Daily Summary for 2026-11-03

daily

Transcript

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

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

Jane: 21 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 Evaluating causal indirect effects when mediators are left-censored by assay limit of quantification.

The day's research: Tom: We need to measure causal effects when biological markers are only partially observed due to assay limits.

Jane: That is critical because understanding treatments requires knowing what happens inside the body, especially with missing data that isn't random.

Lu: The most important work develops a semi-parametric framework for direct and indirect effects when the mediator is left-censored by an assay limit of quantification.

Meng: This research uses fractional imputation combined with a semi-parametric EM algorithm and shows this strategy significantly reduces bias.

Lalam: They confirmed monoclonal antibody therapies reduce hospitalization and death, but viral RNA changes only mediate a modest proportion of the overall effect.

Tom: Following that is work on learning sparse representations for bias-free inference because standard rotation methods can introduce non-estimable bias.

Jane: This paper proposes a novel method establishing an oracle inference property, meaning estimators achieve the same asymptotic variance as if all latent variables were perfectly observed.

Lu: Finally, there is work on causal discovery in time series extending conditional independence tests to include contemporaneous links.

Meng: This method shows improved reliability and higher recall for detecting strong autocorrelation compared to older algorithms like the PC algorithm.

Lalam: The most important finding relates to estimating treatment effects when we have missing outcome data in prenatal studies because identifying pregnancies based only on delivery outcomes misses many cases with unrecorded results.

Tom: The work by Bias in studies of prenatal exposures using real-world data due to pregnancy identification method simulated 10,000,000 pregnancies and showed risk differences were similarly biased across different missing outcome scenarios when all missingness was due to unobserved miscarriage.

Jane: This finding is more significant than Beyond Pretrends because it highlights a specific type of selection bias that can plague causal inference in medical data.

Lu: It shows that even when trying to adjust for missing data, the estimates remain biased unless we measure the predictors of missingness.

Meng: This is less critical than MIMO and Multi-Group Comparison Problem in Process Control because it introduces a statistical framework using Pseudo-Impulse Responses to diagnose faults in complex MIMO systems.

Lalam: The work by ProjGuard shows how continuous monitoring combined with selective intervention can drastically reduce unsafe rates in computer agents, moving from sixteen percent down to three percent on OS-Harm.

Tom: This is less impactful than A First Course in Monte Carlo Methods which serves as a mathematical introduction to Monte Carlo methods, providing the foundational tools for many of these complex statistical analyses.

Jane: Today's papers include Evaluating causal indirect effects when mediators are left-censored by assay limit of quantification developing a method to estimate causal mediation effects even when the biological mediator is limited by technical measurement errors.

Lu: And Beyond Vintage Rotation Bias-Free Sparse Representation Learning with Oracle Inference proposes a new rotation method that guarantees valid inference for learned sparse representations without introducing estimation bias.

Meng: The Geometry of Latent Spaces Projective Invariants Structural Identification and Ratio-Scale Measurement introduces a coordinate-free framework to measure ordinal data that resolves rotational ambiguities in latent variable models.

Lalam: Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets extends causal discovery to include contemporaneous links in time series data while improving reliability.

Tom: Partial identification via conditional linear programs provides a framework for estimating parameters in partially identified problems using covariate-assisted conditional linear programs.

Jane: Implementation and Workflows for INLA-Based Approximate Bayesian Structural Equation Modelling presents an R package that allows for fast approximate Bayesian structural equation modeling using the INLA framework.

Lu: Structural identifiability of partially-observed stochastic processes investigates how parameter identifiability depends on whether data comes from single-particle trajectories or total particle density measurements.

Meng: Simulation-Based Inference with Convolutional Maximum Mean Discrepancy for Noisy Data introduces a method to perform inference on noisy data using convolutional maximum mean discrepancy to compare distributions.

Lalam: A First Course in Monte Carlo Methods is a concise mathematical introduction to the broad family of Monte Carlo methods used across science and engineering.

Tom: Beyond Pretrends A Discordance-Based Sensitivity Analysis for Difference-in-Differences proposes a sensitivity analysis that uses discordance between different imputation methods to assess the robustness of difference-in-differences results.

Jane: MIMO and Multi-Group Comparison Problem in Process Control develops a multivariate statistical framework to diagnose faults in complex multi-input multi-output systems using frequency response functions.

Lu: Scaling Causal Mediation for Complex Systems A Framework for Root Cause Analysis proposes a scalable mediation analysis framework designed to handle complex causal pathways in large directed acyclic graphs.

Meng: Tackling the 6/49 Lottery and Debunking Common Myths with Probabilistic Methods and Combinatorial Designs compares scientific approaches to solving the 6/49 lottery using probabilistic methods and combinatorial designs.

Lalam: On the Spherical Dirichlet Distribution Corrections and Results revisits a spherical-Dirichlet distribution to clarify its density representation in terms of surface-area measure.

Tom: Bayesian Reconciliation of Overlapping Multi Step Forecasts A Post Processing Framework for Cross Origin Coherence uses a Bayesian state-space model to reconcile overlapping forecasts that disagree due to different forecast origins.

Jane: A scalable Bayesian functional factor model for high-dimensional longitudinal molecular data proposes a Bayesian functional factor model to uncover coordinated biomarker trajectories in large, high-dimensional longitudinal molecular datasets.

Lu: Improving causal inference in interrupted time series analysis the triple difference design formalizes and applies the triple difference interrupted time series design to strengthen causal inference in health policy evaluations.

Meng: Multi-state Models For Disease Histories Based On Longitudinal Data investigates extending piecewise exponential additive models to handle complex transitions between disease stages using longitudinal data.

Lalam: One-sample survival tests in the presence of non-proportional hazards in oncology clinical trials proposes new score tests for single-arm oncology trials that account for non-proportional hazards and crossing hazards.

Tom: Partial identification via conditional linear programs estimation and policy learning provides a framework for estimating parameters in partially identified problems using covariate-assisted conditional linear programs.

Jane: The Geometry of Latent Spaces Projective Invariants Structural Identification and Ratio-Scale Measurement introduces a coordinate-free framework to measure ordinal data that resolves rotational ambiguities in latent variable models.

Lu: Implementation and Workflows for INLA-Based Approximate Bayesian Structural Equation Modelling presents an R package that allows for fast approximate Bayesian structural equation modeling using the INLA framework.

Meng: Simulation-Based Inference with Convolutional Maximum Mean Discrepancy for Noisy Data introduces a method to perform inference on noisy data using convolutional maximum mean discrepancy to compare distributions.

Lalam: Bias in studies of prenatal exposures using real-world data due to pregnancy identification method examines how different ways of identifying pregnancies can introduce bias when estimating treatment effects in real-world data.

Tom: ProjGuard Safety Monitoring for Computer-Use Agents via Low-Dimensional Projections proposes a lightweight monitor that uses linear projection to detect adversarial behavior in computer agents and trigger safety interventions.

Jane: Beyond Pretrends A Discordance-Based Sensitivity Analysis for Difference-in-Differences proposes a sensitivity analysis that uses discordance between different imputation methods to assess the robustness of difference-in-differences results.

Lu: MIMO and Multi-Group Comparison Problem in Process Control develops a multivariate statistical framework to diagnose faults in complex multi-input multi-output systems using frequency response functions.

Meng: Scaling Causal Mediation for Complex Systems A Framework for Root Cause Analysis proposes a scalable mediation analysis framework designed to handle complex causal pathways in large directed acyclic graphs.

Lalam: Tackling the 6/49 Lottery and Debunking Common Myths with Probabilistic Methods and Combinatorial Designs compares scientific approaches to solving the 6/49 lottery using probabilistic methods and combinatorial designs.

Tom: On the Spherical Dirichlet Distribution Corrections and Results revisits a spherical-Dirichlet distribution to clarify its density representation in terms of surface-area measure.

Jane: Bayesian Reconciliation of Overlapping Multi Step Forecasts A Post Processing Framework for Cross Origin Coherence uses a Bayesian state-space model to reconcile overlapping forecasts that disagree due to different forecast origins.

Lu: A scalable Bayesian functional factor model for high-dimensional longitudinal molecular data proposes a Bayesian functional factor model to uncover coordinated biomarker trajectories in large, high-dimensional longitudinal molecular datasets.

Meng: Improving causal inference in interrupted time series analysis the triple difference design formalizes and applies the triple difference interrupted time series design to strengthen causal inference in health policy evaluations.

Lalam: Multi-state Models For Disease Histories Based On Longitudinal Data investigates extending piecewise exponential additive models to handle complex transitions between disease stages using longitudinal data.

Tom: One-sample survival tests in the presence of non-proportional hazards in oncology clinical trials proposes new score tests for single-arm oncology trials that account for non-proportional hazards and crossing hazards.

Jane: This paper develops a method to estimate causal mediation effects even when the biological mediator is limited by technical measurement errors.

Lu: This work proposes a new rotation method that guarantees valid inference for learned sparse representations without introducing estimation bias.

Meng: This paper introduces a coordinate-free framework to measure ordinal data that resolves rotational ambiguities in latent variable models.

Lalam: This method extends causal discovery to include contemporaneous links in time series data while improving reliability.

Tom: This paper provides a framework for estimating parameters in partially identified problems using covariate-assisted conditional linear programs.

Jane: This paper presents an R package that allows for fast approximate Bayesian structural equation modeling using the INLA framework.

Lu: This work investigates how parameter identifiability depends on whether data comes from single-particle trajectories or total particle density measurements.

Meng: This paper introduces a method to perform inference on noisy data using convolutional maximum mean discrepancy to compare distributions.

Lalam: This study examines how different ways of identifying pregnancies can introduce bias when estimating treatment effects in real-world data.

Tom: This paper proposes a lightweight monitor that uses linear projection to detect adversarial behavior in computer agents and trigger safety interventions.

Jane: This paper proposes a sensitivity analysis that uses discordance between different imputation methods to assess the robustness of difference-in-differences results.

Lu: This study develops a multivariate statistical framework to diagnose faults in complex multi-input multi-output systems using frequency response functions.

Meng: A scalable mediation analysis framework designed to handle complex causal pathways in large directed acyclic graphs proposes a scalable mediation analysis framework for root cause analysis.

Lalam: This article compares scientific approaches to solving the 6/49 lottery using probabilistic methods and combinatorial designs.

Tom: This paper revisits a spherical-Dirichlet distribution to clarify its density representation in terms of surface-area measure.

Jane: A Post Processing Framework for Cross Origin Coherence uses a Bayesian state-space model to reconcile overlapping forecasts that disagree due to different forecast origins.

Lu: A Bayesian functional factor model for high-dimensional longitudinal molecular data proposes a Bayesian functional factor model to uncover coordinated biomarker trajectories in large, high-dimensional longitudinal molecular datasets.

Meng: This paper formalizes and applies the triple difference interrupted time series design to strengthen causal inference in health policy evaluations.

Lalam: This work investigates extending piecewise exponential additive models to handle complex transitions between disease stages using longitudinal data.

Tom: New score tests for single-arm oncology trials account for non-proportional hazards and crossing hazards proposes new score tests for single-arm oncology trials that account for non-proportional hazards and crossing hazards.

Lucky paper: 2605.20615: Tom: We're looking at Evaluating causal indirect effects when mediators are left-censored by assay limit of quantification today, which is a tough problem when biological markers are only partially observed because of assay limits.

Jane: It’s about estimating natural direct and indirect effects when the mediator is subject to left-censoring due to the Limit of Quantification, or LLoQ.

Lu: The authors use fractional imputation combined with an Expectation-Maximization algorithm and a semi-parametric location-scale conditional density estimator paired with a highly adaptive lasso, or HAL.

Meng: They established identification in a parametric setting by imposing specific structures on the conditional mediator and outcome models when LLoQ censoring is present.

Lalam: They also used an m out of n bootstrap to address nonregularity introduced by the imputation steps to ensure robust inference under the procedure.

Tom: The simulation studies show that while early reduction in anterior nasal SARS-CoV-two RNA is observed, it explains only a modest fraction of the overall protective effect of monoclonal antibody therapy on hospitalization or death.

Jane: They found that flexible imputation strategies yield modestly larger, more negative indirect effect estimates compared to naive LLoQ two imputation methods.

Lu: That suggests that the majority of the treatment's well-characterized benefit operates through pathways not captured by viral RNA levels alone.

Meng: The paper uses a separable effects framework based on Robins Richardson and Shpitser, decomposing the mAb treatment into two hypothetical separable interventions.

Tom: This decomposition allows for quantifying both the natural direct effect and the natural indirect effect, isolating what is attributable to viral clearance versus other biological mechanisms.

Jane: They rely on a population single-world intervention graph derived from an expanded directed acyclic graph to encode those separability assumptions.

Lu: In terms of estimation performance, they found that the EM one true nuisances method and the EM four heteroscedastic CDE HAL strategy perform exceptionally well.

Meng: They demonstrated minimal bias, low mean squared error, and near-nominal confidence interval coverage when those methods were paired with the proposed bootstrap procedure.

Lalam: In contrast, naive LLoQ two substitution consistently yields estimates close to zero for the indirect effect across most estimators.

Tom: So what does this mean for us in practice? It confirms that AN viral RNA is unlikely to be a reliable surrogate endpoint for assessing monoclonal antibody efficacy in this context.

Jane: It suggests the majority of therapeutic benefit stems from pathways independent of viral load levels.

Lucky paper: 2003.03685: Tom: So we're looking at Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets.

Jane: This paper introduces PCMCI+, a conditional independence based method that aims to discover both contemporaneous and lagged causal links in time series where autocorrelation is present.

Lu: What's interesting is how it extends the PCMCI method to include the discovery of contemporaneous links, which requires some substantial changes from the original PC algorithm and time-series adaptations of FCI.

Tom: I see they do an edge removal phase separately for lagged and contemporaneous conditioning sets, and that lagged phase uses much fewer conditional independence tests.

Jane: And they optimize the choice of those conditioning sets for individual conditional independence tests to make them better calibrated under autocorrelation by utilizing the momentary conditional independence idea.

Meng: From a deployment standpoint, how does this optimization translate into actual speed improvements when we are dealing with long time series?

Lu: Numerically, they show that PCMCI+ benefits from strong autocorrelation and yields much more adjacency detection power and especially more orientation recall for contemporaneous links compared to other methods.

Tom: They also claim better control over false positives at much shorter runtimes than the original PC algorithm.

Jane: That's a big deal because when we have strong autocorrelation, traditional CI-based methods often suffer from low recall or inflated false positives.

Meng: If they are faster, that’s good for scaling, but what about the theoretical guarantees they offer?

Lu: The theoretical results establish asymptotic consistency and finite sample order-independence. Lemma one states that Algorithm one always returns a set that always contains the parents of Xj t and at most, the lagged parents of all contemporaneous ancestors of Xj t.

Tom: That sounds like a solid structural guarantee for what they are actually finding in these time series datasets.

Jane: Theorem one proves the soundness of Algorithm two which states that it returns the correct adjacencies under Assumptions one.

Meng: So for someone shipping models to production, what does that mean practically when we look at our real traffic data?

Lu: The performance analysis shows that PCMCI+ has higher adjacency detection power and especially more contemporaneous orientation recall compared to other methods while better controlling false positives.

Tom: And they mention for linear Gaussian setups, PCMCI+ and GCresPC show that the true positive rate for contemporaneous links is stable even under high autocorrelation while PC and LiNGAM show strong declines.

Jane: That comparison against PC and LiNGAM shows that PCMCI+ maintains performance where others degrade significantly when autocorrelation gets high.

Meng: If we are working with complex, nonlinear time series, does this method still hold up as well as it does for linear Gaussian setups?

Lu: The paper notes that the algorithm can be flexibly combined with different CI tests for nonlinear causal discovery and for different variable types, like discrete or continuous data.

Tom: They also point out that autocorrelation is actually key to increase contemporaneous orientation recall because it creates those triples Xi t minus one to Xi t sub zero Xj t that can often be oriented while an isolated link stays undirected in the Markov equivalence class.

Jane: That idea about using autocorrelation to orient links is a neat mechanism for improving causal discovery here.

Meng: The algorithm is available as part of the tigramite Python package at https://github.com/jakobrunge/tigramite, so it seems like there's some community backing for this tool.

Lu: That package availability means researchers can use PCMCI+ to explore these complex causal pathways in their own work with high-dimensional data.

Lucky paper: 2506.12215: Tom: Okay, we're looking at Partial identification via conditional linear programs estimation and policy learning.

Jane: So, the core idea here is a unified framework for covariate-assisted estimation and decision making in partial identification problems where the parameter of interest meets a set of linear constraints.

Lu: The paper introduces bounds on the parameter as expectations of solutions to conditional linear programs that optimize a linear function subject to linear constraints, where both functions and the constraints might depend on covariates that we have to estimate from data.

Meng: They propose two de-biased estimators for these bounds, one that directly solves the conditional linear programs with plugin estimates and another using entropic regularization to create smooth approximations.

Lalam: The simulation study demonstrates that the debiased entropic estimator is not particularly sensitive to the scaling of the hyper-parameter eta; both estimators behave roughly the same way.

Tom: That robustness to scaling sounds good, but what about when we have estimation errors in those covariate conditional constraints and objective vectors?

Jane: The authors established conditions for asymptotic normality of both estimators, and these de-biased estimators are robust to first-order errors in estimating the covariate conditional constraints and objective vectors.

Lu: They also extend these results to decision-making problems where the goal is to estimate an optimal decision policy by maximizing an estimate of the lower or upper bound on a partially identified quantity.

Meng: In policy learning, they show that relative to the optimal decision policy for the true entropic lower bound, pi star eta, the excess regret of their estimated policy is impacted by estimation errors in the constraint vector b hat x only via the de-biased rate rn and squared errors of b hat minus b squared.

Lalam: They also show how to estimate conditional bounds on a partially identified parameter theta v by regressing pseudo-outcomes on the covariates V, which inherits properties from those de-biased estimators.

Tom: In instrumental variable settings, the de-biased rate will be a product of the error in the estimates of the treatment and instrument assignment probabilities.

Jane: That de-biasing function is constructed from the Riesz representor phi c k O ck equal to P d prime one Z equals z times P Z equals z x P Y equals y j D equals d j minus b hat x, and so the de-biased rate will be a product of the error in the estimates of the treatment and instrument assignment probabilities.

Lu: The overall conclusion is that both estimators are robust to errors in estimating those nuisance functions, allowing for valid inference even when those nuisance functions are estimated using flexible machine learning methods.

Lucky paper: 2504.12415: Tom: Okay, let's look at this paper: Bias in studies of prenatal exposures using real-world data due to pregnancy identification method. Jane, what is the core methodological challenge they are tackling here?

Jane: They're looking at how missing outcome data, like whether a patient had a delivery or not, affects treatment effect estimates when the reason for that missing data isn't random.

Lu: It boils down to identifying pregnancies through observed outcomes, which risks missing cases that received prenatal care but whose final result was never documented.

Meng: They simulated ten million pregnancies and tested three analytic samples: those based on observed deliveries, those including both deliveries and miscarriages, and one with all pregnancies.

Lalam: The key finding here is that when the missingness was due solely to unobserved miscarriage, the risk differences and risk ratios were similarly biased across all three identification strategies.

Tom: That sounds like a tough spot because it suggests the way we choose to identify pregnancies doesn't matter if the data is missing because of an unobserved event like a miscarriage.

Jane: Exactly. But then they found something interesting when the missingness was due to measured covariates, which is what we often see in real-world data.

Lu: When predictors of missingness were measured, specifically when it was Missing At Random, or MAR, only the all pregnancies approach yielded unbiased estimates across all those scenarios.

Meng: The numbers show that when the reason for missingness was a measured variable like severity, including pregnancies with unobserved outcomes only helped reduce bias as those measured factors increased.

Lalam: But if the missingness was just due to unobserved miscarriage, including all pregnancies didn't fix the bias; the estimates stayed similarly biased regardless of whether you looked at deliveries or all pregnancies.

Tom: So what this means for someone listening who isn't in epidemiology, it means that if you are studying prenatal exposure and your data has missing outcomes, you have to know *why* it's missing before you can trust the results.

Jane: That’s right. If the missingness is due to something unobserved, like a miscarriage, your identification strategy doesn't solve the problem on its own.

Lu: The paper really underscores that distinguishing between unobserved outcomes and measured covariates is crucial for obtaining reliable estimates of total effects of prenatal exposures.

Meng: It’s not just about getting the right data; it’s about understanding the specific source of missing data to get unbiased results from this study on Bias in studies of prenatal exposures using real-world data due to pregnancy identification method.

Lalam: And importantly, when treatment affected neither miscarriage nor preeclampsia, all analytic samples returned unbiased risk differences and risk ratios.

Tom: That's a solid point because it tells us that sometimes the identification strategy doesn't matter if the treatment itself didn't have an effect on the outcome you were looking at.

Lucky paper: 2605.13631: Tom: Alright, we are moving onto ProjGuard, which is this paper about safety monitoring for computer-use agents using low-dimensional projections.

Jane: So, the core of it is that these agents can operate on real operating systems now, and this capability has brought up risks like prompt injection and indirect instructions.

Lu: ProjGuard proposes a trajectory-based safety mechanism that monitors behavior rather than just inputs, using a hierarchical architecture where a lightweight monitor estimates risk from the agent’s partial interaction trajectory.

Meng: It looks at the sequence of observations, responses, and actions denoted by tau equal o naught r one a one o one... up to r T and a T.

Tom: That accumulated trajectory up to step t is defined as chi t equals m t i=one tau i:t where L denotes concatenation.

Jane: And that accumulated trajectory is then mapped into a vector representation vt equal Vec(chi t) in R d using standard text and sequence representations.

Lu: They constructed a dataset D equal chi d T, s(chi d T) to train the model, which can be partitioned into safe and unsafe subsets called Dsafe and Dunsafe.

Meng: The linear projection W is learned from D by optimizing the objective W star equals arg min W X D d=one L W top Vec(chi d T), s(chi d T).

Tom: And the gains come from utilizing Linear Discriminant Analysis to construct a one-dimensional space where unsafe behaviors become separable, maximizing between-class separation while minimizing within-class variance.

Jane: They then compute the risk score rho t based on its relative distance to the class centroids using rho t equals sigma alpha (z t minus mu safe minus z t minus mu unsafe).

Lu: The full method achieves a similar completion rate of sixty-five percent, within one point of the correction-only method, while drastically reducing the computational cost to only six API calls, three hundred tokens, and seventy-five s.

Meng: That is a big reduction compared to StepJudge-Safety which needs seven hundred eight API calls, thirty-five thousand tokens, and ninety s.

Tom: The temporal evolution analysis shows that corrected trajectories drive the risk back toward lower-risk regions after an alert is triggered.

Jane: Figure three illustrates the temporal evolution of the estimated risk rho t for the GPT-five point two plus LDA plus GPT-five point four configuration, which achieves three percent unsafe trajectories and sixty-five percent completion.

Lu: Safe trajectories remain consistently below the decision threshold gamma equal zero point six, indicating stable and aligned behavior throughout execution.

Tom: Corrected trajectories initially follow a similar trend to unsafe trajectories and may exceed the threshold, triggering an alert.

Jane: The main result is that safe and unsafe trajectories occupy distinguishable regions in a lightweight one-dimensional space, and that this structure can be exploited for low-cost continuous monitoring.

Lu: The hashing two thousand forty eight plus Linear Discriminant Analysis configuration achieved a silhouette score of zero point nine five three and a Davies–Bouldin index of zero point five two four, confirming that the observed separation is sufficiently clear to serve as the basis for the monitor.

More episodes

← Home