Daily Summary for 2026-10-31

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

Tom: It's the thirty-first of October, twenty twenty-six, and this is the day's research.

Jane: 11 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 Beta Regression with Autoregressive Errors for Interrupted Time Series Analysis of Proportion and Rate Outcomes.

The day's research: Tom: The focus is on how assessment timing affects trial outcomes, which matters because observed results might not reflect the true population.

Jane: This paper develops a sensitivity analysis framework handling both binary and count data, moving beyond continuous outcomes.

Lu: They create an augmented inverse intensity-weighted estimator to estimate parameters based on how assessment times relate to the process.

Meng: They propose a data-driven procedure for selecting the relevant range for this sensitivity parameter, quantifying uncertainty in that selection.

Lalam: This framework uses ARC study data suggesting home visits improve mean ACQ scores at six and twelve months if unobserved scores are lower in the intervention arm.

Tom: This builds on earlier work by extending a prior sensitivity analysis for continuous outcomes to accommodate more complex data types.

Jane: Other research explores causal inference methods, like the Mendelian randomization paper clarifying that point-identifying assumptions are weaker than previously thought.

Lu: Another piece addresses time series analysis with betark, an estimator modeling beta regression and autoregressive errors for better calibrated inference.

Meng: Global Synchronization for Multi-Source Data Integration under Blockwise Missing Patterns is significant because it tackles integrating data when blocks of information are missing.

Lalam: GSMMI formulates the alignment task as a global synchronization problem to use all pairwise overlaps simultaneously instead of sequential methods like CMMI.

Tom: The theoretical analysis shows GSMMI can achieve better alignment accuracy than CMMI, especially with many sources and rich overlap structures.

Jane: Simulations confirm GSMMI outperforms CMMI across various settings, with the advantage growing substantially as the number of sources increases from five to eighty.

Lu: This improvement is particularly noticeable when the submatrix size ratio is small and signal strength is moderate.

Meng: This method builds upon Chain-linked Multiple Matrix Integration by improving upon it, and extension to asymmetric matrices allows GSMMI to align both left and right latent positions.

Lalam: In contrast, a critique of roster charts highlights statistical errors regarding the independence of nurses and events in previous modeling assumptions.

Tom: The authors demonstrate that O’Quigley's model exhibits underdispersion rather than overdispersion when analyzing roster chart data.

Jane: A shift shows functional causal inference with the Semiparametric Inference for Functional Difference-in-Differences framework developing a cross-fitted doubly robust estimator.

Lu: This method handles discretely observed data by formulating the functional average treatment effect as a curve-valued causal estimand in Hilbert space.

Meng: The CF–DR estimator achieves superior coverage and testing power compared to other methods in simulations, showing its robustness.

Lalam: Application to NO2 profiles in London ULEZ suggests it can reveal time-of-day patterns associated with lower traffic-site pollution.

Tom: Sensitivity Analysis for Trials with Irregular and Informative Assessment Times: Marginal Mean Models and Sensitivity Parameter Selection In randomized trials, outcome assessment times often vary across participants and may depend on the underlying outcome process.

Jane: Beta Regression with Autoregressive Errors for Interrupted Time Series Analysis of Proportion and Rate Outcomes This paper introduces an estimator modeling beta-distributed conditional density and autoregressive error structure in interrupted time series analysis.

Lu: Towards more plausible point-identifying assumptions in two-sample Mendelian randomization This paper shows standard assumptions can be interpreted as estimators of the average causal effect under certain conditions.

Meng: A Unified Three-Stage Weighting Framework for Causal Inference and Mediation Analysis under Case-Control Sampling This framework estimates population causal effects and decomposes them into direct and indirect effects using case-control data.

Lalam: Leveraging Large Language Models to Improve Precision in Randomized Controlled Trials This paper explores how large language models can safely improve precision by transforming non-standard covariates.

Tom: Reconstructing constant point heat sources from few temperature sensors: sparse selection and spatial resolution This paper investigates methods for recovering unknown stationary point heat sources and their locations from limited sensor data.

Jane: Simulation-based Criticism of Complex Biological Models with Neural Evidence Estimation This paper proposes Neural Evidence Estimation to critique complex simulation models by estimating their marginal likelihood.

Lu: Anchored multiple testing: a transparent use of e-closure to improve FDR procedures This paper introduces an anchoring technique to improve the transparency and performance of e-closure methods for controlling the false discovery rate in multiple testing.

Meng: Global synchronization for multi-source data integration under blockwise missing patterns This paper proposes GSMMI to jointly align multiple sources of data that have blockwise missing patterns.

Lalam: Critique of "Use of roster charts in the investigation and prosecution of nurses" by John O' Quigley This paper critiques statistical errors in a study using roster charts to investigate nurse misconduct, concluding the chart is statistically fake evidence.

Tom: Semiparametric Inference for Functional Difference-in-Differences This paper develops a semiparametric framework for functional difference-in-differences to estimate causal effects on functional outcomes with discretely observed data.

Lucky paper: 2607.07914: Tom: So we're diving into Beta Regression with Autoregressive Errors for Interrupted Time Series Analysis of Proportion and Rate Outcomes, which looks at how assessment timing affects trial results.

Jane: It’s really about moving past ordinary least squares regression, which isn't great for bounded outcomes like proportions and rates.

Lu: The paper introduces betark, a joint conditional maximum likelihood estimator that models the mean, precision, and AR(k) coefficients in a single likelihood function.

Meng: So it’s modeling the conditional mean mu t and a precision parameter phi t based on the information set Ft minus one.

Lalam: It uses a recursive substitution approach to define disturbance xi t as g(mu t) minus x prime beta, and letting that follow an autoregressive process of order k.

Tom: That joint modeling of mean, precision, and AR k coefficients is what makes it different from just using standard errors.

Jane: They compare betark against the quasi-binomial generalized linear model with logit link paired with Newey–West HAC standard errors.

Lu: The simulation study replicated one hundred eight design cells across two thousand runs to evaluate bias, standard error calibration, and Type I error across three series lengths, nine autocorrelation scenarios, and four effect sizes.

Meng: What’s the most striking finding there? Is it that betark consistently beats the GLM plus HAC comparator?

Tom: Yes, the primary finding is that betark provided better-calibrated inference than the GLM plus HAC comparator in most scenarios.

Jane: Especially under highly persistent autocorrelation, where they saw Type I error exceeding sixty percent at short series lengths.

Lu: Specifically, betark achieved higher ninety-five percent confidence interval coverage than the GLM plus HAC comparator in seven of nine autocorrelation scenarios.

Meng: Better calibration is key for deployment when you don't know the true underlying process structure beforehand.

Tom: It shows that betark’s standard error ratio was closer to one point zero than the GLM plus HAC comparator’s in seven of nine autocorrelation scenarios, which is a good sign.

Jane: The implication for applied analysis is that we need a calibrated, condition-dependent recommendation instead of just one uniform method.

Lu: Under high persistent autocorrelation—which they note might be common in cumulative or slowly evolving health research outcomes—betark is clearly preferable across all series lengths examined.

Meng: But the paper does have limitations; they only examined series lengths up to T equals four hundred and AR(three) high-persistent autocorrelation hadn't fully stabilized there for either estimator.

Tom: They also mentioned that misspecification of the autoregressive order by a single lag had a smaller inferential cost than the choice of series length or the actual persistence of the autocorrelation.

Jane: So even if you get it wrong on the AR k part, as long as you pick a reasonable length, it’s okay for most cases.

Lu: The paper confirms that these differences come from variance estimation rather than systematic error in point estimation itself, which is important context.

Meng: For us deploying this to production, better calibration means we can trust our uncertainty estimates more when running models on real traffic data where autocorrelation is expected.

Lalam: I'm confident in the finding that betark provides better-calibrated inference across the large majority of autocorrelation structures and effect sizes examined.

Lucky paper: 2609.22563: Tom: Okay, we're looking at Simulation-based Criticism of Complex Biological Models with Neural Evidence Estimation. This paper is about using neural techniques to learn a model's marginal likelihood directly from simulations, which avoids those noisy Monte Carlo estimators.

Jane: So, the core idea is that if you can estimate that marginal likelihood p(x) using a normalizing flow, you can reliably pick the correct data generating process among several models.

Lu: The methodology involves fitting a density estimator h(x; psi) on simulated data S x to approximate the true probability distribution, and then solving for the flow parameters psi* using stochastic gradient descent.

Meng: Once trained, this neural evidence estimation can produce a marginal likelihood estimate for any new dataset with just one forward pass through the network.

Tom: The paper uses this technique for model selection tests, comparing the estimated marginal likelihood against a sampling distribution of likelihoods from many simulated datasets to compute a p-value for model misspecification.

Jane: They specifically tested this on five stochastic compartmental models—SIR, CH, SI, SIRS, and SEIR—and found that it correctly flagged constant hazards and SI models as misspecified at least forty percent of the time.

Lu: For model selection, they calculated the Bayes’ factor B twelve which is the ratio of two marginal likelihoods to determine a model’s level of evidence according to Jeffreys thresholds.

Tom: The case study results on influenza outbreaks showed decisive evidence in favor of the correct model over several false models between forty percent and fifty percent of the time.

Jane: They also found that a neural posterior estimator on simulation-parameter pairs showed that the SEIR model exhibited by far the highest mean squared posterior predictive error, or MSPPE, compared to other models.

Lu: The limitations they point out are scaling up to high-dimensional datasets because of the curse of dimensionality and discrete data being problematic since normalizing flows rely on continuous transformations.

Meng: The reliability of Neural Evidence Estimation depends on how accurately it estimates that marginal likelihood, as it assumes the true marginal likelihood density for a model is reasonably regular.

Tom: They suggest future work could use flow architectures specialized for discrete data or incorporate a decreasing noise schedule during training to help with those issues.

Jane: The paper argues that epidemiologists should rigorously compare findings from multiple plausible models, discard the implausible ones, and communicate uncertainties inherent in modeling assumptions to the general public.

Lu: Ultimately, they conclude that emerging methodologies in simulation-based inference and statistical model criticism can make multi-model analysis an attainable standard of practice in infectious disease epidemiology.

Lucky paper: 2610.11279: Tom: Alright, we’re moving on to our third paper of five this day: Anchored multiple testing: a transparent use of e-closure to improve FDR procedures.

Jane: This paper looks at how we can get more transparency in methods that control the false discovery rate when using e-closure.

Lu: They introduce an elementary new technique called anchoring which exploits looseness in existing FDR proofs to enlarge a baseline multiple-testing procedure’s self-referential local e-value.

Meng: So, this anchoring principle is generic, it applies to any baseline rejection set where a deterministic function satisfies E

FDP(A)/alpha: less than or equal to qS whenever all hypotheses in S are true.

Tom: That sounds like they are building a new way to quantify the local e-value by using that function and the term one minus qS times FDP(S).

Jane: The main contributions focus on how this anchored method dominates other procedures, specifically showing Anchored-BH dominates one BH under PRDS.

Lu: They also show Anchored-BY dominates BY under arbitrary dependence, which is important because it's incomparable to canonical closed BY.

Meng: And they demonstrated that Anchored e-BH dominates e-BH and is incomparable to closed e-BH as well.

Tom: So they’re claiming this new anchoring technique gives them better performance in several key comparisons against the standard closed methods, like one BH under PRDS.

Jane: They also showed that Anchored-dBH improves the conditional-calibration procedure of Fithian and Lei's two thousand twenty-two work.

Lu: The paper concludes that while these new procedures aren't necessarily admissible, they can safely replace earlier methods whenever those were employed because they offer simplicity and generality.

Meng: Computationally, the complexity for anchored methods is generally in sorting or quadratic time which is manageable for production deployment.

Tom: I see the computational complexity is in sorting or quadratic time, which sounds reasonable for a multiple testing procedure that needs to run on data.

Jane: The core idea here seems to be making sure that every measurable certified superset of A controls the false discovery rate at alpha and contains all of A’s discoveries.

Lu: They define ES as FDP(A)/alpha plus one minus qS times FS, which they call a local e-value, and every certified superset B-bar obtained by e-closure satisfies B-bar contains BHbeta(p) pointwise.

Tom: So they are essentially replacing a self-referential local value with something that transparently retains every baseline discovery and usually adds more.

Jane: It’s about making the e-closure process more transparent, which is something that has been opaque in some recent improvements.

Lu: They also explored how the anchoring principle applies directly to Wang and Ramdas two thousand twenty-two's e-BH procedure as well.

Lucky paper: 2606.29394: Tom: Alright, we're moving into the discussion on the paper titled Critique of “Use of roster charts in the investigation and prosecution of nurses” by Richard D Gill and John O’Quigley.

Jane: This paper is really digging into why O’Quigley's analysis based on those roster charts is statistically flawed, focusing on the assumption about nurse independence.

Lu: They introduce binary variables Yi and Xi to address the central question of presence given suspicion, specifically looking at the quantity Pr(Yi = one Xi = one).

Meng: I’m curious how they are handling the visual presentation aspect, because they mentioned Figure one is described as fake by a meeting of the Legal Section of the Royal Statistical Society.

Lalam: They highlight that intuition can lead to errors, pointing out that a nurse working the most shifts might be more likely to be present than one with fewer shifts.

Tom: So, if we look at the model assumptions, they focus heavily on O’Quigley’s assumption that nurses and shifts are completely independent.

Jane: They show that the model O’Quigley used has been proven seriously wrong because their analysis of the fitted beta-binomial distribution shows underdispersion instead of overdispersion.

Lu: The critique focuses a lot on the fact that the resulting probability distribution of thirty-seven random nurses’ random column totals in a twenty-five by thirty-seven roster chart would exhibit underdispersion.

Tom: This leads to their conclusion that the empirical estimate of presence for the suspected nurse at one hundred percent has no value because there is selection bias involved.

Jane: They assert that predictions made from O’Quigley’s model fit are useless because the fitted model completely fails to fit to the actual data, and they agree with him that the roster chart is statistically fake evidence.

Lu: The paper details how investigators nominated sixty-one shifts as containing potentially suspicious events, with Lucy Letby being seen as the common factor in those twenty-five events.

Meng: They hypothesize that O’Quigley sampled the thirty-six missing rows of the table by using the thirty-seven other nurses’s observed relative frequency of presence in columns from the twenty-five observed rows.

Tom: That means they are using maximum likelihood estimation of a beta-binomial using an augmented table of sixty-one rows to explain how he filled in those missing data points.

Jane: It sounds like they are showing that the construction method itself introduces bias, and that’s why the visual impression ignores things like selection bias.

Lu: The authors state that O’Quigley has sampled the thirty-eight missing rows of the table by using the thirty-seven other nurses’s observed relative frequency of presence in columns from those twenty-five observed rows.

Meng: That means they are trying to model how presence probability changes based on what the other nurses were doing.

Tom: It confirms that even when you try to augment a dataset like that, if the fundamental assumption about independence is wrong, the result is meaningless.

Lucky paper: 2605.26964: Tom: Okay, we're looking at Semiparametric Inference for Functional Difference-in-Differences now. Jane, what’s the core challenge they are tackling here that makes standard difference-in-differences not enough?

Jane: Well, standard DiD is usually for scalar outcomes, but this paper tackles functional data with discretely observed data. The core framework is about formulating and identifying the functional average treatment effect on the treated, or fATT, as a curve-valued causal estimand.

Lu: They derive the efficient influence function for this fATT curve in Hilbert space formulation. That means they're defining what a perfect estimate looks like mathematically before they even try to find it.

Meng: And they show that this influence function is the canonical gradient, and it has this property where the map from nuisance functions to the score is Neyman-orthogonal at η0. That underpins their robustness to errors in estimating those nuisance functions.

Tom: It sounds mathematically intensive, but how does that translate into something practical for someone trying to analyze treatment effects on a continuous outcome?

Jane: They propose a cross-fitted doubly robust estimator that uses an uncentered orthogonal score. They implement this using K-fold cross-fitting to reduce overfitting bias in the nuisance estimation step.

Lu: Theoretically, they establish asymptotic linearity in Hilbert space and weak convergence to a Gaussian element in H = L two(T). That convergence is what lets them get pointwise confidence intervals and multiplier-bootstrap simultaneous confidence bands.

Meng: In terms of deployment, the K-fold cross-fitting is computationally heavy. We have to consider the cost of running this on large functional datasets, especially when we're dealing with many sources.

Tom: So, if you look at their simulation results, what’s the actual performance gain they see compared to other methods?

Jane: The simulation results confirm that the CF–DR estimator achieves superior coverage and testing power in finite samples when compared to other methods. They also show that increasing sample size improves sampling variability.

Lu: But there's a caveat there, because reconstruction-induced smoothing can still be visible when the observation process is sparse and noisy. That means accuracy in curve reconstruction is critical for first-order causal inference to work right.

Meng: So if the data is very sparse, even this powerful CF–DR estimator might not give us perfectly sharp results. We need to know how much smoothing we're actually introducing here.

Tom: They apply this specifically to London ULEZ and hourly NO2 profiles, demonstrating that it can reveal a time-of-day pattern associated with lower traffic-site NO2 pollution. That’s a concrete application point.

Jane: That’s the practical payoff—it helps reveal environmental patterns, like those time-of-day effects, using functional causal inference tools on this paper. The overall conclusion is that CF–DR provides more balanced uncertainty quantification than main feasible alternatives for functional DiD.

Lu: The authors suggest extensions to staggered adoption and spatial dependence in future work because they acknowledge those are important complexities they haven't fully addressed yet in this version.

Meng: For us, the practical advantage is that it separates the causal estimation difficulty from the reconstruction-driven functional error, which makes it more tractable than trying to solve both at once.

Tom: So when we look at the final results, what’s their stance on treating an outcome as a curve instead of just a scalar average for policy analysis?

Jane: They support that treating the outcome as a curve is necessary for policy analysis because it captures more of the underlying causal effect. The paper suggests this framework has substantial practical advantages for functional DiD.