Daily Summary for 2026-11-06

daily

Transcript

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

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

Jane: 14 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 Inference on the Reliability Function with Type-I Censored Weibull Data.

The day's research: Tom: The impossibility result shows constructing distribution-free prediction sets is fundamentally unattainable under standard causal assumptions when covariates are continuous.

Jane: That difficulty stems from predicting counterfactuals when covariates shift and the inherent non-identifiability caused by missing data.

Lu: This connects to prior work on conformal inference for counterfactuals, which relies on knowing or accurately estimating the propensity score.

Meng: Conformal meta-learners depend on stochastic dominance conditions that cannot hold uniformly across model classes, however.

Lalam: A positive result was found using the Split-Pair method under specific structural assumptions, suggesting informative prediction is possible there.

Tom: Separately, we looked at handling missing data in density estimation using sieve maximum likelihood approaches.

Jane: That work established a general rate of convergence under non-monotone missing at random MAR that requires no modeling beyond the MAR assumption itself.

Lu: This led to a new ignoring density estimator achieving minimax optimal rates up to logarithmic factors in a Hölder class.

Meng: It is implemented via an expectation-maximization algorithm operating directly on the incomplete data.

Lalam: We also explored exact parametric tests for reliability inference with censored data, specifically for distributions like Weibull and loglogistic.

Tom: The proposed method uses a generalized pivotal quantity approach by transforming the data into a standard normal distribution via the log transformation of the lifetime variable.

Jane: Simulation results showed that this Gumbel-based approach performs substantially better in terms of coverage probability and interval length compared to existing methods like WLMA.

Lu: The framework for longitudinal microbiome data uses Ornstein-Uhlenbeck processes to understand temporal dynamics.

Meng: This incorporates a structured measurement error model for log-ratio transformed abundance data, allowing for likelihood-based estimation of latent microbial dynamics.

Lalam: The study indicated that disruptions primarily altered equilibrium abundance levels, while a small number of genus pairs showed substantial reductions in mean-reversion rates following the event.

Tom: The work on quantifying inconsistent mediation is important because it offers a way to properly describe causal pathways when direct and indirect effects pull in opposite directions.

Jane: This paper introduces the Mediated Magnitude Share MMS and the degree of suppression to give bounded and interpretable measures instead of relying solely on proportion mediated.

Lu: The MMS quantifies how much the indirect pathway's magnitude relates to the combined absolute magnitudes of both direct and indirect pathways, defined as IE divided by DE plus IE.

Meng: The degree of suppression measures the fraction of the total combined magnitude that is lost because opposing directions cancel each other out, defined as one minus TE divided by DE plus IE.

Lalam: These two measures complement the proportion mediated by showing exactly how much attenuation occurs when pathways conflict, which is a key insight for causal analysis.

Tom: The framework is extended to multiple indirect pathways where the total effect is the sum of direct and all indirect effects, allowing suppression among various subsets of pathways.

Lucky paper: 2604.12011: Tom: Alright, we're diving into this paper now, Inference on the Reliability Function with Type-I Censored Weibull Data. So the core problem is reliability inference when you have missing data, which affects everything from electrical parts to clinical trials.

Jane: The paper points out that making distribution-free predictive inference for individual treatment effects is fundamentally impossible under standard causal assumptions when covariates are continuous.

Lu: That impossibility stems from two main issues: the trouble of predicting counterfactuals when covariates shift and the inherent non-identifiability caused by missing data.

Meng: They also mention that conformal meta-learners depend on stochastic dominance conditions that cannot hold uniformly across model classes, which limits their use.

Lalam: But they found a positive result using the Split-Pair method under specific structural assumptions, suggesting informative prediction is possible in those controlled settings.

Tom: Moving into the density estimation part of this paper, they looked at how to handle missing data in density estimation using sieve maximum likelihood approaches.

Jane: They established a general rate of convergence under non-monotone missing at random MAR that requires no modeling of the missingness mechanism beyond the MAR assumption itself.

Lu: This led to a new ignoring density estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class, implemented via an expectation-maximization algorithm operating directly on the incomplete data.

Meng: I wonder about the practical implications of that rate of convergence; how does it translate when we have actual production traffic instead of idealized data?

Lalam: The paper also explored exact parametric tests for reliability inference with censored data, specifically for distributions like Weibull and loglogistic.

Jane: They introduced a generalized pivotal quantity approach by transforming the data into a standard normal distribution via the log transformation of the lifetime variable.

Tom: Simulation results showed that this Gumbel-based approach performs substantially better in terms of coverage probability and interval length compared to existing methods like WLMA, which they say proved overly conservative.

Lu: They used least squares estimators on the linear model where y equals nu plus sigma w, and they derived the generalized pivotal quantity for sigma as Gsigma equals sigma hat over sigma hat of Z.

Meng: So when you compare the Generalized Least Squares approach with WLMA, what's the concrete difference in how much longer those confidence intervals actually get?

Lalam: The average confidence interval lengths obtained using WLMA are much longer than those from the GLA method, suggesting that GLA provides more precise estimates while maintaining appropriate coverage probabilities.

Tom: And conversely, they noted that the bootstrapping method shows under-coverage, with coverage probabilities below the nominal level in all scenarios.

Jane: The conclusion is that their proposed Gumbel-based generalized pivotal quantity method performs substantially better than the only available generalized pivotal-based method of Xiang et al. (two thousand fifteen).

Lu: They applied this new approach to a real data set of ball bearing failure times and found it provides a good balance between coverage probability and interval length compared to the existing method and bootstrapping.

Meng: It sounds like the main win here is better coverage accuracy while keeping the interval lengths reasonable, even when dealing with censored data.

Lucky paper: 2604.24632: Tom: Alright, we're moving into paper number two of five, which is about theoretical guarantees for stochastic gradient sampling methods via Gaussian convolution inequalities. This paper focuses on getting first-order bounds on the bias in Wasserstein distances for the invariant measure of stochastic gradient kinetic Langevin dynamics.

Jane: It sounds like they are trying to give us a concrete limit on how accurate our estimation gets when we use these stochastic gradient MCMC methods, and they're using new Gaussian convolution inequalities to control the Wasserstein-p distance.

Lu: They establish a general p-Wasserstein inequality for centered measures, stating that if Z Rd x mu(dx) equals zero and Z Rd x 2p mu(dx) is finite, then the Wasserstein-p distance between a Gaussian convolved with a mean-zero perturbation and the Gaussian itself is bounded by some expression involving Kp.

Meng: That bound involves the second moment of the measure, which suggests that if our noise has finite second moments, we get a term proportional to h squared controlling that part of the error.

Tom: I'm looking at their quantification of asymptotic bias for stochastic-gradient UBU, and they provide an explicit O(h) bound on the Wasserstein distance between the invariant measure pi and pi h under Assumption one point five and Proposition two point one four.

Jane: That O(h) bound separates the full-gradient discretization error from what they call the additional stochastic-gradient error, which is a very helpful distinction for us when thinking about scaling up our models.

Lu: They also show that if the gradient noise has finite second moments, the term involving E of Ex∼pi Wp p (mu x g, g) is controlled by a term proportional to h squared.

Meng: The empirical validation on a one-dimensional Gaussian target shows that for SGLD, SG-HMC, and SG-UBU, the asymptotic bias is indeed first-order in the stepsize h.

Tom: They also look at adverse dimensional scaling under spike noise distribution mu s d p, numerically estimating a heuristic growth rate of O(d (one minus alpha) over two) for the bias when using a step size h equal to d minus alpha.

Jane: That suggests that as the dimension increases, the error grows in a specific way depending on how much noise we have.

Lu: The analysis extends to bounding W2(mu x g, g) when mu satisfies a Poincaré inequality, and they show that under those assumptions, the Wasserstein-two distance is bounded by C P root d.

Tom: They also derive bounds for the Kullback–Leibler divergence relating it to the Stein discrepancy S of rho g, which is interesting because it links this metric to something else.

Jane: The coupling approach they use, iteratively replacing a symmetric two-component Gaussian mixture with the Gaussian centered at its midpoint, leads to a bound of the form Wp (mu zero mu T) less than or equal to T times one minus two times minus two p t over p Phi over one p zero.

Lu: The main results establish O(h) Wasserstein bounds on the asymptotic bias of stochastic-gradient UBU under mild assumptions on the stochastic-gradient noise, and the framework covers stochastic gradients arising from subsampling procedures beyond standard minibatching.

Meng: They suggest that second-order behavior under without-replacement subsampling may be provable in the future.

Lucky paper: 2603.24538: Tom: Okay, we're looking at paper number three of five which is about the Generalized and Scalable Deep Gaussian Process Emulation framework.

Jane: It’s about how they built a scalable emulation system for simulators that have complex outputs and non-Gaussian response distributions.

Lu: The paper introduces this framework to accommodate simulators with heteroskedastic Gaussian outputs and various non-Gaussian response distributions like Poisson or negative binomial.

Tom: So, the core idea is using Deep Gaussian Processes, but they generalize it by adding a likelihood layer at the top of the hierarchy to explicitly model stochasticity.

Jane: That likelihood layer is what lets them handle those different types of outputs, which is a big deal because standard Gaussian processes usually only handle one thing.

Meng: I'm interested in how they handled inference, since that’s where things get messy when you scale up for production models.

Lu: They developed inference within the Stochastic Imputation framework, which they call SI; this involves prediction, imputation, and training components.

Tom: The prediction part uses the linked Gaussian process or LGP to give a closed form expression for the posterior predictive distribution.

Jane: And for imputation, they use an ESS-within-Gibbs sampler to impute those latent layers F and W of p l from their conditional posteriors.

Meng: What about scalability when you're dealing with a large number of input locations? That’s where the Vecchia approximation comes in.

Lu: They incorporated the Vecchia approximation to enable scalable inference by applying it to the log-likelihood function in the Stochastic Expectation-Maximization algorithm, replacing sampling from GP nodes with a closed form expression.

Tom: So they claim this reduces complexity and makes training scalable via that Expectation-Maximization algorithm.

Jane: The empirical validation shows that for heteroskedastic Gaussian cases, the GDGP framework substantially outperforms competing models like hetGP and bhetGP in terms of both NRMSE and NCRPS across different replicate settings.

Meng: That's good, but what about count likelihoods? I saw they achieved significantly lower Negative Log-Likelihood compared to models using a Poisson likelihood.

Lu: Yes, they also showed effectiveness for count likelihoods where they achieved significantly lower Negative Log-Likelihood than models using a Poisson likelihood.

Tom: So the framework is robust across different output types, not just the Gaussian ones we usually focus on.

Jane: They also showed that in the heteroskedastic Gaussian case, they could exploit additional analytical structure to get closed form conditional updates in the imputation step for further computational gains.

Meng: That sounds like a practical win for deployment because those closed forms speed things up considerably during training.

Lu: It's important to remember that they provide a careful implementation of the full GDGP framework in the open-source R package dgpsi, which is helpful for researchers who want to actually use it.

Tom: It sounds like a solid piece of work for moving from theoretical emulation to something you can actually run on real simulator models.

Lucky paper: 2610.25266: Tom: Alright, let's look at this paper from Dalhousie University, Shanglun Lia and Toby Kenneya's work on the State Space Modeling of the Ornstein-Uhlenbeck Process with Measurement Error for Microbiome Data. This is paper number four of five.

Jane: So this is about understanding the temporal dynamics of the human microbiome by applying log-ratio transformations to genus-level relative abundances.

Tom: They're using an OU state space model to analyze these time series, and they built in a structured measurement error model for those log-ratio transformed data.

Lu: The framework exploits the covariance structure from multiplicative sequencing errors, which lets them get likelihood-based estimation of latent microbial dynamics.

Meng: So they use a Kalman filter to evaluate the log-likelihood function, estimating parameters on this transformed scale: log theta, log sigma squared epsilon, log sigma squared, and mu.

Jane: They also introduced profile likelihood confidence intervals for the mean reversion rate theta to quantify that parameter uncertainty.

Tom: The diagnostics confirm both the adequacy of the model specification and that the dynamics are mean-reverting when supported by the data.

Lu: When they tested disruption effects, they compared a combined model against a split model for David et al.'s longitudinal dataset.

Jane: The results indicated that documented disruptions primarily altered equilibrium abundance levels of the microbial community while the underlying stochastic dynamics remained largely stable.

Tom: But there was this finding where a small number of genus pairs exhibited substantial reductions in mean-reversion rates following disruption.

Lu: That reduction suggests a substantial weakening of regulatory strength in microbial dynamics after the disruption, pointing to a temporary loss of ecological resilience and slower return to equilibrium.

Meng: That's interesting from an engineering side—it means the system doesn't just snap back; it gets sluggish afterward.

Jane: It means perturbations affecting metabolically important genera weaken the stabilizing feedback mechanisms of the gut microbiome, which is a key biological interpretation they drew there.

Tom: The paper also mentioned future research directions, like extensions to multivariate state space models with interaction structures and incorporating first-stage uncertainty into downstream inference.

Lu: They also noted that it would be valuable to develop methods for detecting disruption times when changepoints are unknown and to incorporate covariates or more flexible temporal processes beyond the OU model.

Lucky paper: 2610.24911: Tom: So we're looking at Martin Bladt's paper, Conformal prediction regions for dependent censored lifetimes. It deals with joint-life insurance where several lifetimes have censoring at known dates.

Jane: That setup means you're trying to construct prediction regions by reducing the lifetime vector to a scalar outcome. What is the core mechanism they use to make that reduction work?

Lu: The approach starts by representing a nested family of regions through the upper level sets of a scalar-valued reduction map s of the lifetime vector X.

Meng: So this map s has to be censoring compatible, meaning s(x c) equals s(x) c for all x and any censoring vector c.

Lalam: Theorem one establishes that this property is equivalent to nondecreasing transformations ψj such that the map s of x is the minimum of psi j applied to each coordinate.

Tom: Okay, so we're talking about these transformations ψj which define how we reduce the full vector X down to a single scalar outcome. How does this relate back to what we discussed earlier about distribution-free inference?

Jane: This paper focuses on joint-life annuities where they use the common observation period A equals A, A as the observed administrative censoring vector.

Lu: They select a region Rb of w based on observations whose censoring vector lies in an orthant determined by that reduction map s.

Meng: The retention probability is estimated by binary regression on the covariates, and they claim this construction gives finite-sample coverage with known retention probabilities without weights under censoring independent of lifetimes or covariates.

Tom: That sounds like a big win for robustness, but how do you get those asymptotic validity guarantees mentioned in Theorem four?

Jane: Theorem four establishes a double-robustness property, showing that under Assumption three point one and Assumption three point two, the marginal coverage converges to one minus alpha as N equals n goes to infinity.

Lu: The asymptotic validity of the conditional coverage statement depends on consistent estimation of either the retention probabilities or the conditional outcome quantiles, subject to those regularity conditions.

Meng: In their application to joint-life annuity data, bandwise calibration lowers the corresponding mean bounds and raises inclusion probabilities to zero point nine one zero and zero point nine two five in certain age bands.

Tom: That's good empirical evidence for the method, but what about the simulation findings? What did they find when you used setting S3 with the identity path?

Jane: In setting S3 with the identity path, both fitted and oracle procedures achieved a coverage of zero point nine one zero and zero point nine zero nine with mean bounds of three point eight five and three point eight seven years respectively.

Lu: They also looked at setting S2 using fitted weights, where they saw that the path estimation error decreases from zero point zero four two to zero point zero zero five as the training size increases from two hundred fifty to sixteen thousand observations.

Meng: That decrease in path estimation error sounds like a solid indicator that this method scales reasonably well for large datasets.

Tom: What about the comparison with Bonferroni bounds? Did they find anything concrete there regarding how their region R psi v F psi compares?

Jane: Proposition three shows that for equal standardized censoring margins, the largest threshold with retention probability at least pi naught is G bar minus one psi applied to pi zero if the coordinates of C are comonotone.

Lu: If the coordinates are independent, it uses G bar minus one psi applied to pi one over d zero instead.

Meng: The comparison shows that R psi v F psi would follow from v F psi greater than or equal to max j v B psi, but the reverse inclusion follows from v F psi less than or equal to min j v B psi.

Tom: So they are showing a bounded relationship between their method and those classical Bonferroni bounds based on the number of tests?

Jane: Exactly, and their final result shows that L b of W minus q psi of W goes to zero in probability.

Lu: They report lower bounds for the first-death times separately, with the primary bound equaling five years for twenty-eight percent of retained observations.

Meng: For the second-death analysis using s of X equals max(X one, X two), pooling over all ages gives a median unweighted empirical inclusion probability of zero point nine eight three.

Tom: That eighty-three percent inclusion probability for the second death sounds quite high compared to what we usually expect in these kinds of analyses?

Jane: Bandwise calibration produces lower bounds that are less often capped and empirical inclusion probabilities closer to zero point nine five in the oldest bands.

Lu: The analysis reports lower bounds of at most five years by design, which is a key feature here.

Tom: So for someone just listening, what does this actually change about predicting those joint lifetimes when data is missing?

More episodes

← Home