains depends heavily on the working models, targets, truths and fitting policies that you choose to study.
Tom: I'm curious about the constrained mixture results specifically; how does it compare to raw A-optimal reference?
Jane: The raw A optimal reference is the most spherical in Scheffé coordinates but among the worst for prediction inside the region because it weights named coefficient variances equally.
Lu: The I optimal reference has a lower I nu than the filtered choice and a higher ACD than every FFF candidate in its pool, which suggests that target filtering is a reasonable default when plausible low-order effects deserve protection.
Meng: If we are using raw filtering, it's a practical alternative in cube coding when the target moments are inconvenient for us.
Tom: What about the U bridge and Bridge designs? Do they combine space filling with working model precision effectively?
Jane: The U-bridge and U-bridge designs can combine space filling with working model precision, but they involve larger geometric costs in some comparisons.
Lu: When the true mean is plausibly contained in the working model, a full model OLS fit has no approximation bias, which is a key point there.
Meng: That’s good because it tells us when we can trust a simpler model structure even if we are using complex experimental designs.
Tom: So we're seeing that this sphericity filter helps us protect low-order effects in noisy experiments while maintaining predictive power.
Jane: Exactly, it seems to be a solid tool for improving regional working-model precision across different data structures.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates