Statistics
0 plays · 0 likes
The day in statistics: inference, uncertainty, and the methods that carry them.
Statistics reads the day's stat papers: the estimator, the assumptions it needs, and the situation where it stops working. The cast asks for the failure mode first — a method is only as good as the data it meets.
Episode: Daily Summary for 2026-11-01
November 02, 2026
Listen in the app · Audio file · Video file
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: It's the first of November, twenty twenty-six, and this is the day's research.
Jane: 9 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 Characteristic function-based tests for spatial randomness.
The day's research: Tom: Population adjusted indirect comparisons help make sense of evidence when different patient groups are involved.
Jane: But these comparisons are not automatically transportable to the populations that matter most.
Lu: The main problem is that adjusting for differences across trials does not guarantee effects hold for our target group.
Meng: We examine this by looking at estimands to distinguish conditional and marginal treatment effect estimates where transportability fails.
Lalam: Some effects on the linear predictor scale are more transportable while hazard ratios often remain dependent on population even with shared effect modifiers.
Tom: This relates to how we compare data from subgroups, right?
Jane: Consistent Bayesian meta-analysis on subgroup specific effects and interactions shows a way to adjust for variation in relative subgroup contribution when combining results.
Lu: That process helps ensure inferences about interactions remain strong no matter which published subgroup data are used.
Meng: On a computational level, least squares variational inference provides a gradient-free method for finding solutions for exponential families.
Lalam: This method can achieve efficient complexity for Gaussian distributions and performs competitively with other methods even without gradients.
Tom: That is what the paper on Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem addresses.
Jane: And we have the work on Consistent Bayesian meta-analysis on subgroup specific effects and interactions too.
Lu: What about the work on Leave-One-Out Neighborhood Smoothing for Graphons?
Meng: That investigates entrywise uncertainty quantification in graphon models by introducing a leave-one-out neighborhood.
Lalam: The paper on Differentially Private Covariate Balancing Causal Inference deals with covariate balancing using observational data.
Tom: And we have the work on Subgroup comparisons within and across studies in meta-analysis summarizing treatment effects for patient subgroups.
Jane: What about Characteristic function-based tests for spatial randomness?
Lu: That introduces a new type of test for complete spatial randomness that applies to mapped point patterns in a rectangle or more.
Meng: Finally, we have the work on Least squares variational inference which introduces LSVI as a gradient-free scheme for fixed-point recursion in exponential families.
Lucky paper: 2504.07946: Jane: So we are looking at the paper titled Characteristic function-based tests for spatial randomness today.
Tom: It introduces a new test for complete spatial randomness that uses characteristic functions and a weighted L2-distance between empirical and uniform characteristic functions.
Lu: The paper says this test is the first of its kind based on characteristic functions that also uses a weighted L2-distance, scaled by n.
Jane: That distance is defined as Delta equals the integral from D over one over n times sigma j equals one to n exp of i t transpose x j minus phi hat X (t) times w (t) d t.
Meng: The weight function w (t) can be chosen in several ways, like the squared modulus of the characteristic function phi zero (t) squared, or a centered Gaussian weight function.
Tom: The paper specifically explores choices like a Bessel-like weight function, which produces a simple rectangular phi r.
Jane: It also mentions that the expectation of the test statistic under the null hypothesis for that Bessel-like weight is E0 (Delta r) equals one minus E0(phi r (xone minus xtwo)).
Lu: And they show that using the Cauchy weight function gives a specific statistic Delta r equal to one over n summed over j and k from one to n exp of negative one over r summed over d equals one to d x j minus x k d factorial minus two summed over j equals one to n ö D d equals one to r n squared minus e of negative x j d divided by r minus e of negative (one minus x j d) over r plus n over n 2r.
Meng: That Cauchy-weighted CF test is chosen because it has a much lower computational cost than the isotropicedge-corrected L-test.
Tom: The simulation results show that the two-tailed test using either method has a type I error rate very close to the theoretical value of zero point zero five, which is good for controlling false discovery rates.
Jane: They compare its power against tests like Zimmerman’s w2-test and Clark Evans test, and the CF test outperforms them in most simulation settings of heterogeneity and aggregation when the sample size is large.
Lu: Theoretically, the asymptotic distribution of Delta converges in distribution to a limit involving infinity times j equals one to infinity lambda j Z two j where Z j is a sequence of independent standard normal variables and lambda j are the eigenvalues of the linear operator T.
Tom: They also derive efficient algorithms for the asymptotic and small-sample null distributions in Section four which is helpful for practical application.
Jane: The paper concludes that while the exact relationship between weight function choice and test power is still unclear, the empirical CF has a unified formula even for non-rectangular study regions like disks or polygons.
Lu: They also noted that the three-dimensional CF statistic can be used to test for space-time CSR, but testing for space-time interaction is usually more interesting in practice than just testing for CSR.
Meng: So, what would we actually deploy with this test? Would it work well on real traffic data where we have non-rectangular regions?
Tom: It seems the computational cost is low enough that it could be viable for large-scale application when dealing with these spatial point patterns.
Lucky paper: 2501.02703: Tom: Let's talk about Full-conformal novelty detection, this is paper number two of two and it focuses on making novelty detection robust with distribution-free FDR control guarantees.
Jane: The core idea seems to be using all reference data during both the fitting and the inference stages to get a more solid measure of evidence.
Lu: Exactly. The paper introduces K-block full conformal novelty detection, or K-FC ND, which mandates using *all* reference data for those full conformal e-values.
Meng: So how does this handle when we only have limited reference datasets? Is there a trade-off in power versus the guaranteed FDR control level?
Tom: They address that with a data-driven model selection strategy that adapts the scoring function, which should help maintain FDR control even when reference datasets are small.
Lalam: The paper also boosts the detection power by using the conditionally calibrated eBH framework, computing a boost exactly without relying on Monte Carlo samples.
Jane: That sounds like a big win for computational stability and statistical rigor. What about extending this method to handle distribution shift?
Lu: They generalize K-FC ND to accommodate distribution shift, which is when the inliers in your test set have a known shift relative to the reference data.
Tom: That extension uses a weighted variant that keeps the finite-sample FDR control while empirically improving power under those shifting conditions.
Meng: When we look at implementation, what are the specific parameters they suggest for K? Is it always K=one for efficiency?
Lu: They mention canonical instantiations like K=one for computational efficiency and K=m to facilitate that data-driven model selection strategy.
Jane: It sounds like they are trying to make a very practical tool here, moving beyond just theoretical guarantees on the e-values.
Tom: The paper also introduces null proportion correction estimators based on Gao two thousand twenty-three which lets you increase the achieved FDR control level to alpha.
Lalam: That modification allows for an increase in the achieved FDR control level up to alpha, which is useful when you need a higher confidence threshold.
Episode: Applying Non-negative Matrix Factorization with Covariates to the Longitudinal Data as Growth Curve Model
In short: The study extends Non-negative Matrix Factorization (NMF) to longitudinal data by incorporating covariates, establishing a formal link to the Growth Curve Model (GCM). This allows NMF to use an optimized basis that improves GCM fits when covariate effects are non-negative. The work provides methods for estimating covariate effects and compares NMF's soft clustering against PCA.
November 02, 2026
Listen in the app · Audio file · Video file
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Applying Non-negative Matrix Factorization with Covariates to the Longitudinal Data as Growth Curve Model".
Jane: Here's the gist of the paper. Applying Non-negative Matrix Factorization with Covariates to the Longitudinal Data as Growth Curve Model is a study that explores how Non-negative Matrix Factorization (NMF) can be extended to model longitudinal data by incorporating covariates, revealing a formal equivalence between this approach and the Growth Curve Model (GCM).
Tom: This correspondence is significant because it allows NMF to utilize an optimized basis that can improve the fit within the GCM framework when covariate effects are non-negative <ref:2403.05359#pg12> Model Formulation and Tri-Factorization The paper introduces NMF with covariates by expressing individual measurements as a product of a basis matrix and an individual-specific coefficient vector, which is then modeled using known covariate matrices to achieve the GCM mean structure <ref:2403.05359#pg3>.
Jane: The observation vector for individual n is approximated as yn ≈ Xbn, where X is the common basis matrix and bn is the individual-specific coefficient vector <ref:2403.05359#pg5>.
Tom: The relationship between the coefficient matrix B and covariates A and parameter matrix Θ is given by B Q×N = Θ Q×R A R×N, leading to Y ≈ XB = XΘA <ref:2403.05359#pg9>.
Paper summary: Jane: So if we look at the title and what we've discussed about "Applying Non-negative Matrix Factorization with Covariates to the Longitudinal Data as Growth Curve Model," what does that mean in plain language for someone listening who might not be a math major?
Tom: It means they took a technique called NMF, which is good at finding patterns in data by breaking it down into components, and they gave it the ability to handle extra information—covariates—and proved that this setup matches the structure of the Growth Curve Model for longitudinal data > Page one <ref:2403.05359#pg1,of the Growth Curve Model>.
Lu: The real significance here is showing that if you optimize those NMF patterns, you get a basis that actually helps fit a GCM when those covariate effects are non-negative, which is a specific and useful condition > Page two.
Jane: And the main practical contribution they offer to us is the statistical inference for the parameter matrix Theta—that’s how we test if those covariates are actually having an effect on the basis components you found, giving us standard errors and Wald-type tests > Page three.
Tom: The authors also point out that while tri-NMF is technically the most general factorization, fixing the covariate matrix makes NMF much more applicable by connecting it to established models like GCM > Page three.
Meng: For me, if I were deploying this, I’d focus on how stable those multiplicative updates are and how much computational overhead it adds compared to simpler methods > Page eight. That's the real hurdle for productionizing any new AI model > Page eight.
Lalam: If you only take one thing away from this paper, it's that NMF with covariates offers a way to combine regression and soft clustering directly, giving you membership probabilities instead of just signed scores > Page two.
Jane: So, in the end, it’s a formal connection between two modeling frameworks that lets us use the power of NMF's optimized basis within the GCM structure for longitudinal data when we have non-negative covariate effects > Page one <ref:2403.05359#pg1>.
Conclusion: Jane: So we've been looking at how they connect this Non-negative Matrix Factorization method to something called the Growth Curve Model for longitudinal data, and now we’re getting to the final thoughts on this paper, "Applying Non-negative Matrix Factorization with Covariates to the Longitudinal Data as Growth Curve Model."
Tom: Yeah, it basically boils down to showing that you can use NMF patterns—those non-negative basis vectors—and then plug them right into a GCM structure when the covariate effects are positive. It’s about formally proving that these two modeling approaches are equivalent in terms of how they set up the data structure.
Lu: The core contribution is that this optimized NMF basis actually improves the fit within the GCM framework, specifically because it handles non-negative covariate effects better than just using a standard GCM setup. It’s about getting a better representation of how those individual measurements change over time.
Meng: For us in production, what matters is that they show how to build that initial structure from the individual data up and then incorporate the covariates through this parameter matrix Theta and the covariate matrix A. We need to see if that process is stable when we try to run it on real, messy datasets.
Jane: And they spend a good chunk of the paper developing a way to get statistical inference for that parameter matrix Theta based on the optimized basis X-hat. That’s where they give us those standard errors and Wald-type tests for those covariate effects.
Tom: That inference part is crucial because it lets us actually test if our assumptions about how those covariates influence the growth trajectory are holding up, not just assume they are working. It moves it from a purely descriptive tool to something we can use for hypothesis testing.
Lu: The paper suggests that while tri-NMF is technically the most flexible factorization, fixing that covariate matrix makes NMF much more useful by tying it directly to established models like the GCM. It broadens its applicability significantly.
Jane: So, if you're listening and only get one thing from this whole study, it’s that this framework lets you use the strength of non-negative patterns from NMF within a familiar statistical model like the Growth Curve Model when those covariate effects are non-negative.
Tom: It’s a solid piece of work because they don't just propose an idea; they show the math behind how to optimize those matrices and then build reliable tests on the results.
Meng: I'm still thinking about how much computational cost that optimization adds when we scale it up, but finding a way to keep those updates numerically stable is a big win for deployment.
Lu: The multiplicative update formulas they derived are pretty clean, which helps with that stability issue you mentioned, Tom.
Jane: This whole thing sets the stage nicely for us to look at how this NMF basis compares against other techniques like PCA or Functional PCA in the next segment.
Episode: The day in conversation for 2026-11-01
November 01, 2026
Listen in the app · Audio file
The discussion: Tom: I think the core issue with that HTA paper is how strongly the estimand dictates whether we can actually move an effect across different patient populations without losing validity. Jane: Exactly. If marginal effects are inherently population-dependent unless you satisfy some strict structural conditions, then relying on pooled estimates becomes a very risky assumption for decision-making. Lu: That reminds me of the earlier discussion we had about heterogeneity in treatment effect estimation; I wonder if this paper provides a more rigorous framework than what we’ve used when combining those subgroups. Meng: But are we really suggesting that *all* indirect comparisons fail transportability if the conditions aren't met, or is it just that the sensitivity to population structure changes? Lalam: It seems like the specific nature of the effect—is it additive or multiplicative—is a huge factor determining if those conditions are even remotely plausible. Tom: Speaking of combining data, I found the Bayesian meta-analysis paper interesting because it specifically tackles how to recover interaction inferences when we only have published subgroup data that might not perfectly align with our current modeling assumptions. Jane: That adjustment for variation is key, but does that method introduce its own set of biases when estimating the baseline effect before accounting for the subgroup contribution? Lu: I’m concerned about the computational cost of running those complex adjustments across dozens of trials; how scalable is this approach for a larger evidence base? Meng: Maybe the privacy paper offers a practical angle, though. Differentially Private Covariate Balancing C
Episode: Daily Summary for 2026-11-01
November 01, 2026
Episode: Least squares variational inference
In short: Least Squares Variational Inference (LSVI) is a gradient-free method for fixed-point recursion in exponential families. It performs ordinary least squares regression on tempered log-target evaluations under a variational approximation. This scheme avoids inverting the Fisher information matrix for Gaussian distributions, offering efficient complexity and competitive performance compared to standard methods.
November 01, 2026
Episode: Characteristic function-based tests for spatial randomness
In short: A new test for complete spatial randomness is introduced using characteristic functions and a weighted L2-distance between empirical and uniform characteristic functions. This test connects to Ripley's K-function and Zimmerman's w2 statistic, offering a simple, edge-effect-free method that performs well in simulations across various spatial patterns.
November 01, 2026
Episode: Subgroup comparisons within and across studies in meta-analysis
In short: The study investigates methods for accurately estimating subgroup treatment effects and interactions in meta-analysis, addressing inconsistencies caused by unbalanced subgroup distributions across trials. The authors propose using SWADA weights as a practical default strategy to resolve these issues, leading to better and more robust estimates.
November 01, 2026
Episode: Consistent Bayesian meta-analysis on subgroup specific effects and interactions
In short: The CAMS method adjusts meta-analyses of subgroup effects by incorporating study-specific contribution (information fraction) to ensure consistent interaction estimates. This addresses inconsistencies arising from varying subgroup weighting in trials, providing reliable interaction decisions when published subgroup data are available.
November 01, 2026
Episode: Leave-One-Out Neighborhood Smoothing for Graphons: Berry--Esseen Bounds and Edge-Wise Confidence Intervals
In short: This work introduces a leave-one-out neighborhood smoother for graphon models to quantify uncertainty for every individual edge probability Pij. It achieves finite-sample concentration bounds and Berry–Esseen normal approximations for each pair, providing rigorous, entrywise confidence intervals based on local network structure.
November 01, 2026
Episode: Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem: An Estimand-Based Perspective and Implications for Health Technology Assessment
In short: The paper reframes Population-Adjusted Indirect Comparisons (PAICs) as a transportability problem: can an effect from one patient group be validly applied to another? It distinguishes between within-population properties and cross-population validity, establishing strict criteria for when PAIC estimates are directly transportable and when they must be treated as population-specific.
November 01, 2026
Episode: Differentially Private Covariate Balancing Causal Inference
In short: The paper introduces a method to perform causal inference from observational data while protecting privacy. It uses a differentially private two-stage weighting estimator based on covariate-balancing rules to ensure good balance between treatment groups, even when releasing statistics compromises privacy. This allows for reliable causal effect estimation in sensitive settings.
November 01, 2026
Episode: The day in conversation for 2026-10-31
November 01, 2026
Episode: Daily Summary for 2026-10-31
October 31, 2026
Episode: Semiparametric Inference for Functional Difference-in-Differences
October 31, 2026
Episode: Critique of "Use of roster charts in the investigation and prosecution of nurses..." by John O' Quigley
In short: The paper critiques John O'Quigley's statistical analysis of a nurse roster chart used to investigate misconduct. The authors demonstrate serious statistical errors, arguing that O'Quigley’s core assumptions about nurse independence are fundamentally flawed. They conclude the roster chart is scientifically fake evidence and cannot support the hypothesis that a specific nurse is responsible for deaths.
October 31, 2026
Episode: Global synchronization for multi-source data integration under blockwise missing patterns
In short: This work introduces Global Synchronized Multiple Matrix Integration (GSMMI) to solve multi-source data integration problems where data blocks are missing across different sources. GSMMI formulates alignment as a global synchronization problem, aligning all overlapping submatrices simultaneously using all pairwise overlaps. This method outperforms existing methods like CMMI by exploiting richer overlap information and achieving better alignment accuracy.
October 31, 2026
Episode: Sensitivity Analysis for Trials with Irregular and Informative Assessment Times: Marginal Mean Models and Sensitivity Parameter Selection
October 31, 2026
Episode: Leveraging Large Language Models to Improve Precision in Randomized Controlled Trials
In short: The study investigated using Large Language Models (LLMs) to safely improve the precision of Randomized Controlled Trials (RCTs). The researchers developed a method to use LLM predictions as covariates, primarily by transforming non-standard data like text into numerical representations. Results showed that this approach can significantly boost precision, especially when standard covariates are lacking or when using text data as input, though the benefit varies by case study.
October 31, 2026
Episode: A Unified Three-Stage Weighting Framework for Causal Inference and Mediation Analysis under Case-Control Sampling
In short: The framework estimates total and pathway-specific causal effects from case-control studies without needing to know population outcome prevalence. It achieves this through a three-stage process: recovering prevalence, reconstructing a pseudo-population using design weights, and applying stabilized causal weights. This method provides robust mediation analysis under sampling bias.
October 31, 2026
Episode: Simulation-based Criticism of Complex Biological Models with Neural Evidence Estimation
In short: The Neural Evidence Estimation (NEE) method uses normalizing flows to learn a complex model's marginal likelihood directly from simulations, bypassing noisy Monte Carlo methods. This allows AI to reliably select the best biological model among several possibilities and flag models that are poorly specified, offering a rigorous way to test scientific assumptions.
October 31, 2026
Episode: Beta Regression with Autoregressive Errors for Interrupted Time Series Analysis of Proportion and Rate Outcomes: A Simulation Study
In short: This study introduced 'betark,' a new joint maximum likelihood estimator for beta regression with autoregressive errors, to analyze proportion and rate outcomes in interrupted time series. It compared betark against the standard quasi-binomial GLM plus Newey–West HAC errors. Betark provided better-calibrated inference than the comparator across most conditions, especially when high autocorrelation was present.
October 31, 2026
Episode: Towards more plausible point-identifying assumptions in two-sample Mendelian randomization
In short: Two-sample Mendelian randomization methods are reinterpreted as estimators of average causal effects under weaker identification assumptions than previously thought. The authors derive conditions, like Non Simultaneous Heterogeneity (NOSH), that imply summary-level homogeneity, clarifying the point-identifying assumptions required for these methods and improving their practical interpretation.
October 31, 2026
Episode: Reconstructing constant point heat sources from few temperature sensors: sparse selection and spatial resolution
In short: The framework reconstructs unknown constant heat sources from limited temperature sensor data by combining sparse dictionary selection with continuous constrained fitting. It uses methods like Orthogonal Matching Pursuit to select candidate source locations and then employs a bounded trust-region least-squares solver to recover the number and strengths of the sources, providing a local interpretation of which source information the observations support.
October 31, 2026
Episode: Daily Summary for 2026-10-27
October 29, 2026
Episode: Daily Summary for 2026-10-16
October 29, 2026
Episode: Daily Summary for 2026-10-17
October 29, 2026
Episode: Daily Summary for 2026-10-18
October 29, 2026
Episode: Daily Summary for 2026-10-19
October 29, 2026
Episode: Daily Summary for 2026-10-20
October 29, 2026
Episode: Daily Summary for 2026-10-21
October 29, 2026
Episode: Daily Summary for 2026-10-23
October 29, 2026
Episode: Daily Summary for 2026-10-27
October 29, 2026
Episode: The day in conversation for 2026-10-14
October 28, 2026
Episode: The day in conversation for 2026-10-16
October 28, 2026
Episode: The day in conversation for 2026-10-17
October 28, 2026
Episode: The day in conversation for 2026-10-18
October 28, 2026
Episode: The day in conversation for 2026-10-19
October 28, 2026
Episode: The day in conversation for 2026-10-20
October 28, 2026
Episode: The day in conversation for 2026-10-21
October 28, 2026
Episode: The day in conversation for 2026-10-23
October 28, 2026
Episode: The day in conversation for 2026-10-27
October 28, 2026